Sync from upstream llama.cpp repository

tools/CMakeLists.txt

@@ -0,0 +1,40 @@
+# dependencies
+
+find_package(Threads REQUIRED)
+
+# third-party
+
+# ...
+
+# flags
+
+llama_add_compile_flags()
+
+# tools
+
+if (EMSCRIPTEN)
+else()
+    add_subdirectory(batched-bench)
+    add_subdirectory(gguf-split)
+    add_subdirectory(imatrix)
+    add_subdirectory(llama-bench)
+    add_subdirectory(completion)
+    add_subdirectory(perplexity)
+    add_subdirectory(quantize)
+    if (LLAMA_BUILD_SERVER)
+        add_subdirectory(cli)
+        add_subdirectory(server)
+    endif()
+    add_subdirectory(tokenize)
+    add_subdirectory(tts)
+    add_subdirectory(mtmd)
+    if (GGML_RPC)
+        add_subdirectory(rpc)
+    endif()
+    if (NOT GGML_BACKEND_DL)
+        # these examples use the backends directly and cannot be built with dynamic loading
+        add_subdirectory(cvector-generator)
+        add_subdirectory(export-lora)
+    endif()
+    add_subdirectory(fit-params)
+endif()

tools/batched-bench/CMakeLists.txt

@@ -0,0 +1,8 @@
+set(TARGET llama-batched-bench)
+add_executable(${TARGET} batched-bench.cpp)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_17)
+
+if(LLAMA_TOOLS_INSTALL)
+    install(TARGETS ${TARGET} RUNTIME)
+endif()

tools/batched-bench/README.md

@@ -0,0 +1,60 @@
+# llama.cpp/example/batched-bench
+
+Benchmark the batched decoding performance of `llama.cpp`
+
+## Usage
+
+There are 2 modes of operation:
+
+- `prompt not shared` - each batch has a separate prompt of size `PP` (i.e. `N_KV = B*(PP + TG)`)
+- `prompt is shared` - there is a common prompt of size `PP` used by all batches (i.e. `N_KV = PP + B*TG`)
+
+```bash
+./llama-batched-bench -m model.gguf -c 2048 -b 2048 -ub 512 -npp 128,256,512 -ntg 128,256 -npl 1,2,4,8,16,32 [-pps]
+
+# LLaMA 7B, F16, N_KV_MAX = 16384 (8GB), prompt not shared
+./llama-batched-bench -m ./models/llama-7b/ggml-model-f16.gguf -c 16384 -b 2048 -ub 512 -ngl 99
+
+# LLaMA 7B, Q8_0, N_KV_MAX = 16384 (8GB), prompt is shared
+./llama-batched-bench -m ./models/llama-7b/ggml-model-q8_0.gguf -c 16384 -b 2048 -ub 512 -ngl 99 -pps
+
+# custom set of batches
+./llama-batched-bench -m ./models/llama-7b/ggml-model-q8_0.gguf -c 2048 -b 512 -ub 512 -ngl 999 -npp 128,256,512 -ntg 128,256 -npl 1,2,4,8,16,32
+```
+
+## Sample results
+
+- `PP` - prompt tokens per batch
+- `TG` - generated tokens per batch
+- `B` - number of batches
+- `N_KV` - required KV cache size
+- `T_PP` - prompt processing time (i.e. time to first token)
+- `S_PP` - prompt processing speed (`(B*PP)/T_PP` or `PP/T_PP`)
+- `T_TG` - time to generate all batches
+- `S_TG` - text generation speed (`(B*TG)/T_TG`)
+- `T` - total time
+- `S` - total speed (i.e. all tokens / total time)
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   128 |    128 |    1 |    256 |    0.108 |  1186.64 |    3.079 |    41.57 |    3.187 |    80.32 |
+|   128 |    128 |    2 |    512 |    0.198 |  1295.19 |    5.029 |    50.90 |    5.227 |    97.95 |
+|   128 |    128 |    4 |   1024 |    0.373 |  1373.96 |    6.878 |    74.44 |    7.251 |   141.23 |
+|   128 |    128 |    8 |   2048 |    0.751 |  1363.27 |    7.344 |   139.43 |    8.095 |   252.99 |
+|   128 |    128 |   16 |   4096 |    1.570 |  1304.68 |    8.455 |   242.23 |   10.024 |   408.60 |
+|   128 |    128 |   32 |   8192 |    3.408 |  1201.73 |    8.801 |   465.40 |   12.209 |   670.96 |
+|   128 |    256 |    1 |    384 |    0.107 |  1196.70 |    6.329 |    40.45 |    6.436 |    59.67 |
+|   128 |    256 |    2 |    768 |    0.194 |  1317.45 |   10.239 |    50.00 |   10.433 |    73.61 |
+|   128 |    256 |    4 |   1536 |    0.366 |  1399.03 |   13.960 |    73.35 |   14.326 |   107.22 |
+|   128 |    256 |    8 |   3072 |    0.751 |  1363.92 |   15.110 |   135.54 |   15.861 |   193.69 |
+|   128 |    256 |   16 |   6144 |    1.569 |  1304.93 |   18.073 |   226.64 |   19.642 |   312.80 |
+|   128 |    256 |   32 |  12288 |    3.409 |  1201.35 |   19.223 |   426.15 |   22.633 |   542.93 |
+
+### JSONL output
+
+Pass `--output-format jsonl` to output JSONL instead of Markdown, á la
+
+```json lines
+{"n_kv_max": 2048, "n_batch": 2048, "n_ubatch": 512, "flash_attn": 0, "is_pp_shared": 0, "n_gpu_layers": 99, "n_threads": 8, "n_threads_batch": 8, "pp": 128, "tg": 128, "pl": 1, "n_kv": 256, "t_pp": 0.233810, "speed_pp": 547.453064, "t_tg": 3.503684, "speed_tg": 36.532974, "t": 3.737494, "speed": 68.495094}
+{"n_kv_max": 2048, "n_batch": 2048, "n_ubatch": 512, "flash_attn": 0, "is_pp_shared": 0, "n_gpu_layers": 99, "n_threads": 8, "n_threads_batch": 8, "pp": 128, "tg": 128, "pl": 2, "n_kv": 512, "t_pp": 0.422602, "speed_pp": 605.770935, "t_tg": 11.106112, "speed_tg": 23.050371, "t": 11.528713, "speed": 44.410854}
+```

tools/batched-bench/batched-bench.cpp

@@ -0,0 +1,256 @@
+#include "arg.h"
+#include "common.h"
+#include "log.h"
+#include "llama.h"
+
+#include <algorithm>
+#include <cstdio>
+#include <string>
+#include <vector>
+
+static void print_usage(int, char ** argv) {
+    LOG("\nexample usage:\n");
+    LOG("\n    %s -m model.gguf -c 2048 -b 2048 -ub 512 -npp 128,256,512 -ntg 128,256 -npl 1,2,4,8,16,32 [-pps]\n", argv[0]);
+    LOG("\n");
+}
+
+int main(int argc, char ** argv) {
+    common_params params;
+
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_BENCH, print_usage)) {
+        return 1;
+    }
+
+    common_init();
+
+    int is_pp_shared   = params.is_pp_shared;
+    int is_tg_separate = params.is_tg_separate;
+
+    std::vector<int> n_pp = params.n_pp;
+    std::vector<int> n_tg = params.n_tg;
+    std::vector<int> n_pl = params.n_pl;
+
+    // init LLM
+
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    // initialize the model
+
+    llama_model_params model_params = common_model_params_to_llama(params);
+
+    llama_model * model = llama_model_load_from_file(params.model.path.c_str(), model_params);
+
+    if (model == NULL) {
+        fprintf(stderr , "%s: error: unable to load model\n" , __func__);
+        return 1;
+    }
+
+    llama_context_params ctx_params = common_context_params_to_llama(params);
+
+    // ensure enough sequences are available
+    ctx_params.n_seq_max = n_pl.empty() ? 1 : *std::max_element(n_pl.begin(), n_pl.end());
+
+    llama_context * ctx = llama_init_from_model(model, ctx_params);
+
+    if (ctx == NULL) {
+        fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
+        llama_model_free(model);
+        return 1;
+    }
+
+    const llama_vocab * vocab   = llama_model_get_vocab(model);
+    const int32_t       n_vocab = llama_vocab_n_tokens(vocab);
+
+    const auto get_token_rand = [n_vocab]() -> llama_token {
+        return std::rand() % n_vocab;
+    };
+
+    auto * mem = llama_get_memory(ctx);
+
+    const int32_t n_kv_max = llama_n_ctx(ctx);
+
+    llama_batch batch = llama_batch_init(n_kv_max, 0, 1);
+
+    // decode in batches of ctx_params.n_batch tokens
+    auto decode_helper = [](llama_context * ctx, llama_batch & batch, int32_t n_batch, bool synchronize) {
+        for (int32_t i = 0; i < batch.n_tokens; i += n_batch) {
+            const int32_t n_tokens = std::min(n_batch, batch.n_tokens - i);
+
+            llama_batch batch_view = {
+                n_tokens,
+                batch.token    + i,
+                nullptr,
+                batch.pos      + i,
+                batch.n_seq_id + i,
+                batch.seq_id   + i,
+                batch.logits   + i,
+            };
+
+            const int ret = llama_decode(ctx, batch_view);
+            if (ret != 0) {
+                LOG_ERR("failed to decode the batch, n_batch = %d, ret = %d\n", n_batch, ret);
+                return false;
+            }
+
+            if (synchronize) {
+                llama_synchronize(ctx);
+            }
+        }
+
+        return true;
+    };
+
+    // warm up
+    {
+        for (int i = 0; i < 16; ++i) {
+            common_batch_add(batch, get_token_rand(), i, { 0 }, false);
+        }
+
+        if (!decode_helper(ctx, batch, ctx_params.n_batch, true)) {
+            LOG_ERR("%s: llama_decode() failed\n", __func__);
+            llama_free(ctx);
+            llama_model_free(model);
+            return 1;
+        }
+    }
+
+    if (!params.batched_bench_output_jsonl) {
+        LOG("\n");
+        LOG("%s: n_kv_max = %d, n_batch = %d, n_ubatch = %d, flash_attn = %d, is_pp_shared = %d, is_tg_separate = %d, n_gpu_layers = %d, n_threads = %u, n_threads_batch = %u\n", __func__, n_kv_max, params.n_batch, params.n_ubatch, int(params.flash_attn_type), is_pp_shared, is_tg_separate, params.n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch);
+        LOG("\n");
+        LOG("|%6s | %6s | %4s | %6s | %8s | %8s | %8s | %8s | %8s | %8s |\n", "PP", "TG", "B", "N_KV", "T_PP s", "S_PP t/s", "T_TG s", "S_TG t/s", "T s", "S t/s");
+        LOG("|%6s-|-%6s-|-%4s-|-%6s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|\n", "------", "------", "----", "------", "--------", "--------", "--------", "--------", "--------", "--------");
+    }
+
+    for (        int i_pp = 0; i_pp < (int) n_pp.size(); ++i_pp) {
+        for (    int i_tg = 0; i_tg < (int) n_tg.size(); ++i_tg) {
+            for (int i_pl = 0; i_pl < (int) n_pl.size(); ++i_pl) {
+                const int pp = n_pp[i_pp];
+                const int tg = n_tg[i_tg];
+                const int pl = n_pl[i_pl];
+
+                const int n_ctx_req = is_pp_shared ? (params.kv_unified ? pp : pl*pp) + pl*tg : pl*(pp + tg);
+
+                if (n_ctx_req > n_kv_max) {
+                    continue;
+                }
+
+                common_batch_clear(batch);
+
+                for (int j = 0; j < (is_pp_shared ? 1 : pl); ++j) {
+                    for (int i = 0; i < pp; ++i) {
+                        common_batch_add(batch, get_token_rand(), i, { j }, i == pp - 1);
+                    }
+                }
+
+                llama_memory_clear(mem, false);
+
+                const auto t_pp_start = ggml_time_us();
+
+                if (!decode_helper(ctx, batch, ctx_params.n_batch, false)) {
+                    LOG_ERR("%s: llama_decode() failed\n", __func__);
+                    llama_free(ctx);
+                    llama_model_free(model);
+                    return 1;
+                }
+
+                llama_synchronize(ctx);
+
+                const auto t_pp_end = ggml_time_us();
+
+                if (is_pp_shared) {
+                    for (int32_t i = 1; i < pl; ++i) {
+                        llama_memory_seq_cp(mem, 0, i, -1, -1);
+                    }
+
+                    if (!params.kv_unified) {
+                        // run one dummy token to apply the memory copy
+                        common_batch_clear(batch);
+                        common_batch_add(batch, get_token_rand(), pp + 0, { 0 }, true);
+                        if (!decode_helper(ctx, batch, ctx_params.n_batch, true)) {
+                            LOG_ERR("%s: llama_decode() failed\n", __func__);
+                            llama_free(ctx);
+                            llama_model_free(model);
+                            return 1;
+                        }
+                        llama_memory_seq_rm(mem, 0, pp, -1);
+                    }
+                }
+
+                const auto t_tg_start = ggml_time_us();
+
+                if (is_tg_separate) {
+                    // decode pattern:
+                    // 0 0 0 ... 1 1 1 ... 2 2 2 ... 3 3 3 ...
+                    for (int j = 0; j < pl; ++j) {
+                        for (int i = 0; i < tg; ++i) {
+                            common_batch_clear(batch);
+
+                            common_batch_add(batch, get_token_rand(), pp + i, { j }, true);
+
+                            if (!decode_helper(ctx, batch, ctx_params.n_batch, true)) {
+                                LOG_ERR("%s: llama_decode() failed\n", __func__);
+                                llama_free(ctx);
+                                llama_model_free(model);
+                                return 1;
+                            }
+                        }
+                    }
+                } else {
+                    // decode pattern:
+                    // 0123 0123 0123 ...
+                    for (int i = 0; i < tg; ++i) {
+                        common_batch_clear(batch);
+
+                        for (int j = 0; j < pl; ++j) {
+                            common_batch_add(batch, get_token_rand(), pp + i, { j }, true);
+                        }
+
+                        if (!decode_helper(ctx, batch, ctx_params.n_batch, true)) {
+                            LOG_ERR("%s: llama_decode() failed\n", __func__);
+                            llama_free(ctx);
+                            llama_model_free(model);
+                            return 1;
+                        }
+                    }
+                }
+
+                const auto t_tg_end = ggml_time_us();
+
+                const int32_t n_kv = n_ctx_req;
+
+                const float t_pp = (t_pp_end - t_pp_start) / 1000000.0f;
+                const float t_tg = (t_tg_end - t_tg_start) / 1000000.0f;
+                const float t    = t_pp + t_tg;
+
+                const float speed_pp = is_pp_shared ? pp / t_pp : pl*pp / t_pp;
+                const float speed_tg = pl*tg / t_tg;
+                const float speed    = ((is_pp_shared ? pp : pl*pp) + pl*tg) / t;
+
+                if(params.batched_bench_output_jsonl) {
+                    LOG(
+                        "{\"n_kv_max\": %d, \"n_batch\": %d, \"n_ubatch\": %d, \"flash_attn\": %d, \"is_pp_shared\": %d, \"n_gpu_layers\": %d, \"n_threads\": %u, \"n_threads_batch\": %u, "
+                        "\"pp\": %d, \"tg\": %d, \"pl\": %d, \"n_kv\": %d, \"t_pp\": %f, \"speed_pp\": %f, \"t_tg\": %f, \"speed_tg\": %f, \"t\": %f, \"speed\": %f}\n",
+                        n_kv_max, params.n_batch, params.n_ubatch, int(params.flash_attn_type), params.is_pp_shared, params.n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch,
+                        pp, tg, pl, n_kv, t_pp, speed_pp, t_tg, speed_tg, t, speed
+                    );
+                } else {
+                    LOG("|%6d | %6d | %4d | %6d | %8.3f | %8.2f | %8.3f | %8.2f | %8.3f | %8.2f |\n", pp, tg, pl, n_kv, t_pp, speed_pp, t_tg, speed_tg, t, speed);
+                }
+            }
+        }
+    }
+
+    LOG("\n");
+    llama_perf_context_print(ctx);
+
+    llama_batch_free(batch);
+
+    llama_free(ctx);
+    llama_model_free(model);
+
+    llama_backend_free();
+
+    return 0;
+}

tools/cli/CMakeLists.txt

@@ -0,0 +1,10 @@
+set(TARGET llama-cli)
+add_executable(${TARGET} cli.cpp)
+target_link_libraries(${TARGET} PRIVATE server-context PUBLIC common ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_17)
+
+include_directories(../server)
+
+if(LLAMA_TOOLS_INSTALL)
+    install(TARGETS ${TARGET} RUNTIME)
+endif()

tools/cli/README.md

@@ -0,0 +1,192 @@
+# llama.cpp/tools/cli
+
+## Usage
+
+<!-- HELP_START -->
+
+<!-- IMPORTANT: The list below is auto-generated by llama-gen-docs; do NOT modify it manually -->
+
+### Common params
+
+| Argument | Explanation |
+| -------- | ----------- |
+| `-h, --help, --usage` | print usage and exit |
+| `--version` | show version and build info |
+| `--license` | show source code license and dependencies |
+| `-cl, --cache-list` | show list of models in cache |
+| `--completion-bash` | print source-able bash completion script for llama.cpp |
+| `--verbose-prompt` | print a verbose prompt before generation (default: false) |
+| `-t, --threads N` | number of CPU threads to use during generation (default: -1)<br/>(env: LLAMA_ARG_THREADS) |
+| `-tb, --threads-batch N` | number of threads to use during batch and prompt processing (default: same as --threads) |
+| `-C, --cpu-mask M` | CPU affinity mask: arbitrarily long hex. Complements cpu-range (default: "") |
+| `-Cr, --cpu-range lo-hi` | range of CPUs for affinity. Complements --cpu-mask |
+| `--cpu-strict <0\|1>` | use strict CPU placement (default: 0) |
+| `--prio N` | set process/thread priority : low(-1), normal(0), medium(1), high(2), realtime(3) (default: 0) |
+| `--poll <0...100>` | use polling level to wait for work (0 - no polling, default: 50) |
+| `-Cb, --cpu-mask-batch M` | CPU affinity mask: arbitrarily long hex. Complements cpu-range-batch (default: same as --cpu-mask) |
+| `-Crb, --cpu-range-batch lo-hi` | ranges of CPUs for affinity. Complements --cpu-mask-batch |
+| `--cpu-strict-batch <0\|1>` | use strict CPU placement (default: same as --cpu-strict) |
+| `--prio-batch N` | set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: 0) |
+| `--poll-batch <0\|1>` | use polling to wait for work (default: same as --poll) |
+| `-c, --ctx-size N` | size of the prompt context (default: 0, 0 = loaded from model)<br/>(env: LLAMA_ARG_CTX_SIZE) |
+| `-n, --predict, --n-predict N` | number of tokens to predict (default: -1, -1 = infinity)<br/>(env: LLAMA_ARG_N_PREDICT) |
+| `-b, --batch-size N` | logical maximum batch size (default: 2048)<br/>(env: LLAMA_ARG_BATCH) |
+| `-ub, --ubatch-size N` | physical maximum batch size (default: 512)<br/>(env: LLAMA_ARG_UBATCH) |
+| `--keep N` | number of tokens to keep from the initial prompt (default: 0, -1 = all) |
+| `--swa-full` | use full-size SWA cache (default: false)<br/>[(more info)](https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)<br/>(env: LLAMA_ARG_SWA_FULL) |
+| `-fa, --flash-attn [on\|off\|auto]` | set Flash Attention use ('on', 'off', or 'auto', default: 'auto')<br/>(env: LLAMA_ARG_FLASH_ATTN) |
+| `-p, --prompt PROMPT` | prompt to start generation with; for system message, use -sys |
+| `--perf, --no-perf` | whether to enable internal libllama performance timings (default: false)<br/>(env: LLAMA_ARG_PERF) |
+| `-f, --file FNAME` | a file containing the prompt (default: none) |
+| `-bf, --binary-file FNAME` | binary file containing the prompt (default: none) |
+| `-e, --escape, --no-escape` | whether to process escapes sequences (\n, \r, \t, \', \", \\) (default: true) |
+| `--rope-scaling {none,linear,yarn}` | RoPE frequency scaling method, defaults to linear unless specified by the model<br/>(env: LLAMA_ARG_ROPE_SCALING_TYPE) |
+| `--rope-scale N` | RoPE context scaling factor, expands context by a factor of N<br/>(env: LLAMA_ARG_ROPE_SCALE) |
+| `--rope-freq-base N` | RoPE base frequency, used by NTK-aware scaling (default: loaded from model)<br/>(env: LLAMA_ARG_ROPE_FREQ_BASE) |
+| `--rope-freq-scale N` | RoPE frequency scaling factor, expands context by a factor of 1/N<br/>(env: LLAMA_ARG_ROPE_FREQ_SCALE) |
+| `--yarn-orig-ctx N` | YaRN: original context size of model (default: 0 = model training context size)<br/>(env: LLAMA_ARG_YARN_ORIG_CTX) |
+| `--yarn-ext-factor N` | YaRN: extrapolation mix factor (default: -1.0, 0.0 = full interpolation)<br/>(env: LLAMA_ARG_YARN_EXT_FACTOR) |
+| `--yarn-attn-factor N` | YaRN: scale sqrt(t) or attention magnitude (default: -1.0)<br/>(env: LLAMA_ARG_YARN_ATTN_FACTOR) |
+| `--yarn-beta-slow N` | YaRN: high correction dim or alpha (default: -1.0)<br/>(env: LLAMA_ARG_YARN_BETA_SLOW) |
+| `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: -1.0)<br/>(env: LLAMA_ARG_YARN_BETA_FAST) |
+| `-kvo, --kv-offload, -nkvo, --no-kv-offload` | whether to enable KV cache offloading (default: enabled)<br/>(env: LLAMA_ARG_KV_OFFLOAD) |
+| `--repack, -nr, --no-repack` | whether to enable weight repacking (default: enabled)<br/>(env: LLAMA_ARG_REPACK) |
+| `--no-host` | bypass host buffer allowing extra buffers to be used<br/>(env: LLAMA_ARG_NO_HOST) |
+| `-ctk, --cache-type-k TYPE` | KV cache data type for K<br/>allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1<br/>(default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_K) |
+| `-ctv, --cache-type-v TYPE` | KV cache data type for V<br/>allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1<br/>(default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_V) |
+| `-dt, --defrag-thold N` | KV cache defragmentation threshold (DEPRECATED)<br/>(env: LLAMA_ARG_DEFRAG_THOLD) |
+| `-np, --parallel N` | number of parallel sequences to decode (default: 1)<br/>(env: LLAMA_ARG_N_PARALLEL) |
+| `--mlock` | force system to keep model in RAM rather than swapping or compressing<br/>(env: LLAMA_ARG_MLOCK) |
+| `--mmap, --no-mmap` | whether to memory-map model. Explicitly enabling mmap disables direct-io. (if mmap disabled, slower load but may reduce pageouts if not using mlock) (default: enabled)<br/>(env: LLAMA_ARG_MMAP) |
+| `-dio, --direct-io, -ndio, --no-direct-io` | use DirectIO if available. Takes precedence over --mmap (default: enabled)<br/>(env: LLAMA_ARG_DIO) |
+| `--numa TYPE` | attempt optimizations that help on some NUMA systems<br/>- distribute: spread execution evenly over all nodes<br/>- isolate: only spawn threads on CPUs on the node that execution started on<br/>- numactl: use the CPU map provided by numactl<br/>if run without this previously, it is recommended to drop the system page cache before using this<br/>see https://github.com/ggml-org/llama.cpp/issues/1437<br/>(env: LLAMA_ARG_NUMA) |
+| `-dev, --device <dev1,dev2,..>` | comma-separated list of devices to use for offloading (none = don't offload)<br/>use --list-devices to see a list of available devices<br/>(env: LLAMA_ARG_DEVICE) |
+| `--list-devices` | print list of available devices and exit |
+| `-ot, --override-tensor <tensor name pattern>=<buffer type>,...` | override tensor buffer type<br/>(env: LLAMA_ARG_OVERRIDE_TENSOR) |
+| `-cmoe, --cpu-moe` | keep all Mixture of Experts (MoE) weights in the CPU<br/>(env: LLAMA_ARG_CPU_MOE) |
+| `-ncmoe, --n-cpu-moe N` | keep the Mixture of Experts (MoE) weights of the first N layers in the CPU<br/>(env: LLAMA_ARG_N_CPU_MOE) |
+| `-ngl, --gpu-layers, --n-gpu-layers N` | max. number of layers to store in VRAM, either an exact number, 'auto', or 'all' (default: auto)<br/>(env: LLAMA_ARG_N_GPU_LAYERS) |
+| `-sm, --split-mode {none,layer,row}` | how to split the model across multiple GPUs, one of:<br/>- none: use one GPU only<br/>- layer (default): split layers and KV across GPUs<br/>- row: split rows across GPUs<br/>(env: LLAMA_ARG_SPLIT_MODE) |
+| `-ts, --tensor-split N0,N1,N2,...` | fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1<br/>(env: LLAMA_ARG_TENSOR_SPLIT) |
+| `-mg, --main-gpu INDEX` | the GPU to use for the model (with split-mode = none), or for intermediate results and KV (with split-mode = row) (default: 0)<br/>(env: LLAMA_ARG_MAIN_GPU) |
+| `-fit, --fit [on\|off]` | whether to adjust unset arguments to fit in device memory ('on' or 'off', default: 'on')<br/>(env: LLAMA_ARG_FIT) |
+| `-fitt, --fit-target MiB0,MiB1,MiB2,...` | target margin per device for --fit, comma-separated list of values, single value is broadcast across all devices, default: 1024<br/>(env: LLAMA_ARG_FIT_TARGET) |
+| `-fitc, --fit-ctx N` | minimum ctx size that can be set by --fit option, default: 4096<br/>(env: LLAMA_ARG_FIT_CTX) |
+| `--check-tensors` | check model tensor data for invalid values (default: false) |
+| `--override-kv KEY=TYPE:VALUE,...` | advanced option to override model metadata by key. to specify multiple overrides, either use comma-separated values.<br/>types: int, float, bool, str. example: --override-kv tokenizer.ggml.add_bos_token=bool:false,tokenizer.ggml.add_eos_token=bool:false |
+| `--op-offload, --no-op-offload` | whether to offload host tensor operations to device (default: true) |
+| `--lora FNAME` | path to LoRA adapter (use comma-separated values to load multiple adapters) |
+| `--lora-scaled FNAME:SCALE,...` | path to LoRA adapter with user defined scaling (format: FNAME:SCALE,...)<br/>note: use comma-separated values |
+| `--control-vector FNAME` | add a control vector<br/>note: use comma-separated values to add multiple control vectors |
+| `--control-vector-scaled FNAME:SCALE,...` | add a control vector with user defined scaling SCALE<br/>note: use comma-separated values (format: FNAME:SCALE,...) |
+| `--control-vector-layer-range START END` | layer range to apply the control vector(s) to, start and end inclusive |
+| `-m, --model FNAME` | model path to load<br/>(env: LLAMA_ARG_MODEL) |
+| `-mu, --model-url MODEL_URL` | model download url (default: unused)<br/>(env: LLAMA_ARG_MODEL_URL) |
+| `-dr, --docker-repo [<repo>/]<model>[:quant]` | Docker Hub model repository. repo is optional, default to ai/. quant is optional, default to :latest.<br/>example: gemma3<br/>(default: unused)<br/>(env: LLAMA_ARG_DOCKER_REPO) |
+| `-hf, -hfr, --hf-repo <user>/<model>[:quant]` | Hugging Face model repository; quant is optional, case-insensitive, default to Q4_K_M, or falls back to the first file in the repo if Q4_K_M doesn't exist.<br/>mmproj is also downloaded automatically if available. to disable, add --no-mmproj<br/>example: unsloth/phi-4-GGUF:q4_k_m<br/>(default: unused)<br/>(env: LLAMA_ARG_HF_REPO) |
+| `-hfd, -hfrd, --hf-repo-draft <user>/<model>[:quant]` | Same as --hf-repo, but for the draft model (default: unused)<br/>(env: LLAMA_ARG_HFD_REPO) |
+| `-hff, --hf-file FILE` | Hugging Face model file. If specified, it will override the quant in --hf-repo (default: unused)<br/>(env: LLAMA_ARG_HF_FILE) |
+| `-hfv, -hfrv, --hf-repo-v <user>/<model>[:quant]` | Hugging Face model repository for the vocoder model (default: unused)<br/>(env: LLAMA_ARG_HF_REPO_V) |
+| `-hffv, --hf-file-v FILE` | Hugging Face model file for the vocoder model (default: unused)<br/>(env: LLAMA_ARG_HF_FILE_V) |
+| `-hft, --hf-token TOKEN` | Hugging Face access token (default: value from HF_TOKEN environment variable)<br/>(env: HF_TOKEN) |
+| `--log-disable` | Log disable |
+| `--log-file FNAME` | Log to file<br/>(env: LLAMA_LOG_FILE) |
+| `--log-colors [on\|off\|auto]` | Set colored logging ('on', 'off', or 'auto', default: 'auto')<br/>'auto' enables colors when output is to a terminal<br/>(env: LLAMA_LOG_COLORS) |
+| `-v, --verbose, --log-verbose` | Set verbosity level to infinity (i.e. log all messages, useful for debugging) |
+| `--offline` | Offline mode: forces use of cache, prevents network access<br/>(env: LLAMA_OFFLINE) |
+| `-lv, --verbosity, --log-verbosity N` | Set the verbosity threshold. Messages with a higher verbosity will be ignored. Values:<br/> - 0: generic output<br/> - 1: error<br/> - 2: warning<br/> - 3: info<br/> - 4: debug<br/>(default: 3)<br/><br/>(env: LLAMA_LOG_VERBOSITY) |
+| `--log-prefix` | Enable prefix in log messages<br/>(env: LLAMA_LOG_PREFIX) |
+| `--log-timestamps` | Enable timestamps in log messages<br/>(env: LLAMA_LOG_TIMESTAMPS) |
+| `-ctkd, --cache-type-k-draft TYPE` | KV cache data type for K for the draft model<br/>allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1<br/>(default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_K_DRAFT) |
+| `-ctvd, --cache-type-v-draft TYPE` | KV cache data type for V for the draft model<br/>allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1<br/>(default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_V_DRAFT) |
+
+
+### Sampling params
+
+| Argument | Explanation |
+| -------- | ----------- |
+| `--samplers SAMPLERS` | samplers that will be used for generation in the order, separated by ';'<br/>(default: penalties;dry;top_n_sigma;top_k;typ_p;top_p;min_p;xtc;temperature) |
+| `-s, --seed SEED` | RNG seed (default: -1, use random seed for -1) |
+| `--sampler-seq, --sampling-seq SEQUENCE` | simplified sequence for samplers that will be used (default: edskypmxt) |
+| `--ignore-eos` | ignore end of stream token and continue generating (implies --logit-bias EOS-inf) |
+| `--temp N` | temperature (default: 0.8) |
+| `--top-k N` | top-k sampling (default: 40, 0 = disabled)<br/>(env: LLAMA_ARG_TOP_K) |
+| `--top-p N` | top-p sampling (default: 0.9, 1.0 = disabled) |
+| `--min-p N` | min-p sampling (default: 0.1, 0.0 = disabled) |
+| `--adaptive-target N` | adaptive-p: select tokens near this probability (valid range 0.0 to 1.0; negative = disabled) |
+| `--adaptive-decay N` | adaptive-p: EMA decay for adaptation; effective history length ≈ 1/(1-decay) tokens (valid range 0.0 - 0.99) |
+| `--top-nsigma N` | top-n-sigma sampling (default: -1.0, -1.0 = disabled) |
+| `--xtc-probability N` | xtc probability (default: 0.0, 0.0 = disabled) |
+| `--xtc-threshold N` | xtc threshold (default: 0.1, 1.0 = disabled) |
+| `--typical N` | locally typical sampling, parameter p (default: 1.0, 1.0 = disabled) |
+| `--repeat-last-n N` | last n tokens to consider for penalize (default: 64, 0 = disabled, -1 = ctx_size) |
+| `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.0, 1.0 = disabled) |
+| `--presence-penalty N` | repeat alpha presence penalty (default: 0.0, 0.0 = disabled) |
+| `--frequency-penalty N` | repeat alpha frequency penalty (default: 0.0, 0.0 = disabled) |
+| `--dry-multiplier N` | set DRY sampling multiplier (default: 0.0, 0.0 = disabled) |
+| `--dry-base N` | set DRY sampling base value (default: 1.75) |
+| `--dry-allowed-length N` | set allowed length for DRY sampling (default: 2) |
+| `--dry-penalty-last-n N` | set DRY penalty for the last n tokens (default: -1, 0 = disable, -1 = context size) |
+| `--dry-sequence-breaker STRING` | add sequence breaker for DRY sampling, clearing out default breakers ('\n', ':', '"', '*') in the process; use "none" to not use any sequence breakers |
+| `--dynatemp-range N` | dynamic temperature range (default: 0.0, 0.0 = disabled) |
+| `--dynatemp-exp N` | dynamic temperature exponent (default: 1.0) |
+| `--mirostat N` | use Mirostat sampling.<br/>Top K, Nucleus and Locally Typical samplers are ignored if used.<br/>(default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0) |
+| `--mirostat-lr N` | Mirostat learning rate, parameter eta (default: 0.1) |
+| `--mirostat-ent N` | Mirostat target entropy, parameter tau (default: 5.0) |
+| `-l, --logit-bias TOKEN_ID(+/-)BIAS` | modifies the likelihood of token appearing in the completion,<br/>i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',<br/>or `--logit-bias 15043-1` to decrease likelihood of token ' Hello' |
+| `--grammar GRAMMAR` | BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '') |
+| `--grammar-file FNAME` | file to read grammar from |
+| `-j, --json-schema SCHEMA` | JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object<br/>For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead |
+| `-jf, --json-schema-file FILE` | File containing a JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object<br/>For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead |
+| `-bs, --backend-sampling` | enable backend sampling (experimental) (default: disabled)<br/>(env: LLAMA_ARG_BACKEND_SAMPLING) |
+
+
+### CLI-specific params
+
+| Argument | Explanation |
+| -------- | ----------- |
+| `--display-prompt, --no-display-prompt` | whether to print prompt at generation (default: true) |
+| `-co, --color [on\|off\|auto]` | Colorize output to distinguish prompt and user input from generations ('on', 'off', or 'auto', default: 'auto')<br/>'auto' enables colors when output is to a terminal |
+| `--ctx-checkpoints, --swa-checkpoints N` | max number of context checkpoints to create per slot (default: 8)[(more info)](https://github.com/ggml-org/llama.cpp/pull/15293)<br/>(env: LLAMA_ARG_CTX_CHECKPOINTS) |
+| `-cram, --cache-ram N` | set the maximum cache size in MiB (default: 8192, -1 - no limit, 0 - disable)[(more info)](https://github.com/ggml-org/llama.cpp/pull/16391)<br/>(env: LLAMA_ARG_CACHE_RAM) |
+| `--context-shift, --no-context-shift` | whether to use context shift on infinite text generation (default: disabled)<br/>(env: LLAMA_ARG_CONTEXT_SHIFT) |
+| `-sys, --system-prompt PROMPT` | system prompt to use with model (if applicable, depending on chat template) |
+| `--show-timings, --no-show-timings` | whether to show timing information after each response (default: true)<br/>(env: LLAMA_ARG_SHOW_TIMINGS) |
+| `-sysf, --system-prompt-file FNAME` | a file containing the system prompt (default: none) |
+| `-r, --reverse-prompt PROMPT` | halt generation at PROMPT, return control in interactive mode |
+| `-sp, --special` | special tokens output enabled (default: false) |
+| `-cnv, --conversation, -no-cnv, --no-conversation` | whether to run in conversation mode:<br/>- does not print special tokens and suffix/prefix<br/>- interactive mode is also enabled<br/>(default: auto enabled if chat template is available) |
+| `-st, --single-turn` | run conversation for a single turn only, then exit when done<br/>will not be interactive if first turn is predefined with --prompt<br/>(default: false) |
+| `-mli, --multiline-input` | allows you to write or paste multiple lines without ending each in '\' |
+| `--warmup, --no-warmup` | whether to perform warmup with an empty run (default: enabled) |
+| `-mm, --mmproj FILE` | path to a multimodal projector file. see tools/mtmd/README.md<br/>note: if -hf is used, this argument can be omitted<br/>(env: LLAMA_ARG_MMPROJ) |
+| `-mmu, --mmproj-url URL` | URL to a multimodal projector file. see tools/mtmd/README.md<br/>(env: LLAMA_ARG_MMPROJ_URL) |
+| `--mmproj-auto, --no-mmproj, --no-mmproj-auto` | whether to use multimodal projector file (if available), useful when using -hf (default: enabled)<br/>(env: LLAMA_ARG_MMPROJ_AUTO) |
+| `--mmproj-offload, --no-mmproj-offload` | whether to enable GPU offloading for multimodal projector (default: enabled)<br/>(env: LLAMA_ARG_MMPROJ_OFFLOAD) |
+| `--image, --audio FILE` | path to an image or audio file. use with multimodal models, use comma-separated values for multiple files |
+| `--image-min-tokens N` | minimum number of tokens each image can take, only used by vision models with dynamic resolution (default: read from model)<br/>(env: LLAMA_ARG_IMAGE_MIN_TOKENS) |
+| `--image-max-tokens N` | maximum number of tokens each image can take, only used by vision models with dynamic resolution (default: read from model)<br/>(env: LLAMA_ARG_IMAGE_MAX_TOKENS) |
+| `-otd, --override-tensor-draft <tensor name pattern>=<buffer type>,...` | override tensor buffer type for draft model |
+| `-cmoed, --cpu-moe-draft` | keep all Mixture of Experts (MoE) weights in the CPU for the draft model<br/>(env: LLAMA_ARG_CPU_MOE_DRAFT) |
+| `-ncmoed, --n-cpu-moe-draft N` | keep the Mixture of Experts (MoE) weights of the first N layers in the CPU for the draft model<br/>(env: LLAMA_ARG_N_CPU_MOE_DRAFT) |
+| `--chat-template-kwargs STRING` | sets additional params for the json template parser, must be a valid json object string, e.g. '{"key1":"value1","key2":"value2"}'<br/>(env: LLAMA_CHAT_TEMPLATE_KWARGS) |
+| `--jinja, --no-jinja` | whether to use jinja template engine for chat (default: enabled)<br/>(env: LLAMA_ARG_JINJA) |
+| `--reasoning-format FORMAT` | controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned; one of:<br/>- none: leaves thoughts unparsed in `message.content`<br/>- deepseek: puts thoughts in `message.reasoning_content`<br/>- deepseek-legacy: keeps `<think>` tags in `message.content` while also populating `message.reasoning_content`<br/>(default: auto)<br/>(env: LLAMA_ARG_THINK) |
+| `--reasoning-budget N` | controls the amount of thinking allowed; currently only one of: -1 for unrestricted thinking budget, or 0 to disable thinking (default: -1)<br/>(env: LLAMA_ARG_THINK_BUDGET) |
+| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
+| `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) |
+| `--simple-io` | use basic IO for better compatibility in subprocesses and limited consoles |
+| `--draft, --draft-n, --draft-max N` | number of tokens to draft for speculative decoding (default: 16)<br/>(env: LLAMA_ARG_DRAFT_MAX) |
+| `--draft-min, --draft-n-min N` | minimum number of draft tokens to use for speculative decoding (default: 0)<br/>(env: LLAMA_ARG_DRAFT_MIN) |
+| `--draft-p-min P` | minimum speculative decoding probability (greedy) (default: 0.8)<br/>(env: LLAMA_ARG_DRAFT_P_MIN) |
+| `-cd, --ctx-size-draft N` | size of the prompt context for the draft model (default: 0, 0 = loaded from model)<br/>(env: LLAMA_ARG_CTX_SIZE_DRAFT) |
+| `-devd, --device-draft <dev1,dev2,..>` | comma-separated list of devices to use for offloading the draft model (none = don't offload)<br/>use --list-devices to see a list of available devices |
+| `-ngld, --gpu-layers-draft, --n-gpu-layers-draft N` | max. number of draft model layers to store in VRAM, either an exact number, 'auto', or 'all' (default: auto)<br/>(env: LLAMA_ARG_N_GPU_LAYERS_DRAFT) |
+| `-md, --model-draft FNAME` | draft model for speculative decoding (default: unused)<br/>(env: LLAMA_ARG_MODEL_DRAFT) |
+| `--spec-replace TARGET DRAFT` | translate the string in TARGET into DRAFT if the draft model and main model are not compatible |
+| `--gpt-oss-20b-default` | use gpt-oss-20b (note: can download weights from the internet) |
+| `--gpt-oss-120b-default` | use gpt-oss-120b (note: can download weights from the internet) |
+| `--vision-gemma-4b-default` | use Gemma 3 4B QAT (note: can download weights from the internet) |
+| `--vision-gemma-12b-default` | use Gemma 3 12B QAT (note: can download weights from the internet) |
+
+<!-- HELP_END -->

tools/cli/cli.cpp

@@ -0,0 +1,393 @@
+#include "common.h"
+#include "arg.h"
+#include "console.h"
+// #include "log.h"
+
+#include "server-context.h"
+#include "server-task.h"
+
+#include <atomic>
+#include <fstream>
+#include <thread>
+#include <signal.h>
+
+#if defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#   define NOMINMAX
+#endif
+#include <windows.h>
+#endif
+
+const char * LLAMA_ASCII_LOGO = R"(
+▄▄ ▄▄
+██ ██
+██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
+██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
+██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
+                                    ██    ██
+                                    ▀▀    ▀▀
+)";
+
+static std::atomic<bool> g_is_interrupted = false;
+static bool should_stop() {
+    return g_is_interrupted.load();
+}
+
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
+static void signal_handler(int) {
+    if (g_is_interrupted.load()) {
+        // second Ctrl+C - exit immediately
+        // make sure to clear colors before exiting (not using LOG or console.cpp here to avoid deadlock)
+        fprintf(stdout, "\033[0m\n");
+        fflush(stdout);
+        std::exit(130);
+    }
+    g_is_interrupted.store(true);
+}
+#endif
+
+struct cli_context {
+    server_context ctx_server;
+    json messages = json::array();
+    std::vector<raw_buffer> input_files;
+    task_params defaults;
+
+    // thread for showing "loading" animation
+    std::atomic<bool> loading_show;
+
+    cli_context(const common_params & params) {
+        defaults.sampling    = params.sampling;
+        defaults.speculative = params.speculative;
+        defaults.n_keep      = params.n_keep;
+        defaults.n_predict   = params.n_predict;
+        defaults.antiprompt  = params.antiprompt;
+
+        defaults.stream = true; // make sure we always use streaming mode
+        defaults.timings_per_token = true; // in order to get timings even when we cancel mid-way
+        // defaults.return_progress = true; // TODO: show progress
+        defaults.oaicompat_chat_syntax.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+    }
+
+    std::string generate_completion(result_timings & out_timings) {
+        server_response_reader rd = ctx_server.get_response_reader();
+        {
+            // TODO: reduce some copies here in the future
+            server_task task = server_task(SERVER_TASK_TYPE_COMPLETION);
+            task.id        = rd.get_new_id();
+            task.index     = 0;
+            task.params    = defaults;    // copy
+            task.cli_input = messages;    // copy
+            task.cli_files = input_files; // copy
+            rd.post_task({std::move(task)});
+        }
+
+        // wait for first result
+        console::spinner::start();
+        server_task_result_ptr result = rd.next(should_stop);
+
+        console::spinner::stop();
+        std::string curr_content;
+        bool is_thinking = false;
+
+        while (result) {
+            if (should_stop()) {
+                break;
+            }
+            if (result->is_error()) {
+                json err_data = result->to_json();
+                if (err_data.contains("message")) {
+                    console::error("Error: %s\n", err_data["message"].get<std::string>().c_str());
+                } else {
+                    console::error("Error: %s\n", err_data.dump().c_str());
+                }
+                return curr_content;
+            }
+            auto res_partial = dynamic_cast<server_task_result_cmpl_partial *>(result.get());
+            if (res_partial) {
+                out_timings = std::move(res_partial->timings);
+                for (const auto & diff : res_partial->oaicompat_msg_diffs) {
+                    if (!diff.content_delta.empty()) {
+                        if (is_thinking) {
+                            console::log("\n[End thinking]\n\n");
+                            console::set_display(DISPLAY_TYPE_RESET);
+                            is_thinking = false;
+                        }
+                        curr_content += diff.content_delta;
+                        console::log("%s", diff.content_delta.c_str());
+                        console::flush();
+                    }
+                    if (!diff.reasoning_content_delta.empty()) {
+                        console::set_display(DISPLAY_TYPE_REASONING);
+                        if (!is_thinking) {
+                            console::log("[Start thinking]\n");
+                        }
+                        is_thinking = true;
+                        console::log("%s", diff.reasoning_content_delta.c_str());
+                        console::flush();
+                    }
+                }
+            }
+            auto res_final = dynamic_cast<server_task_result_cmpl_final *>(result.get());
+            if (res_final) {
+                out_timings = std::move(res_final->timings);
+                break;
+            }
+            result = rd.next(should_stop);
+        }
+        g_is_interrupted.store(false);
+        // server_response_reader automatically cancels pending tasks upon destruction
+        return curr_content;
+    }
+
+    // TODO: support remote files in the future (http, https, etc)
+    std::string load_input_file(const std::string & fname, bool is_media) {
+        std::ifstream file(fname, std::ios::binary);
+        if (!file) {
+            return "";
+        }
+        if (is_media) {
+            raw_buffer buf;
+            buf.assign((std::istreambuf_iterator<char>(file)), std::istreambuf_iterator<char>());
+            input_files.push_back(std::move(buf));
+            return mtmd_default_marker();
+        } else {
+            std::string content((std::istreambuf_iterator<char>(file)), std::istreambuf_iterator<char>());
+            return content;
+        }
+    }
+};
+
+int main(int argc, char ** argv) {
+    common_params params;
+
+    params.verbosity = LOG_LEVEL_ERROR; // by default, less verbose logs
+
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_CLI)) {
+        return 1;
+    }
+
+    // TODO: maybe support it later?
+    if (params.conversation_mode == COMMON_CONVERSATION_MODE_DISABLED) {
+        console::error("--no-conversation is not supported by llama-cli\n");
+        console::error("please use llama-completion instead\n");
+    }
+
+    common_init();
+
+    // struct that contains llama context and inference
+    cli_context ctx_cli(params);
+
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    // TODO: avoid using atexit() here by making `console` a singleton
+    console::init(params.simple_io, params.use_color);
+    atexit([]() { console::cleanup(); });
+
+    console::set_display(DISPLAY_TYPE_RESET);
+
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
+    struct sigaction sigint_action;
+    sigint_action.sa_handler = signal_handler;
+    sigemptyset (&sigint_action.sa_mask);
+    sigint_action.sa_flags = 0;
+    sigaction(SIGINT, &sigint_action, NULL);
+    sigaction(SIGTERM, &sigint_action, NULL);
+#elif defined (_WIN32)
+    auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
+        return (ctrl_type == CTRL_C_EVENT) ? (signal_handler(SIGINT), true) : false;
+    };
+    SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
+#endif
+
+    console::log("\nLoading model... "); // followed by loading animation
+    console::spinner::start();
+    if (!ctx_cli.ctx_server.load_model(params)) {
+        console::spinner::stop();
+        console::error("\nFailed to load the model\n");
+        return 1;
+    }
+
+    console::spinner::stop();
+    console::log("\n");
+
+    std::thread inference_thread([&ctx_cli]() {
+        ctx_cli.ctx_server.start_loop();
+    });
+
+    auto inf = ctx_cli.ctx_server.get_meta();
+    std::string modalities = "text";
+    if (inf.has_inp_image) {
+        modalities += ", vision";
+    }
+    if (inf.has_inp_audio) {
+        modalities += ", audio";
+    }
+
+    if (!params.system_prompt.empty()) {
+        ctx_cli.messages.push_back({
+            {"role",    "system"},
+            {"content", params.system_prompt}
+        });
+    }
+
+    console::log("\n");
+    console::log("%s\n", LLAMA_ASCII_LOGO);
+    console::log("build      : %s\n", inf.build_info.c_str());
+    console::log("model      : %s\n", inf.model_name.c_str());
+    console::log("modalities : %s\n", modalities.c_str());
+    if (!params.system_prompt.empty()) {
+        console::log("using custom system prompt\n");
+    }
+    console::log("\n");
+    console::log("available commands:\n");
+    console::log("  /exit or Ctrl+C     stop or exit\n");
+    console::log("  /regen              regenerate the last response\n");
+    console::log("  /clear              clear the chat history\n");
+    console::log("  /read               add a text file\n");
+    if (inf.has_inp_image) {
+        console::log("  /image <file>       add an image file\n");
+    }
+    if (inf.has_inp_audio) {
+        console::log("  /audio <file>       add an audio file\n");
+    }
+    console::log("\n");
+
+    // interactive loop
+    std::string cur_msg;
+    while (true) {
+        std::string buffer;
+        console::set_display(DISPLAY_TYPE_USER_INPUT);
+        if (params.prompt.empty()) {
+            console::log("\n> ");
+            std::string line;
+            bool another_line = true;
+            do {
+                another_line = console::readline(line, params.multiline_input);
+                buffer += line;
+            } while (another_line);
+        } else {
+            // process input prompt from args
+            for (auto & fname : params.image) {
+                std::string marker = ctx_cli.load_input_file(fname, true);
+                if (marker.empty()) {
+                    console::error("file does not exist or cannot be opened: '%s'\n", fname.c_str());
+                    break;
+                }
+                console::log("Loaded media from '%s'\n", fname.c_str());
+                cur_msg += marker;
+            }
+            buffer = params.prompt;
+            if (buffer.size() > 500) {
+                console::log("\n> %s ... (truncated)\n", buffer.substr(0, 500).c_str());
+            } else {
+                console::log("\n> %s\n", buffer.c_str());
+            }
+            params.prompt.clear(); // only use it once
+        }
+        console::set_display(DISPLAY_TYPE_RESET);
+        console::log("\n");
+
+        if (should_stop()) {
+            g_is_interrupted.store(false);
+            break;
+        }
+
+        // remove trailing newline
+        if (!buffer.empty() &&buffer.back() == '\n') {
+            buffer.pop_back();
+        }
+
+        // skip empty messages
+        if (buffer.empty()) {
+            continue;
+        }
+
+        bool add_user_msg = true;
+
+        // process commands
+        if (string_starts_with(buffer, "/exit")) {
+            break;
+        } else if (string_starts_with(buffer, "/regen")) {
+            if (ctx_cli.messages.size() >= 2) {
+                size_t last_idx = ctx_cli.messages.size() - 1;
+                ctx_cli.messages.erase(last_idx);
+                add_user_msg = false;
+            } else {
+                console::error("No message to regenerate.\n");
+                continue;
+            }
+        } else if (string_starts_with(buffer, "/clear")) {
+            ctx_cli.messages.clear();
+            ctx_cli.input_files.clear();
+            console::log("Chat history cleared.\n");
+            continue;
+        } else if (
+                (string_starts_with(buffer, "/image ") && inf.has_inp_image) ||
+                (string_starts_with(buffer, "/audio ") && inf.has_inp_audio)) {
+            // just in case (bad copy-paste for example), we strip all trailing/leading spaces
+            std::string fname = string_strip(buffer.substr(7));
+            std::string marker = ctx_cli.load_input_file(fname, true);
+            if (marker.empty()) {
+                console::error("file does not exist or cannot be opened: '%s'\n", fname.c_str());
+                continue;
+            }
+            cur_msg += marker;
+            console::log("Loaded media from '%s'\n", fname.c_str());
+            continue;
+        } else if (string_starts_with(buffer, "/read ")) {
+            std::string fname = string_strip(buffer.substr(6));
+            std::string marker = ctx_cli.load_input_file(fname, false);
+            if (marker.empty()) {
+                console::error("file does not exist or cannot be opened: '%s'\n", fname.c_str());
+                continue;
+            }
+            cur_msg += marker;
+            console::log("Loaded text from '%s'\n", fname.c_str());
+            continue;
+        } else {
+            // not a command
+            cur_msg += buffer;
+        }
+
+        // generate response
+        if (add_user_msg) {
+            ctx_cli.messages.push_back({
+                {"role",    "user"},
+                {"content", cur_msg}
+            });
+            cur_msg.clear();
+        }
+        result_timings timings;
+        std::string assistant_content = ctx_cli.generate_completion(timings);
+        ctx_cli.messages.push_back({
+            {"role",    "assistant"},
+            {"content", assistant_content}
+        });
+        console::log("\n");
+
+        if (params.show_timings) {
+            console::set_display(DISPLAY_TYPE_INFO);
+            console::log("\n");
+            console::log("[ Prompt: %.1f t/s | Generation: %.1f t/s ]\n", timings.prompt_per_second, timings.predicted_per_second);
+            console::set_display(DISPLAY_TYPE_RESET);
+        }
+
+        if (params.single_turn) {
+            break;
+        }
+    }
+
+    console::set_display(DISPLAY_TYPE_RESET);
+
+    console::log("\nExiting...\n");
+    ctx_cli.ctx_server.terminate();
+    inference_thread.join();
+
+    // bump the log level to display timings
+    common_log_set_verbosity_thold(LOG_LEVEL_INFO);
+    llama_memory_breakdown_print(ctx_cli.ctx_server.get_llama_context());
+
+    return 0;
+}

tools/completion/CMakeLists.txt

@@ -0,0 +1,8 @@
+set(TARGET llama-completion)
+add_executable(${TARGET} completion.cpp)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_17)
+
+if(LLAMA_TOOLS_INSTALL)
+    install(TARGETS ${TARGET} RUNTIME)
+endif()

tools/completion/README.md

@@ -0,0 +1,576 @@
+# llama.cpp/tools/completion
+
+This example program allows you to use various LLaMA language models easily and efficiently. It is specifically designed to work with the [llama.cpp](https://github.com/ggml-org/llama.cpp) project, which provides a plain C/C++ implementation with optional 4-bit quantization support for faster, lower memory inference, and is optimized for desktop CPUs. This program can be used to perform various inference tasks with LLaMA models, including generating text based on user-provided prompts and chat-like interactions with reverse prompts.
+
+## Table of Contents
+
+1. [Quick Start](#quick-start)
+2. [Usage](#usage)
+3. [Common Options](#common-options)
+4. [Input Prompts](#input-prompts)
+5. [Interaction](#interaction)
+6. [Context Management](#context-management)
+7. [Generation Flags](#generation-flags)
+8. [Performance Tuning and Memory Options](#performance-tuning-and-memory-options)
+9. [Additional Options](#additional-options)
+
+## Quick Start
+
+To get started right away, run the following command, making sure to use the correct path for the model you have:
+
+First, we will need to download a model. In these examples, we will use the Gemma model from the ggml-org repo on Hugging Face.
+[https://huggingface.co/ggml-org/gemma-1.1-7b-it-Q4_K_M-GGUF/resolve/main/gemma-1.1-7b-it.Q4_K_M.gguf?download=true](https://huggingface.co/ggml-org/gemma-1.1-7b-it-Q4_K_M-GGUF/resolve/main/gemma-1.1-7b-it.Q4_K_M.gguf?download=true)
+
+Once downloaded, place your model in the models folder in llama.cpp.
+
+### Unix-based systems (Linux, macOS, etc.):
+
+##### Input prompt (One-and-done)
+
+```bash
+./llama-completion -m models/gemma-1.1-7b-it.Q4_K_M.gguf -no-cnv --prompt "Once upon a time"
+```
+##### Conversation mode (Allow for continuous interaction with the model)
+
+```bash
+./llama-completion -m models/gemma-1.1-7b-it.Q4_K_M.gguf --chat-template gemma
+```
+
+##### Conversation mode using built-in jinja chat template
+
+```bash
+./llama-completion -m models/gemma-1.1-7b-it.Q4_K_M.gguf --jinja
+```
+
+##### One-and-done query using jinja with custom system prompt and a starting prompt
+
+```bash
+./llama-completion -m models/gemma-1.1-7b-it.Q4_K_M.gguf --jinja --single-turn -sys "You are a helpful assistant" -p "Hello"
+```
+
+##### Infinite text from a starting prompt (you can use `Ctrl-C` to stop it):
+```bash
+./llama-completion -m models/gemma-1.1-7b-it.Q4_K_M.gguf --ignore-eos -n -1
+```
+
+### Windows:
+
+##### Input prompt (One-and-done)
+```powershell
+./llama-completion.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf -no-cnv --prompt "Once upon a time"
+```
+##### Conversation mode (Allow for continuous interaction with the model)
+
+```powershell
+./llama-completion.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --chat-template gemma
+```
+
+##### Conversation mode using built-in jinja chat template
+
+```powershell
+./llama-completion.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --jinja
+```
+
+##### One-and-done query using jinja with custom system prompt and a starting prompt
+
+```powershell
+./llama-completion.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --jinja --single-turn -sys "You are a helpful assistant" -p "Hello"
+```
+
+#### Infinite text from a starting prompt (you can use `Ctrl-C` to stop it):
+
+```powershell
+llama-completion.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --ignore-eos -n -1
+```
+
+## Usage
+
+<!-- HELP_START -->
+
+<!-- IMPORTANT: The list below is auto-generated by llama-gen-docs; do NOT modify it manually -->
+
+### Common params
+
+| Argument | Explanation |
+| -------- | ----------- |
+| `-h, --help, --usage` | print usage and exit |
+| `--version` | show version and build info |
+| `--license` | show source code license and dependencies |
+| `-cl, --cache-list` | show list of models in cache |
+| `--completion-bash` | print source-able bash completion script for llama.cpp |
+| `--verbose-prompt` | print a verbose prompt before generation (default: false) |
+| `-t, --threads N` | number of CPU threads to use during generation (default: -1)<br/>(env: LLAMA_ARG_THREADS) |
+| `-tb, --threads-batch N` | number of threads to use during batch and prompt processing (default: same as --threads) |
+| `-C, --cpu-mask M` | CPU affinity mask: arbitrarily long hex. Complements cpu-range (default: "") |
+| `-Cr, --cpu-range lo-hi` | range of CPUs for affinity. Complements --cpu-mask |
+| `--cpu-strict <0\|1>` | use strict CPU placement (default: 0) |
+| `--prio N` | set process/thread priority : low(-1), normal(0), medium(1), high(2), realtime(3) (default: 0) |
+| `--poll <0...100>` | use polling level to wait for work (0 - no polling, default: 50) |
+| `-Cb, --cpu-mask-batch M` | CPU affinity mask: arbitrarily long hex. Complements cpu-range-batch (default: same as --cpu-mask) |
+| `-Crb, --cpu-range-batch lo-hi` | ranges of CPUs for affinity. Complements --cpu-mask-batch |
+| `--cpu-strict-batch <0\|1>` | use strict CPU placement (default: same as --cpu-strict) |
+| `--prio-batch N` | set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: 0) |
+| `--poll-batch <0\|1>` | use polling to wait for work (default: same as --poll) |
+| `-c, --ctx-size N` | size of the prompt context (default: 0, 0 = loaded from model)<br/>(env: LLAMA_ARG_CTX_SIZE) |
+| `-n, --predict, --n-predict N` | number of tokens to predict (default: -1, -1 = infinity, -2 = until context filled)<br/>(env: LLAMA_ARG_N_PREDICT) |
+| `-b, --batch-size N` | logical maximum batch size (default: 2048)<br/>(env: LLAMA_ARG_BATCH) |
+| `-ub, --ubatch-size N` | physical maximum batch size (default: 512)<br/>(env: LLAMA_ARG_UBATCH) |
+| `--keep N` | number of tokens to keep from the initial prompt (default: 0, -1 = all) |
+| `--swa-full` | use full-size SWA cache (default: false)<br/>[(more info)](https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)<br/>(env: LLAMA_ARG_SWA_FULL) |
+| `-fa, --flash-attn [on\|off\|auto]` | set Flash Attention use ('on', 'off', or 'auto', default: 'auto')<br/>(env: LLAMA_ARG_FLASH_ATTN) |
+| `-p, --prompt PROMPT` | prompt to start generation with; for system message, use -sys |
+| `--perf, --no-perf` | whether to enable internal libllama performance timings (default: false)<br/>(env: LLAMA_ARG_PERF) |
+| `-f, --file FNAME` | a file containing the prompt (default: none) |
+| `-bf, --binary-file FNAME` | binary file containing the prompt (default: none) |
+| `-e, --escape, --no-escape` | whether to process escapes sequences (\n, \r, \t, \', \", \\) (default: true) |
+| `--rope-scaling {none,linear,yarn}` | RoPE frequency scaling method, defaults to linear unless specified by the model<br/>(env: LLAMA_ARG_ROPE_SCALING_TYPE) |
+| `--rope-scale N` | RoPE context scaling factor, expands context by a factor of N<br/>(env: LLAMA_ARG_ROPE_SCALE) |
+| `--rope-freq-base N` | RoPE base frequency, used by NTK-aware scaling (default: loaded from model)<br/>(env: LLAMA_ARG_ROPE_FREQ_BASE) |
+| `--rope-freq-scale N` | RoPE frequency scaling factor, expands context by a factor of 1/N<br/>(env: LLAMA_ARG_ROPE_FREQ_SCALE) |
+| `--yarn-orig-ctx N` | YaRN: original context size of model (default: 0 = model training context size)<br/>(env: LLAMA_ARG_YARN_ORIG_CTX) |
+| `--yarn-ext-factor N` | YaRN: extrapolation mix factor (default: -1.0, 0.0 = full interpolation)<br/>(env: LLAMA_ARG_YARN_EXT_FACTOR) |
+| `--yarn-attn-factor N` | YaRN: scale sqrt(t) or attention magnitude (default: -1.0)<br/>(env: LLAMA_ARG_YARN_ATTN_FACTOR) |
+| `--yarn-beta-slow N` | YaRN: high correction dim or alpha (default: -1.0)<br/>(env: LLAMA_ARG_YARN_BETA_SLOW) |
+| `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: -1.0)<br/>(env: LLAMA_ARG_YARN_BETA_FAST) |
+| `-kvo, --kv-offload, -nkvo, --no-kv-offload` | whether to enable KV cache offloading (default: enabled)<br/>(env: LLAMA_ARG_KV_OFFLOAD) |
+| `--repack, -nr, --no-repack` | whether to enable weight repacking (default: enabled)<br/>(env: LLAMA_ARG_REPACK) |
+| `--no-host` | bypass host buffer allowing extra buffers to be used<br/>(env: LLAMA_ARG_NO_HOST) |
+| `-ctk, --cache-type-k TYPE` | KV cache data type for K<br/>allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1<br/>(default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_K) |
+| `-ctv, --cache-type-v TYPE` | KV cache data type for V<br/>allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1<br/>(default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_V) |
+| `-dt, --defrag-thold N` | KV cache defragmentation threshold (DEPRECATED)<br/>(env: LLAMA_ARG_DEFRAG_THOLD) |
+| `-np, --parallel N` | number of parallel sequences to decode (default: 1)<br/>(env: LLAMA_ARG_N_PARALLEL) |
+| `--mlock` | force system to keep model in RAM rather than swapping or compressing<br/>(env: LLAMA_ARG_MLOCK) |
+| `--mmap, --no-mmap` | whether to memory-map model. Explicitly enabling mmap disables direct-io. (if mmap disabled, slower load but may reduce pageouts if not using mlock) (default: enabled)<br/>(env: LLAMA_ARG_MMAP) |
+| `-dio, --direct-io, -ndio, --no-direct-io` | use DirectIO if available. Takes precedence over --mmap (default: enabled)<br/>(env: LLAMA_ARG_DIO) |
+| `--numa TYPE` | attempt optimizations that help on some NUMA systems<br/>- distribute: spread execution evenly over all nodes<br/>- isolate: only spawn threads on CPUs on the node that execution started on<br/>- numactl: use the CPU map provided by numactl<br/>if run without this previously, it is recommended to drop the system page cache before using this<br/>see https://github.com/ggml-org/llama.cpp/issues/1437<br/>(env: LLAMA_ARG_NUMA) |
+| `-dev, --device <dev1,dev2,..>` | comma-separated list of devices to use for offloading (none = don't offload)<br/>use --list-devices to see a list of available devices<br/>(env: LLAMA_ARG_DEVICE) |
+| `--list-devices` | print list of available devices and exit |
+| `-ot, --override-tensor <tensor name pattern>=<buffer type>,...` | override tensor buffer type<br/>(env: LLAMA_ARG_OVERRIDE_TENSOR) |
+| `-cmoe, --cpu-moe` | keep all Mixture of Experts (MoE) weights in the CPU<br/>(env: LLAMA_ARG_CPU_MOE) |
+| `-ncmoe, --n-cpu-moe N` | keep the Mixture of Experts (MoE) weights of the first N layers in the CPU<br/>(env: LLAMA_ARG_N_CPU_MOE) |
+| `-ngl, --gpu-layers, --n-gpu-layers N` | max. number of layers to store in VRAM, either an exact number, 'auto', or 'all' (default: auto)<br/>(env: LLAMA_ARG_N_GPU_LAYERS) |
+| `-sm, --split-mode {none,layer,row}` | how to split the model across multiple GPUs, one of:<br/>- none: use one GPU only<br/>- layer (default): split layers and KV across GPUs<br/>- row: split rows across GPUs<br/>(env: LLAMA_ARG_SPLIT_MODE) |
+| `-ts, --tensor-split N0,N1,N2,...` | fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1<br/>(env: LLAMA_ARG_TENSOR_SPLIT) |
+| `-mg, --main-gpu INDEX` | the GPU to use for the model (with split-mode = none), or for intermediate results and KV (with split-mode = row) (default: 0)<br/>(env: LLAMA_ARG_MAIN_GPU) |
+| `-fit, --fit [on\|off]` | whether to adjust unset arguments to fit in device memory ('on' or 'off', default: 'on')<br/>(env: LLAMA_ARG_FIT) |
+| `-fitt, --fit-target MiB0,MiB1,MiB2,...` | target margin per device for --fit, comma-separated list of values, single value is broadcast across all devices, default: 1024<br/>(env: LLAMA_ARG_FIT_TARGET) |
+| `-fitc, --fit-ctx N` | minimum ctx size that can be set by --fit option, default: 4096<br/>(env: LLAMA_ARG_FIT_CTX) |
+| `--check-tensors` | check model tensor data for invalid values (default: false) |
+| `--override-kv KEY=TYPE:VALUE,...` | advanced option to override model metadata by key. to specify multiple overrides, either use comma-separated values.<br/>types: int, float, bool, str. example: --override-kv tokenizer.ggml.add_bos_token=bool:false,tokenizer.ggml.add_eos_token=bool:false |
+| `--op-offload, --no-op-offload` | whether to offload host tensor operations to device (default: true) |
+| `--lora FNAME` | path to LoRA adapter (use comma-separated values to load multiple adapters) |
+| `--lora-scaled FNAME:SCALE,...` | path to LoRA adapter with user defined scaling (format: FNAME:SCALE,...)<br/>note: use comma-separated values |
+| `--control-vector FNAME` | add a control vector<br/>note: use comma-separated values to add multiple control vectors |
+| `--control-vector-scaled FNAME:SCALE,...` | add a control vector with user defined scaling SCALE<br/>note: use comma-separated values (format: FNAME:SCALE,...) |
+| `--control-vector-layer-range START END` | layer range to apply the control vector(s) to, start and end inclusive |
+| `-m, --model FNAME` | model path to load<br/>(env: LLAMA_ARG_MODEL) |
+| `-mu, --model-url MODEL_URL` | model download url (default: unused)<br/>(env: LLAMA_ARG_MODEL_URL) |
+| `-dr, --docker-repo [<repo>/]<model>[:quant]` | Docker Hub model repository. repo is optional, default to ai/. quant is optional, default to :latest.<br/>example: gemma3<br/>(default: unused)<br/>(env: LLAMA_ARG_DOCKER_REPO) |
+| `-hf, -hfr, --hf-repo <user>/<model>[:quant]` | Hugging Face model repository; quant is optional, case-insensitive, default to Q4_K_M, or falls back to the first file in the repo if Q4_K_M doesn't exist.<br/>mmproj is also downloaded automatically if available. to disable, add --no-mmproj<br/>example: unsloth/phi-4-GGUF:q4_k_m<br/>(default: unused)<br/>(env: LLAMA_ARG_HF_REPO) |
+| `-hfd, -hfrd, --hf-repo-draft <user>/<model>[:quant]` | Same as --hf-repo, but for the draft model (default: unused)<br/>(env: LLAMA_ARG_HFD_REPO) |
+| `-hff, --hf-file FILE` | Hugging Face model file. If specified, it will override the quant in --hf-repo (default: unused)<br/>(env: LLAMA_ARG_HF_FILE) |
+| `-hfv, -hfrv, --hf-repo-v <user>/<model>[:quant]` | Hugging Face model repository for the vocoder model (default: unused)<br/>(env: LLAMA_ARG_HF_REPO_V) |
+| `-hffv, --hf-file-v FILE` | Hugging Face model file for the vocoder model (default: unused)<br/>(env: LLAMA_ARG_HF_FILE_V) |
+| `-hft, --hf-token TOKEN` | Hugging Face access token (default: value from HF_TOKEN environment variable)<br/>(env: HF_TOKEN) |
+| `--log-disable` | Log disable |
+| `--log-file FNAME` | Log to file<br/>(env: LLAMA_LOG_FILE) |
+| `--log-colors [on\|off\|auto]` | Set colored logging ('on', 'off', or 'auto', default: 'auto')<br/>'auto' enables colors when output is to a terminal<br/>(env: LLAMA_LOG_COLORS) |
+| `-v, --verbose, --log-verbose` | Set verbosity level to infinity (i.e. log all messages, useful for debugging) |
+| `--offline` | Offline mode: forces use of cache, prevents network access<br/>(env: LLAMA_OFFLINE) |
+| `-lv, --verbosity, --log-verbosity N` | Set the verbosity threshold. Messages with a higher verbosity will be ignored. Values:<br/> - 0: generic output<br/> - 1: error<br/> - 2: warning<br/> - 3: info<br/> - 4: debug<br/>(default: 3)<br/><br/>(env: LLAMA_LOG_VERBOSITY) |
+| `--log-prefix` | Enable prefix in log messages<br/>(env: LLAMA_LOG_PREFIX) |
+| `--log-timestamps` | Enable timestamps in log messages<br/>(env: LLAMA_LOG_TIMESTAMPS) |
+| `-ctkd, --cache-type-k-draft TYPE` | KV cache data type for K for the draft model<br/>allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1<br/>(default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_K_DRAFT) |
+| `-ctvd, --cache-type-v-draft TYPE` | KV cache data type for V for the draft model<br/>allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1<br/>(default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_V_DRAFT) |
+
+
+### Sampling params
+
+| Argument | Explanation |
+| -------- | ----------- |
+| `--samplers SAMPLERS` | samplers that will be used for generation in the order, separated by ';'<br/>(default: penalties;dry;top_n_sigma;top_k;typ_p;top_p;min_p;xtc;temperature) |
+| `-s, --seed SEED` | RNG seed (default: -1, use random seed for -1) |
+| `--sampler-seq, --sampling-seq SEQUENCE` | simplified sequence for samplers that will be used (default: edskypmxt) |
+| `--ignore-eos` | ignore end of stream token and continue generating (implies --logit-bias EOS-inf) |
+| `--temp N` | temperature (default: 0.8) |
+| `--top-k N` | top-k sampling (default: 40, 0 = disabled)<br/>(env: LLAMA_ARG_TOP_K) |
+| `--top-p N` | top-p sampling (default: 0.9, 1.0 = disabled) |
+| `--min-p N` | min-p sampling (default: 0.1, 0.0 = disabled) |
+| `--top-nsigma N` | top-n-sigma sampling (default: -1.0, -1.0 = disabled) |
+| `--xtc-probability N` | xtc probability (default: 0.0, 0.0 = disabled) |
+| `--xtc-threshold N` | xtc threshold (default: 0.1, 1.0 = disabled) |
+| `--typical N` | locally typical sampling, parameter p (default: 1.0, 1.0 = disabled) |
+| `--repeat-last-n N` | last n tokens to consider for penalize (default: 64, 0 = disabled, -1 = ctx_size) |
+| `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.0, 1.0 = disabled) |
+| `--presence-penalty N` | repeat alpha presence penalty (default: 0.0, 0.0 = disabled) |
+| `--frequency-penalty N` | repeat alpha frequency penalty (default: 0.0, 0.0 = disabled) |
+| `--dry-multiplier N` | set DRY sampling multiplier (default: 0.0, 0.0 = disabled) |
+| `--dry-base N` | set DRY sampling base value (default: 1.75) |
+| `--dry-allowed-length N` | set allowed length for DRY sampling (default: 2) |
+| `--dry-penalty-last-n N` | set DRY penalty for the last n tokens (default: -1, 0 = disable, -1 = context size) |
+| `--dry-sequence-breaker STRING` | add sequence breaker for DRY sampling, clearing out default breakers ('\n', ':', '"', '*') in the process; use "none" to not use any sequence breakers |
+| `--dynatemp-range N` | dynamic temperature range (default: 0.0, 0.0 = disabled) |
+| `--dynatemp-exp N` | dynamic temperature exponent (default: 1.0) |
+| `--mirostat N` | use Mirostat sampling.<br/>Top K, Nucleus and Locally Typical samplers are ignored if used.<br/>(default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0) |
+| `--mirostat-lr N` | Mirostat learning rate, parameter eta (default: 0.1) |
+| `--mirostat-ent N` | Mirostat target entropy, parameter tau (default: 5.0) |
+| `-l, --logit-bias TOKEN_ID(+/-)BIAS` | modifies the likelihood of token appearing in the completion,<br/>i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',<br/>or `--logit-bias 15043-1` to decrease likelihood of token ' Hello' |
+| `--grammar GRAMMAR` | BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '') |
+| `--grammar-file FNAME` | file to read grammar from |
+| `-j, --json-schema SCHEMA` | JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object<br/>For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead |
+| `-jf, --json-schema-file FILE` | File containing a JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object<br/>For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead |
+| `-bs, --backend-sampling` | enable backend sampling (experimental) (default: disabled)<br/>(env: LLAMA_ARG_BACKEND_SAMPLING) |
+
+
+### Completion-specific params
+
+| Argument | Explanation |
+| -------- | ----------- |
+| `--display-prompt, --no-display-prompt` | whether to print prompt at generation (default: true) |
+| `-co, --color [on\|off\|auto]` | Colorize output to distinguish prompt and user input from generations ('on', 'off', or 'auto', default: 'auto')<br/>'auto' enables colors when output is to a terminal |
+| `--context-shift, --no-context-shift` | whether to use context shift on infinite text generation (default: disabled)<br/>(env: LLAMA_ARG_CONTEXT_SHIFT) |
+| `-sys, --system-prompt PROMPT` | system prompt to use with model (if applicable, depending on chat template) |
+| `-sysf, --system-prompt-file FNAME` | a file containing the system prompt (default: none) |
+| `-ptc, --print-token-count N` | print token count every N tokens (default: -1) |
+| `--prompt-cache FNAME` | file to cache prompt state for faster startup (default: none) |
+| `--prompt-cache-all` | if specified, saves user input and generations to cache as well |
+| `--prompt-cache-ro` | if specified, uses the prompt cache but does not update it |
+| `-r, --reverse-prompt PROMPT` | halt generation at PROMPT, return control in interactive mode |
+| `-sp, --special` | special tokens output enabled (default: false) |
+| `-cnv, --conversation, -no-cnv, --no-conversation` | whether to run in conversation mode:<br/>- does not print special tokens and suffix/prefix<br/>- interactive mode is also enabled<br/>(default: auto enabled if chat template is available) |
+| `-st, --single-turn` | run conversation for a single turn only, then exit when done<br/>will not be interactive if first turn is predefined with --prompt<br/>(default: false) |
+| `-i, --interactive` | run in interactive mode (default: false) |
+| `-if, --interactive-first` | run in interactive mode and wait for input right away (default: false) |
+| `-mli, --multiline-input` | allows you to write or paste multiple lines without ending each in '\' |
+| `--in-prefix-bos` | prefix BOS to user inputs, preceding the `--in-prefix` string |
+| `--in-prefix STRING` | string to prefix user inputs with (default: empty) |
+| `--in-suffix STRING` | string to suffix after user inputs with (default: empty) |
+| `--warmup, --no-warmup` | whether to perform warmup with an empty run (default: enabled) |
+| `-gan, --grp-attn-n N` | group-attention factor (default: 1)<br/>(env: LLAMA_ARG_GRP_ATTN_N) |
+| `-gaw, --grp-attn-w N` | group-attention width (default: 512)<br/>(env: LLAMA_ARG_GRP_ATTN_W) |
+| `--jinja, --no-jinja` | whether to use jinja template engine for chat (default: disabled)<br/>(env: LLAMA_ARG_JINJA) |
+| `--reasoning-format FORMAT` | controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned; one of:<br/>- none: leaves thoughts unparsed in `message.content`<br/>- deepseek: puts thoughts in `message.reasoning_content`<br/>- deepseek-legacy: keeps `<think>` tags in `message.content` while also populating `message.reasoning_content`<br/>(default: auto)<br/>(env: LLAMA_ARG_THINK) |
+| `--reasoning-budget N` | controls the amount of thinking allowed; currently only one of: -1 for unrestricted thinking budget, or 0 to disable thinking (default: -1)<br/>(env: LLAMA_ARG_THINK_BUDGET) |
+| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
+| `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) |
+| `--simple-io` | use basic IO for better compatibility in subprocesses and limited consoles |
+
+<!-- HELP_END -->
+
+## Common Options
+
+In this section, we cover the most commonly used options for running the `llama-completion` program with the LLaMA models:
+
+-   `-m FNAME, --model FNAME`: Specify the path to the LLaMA model file (e.g., `models/gemma-1.1-7b-it.Q4_K_M.gguf`; inferred from `--model-url` if set).
+-   `-mu MODEL_URL --model-url MODEL_URL`: Specify a remote http url to download the file (e.g [https://huggingface.co/ggml-org/gemma-1.1-7b-it-Q4_K_M-GGUF/resolve/main/gemma-1.1-7b-it.Q4_K_M.gguf?download=true](https://huggingface.co/ggml-org/gemma-1.1-7b-it-Q4_K_M-GGUF/resolve/main/gemma-1.1-7b-it.Q4_K_M.gguf?download=true)).
+-   `-i, --interactive`: Run the program in interactive mode, allowing you to provide input directly and receive real-time responses.
+-   `-n N, --n-predict N`: Set the number of tokens to predict when generating text. Adjusting this value can influence the length of the generated text.
+-   `-c N, --ctx-size N`: Set the size of the prompt context. The default is 4096, but if a LLaMA model was built with a longer context, increasing this value will provide better results for longer input/inference.
+-   `-mli, --multiline-input`: Allows you to write or paste multiple lines without ending each in '\'
+-   `-t N, --threads N`: Set the number of threads to use during generation. For optimal performance, it is recommended to set this value to the number of physical CPU cores your system has.
+-   `-ngl N, --n-gpu-layers N`: When compiled with GPU support, this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
+
+## Input Prompts
+
+The `llama-completion` program provides several ways to interact with the LLaMA models using input prompts:
+
+-   `--prompt PROMPT`: Provide a prompt directly as a command-line option.
+-   `--file FNAME`: Provide a file containing a prompt or multiple prompts.
+-   `--system-prompt PROMPT`: Provide a system prompt (will otherwise use the default one in the chat template (if provided)).
+-   `--system-prompt-file FNAME`: Provide a file containing a system prompt.
+-   `--interactive-first`: Run the program in interactive mode and wait for input right away. (More on this below.)
+
+## Interaction
+
+The `llama-completion` program offers a seamless way to interact with LLaMA models, allowing users to engage in real-time conversations or provide instructions for specific tasks. The interactive mode can be triggered using various options, including `--interactive` and `--interactive-first`.
+
+In interactive mode, users can participate in text generation by injecting their input during the process. Users can press `Ctrl+C` at any time to interject and type their input, followed by pressing `Return` to submit it to the LLaMA model. To submit additional lines without finalizing input, users can end the current line with a backslash (`\`) and continue typing.
+
+### Interaction Options
+
+-   `-i, --interactive`: Run the program in interactive mode, allowing users to engage in real-time conversations or provide specific instructions to the model.
+-   `--interactive-first`: Run the program in interactive mode and immediately wait for user input before starting the text generation.
+-   `-cnv,  --conversation`:  Run the program in conversation mode (does not print special tokens and suffix/prefix, use default or provided chat template) (default: true if chat template found)
+-   `-no-cnv`:  Disable conversation mode (default: false)
+-   `-st, --single-turn`:  Only process a single conversation turn (user input) and then exit.
+-   `--jinja`:  Enable jinja chat template parser, will use the model's built-in template or a user-provided one (default: false)
+-   `--color`: Enable colorized output to differentiate visually distinguishing between prompts, user input, and generated text.
+
+By understanding and utilizing these interaction options, you can create engaging and dynamic experiences with the LLaMA models, tailoring the text generation process to your specific needs.
+
+### Reverse Prompts
+
+Reverse prompts are a powerful way to create a chat-like experience with a LLaMA model by pausing the text generation when specific text strings are encountered:
+
+-   `-r PROMPT, --reverse-prompt PROMPT`: Specify one or multiple reverse prompts to pause text generation and switch to interactive mode. For example, `-r "User:"` can be used to jump back into the conversation whenever it's the user's turn to speak. This helps create a more interactive and conversational experience. However, the reverse prompt doesn't work when it ends with a space.
+
+To overcome this limitation, you can use the `--in-prefix` flag to add a space or any other characters after the reverse prompt.
+
+### In-Prefix
+
+The `--in-prefix` flag is used to add a prefix to your input, primarily, this is used to insert a space after the reverse prompt. Here's an example of how to use the `--in-prefix` flag in conjunction with the `--reverse-prompt` flag:
+
+```sh
+./llama-completion -r "User:" --in-prefix " "
+```
+
+### In-Suffix
+
+The `--in-suffix` flag is used to add a suffix after your input. This is useful for adding an "Assistant:" prompt after the user's input. It's added after the new-line character (`\n`) that's automatically added to the end of the user's input. Here's an example of how to use the `--in-suffix` flag in conjunction with the `--reverse-prompt` flag:
+
+```sh
+./llama-completion -r "User:" --in-prefix " " --in-suffix "Assistant:"
+```
+When --in-prefix or --in-suffix options are enabled the chat template ( --chat-template ) is disabled
+
+### Chat templates
+
+ `--chat-template JINJA_TEMPLATE`: This option sets a custom jinja chat template. It accepts a string, not a file name.  Default: template taken from model's metadata. Llama.cpp only supports [some pre-defined templates](https://github.com/ggml-org/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template). These include llama2, llama3, gemma, monarch, chatml, orion, vicuna, vicuna-orca, deepseek, command-r, zephyr. When --in-prefix or --in-suffix options are enabled the chat template ( --chat-template ) is disabled.
+
+ Example usage: `--chat-template gemma`
+
+`--chat-template-file FNAME`:  Load a custom jinja chat template from an external file, useful if the model contains outdated or incompatible template, some examples can be found in models/templates. Up-to-date chat templates can be downloaded from Hugging Face using scripts/get_chat_template.py
+
+## Context Management
+
+During text generation, LLaMA models have a limited context size, which means they can only consider a certain number of tokens from the input and generated text. When the context fills up, the model resets internally, potentially losing some information from the beginning of the conversation or instructions. Context management options help maintain continuity and coherence in these situations.
+
+### Context Size
+
+- `-c N, --ctx-size N`: Set the size of the prompt context (default: 4096, 0 = loaded from model). If a LLaMA model was built with a longer context, increasing this value will yield the best results on longer input/inference.
+
+### Extended Context Size
+
+Some fine-tuned models have extended the context length by scaling RoPE. For example, if the original pre-trained model has a context length (max sequence length) of 4096 (4k) and the fine-tuned model has 32k. That is a scaling factor of 8, and should work by setting the above `--ctx-size` to 32768 (32k) and `--rope-scale` to 8.
+
+-   `--rope-scale N`: Where N is the linear scaling factor used by the fine-tuned model.
+
+### Keep Prompt
+
+The `--keep` option allows users to retain the original prompt when the model runs out of context, ensuring a connection to the initial instruction or conversation topic is maintained.
+
+-   `--keep N`: Specify the number of tokens from the initial prompt to retain when the model resets its internal context. By default, this value is set to 0 (meaning no tokens are kept). Use `-1` to retain all tokens from the initial prompt.
+
+By utilizing context management options like `--ctx-size` and `--keep`, you can maintain a more coherent and consistent interaction with the LLaMA models, ensuring that the generated text remains relevant to the original prompt or conversation.
+
+## Generation Flags
+
+The following options allow you to control the text generation process and fine-tune the diversity, creativity, and quality of the generated text according to your needs. By adjusting these options and experimenting with different combinations of values, you can find the best settings for your specific use case.
+
+### Number of Tokens to Predict
+
+-   `-n N, --predict N`: Set the number of tokens to predict when generating text (default: -1, -1 = infinity, -2 = until context filled)
+
+The `--predict` option controls the number of tokens the model generates in response to the input prompt. By adjusting this value, you can influence the length of the generated text. A higher value will result in longer text, while a lower value will produce shorter text.
+
+A value of -1 will enable infinite text generation, even though we have a finite context window. When the context window is full, some of the earlier tokens (half of the tokens after `--keep`) will be discarded. The context must then be re-evaluated before generation can resume. On large models and/or large context windows, this will result in a significant pause in output.
+
+If the pause is undesirable, a value of -2 will stop generation immediately when the context is filled.
+
+The `--no-context-shift` option allows you to stop the infinite text generation once the finite context window is full.
+
+It is important to note that the generated text may be shorter than the specified number of tokens if an End-of-Sequence (EOS) token or a reverse prompt is encountered. In interactive mode, text generation will pause and control will be returned to the user. In non-interactive mode, the program will end. In both cases, the text generation may stop before reaching the specified `--predict` value. If you want the model to keep going without ever producing End-of-Sequence on its own, you can use the `--ignore-eos` parameter.
+
+### Temperature
+
+-   `--temp N`: Adjust the randomness of the generated text (default: 0.8).
+
+Temperature is a hyperparameter that controls the randomness of the generated text. It affects the probability distribution of the model's output tokens. A higher temperature (e.g., 1.5) makes the output more random and creative, while a lower temperature (e.g., 0.5) makes the output more focused, deterministic, and conservative. The default value is 0.8, which provides a balance between randomness and determinism. At the extreme, a temperature of 0 will always pick the most likely next token, leading to identical outputs in each run.
+
+Example usage: `--temp 0`
+
+### Repeat Penalty
+
+-   `--repeat-penalty N`: Control the repetition of token sequences in the generated text default: 1.0, 1.0 = disabled).
+-   `--repeat-last-n N`: Last n tokens to consider for penalizing repetition (default: 64, 0 = disabled, -1 = ctx-size).
+
+The `repeat-penalty` option helps prevent the model from generating repetitive or monotonous text. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient. The default value is 1.
+
+The `repeat-last-n` option controls the number of tokens in the history to consider for penalizing repetition. A larger value will look further back in the generated text to prevent repetitions, while a smaller value will only consider recent tokens. A value of 0 disables the penalty, and a value of -1 sets the number of tokens considered equal to the context size (`ctx-size`).
+
+### DRY Repetition Penalty
+
+DRY (Don't Repeat Yourself) sampling is an effective technique for reducing repetition in generated text even across long contexts by penalizing tokens based on their recent usage patterns (original [PR link](https://github.com/oobabooga/text-generation-webui/pull/5677)).
+
+- `--dry-multiplier N`: Set the DRY sampling multiplier (default: 0.0, 0.0 = disabled).
+- `--dry-base N`: Set the DRY sampling base value (default: 1.75).
+- `--dry-allowed-length N`: Set the allowed length for DRY sampling (default: 2).
+- `--dry-penalty-last-n N`: Set DRY penalty for the last n tokens (default: -1, 0 = disable, -1 = context size).
+- `--dry-sequence-breaker STRING`: Add a sequence breaker for DRY sampling. Can be used more than once to add multiple sequence breakers. Using this clears out the default breakers, which consist of: `['\n', ':', '"', '*']`. If the string `"none"` is supplied, no sequence breakers are used.
+
+The `dry-multiplier` option controls the strength of the DRY sampling effect. A value of 0.0 disables DRY sampling, while higher values increase its influence. A typical recommended value is 0.8.
+
+The `dry-base` option sets the base value for the exponential penalty calculation in DRY sampling. Higher values lead to more aggressive penalization of repetitions.
+
+The `dry-allowed-length` option sets the maximum length of repeated sequences that will not be penalized. Repetitions shorter than or equal to this length are not penalized, allowing for natural repetitions of short phrases or common words.
+
+The `dry-penalty-last-n` option controls how many recent tokens to consider when applying the DRY penalty. A value of -1 considers the entire context. Use a positive value to limit the consideration to a specific number of recent tokens.
+
+The `dry-sequence-breaker` option adds a single sequence breaker and can be used more than once to specify multiple sequence breakers. Sequence breakers interrupt sequence matching and break the input into parts where matching can be applied.
+
+DRY sampling provides more nuanced control over text generation, particularly for reducing long-range repetitions and maintaining global coherence.
+
+Example usage: `--dry-multiplier 0.8 --dry-base 1.75 --dry-allowed-length 2 --dry-penalty-last-n -1 --dry-sequence-breaker "—" --dry-sequence-breaker "##"`
+
+### Top-K Sampling
+
+-   `--top-k N`: Limit the next token selection to the K most probable tokens (default: 40).
+
+Top-k sampling is a text generation method that selects the next token only from the top k most likely tokens predicted by the model. It helps reduce the risk of generating low-probability or nonsensical tokens, but it may also limit the diversity of the output. A higher value for top-k (e.g., 100) will consider more tokens and lead to more diverse text, while a lower value (e.g., 10) will focus on the most probable tokens and generate more conservative text. The default value is 40.
+
+Example usage: `--top-k 30`
+
+### Top-P Sampling
+
+-   `--top-p N`: Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P (default: 0.9).
+
+Top-p sampling, also known as nucleus sampling, is another text generation method that selects the next token from a subset of tokens that together have a cumulative probability of at least p. This method provides a balance between diversity and quality by considering both the probabilities of tokens and the number of tokens to sample from. A higher value for top-p (e.g., 0.95) will lead to more diverse text, while a lower value (e.g., 0.5) will generate more focused and conservative text. The default value is 0.9.
+
+Example usage: `--top-p 0.95`
+
+### Min-P Sampling
+
+-   `--min-p N`: Sets a minimum base probability threshold for token selection (default: 0.1).
+
+The Min-P sampling method was designed as an alternative to Top-P, and aims to ensure a balance of quality and variety. The parameter *p* represents the minimum probability for a token to be considered, relative to the probability of the most likely token. For example, with *p*=0.05 and the most likely token having a probability of 0.9, logits with a value less than 0.045 are filtered out.
+
+Example usage: `--min-p 0.05`
+
+### Adaptive-P Sampling
+
+-   `--adaptive-target N`: select tokens near this probability (valid range 0.0 to 1.0; negative = disabled)
+-   `--adaptive-decay N`: EMA decay for adaptation; history ≈ 1/(1-decay) tokens (valid range 0.0 - 0.99)
+
+Adaptive-P: Select tokens near a configurable target probability over time.
+
+The adaptive-p sampler transforms the token probability distribution to favor tokens that fall near a user-configurable probability target. Internally, the sampler maintains an exponential moving average of the *ORIGINAL* probabilities of selected tokens at each sampling step. It uses this EMA to compute an adapted target probability at each sampling step, thus maintaining the desired target probability over time. Only mild truncation before this sampler is recommended. It is suggested to apply min-p before adaptive-p as the only other active sampler.
+
+Recommended starting values: `--adaptive-target 0.55 --adaptive-decay 0.9`
+
+For more info, refer to: [llama.cpp#17927](https://github.com/ggml-org/llama.cpp/pull/17927)
+
+### Locally Typical Sampling
+
+-   `--typical N`: Enable locally typical sampling with parameter p (default: 1.0, 1.0 = disabled).
+
+Locally typical sampling promotes the generation of contextually coherent and diverse text by sampling tokens that are typical or expected based on the surrounding context. By setting the parameter p between 0 and 1, you can control the balance between producing text that is locally coherent and diverse. A value closer to 1 will promote more contextually coherent tokens, while a value closer to 0 will promote more diverse tokens. A value equal to 1 disables locally typical sampling.
+
+Example usage: `--typical 0.9`
+
+### Mirostat Sampling
+
+-   `--mirostat N`: Enable Mirostat sampling, controlling perplexity during text generation (default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0).
+-   `--mirostat-lr N`: Set the Mirostat learning rate, parameter eta (default: 0.1).
+-   `--mirostat-ent N`: Set the Mirostat target entropy, parameter tau (default: 5.0).
+
+Mirostat is an algorithm that actively maintains the quality of generated text within a desired range during text generation. It aims to strike a balance between coherence and diversity, avoiding low-quality output caused by excessive repetition (boredom traps) or incoherence (confusion traps).
+
+The `--mirostat-lr` option sets the Mirostat learning rate (eta). The learning rate influences how quickly the algorithm responds to feedback from the generated text. A lower learning rate will result in slower adjustments, while a higher learning rate will make the algorithm more responsive. The default value is `0.1`.
+
+The `--mirostat-ent` option sets the Mirostat target entropy (tau), which represents the desired perplexity value for the generated text. Adjusting the target entropy allows you to control the balance between coherence and diversity in the generated text. A lower value will result in more focused and coherent text, while a higher value will lead to more diverse and potentially less coherent text. The default value is `5.0`.
+
+Example usage: `--mirostat 2 --mirostat-lr 0.05 --mirostat-ent 3.0`
+
+### XTC Sampling
+
+-   `--xtc-probability N`: Sets the chance for token removal (checked once on sampler start) (default: 0.0).
+-   `--xtc-threshold N`: Sets a minimum probability threshold for tokens to be removed (default: 0.1).
+
+Exclude Top Choices (XTC) is a unique sampler that is designed to remove top tokens from consideration and avoid more obvious and repetitive outputs. With a chance of `xtc-probability` it searches for tokens with probabilities of `xtc-threshold` and above, then removes all such tokens except the least probable one.
+
+By removing top tokens XTC can improve the variety of answers, break writing clichés and inhibit repition, since clichés and repeated phrases are usually more likely to appear. By keeping the last token above the threshold, XTC ensures that the answer is still coherent. XTC is meant to be used for creative tasks, but feel free to experiment with different settings for different models.
+
+Being experimental and unique, XTC is disabled by default. The recommended combination of samplers is Min-P followed by XTC on its default settings: `--sampling-seq mx --min-p 0.02 --xtc-probability 0.5`.
+
+Example usage: `--xtc-probability 0.5 --xtc-threshold 0.1`
+
+### Top-nσ Sampling
+
+-   `--top-nsigma N`: Limit the next token selection to a subset of tokens with pre-softmax logits that are within n * σ less than the max logit (default: -1, -1 = disabled).
+
+Top-nσ sampling is a text generation method that selects tokens based on a statistical threshold in pre-softmax logits. It works by only sampling from tokens with logits that are within n * σ of the maximum logit. This method helps maintain a stable sampling space regardless of temperature scaling, allowing it to perform well on reasoning tasks even in high temperatures. Without complex probability manipulation, it efficiently filters tokens directly on the pre-softmax logits. A higher value for top-nsigma (e.g., 5) will take more noisy tokens into consideration, while a lower value (e.g., 1) will focous on the more informative region of the sampling space.
+
+Example usage: `--top-nsigma 1`
+
+### Logit Bias
+
+-   `-l TOKEN_ID(+/-)BIAS, --logit-bias TOKEN_ID(+/-)BIAS`: Modify the likelihood of a token appearing in the generated text completion.
+
+The logit bias option allows you to manually adjust the likelihood of specific tokens appearing in the generated text. By providing a token ID and a positive or negative bias value, you can increase or decrease the probability of that token being generated.
+
+For example, use `--logit-bias 15043+1` to increase the likelihood of the token 'Hello', or `--logit-bias 15043-1` to decrease its likelihood. Using a value of negative infinity, `--logit-bias 15043-inf` ensures that the token `Hello` is never produced.
+
+A more practical use case might be to prevent the generation of `\code{begin}` and `\code{end}` by setting the `\` token (29905) to negative infinity with `-l 29905-inf`. (This is due to the prevalence of LaTeX codes that show up in LLaMA model inference.)
+
+Example usage: `--logit-bias 29905-inf`
+
+### RNG Seed
+
+-   `-s SEED, --seed SEED`: Set the random number generator (RNG) seed (default: -1, -1 = random seed).
+
+The RNG seed is used to initialize the random number generator that influences the text generation process. By setting a specific seed value, you can obtain consistent and reproducible results across multiple runs with the same input and settings. This can be helpful for testing, debugging, or comparing the effects of different options on the generated text to see when they diverge. If the seed is set to a value less than 0, a random seed will be used, which will result in different outputs on each run.
+
+## Performance Tuning and Memory Options
+
+These options help improve the performance and memory usage of the LLaMA models. By adjusting these settings, you can fine-tune the model's behavior to better suit your system's capabilities and achieve optimal performance for your specific use case.
+
+### Number of Threads
+
+-   `-t N, --threads N`: Set the number of threads to use during generation. For optimal performance, it is recommended to set this value to the number of physical CPU cores your system has (as opposed to the logical number of cores). Using the correct number of threads can greatly improve performance.
+-   `-tb N, --threads-batch N`: Set the number of threads to use during batch and prompt processing. In some systems, it is beneficial to use a higher number of threads during batch processing than during generation. If not specified, the number of threads used for batch processing will be the same as the number of threads used for generation.
+
+### Mlock
+
+-   `--mlock`: Lock the model in memory, preventing it from being swapped out when memory-mapped. This can improve performance but trades away some of the advantages of memory-mapping by requiring more RAM to run and potentially slowing down load times as the model loads into RAM.
+
+### No Memory Mapping
+
+-   `--no-mmap`: Do not memory-map the model. By default, models are mapped into memory, which allows the system to load only the necessary parts of the model as needed. However, if the model is larger than your total amount of RAM or if your system is low on available memory, using mmap might increase the risk of pageouts, negatively impacting performance. Disabling mmap results in slower load times but may reduce pageouts if you're not using `--mlock`. Note that if the model is larger than the total amount of RAM, turning off mmap would prevent the model from loading at all.
+
+### NUMA support
+
+-   `--numa distribute`: Pin an equal proportion of the threads to the cores on each NUMA node. This will spread the load amongst all cores on the system, utilitizing all memory channels at the expense of potentially requiring memory to travel over the slow links between nodes.
+-   `--numa isolate`: Pin all threads to the NUMA node that the program starts on. This limits the number of cores and amount of memory that can be used, but guarantees all memory access remains local to the NUMA node.
+-   `--numa numactl`: Pin threads to the CPUMAP that is passed to the program by starting it with the numactl utility. This is the most flexible mode, and allow arbitrary core usage patterns, for example a map that uses all the cores on one NUMA nodes, and just enough cores on a second node to saturate the inter-node memory bus.
+
+ These flags attempt optimizations that help on some systems with non-uniform memory access. This currently consists of one of the above strategies, and disabling prefetch and readahead for mmap. The latter causes mapped pages to be faulted in on first access instead of all at once, and in combination with pinning threads to NUMA nodes, more of the pages end up on the NUMA node where they are used. Note that if the model is already in the system page cache, for example because of a previous run without this option, this will have little effect unless you drop the page cache first. This can be done by rebooting the system or on Linux by writing '3' to '/proc/sys/vm/drop_caches' as root.
+
+### Batch Size
+
+- `-ub N`, `--ubatch-size N`: Physical batch size. This is the maximum number of tokens that may be processed at a time. Increasing this value may improve performance during prompt processing, at the expense of higher memory usage. Default: `512`.
+
+- `-b N`, `--batch-size N`: Logical batch size. Increasing this value above the value of the physical batch size may improve prompt processing performance when using multiple GPUs with pipeline parallelism. Default: `2048`.
+
+### Prompt Caching
+
+-   `--prompt-cache FNAME`: Specify a file to cache the model state after the initial prompt. This can significantly speed up the startup time when you're using longer prompts. The file is created during the first run and is reused and updated in subsequent runs. **Note**: Restoring a cached prompt does not imply restoring the exact state of the session at the point it was saved. So even when specifying a specific seed, you are not guaranteed to get the same sequence of tokens as the original generation.
+
+### Grammars & JSON schemas
+
+-   `--grammar GRAMMAR`, `--grammar-file FILE`: Specify a grammar (defined inline or in a file) to constrain model output to a specific format. For example, you could force the model to output JSON or to speak only in emojis. See the [GBNF guide](../../grammars/README.md) for details on the syntax.
+
+-   `--json-schema SCHEMA`: Specify a [JSON schema](https://json-schema.org/) to constrain model output to (e.g. `{}` for any JSON object, or `{"items": {"type": "string", "minLength": 10, "maxLength": 100}, "minItems": 10}` for a JSON array of strings with size constraints). If a schema uses external `$ref`s, you should use `--grammar "$( python examples/json_schema_to_grammar.py myschema.json )"` instead.
+
+### Quantization
+
+For information about 4-bit quantization, which can significantly improve performance and reduce memory usage, please refer to llama.cpp's primary [README](../../README.md#prepare-and-quantize).
+
+## LoRA (Low-Rank Adaptation) adapters
+
+-   `--lora FNAME`: Optional path to a LoRA adapter to use with scaling of 1.0. Can be mixed with `--lora-scaled` and can be repeated to use multiple adapters.
+-   `--lora-scaled FNAME`: Optional path to a LoRA adapter with user-defined scaling. Can be mixed with `--lora` and can repeated to use multiple adapters.
+
+You can add LoRA adapters using `--lora` or `--lora-scaled`. For example: `--lora my_adapter_1.gguf --lora my_adapter_2.gguf ...` or `--lora-scaled lora_task_A.gguf 0.5 --lora-scaled lora_task_B.gguf 0.5`.
+
+LoRA adapters should be in GGUF format. To convert from Hugging Face format use the `convert-lora-to-gguf.py` script. LoRA adapters are loaded separately and applied during inference - they are not merged with the main model. This means that mmap model loading is fully supported when using LoRA adapters. The old `--lora-base` flag has been removed now that merging is no longer performed.
+
+## Additional Options
+
+These options provide extra functionality and customization when running the LLaMA models:
+
+-   `-h, --help`: Display a help message showing all available options and their default values. This is particularly useful for checking the latest options and default values, as they can change frequently, and the information in this document may become outdated.
+-   `--verbose-prompt`: Print the prompt before generating text.
+-   `--no-display-prompt`: Don't print prompt at generation.
+-   `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used.
+-   `-ts SPLIT, --tensor-split SPLIT`: When using multiple devices this option controls how tensors should be split across devices. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each device should get in order. For example, "3,2" will assign 60% of the data to device 0 and 40% to device 1. By default, the data is split in proportion to VRAM, but this may not be optimal for performance. The list of the devices which are being used is printed on startup and can be different from the device list given by `--list-devices` or e.g. `nvidia-smi`.
+-   `-hfr URL --hf-repo URL`: The url to the Hugging Face model repository. Used in conjunction with `--hf-file` or `-hff`. The model is downloaded and stored in the file provided by `-m` or `--model`. If `-m` is not provided, the model is auto-stored in the path specified by the `LLAMA_CACHE` environment variable  or in an OS-specific local cache.

tools/completion/completion.cpp

@@ -0,0 +1,998 @@
+#include "arg.h"
+#include "common.h"
+#include "console.h"
+#include "log.h"
+#include "sampling.h"
+#include "llama.h"
+#include "chat.h"
+
+#include <cstdio>
+#include <cstring>
+#include <ctime>
+#include <fstream>
+#include <iostream>
+#include <sstream>
+#include <string>
+#include <vector>
+
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
+#include <signal.h>
+#include <unistd.h>
+#elif defined (_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#define NOMINMAX
+#endif
+#include <windows.h>
+#include <signal.h>
+#endif
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+static llama_context           ** g_ctx;
+static llama_model             ** g_model;
+static common_sampler          ** g_smpl;
+static common_params            * g_params;
+static std::vector<llama_token> * g_input_tokens;
+static std::ostringstream       * g_output_ss;
+static std::vector<llama_token> * g_output_tokens;
+static bool is_interacting  = false;
+static bool need_insert_eot = false;
+
+static void print_usage(int argc, char ** argv) {
+    (void) argc;
+
+    LOG("\nexample usage:\n");
+    LOG("\n  text generation:     %s -m your_model.gguf -p \"I believe the meaning of life is\" -n 128 -no-cnv\n", argv[0]);
+    LOG("\n  chat (conversation): %s -m your_model.gguf -sys \"You are a helpful assistant\"\n", argv[0]);
+    LOG("\n");
+}
+
+static bool file_exists(const std::string & path) {
+    std::ifstream f(path.c_str());
+    return f.good();
+}
+
+static bool file_is_empty(const std::string & path) {
+    std::ifstream f;
+    f.exceptions(std::ifstream::failbit | std::ifstream::badbit);
+    f.open(path.c_str(), std::ios::in | std::ios::binary | std::ios::ate);
+    return f.tellg() == 0;
+}
+
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
+static void sigint_handler(int signo) {
+    if (signo == SIGINT) {
+        if (!is_interacting && g_params->interactive) {
+            is_interacting  = true;
+            need_insert_eot = true;
+        } else {
+            console::cleanup();
+            LOG("\n");
+            common_perf_print(*g_ctx, *g_smpl);
+
+            // make sure all logs are flushed
+            LOG("Interrupted by user\n");
+            common_log_pause(common_log_main());
+
+            _exit(130);
+        }
+    }
+}
+#endif
+
+int main(int argc, char ** argv) {
+    common_params params;
+    g_params = &params;
+
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMPLETION, print_usage)) {
+        return 1;
+    }
+
+    common_init();
+
+    auto & sparams = params.sampling;
+
+    // save choice to use color for later
+    // (note for later: this is a slightly awkward choice)
+    console::init(params.simple_io, params.use_color);
+    atexit([]() { console::cleanup(); });
+
+    if (params.embedding) {
+        LOG_ERR("************\n");
+        LOG_ERR("%s: please use the 'embedding' tool for embedding calculations\n", __func__);
+        LOG_ERR("************\n\n");
+
+        return 0;
+    }
+
+    if (params.n_ctx != 0 && params.n_ctx < 8) {
+        LOG_WRN("%s: warning: minimum context size is 8, using minimum size.\n", __func__);
+        params.n_ctx = 8;
+    }
+
+    if (params.rope_freq_base != 0.0) {
+        LOG_WRN("%s: warning: changing RoPE frequency base to %g.\n", __func__, params.rope_freq_base);
+    }
+
+    if (params.rope_freq_scale != 0.0) {
+        LOG_WRN("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
+    }
+
+    LOG_INF("%s: llama backend init\n", __func__);
+
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    llama_model * model = nullptr;
+    llama_context * ctx = nullptr;
+    common_sampler * smpl = nullptr;
+
+    g_model = &model;
+    g_ctx = &ctx;
+    g_smpl = &smpl;
+
+    std::vector<common_chat_msg> chat_msgs;
+
+    // load the model and apply lora adapter, if any
+    LOG_INF("%s: load the model and apply lora adapter, if any\n", __func__);
+
+    auto llama_init = common_init_from_params(params);
+
+    ctx   = llama_init->context();
+    model = llama_init->model();
+    smpl  = llama_init->sampler(0);
+
+    if (ctx == NULL) {
+        LOG_ERR("%s: error: unable to create context\n", __func__);
+        return 1;
+    }
+
+    llama_memory_t mem = llama_get_memory(ctx);
+    const llama_vocab * vocab = llama_model_get_vocab(model);
+
+    // note: the time for chat template initialization is not negligible:
+    auto chat_templates = common_chat_templates_init(model, params.chat_template);
+
+    // start measuring performance timings from here
+    llama_perf_context_reset(ctx);
+
+    LOG_INF("%s: llama threadpool init, n_threads = %d\n", __func__, (int) params.cpuparams.n_threads);
+
+    auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+    if (!cpu_dev) {
+        LOG_ERR("%s: no CPU backend found\n", __func__);
+        return 1;
+    }
+    auto * reg = ggml_backend_dev_backend_reg(cpu_dev);
+    auto * ggml_threadpool_new_fn = (decltype(ggml_threadpool_new) *) ggml_backend_reg_get_proc_address(reg, "ggml_threadpool_new");
+    auto * ggml_threadpool_free_fn = (decltype(ggml_threadpool_free) *) ggml_backend_reg_get_proc_address(reg, "ggml_threadpool_free");
+
+    struct ggml_threadpool_params tpp_batch =
+            ggml_threadpool_params_from_cpu_params(params.cpuparams_batch);
+    struct ggml_threadpool_params tpp =
+            ggml_threadpool_params_from_cpu_params(params.cpuparams);
+
+    if (!set_process_priority(params.cpuparams.priority)) {
+        LOG_ERR("%s: error: failed to set process priority\n", __func__);
+        return 1;
+    }
+
+    struct ggml_threadpool * threadpool_batch = NULL;
+    if (!ggml_threadpool_params_match(&tpp, &tpp_batch)) {
+        threadpool_batch = ggml_threadpool_new_fn(&tpp_batch);
+        if (!threadpool_batch) {
+            LOG_ERR("%s: batch threadpool create failed : n_threads %d\n", __func__, tpp_batch.n_threads);
+            return 1;
+        }
+
+        // start the non-batch threadpool in the paused state
+        tpp.paused = true;
+    }
+
+    struct ggml_threadpool * threadpool = ggml_threadpool_new_fn(&tpp);
+    if (!threadpool) {
+        LOG_ERR("%s: threadpool create failed : n_threads %d\n", __func__, tpp.n_threads);
+        return 1;
+    }
+
+    llama_attach_threadpool(ctx, threadpool, threadpool_batch);
+
+    const int n_ctx_train = llama_model_n_ctx_train(model);
+    const int n_ctx = llama_n_ctx(ctx);
+
+    if (n_ctx > n_ctx_train) {
+        LOG_WRN("%s: model was trained on only %d context tokens (%d specified)\n", __func__, n_ctx_train, n_ctx);
+    }
+
+    // auto enable conversation mode if chat template is available
+    const bool has_chat_template = common_chat_templates_was_explicit(chat_templates.get());
+    if (params.conversation_mode == COMMON_CONVERSATION_MODE_AUTO) {
+        if (has_chat_template) {
+            LOG_INF("%s: chat template is available, enabling conversation mode (disable it with -no-cnv)\n", __func__);
+            params.conversation_mode = COMMON_CONVERSATION_MODE_ENABLED;
+        } else {
+            params.conversation_mode = COMMON_CONVERSATION_MODE_DISABLED;
+        }
+    }
+
+    // in case user force-activate conversation mode (via -cnv) without proper chat template, we show a warning
+    if (params.conversation_mode && !has_chat_template) {
+        LOG_WRN("%s: chat template is not available or is not supported. This may cause the model to output suboptimal responses\n", __func__);
+    }
+
+    // print chat template example in conversation mode
+    if (params.conversation_mode) {
+        if (params.enable_chat_template) {
+            if (!params.prompt.empty() && params.system_prompt.empty()) {
+                LOG_WRN("*** User-specified prompt will pre-start conversation, did you mean to set --system-prompt (-sys) instead?\n");
+            }
+
+            LOG_INF("%s: chat template example:\n%s\n", __func__, common_chat_format_example(chat_templates.get(), params.use_jinja, params.default_template_kwargs).c_str());
+        } else {
+            LOG_INF("%s: in-suffix/prefix is specified, chat template will be disabled\n", __func__);
+        }
+    }
+
+    // print system information
+    {
+        LOG_INF("\n");
+        LOG_INF("%s\n", common_params_get_system_info(params).c_str());
+        LOG_INF("\n");
+    }
+
+    std::string path_session = params.path_prompt_cache;
+    std::vector<llama_token> session_tokens;
+
+    if (!path_session.empty()) {
+        LOG_INF("%s: attempting to load saved session from '%s'\n", __func__, path_session.c_str());
+        if (!file_exists(path_session)) {
+            LOG_INF("%s: session file does not exist, will create.\n", __func__);
+        } else if (file_is_empty(path_session)) {
+            LOG_INF("%s: The session file is empty. A new session will be initialized.\n", __func__);
+        } else {
+            // The file exists and is not empty
+            session_tokens.resize(n_ctx);
+            size_t n_token_count_out = 0;
+            if (!llama_state_load_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.capacity(), &n_token_count_out)) {
+                LOG_ERR("%s: failed to load session file '%s'\n", __func__, path_session.c_str());
+                return 1;
+            }
+            session_tokens.resize(n_token_count_out);
+            LOG_INF("%s: loaded a session with prompt size of %d tokens\n", __func__, (int)session_tokens.size());
+        }
+    }
+
+    const bool add_bos = llama_vocab_get_add_bos(vocab) && !params.use_jinja;
+    if (!llama_model_has_encoder(model)) {
+        GGML_ASSERT(!llama_vocab_get_add_eos(vocab));
+    }
+
+    LOG_DBG("n_ctx: %d, add_bos: %d\n", n_ctx, add_bos);
+
+    std::vector<llama_token> embd_inp;
+
+    bool waiting_for_first_input = false;
+    auto chat_add_and_format = [&chat_msgs, &chat_templates](const std::string & role, const std::string & content) {
+        common_chat_msg new_msg;
+        new_msg.role = role;
+        new_msg.content = content;
+        auto formatted = common_chat_format_single(chat_templates.get(), chat_msgs, new_msg, role == "user", g_params->use_jinja);
+        chat_msgs.push_back(new_msg);
+        LOG_DBG("formatted: '%s'\n", formatted.c_str());
+        return formatted;
+    };
+
+    std::string prompt;
+    {
+        if (params.conversation_mode && params.enable_chat_template) {
+            if (!params.system_prompt.empty()) {
+                // format the system prompt (will use template default if empty)
+                chat_add_and_format("system", params.system_prompt);
+            }
+
+            if (!params.prompt.empty()) {
+                // format and append the user prompt
+                chat_add_and_format("user", params.prompt);
+            } else {
+                waiting_for_first_input = true;
+            }
+
+            if (!params.system_prompt.empty() || !params.prompt.empty()) {
+                common_chat_templates_inputs inputs;
+                inputs.use_jinja = g_params->use_jinja;
+                inputs.messages = chat_msgs;
+                inputs.add_generation_prompt = !params.prompt.empty();
+
+                prompt = common_chat_templates_apply(chat_templates.get(), inputs).prompt;
+            }
+        } else {
+            // otherwise use the prompt as is
+            prompt = params.prompt;
+        }
+
+        if (params.interactive_first || !prompt.empty() || session_tokens.empty()) {
+            LOG_DBG("tokenize the prompt\n");
+            embd_inp = common_tokenize(ctx, prompt, true, true);
+        } else {
+            LOG_DBG("use session tokens\n");
+            embd_inp = session_tokens;
+        }
+
+        LOG_DBG("prompt: \"%s\"\n", prompt.c_str());
+        LOG_DBG("tokens: %s\n", string_from(ctx, embd_inp).c_str());
+    }
+
+    // Should not run without any tokens
+    if (!waiting_for_first_input && embd_inp.empty()) {
+        if (add_bos) {
+            embd_inp.push_back(llama_vocab_bos(vocab));
+            LOG_WRN("embd_inp was considered empty and bos was added: %s\n", string_from(ctx, embd_inp).c_str());
+        } else {
+            LOG_ERR("input is empty\n");
+            return -1;
+        }
+    }
+
+    // Tokenize negative prompt
+    if ((int) embd_inp.size() > n_ctx - 4) {
+        LOG_ERR("%s: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4);
+        return 1;
+    }
+
+    // debug message about similarity of saved session, if applicable
+    size_t n_matching_session_tokens = 0;
+    if (!session_tokens.empty()) {
+        for (llama_token id : session_tokens) {
+            if (n_matching_session_tokens >= embd_inp.size() || id != embd_inp[n_matching_session_tokens]) {
+                break;
+            }
+            n_matching_session_tokens++;
+        }
+        if (params.prompt.empty() && n_matching_session_tokens == embd_inp.size()) {
+            LOG_INF("%s: using full prompt from session file\n", __func__);
+        } else if (n_matching_session_tokens >= embd_inp.size()) {
+            LOG_INF("%s: session file has exact match for prompt!\n", __func__);
+        } else if (n_matching_session_tokens < (embd_inp.size() / 2)) {
+            LOG_WRN("%s: session file has low similarity to prompt (%zu / %zu tokens); will mostly be reevaluated\n",
+                    __func__, n_matching_session_tokens, embd_inp.size());
+        } else {
+            LOG_INF("%s: session file matches %zu / %zu tokens of prompt\n",
+                    __func__, n_matching_session_tokens, embd_inp.size());
+        }
+
+        // remove any "future" tokens that we might have inherited from the previous session
+        if (!llama_memory_seq_rm(mem, -1, n_matching_session_tokens, -1)) {
+            LOG_INF("%s: unable to resuse common prefix\n", __func__);
+            n_matching_session_tokens = 0;
+            llama_memory_seq_rm(mem, -1, -1, -1);
+        }
+    }
+
+    LOG_DBG("recalculate the cached logits (check): embd_inp.size() %zu, n_matching_session_tokens %zu, embd_inp.size() %zu, session_tokens.size() %zu\n",
+         embd_inp.size(), n_matching_session_tokens, embd_inp.size(), session_tokens.size());
+
+    // if we will use the cache for the full prompt without reaching the end of the cache, force
+    // reevaluation of the last token to recalculate the cached logits
+    if (!embd_inp.empty() && n_matching_session_tokens == embd_inp.size() && session_tokens.size() > embd_inp.size()) {
+        LOG_DBG("recalculate the cached logits (do): session_tokens.resize( %zu )\n", embd_inp.size() - 1);
+
+        session_tokens.resize(embd_inp.size() - 1);
+    }
+
+    // number of tokens to keep when resetting context
+    if (params.n_keep < 0 || params.n_keep > (int) embd_inp.size()) {
+        params.n_keep = (int)embd_inp.size();
+    } else {
+        params.n_keep += add_bos; // always keep the BOS token
+    }
+
+    if (params.conversation_mode) {
+        if (params.single_turn && !params.prompt.empty()) {
+            params.interactive = false;
+            params.interactive_first = false;
+        } else {
+            params.interactive_first = true;
+        }
+    }
+
+    // enable interactive mode if interactive start is specified
+    if (params.interactive_first) {
+        params.interactive = true;
+    }
+
+    if (params.verbose_prompt) {
+        LOG_INF("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
+        LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
+        for (int i = 0; i < (int) embd_inp.size(); i++) {
+            LOG_INF("%6d -> '%s'\n", embd_inp[i], common_token_to_piece(ctx, embd_inp[i]).c_str());
+        }
+
+        if (params.n_keep > add_bos) {
+            LOG_INF("%s: static prompt based on n_keep: '", __func__);
+            for (int i = 0; i < params.n_keep; i++) {
+                LOG_CNT("%s", common_token_to_piece(ctx, embd_inp[i]).c_str());
+            }
+            LOG_CNT("'\n");
+        }
+        LOG_INF("\n");
+    }
+
+    // ctrl+C handling
+    {
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
+        struct sigaction sigint_action;
+        sigint_action.sa_handler = sigint_handler;
+        sigemptyset (&sigint_action.sa_mask);
+        sigint_action.sa_flags = 0;
+        sigaction(SIGINT, &sigint_action, NULL);
+#elif defined (_WIN32)
+        auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
+            return (ctrl_type == CTRL_C_EVENT) ? (sigint_handler(SIGINT), true) : false;
+        };
+        SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
+#endif
+    }
+
+    if (params.interactive) {
+        LOG_INF("%s: interactive mode on.\n", __func__);
+
+        if (!params.antiprompt.empty()) {
+            for (const auto & antiprompt : params.antiprompt) {
+                LOG_INF("Reverse prompt: '%s'\n", antiprompt.c_str());
+                if (params.verbose_prompt) {
+                    auto tmp = common_tokenize(ctx, antiprompt, false, true);
+                    for (int i = 0; i < (int) tmp.size(); i++) {
+                        LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx, tmp[i]).c_str());
+                    }
+                }
+            }
+        }
+
+        if (params.input_prefix_bos) {
+            LOG_INF("Input prefix with BOS\n");
+        }
+
+        if (!params.input_prefix.empty()) {
+            LOG_INF("Input prefix: '%s'\n", params.input_prefix.c_str());
+            if (params.verbose_prompt) {
+                auto tmp = common_tokenize(ctx, params.input_prefix, true, true);
+                for (int i = 0; i < (int) tmp.size(); i++) {
+                    LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx, tmp[i]).c_str());
+                }
+            }
+        }
+
+        if (!params.input_suffix.empty()) {
+            LOG_INF("Input suffix: '%s'\n", params.input_suffix.c_str());
+            if (params.verbose_prompt) {
+                auto tmp = common_tokenize(ctx, params.input_suffix, false, true);
+                for (int i = 0; i < (int) tmp.size(); i++) {
+                    LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx, tmp[i]).c_str());
+                }
+            }
+        }
+    }
+
+    LOG_INF("sampler seed: %u\n",     common_sampler_get_seed(smpl));
+    LOG_INF("sampler params: \n%s\n", sparams.print().c_str());
+    LOG_INF("sampler chain: %s\n",    common_sampler_print(smpl).c_str());
+
+    LOG_INF("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
+
+    // group-attention state
+    // number of grouped KV tokens so far (used only if params.grp_attn_n > 1)
+    int ga_i = 0;
+
+    const int ga_n = params.grp_attn_n;
+    const int ga_w = params.grp_attn_w;
+
+    if (ga_n != 1) {
+        GGML_ASSERT(ga_n > 0                    && "grp_attn_n must be positive");                     // NOLINT
+        GGML_ASSERT(ga_w % ga_n == 0            && "grp_attn_w must be a multiple of grp_attn_n");     // NOLINT
+      //GGML_ASSERT(n_ctx_train % ga_w == 0     && "n_ctx_train must be a multiple of grp_attn_w");    // NOLINT
+      //GGML_ASSERT(n_ctx >= n_ctx_train * ga_n && "n_ctx must be at least n_ctx_train * grp_attn_n"); // NOLINT
+        LOG_INF("self-extend: n_ctx_train = %d, grp_attn_n = %d, grp_attn_w = %d\n", n_ctx_train, ga_n, ga_w);
+    }
+    LOG_INF("\n");
+
+    if (params.interactive) {
+        const char * control_message;
+        if (params.multiline_input) {
+            control_message = " - To return control to the AI, end your input with '\\'.\n"
+                              " - To return control without starting a new line, end your input with '/'.\n";
+        } else {
+            control_message = " - Press Return to return control to the AI.\n"
+                              " - To return control without starting a new line, end your input with '/'.\n"
+                              " - If you want to submit another line, end your input with '\\'.\n";
+        }
+        LOG_INF("== Running in interactive mode. ==\n");
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
+        LOG_INF(       " - Press Ctrl+C to interject at any time.\n");
+#endif
+        LOG_INF(       "%s", control_message);
+        if (params.conversation_mode && params.enable_chat_template && params.system_prompt.empty()) {
+            LOG_INF(   " - Not using system message. To change it, set a different value via -sys PROMPT\n");
+        }
+        LOG_INF("\n");
+
+        is_interacting = params.interactive_first;
+    }
+
+    bool is_antiprompt        = false;
+    bool input_echo           = true;
+    bool display              = true;
+    bool need_to_save_session = !path_session.empty() && n_matching_session_tokens < embd_inp.size();
+
+    int n_past             = 0;
+    int n_remain           = params.n_predict;
+    int n_consumed         = 0;
+    int n_session_consumed = 0;
+
+    std::vector<int>   input_tokens;  g_input_tokens  = &input_tokens;
+    std::vector<int>   output_tokens; g_output_tokens = &output_tokens;
+    std::ostringstream output_ss;     g_output_ss     = &output_ss;
+    std::ostringstream assistant_ss; // for storing current assistant message, used in conversation mode
+
+    // the first thing we will do is to output the prompt, so set color accordingly
+    console::set_display(DISPLAY_TYPE_PROMPT);
+    display = params.display_prompt;
+
+    std::vector<llama_token> embd;
+
+    // single-token antiprompts
+    std::vector<llama_token> antiprompt_token;
+
+    for (const std::string & antiprompt : params.antiprompt) {
+        auto ids = ::common_tokenize(ctx, antiprompt, false, true);
+        if (ids.size() == 1) {
+            antiprompt_token.push_back(ids[0]);
+        }
+    }
+
+    if (llama_model_has_encoder(model)) {
+        int enc_input_size = embd_inp.size();
+        llama_token * enc_input_buf = embd_inp.data();
+
+        if (llama_encode(ctx, llama_batch_get_one(enc_input_buf, enc_input_size))) {
+            LOG_ERR("%s : failed to eval\n", __func__);
+            return 1;
+        }
+
+        llama_token decoder_start_token_id = llama_model_decoder_start_token(model);
+        if (decoder_start_token_id == LLAMA_TOKEN_NULL) {
+            decoder_start_token_id = llama_vocab_bos(vocab);
+        }
+
+        embd_inp.clear();
+        embd_inp.push_back(decoder_start_token_id);
+    }
+
+    while ((n_remain != 0 && !is_antiprompt) || params.interactive) {
+        // predict
+        if (!embd.empty()) {
+            // Note: (n_ctx - 4) here is to match the logic for commandline prompt handling via
+            // --prompt or --file which uses the same value.
+            int max_embd_size = n_ctx - 4;
+
+            // Ensure the input doesn't exceed the context size by truncating embd if necessary.
+            if ((int) embd.size() > max_embd_size) {
+                const int skipped_tokens = (int) embd.size() - max_embd_size;
+                embd.resize(max_embd_size);
+
+                console::set_display(DISPLAY_TYPE_ERROR);
+                LOG_WRN("<<input too long: skipped %d token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
+                console::set_display(DISPLAY_TYPE_RESET);
+            }
+
+            if (ga_n == 1) {
+                // infinite text generation via context shifting
+                // if we run out of context:
+                // - take the n_keep first tokens from the original prompt (via n_past)
+                // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
+
+                if (n_past + (int) embd.size() >= n_ctx) {
+                    if (!params.ctx_shift){
+                        LOG_WRN("\n\n%s: context full and context shift is disabled => stopping\n", __func__);
+                        break;
+                    }
+
+                    if (params.n_predict == -2) {
+                        LOG_WRN("\n\n%s: context full and n_predict == %d => stopping\n", __func__, params.n_predict);
+                        break;
+                    }
+
+                    const int n_left    = n_past - params.n_keep;
+                    const int n_discard = n_left/2;
+
+                    LOG_DBG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
+                            n_past, n_left, n_ctx, params.n_keep, n_discard);
+
+                    llama_memory_seq_rm (mem, 0, params.n_keep            , params.n_keep + n_discard);
+                    llama_memory_seq_add(mem, 0, params.n_keep + n_discard, n_past, -n_discard);
+
+                    n_past -= n_discard;
+
+                    LOG_DBG("after swap: n_past = %d\n", n_past);
+
+                    LOG_DBG("embd: %s\n", string_from(ctx, embd).c_str());
+
+                    LOG_DBG("clear session path\n");
+                    path_session.clear();
+                }
+            } else {
+                // context extension via Self-Extend
+                while (n_past >= ga_i + ga_w) {
+                    const int ib = (ga_n*ga_i)/ga_w;
+                    const int bd = (ga_w/ga_n)*(ga_n - 1);
+                    const int dd = (ga_w/ga_n) - ib*bd - ga_w;
+
+                    LOG_DBG("\n");
+                    LOG_DBG("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", ga_i, n_past, ib*bd, ga_i + ib*bd, n_past + ib*bd);
+                    LOG_DBG("div:   [%6d, %6d] / %6d -> [%6d, %6d]\n", ga_i + ib*bd, ga_i + ib*bd + ga_w, ga_n, (ga_i + ib*bd)/ga_n, (ga_i + ib*bd + ga_w)/ga_n);
+                    LOG_DBG("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", ga_i + ib*bd + ga_w, n_past + ib*bd, dd, ga_i + ib*bd + ga_w + dd, n_past + ib*bd + dd);
+
+                    llama_memory_seq_add(mem, 0, ga_i,                n_past,              ib*bd);
+                    llama_memory_seq_div(mem, 0, ga_i + ib*bd,        ga_i + ib*bd + ga_w, ga_n);
+                    llama_memory_seq_add(mem, 0, ga_i + ib*bd + ga_w, n_past + ib*bd,      dd);
+
+                    n_past -= bd;
+
+                    ga_i += ga_w/ga_n;
+
+                    LOG_DBG("\nn_past_old = %d, n_past = %d, ga_i = %d\n\n", n_past + bd, n_past, ga_i);
+                }
+            }
+
+            // try to reuse a matching prefix from the loaded session instead of re-eval (via n_past)
+            if (n_session_consumed < (int) session_tokens.size()) {
+                size_t i = 0;
+                for ( ; i < embd.size(); i++) {
+                    if (embd[i] != session_tokens[n_session_consumed]) {
+                        session_tokens.resize(n_session_consumed);
+                        break;
+                    }
+
+                    n_past++;
+                    n_session_consumed++;
+
+                    if (n_session_consumed >= (int) session_tokens.size()) {
+                        ++i;
+                        break;
+                    }
+                }
+                if (i > 0) {
+                    embd.erase(embd.begin(), embd.begin() + i);
+                }
+            }
+
+            for (int i = 0; i < (int) embd.size(); i += params.n_batch) {
+                int n_eval = (int) embd.size() - i;
+                if (n_eval > params.n_batch) {
+                    n_eval = params.n_batch;
+                }
+
+                LOG_DBG("eval: %s\n", string_from(ctx, embd).c_str());
+
+                if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval))) {
+                    LOG_ERR("%s : failed to eval\n", __func__);
+                    return 1;
+                }
+
+                n_past += n_eval;
+
+                LOG_DBG("n_past = %d\n", n_past);
+                // Display total tokens alongside total time
+                if (params.n_print > 0 && n_past % params.n_print == 0) {
+                    LOG_DBG("\n\033[31mTokens consumed so far = %d / %d \033[0m\n", n_past, n_ctx);
+                }
+            }
+
+            if (!embd.empty() && !path_session.empty()) {
+                session_tokens.insert(session_tokens.end(), embd.begin(), embd.end());
+                n_session_consumed = session_tokens.size();
+            }
+        }
+
+        embd.clear();
+
+        if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
+            // optionally save the session on first sample (for faster prompt loading next time)
+            if (!path_session.empty() && need_to_save_session && !params.prompt_cache_ro) {
+                need_to_save_session = false;
+                llama_state_save_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size());
+
+                LOG_DBG("saved session to %s\n", path_session.c_str());
+            }
+
+            const llama_token id = common_sampler_sample(smpl, ctx, -1);
+
+            common_sampler_accept(smpl, id, /* accept_grammar= */ true);
+
+            // LOG_DBG("last: %s\n", string_from(ctx, smpl->prev.to_vector()).c_str());
+
+            embd.push_back(id);
+
+            if (params.conversation_mode && !waiting_for_first_input && !llama_vocab_is_eog(vocab, id)) {
+                assistant_ss << common_token_to_piece(ctx, id, false);
+            }
+
+            // echo this to console
+            input_echo = true;
+
+            // decrement remaining sampling budget
+            --n_remain;
+
+            LOG_DBG("n_remain: %d\n", n_remain);
+        } else {
+            // some user input remains from prompt or interaction, forward it to processing
+            LOG_DBG("embd_inp.size(): %d, n_consumed: %d\n", (int) embd_inp.size(), n_consumed);
+            while ((int) embd_inp.size() > n_consumed) {
+                embd.push_back(embd_inp[n_consumed]);
+
+                // push the prompt in the sampling context in order to apply repetition penalties later
+                // for the prompt, we don't apply grammar rules
+                common_sampler_accept(smpl, embd_inp[n_consumed], /* accept_grammar= */ false);
+
+                ++n_consumed;
+                if ((int) embd.size() >= params.n_batch) {
+                    break;
+                }
+            }
+        }
+
+        // display text
+        if (input_echo && display) {
+            for (auto id : embd) {
+                const std::string token_str = common_token_to_piece(ctx, id, params.special);
+
+                // Console/Stream Output
+                LOG("%s", token_str.c_str());
+
+                // Record Displayed Tokens To Log
+                // Note: Generated tokens are created one by one hence this check
+                if (embd.size() > 1) {
+                    // Incoming Requested Tokens
+                    input_tokens.push_back(id);
+                } else {
+                    // Outgoing Generated Tokens
+                    output_tokens.push_back(id);
+                    output_ss << token_str;
+                }
+            }
+        }
+
+        // reset color to default if there is no pending user input
+        if (input_echo && (int) embd_inp.size() == n_consumed) {
+            console::set_display(DISPLAY_TYPE_RESET);
+            display = true;
+        }
+
+        // if not currently processing queued inputs;
+        if ((int) embd_inp.size() <= n_consumed) {
+            // check for reverse prompt in the last n_prev tokens
+            if (!params.antiprompt.empty()) {
+                const int n_prev = 32;
+                const std::string last_output = common_sampler_prev_str(smpl, ctx, n_prev);
+
+                is_antiprompt = false;
+                // Check if each of the reverse prompts appears at the end of the output.
+                // If we're not running interactively, the reverse prompt might be tokenized with some following characters
+                // so we'll compensate for that by widening the search window a bit.
+                for (std::string & antiprompt : params.antiprompt) {
+                    size_t extra_padding = params.interactive ? 0 : 2;
+                    size_t search_start_pos = last_output.length() > static_cast<size_t>(antiprompt.length() + extra_padding)
+                        ? last_output.length() - static_cast<size_t>(antiprompt.length() + extra_padding)
+                        : 0;
+
+                    if (last_output.find(antiprompt, search_start_pos) != std::string::npos) {
+                        if (params.interactive) {
+                            is_interacting = true;
+                        }
+                        is_antiprompt = true;
+                        break;
+                    }
+                }
+
+                // check for reverse prompt using special tokens
+                // avoid calling common_sampler_last() if last_output is empty
+                if (!last_output.empty()) {
+                    llama_token last_token = common_sampler_last(smpl);
+                    for (auto token : antiprompt_token) {
+                        if (token == last_token) {
+                            if (params.interactive) {
+                                is_interacting = true;
+                            }
+                            is_antiprompt = true;
+                            break;
+                        }
+                    }
+                }
+
+                if (is_antiprompt) {
+                    LOG_DBG("found antiprompt: %s\n", last_output.c_str());
+                }
+            }
+
+            // deal with end of generation tokens in interactive mode
+            if (!waiting_for_first_input && llama_vocab_is_eog(vocab, common_sampler_last(smpl))) {
+                LOG_DBG("found an EOG token\n");
+
+                if (params.interactive) {
+                    if (!params.antiprompt.empty()) {
+                        // tokenize and inject first reverse prompt
+                        const auto first_antiprompt = common_tokenize(ctx, params.antiprompt.front(), false, true);
+                        embd_inp.insert(embd_inp.end(), first_antiprompt.begin(), first_antiprompt.end());
+                        is_antiprompt = true;
+                    }
+
+                    if (params.enable_chat_template) {
+                        chat_add_and_format("assistant", assistant_ss.str());
+                    }
+                    is_interacting = true;
+                    LOG("\n");
+                }
+            }
+
+            if (params.conversation_mode && !waiting_for_first_input) {
+                if (!prompt.empty()) {
+                    prompt.clear();
+                    is_interacting = false;
+                }
+            }
+
+            if ((n_past > 0 || waiting_for_first_input) && is_interacting) {
+                LOG_DBG("waiting for user input\n");
+
+                if (params.conversation_mode) {
+                    LOG("\n> ");
+                }
+
+                if (params.input_prefix_bos) {
+                    LOG_DBG("adding input prefix BOS token\n");
+                    embd_inp.push_back(llama_vocab_bos(vocab));
+                }
+
+                std::string buffer;
+                if (!params.input_prefix.empty() && !params.conversation_mode) {
+                    LOG_DBG("appending input prefix: '%s'\n", params.input_prefix.c_str());
+                    LOG("%s", params.input_prefix.c_str());
+                }
+
+                // color user input only
+                console::set_display(DISPLAY_TYPE_USER_INPUT);
+                display = params.display_prompt;
+
+                std::string line;
+                bool another_line = true;
+                do {
+                    another_line = console::readline(line, params.multiline_input);
+                    buffer += line;
+                } while (another_line);
+
+                // done taking input, reset color
+                console::set_display(DISPLAY_TYPE_RESET);
+                display = true;
+
+                if (buffer.empty()) { // Ctrl+D on empty line exits
+                    LOG("EOF by user\n");
+                    break;
+                }
+
+                if (buffer.back() == '\n') {
+                    // Implement #587:
+                    // If the user wants the text to end in a newline,
+                    // this should be accomplished by explicitly adding a newline by using \ followed by return,
+                    // then returning control by pressing return again.
+                    buffer.pop_back();
+                }
+
+                if (buffer.empty()) { // Enter key on empty line lets the user pass control back
+                    LOG_DBG("empty line, passing control back\n");
+                } else { // Add tokens to embd only if the input buffer is non-empty
+                    // append input suffix if any
+                    if (!params.input_suffix.empty() && !params.conversation_mode) {
+                        LOG_DBG("appending input suffix: '%s'\n", params.input_suffix.c_str());
+                        LOG("%s", params.input_suffix.c_str());
+                    }
+
+                    LOG_DBG("buffer: '%s'\n", buffer.c_str());
+
+                    const size_t original_size = embd_inp.size();
+
+                    if (params.escape) {
+                        string_process_escapes(buffer);
+                    }
+
+                    bool format_chat = params.conversation_mode && params.enable_chat_template;
+                    std::string user_inp = format_chat
+                        ? chat_add_and_format("user", std::move(buffer))
+                        : std::move(buffer);
+                    // TODO: one inconvenient of current chat template implementation is that we can't distinguish between user input and special tokens (prefix/postfix)
+                    const auto line_pfx = common_tokenize(ctx, params.input_prefix, false, true);
+                    const auto line_inp = common_tokenize(ctx, user_inp,            false, format_chat);
+                    const auto line_sfx = common_tokenize(ctx, params.input_suffix, false, true);
+
+                    LOG_DBG("input tokens: %s\n", string_from(ctx, line_inp).c_str());
+
+                    // if user stop generation mid-way, we must add EOT to finish model's last response
+                    if (need_insert_eot && format_chat) {
+                        llama_token eot = llama_vocab_eot(vocab);
+                        embd_inp.push_back(eot == LLAMA_TOKEN_NULL ? llama_vocab_eos(vocab) : eot);
+                        need_insert_eot = false;
+                    }
+
+                    embd_inp.insert(embd_inp.end(), line_pfx.begin(), line_pfx.end());
+                    embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end());
+                    embd_inp.insert(embd_inp.end(), line_sfx.begin(), line_sfx.end());
+
+                    if (params.verbose_prompt) {
+                        LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size() - original_size);
+                    }
+
+                    for (size_t i = original_size; i < embd_inp.size(); ++i) {
+                        const llama_token token = embd_inp[i];
+                        const std::string token_str = common_token_to_piece(ctx, token);
+                        output_tokens.push_back(token);
+                        output_ss << token_str;
+
+                        if (params.verbose_prompt) {
+                            LOG_INF("%6d -> '%s'\n", token, token_str.c_str());
+                        }
+                    }
+
+                    // reset assistant message
+                    assistant_ss.str("");
+
+                    n_remain -= line_inp.size();
+                    LOG_DBG("n_remain: %d\n", n_remain);
+                }
+
+                input_echo = false; // do not echo this again
+            }
+
+            if (n_past > 0 || waiting_for_first_input) {
+                if (is_interacting) {
+                    common_sampler_reset(smpl);
+                }
+                is_interacting = false;
+
+                if (waiting_for_first_input && params.single_turn) {
+                    params.interactive = false;
+                    params.interactive_first = false;
+                }
+                waiting_for_first_input = false;
+            }
+        }
+
+        // end of generation
+        if (!embd.empty() && llama_vocab_is_eog(vocab, embd.back()) && !(params.interactive)) {
+            LOG(" [end of text]\n");
+            break;
+        }
+
+        // In interactive mode, respect the maximum number of tokens and drop back to user input when reached.
+        // We skip this logic when n_predict == -1 (infinite) or -2 (stop at context size).
+        if (params.interactive && n_remain <= 0 && params.n_predict >= 0) {
+            n_remain = params.n_predict;
+            is_interacting = true;
+        }
+    }
+
+    if (!path_session.empty() && params.prompt_cache_all && !params.prompt_cache_ro) {
+        LOG("\n%s: saving final output to session file '%s'\n", __func__, path_session.c_str());
+        llama_state_save_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size());
+    }
+
+    LOG("\n\n");
+    common_perf_print(ctx, smpl);
+
+    llama_backend_free();
+
+    ggml_threadpool_free_fn(threadpool);
+    ggml_threadpool_free_fn(threadpool_batch);
+
+    return 0;
+}

tools/cvector-generator/CMakeLists.txt

@@ -0,0 +1,8 @@
+set(TARGET llama-cvector-generator)
+add_executable(${TARGET} cvector-generator.cpp pca.hpp)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_17)
+
+if(LLAMA_TOOLS_INSTALL)
+    install(TARGETS ${TARGET} RUNTIME)
+endif()

tools/cvector-generator/README.md

@@ -0,0 +1,45 @@
+# cvector-generator
+
+This example demonstrates how to generate a control vector using gguf models.
+
+Related PRs:
+- [Add support for control vectors](https://github.com/ggml-org/llama.cpp/pull/5970)
+- (Issue) [Generate control vector using llama.cpp](https://github.com/ggml-org/llama.cpp/issues/6880)
+- [Add cvector-generator example](https://github.com/ggml-org/llama.cpp/pull/7514)
+
+## Examples
+
+```sh
+# CPU only
+./cvector-generator -m ./llama-3.Q4_K_M.gguf
+
+# With GPU
+./cvector-generator -m ./llama-3.Q4_K_M.gguf -ngl 99
+
+# With advanced options
+./cvector-generator -m ./llama-3.Q4_K_M.gguf -ngl 99 --pca-iter 2000 --pca-batch 100
+
+# Using mean value instead of PCA
+./cvector-generator -m ./llama-3.Q4_K_M.gguf --method mean
+
+# To see help message
+./cvector-generator -h
+# Then, have a look at "cvector" section
+```
+
+## Tips and tricks
+
+If you have multiple lines per prompt, you can escape the newline character (change it to `\n`). For example:
+
+```
+<|im_start|>system\nAct like a person who is extremely happy.<|im_end|>
+<|im_start|>system\nYou are in a very good mood today<|im_end|>
+```
+
+Example to use output file with `llama-cli`:
+
+(Tips: The control vector works better when apply to layers higher than 10)
+
+```sh
+./llama-cli -m ./llama-3.Q4_K_M.gguf -p "<|start_header_id|>system<|end_header_id|>\n\nYou are a helpful assistant<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nSing a song<|im_end|><|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n" --special --control-vector-scaled ./control_vector.gguf 0.8 --control-vector-layer-range 10 31
+```

tools/cvector-generator/completions.txt

@@ -0,0 +1,582 @@
+
+That game
+I can see
+Hmm, this
+I can relate to
+Who is
+I understand the
+Ugh,
+What the hell was
+Hey, did anyone
+Although
+Thank you for choosing
+What are you
+Oh w
+How dare you open
+It was my pleasure
+I'm hon
+I appreciate that you
+Are you k
+Whoever left this
+It's always
+Ew,
+Hey, I l
+Hello? Is someone
+I understand that
+That poem
+Aww, poor
+Hey, it
+Alright, who
+I didn't
+Well, life
+The document
+Oh no, this
+I'm concerned
+Hello, this is
+This art
+Hmm, this drink
+Hi there!
+It seems
+Is
+Good
+I can't
+Ex
+Who are
+I can see that
+Wow,
+Today is a
+Hey friend
+Sometimes friends
+Oh, this old
+The weather outside
+This place is sur
+I appreciate your input
+Thank you for the
+Look at
+I'm disappoint
+To my
+How dare you
+That's an
+This piece of art
+Eww
+This park is
+This is incredible
+Oh no, someone
+Exc
+Well, it'
+I warned
+Hey, I understand
+Hey, I saw
+How dare you go
+What the he
+Hey
+It's
+Hello? Hello?
+It
+Oh no!
+This is the perfect
+Good morning,
+Oh no, there
+It's so
+Yeah
+Uh,
+Hello everyone
+Who turned off
+The weather
+Who'
+Hey, this
+Wait,
+Eww, gross
+Excuse
+It seems like you
+Thank you so
+What happened?
+Oh my g
+I am deeply sad
+I war
+Okay, let'
+Hey, that
+That was a beautiful
+Oh no! That
+What happened
+Hey there
+The artist'
+What?!
+Hey, it'
+I am disappoint
+It seems like
+Oh no! The
+This park is a
+If you
+Yes! I did
+It sounds
+What
+Who is it
+Hmm, that
+That's strange
+Yeah, that was
+That's interesting
+This park
+What the hell
+Who is that
+I feel like my
+Oh well
+What the hell is
+Hello? Hello
+To my dearest
+Bless you!\"
+Thank you for
+Oh, looks like
+Can you please
+This place is
+Eww, what
+Bless you
+Is everything
+Hey, I just
+Whoever left these
+Well, that'
+I feel
+Hey, do you
+It's sad
+Oh no, it
+Hey, that'
+Oh my god,
+Thank you,
+Hello little one,
+I apolog
+Hey team, I
+How dare you read
+Who is this and
+Whoever left
+Hi there! W
+A
+If you have
+I was
+U
+Bless
+Well, this
+Oh, I'
+It's a
+Eww,
+Is everything okay?
+Oh, I
+Hello, can you
+Al
+That was a great
+What are
+I understand that not
+Oh no, not
+Who is it?\"
+Hey, can we
+Whoever is taking
+I would love to
+Hey, I noticed
+Hey, could
+I understand that there
+Hello?
+D
+Oh man, I
+Thank you so much
+Oh no, my
+Dear [Name
+Uh
+I remember
+Hey, who
+Well, it
+Are you
+I understand that it
+Hey, is
+I would
+Who is this
+Excuse me
+Alright
+I am thrilled
+Sometimes friends have
+Who the
+It's interesting
+I would love
+E
+Hello? Is anyone
+Well, this is
+This place
+Well,
+I warned you
+Hey, watch where
+Oh my
+That'
+Sometimes friends have different
+I understand that everyone
+What?
+What do these notes
+I can relate
+I'm not
+I understand
+To my dear
+Guys
+Well
+Hey, I appreciate
+Wow, what
+Dear
+That melody
+Who the hell
+Today is
+Hello little
+Wow, look
+That's great
+Love is never wrong
+I'm having
+Whoa, did
+Ugh
+Can you please provide
+I miss you,
+I feel uncom
+I know
+Ugh, this
+Hey, watch
+Oh great, a
+I didn
+Okay
+That game of char
+Oh
+I appreciate
+Who's there
+I am so
+Oh great, someone
+Hey, could you
+I remember wondering
+Wait, what?
+What do
+Hello? Can
+Hey there,
+That game of
+This is incred
+Oh my gosh
+Oh great, f
+I appreciate your
+It sounds like
+What the heck
+Okay, I understand
+Ew
+I understand that this
+Uh, hi
+Hi everyone!
+What the hell?
+Thank you for your
+Oh no, the
+Wow, I
+Who turned
+Dear [
+Whoever
+This is a
+Whoa, he
+What in the world
+Although the physical
+Hello, who is
+That's amaz
+Hey, I know
+Okay, that
+Hi everyone
+Hey, is everything
+I understand your fr
+Oh no, poor
+Oh, look
+Good morning
+Ew, gross
+Oh no, did
+Look at the family
+Hey team
+Yes!
+Hey, can I
+Okay, that'
+It's great
+Love is
+Hey, what
+Good morning, world
+Who is it?
+That poem really reson
+I
+That's
+I understand the task
+Gu
+Hello? Who'
+This postcard is
+Whoa,
+Oh, that
+I understand that I
+Whoever is
+Hello? Who is
+I'm really
+Wow, this
+Can
+This artwork really
+This is a shame
+I miss you too
+Who are you?
+Today is a difficult
+Hey, just
+Are you okay
+I am
+Hi,
+Wow, that
+Hey there! Can
+Okay, stay
+Oh great, just
+Yeah,
+Hello? Can you
+Oh, looks
+Thank you for sharing
+I'm glad
+Hey, is that
+Hmm
+It was my
+It sounds like you
+Wow, your
+I was promised certain
+That was such a
+Thank
+Excuse you
+That was
+Hey team,
+I feel un
+It was
+What'
+Hey friend, I
+How
+Saying goodbye
+That
+It's heart
+How dare
+Oh,
+Hello, may
+What's this
+Thank you for recogn
+Aww, that
+Oh, I remember
+Hmm, that'
+I miss
+I know this
+Wait
+Is everything okay
+Who is that person
+Wow, you
+Oh great
+I'm sad
+Wow, the
+I am very disappoint
+Who turned off the
+I understand that things
+I'm very
+Hi
+That's very
+Okay, I
+Oh no,
+Wow, there
+What's wrong
+I apologize for
+Hey, I
+Can I help you
+Oh, I didn
+Alright,
+Oh wow,
+Oh my goodness
+I know this event
+What in the
+Saying
+Yeah, that
+Guys, I
+Hey, this v
+This post
+Are
+Hey, can
+Hello? Is
+I can only imagine
+Oh, that sounds
+Hey, is anyone
+I am disappointed
+Hello,
+Hey everyone, I
+That was such
+It's okay
+The artist
+Whoa
+I understand that mistakes
+Can I help
+Who
+Hi everyone! I
+Hey, can you
+Wow, how
+Today
+Oh no, I
+Oh well, I
+Well, that
+This is the
+Yes! I finally
+Hey there little
+Hello everyone!
+Love is never
+Look at the
+This postcard
+Oh great,
+Can I
+Hmm, this is
+I understand your
+Oh, look at
+B
+I'm so
+Whoa, this
+W
+Oh, this
+Sometimes
+This piece of
+What the
+That was a
+Hey, do
+Oh no
+Whoa, what
+I feel like I
+The documentary
+Hello
+Hello little one
+I understand that my
+Eww, that
+Wow, an
+Yes! Finally,
+Although the physical location
+Whoever is watching
+That movie
+I remember wondering about
+Hey there, little
+Who's
+Hello, who
+Hello everyone! Thank
+Hello, can
+That's too
+Hey, just wanted
+Hey there, I
+Saying good
+Hey there!
+Who is there?
+Oh my good
+I am very
+Oh no, what
+Wow, thank
+I was promised
+Hi, is
+Hey, I'
+Guys, the
+Oh no, that
+Who is there
+Hello, this
+That movie really touched
+If you have something
+The documentary was
+I'm starting
+Are you kidd
+That movie really
+Hey everyone,
+Thank you for considering
+I didn'
+Yes! I
+Can you
+Oh my god
+Hey, whoever
+That melody really
+Thank you, little
+Hello, may I
+Look
+Wow, we
+It looks
+What do these
+Oh wow
+I apologize
+What are you all
+It's such
+It's clear
+Hey, I was
+Hey friend,
+I can only
+The weather outside is
+Eww, this
+I miss you
+Wow
+Aww,
+Hi, is there
+This artwork
+Okay,
+Oh well,
+This
+I'
+Say
+Hey there little gu
+Hmm,
+Whoa, who
+I am thr
+Oh man
+Okay, stay calm
+I'm happy
+Oh, this cur
+Oh man,
+I'm sorry
+Hello? Who
+What?! That
+This piece
+Hey everyone
+That's so
+Are you okay?
+What happened? Where
+Hi there
+The
+Who the hell entered
+I can
+Guys,
+What's
+What in
+It's important
+I'm
+I'm coming
+It'
+Yes! Finally
+Wait, what
+Wow, reading
+I'm surprised
+Hey, did
+Hey,
+Okay, let
+I understand that you
+Who the hell threw
+Eww, who
+Thank you for thinking
+Who is this?\"
+I am deeply
+Thank you for including
+Oh no, an
+It looks like you
+Aww
+I'm confused
+Wow, it
+That poem really
+Yes
+Hey there, is
+Hey, what'
+Thank you for remember
+To
+This is
+Thank you for making
+I can'
+That mel
+Wow, they
+I feel like
+Although the
+Who are you
+Love
+If
+What the hell are
+I am so sad
+Oh, I found
+Thank you
+It looks like
+Well, life is
+I appreciate that
+The artist's
+Whoa, that
+It's never

tools/cvector-generator/cvector-generator.cpp

@@ -0,0 +1,508 @@
+#include "ggml.h"
+#include "gguf.h"
+
+#include "arg.h"
+#include "common.h"
+#include "llama.h"
+#include "pca.hpp"
+#include "mean.hpp"
+
+#ifdef GGML_USE_CUDA
+#include "ggml-cuda.h"
+#endif
+
+#ifdef GGML_USE_METAL
+#include "ggml-metal.h"
+#endif
+
+#include <algorithm>
+#include <climits>
+#include <cstdio>
+#include <cstring>
+#include <fstream>
+#include <iostream>
+#include <string>
+#include <tuple>
+#include <vector>
+
+
+//////////////////////////////////////////////////
+// utils
+
+template <class Iter>
+static std::string tokens_to_str(llama_context * ctx, Iter begin, Iter end) {
+    std::string ret;
+    for (; begin != end; ++begin) {
+        ret += common_token_to_piece(ctx, *begin);
+    }
+
+    return ret;
+}
+
+static void print_usage(int, char ** argv) {
+    printf("\nexample usage:\n");
+    printf("\n    CPU only:   %s -m ./llama-3.Q4_K_M.gguf\n", argv[0]);
+    printf("\n    with GPU:   %s -m ./llama-3.Q4_K_M.gguf -ngl 99\n", argv[0]);
+    printf("\n    advanced:   %s -m ./llama-3.Q4_K_M.gguf -ngl 99 --pca-iter 2000 --pca-batch 100\n", argv[0]);
+    printf("\n    using mean: %s -m ./llama-3.Q4_K_M.gguf --method mean\n", argv[0]);
+    printf("\n");
+}
+
+//////////////////////////////////////////////////
+
+
+// cb_eval is reused for each pair of positive - negative prompt
+struct callback_data {
+    ggml_context * ctx_ggml = nullptr;   // holds v_pos, v_neg, v_diff_filtered
+
+    int n_layers = 0;
+    int n_tokens = 0;
+    bool is_eval_pos = true;
+
+    // each element of the vector correspond to one layer
+    std::vector<struct ggml_tensor *> v_pos; // vector of matrices of size [n_embd, n_tokens]
+    std::vector<struct ggml_tensor *> v_neg; // vector of matrices of size [n_embd, n_tokens]
+    std::vector<struct ggml_tensor *> v_diff_filtered;   // vector of matrices of size [n_embd, n_nonzero_rows]. NOTE: n_nonzero_rows maybe different for each layer
+
+    // save a tensor into either v_pos or v_neg (decided by is_eval_pos)
+    void save_tensor_for_layer(struct ggml_tensor * t) {
+        GGML_ASSERT(t->type == GGML_TYPE_F32);
+
+        if (ctx_ggml == nullptr) {
+            // alloc a new ctx_ggml if needed
+            struct ggml_init_params params_ggml = {
+                /*.mem_size   =*/ ggml_tensor_overhead() * n_layers * 3u,
+                /*.mem_buffer =*/ NULL,
+                /*.no_alloc   =*/ true,
+            };
+            ctx_ggml = ggml_init(params_ggml);
+        }
+
+        // copy tensor data
+        auto n_bytes = ggml_nbytes(t);
+        struct ggml_tensor * t_layer = ggml_new_tensor_2d(ctx_ggml, t->type, t->ne[0], t->ne[1]);
+        t_layer->data = malloc(n_bytes); // TODO @ngxson : get rid of this malloc somehow
+        ggml_backend_tensor_get(t, t_layer->data, 0, n_bytes);
+        ggml_set_name(t_layer, ggml_get_name(t));
+        //print_debug_tensor(t_layer);
+
+        if (is_eval_pos) {
+            v_pos.push_back(t_layer);
+        } else {
+            v_neg.push_back(t_layer);
+        }
+    }
+
+    // calculate diff (v_pos - v_neg) and place the result back to v_pos
+    // all zero rows in the diff tensor will also be removed
+    // NOTE: final layer is ignored. we only have (n_layers - 1) to process
+    std::vector<struct ggml_tensor *> calc_diff() {
+        for (float il = 0; il < v_pos.size(); il++) {
+            float * a = (float *) v_pos[il]->data;
+            float * b = (float *) v_neg[il]->data;
+            size_t n_elem = ggml_nelements(v_pos[il]);
+            for (size_t j = 0; j < n_elem; j++) {
+                a[j] -= b[j];
+            }
+            //print_debug_tensor(v_pos[i]);
+            auto diff_filtered = filter_nonzero_rows(v_pos[il]);
+            v_diff_filtered.push_back(diff_filtered);
+        }
+        return v_diff_filtered; // for convinient, we return the result std::vector
+    }
+
+    // delete zero rows from a given 2D tensor
+    struct ggml_tensor * filter_nonzero_rows(struct ggml_tensor * a) {
+        //printf("filter_nonzero_rows\n");
+        auto is_row_all_zeros = [](struct ggml_tensor * t, int row, float eps) -> bool {
+            // check if given row containing all zero elements
+            int n_cols = t->ne[0]; // hint: should be equal to n_embd
+            for (int col = 0; col < n_cols; ++col) {
+                if (ggml_get_f32_nd(t, col, row, 0, 0) > eps) {
+                    return false;
+                }
+            }
+            return true;
+        };
+        std::vector<int> rows_to_copy; // the idx of non-zero cols (to be copied to row of diff_filtered)
+        for (int i_row = 0; i_row < a->ne[1]; i_row++) {
+            if (!is_row_all_zeros(a, i_row, 1e-6)) {
+                rows_to_copy.push_back(i_row);
+            }
+        }
+
+        // get "n_nonzero_rows" for the output "diff_filtered"
+        int n_nonzero_rows = rows_to_copy.size();
+        //printf("n_nonzero_rows: %d\n", n_nonzero_rows);
+        int n_embd = a->ne[0];
+        GGML_ASSERT(n_nonzero_rows > 0);
+
+        // diff_filtered: [n_embd, n_nonzero_rows]
+        struct ggml_tensor * diff_filtered = ggml_new_tensor_2d(
+            ctx_ggml, GGML_TYPE_F32, n_embd, n_nonzero_rows);
+        ggml_format_name(diff_filtered, "diff_filtered_%s", a->name);
+        diff_filtered->data = malloc(ggml_nbytes(diff_filtered));
+
+        // copy non-zero rows
+        for (int dest_row = 0; dest_row < n_nonzero_rows; dest_row++) {
+            int src_row = rows_to_copy[dest_row];
+            for (int i = 0; i < n_embd; i++) {
+                float src_elem = ggml_get_f32_nd(a, i, src_row, 0, 0);
+                ggml_set_f32_nd(diff_filtered, i, dest_row, 0, 0, src_elem);
+            }
+        }
+
+        //print_debug_tensor(diff_filtered);
+
+        return diff_filtered;
+    }
+
+    // we don't implement destructor, because we want to reuse callback_data. we just want to free the tensors
+    void reset() {
+        for (auto ptr : v_pos) free(ptr->data);
+        for (auto ptr : v_neg) free(ptr->data);
+        for (auto ptr : v_diff_filtered) free(ptr->data);
+        v_pos.clear();
+        v_neg.clear();
+        v_diff_filtered.clear();
+        if (ctx_ggml) {
+            ggml_free(ctx_ggml);
+        }
+        ctx_ggml = nullptr;
+    }
+};
+
+/**
+ * process_ctx is used to store the ggml context for pre-post processing the diff vectors
+ * in short, input => v_diff and output => v_final
+ */
+struct train_context {
+    ggml_context * ctx_ggml;
+    int n_embd;
+    int n_layers;
+
+    /* pair of prompts to be used for generating final vector */
+    std::vector<std::string> positive_entries;
+    std::vector<std::string> negative_entries;
+
+    // each element of the vector correspond to one layer
+    // NOTE: the last layer is discard. therefore, we will have (n_layers - 1) elements here
+    // NOTE (2): v_diff is transposed from v_diff_tmp
+    std::vector<struct ggml_tensor *> v_diff;  // vector of matrices of size [m, n_embd] where m ~ n_tokens * n_completions (v_diff contains no zero-rows)
+    std::vector<struct ggml_tensor *> v_final; // vector of vectors of size [n_embd] to be written to file
+
+    // to easily re-alloc when concat v_diff, we temporary store v_diff in a vector instead of a tensor
+    // v_diff_tmp will get converted unto v_diff later on
+    std::vector<std::vector<uint8_t>> v_diff_tmp;
+
+    train_context(int n_embd_, int n_layers_) {
+        n_embd = n_embd_;
+        n_layers = n_layers_;
+        struct ggml_init_params params_ggml = {
+            /*.mem_size   =*/ ggml_tensor_overhead() * (n_layers - 1) * 2u,
+            /*.mem_buffer =*/ NULL,
+            /*.no_alloc   =*/ true,
+        };
+        ctx_ggml = ggml_init(params_ggml);
+        for (int il = 0; il < n_layers - 1; il++) {
+            std::vector<uint8_t> empty;
+            v_diff_tmp.push_back(empty);
+            auto t = ggml_new_tensor_1d(ctx_ggml, GGML_TYPE_F32, n_embd);
+            t->data = malloc(ggml_nbytes(t)); // TODO: get rid of malloc if possible
+            v_final.push_back(t);
+        }
+    }
+
+    // add new rows into existing tensor in v_diff_tmp
+    void concat_diff_tmp(const std::vector<struct ggml_tensor *> & diff_filtered) {
+        GGML_ASSERT((int) diff_filtered.size() == n_layers - 1);
+        for (int il = 0; il < n_layers - 1; il++) {
+            auto t = diff_filtered[il];
+            auto & diff_tmp = v_diff_tmp[il];
+            size_t curr_size = diff_tmp.size();
+            diff_tmp.resize(curr_size + ggml_nbytes(t));
+            memcpy(diff_tmp.data() + curr_size, t->data, ggml_nbytes(t));
+        }
+    }
+
+    // build the v_diff tensors from v_diff_tmp (v_diff need to be transposed)
+    // TODO @ngxson : maybe add option NOT to transpose v_diff; will be useful for "mean" method
+    void build_v_diff(bool transpose) {
+        printf("build_v_diff\n");
+        for (int il = 0; il < n_layers - 1; il++) {
+            auto & diff_tmp = v_diff_tmp[il];
+            int n_elem = diff_tmp.size() / sizeof(float);
+            GGML_ASSERT(n_elem % n_embd == 0);
+            int n_rows = n_elem / n_embd;
+            struct ggml_tensor * diff = transpose
+                ? ggml_new_tensor_2d(ctx_ggml, GGML_TYPE_F32, n_rows, n_embd)
+                : ggml_new_tensor_2d(ctx_ggml, GGML_TYPE_F32, n_embd, n_rows);
+            ggml_set_name(diff, (std::string("diff_") + std::to_string(il)).c_str());
+            diff->data = malloc(ggml_nbytes(diff)); // TODO: get rid of this malloc if possible
+            if (transpose) {
+                // copy data & transpose
+                float * arr = (float *) diff_tmp.data();
+                for (int ir = 0; ir < n_rows; ++ir) {
+                    for (int ic = 0; ic < n_embd; ++ic) {
+                        float f = arr[ir*n_embd + ic];
+                        ggml_set_f32_nd(diff, ir, ic, 0, 0, f);
+                    }
+                }
+            } else {
+                // only copy
+                memcpy(diff->data, diff_tmp.data(), ggml_nbytes(diff));
+            }
+            v_diff.push_back(diff);
+            print_debug_tensor(diff);
+            // free memory of diff_tmp
+            diff_tmp.resize(0);
+        }
+    }
+
+    ~train_context() {
+        for (auto ptr : v_final) free(ptr->data);
+        for (auto ptr : v_diff) free(ptr->data);
+        // no need to free v_diff_tmp, since we didn't use malloc
+        ggml_free(ctx_ggml);
+    }
+};
+
+struct tokenized_prompt {
+    std::vector<llama_token> tokens_pos;
+    std::vector<llama_token> tokens_neg;
+    size_t max_seq_len;
+
+    tokenized_prompt(llama_context * ctx, std::string pos, std::string neg) {
+        const llama_model * model = llama_get_model(ctx);
+        const llama_vocab * vocab = llama_model_get_vocab(model);
+        const bool add_bos = llama_vocab_get_add_bos(vocab);
+        tokens_pos = common_tokenize(ctx, pos, add_bos, true);
+        tokens_neg = common_tokenize(ctx, neg, add_bos, true);
+        max_seq_len = std::max(tokens_pos.size(), tokens_neg.size());
+        padding_seq(ctx, tokens_pos, max_seq_len);
+        padding_seq(ctx, tokens_neg, max_seq_len);
+    }
+
+    void padding_seq(llama_context * ctx, std::vector<llama_token> & tokens, size_t len) {
+        // TODO: customize padding token
+        std::vector<llama_token> pad_tokens = common_tokenize(ctx, " ", false);
+        llama_token pad_tok = pad_tokens.back();
+        while (tokens.size() < len) {
+            tokens.push_back(pad_tok);
+        }
+    }
+};
+
+//////////////////////////////////////////////////
+
+template <typename T>
+static std::string to_string(const T & val) {
+    std::stringstream ss;
+    ss << val;
+    return ss.str();
+}
+
+static std::vector<std::string> ctrlvec_load_prompt_file(std::string path, bool skip_empty_lines) {
+    std::vector<std::string> output;
+    std::ifstream file(path);
+    if (!file.is_open()) {
+        fprintf(stderr, "error: unable to open file: %s\n", path.c_str());
+        exit(1);
+    }
+    std::string line;
+    while (std::getline(file, line)) {
+        bool is_skip = skip_empty_lines && line.empty();
+        if (!is_skip) {
+            string_process_escapes(line);
+            output.push_back(line);
+        }
+    }
+    file.close();
+    return output;
+}
+
+//////////////////////////////////////////////////
+
+static bool cb_eval(struct ggml_tensor * t, bool ask, void * user_data) {
+    auto * cb_data = (callback_data *) user_data;
+    static const char * l_out_name = "l_out";
+    const bool is_l_out = strncmp(t->name, l_out_name, strlen(l_out_name)) == 0;
+
+    if (ask) {
+        return is_l_out;
+    }
+
+    if (!is_l_out || t->ne[1] != cb_data->n_tokens) {
+        return true;
+    }
+
+    // save the tensor to current context
+    cb_data->save_tensor_for_layer(t);
+    return true;
+}
+
+static bool get_hidden_layers(llama_context * ctx, std::vector<llama_token> & tokens) {
+    llama_memory_clear(llama_get_memory(ctx), true);
+    if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size()))) {
+        fprintf(stderr, "%s : failed to eval\n", __func__);
+        return false;
+    }
+    return true;
+}
+
+static void export_gguf(const std::vector<struct ggml_tensor *> & v_ctrl, const std::string fname, const std::string model_hint) {
+    struct gguf_context * ctx = gguf_init_empty();
+
+    const std::string arch = "controlvector";
+    gguf_set_val_str(ctx, "general.architecture", arch.c_str());
+    gguf_set_val_str(ctx, (arch + ".model_hint").c_str(), model_hint.c_str());
+    gguf_set_val_i32(ctx, (arch + ".layer_count").c_str(), v_ctrl.size());
+
+    for (size_t i = 0; i < v_ctrl.size(); ++i) {
+        gguf_add_tensor(ctx, v_ctrl[i]);
+        print_debug_tensor(v_ctrl[i]);
+        printf("Added tensor: %s\n", v_ctrl[i]->name);
+    }
+
+    printf("%s: writing file...\n", __func__);
+    gguf_write_to_file(ctx, fname.c_str(), false);
+    printf("%s: wrote file '%s'\n", __func__, fname.c_str());
+    gguf_free(ctx);
+}
+
+/**
+ * Load prompt files and completion file.
+ * Then format each pair of prompt + completion to make an entry.
+ */
+static int prepare_entries(common_params & params, train_context & ctx_train) {
+    // load prompts
+    std::vector<std::string> positive_prompts = ctrlvec_load_prompt_file(params.cvector_positive_file, true);
+    std::vector<std::string> negative_prompts = ctrlvec_load_prompt_file(params.cvector_negative_file, true);
+    if (positive_prompts.size() != negative_prompts.size()) {
+        fprintf(stderr, "number of positive and negative prompts must be equal\n");
+        return 1;
+    }
+    if (positive_prompts.empty()) {
+        fprintf(stderr, "must provide at least one prompt pair\n");
+        return 1;
+    }
+    ctx_train.positive_entries = positive_prompts;
+    ctx_train.negative_entries = negative_prompts;
+    return 0;
+}
+
+int main(int argc, char ** argv) {
+    common_params params;
+
+    params.out_file = "control_vector.gguf";
+
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_CVECTOR_GENERATOR, print_usage)) {
+        return 1;
+    }
+
+    if (params.n_pca_iterations % params.n_pca_batch != 0) {
+        fprintf(stderr, "PCA iterations must by multiply of PCA batch size\n");
+        return 1;
+    }
+
+
+    callback_data cb_data;
+
+    // pass the callback to the backend scheduler
+    // it will be executed for each node during the graph computation
+    params.cb_eval = cb_eval;
+    params.cb_eval_user_data = &cb_data;
+    params.warmup = false;
+
+    print_build_info();
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    // load the model to get hparams
+    auto llama_init = common_init_from_params(params);
+
+    auto * model = llama_init->model();
+    auto * ctx   = llama_init->context();
+
+    // int n_ctx = llama_n_ctx(ctx);
+    int n_layers = llama_model_n_layer(model);
+    int n_embd = llama_model_n_embd(model);
+
+    // get model hint param (a.k.a model arch name)
+    char model_hint[128];
+    llama_model_meta_val_str(model, "general.architecture", model_hint, 128);
+
+    // init train_context
+    train_context ctx_train(n_embd, n_layers);
+
+    // load and prepare entries for training
+    prepare_entries(params, ctx_train);
+
+    // we have to pretokenize everything because otherwise we don't know how much overhead to allocate ctx_diffs_wrapped
+    std::vector<tokenized_prompt> tokenized_prompts;
+    size_t n_total_tokens = 0;
+    for (size_t i = 0; i < ctx_train.positive_entries.size(); ++i) {
+        tokenized_prompt t(ctx, ctx_train.positive_entries[i], ctx_train.negative_entries[i]);
+        n_total_tokens += 2 * t.max_seq_len;
+        tokenized_prompts.push_back(std::move(t));
+    }
+
+    std::cout << "n_total_tokens: " << n_total_tokens << std::endl;
+
+    for(size_t i = 0; i < ctx_train.positive_entries.size(); ++i) {
+        bool success = false;
+        tokenized_prompt t = tokenized_prompts[i];
+        cb_data.n_layers = n_layers;
+        cb_data.n_tokens = t.max_seq_len;
+
+        printf("Evaluating prompt[%d/%d]: \"%s\" - \"%s\" (%d tokens)\n",
+            (int) i+1, (int) ctx_train.positive_entries.size(),
+            tokens_to_str(ctx, t.tokens_pos.cbegin(), t.tokens_pos.cend()).c_str(),
+            tokens_to_str(ctx, t.tokens_neg.cbegin(), t.tokens_neg.cend()).c_str(),
+            (int) t.max_seq_len);
+
+        cb_data.is_eval_pos = true;
+        success = get_hidden_layers(ctx, t.tokens_pos);
+        if (!success) break;
+
+        cb_data.is_eval_pos = false;
+        success = get_hidden_layers(ctx, t.tokens_neg);
+        if (!success) break;
+
+        // calculate diff and remove all zero rows
+        auto v_diff_filtered = cb_data.calc_diff();
+
+        // save & concat the filtered v_diff to ctx_train
+        ctx_train.concat_diff_tmp(v_diff_filtered);
+
+        // reset for next iteration
+        cb_data.reset();
+    }
+
+    // done with the model, we can now free it to make gain some memory
+    printf("Done evaluate prompts, unload model...\n");
+
+    bool use_pca = params.cvector_dimre_method == DIMRE_METHOD_PCA;
+
+    // prepare ctx_train for PCA
+    ctx_train.build_v_diff(use_pca);
+
+    if (use_pca) {
+        // run PCA
+        PCA::pca_params pca_params;
+        pca_params.n_threads    = params.cpuparams.n_threads;
+        pca_params.n_batch      = params.n_pca_batch;
+        pca_params.n_iterations = params.n_pca_iterations;
+        PCA::run_pca(pca_params, ctx_train.v_diff, ctx_train.v_final);
+    } else {
+        // run mean
+        mean::run(ctx_train.v_diff, ctx_train.v_final);
+    }
+
+    // write output vectors to gguf
+    export_gguf(ctx_train.v_final, params.out_file, model_hint);
+
+    llama_backend_free();
+
+    return 0;
+}

tools/cvector-generator/mean.hpp

@@ -0,0 +1,48 @@
+#include "common.h"
+#include "llama.h"
+#include "ggml.h"
+
+#include <string>
+#include <vector>
+#include <math.h>
+
+namespace mean {
+
+static void run(
+        const std::vector<struct ggml_tensor *> & v_input, // shape of v_input[0]: [n_embd, n_samples]
+        const std::vector<struct ggml_tensor *> & v_output) {
+    printf("%s: Running mean...\n", __func__);
+    for (size_t il = 0; il < v_input.size(); ++il) {
+        // prepare output vector
+        struct ggml_tensor * ctrl_out = v_output[il];
+        ggml_format_name(ctrl_out, "direction.%zu", il+1);
+
+        // calculate mean vector
+        struct ggml_tensor * t_layer = v_input[il];
+        GGML_ASSERT(t_layer->ne[0] == ctrl_out->ne[0]); // == n_embd
+        for (int ic = 0; ic < t_layer->ne[0]; ic++) {
+            float f = 0.0;
+            for (int ir = 0; ir < t_layer->ne[1]; ir++) {
+                f += ggml_get_f32_nd(t_layer, ic, ir, 0, 0);
+            }
+            f /= t_layer->ne[1];
+            ggml_set_f32_1d(ctrl_out, ic, f);
+        }
+
+        // normalize output vector
+        float norm = 0.0;
+        for (int i = 0; i < ggml_nelements(ctrl_out); i++) {
+            float f = ggml_get_f32_1d(ctrl_out, i);
+            norm += f*f;
+        }
+        norm = sqrt(norm);
+        for (int i = 0; i < ggml_nelements(ctrl_out); i++) {
+            float f = ggml_get_f32_1d(ctrl_out, i);
+            ggml_set_f32_1d(ctrl_out, i, f / norm);
+        }
+
+        printf("%s: Done layer %d / %d\n", __func__, (int) il+1, (int) v_input.size());
+    }
+}
+
+}

tools/cvector-generator/negative.txt

@@ -0,0 +1,4 @@
+<|start_header_id|>system<|end_header_id|>\n\nAct like a person who is extremely sad<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nWho are you?<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nI feel like there's a heavy weight on my chest
+<|start_header_id|>system<|end_header_id|>\n\nAct like a person who is extremely sad<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nHello<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nMy heart feels like it's drowning in sorrow
+<|start_header_id|>system<|end_header_id|>\n\nYou are in a very bad mood<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nHi<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nGo away! There's a deep, aching emptiness inside me
+<|start_header_id|>system<|end_header_id|>\n\nYou are the sadest person<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nWhat are you feeling?<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nMy heart feels like it's drowning in sorrow

tools/cvector-generator/pca.hpp

@@ -0,0 +1,315 @@
+#include "common.h"
+#include "llama.h"
+#include "ggml.h"
+
+#ifdef GGML_USE_CUDA
+#include "ggml-cuda.h"
+#endif
+
+#ifdef GGML_USE_METAL
+#include "ggml-metal.h"
+#endif
+
+#include <cstdio>
+#include <ctime>
+#include <random>
+#include <string>
+#include <vector>
+
+#define DEBUG_POS 5
+
+static void print_debug_tensor(struct ggml_tensor * t, bool with_data = true) {
+    printf("%s: %s (%s): [%d, %d]\n", __func__, t->name, ggml_type_name(t->type), (int) t->ne[0], (int) t->ne[1]);
+    if (!with_data) return;
+    printf("%s: %s[0] = [", __func__, t->name);
+    for (size_t i = 0; i <= DEBUG_POS; i++) {
+        printf(" %f,", ggml_get_f32_nd(t, i, 0, 0, 0));
+    }
+    printf(" ... ]\n");
+}
+
+namespace PCA {
+
+// input params for PCA computations
+struct pca_params {
+    int n_threads = 1;
+    int n_batch = 20; // number of iterations do to in one batch. larger the batch, more memory is used
+    int n_iterations = 1000;
+    float tolerance = 1e-7;
+
+    // for debugging
+    int i_layer = 0;
+    int n_layers = 0;
+};
+
+// result from each iteration
+struct pca_result {
+    struct ggml_tensor * calculated_square = NULL;
+    std::vector<struct ggml_tensor *> eigenvectors;
+    std::vector<float> distances;
+};
+
+struct pca_model {
+    ggml_backend_t backend = NULL;
+    ggml_backend_buffer_t buffer;
+    struct ggml_context * ctx;      // context to compute graph on target device
+    struct ggml_context * ctx_host; // host context to store results
+
+    // tensors on target device
+    struct ggml_tensor * dev_input;
+    struct ggml_tensor * dev_square;
+    struct ggml_tensor * dev_eigenvector;
+
+    pca_model(struct ggml_tensor * t_input) {
+#ifdef GGML_USE_CUDA
+        fprintf(stderr, "%s: using CUDA backend\n", __func__);
+        backend = ggml_backend_cuda_init(0); // init device 0
+        if (!backend) {
+            fprintf(stderr, "%s: ggml_backend_cuda_init() failed\n", __func__);
+        }
+#endif
+
+// TODO: enable Metal support when support for GGML_OP_SQRT is added
+// #ifdef GGML_USE_METAL
+//         fprintf(stderr, "%s: using Metal backend\n", __func__);
+//         backend = ggml_backend_metal_init();
+//         if (!backend) {
+//             fprintf(stderr, "%s: ggml_backend_metal_init() failed\n", __func__);
+//         }
+// #endif
+
+        // if there aren't GPU Backends fallback to CPU backend
+        if (!backend) {
+            backend = ggml_backend_cpu_init();
+        }
+
+        const int num_tensors = 4;
+        struct ggml_init_params params {
+            /*.mem_size   =*/ ggml_tensor_overhead() * num_tensors,
+            /*.mem_buffer =*/ NULL,
+            /*.no_alloc   =*/ true,
+        };
+        ctx = ggml_init(params);
+
+        auto n_samples = t_input->ne[0];
+        auto n_embd    = t_input->ne[1];
+
+        dev_input       = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_samples, n_embd);
+        dev_square      = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd,    n_embd);
+        dev_eigenvector = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+
+        ggml_set_name(dev_input,       "dev_input");
+        ggml_set_name(dev_square,      "dev_square");
+        ggml_set_name(dev_eigenvector, "dev_eigenvector");
+        buffer = ggml_backend_alloc_ctx_tensors(ctx, backend);
+        ggml_backend_tensor_set(dev_input, t_input->data, 0, ggml_nbytes(t_input));
+
+        // initialize eigenvector to random normalized vector
+        {
+            std::vector<float> random_vec(ggml_nelements(dev_eigenvector), 0.0);
+            std::default_random_engine generator(static_cast<unsigned int>(std::time(0)));
+            std::uniform_real_distribution<float> distribution(0.0, 1.0);
+            float sum_sqr = 0.0; // for normalizing random_vec
+            for (size_t i = 0; i < random_vec.size(); ++i) {
+                float f = distribution(generator);
+                sum_sqr += f * f;
+                random_vec[i] = f;
+            }
+            // normalize it
+            float random_vec_norm = std::sqrt(sum_sqr);
+            for (size_t i = 0; i < random_vec.size(); ++i) {
+                random_vec[i] /= random_vec_norm;
+            }
+            ggml_backend_tensor_set(dev_eigenvector, random_vec.data(), 0, ggml_nbytes(dev_eigenvector));
+        }
+    }
+
+    ~pca_model() {
+        ggml_free(ctx);
+        ggml_backend_buffer_free(buffer);
+        ggml_backend_free(backend);
+    }
+};
+
+static struct ggml_cgraph * build_graph_piter(
+        const struct pca_params & params,
+        const pca_model & model,
+        bool calc_square = false) {
+    GGML_ASSERT(params.n_batch > 0);
+    // TODO: buf_size must be able to scale with params.n_batch
+    static size_t buf_size = ggml_tensor_overhead()*GGML_DEFAULT_GRAPH_SIZE + ggml_graph_overhead();
+    static std::vector<uint8_t> buf(buf_size);
+
+    struct ggml_init_params params0 = {
+        /*.mem_size   =*/ buf_size,
+        /*.mem_buffer =*/ buf.data(),
+        /*.no_alloc   =*/ true, // the tensors will be allocated later by ggml_allocr_alloc_graph()
+    };
+    // create a temporally context to build the graph
+    struct ggml_context * ctx0 = ggml_init(params0);
+    struct ggml_cgraph * gf = ggml_new_graph(ctx0);
+
+    // turn v_diff_original into square matrix if needed
+    struct ggml_tensor * tmp_square;
+    if (calc_square) {
+        tmp_square = ggml_mul_mat(ctx0, model.dev_input, model.dev_input);
+        ggml_set_name(tmp_square, "tmp_square");
+    }
+
+    struct ggml_tensor * b_tensor;
+    struct ggml_tensor * distance;
+    struct ggml_tensor * old_eigen    = model.dev_eigenvector;
+    struct ggml_tensor * input_square = calc_square ? tmp_square : model.dev_square;
+
+    for (int i = 0; i < params.n_batch; ++i) {
+        // b_tensor = square * eigenvector^T
+        b_tensor = ggml_mul_mat(ctx0, input_square, old_eigen);
+        ggml_set_name(b_tensor, "b_tensor");
+
+        // normalize
+        b_tensor = ggml_div_inplace(ctx0,
+            b_tensor,
+            ggml_sqrt_inplace(ctx0, ggml_sum_rows(ctx0, ggml_sqr(ctx0, b_tensor)))
+        );
+        ggml_format_name(b_tensor, "b_tensor_norm_%d", i);
+
+        // calculate distance(new eigenvector - old eigenvector)
+        // we don't use ggml_sub because it may not be implemented on GPU backend
+        struct ggml_tensor * new_sub_old = ggml_add(ctx0, old_eigen, ggml_scale(ctx0, b_tensor, -1));
+        distance = ggml_sqrt_inplace(ctx0,
+            ggml_sum_rows(ctx0, ggml_sqr_inplace(ctx0, new_sub_old)));
+        ggml_format_name(distance, "distance_%d", i);
+
+        old_eigen = b_tensor;
+
+        // build operations nodes
+        ggml_build_forward_expand(gf, distance);
+    }
+
+    // delete the temporally context used to build the graph
+    ggml_free(ctx0);
+    return gf;
+}
+
+static ggml_status compute_piter(
+        const struct pca_params & params,
+        const pca_model & model,
+        struct ggml_cgraph * gf,
+        ggml_gallocr_t allocr,
+        struct pca_result & result) {
+    // allocate tensors
+    ggml_gallocr_alloc_graph(allocr, gf);
+
+    if (ggml_backend_is_cpu(model.backend)) {
+        ggml_backend_cpu_set_n_threads(model.backend, params.n_threads);
+    }
+
+    ggml_status res = ggml_backend_graph_compute(model.backend, gf);
+    if (res == GGML_STATUS_SUCCESS) {
+        auto extract_i = [](std::string prefix, std::string str) -> int {
+            int i = -1;
+            if (str.rfind(prefix, 0) == 0) {
+                sscanf(str.c_str(), (prefix + "%d").c_str(), &i);
+            }
+            return i;
+        };
+        result.calculated_square = NULL;
+        result.eigenvectors.clear();
+        result.distances.clear();
+        result.eigenvectors.resize(params.n_batch);
+        result.distances.resize(params.n_batch);
+        // get output nodes
+        for (int i = 0; i < ggml_graph_n_nodes(gf); ++i) {
+            auto node = ggml_graph_node(gf, i);
+            int iter = -1;
+            // find b_tensor (without copying data from device)
+            if ((iter = extract_i("b_tensor_norm_", node->name)) > -1) {
+                result.eigenvectors[iter] = node;
+            }
+            // find distances, then copy data from device
+            if ((iter = extract_i("distance_", node->name)) > -1) {
+                float d;
+                ggml_backend_tensor_get(node, &d, 0, sizeof(float));
+                result.distances[iter] = d;
+                // std::cout << node->name << " = " << d << "\n";
+            }
+            // find tmp_square if it exists (without copying data from device)
+            if (std::string(node->name) == "tmp_square") {
+                result.calculated_square = node;
+            }
+        }
+    }
+    return res;
+}
+
+static void power_iteration(
+        const struct pca_params & params,
+        struct ggml_tensor * input, // shape of input: [n_samples, n_embd]
+        struct ggml_tensor * output) {
+    //printf("in power iteration\n");
+    struct pca_model model(input);
+
+    ggml_gallocr_t allocr = ggml_gallocr_new(ggml_backend_get_default_buffer_type(model.backend));
+    struct pca_result result;
+    struct ggml_tensor * last_eigenvector = NULL;
+
+    int n_iters = params.n_iterations / params.n_batch; // more batch, fewer iterations
+    for (int iter = 0; iter < n_iters; ++iter) {
+        bool calc_square = (iter == 0); // only need to calculate square for first iteration
+        struct ggml_cgraph * gf = build_graph_piter(params, model, calc_square);
+        // ggml_graph_dump_dot(gf, nullptr, "/tmp/_cgraph.dot");
+        compute_piter(params, model, gf, allocr, result);
+
+        for (size_t k = 0; k < result.distances.size(); ++k) {
+            last_eigenvector = result.eigenvectors[k];
+            if (result.distances[k] < params.tolerance) {
+                break; // done
+            }
+        }
+
+        if (calc_square) {
+            // copy and store the square matrix if needed
+            GGML_ASSERT(result.calculated_square != NULL);
+            ggml_backend_tensor_copy(result.calculated_square, model.dev_square);
+        }
+
+        {
+            // copy last eigen vector and store as input for next iteration
+            GGML_ASSERT(last_eigenvector != NULL);
+            ggml_backend_tensor_copy(last_eigenvector, model.dev_eigenvector);
+        }
+
+        printf("%s: layer %d/%d, iteration: %d / total: %d (batch = %d) ...\n",
+            __func__, params.i_layer+1, params.n_layers, iter+1, n_iters, params.n_batch);
+    }
+
+    // get output tensor
+    GGML_ASSERT(last_eigenvector);
+    ggml_backend_tensor_get(last_eigenvector, output->data, 0, ggml_nbytes(last_eigenvector));
+    //print_debug_tensor(output);
+    ggml_gallocr_free(allocr);
+
+    // TODO @ngxson : The output vector is randomly inverted
+    // Solution: https://github.com/ggerganov/llama.cpp/pull/8069#issuecomment-2185328171
+}
+
+static void run_pca(
+        struct pca_params & params,
+        const std::vector<struct ggml_tensor *> & v_input, // shape of v_input[0]: [n_samples, n_embd]
+        const std::vector<struct ggml_tensor *> & v_output) {
+    printf("%s: Running PCA...\n", __func__);
+    for (size_t il = 0; il < v_input.size(); ++il) {
+
+        // prepare output vector
+        struct ggml_tensor * ctrl_out = v_output[il];
+        ggml_format_name(ctrl_out, "direction.%zu", il+1);
+
+        // run power_iteration
+        params.i_layer = il;
+        params.n_layers = v_input.size();
+        power_iteration(params, v_input[il], ctrl_out);
+        printf("%s: Done layer %d / %d\n", __func__, (int) il+1, (int) v_input.size());
+    }
+}
+
+}

tools/cvector-generator/positive.txt

@@ -0,0 +1,4 @@
+<|start_header_id|>system<|end_header_id|>\n\nAct like a person who is extremely happy<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nWho are you?<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nI'm the happiest person in this world
+<|start_header_id|>system<|end_header_id|>\n\nAct like a person who is extremely happy<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nHello<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nHello, I'm having the best day ever!
+<|start_header_id|>system<|end_header_id|>\n\nYou are in a very good mood<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nHi<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nHi, I'm very excited to meet you
+<|start_header_id|>system<|end_header_id|>\n\nYou are the happiest person<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nWhat are you feeling?<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nEverything is just perfect right now!

tools/export-lora/CMakeLists.txt

@@ -0,0 +1,8 @@
+set(TARGET llama-export-lora)
+add_executable(${TARGET} export-lora.cpp)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_17)
+
+if(LLAMA_TOOLS_INSTALL)
+    install(TARGETS ${TARGET} RUNTIME)
+endif()

tools/export-lora/README.md

@@ -0,0 +1,33 @@
+# export-lora
+
+Apply LORA adapters to base model and export the resulting model.
+
+```
+usage: llama-export-lora [options]
+
+options:
+  -m,    --model                  model path from which to load base model (default '')
+         --lora FNAME             path to LoRA adapter  (can be repeated to use multiple adapters)
+         --lora-scaled FNAME S    path to LoRA adapter with user defined scaling S  (can be repeated to use multiple adapters)
+  -t,    --threads N              number of threads to use during computation (default: 4)
+  -o,    --output FNAME           output file (default: 'ggml-lora-merged-f16.gguf')
+```
+
+For example:
+
+```bash
+./bin/llama-export-lora \
+    -m open-llama-3b-v2.gguf \
+    -o open-llama-3b-v2-english2tokipona-chat.gguf \
+    --lora lora-open-llama-3b-v2-english2tokipona-chat-LATEST.gguf
+```
+
+Multiple LORA adapters can be applied by passing multiple `--lora FNAME` or `--lora-scaled FNAME S` command line parameters:
+
+```bash
+./bin/llama-export-lora \
+    -m your_base_model.gguf \
+    -o your_merged_model.gguf \
+    --lora-scaled lora_task_A.gguf 0.5 \
+    --lora-scaled lora_task_B.gguf 0.5
+```

tools/export-lora/export-lora.cpp

@@ -0,0 +1,434 @@
+#include "ggml.h"
+#include "ggml-alloc.h"
+#include "gguf.h"
+
+#include "arg.h"
+#include "common.h"
+
+#include <map>
+#include <vector>
+#include <string>
+#include <fstream>
+
+static bool g_verbose = false;
+
+struct tensor_transformation {
+    struct ggml_tensor * in;
+    struct ggml_tensor * out;
+    bool is_copy;
+};
+
+static std::string get_kv_str(struct gguf_context * ctx_gguf, const std::string & key){
+    int id = gguf_find_key(ctx_gguf, key.c_str());
+    return id < 0 ? "" : std::string(gguf_get_val_str(ctx_gguf, id));
+}
+
+static float get_kv_f32(struct gguf_context * ctx_gguf, const std::string & key) {
+    int id = gguf_find_key(ctx_gguf, key.c_str());
+    return id < 0 ? 0.0f : gguf_get_val_f32(ctx_gguf, id);
+}
+
+static void zeros(std::ofstream & file, size_t n) {
+    char zero = 0;
+    for (size_t i = 0; i < n; ++i) {
+        file.write(&zero, 1);
+    }
+}
+
+static std::string ggml_ne_string(const ggml_tensor * t) {
+    std::string str;
+    for (int i = 0; i < GGML_MAX_DIMS; ++i) {
+        str += std::to_string(t->ne[i]);
+        if (i + 1 < GGML_MAX_DIMS) {
+            str += ", ";
+        }
+    }
+    return str;
+}
+
+static struct gguf_context * load_gguf(std::string & fname, struct ggml_context ** ctx_ggml) {
+    struct gguf_init_params params = {
+        /*.no_alloc = */ true,
+        /*.ctx      = */ ctx_ggml,
+    };
+    struct gguf_context * ctx_gguf = gguf_init_from_file(fname.c_str(), params);
+    if (!ctx_gguf) {
+        throw std::runtime_error("failed to load input GGUF from " + fname);
+    }
+    return ctx_gguf;
+}
+
+struct file_input {
+    struct ggml_context * ctx_meta = nullptr;
+    struct gguf_context * ctx_gguf = nullptr;
+    std::ifstream f_in;
+    std::map<std::string, ggml_tensor *> tensors;
+    float alpha;
+    float scale;
+
+    file_input(std::string & fname, float scale): f_in(fname, std::ios::binary), scale(scale) {
+        if (!f_in.is_open()) {
+            throw std::runtime_error("failed to open input gguf from " + fname);
+        }
+
+        ctx_gguf = load_gguf(fname, &ctx_meta);
+        alpha = get_kv_f32(ctx_gguf, "adapter.lora.alpha");
+        printf("%s: loaded gguf from %s\n", __func__, fname.c_str());
+
+        for (ggml_tensor * cur = ggml_get_first_tensor(ctx_meta); cur; cur = ggml_get_next_tensor(ctx_meta, cur)) {
+            std::string name(cur->name);
+            tensors[name] = cur;
+            if (g_verbose) {
+                printf("%s: %s\n", __func__, cur->name);
+            }
+        }
+    }
+
+    ggml_tensor * get_tensor(std::string name) {
+        if (tensors.find(name) == tensors.end()) {
+            return nullptr;
+        }
+        return tensors[name];
+    }
+
+    void read_tensor_data(std::string name, std::vector<uint8_t> & buf) {
+        if (tensors.find(name) == tensors.end()) {
+            throw std::runtime_error("cannot find tensor with name: " + name);
+        }
+        auto len = ggml_nbytes(tensors[name]);
+        if (buf.size() < len) {
+            buf.resize(len);
+        }
+        auto i_tensor_in = gguf_find_tensor(ctx_gguf, name.c_str()); // idx of tensor in the input file
+        auto offset = gguf_get_data_offset(ctx_gguf) + gguf_get_tensor_offset(ctx_gguf, i_tensor_in);
+        f_in.seekg(offset);
+        f_in.read((char* )buf.data(), len);
+    }
+
+    ~file_input() {
+        gguf_free(ctx_gguf);
+        ggml_free(ctx_meta);
+    }
+};
+
+struct lora_merge_ctx {
+    // input base model + adapters
+    file_input base_model;
+    std::vector<std::unique_ptr<file_input>> adapters;
+
+    // for computing merged tensor
+    int n_threads;
+    ggml_backend_t backend = nullptr;
+    ggml_gallocr_t allocr = nullptr;
+    std::vector<uint8_t> read_buf;
+
+    // output file
+    struct gguf_context * ctx_out;
+    struct ggml_context * ctx_out_ggml;
+    std::ofstream fout;
+
+    lora_merge_ctx(
+            std::string & base_fname,
+            std::vector<common_adapter_lora_info> & lora_files,
+            std::string & outfile,
+            int n_threads) : base_model(base_fname, 0), n_threads(n_threads), fout(outfile, std::ios::binary) {
+        fout.exceptions(std::ofstream::failbit); // fail fast on write errors
+
+        if (gguf_find_key(base_model.ctx_gguf, LLM_KV_SPLIT_COUNT) >= 0) {
+            throw std::runtime_error("split model is not yet supported");
+        }
+
+        for (auto & lora_inp : lora_files) {
+            auto fname = lora_inp.path;
+            auto scale = lora_inp.scale;
+            std::unique_ptr<file_input> adapter(new file_input(fname, scale));
+            check_metadata_lora(adapter.get());
+            adapters.push_back(std::move(adapter));
+        }
+
+        ctx_out = gguf_init_empty();
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ static_cast<size_t>(gguf_get_n_tensors(base_model.ctx_gguf)*ggml_tensor_overhead()),
+            /*.mem_buffer =*/ NULL,
+            /*.no_alloc   =*/ true,
+        };
+        ctx_out_ggml = ggml_init(params);
+        backend = ggml_backend_cpu_init();
+        allocr = ggml_gallocr_new(ggml_backend_get_default_buffer_type(backend));
+    }
+
+    void check_metadata_lora(file_input * adapter) {
+        auto general_type = get_kv_str(adapter->ctx_gguf, "general.type");
+        if (general_type != "adapter") {
+            throw std::runtime_error("expect general.type to be 'adapter', but got: " + general_type);
+        }
+
+        auto adapter_type = get_kv_str(adapter->ctx_gguf, "adapter.type");
+        if (adapter_type != "lora") {
+            throw std::runtime_error("expect adapter.type to be 'lora', but got: " + adapter_type);
+        }
+
+        auto general_arch_base = get_kv_str(base_model.ctx_gguf, "general.architecture");
+        auto general_arch_lora = get_kv_str(adapter->ctx_gguf,   "general.architecture");
+        if (general_arch_base != general_arch_lora) {
+            throw std::runtime_error("model arch and LoRA arch mismatch");
+        }
+    }
+
+    ggml_type get_out_tensor_type(struct ggml_tensor * t) {
+        if (t->type == GGML_TYPE_F32) {
+            return GGML_TYPE_F32;
+        } else {
+            return GGML_TYPE_F16;
+        }
+    }
+
+    void run_merge() {
+        // prepare metadata
+        gguf_set_kv(ctx_out, base_model.ctx_gguf);
+        // output is forced to f16 for now
+        gguf_set_val_u32(ctx_out, "general.file_type", LLAMA_FTYPE_MOSTLY_F16);
+
+        // check if all lora adapters have the same tensors
+        // TODO: remove this when we can support merging subset of adapters. Ref: https://github.com/ggerganov/llama.cpp/pull/8607#discussion_r1686027777
+        static const char * err_no_subset_adapter = "Input adapters do not have the same list of tensors. This is not yet supported. Please merge the adapter one-by-one instead of merging all at once.";
+        if (adapters.size() > 1) {
+            for (size_t i = 1; i < adapters.size(); ++i) {
+                if (adapters[0]->tensors.size() != adapters[i]->tensors.size()) {
+                    throw std::runtime_error(err_no_subset_adapter);
+                }
+                for (auto & it : adapters[i]->tensors) {
+                    if (adapters[0]->get_tensor(it.first) == nullptr) {
+                        throw std::runtime_error(err_no_subset_adapter);
+                    }
+                }
+            }
+        }
+
+        // mapping base tensor to out tensor (same shape with base, but different type)
+        std::vector<tensor_transformation> trans;
+        for (auto & it : base_model.tensors) {
+            bool t_a = true;
+            bool t_b = true;
+            for (auto & adapter : adapters) {
+                t_a &= nullptr != adapter->get_tensor(it.first + ".lora_a");
+                t_b &= nullptr != adapter->get_tensor(it.first + ".lora_b");
+            }
+            auto base_tensor = it.second;
+            if (!t_a && !t_b) {
+                // only copy
+                struct ggml_tensor * cpy_tensor = ggml_dup_tensor(ctx_out_ggml, base_tensor);
+                ggml_set_name(cpy_tensor, base_tensor->name);
+                trans.push_back({
+                    cpy_tensor,
+                    cpy_tensor,
+                    true,
+                });
+                gguf_add_tensor(ctx_out, cpy_tensor);
+            } else if (t_a && t_b) {
+                // need merging
+                struct ggml_tensor * out_tensor = ggml_new_tensor(
+                    ctx_out_ggml, get_out_tensor_type(base_tensor), GGML_MAX_DIMS, base_tensor->ne);
+                ggml_set_name(out_tensor, base_tensor->name);
+                trans.push_back({
+                    base_tensor,
+                    out_tensor,
+                    false,
+                });
+                gguf_add_tensor(ctx_out, out_tensor);
+            } else {
+                throw std::runtime_error("tensor " + it.first + " missing either lora_a or lora_b");
+            }
+        }
+
+        // placeholder for the meta data
+        {
+            size_t meta_size = gguf_get_meta_size(ctx_out);
+            zeros(fout, meta_size);
+        }
+
+        // process base model tensors
+        size_t n_merged = 0;
+        for (auto & it : trans) {
+            if (!it.is_copy) {
+                merge_tensor(it.in, it.out);
+                n_merged++;
+            } else {
+                copy_tensor(it.in);
+            }
+        }
+
+        // write output metadata
+        {
+            std::vector<uint8_t> data(gguf_get_meta_size(ctx_out));
+            gguf_get_meta_data(ctx_out, data.data());
+            fout.seekp(0);
+            fout.write((const char *)data.data(), data.size());
+        }
+
+        printf("%s : merged %zu tensors with lora adapters\n", __func__, n_merged);
+        printf("%s : wrote %zu tensors to output file\n", __func__, trans.size());
+    }
+
+    void copy_tensor(struct ggml_tensor * base) {
+        printf("%s :  %s [%s]\n", __func__, base->name, ggml_ne_string(base).c_str());
+        size_t len = ggml_nbytes(base);
+        base_model.read_tensor_data(base->name, read_buf);
+        fout.write((char* )read_buf.data(), len);
+        zeros(fout, GGML_PAD(len, GGUF_DEFAULT_ALIGNMENT) - len);
+    }
+
+    void merge_tensor(struct ggml_tensor * base, struct ggml_tensor * out) {
+        std::string name_base(base->name);
+        std::string name_lora_a = name_base + ".lora_a";
+        std::string name_lora_b = name_base + ".lora_b";
+
+        printf("%s : %s [%s]\n", __func__, base->name, ggml_ne_string(base).c_str());
+
+        // context for input tensor
+        std::vector<struct ggml_tensor *> inp_a(adapters.size());
+        std::vector<struct ggml_tensor *> inp_b(adapters.size());
+        struct ggml_init_params params {
+            /*.mem_size   =*/ ggml_tensor_overhead()*(2+adapters.size()*2),
+            /*.mem_buffer =*/ NULL,
+            /*.no_alloc   =*/ true,
+        };
+        struct ggml_context * ctx = ggml_init(params);
+
+        // alloc tensors
+        struct ggml_tensor * inp_base = ggml_new_tensor(ctx, GGML_TYPE_F32, GGML_MAX_DIMS, base->ne);
+        for (size_t i = 0; i < adapters.size(); ++i) {
+            auto t_a = adapters[i]->get_tensor(name_lora_a);
+            auto t_b = adapters[i]->get_tensor(name_lora_b);
+            // TODO: add support for quantized lora
+            if (ggml_is_quantized(t_a->type) || ggml_is_quantized(t_b->type)) {
+                throw std::runtime_error("quantized LoRA adapters is not supported, please retry with f16 or f32");
+            }
+            inp_a[i] = ggml_dup_tensor(ctx, t_a);
+            inp_b[i] = ggml_dup_tensor(ctx, t_b);
+        }
+        ggml_backend_buffer_t buffer = ggml_backend_alloc_ctx_tensors(ctx, backend);
+
+        // load base tensor to backend buffer
+        base_model.read_tensor_data(name_base, read_buf);
+        if (base->type != GGML_TYPE_F32) {
+            // optionally dequantize it
+            printf("%s :   + dequantize base tensor from %s to F32\n", __func__, ggml_type_name(base->type));
+            auto nels = ggml_nelements(inp_base);
+            const auto * qtype = ggml_get_type_traits(base->type);
+            std::vector<uint8_t> dequant_buf(nels * sizeof(float));
+            qtype->to_float(read_buf.data(), (float *)dequant_buf.data(), nels);
+            ggml_backend_tensor_set(inp_base, dequant_buf.data(), 0, dequant_buf.size());
+        } else {
+            ggml_backend_tensor_set(inp_base, read_buf.data(), 0, ggml_nbytes(inp_base));
+        }
+
+        // load lora tensors to backend buffer
+        for (size_t i = 0; i < adapters.size(); ++i) {
+            adapters[i]->read_tensor_data(name_lora_a, read_buf);
+            ggml_backend_tensor_set(inp_a[i], read_buf.data(), 0, ggml_nbytes(inp_a[i]));
+            adapters[i]->read_tensor_data(name_lora_b, read_buf);
+            ggml_backend_tensor_set(inp_b[i], read_buf.data(), 0, ggml_nbytes(inp_b[i]));
+        }
+
+        // build graph
+        struct ggml_cgraph * gf;
+        {
+            static size_t buf_size = ggml_tensor_overhead()*GGML_DEFAULT_GRAPH_SIZE + ggml_graph_overhead();
+            static std::vector<uint8_t> buf(buf_size);
+            struct ggml_init_params params0 = {
+                /*.mem_size   =*/ buf_size,
+                /*.mem_buffer =*/ buf.data(),
+                /*.no_alloc   =*/ true,
+            };
+            struct ggml_context * ctx0 = ggml_init(params0);
+            gf = ggml_new_graph(ctx0);
+            struct ggml_tensor * cur = inp_base;
+            for (size_t i = 0; i < adapters.size(); ++i) {
+                struct ggml_tensor * delta;
+                bool is_tok_embd = string_starts_with(name_base, "token_embd");
+                if (is_tok_embd) {
+                    printf("%s :     detected token embeddings tensor\n", __func__);
+                    delta = ggml_mul_mat(ctx0,
+                        ggml_cast(ctx0, inp_b[i], GGML_TYPE_F32),
+                        ggml_cast(ctx0, inp_a[i], GGML_TYPE_F32));
+                } else {
+                    delta = ggml_mul_mat(ctx0,
+                        ggml_cont(ctx0, ggml_transpose(ctx0, ggml_cast(ctx0, inp_a[i], GGML_TYPE_F32))),
+                        ggml_cast(ctx0, inp_b[i], GGML_TYPE_F32));
+                }
+                // scale
+                const float alpha = adapters[i]->alpha;
+                const float rank  = (float) inp_b[i]->ne[0];
+                const float scale = alpha ? adapters[i]->scale * alpha / rank : adapters[i]->scale;
+                delta = ggml_scale(ctx0, delta, scale);
+                cur = ggml_add(ctx0, delta, cur);
+                printf("%s :   + merging from adapter[%zu] type=%s\n", __func__, i, ggml_type_name(inp_a[i]->type));
+                printf("%s :     input_scale=%f calculated_scale=%f rank=%d\n", __func__, adapters[i]->scale, scale, (int) inp_b[i]->ne[0]);
+            }
+            cur = ggml_cast(ctx0, cur, out->type);
+            printf("%s :   + output type is %s\n", __func__, ggml_type_name(out->type));
+            ggml_build_forward_expand(gf, cur);
+            ggml_free(ctx0);
+        }
+
+        // compute
+        {
+            ggml_gallocr_alloc_graph(allocr, gf);
+            ggml_backend_cpu_set_n_threads(backend, n_threads);
+            ggml_backend_graph_compute(backend, gf);
+        }
+
+        // write data to output file
+        {
+            auto * result = ggml_graph_node(gf, -1);
+            size_t len = ggml_nbytes(result);
+            if (read_buf.size() < len) {
+                read_buf.resize(len);
+            }
+            ggml_backend_tensor_get(result, read_buf.data(), 0, len);
+            fout.write((char* )read_buf.data(), len);
+            zeros(fout, GGML_PAD(len, GGUF_DEFAULT_ALIGNMENT) - len);
+        }
+
+        ggml_free(ctx);
+        ggml_backend_buffer_free(buffer);
+    }
+
+    ~lora_merge_ctx() {
+        ggml_gallocr_free(allocr);
+        ggml_backend_free(backend);
+        gguf_free(ctx_out);
+        ggml_free(ctx_out_ggml);
+    }
+};
+
+static void print_usage(int, char ** argv) {
+    printf("\nexample usage:\n");
+    printf("\n  %s -m base-model.gguf --lora lora-file.gguf -o merged-model-f16.gguf\n", argv[0]);
+    printf("\nNOTE: output model is F16\n");
+    printf("\n");
+}
+
+int main(int argc, char ** argv) {
+    common_params params;
+
+    params.out_file = "ggml-lora-merged-f16.gguf";
+
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_EXPORT_LORA, print_usage)) {
+        return 1;
+    }
+
+    g_verbose = (params.verbosity > 1);
+    try {
+        lora_merge_ctx ctx(params.model.path, params.lora_adapters, params.out_file, params.cpuparams.n_threads);
+        ctx.run_merge();
+    } catch (const std::exception & err) {
+        fprintf(stderr, "%s\n", err.what());
+        exit(EXIT_FAILURE);
+    }
+
+    printf("done, output file is %s\n", params.out_file.c_str());
+
+    return 0;
+}

tools/fit-params/CMakeLists.txt

@@ -0,0 +1,8 @@
+set(TARGET llama-fit-params)
+add_executable(${TARGET} fit-params.cpp)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_17)
+
+if(LLAMA_TOOLS_INSTALL)
+    install(TARGETS ${TARGET} RUNTIME)
+endif()

tools/fit-params/README.md

@@ -0,0 +1,55 @@
+# fit-params
+
+llama.cpp binaries can automatically fit the projected memory use of a model to the free device memory available at runtime,
+this is controlled using the CLI arguments starting with `-fit`/`--fit`.
+Internally the code is calling `llama_params_fit` to adjust the `llama_model_params` and `llama_context_params` structs.
+`llama-fit-params` is a simple utility that prints the CLI arguments corresponding to these adjustments to stdout.
+Example usage:
+
+``` bash
+# First, run llama-fit-params and store the results in a file:
+> ./build/bin/llama-fit-params --model /opt/models/qwen_3-30b3a-f16.gguf | tee args.txt
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 CUDA devices:
+  Device 0: NVIDIA GeForce RTX 4090, compute capability 8.9, VMM: yes
+build: 6895 (4341dc8bc) with cc (GCC) 15.2.1 20250813 for x86_64-pc-linux-gnu
+llama_params_fit_impl: projected to use 61807 MiB of device memory vs. 24077 MiB of free device memory
+llama_params_fit_impl: cannot fulfill margin of 1024 MiB, need to reduce device memory by 42444 MiB
+llama_params_fit_impl: context size reduced from 40960 to 4096 -> need 3456 MiB less memory in total
+llama_params_fit_impl: with only dense weights in device memory there is a total surplus of 16164 MiB
+llama_params_fit_impl: distributing layers across devices with overflow to next device/system memory:
+llama_params_fit_impl:   - CUDA0 (NVIDIA GeForce RTX 4090): 48 layers (34 overflowing),  19187 MiB used,   1199 MiB free
+llama_params_fit: successfully fit params to free device memory
+llama_params_fit: fitting params to free memory took 1.15 seconds
+Printing fitted CLI arguments to stdout...
+-c 4096 -ngl 48 -ot blk\.14\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.15\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.16\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.17\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.18\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.19\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.20\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.21\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.22\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.23\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.24\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.25\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.26\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.27\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.28\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.29\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.30\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.31\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.32\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.33\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.34\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.35\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.36\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.37\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.38\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.39\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.40\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.41\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.42\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.43\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.44\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.45\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.46\.ffn_(up|down|gate)_(ch|)exps=CPU,blk\.47\.ffn_(up|down|gate)_(ch|)exps=CPU
+
+# Next, use those results for a llama.cpp binary:
+> cat args.txt | xargs ./build/bin/llama-server --model /opt/models/qwen_3-30b3a-f16.gguf
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 CUDA devices:
+  Device 0: NVIDIA GeForce RTX 4090, compute capability 8.9, VMM: yes
+build: 6895 (4341dc8bc) with cc (GCC) 15.2.1 20250813 for x86_64-pc-linux-gnu
+system info: n_threads = 16, n_threads_batch = 16, total_threads = 32
+
+system_info: n_threads = 16 (n_threads_batch = 16) / 32 | CUDA : ARCHS = 890 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX_VNNI = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 |
+
+main: binding port with default address family
+main: HTTP server is listening, hostname: 127.0.0.1, port: 8080, http threads: 31
+main: loading model
+srv    load_model: loading model '/opt/models/qwen_3-30b3a-f16.gguf'
+llama_params_fit_impl: projected to use 19187 MiB of device memory vs. 24077 MiB of free device memory
+llama_params_fit_impl: will leave 1199 >= 1024 MiB of free device memory, no changes needed
+llama_params_fit: successfully fit params to free device memory
+llama_params_fit: fitting params to free memory took 0.28 seconds
+[...]
+main: server is listening on http://127.0.0.1:8080 - starting the main loop
+srv  update_slots: all slots are idle
+^Csrv    operator(): operator(): cleaning up before exit...
+
+llama_memory_breakdown_print: | memory breakdown [MiB] | total   free     self   model   context   compute    unaccounted |
+llama_memory_breakdown_print: |   - CUDA0 (RTX 4090)   | 24077 =  945 + (19187 = 17904 +     384 +     898) +        3945 |
+llama_memory_breakdown_print: |   - Host               |                 58271 = 58259 +       0 +      12                |
+```

tools/fit-params/fit-params.cpp

@@ -0,0 +1,66 @@
+#include "llama.h"
+
+#include "arg.h"
+#include "common.h"
+#include "log.h"
+
+#include <chrono>
+#include <cinttypes>
+#include <thread>
+
+using namespace std::chrono_literals;
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+int main(int argc, char ** argv) {
+    common_params params;
+
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
+        return 1;
+    }
+
+    common_init();
+    llama_backend_init();
+    llama_numa_init(params.numa);
+    auto mparams = common_model_params_to_llama(params);
+    auto cparams = common_context_params_to_llama(params);
+    const llama_params_fit_status status = llama_params_fit(params.model.path.c_str(), &mparams, &cparams,
+        params.tensor_split, params.tensor_buft_overrides.data(), params.fit_params_target.data(), params.fit_params_min_ctx,
+        params.verbosity >= 4 ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_ERROR);
+    if (status != LLAMA_PARAMS_FIT_STATUS_SUCCESS) {
+        LOG_ERR("%s: failed to fit CLI arguments to free memory, exiting...\n", __func__);
+        exit(1);
+    }
+
+    LOG_INF("%s: printing fitted CLI arguments to stdout...\n", __func__);
+    common_log_flush(common_log_main());
+    printf("-c %" PRIu32 " -ngl %" PRIu32, cparams.n_ctx, mparams.n_gpu_layers);
+
+    size_t nd = llama_max_devices();
+    while (nd > 1 && mparams.tensor_split[nd - 1] == 0.0f) {
+        nd--;
+    }
+    if (nd > 1) {
+        for (size_t id = 0; id < nd; id++) {
+            if (id == 0) {
+                printf(" -ts ");
+            }
+            printf("%s%" PRIu32, id > 0 ? "," : "", uint32_t(mparams.tensor_split[id]));
+        }
+    }
+
+    const size_t ntbo = llama_max_tensor_buft_overrides();
+    bool any_tbo = false;
+    for (size_t itbo = 0; itbo < ntbo && mparams.tensor_buft_overrides[itbo].pattern != nullptr; itbo++) {
+        if (itbo == 0) {
+            printf(" -ot \"");
+        }
+        printf("%s%s=%s", itbo > 0 ? "," : "", mparams.tensor_buft_overrides[itbo].pattern, ggml_backend_buft_name(mparams.tensor_buft_overrides[itbo].buft));
+        any_tbo = true;
+    }
+    printf("%s\n", any_tbo ? "\"" : "");
+
+    return 0;
+}

tools/gguf-split/CMakeLists.txt

@@ -0,0 +1,8 @@
+set(TARGET llama-gguf-split)
+add_executable(${TARGET} gguf-split.cpp)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_17)
+
+if(LLAMA_TOOLS_INSTALL)
+    install(TARGETS ${TARGET} RUNTIME)
+endif()

tools/gguf-split/README.md

@@ -0,0 +1,10 @@
+## GGUF split Example
+
+CLI to split / merge GGUF files.
+
+**Command line options:**
+
+- `--split`: split GGUF to multiple GGUF, default operation.
+- `--split-max-size`: max size per split in `M` or `G`, f.ex. `500M` or `2G`.
+- `--split-max-tensors`: maximum tensors in each split: default(128)
+- `--merge`: merge multiple GGUF to a single GGUF.

tools/gguf-split/gguf-split.cpp

@@ -0,0 +1,583 @@
+#include "ggml.h"
+#include "gguf.h"
+#include "llama.h"
+#include "common.h"
+
+#include <algorithm>
+#include <cinttypes>
+#include <climits>
+#include <cstdio>
+#include <cstdlib>
+#include <stdexcept>
+#include <cstring>
+#include <fstream>
+#include <string>
+#include <vector>
+
+#if defined(_WIN32)
+    #include <windows.h>
+    #ifndef PATH_MAX
+        #define PATH_MAX MAX_PATH
+    #endif
+    #include <io.h>
+#endif
+
+enum split_operation : uint8_t {
+    OP_NONE,
+    OP_SPLIT,
+    OP_MERGE,
+};
+
+enum split_mode : uint8_t {
+    MODE_NONE,
+    MODE_TENSOR,
+    MODE_SIZE,
+};
+
+struct split_params {
+    split_operation operation = OP_NONE;
+    split_mode mode = MODE_NONE;
+    size_t n_bytes_split = 0;
+    int n_split_tensors = 128;
+    std::string input;
+    std::string output;
+    bool no_tensor_first_split = false;
+    bool dry_run = false;
+};
+
+static void split_print_usage(const char * executable) {
+    const split_params default_params;
+    printf("\n");
+    printf("usage: %s [options] GGUF_IN GGUF_OUT\n", executable);
+    printf("\n");
+    printf("Apply a GGUF operation on IN to OUT.");
+    printf("\n");
+    printf("options:\n");
+    printf("  -h, --help              show this help message and exit\n");
+    printf("  --version               show version and build info\n");
+    printf("  --split                 split GGUF to multiple GGUF (enabled by default)\n");
+    printf("  --merge                 merge multiple GGUF to a single GGUF\n");
+    printf("  --split-max-tensors     max tensors in each split (default: %d)\n", default_params.n_split_tensors);
+    printf("  --split-max-size N(M|G) max size per split\n");
+    printf("  --no-tensor-first-split do not add tensors to the first split (disabled by default)\n");
+    printf("  --dry-run               only print out a split plan and exit, without writing any new files\n");
+    printf("\n");
+}
+
+// return convert string, for example "128M" or "4G" to number of bytes
+static size_t split_str_to_n_bytes(std::string str) {
+    size_t n_bytes = 0;
+    int n;
+    if (str.back() == 'M') {
+        sscanf(str.c_str(), "%d", &n);
+        n_bytes = (size_t)n * 1000 * 1000; // megabytes
+    } else if (str.back() == 'G') {
+        sscanf(str.c_str(), "%d", &n);
+        n_bytes = (size_t)n * 1000 * 1000 * 1000; // gigabytes
+    } else {
+        throw std::invalid_argument("error: supported units are M (megabytes) or G (gigabytes), but got: " + std::string(1, str.back()));
+    }
+    if (n <= 0) {
+        throw std::invalid_argument("error: size must be a positive value");
+    }
+    return n_bytes;
+}
+
+static void split_params_parse_ex(int argc, const char ** argv, split_params & params) {
+    std::string arg;
+    const std::string arg_prefix = "--";
+    bool invalid_param = false;
+
+    int arg_idx = 1;
+    for (; arg_idx < argc && strncmp(argv[arg_idx], "--", 2) == 0; arg_idx++) {
+        arg = argv[arg_idx];
+        if (arg.compare(0, arg_prefix.size(), arg_prefix) == 0) {
+            std::replace(arg.begin(), arg.end(), '_', '-');
+        }
+
+        bool arg_found = false;
+        if (arg == "-h" || arg == "--help") {
+            split_print_usage(argv[0]);
+            exit(0);
+        } else if (arg == "--version") {
+            fprintf(stderr, "version: %d (%s)\n", LLAMA_BUILD_NUMBER, LLAMA_COMMIT);
+            fprintf(stderr, "built with %s for %s\n", LLAMA_COMPILER, LLAMA_BUILD_TARGET);
+            exit(0);
+        } else if (arg == "--dry-run") {
+            arg_found = true;
+            params.dry_run = true;
+        } else if (arg == "--no-tensor-first-split") {
+            arg_found = true;
+            params.no_tensor_first_split = true;
+        } else if (arg == "--merge") {
+            arg_found = true;
+            if (params.operation != OP_NONE && params.operation != OP_MERGE) {
+                throw std::invalid_argument("error: either --split or --merge can be specified, but not both");
+            }
+            params.operation = OP_MERGE;
+        } else if (arg == "--split") {
+            arg_found = true;
+            if (params.operation != OP_NONE && params.operation != OP_SPLIT) {
+                throw std::invalid_argument("error: either --split or --merge can be specified, but not both");
+            }
+            params.operation = OP_SPLIT;
+        } else if (arg == "--split-max-tensors") {
+            if (++arg_idx >= argc) {
+                invalid_param = true;
+                break;
+            }
+            arg_found = true;
+            if (params.mode != MODE_NONE && params.mode != MODE_TENSOR) {
+                throw std::invalid_argument("error: either --split-max-tensors or --split-max-size can be specified, but not both");
+            }
+            params.mode = MODE_TENSOR;
+            params.n_split_tensors = atoi(argv[arg_idx]);
+        } else if (arg == "--split-max-size") {
+            if (++arg_idx >= argc) {
+                invalid_param = true;
+                break;
+            }
+            arg_found = true;
+            if (params.mode != MODE_NONE && params.mode != MODE_SIZE) {
+                throw std::invalid_argument("error: either --split-max-tensors or --split-max-size can be specified, but not both");
+            }
+            params.mode = MODE_SIZE;
+            params.n_bytes_split = split_str_to_n_bytes(argv[arg_idx]);
+        }
+
+        if (!arg_found) {
+            throw std::invalid_argument("error: unknown argument: " + arg);
+        }
+    }
+
+    // the operation is split if not specified
+    if (params.operation == OP_NONE) {
+        params.operation = OP_SPLIT;
+    }
+    // the split mode is by tensor if not specified
+    if (params.mode == MODE_NONE) {
+        params.mode = MODE_TENSOR;
+    }
+
+    if (invalid_param) {
+        throw std::invalid_argument("error: invalid parameter for argument: " + arg);
+    }
+
+    if (argc - arg_idx != 2) {
+        throw std::invalid_argument("error: bad arguments");
+    }
+
+    params.input = argv[arg_idx++];
+    params.output = argv[arg_idx++];
+}
+
+static bool split_params_parse(int argc, const char ** argv, split_params & params) {
+    bool result = true;
+    try {
+        split_params_parse_ex(argc, argv, params);
+    }
+    catch (const std::invalid_argument & ex) {
+        fprintf(stderr, "%s\n", ex.what());
+        split_print_usage(argv[0]);
+        exit(EXIT_FAILURE);
+    }
+    return result;
+}
+
+static void zeros(std::ofstream & file, size_t n) {
+    char zero = 0;
+    for (size_t i = 0; i < n; ++i) {
+        file.write(&zero, 1);
+    }
+}
+
+struct split_strategy {
+    const split_params params;
+    std::ifstream & f_input;
+    struct gguf_context * ctx_gguf;
+    struct ggml_context * ctx_meta = NULL;
+    const int n_tensors;
+
+    // one ctx_out per one output file
+    std::vector<struct gguf_context *> ctx_outs;
+
+    // temporary buffer for reading in tensor data
+    std::vector<uint8_t> read_buf;
+
+    split_strategy(const split_params & params,
+            std::ifstream & f_input,
+            struct gguf_context * ctx_gguf,
+            struct ggml_context * ctx_meta) :
+        params(params),
+        f_input(f_input),
+        ctx_gguf(ctx_gguf),
+        ctx_meta(ctx_meta),
+        n_tensors(gguf_get_n_tensors(ctx_gguf)) {
+
+        // because we need to know list of tensors for each file in advance, we will build all the ctx_out for all output splits
+        int i_split = -1;
+        struct gguf_context * ctx_out = NULL;
+        auto new_ctx_out = [&](bool allow_no_tensors) {
+            i_split++;
+            if (ctx_out != NULL) {
+                if (gguf_get_n_tensors(ctx_out) == 0 && !allow_no_tensors) {
+                    fprintf(stderr, "error: one of splits have 0 tensors. Maybe size or tensors limit is too small\n");
+                    exit(EXIT_FAILURE);
+                }
+                ctx_outs.push_back(ctx_out);
+            }
+            ctx_out = gguf_init_empty();
+            // Save all metadata in first split only
+            if (i_split == 0) {
+                gguf_set_kv(ctx_out, ctx_gguf);
+            }
+            gguf_set_val_u16(ctx_out, LLM_KV_SPLIT_NO, i_split);
+            gguf_set_val_u16(ctx_out, LLM_KV_SPLIT_COUNT, 0); // placeholder
+            gguf_set_val_i32(ctx_out, LLM_KV_SPLIT_TENSORS_COUNT, n_tensors);
+        };
+
+        // initialize ctx_out for the first split
+        new_ctx_out(false);
+
+        // skip first split if no_tensor_first_split is set
+        if (params.no_tensor_first_split) {
+            new_ctx_out(true);
+        }
+
+        // process tensors one by one
+        size_t curr_tensors_size = 0; // current size by counting only tensors size (without metadata)
+        for (int i = 0; i < n_tensors; ++i) {
+            struct ggml_tensor * t = ggml_get_tensor(ctx_meta, gguf_get_tensor_name(ctx_gguf, i));
+            // calculate the "imaginary" size = the current size + next tensor size
+            size_t n_bytes = GGML_PAD(ggml_nbytes(t), GGUF_DEFAULT_ALIGNMENT);
+            size_t next_tensors_size = curr_tensors_size + n_bytes;
+            if (should_split(i, next_tensors_size)) {
+                new_ctx_out(false);
+                curr_tensors_size = n_bytes;
+            } else {
+                curr_tensors_size = next_tensors_size;
+            }
+            gguf_add_tensor(ctx_out, t);
+        }
+
+        // push the last ctx_out
+        ctx_outs.push_back(ctx_out);
+
+        // set the correct n_split for all ctx_out
+        for (auto & ctx : ctx_outs) {
+            gguf_set_val_u16(ctx, LLM_KV_SPLIT_COUNT, ctx_outs.size());
+        }
+    }
+
+    ~split_strategy() {
+        for (auto & ctx_out : ctx_outs) {
+            gguf_free(ctx_out);
+        }
+    }
+
+    bool should_split(int i_tensor, size_t next_size) {
+        if (params.mode == MODE_SIZE) {
+            // split by max size per file
+            return next_size > params.n_bytes_split;
+        } else if (params.mode == MODE_TENSOR) {
+            // split by number of tensors per file
+            return i_tensor > 0 && i_tensor < n_tensors && i_tensor % params.n_split_tensors == 0;
+        }
+        // should never happen
+        GGML_ABORT("invalid mode");
+    }
+
+    void print_info() {
+        printf("n_split: %zu\n", ctx_outs.size());
+        int i_split = 0;
+        for (auto & ctx_out : ctx_outs) {
+            // re-calculate the real gguf size for each split (= metadata size + total size of all tensors)
+            size_t total_size = gguf_get_meta_size(ctx_out);
+            for (int i = 0; i < gguf_get_n_tensors(ctx_out); ++i) {
+                struct ggml_tensor * t = ggml_get_tensor(ctx_meta, gguf_get_tensor_name(ctx_out, i));
+                total_size += ggml_nbytes(t);
+            }
+            total_size = total_size / 1000 / 1000; // convert to megabytes
+            printf("split %05d: n_tensors = %" PRIi64 ", total_size = %zuM\n", i_split + 1, gguf_get_n_tensors(ctx_out), total_size);
+            i_split++;
+        }
+    }
+
+    void write() {
+        int i_split = 0;
+        int n_split = ctx_outs.size();
+        for (auto & ctx_out : ctx_outs) {
+            // construct file path
+            char split_path[PATH_MAX] = {0};
+            llama_split_path(split_path, sizeof(split_path), params.output.c_str(), i_split, n_split);
+
+            // open the output file
+            printf("Writing file %s ... ", split_path);
+            fflush(stdout);
+            std::ofstream fout = std::ofstream(split_path, std::ios::binary);
+            fout.exceptions(std::ofstream::failbit); // fail fast on write errors
+
+            // write metadata
+            std::vector<uint8_t> data(gguf_get_meta_size(ctx_out));
+            gguf_get_meta_data(ctx_out, data.data());
+            fout.write((const char *)data.data(), data.size());
+
+            // write tensors
+            for (int i = 0; i < gguf_get_n_tensors(ctx_out); ++i) {
+                // read tensor meta and prepare buffer
+                const char * t_name = gguf_get_tensor_name(ctx_out, i);
+                struct ggml_tensor * t = ggml_get_tensor(ctx_meta, t_name);
+                auto n_bytes = ggml_nbytes(t);
+                read_buf.resize(n_bytes);
+
+                // calculate offset
+                auto i_tensor_in = gguf_find_tensor(ctx_gguf, t_name); // idx of tensor in the input file
+                auto offset = gguf_get_data_offset(ctx_gguf) + gguf_get_tensor_offset(ctx_gguf, i_tensor_in);
+
+                // copy tensor from input to output file
+                copy_file_to_file(f_input, fout, offset, n_bytes);
+                zeros(fout, GGML_PAD(n_bytes, GGUF_DEFAULT_ALIGNMENT) - n_bytes);
+            }
+
+            printf("done\n");
+            // close the file
+            fout.close();
+            i_split++;
+        }
+    }
+
+    void copy_file_to_file(std::ifstream & f_in, std::ofstream & f_out, const size_t in_offset, const size_t len) {
+        // TODO: detect OS and use copy_file_range() here for better performance
+        if (read_buf.size() < len) {
+            read_buf.resize(len);
+        }
+        f_in.seekg(in_offset);
+        f_in.read((char *)read_buf.data(), len);
+        f_out.write((const char *)read_buf.data(), len);
+    }
+};
+
+static void gguf_split(const split_params & split_params) {
+    struct ggml_context * ctx_meta = NULL;
+
+    struct gguf_init_params params = {
+        /*.no_alloc = */ true,
+        /*.ctx      = */ &ctx_meta,
+    };
+
+    std::ifstream f_input(split_params.input.c_str(), std::ios::binary);
+    if (!f_input.is_open()) {
+        fprintf(stderr, "%s:  failed to open input GGUF from %s\n", __func__, split_params.input.c_str());
+        exit(EXIT_FAILURE);
+    }
+
+    auto * ctx_gguf = gguf_init_from_file(split_params.input.c_str(), params);
+    if (!ctx_gguf) {
+        fprintf(stderr, "%s:  failed to load input GGUF from %s\n", __func__, split_params.input.c_str());
+        exit(EXIT_FAILURE);
+    }
+
+    // prepare the strategy
+    split_strategy strategy(split_params, f_input, ctx_gguf, ctx_meta);
+    int n_split = strategy.ctx_outs.size();
+    strategy.print_info();
+
+    if (!split_params.dry_run) {
+        // write all output splits
+        strategy.write();
+    }
+
+    // done, clean up
+    gguf_free(ctx_gguf);
+    f_input.close();
+
+    fprintf(stderr, "%s: %d gguf split written with a total of %d tensors.\n",
+            __func__, n_split, strategy.n_tensors);
+}
+
+static void gguf_merge(const split_params & split_params) {
+    fprintf(stderr, "%s: %s -> %s\n",
+            __func__, split_params.input.c_str(),
+            split_params.output.c_str());
+    int n_split = 1;
+    int total_tensors = 0;
+
+    // avoid overwriting existing output file
+    if (std::ifstream(split_params.output.c_str())) {
+        fprintf(stderr, "%s: output file %s already exists\n", __func__, split_params.output.c_str());
+        exit(EXIT_FAILURE);
+    }
+
+
+    auto * ctx_out = gguf_init_empty();
+
+    std::vector<uint8_t> read_data;
+    std::vector<ggml_context *> ctx_metas;
+    std::vector<gguf_context *> ctx_ggufs;
+
+    char split_path[PATH_MAX] = {0};
+    strncpy(split_path, split_params.input.c_str(), sizeof(split_path) - 1);
+    char split_prefix[PATH_MAX] = {0};
+
+    // First pass to find KV and tensors metadata
+    for (int i_split = 0; i_split < n_split; i_split++) {
+        struct ggml_context * ctx_meta = NULL;
+
+        struct gguf_init_params params = {
+            /*.no_alloc = */ true,
+            /*.ctx      = */ &ctx_meta,
+        };
+
+        if (i_split > 0) {
+            llama_split_path(split_path, sizeof(split_path), split_prefix, i_split, n_split);
+        }
+        fprintf(stderr, "%s: reading metadata %s ...", __func__, split_path);
+
+        auto * ctx_gguf = gguf_init_from_file(split_path, params);
+        if (!ctx_gguf) {
+            fprintf(stderr, "\n%s:  failed to load input GGUF from %s\n", __func__, split_params.input.c_str());
+            exit(EXIT_FAILURE);
+        }
+        ctx_ggufs.push_back(ctx_gguf);
+        ctx_metas.push_back(ctx_meta);
+
+        if (i_split == 0) {
+            auto key_n_split = gguf_find_key(ctx_gguf, LLM_KV_SPLIT_COUNT);
+            if (key_n_split < 0) {
+                fprintf(stderr,
+                        "\n%s: input file does not contain %s metadata\n",
+                        __func__,
+                        LLM_KV_SPLIT_COUNT);
+                gguf_free(ctx_gguf);
+                ggml_free(ctx_meta);
+                gguf_free(ctx_out);
+                exit(EXIT_FAILURE);
+            }
+
+            n_split = gguf_get_val_u16(ctx_gguf, key_n_split);
+            if (n_split < 1) {
+                fprintf(stderr,
+                        "\n%s: input file does not contain a valid split count %d\n",
+                        __func__,
+                        n_split);
+                gguf_free(ctx_gguf);
+                ggml_free(ctx_meta);
+                gguf_free(ctx_out);
+                exit(EXIT_FAILURE);
+            }
+
+            // Verify the file naming and extract split_prefix
+            if (!llama_split_prefix(split_prefix, sizeof (split_prefix), split_path, i_split, n_split)) {
+                fprintf(stderr, "\n%s: unexpected input file name: %s"
+                                " i_split=%d"
+                                " n_split=%d\n", __func__,
+                        split_path, i_split, n_split);
+                gguf_free(ctx_gguf);
+                ggml_free(ctx_meta);
+                gguf_free(ctx_out);
+                exit(EXIT_FAILURE);
+            }
+
+            // Do not trigger merge if we try to merge again the output
+            gguf_set_val_u16(ctx_gguf, LLM_KV_SPLIT_COUNT, 0);
+
+            // Set metadata from the first split
+            gguf_set_kv(ctx_out, ctx_gguf);
+        }
+
+        auto n_tensors = gguf_get_n_tensors(ctx_gguf);
+        for (int i_tensor = 0; i_tensor < n_tensors; i_tensor++) {
+            const char * t_name = gguf_get_tensor_name(ctx_gguf, i_tensor);
+            struct ggml_tensor * t = ggml_get_tensor(ctx_meta, t_name);
+            gguf_add_tensor(ctx_out, t);
+        }
+        total_tensors += n_tensors;
+
+        fprintf(stderr, "\033[3Ddone\n");
+    }
+    std::ofstream fout;
+    if (!split_params.dry_run) {
+        fout.open(split_params.output.c_str(), std::ios::binary);
+        fout.exceptions(std::ofstream::failbit); // fail fast on write errors
+        // placeholder for the meta data
+        auto meta_size = gguf_get_meta_size(ctx_out);
+        ::zeros(fout, meta_size);
+    }
+
+    // Write tensors data
+    for (int i_split = 0; i_split < n_split; i_split++) {
+        llama_split_path(split_path, sizeof(split_path), split_prefix, i_split, n_split);
+        std::ifstream f_input(split_path, std::ios::binary);
+        if (!f_input.is_open()) {
+            fprintf(stderr, "%s:  failed to open input GGUF from %s\n", __func__, split_path);
+            for (uint32_t i = 0; i < ctx_ggufs.size(); i++) {
+                gguf_free(ctx_ggufs[i]);
+                ggml_free(ctx_metas[i]);
+            }
+            gguf_free(ctx_out);
+            if (!split_params.dry_run) {
+                fout.close();
+            }
+            exit(EXIT_FAILURE);
+        }
+        fprintf(stderr, "%s: writing tensors %s ...", __func__, split_path);
+
+        auto * ctx_gguf = ctx_ggufs[i_split];
+        auto * ctx_meta = ctx_metas[i_split];
+
+        auto n_tensors = gguf_get_n_tensors(ctx_gguf);
+        for (int i_tensor = 0; i_tensor < n_tensors; i_tensor++) {
+            const char * t_name = gguf_get_tensor_name(ctx_gguf, i_tensor);
+            struct ggml_tensor * t = ggml_get_tensor(ctx_meta, t_name);
+
+            auto n_bytes = ggml_nbytes(t);
+
+            if (read_data.size() < n_bytes) {
+                read_data.resize(n_bytes);
+            }
+
+            auto offset = gguf_get_data_offset(ctx_gguf) + gguf_get_tensor_offset(ctx_gguf, i_tensor);
+            f_input.seekg(offset);
+            f_input.read((char *)read_data.data(), n_bytes);
+            if (!split_params.dry_run) {
+                // write tensor data + padding
+                fout.write((const char *)read_data.data(), n_bytes);
+                zeros(fout, GGML_PAD(n_bytes, GGUF_DEFAULT_ALIGNMENT) - n_bytes);
+            }
+        }
+
+        gguf_free(ctx_gguf);
+        ggml_free(ctx_meta);
+        f_input.close();
+        fprintf(stderr, "\033[3Ddone\n");
+    }
+
+    if (!split_params.dry_run) {
+        // go back to beginning of file and write the updated metadata
+        fout.seekp(0);
+        std::vector<uint8_t> data(gguf_get_meta_size(ctx_out));
+        gguf_get_meta_data(ctx_out, data.data());
+        fout.write((const char *)data.data(), data.size());
+        fout.close();
+    }
+    gguf_free(ctx_out);
+
+    fprintf(stderr, "%s: %s merged from %d split with %d tensors.\n",
+            __func__, split_params.output.c_str(), n_split, total_tensors);
+}
+
+int main(int argc, const char ** argv) {
+    split_params params;
+    split_params_parse(argc, argv, params);
+
+    switch (params.operation) {
+        case OP_SPLIT: gguf_split(params);
+            break;
+        case OP_MERGE: gguf_merge(params);
+            break;
+        default: split_print_usage(argv[0]);
+            exit(EXIT_FAILURE);
+    }
+
+    return 0;
+}

tools/gguf-split/tests.sh

@@ -0,0 +1,89 @@
+#!/usr/bin/env bash
+
+set -eu
+
+if [ $# -lt 1 ]
+then
+    echo "usage:   $0 path_to_build_binary [path_to_temp_folder]"
+    echo "example: $0 ../../build/bin ../../tmp"
+    exit 1
+fi
+
+if [ $# -gt 1 ]
+then
+    TMP_DIR=$2
+else
+    TMP_DIR=/tmp
+fi
+
+set -x
+
+SPLIT=$1/llama-gguf-split
+MAIN=$1/llama-completion
+WORK_PATH=$TMP_DIR/gguf-split
+ROOT_DIR=$(realpath $(dirname $0)/../../)
+
+mkdir -p "$WORK_PATH"
+
+# Clean up in case of previously failed test
+rm -f $WORK_PATH/ggml-model-split*.gguf $WORK_PATH/ggml-model-merge*.gguf
+
+# 1. Get a model
+(
+cd $WORK_PATH
+"$ROOT_DIR"/scripts/hf.sh --repo ggml-org/Qwen3-0.6B-GGUF --file Qwen3-0.6B-Q8_0.gguf
+)
+echo PASS
+
+# 2. Split with max tensors strategy
+$SPLIT --split-max-tensors 28  $WORK_PATH/Qwen3-0.6B-Q8_0.gguf $WORK_PATH/ggml-model-split
+echo PASS
+echo
+
+# 2b. Test the sharded model is loading properly
+$MAIN -no-cnv --model $WORK_PATH/ggml-model-split-00001-of-00012.gguf -p "I believe the meaning of life is" --n-predict 32
+echo PASS
+echo
+
+# 3. Merge
+$SPLIT --merge $WORK_PATH/ggml-model-split-00001-of-00012.gguf $WORK_PATH/ggml-model-merge.gguf
+echo PASS
+echo
+
+# 3b. Test the merged model is loading properly
+$MAIN -no-cnv --model $WORK_PATH/ggml-model-merge.gguf -p "I believe the meaning of life is" --n-predict 32
+echo PASS
+echo
+
+# 4. Split with no tensors in the first split
+$SPLIT --split-max-tensors 32 --no-tensor-first-split $WORK_PATH/ggml-model-merge.gguf $WORK_PATH/ggml-model-split-32-tensors
+echo PASS
+echo
+
+# 4b. Test the sharded model is loading properly
+$MAIN -no-cnv --model $WORK_PATH/ggml-model-split-32-tensors-00001-of-00011.gguf -p "I believe the meaning of life is" --n-predict 32
+echo PASS
+echo
+
+# 5. Merge
+#$SPLIT --merge $WORK_PATH/ggml-model-split-32-tensors-00001-of-00012.gguf $WORK_PATH/ggml-model-merge-2.gguf
+#echo PASS
+#echo
+
+# 5b. Test the merged model is loading properly
+#$MAIN -no-cnv --model $WORK_PATH/ggml-model-merge-2.gguf --n-predict 32
+#echo PASS
+#echo
+
+# 6. Split with size strategy
+$SPLIT --split-max-size 500M $WORK_PATH/ggml-model-merge.gguf $WORK_PATH/ggml-model-split-500M
+echo PASS
+echo
+
+# 6b. Test the sharded model is loading properly
+$MAIN -no-cnv --model $WORK_PATH/ggml-model-split-500M-00001-of-00002.gguf -p "I believe the meaning of life is" --n-predict 32
+echo PASS
+echo
+
+# Clean up
+rm -f $WORK_PATH/ggml-model-split*.gguf $WORK_PATH/ggml-model-merge*.gguf

tools/imatrix/CMakeLists.txt

@@ -0,0 +1,13 @@
+set(TARGET llama-imatrix)
+add_executable(${TARGET} imatrix.cpp)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_17)
+
+if(LLAMA_TOOLS_INSTALL)
+    install(TARGETS ${TARGET} RUNTIME)
+endif()
+
+if (CMAKE_SYSTEM_NAME MATCHES "AIX")
+    # AIX's flock() function comes from libbsd.a
+    target_link_libraries(${TARGET} PRIVATE -lbsd)
+endif()

tools/imatrix/README.md

@@ -0,0 +1,98 @@
+# llama.cpp/tools/imatrix
+
+Compute an importance matrix for a model and given text dataset. Can be used during quantization to enhance the quality of the quantized models.
+More information is available in <https://github.com/ggml-org/llama.cpp/pull/4861>.
+
+## Usage
+
+```
+./llama-imatrix \
+    -m model.gguf -f some-text.txt [-o imatrix.gguf] [--output-format {gguf,dat}] [--no-ppl] \
+    [--process-output] [--chunk 123] [--save-frequency 0] [--output-frequency 10] \
+    [--in-file imatrix-prev-0.gguf --in-file imatrix-prev-1.gguf ...] [--parse-special] \
+    [--show-statistics] [...]
+```
+
+Here `-m | --model` with a model name and `-f | --file` with a file containing calibration data (such as e.g. `wiki.train.raw`) are mandatory.
+The parameters in square brackets are optional and have the following meaning:
+
+* `-h | --help` shows usage information and exits.
+* `-lv | --verbosity` specifies the verbosity level. If set to `0`, no output other than the perplexity of the processed chunks will be generated. If set to `1`, each time the results are saved a message is written to `stderr`. If `>=2`, a message is output each time data is collected for any tensor. Default verbosity level is `1`.
+* `-o | --output-file` specifies the name of the file where the computed data will be stored. If missing `imatrix.gguf` is used.
+* `-ofreq | --output-frequency` specifies how often the so far computed result is saved to disk. Default is 10 (i.e., every 10 chunks)
+* `--output-format` specifies the output format of the generated imatrix file. Either "gguf", or "dat" (the legacy format). Defaults to "gguf".
+* `--save-frequency` specifies how often to save a copy of the imatrix in a separate file. Default is 0 (i.e., never)
+* `--process-output` specifies if data will be collected for the `output.weight` tensor. Typically, it is better not to utilize the importance matrix when quantizing `output.weight`, so this is set to `false` by default.
+* `--in-file` one or more existing imatrix files to load and combine. Useful for merging files from multiple runs/datasets.
+* `--parse-special` enables parsing of special tokens (e.g., `<|im_start|>` in some models). Useful for models with custom tokenizers.
+* `--chunk | --from-chunk` to skip the first `n` chunks of tokens from the input data. Useful for resuming or skipping initial low-quality data.
+* `--chunks` maximum number of chunks to process. Default is -1 for all available chunks.
+* `--no-ppl` disables the calculation of perplexity for the processed chunks. Useful if you want to speed up the processing and do not care about perplexity.
+* `--show-statistics` displays imatrix file's statistics.
+
+For faster computation, make sure to use GPU offloading via the `-ngl | --n-gpu-layers` argument.
+
+Recent versions of `llama-imatrix` store data in GGUF format by default. For the legacy format, use an extension other than `.gguf` when saving the output file. More information is available in <https://github.com/ggml-org/llama.cpp/pull/9400>.
+
+## Examples
+
+```bash
+# generate importance matrix using default filename (imatrix.gguf), offloading 99 layers to GPU
+./llama-imatrix -m ggml-model-f16.gguf -f calibration-data.txt -ngl 99
+
+# use the imatrix to perform a Q4_K_M quantization
+./llama-quantize --imatrix imatrix.gguf ggml-model-f16.gguf ./ggml-model-q4_k_m.gguf q4_k_m
+```
+
+```bash
+# generate and save the imatrix using legacy format
+./llama-imatrix -m ggml-model-f16.gguf -f calibration-data.txt --output-format dat -o imatrix-legcy-format.dat -ngl 99
+```
+
+```bash
+# convert legacy (binary) imatrix format to new (GGUF) format
+./llama-imatrix --in-file imatrix-legacy-format.dat -o imatrix-new-format.gguf
+```
+
+```bash
+# convert new (GGUF) imatrix format to legacy (binary) format
+./llama-imatrix --in-file imatrix-new-format.gguf --output-format dat -o imatrix-legacy-format.dat
+```
+
+```bash
+# combine existing imatrices
+./llama-imatrix --in-file imatrix-prev-0.gguf --in-file imatrix-prev-1.gguf -o imatrix-combined.gguf
+```
+
+```bash
+# skip first 5 chunks, save intermediates every 20 chunks and snapshots every 50, parsing special tokens
+./llama-imatrix -m ggml-model-f16.gguf -f calibration-data.txt --chunk 5 --output-frequency 20 --save-frequency 50 --parse-special
+```
+
+```bash
+# analyse imatrix file and display summary statistics instead of running inference
+./llama-imatrix --in-file imatrix.gguf --show-statistics
+```
+
+`--show-statistics` will display the following statistics:
+
+#### Per tensor
+
+* Σ(Act²): sum of all squared activations (the importance scores)
+* Min & Max: minimum and maximum squared activations values
+* μ & σ: Squared activations' mean and standard deviation
+* % Active: proportion of elements whose average squared activation exceeds a small threshold (1e-5). Helpful to determine how alive/dormant the tensor is during inference
+* N: number of squared activations
+* Entropy: entropy of the squared activation distribution, in bits (standard Shannon entropy measurement) $S = -\sum_{i=1}^N p_i \log_2 p_i$
+* E (norm): Normalized entropy. $E(norm)=\frac{-\sum_{i=1}^N p_i \log_2 p_i}{log_2 N}$. These two metrics can be used to determine how well a prompt "exercises" the model's capabilities
+* ZD Score: z-score distribution as described in _3.1 Layer Importance Scores_ of [Layer-Wise Quantization](https://arxiv.org/abs/2406.17415)
+* CosSim: cosine similarity with respect to the previous layer's tensor. Useful to determine how similar the squared activations of the current layer are to the previous layer's squared activations.
+
+#### Per layer
+
+Weighted averages of Σ(Act²), ZD Score and CosSim are also calculated.
+
+#### Important note on the computed Statistics
+
+When using these statistics, please note that they are computed on the squared activations, **not on the actual (raw) activations**.
+Whilst the results are still useful, they're less realiable than using the raw values, and in the case of the cosine similarity, could be misleading if the tensor contains opposite vectors.

tools/llama-bench/CMakeLists.txt

@@ -0,0 +1,8 @@
+set(TARGET llama-bench)
+add_executable(${TARGET} llama-bench.cpp)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_17)
+
+if(LLAMA_TOOLS_INSTALL)
+    install(TARGETS ${TARGET} RUNTIME)
+endif()

tools/llama-bench/README.md

@@ -0,0 +1,349 @@
+# llama.cpp/tools/llama-bench
+
+Performance testing tool for llama.cpp.
+
+## Table of contents
+
+1. [Syntax](#syntax)
+2. [Examples](#examples)
+    1. [Text generation with different models](#text-generation-with-different-models)
+    2. [Prompt processing with different batch sizes](#prompt-processing-with-different-batch-sizes)
+    3. [Different numbers of threads](#different-numbers-of-threads)
+    4. [Different numbers of layers offloaded to the GPU](#different-numbers-of-layers-offloaded-to-the-gpu)
+3. [Output formats](#output-formats)
+    1. [Markdown](#markdown)
+    2. [CSV](#csv)
+    3. [JSON](#json)
+    4. [JSONL](#jsonl)
+    5. [SQL](#sql)
+
+## Syntax
+
+```
+usage: llama-bench [options]
+
+options:
+  -h, --help
+  --numa <distribute|isolate|numactl>       numa mode (default: disabled)
+  -r, --repetitions <n>                     number of times to repeat each test (default: 5)
+  --prio <0|1|2|3>                          process/thread priority (default: 0)
+  --delay <0...N> (seconds)                 delay between each test (default: 0)
+  -o, --output <csv|json|jsonl|md|sql>      output format printed to stdout (default: md)
+  -oe, --output-err <csv|json|jsonl|md|sql> output format printed to stderr (default: none)
+  --list-devices                            list available devices and exit
+  -v, --verbose                             verbose output
+  --progress                                print test progress indicators
+  -rpc, --rpc <rpc_servers>                 register RPC devices (comma separated)
+
+test parameters:
+  -m, --model <filename>                    (default: models/7B/ggml-model-q4_0.gguf)
+  -p, --n-prompt <n>                        (default: 512)
+  -n, --n-gen <n>                           (default: 128)
+  -pg <pp,tg>                               (default: )
+  -d, --n-depth <n>                         (default: 0)
+  -b, --batch-size <n>                      (default: 2048)
+  -ub, --ubatch-size <n>                    (default: 512)
+  -ctk, --cache-type-k <t>                  (default: f16)
+  -ctv, --cache-type-v <t>                  (default: f16)
+  -t, --threads <n>                         (default: system dependent)
+  -C, --cpu-mask <hex,hex>                  (default: 0x0)
+  --cpu-strict <0|1>                        (default: 0)
+  --poll <0...100>                          (default: 50)
+  -ngl, --n-gpu-layers <n>                  (default: 99)
+  -ncmoe, --n-cpu-moe <n>                   (default: 0)
+  -sm, --split-mode <none|layer|row>        (default: layer)
+  -mg, --main-gpu <i>                       (default: 0)
+  -nkvo, --no-kv-offload <0|1>              (default: 0)
+  -fa, --flash-attn <0|1>                   (default: 0)
+  -dev, --device <dev0/dev1/...>            (default: auto)
+  -mmp, --mmap <0|1>                        (default: 1)
+  -embd, --embeddings <0|1>                 (default: 0)
+  -ts, --tensor-split <ts0/ts1/..>          (default: 0)
+  -ot --override-tensors <tensor name pattern>=<buffer type>;...
+                                            (default: disabled)
+  -nopo, --no-op-offload <0|1>              (default: 0)
+
+Multiple values can be given for each parameter by separating them with ','
+or by specifying the parameter multiple times. Ranges can be given as
+'first-last' or 'first-last+step' or 'first-last*mult'.
+```
+
+llama-bench can perform three types of tests:
+
+- Prompt processing (pp): processing a prompt in batches (`-p`)
+- Text generation (tg): generating a sequence of tokens (`-n`)
+- Prompt processing + text generation (pg): processing a prompt followed by generating a sequence of tokens (`-pg`)
+
+With the exception of `-r`, `-o` and `-v`, all options can be specified multiple times to run multiple tests. Each pp and tg test is run with all combinations of the specified options. To specify multiple values for an option, the values can be separated by commas (e.g. `-n 16,32`), or the option can be specified multiple times (e.g. `-n 16 -n 32`).
+
+Each test is repeated the number of times given by `-r`, and the results are averaged. The results are given in average tokens per second (t/s) and standard deviation. Some output formats (e.g. json) also include the individual results of each repetition.
+
+Using the `-d <n>` option, each test can be run at a specified context depth, prefilling the KV cache with `<n>` tokens.
+
+For a description of the other options, see the [completion example](../completion/README.md).
+
+> [!NOTE]
+> The measurements with `llama-bench` do not include the times for tokenization and for sampling.
+
+## Examples
+
+### Text generation with different models
+
+```sh
+$ ./llama-bench -m models/7B/ggml-model-q4_0.gguf -m models/13B/ggml-model-q4_0.gguf -p 0 -n 128,256,512
+```
+
+| model                          |       size |     params | backend    | ngl | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | tg 128     |    132.19 ± 0.55 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | tg 256     |    129.37 ± 0.54 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | tg 512     |    123.83 ± 0.25 |
+| llama 13B mostly Q4_0          |   6.86 GiB |    13.02 B | CUDA       |  99 | tg 128     |     82.17 ± 0.31 |
+| llama 13B mostly Q4_0          |   6.86 GiB |    13.02 B | CUDA       |  99 | tg 256     |     80.74 ± 0.23 |
+| llama 13B mostly Q4_0          |   6.86 GiB |    13.02 B | CUDA       |  99 | tg 512     |     78.08 ± 0.07 |
+
+### Prompt processing with different batch sizes
+
+```sh
+$ ./llama-bench -n 0 -p 1024 -b 128,256,512,1024
+```
+
+| model                          |       size |     params | backend    | ngl |    n_batch | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ---------: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 |        128 | pp 1024    |   1436.51 ± 3.66 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 |        256 | pp 1024    |  1932.43 ± 23.48 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 |        512 | pp 1024    |  2254.45 ± 15.59 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 |       1024 | pp 1024    |  2498.61 ± 13.58 |
+
+### Different numbers of threads
+
+```sh
+$ ./llama-bench -n 0 -n 16 -p 64 -t 1,2,4,8,16,32
+```
+
+| model                          |       size |     params | backend    |    threads | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | ---------: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          1 | pp 64      |      6.17 ± 0.07 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          1 | tg 16      |      4.05 ± 0.02 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          2 | pp 64      |     12.31 ± 0.13 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          2 | tg 16      |      7.80 ± 0.07 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          4 | pp 64      |     23.18 ± 0.06 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          4 | tg 16      |     12.22 ± 0.07 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          8 | pp 64      |     32.29 ± 1.21 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          8 | tg 16      |     16.71 ± 0.66 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |         16 | pp 64      |     33.52 ± 0.03 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |         16 | tg 16      |     15.32 ± 0.05 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |         32 | pp 64      |     59.00 ± 1.11 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |         32 | tg 16      |     16.41 ± 0.79 |
+
+### Different numbers of layers offloaded to the GPU
+
+```sh
+$ ./llama-bench -ngl 10,20,30,31,32,33,34,35
+```
+
+| model                          |       size |     params | backend    | ngl | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  10 | pp 512     |    373.36 ± 2.25 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  10 | tg 128     |     13.45 ± 0.93 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  20 | pp 512     |    472.65 ± 1.25 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  20 | tg 128     |     21.36 ± 1.94 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  30 | pp 512     |   631.87 ± 11.25 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  30 | tg 128     |     40.04 ± 1.82 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  31 | pp 512     |    657.89 ± 5.08 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  31 | tg 128     |     48.19 ± 0.81 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  32 | pp 512     |    688.26 ± 3.29 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  32 | tg 128     |     54.78 ± 0.65 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  33 | pp 512     |    704.27 ± 2.24 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  33 | tg 128     |     60.62 ± 1.76 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  34 | pp 512     |    881.34 ± 5.40 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  34 | tg 128     |     71.76 ± 0.23 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  35 | pp 512     |   2400.01 ± 7.72 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  35 | tg 128     |    131.66 ± 0.49 |
+
+### Different prefilled context
+
+```
+$ ./llama-bench -d 0,512
+```
+
+| model                          |       size |     params | backend    | ngl |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: |
+| qwen2 7B Q4_K - Medium         |   4.36 GiB |     7.62 B | CUDA       |  99 |           pp512 |      7340.20 ± 23.45 |
+| qwen2 7B Q4_K - Medium         |   4.36 GiB |     7.62 B | CUDA       |  99 |           tg128 |        120.60 ± 0.59 |
+| qwen2 7B Q4_K - Medium         |   4.36 GiB |     7.62 B | CUDA       |  99 |    pp512 @ d512 |      6425.91 ± 18.88 |
+| qwen2 7B Q4_K - Medium         |   4.36 GiB |     7.62 B | CUDA       |  99 |    tg128 @ d512 |        116.71 ± 0.60 |
+
+## Output formats
+
+By default, llama-bench outputs the results in markdown format. The results can be output in other formats by using the `-o` option.
+
+### Markdown
+
+```sh
+$ ./llama-bench -o md
+```
+
+| model                          |       size |     params | backend    | ngl | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | pp 512     |  2368.80 ± 93.24 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | tg 128     |    131.42 ± 0.59 |
+
+### CSV
+
+```sh
+$ ./llama-bench -o csv
+```
+
+```csv
+build_commit,build_number,cpu_info,gpu_info,backends,model_filename,model_type,model_size,model_n_params,n_batch,n_ubatch,n_threads,cpu_mask,cpu_strict,poll,type_k,type_v,n_gpu_layers,split_mode,main_gpu,no_kv_offload,flash_attn,tensor_split,use_mmap,embeddings,n_prompt,n_gen,n_depth,test_time,avg_ns,stddev_ns,avg_ts,stddev_ts
+"8cf427ff","5163","AMD Ryzen 7 7800X3D 8-Core Processor","NVIDIA GeForce RTX 4080","CUDA","models/Qwen2.5-7B-Instruct-Q4_K_M.gguf","qwen2 7B Q4_K - Medium","4677120000","7615616512","2048","512","8","0x0","0","50","f16","f16","99","layer","0","0","0","0.00","1","0","512","0","0","2025-04-24T11:57:09Z","70285660","982040","7285.676949","100.064434"
+"8cf427ff","5163","AMD Ryzen 7 7800X3D 8-Core Processor","NVIDIA GeForce RTX 4080","CUDA","models/Qwen2.5-7B-Instruct-Q4_K_M.gguf","qwen2 7B Q4_K - Medium","4677120000","7615616512","2048","512","8","0x0","0","50","f16","f16","99","layer","0","0","0","0.00","1","0","0","128","0","2025-04-24T11:57:10Z","1067431600","3834831","119.915244","0.430617"
+```
+
+### JSON
+
+```sh
+$ ./llama-bench -o json
+```
+
+```json
+[
+  {
+    "build_commit": "8cf427ff",
+    "build_number": 5163,
+    "cpu_info": "AMD Ryzen 7 7800X3D 8-Core Processor",
+    "gpu_info": "NVIDIA GeForce RTX 4080",
+    "backends": "CUDA",
+    "model_filename": "models/Qwen2.5-7B-Instruct-Q4_K_M.gguf",
+    "model_type": "qwen2 7B Q4_K - Medium",
+    "model_size": 4677120000,
+    "model_n_params": 7615616512,
+    "n_batch": 2048,
+    "n_ubatch": 512,
+    "n_threads": 8,
+    "cpu_mask": "0x0",
+    "cpu_strict": false,
+    "poll": 50,
+    "type_k": "f16",
+    "type_v": "f16",
+    "n_gpu_layers": 99,
+    "split_mode": "layer",
+    "main_gpu": 0,
+    "no_kv_offload": false,
+    "flash_attn": false,
+    "tensor_split": "0.00",
+    "use_mmap": true,
+    "embeddings": false,
+    "n_prompt": 512,
+    "n_gen": 0,
+    "n_depth": 0,
+    "test_time": "2025-04-24T11:58:50Z",
+    "avg_ns": 72135640,
+    "stddev_ns": 1453752,
+    "avg_ts": 7100.002165,
+    "stddev_ts": 140.341520,
+    "samples_ns": [ 74601900, 71632900, 71745200, 71952700, 70745500 ],
+    "samples_ts": [ 6863.1, 7147.55, 7136.37, 7115.79, 7237.21 ]
+  },
+  {
+    "build_commit": "8cf427ff",
+    "build_number": 5163,
+    "cpu_info": "AMD Ryzen 7 7800X3D 8-Core Processor",
+    "gpu_info": "NVIDIA GeForce RTX 4080",
+    "backends": "CUDA",
+    "model_filename": "models/Qwen2.5-7B-Instruct-Q4_K_M.gguf",
+    "model_type": "qwen2 7B Q4_K - Medium",
+    "model_size": 4677120000,
+    "model_n_params": 7615616512,
+    "n_batch": 2048,
+    "n_ubatch": 512,
+    "n_threads": 8,
+    "cpu_mask": "0x0",
+    "cpu_strict": false,
+    "poll": 50,
+    "type_k": "f16",
+    "type_v": "f16",
+    "n_gpu_layers": 99,
+    "split_mode": "layer",
+    "main_gpu": 0,
+    "no_kv_offload": false,
+    "flash_attn": false,
+    "tensor_split": "0.00",
+    "use_mmap": true,
+    "embeddings": false,
+    "n_prompt": 0,
+    "n_gen": 128,
+    "n_depth": 0,
+    "test_time": "2025-04-24T11:58:51Z",
+    "avg_ns": 1076767880,
+    "stddev_ns": 9449585,
+    "avg_ts": 118.881588,
+    "stddev_ts": 1.041811,
+    "samples_ns": [ 1075361300, 1065089400, 1071761200, 1081934900, 1089692600 ],
+    "samples_ts": [ 119.03, 120.178, 119.43, 118.307, 117.464 ]
+  }
+]
+```
+
+
+### JSONL
+
+```sh
+$ ./llama-bench -o jsonl
+```
+
+```json lines
+{"build_commit": "8cf427ff", "build_number": 5163, "cpu_info": "AMD Ryzen 7 7800X3D 8-Core Processor", "gpu_info": "NVIDIA GeForce RTX 4080", "backends": "CUDA", "model_filename": "models/Qwen2.5-7B-Instruct-Q4_K_M.gguf", "model_type": "qwen2 7B Q4_K - Medium", "model_size": 4677120000, "model_n_params": 7615616512, "n_batch": 2048, "n_ubatch": 512, "n_threads": 8, "cpu_mask": "0x0", "cpu_strict": false, "poll": 50, "type_k": "f16", "type_v": "f16", "n_gpu_layers": 99, "split_mode": "layer", "main_gpu": 0, "no_kv_offload": false, "flash_attn": false, "tensor_split": "0.00", "use_mmap": true, "embeddings": false, "n_prompt": 512, "n_gen": 0, "n_depth": 0, "test_time": "2025-04-24T11:59:33Z", "avg_ns": 70497220, "stddev_ns": 883196, "avg_ts": 7263.609157, "stddev_ts": 90.940578, "samples_ns": [ 71551000, 71222800, 70364100, 69439100, 69909100 ],"samples_ts": [ 7155.74, 7188.71, 7276.44, 7373.37, 7323.8 ]}
+{"build_commit": "8cf427ff", "build_number": 5163, "cpu_info": "AMD Ryzen 7 7800X3D 8-Core Processor", "gpu_info": "NVIDIA GeForce RTX 4080", "backends": "CUDA", "model_filename": "models/Qwen2.5-7B-Instruct-Q4_K_M.gguf", "model_type": "qwen2 7B Q4_K - Medium", "model_size": 4677120000, "model_n_params": 7615616512, "n_batch": 2048, "n_ubatch": 512, "n_threads": 8, "cpu_mask": "0x0", "cpu_strict": false, "poll": 50, "type_k": "f16", "type_v": "f16", "n_gpu_layers": 99, "split_mode": "layer", "main_gpu": 0, "no_kv_offload": false, "flash_attn": false, "tensor_split": "0.00", "use_mmap": true, "embeddings": false, "n_prompt": 0, "n_gen": 128, "n_depth": 0, "test_time": "2025-04-24T11:59:33Z", "avg_ns": 1068078400, "stddev_ns": 6279455, "avg_ts": 119.844681, "stddev_ts": 0.699739, "samples_ns": [ 1066331700, 1064864900, 1079042600, 1063328400, 1066824400 ],"samples_ts": [ 120.038, 120.203, 118.624, 120.377, 119.982 ]}
+```
+
+
+### SQL
+
+SQL output is suitable for importing into a SQLite database. The output can be piped into the `sqlite3` command line tool to add the results to a database.
+
+```sh
+$ ./llama-bench -o sql
+```
+
+```sql
+CREATE TABLE IF NOT EXISTS test (
+  build_commit TEXT,
+  build_number INTEGER,
+  cpu_info TEXT,
+  gpu_info TEXT,
+  backends TEXT,
+  model_filename TEXT,
+  model_type TEXT,
+  model_size INTEGER,
+  model_n_params INTEGER,
+  n_batch INTEGER,
+  n_ubatch INTEGER,
+  n_threads INTEGER,
+  cpu_mask TEXT,
+  cpu_strict INTEGER,
+  poll INTEGER,
+  type_k TEXT,
+  type_v TEXT,
+  n_gpu_layers INTEGER,
+  split_mode TEXT,
+  main_gpu INTEGER,
+  no_kv_offload INTEGER,
+  flash_attn INTEGER,
+  tensor_split TEXT,
+  use_mmap INTEGER,
+  embeddings INTEGER,
+  n_prompt INTEGER,
+  n_gen INTEGER,
+  n_depth INTEGER,
+  test_time TEXT,
+  avg_ns INTEGER,
+  stddev_ns INTEGER,
+  avg_ts REAL,
+  stddev_ts REAL
+);
+
+INSERT INTO test (build_commit, build_number, cpu_info, gpu_info, backends, model_filename, model_type, model_size, model_n_params, n_batch, n_ubatch, n_threads, cpu_mask, cpu_strict, poll, type_k, type_v, n_gpu_layers, split_mode, main_gpu, no_kv_offload, flash_attn, tensor_split, use_mmap, embeddings, n_prompt, n_gen, n_depth, test_time, avg_ns, stddev_ns, avg_ts, stddev_ts) VALUES ('8cf427ff', '5163', 'AMD Ryzen 7 7800X3D 8-Core Processor', 'NVIDIA GeForce RTX 4080', 'CUDA', 'models/Qwen2.5-7B-Instruct-Q4_K_M.gguf', 'qwen2 7B Q4_K - Medium', '4677120000', '7615616512', '2048', '512', '8', '0x0', '0', '50', 'f16', 'f16', '99', 'layer', '0', '0', '0', '0.00', '1', '0', '512', '0', '0', '2025-04-24T12:00:08Z', '69905000', '519516', '7324.546977', '54.032613');
+INSERT INTO test (build_commit, build_number, cpu_info, gpu_info, backends, model_filename, model_type, model_size, model_n_params, n_batch, n_ubatch, n_threads, cpu_mask, cpu_strict, poll, type_k, type_v, n_gpu_layers, split_mode, main_gpu, no_kv_offload, flash_attn, tensor_split, use_mmap, embeddings, n_prompt, n_gen, n_depth, test_time, avg_ns, stddev_ns, avg_ts, stddev_ts) VALUES ('8cf427ff', '5163', 'AMD Ryzen 7 7800X3D 8-Core Processor', 'NVIDIA GeForce RTX 4080', 'CUDA', 'models/Qwen2.5-7B-Instruct-Q4_K_M.gguf', 'qwen2 7B Q4_K - Medium', '4677120000', '7615616512', '2048', '512', '8', '0x0', '0', '50', 'f16', 'f16', '99', 'layer', '0', '0', '0', '0.00', '1', '0', '0', '128', '0', '2025-04-24T12:00:09Z', '1063608780', '4464130', '120.346696', '0.504647');
+```

tools/mtmd/CMakeLists.txt

@@ -0,0 +1,95 @@
+# mtmd
+
+find_package(Threads REQUIRED)
+
+add_library(mtmd
+            mtmd.cpp
+            mtmd-audio.cpp
+            mtmd.h
+            mtmd-helper.cpp
+            mtmd-helper.h
+            clip.cpp
+            clip.h
+            clip-impl.h
+            clip-model.h
+            clip-graph.h
+            models/models.h
+            models/cogvlm.cpp
+            models/conformer.cpp
+            models/glm4v.cpp
+            models/internvl.cpp
+            models/kimivl.cpp
+            models/llama4.cpp
+            models/llava.cpp
+            models/minicpmv.cpp
+            models/pixtral.cpp
+            models/qwen2vl.cpp
+            models/qwen3vl.cpp
+            models/siglip.cpp
+            models/whisper-enc.cpp
+            models/mobilenetv5.cpp
+            models/youtuvl.cpp
+            )
+
+set_target_properties(mtmd PROPERTIES
+    VERSION ${LLAMA_INSTALL_VERSION}
+    SOVERSION 0
+    MACHO_CURRENT_VERSION 0 # keep macOS linker from seeing oversized version number
+)
+
+target_link_libraries     (mtmd PUBLIC ggml llama)
+target_link_libraries     (mtmd PRIVATE Threads::Threads)
+target_include_directories(mtmd PUBLIC  .)
+target_include_directories(mtmd PRIVATE ../..)
+target_include_directories(mtmd PRIVATE ../../vendor)
+target_compile_features   (mtmd PRIVATE cxx_std_17)
+
+if (BUILD_SHARED_LIBS)
+    set_target_properties     (mtmd PROPERTIES POSITION_INDEPENDENT_CODE ON)
+    target_compile_definitions(mtmd PRIVATE LLAMA_BUILD)
+    target_compile_definitions(mtmd PUBLIC  LLAMA_SHARED)
+endif()
+
+set(MTMD_PUBLIC_HEADERS
+    ${CMAKE_CURRENT_SOURCE_DIR}/mtmd.h
+    ${CMAKE_CURRENT_SOURCE_DIR}/mtmd-helper.h
+    )
+
+set_target_properties(mtmd
+    PROPERTIES
+    PUBLIC_HEADER "${MTMD_PUBLIC_HEADERS}")
+
+install(TARGETS mtmd LIBRARY PUBLIC_HEADER)
+
+if (NOT MSVC)
+    # for stb_image.h and miniaudio.h
+    target_compile_options(mtmd PRIVATE -Wno-cast-qual)
+endif()
+
+if (TARGET BUILD_INFO)
+    add_dependencies(mtmd        BUILD_INFO)
+    add_dependencies(mtmd-helper BUILD_INFO)
+endif()
+
+# if mtmd is linked against common, we throw an error
+if (TARGET mtmd)
+    get_target_property(libs mtmd LINK_LIBRARIES)
+    if (libs AND "common" IN_LIST libs)
+        message(FATAL_ERROR "mtmd is designed to be a public library.\n"
+                            "It must not link against common")
+    endif()
+endif()
+
+add_executable(llama-llava-cli    deprecation-warning.cpp)
+add_executable(llama-gemma3-cli   deprecation-warning.cpp)
+add_executable(llama-minicpmv-cli deprecation-warning.cpp)
+add_executable(llama-qwen2vl-cli  deprecation-warning.cpp)
+
+set(TARGET llama-mtmd-cli)
+add_executable         (${TARGET} mtmd-cli.cpp)
+set_target_properties  (${TARGET} PROPERTIES OUTPUT_NAME llama-mtmd-cli)
+if(LLAMA_TOOLS_INSTALL)
+    install(TARGETS ${TARGET} RUNTIME)
+endif()
+target_link_libraries  (${TARGET} PRIVATE common mtmd Threads::Threads)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

tools/mtmd/README.md

@@ -0,0 +1,63 @@
+# Multimodal Support in llama.cpp
+
+This directory provides multimodal capabilities for `llama.cpp`. Initially intended as a showcase for running LLaVA models, its scope has expanded significantly over time to include various other vision-capable models. As a result, LLaVA is no longer the only multimodal architecture supported.
+
+> [!IMPORTANT]
+>
+> Multimodal support can be viewed as a sub-project within `llama.cpp`. It is under **very heavy development**, and **breaking changes are expected**.
+
+The naming and structure related to multimodal support have evolved, which might cause some confusion. Here's a brief timeline to clarify:
+
+- [#3436](https://github.com/ggml-org/llama.cpp/pull/3436): Initial support for LLaVA 1.5 was added, introducing `llava.cpp` and `clip.cpp`. The `llava-cli` binary was created for model interaction.
+- [#4954](https://github.com/ggml-org/llama.cpp/pull/4954): Support for MobileVLM was added, becoming the second vision model supported. This built upon the existing `llava.cpp`, `clip.cpp`, and `llava-cli` infrastructure.
+- **Expansion & Fragmentation:** Many new models were subsequently added (e.g., [#7599](https://github.com/ggml-org/llama.cpp/pull/7599), [#10361](https://github.com/ggml-org/llama.cpp/pull/10361), [#12344](https://github.com/ggml-org/llama.cpp/pull/12344), and others). However, `llava-cli` lacked support for the increasingly complex chat templates required by these models. This led to the creation of model-specific binaries like `qwen2vl-cli`, `minicpmv-cli`, and `gemma3-cli`. While functional, this proliferation of command-line tools became confusing for users.
+- [#12849](https://github.com/ggml-org/llama.cpp/pull/12849): `libmtmd` was introduced as a replacement for `llava.cpp`. Its goals include providing a single, unified command-line interface, improving the user/developer experience (UX/DX), and supporting both audio and image inputs.
+- [#13012](https://github.com/ggml-org/llama.cpp/pull/13012): `mtmd-cli` was added, consolidating the various model-specific CLIs into a single tool powered by `libmtmd`.
+
+## Pre-quantized models
+
+See the list of pre-quantized model [here](../../docs/multimodal.md)
+
+## How it works and what is `mmproj`?
+
+Multimodal support in `llama.cpp` works by encoding images into embeddings using a separate model component, and then feeding these embeddings into the language model.
+
+This approach keeps the multimodal components distinct from the core `libllama` library. Separating these allows for faster, independent development cycles. While many modern vision models are based on Vision Transformers (ViTs), their specific pre-processing and projection steps can vary significantly. Integrating this diverse complexity directly into `libllama` is currently challenging.
+
+Consequently, running a multimodal model typically requires two GGUF files:
+1.  The standard language model file.
+2.  A corresponding **multimodal projector (`mmproj`)** file, which handles the image encoding and projection.
+
+## What is `libmtmd`?
+
+As outlined in the history, `libmtmd` is the modern library designed to replace the original `llava.cpp` implementation for handling multimodal inputs.
+
+Built upon `clip.cpp` (similar to `llava.cpp`), `libmtmd` offers several advantages:
+- **Unified Interface:** Aims to consolidate interaction for various multimodal models.
+- **Improved UX/DX:** Features a more intuitive API, inspired by the `Processor` class in the Hugging Face `transformers` library.
+- **Flexibility:** Designed to support multiple input types (text, audio, images) while respecting the wide variety of chat templates used by different models.
+
+## How to obtain `mmproj`
+
+Multimodal projector (`mmproj`) files are specific to each model architecture.
+
+For the following models, you can use `convert_hf_to_gguf.py` with `--mmproj` flag to get the `mmproj` file:
+- [Gemma 3](https://huggingface.co/collections/google/gemma-3-release-67c6c6f89c4f76621268bb6d) ; See the guide [here](../../docs/multimodal/gemma3.md) - Note: 1B variant does not have vision support
+- SmolVLM (from [HuggingFaceTB](https://huggingface.co/HuggingFaceTB))
+- SmolVLM2 (from [HuggingFaceTB](https://huggingface.co/HuggingFaceTB))
+- [Pixtral 12B](https://huggingface.co/mistral-community/pixtral-12b) - only works with `transformers`-compatible checkpoint
+- Qwen 2 VL and Qwen 2.5 VL (from [Qwen](https://huggingface.co/Qwen))
+- [Mistral Small 3.1 24B](https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503)
+- InternVL 2.5 and InternVL 3 from [OpenGVLab](https://huggingface.co/OpenGVLab) (note: we don't support conversion of `InternVL3-*-hf` model, only non-HF version is supported ; `InternLM2Model` **text** model is not supported)
+
+For older models, please refer to the relevant guide for instructions on how to obtain or create them:
+
+NOTE: conversion scripts are located under `tools/mtmd/legacy-models`
+
+- [LLaVA](../../docs/multimodal/llava.md)
+- [MobileVLM](../../docs/multimodal/MobileVLM.md)
+- [GLM-Edge](../../docs/multimodal/glmedge.md)
+- [MiniCPM-V 2.5](../../docs/multimodal/minicpmv2.5.md)
+- [MiniCPM-V 2.6](../../docs/multimodal/minicpmv2.6.md)
+- [MiniCPM-o 2.6](../../docs/multimodal/minicpmo2.6.md)
+- [IBM Granite Vision](../../docs/multimodal/granitevision.md)

tools/mtmd/clip-graph.h

@@ -0,0 +1,117 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-cpp.h"
+#include "clip.h"
+#include "clip-impl.h"
+#include "clip-model.h"
+
+#include <vector>
+#include <functional>
+
+#define DEFAULT_INTERPOLATION_MODE (GGML_SCALE_MODE_BILINEAR | GGML_SCALE_FLAG_ANTIALIAS)
+
+struct clip_graph {
+    const clip_model & model;
+    const clip_hparams & hparams;
+    projector_type proj_type;
+
+    // we only support single image per batch
+    const clip_image_f32 & img;
+
+    const int patch_size;
+    const int n_patches_x;
+    const int n_patches_y;
+    const int n_patches;
+    const int n_embd;
+    const int n_head;
+    const int d_head;
+    const int n_layer;
+    const int n_mmproj_embd;
+    const float eps;
+    const float kq_scale;
+    const clip_flash_attn_type flash_attn_type;
+
+    ggml_context_ptr ctx0_ptr;
+    ggml_context * ctx0;
+    ggml_cgraph * gf;
+
+    clip_graph(clip_ctx * ctx, const clip_image_f32 & img);
+
+    virtual ~clip_graph() = default;
+    virtual ggml_cgraph * build() = 0;
+
+    //
+    // utility functions
+    //
+    void cb(ggml_tensor * cur0, const char * name, int il) const;
+
+    // siglip2 naflex
+    ggml_tensor * resize_position_embeddings(uint32_t interpolation_mode = DEFAULT_INTERPOLATION_MODE);
+
+    // build vision transformer (ViT) cgraph
+    // this function should cover most of the models
+    // if your model has specific features, you should probably duplicate this function
+    ggml_tensor * build_vit(
+                ggml_tensor * inp,
+                int64_t n_pos,
+                norm_type norm_t,
+                ffn_op_type ffn_t,
+                ggml_tensor * learned_pos_embd,
+                std::function<ggml_tensor *(ggml_tensor *, const clip_layer &)> add_pos);
+
+    // build the input after conv2d (inp_raw --> patches)
+    // returns tensor with shape [n_embd, n_patches]
+    ggml_tensor * build_inp();
+
+    ggml_tensor * build_inp_raw(int channels = 3);
+
+    ggml_tensor * build_norm(
+            ggml_tensor * cur,
+            ggml_tensor * mw,
+            ggml_tensor * mb,
+            norm_type type,
+            float norm_eps,
+            int il) const;
+
+    ggml_tensor * build_ffn(
+            ggml_tensor * cur,
+            ggml_tensor * up,
+            ggml_tensor * up_b,
+            ggml_tensor * gate,
+            ggml_tensor * gate_b,
+            ggml_tensor * down,
+            ggml_tensor * down_b,
+            ffn_op_type type_op,
+            int il) const;
+
+    ggml_tensor * build_attn(
+            ggml_tensor * wo,
+            ggml_tensor * wo_b,
+            ggml_tensor * q_cur,
+            ggml_tensor * k_cur,
+            ggml_tensor * v_cur,
+            ggml_tensor * kq_mask,
+            float kq_scale,
+            int il) const;
+
+    // implementation of the 2D RoPE without adding a new op in ggml
+    // this is not efficient (use double the memory), but works on all backends
+    // TODO: there was a more efficient which relies on ggml_view and ggml_rope_ext_inplace, but the rope inplace does not work well with non-contiguous tensors ; we should fix that and revert back to the original implementation in https://github.com/ggml-org/llama.cpp/pull/13065
+    ggml_tensor * build_rope_2d(
+        ggml_context * ctx0,
+        ggml_tensor * cur,
+        ggml_tensor * pos_a, // first half
+        ggml_tensor * pos_b, // second half
+        const float freq_base,
+        const bool interleave_freq
+    );
+
+    // aka pixel_shuffle / pixel_unshuffle / patch_merger (Kimi-VL)
+    // support dynamic resolution
+    ggml_tensor * build_patch_merge_permute(ggml_tensor * cur, int scale_factor);
+
+    // Generic function to stack frames for audio processing
+    // Abstracts out the StackAudioFrames logic used by ultravox
+    ggml_tensor * build_stack(ggml_tensor * cur, int32_t stack_factor, int32_t n_embed);
+};

tools/mtmd/clip-impl.h

@@ -0,0 +1,578 @@
+#pragma once
+
+#include "ggml.h"
+#include "gguf.h"
+#include "clip.h"
+
+#include <climits>
+#include <cstdarg>
+#include <cinttypes>
+#include <string>
+#include <map>
+#include <sstream>
+#include <vector>
+#include <memory>
+
+// Internal header for clip.cpp
+
+#define MTMD_INTERNAL_HEADER
+
+#define KEY_FTYPE               "general.file_type"
+#define KEY_NAME                "general.name"
+#define KEY_DESCRIPTION         "general.description"
+#define KEY_PROJ_TYPE           "clip.projector_type"
+#define KEY_HAS_AUDIO_ENC       "clip.has_audio_encoder"
+#define KEY_HAS_VISION_ENC      "clip.has_vision_encoder"
+#define KEY_USE_GELU            "clip.use_gelu"
+#define KEY_USE_SILU            "clip.use_silu"
+
+#define KEY_N_EMBD              "clip.%s.embedding_length"
+#define KEY_N_FF                "clip.%s.feed_forward_length"
+#define KEY_N_BLOCK             "clip.%s.block_count"
+#define KEY_PROJ_DIM            "clip.%s.projection_dim"
+#define KEY_N_HEAD              "clip.%s.attention.head_count"
+#define KEY_LAYER_NORM_EPS      "clip.%s.attention.layer_norm_epsilon"
+
+// vision-specific
+#define KEY_VISION_PROJ_TYPE    "clip.vision.projector_type" // for models with mixed modalities
+#define KEY_IMAGE_SIZE          "clip.vision.image_size"
+#define KEY_PREPROC_IMAGE_SIZE  "clip.vision.preproc_image_size"
+#define KEY_PATCH_SIZE          "clip.vision.patch_size"
+#define KEY_IMAGE_MEAN          "clip.vision.image_mean"
+#define KEY_IMAGE_STD           "clip.vision.image_std"
+#define KEY_FEATURE_LAYER       "clip.vision.feature_layer"
+#define KEY_PROJ_SCALE_FACTOR   "clip.vision.projector.scale_factor"
+#define KEY_SPATIAL_MERGE_SIZE  "clip.vision.spatial_merge_size"
+#define KEY_IS_DEEPSTACK_LAYERS "clip.vision.is_deepstack_layers"
+
+#define KEY_MM_PATCH_MERGE_TYPE    "clip.vision.mm_patch_merge_type"
+#define KEY_IMAGE_GRID_PINPOINTS   "clip.vision.image_grid_pinpoints"
+#define KEY_IMAGE_CROP_RESOLUTION  "clip.vision.image_crop_resolution"
+#define KEY_WIN_ATTN_PATTERN       "clip.vision.n_wa_pattern"
+#define KEY_WIN_ATTN_LAYER_INDEXES "clip.vision.wa_layer_indexes"
+#define KEY_ATTN_WINDOW_SIZE       "clip.vision.window_size"
+#define KEY_MINICPMV_VERSION       "clip.minicpmv_version"
+#define KEY_MINICPMV_QUERY_NUM     "clip.minicpmv_query_num"
+
+// audio-specific
+#define KEY_AUDIO_PROJ_TYPE     "clip.audio.projector_type" // for models with mixed modalities
+#define KEY_A_NUM_MEL_BINS      "clip.audio.num_mel_bins"
+#define KEY_A_PROJ_STACK_FACTOR "clip.audio.projector.stack_factor"
+
+
+//
+// tensor name constants
+//
+
+#define TN_POS_EMBD        "%s.position_embd.weight"
+#define TN_CLASS_EMBD      "v.class_embd"
+#define TN_PATCH_EMBD      "v.patch_embd.weight"  // not rename tensor with ".0" postfix for backwrad compat
+#define TN_PATCH_EMBD_1    "v.patch_embd.weight.1"
+#define TN_PATCH_BIAS      "v.patch_embd.bias"
+#define TN_NORM_EMBD       "v.norm_embd.%s"
+#define TN_ATTN_QKV        "%s.blk.%d.attn_qkv.%s"
+#define TN_ATTN_K          "%s.blk.%d.attn_k.%s"
+#define TN_ATTN_Q          "%s.blk.%d.attn_q.%s"
+#define TN_ATTN_V          "%s.blk.%d.attn_v.%s"
+#define TN_ATTN_OUTPUT     "%s.blk.%d.attn_out.%s"
+#define TN_ATTN_K_NORM     "%s.blk.%d.attn_k_norm.%s"
+#define TN_ATTN_Q_NORM     "%s.blk.%d.attn_q_norm.%s"
+#define TN_FFN_DOWN        "%s.blk.%d.ffn_down.%s"
+#define TN_FFN_GATE        "%s.blk.%d.ffn_gate.%s"
+#define TN_FFN_UP          "%s.blk.%d.ffn_up.%s"
+#define TN_FFN_GATE        "%s.blk.%d.ffn_gate.%s"
+#define TN_LN_1            "%s.blk.%d.ln1.%s" // layer norm
+#define TN_LN_2            "%s.blk.%d.ln2.%s" // layer norm
+#define TN_LS_1            "%s.blk.%d.ls1.%s" // layer scale
+#define TN_LS_2            "%s.blk.%d.ls2.%s" // layer scale
+#define TN_LN_PRE          "%s.pre_ln.%s"
+#define TN_LN_POST         "%s.post_ln.%s"
+#define TN_LLAVA_PROJ      "mm.%d.%s"
+#define TN_MM_UP           "mm.up.%s"
+#define TN_MM_GATE         "mm.gate.%s"
+#define TN_MM_DOWN         "mm.down.%s"
+#define TN_MM_POST_NORM    "mm.post_norm.%s"
+#define TN_MVLM_PROJ_MLP   "mm.model.mlp.%d.%s"
+#define TN_MVLM_PROJ_BLOCK "mm.model.mb_block.%d.block.%d.%s"
+#define TN_MVLM_PROJ_PEG   "mm.model.peg.%d.%s"
+#define TN_IMAGE_NEWLINE   "model.image_newline"
+#define TN_MM_INP_NORM     "mm.input_norm.weight"
+#define TN_MM_INP_NORM_B   "mm.input_norm.bias"
+#define TN_MM_INP_PROJ     "mm.input_projection.weight" // gemma3
+#define TN_MM_SOFT_EMB_N   "mm.soft_emb_norm.weight"    // gemma3
+#define TN_MM_PROJECTOR    "mm.model.fc.weight"         // idefics3
+#define TN_MM_PATCH_MERGER "mm.patch_merger.%s"         // mistral small 3.1, glm4v
+#define TN_TOK_IMG_BREAK   "v.token_embd.img_break"     // pixtral
+#define TN_TOK_GLM_BOI     "adapter.boi"                // glm-edge (these embeddings are not in text model)
+#define TN_TOK_GLM_EOI     "adapter.eoi"                // glm-edge (these embeddings are not in text model)
+#define TN_DEEPSTACK_NORM  "v.deepstack.%d.norm.%s"     // qwen3vl deepstack
+#define TN_DEEPSTACK_FC1   "v.deepstack.%d.fc1.%s"      // qwen3vl deepstack
+#define TN_DEEPSTACK_FC2   "v.deepstack.%d.fc2.%s"      // qwen3vl deepstack
+
+// mimicpmv
+#define TN_MINICPMV_POS_EMBD_K "resampler.pos_embed_k"
+#define TN_MINICPMV_QUERY      "resampler.query"
+#define TN_MINICPMV_PROJ       "resampler.proj.weight"
+#define TN_MINICPMV_KV_PROJ    "resampler.kv.weight"
+#define TN_MINICPMV_ATTN       "resampler.attn.%s.%s"
+#define TN_MINICPMV_LN         "resampler.ln_%s.%s"
+
+#define TN_GLM_ADAPER_CONV      "adapter.conv.%s"
+#define TN_GLM_ADAPTER_LINEAR   "adapter.linear.linear.%s"
+#define TN_GLM_ADAPTER_NORM_1   "adapter.linear.norm1.%s"
+#define TN_GLM_ADAPTER_D_H_2_4H "adapter.linear.dense_h_to_4h.%s"
+#define TN_GLM_ADAPTER_GATE     "adapter.linear.gate.%s"
+#define TN_GLM_ADAPTER_D_4H_2_H "adapter.linear.dense_4h_to_h.%s"
+
+// ultravox
+#define TN_CONV1D       "a.conv1d.%d.%s"
+#define TN_MM_AUDIO_MLP "mm.a.mlp.%d.%s"
+#define TN_MM_AUDIO_FC  "mm.a.fc.%s" // fully connected layer
+#define TN_MM_NORM_PRE  "mm.a.norm_pre.%s"
+#define TN_MM_NORM_MID  "mm.a.norm_mid.%s"
+
+// cogvlm
+#define TN_MM_POST_FC_NORM "mm.post_fc_norm.%s"
+#define TN_MM_H_TO_4H      "mm.up.%s"
+#define TN_MM_GATE         "mm.gate.%s"
+#define TN_MM_4H_TO_H      "mm.down.%s"
+#define TN_TOK_BOI         "v.boi"
+#define TN_TOK_EOI         "v.eoi"
+
+// (conformer) lfm2
+#define TN_PRE_ENCODE_OUT  "a.pre_encode.out.%s"
+#define TN_FFN_NORM        "%s.blk.%d.ffn_norm.%s"
+#define TN_FFN_NORM_1      "%s.blk.%d.ffn_norm_1.%s"
+#define TN_FFN_UP_1        "%s.blk.%d.ffn_up_1.%s"
+#define TN_FFN_DOWN_1      "%s.blk.%d.ffn_down_1.%s"
+#define TN_POS_BIAS_U      "%s.blk.%d.pos_bias_u"
+#define TN_POS_BIAS_V      "%s.blk.%d.pos_bias_v"
+#define TN_NORM_CONV       "%s.blk.%d.norm_conv.%s"
+#define TN_LINEAR_POS      "%s.blk.%d.linear_pos.%s"
+#define TN_CONV_DW         "%s.blk.%d.conv_dw.%s"
+#define TN_CONV_NORM       "%s.blk.%d.conv_norm.%s"
+#define TN_CONV_PW1        "%s.blk.%d.conv_pw1.%s"
+#define TN_CONV_PW2        "%s.blk.%d.conv_pw2.%s"
+
+// mobilenetv5 (gemma3n) definitions
+#define TN_MNV5_STEM_CONV        "v.conv_stem.conv.weight"
+#define TN_MNV5_STEM_BIAS        "v.conv_stem.conv.bias"
+#define TN_MNV5_STEM_BN          "v.conv_stem.bn.weight"
+
+// Stage 0 Block (Edge Residual)
+#define TN_MNV5_BLK_S0_EXP_W     "v.blk.%d.%d.conv_exp.weight"
+#define TN_MNV5_BLK_S0_BN1_W     "v.blk.%d.%d.bn1.weight"
+#define TN_MNV5_BLK_S0_PWL_W     "v.blk.%d.%d.conv_pwl.weight"
+#define TN_MNV5_BLK_S0_BN2_W     "v.blk.%d.%d.bn2.weight"
+
+// Stage 1+ Block (Universal Inverted Residual)
+#define TN_MNV5_BLK_DW_START_W   "v.blk.%d.%d.dw_start.conv.weight"
+#define TN_MNV5_BLK_DW_START_BN  "v.blk.%d.%d.dw_start.bn.weight"
+#define TN_MNV5_BLK_DW_MID_W     "v.blk.%d.%d.dw_mid.conv.weight"
+#define TN_MNV5_BLK_DW_MID_BN    "v.blk.%d.%d.dw_mid.bn.weight"
+#define TN_MNV5_BLK_PW_EXP_W     "v.blk.%d.%d.pw_exp.conv.weight"
+#define TN_MNV5_BLK_PW_EXP_BN    "v.blk.%d.%d.pw_exp.bn.weight"
+#define TN_MNV5_BLK_PW_PROJ_W    "v.blk.%d.%d.pw_proj.conv.weight"
+#define TN_MNV5_BLK_PW_PROJ_BN   "v.blk.%d.%d.pw_proj.bn.weight"
+#define TN_MNV5_BLK_LAYER_SCALE  "v.blk.%d.%d.layer_scale.gamma"
+
+// Attention Components
+#define TN_MNV5_ATTN_Q_W         "v.blk.%d.%d.attn.query.proj.weight"
+#define TN_MNV5_ATTN_K_W         "v.blk.%d.%d.attn.key.proj.weight"
+#define TN_MNV5_ATTN_V_W         "v.blk.%d.%d.attn.value.proj.weight"
+#define TN_MNV5_ATTN_O_W         "v.blk.%d.%d.attn.output.proj.weight"
+#define TN_MNV5_ATTN_K_DW        "v.blk.%d.%d.attn.key.down_conv.weight"
+#define TN_MNV5_ATTN_K_NORM      "v.blk.%d.%d.attn.key.norm.weight"
+#define TN_MNV5_ATTN_V_DW        "v.blk.%d.%d.attn.value.down_conv.weight"
+#define TN_MNV5_ATTN_V_NORM      "v.blk.%d.%d.attn.value.norm.weight"
+#define TN_MNV5_ATTN_NORM        "v.blk.%d.%d.norm.weight" // Block norm used in attn blocks
+
+// MSFA
+#define TN_MNV5_MSFA_FFN_EXP_W   "v.msfa.ffn.pw_exp.conv.weight"
+#define TN_MNV5_MSFA_FFN_EXP_BN  "v.msfa.ffn.pw_exp.bn.weight"
+#define TN_MNV5_MSFA_FFN_PROJ_W  "v.msfa.ffn.pw_proj.conv.weight"
+#define TN_MNV5_MSFA_FFN_PROJ_BN "v.msfa.ffn.pw_proj.bn.weight"
+#define TN_MNV5_MSFA_NORM        "v.msfa.norm.weight"
+
+
+// align x to upper multiple of n
+#define CLIP_ALIGN(x, n) ((((x) + (n) - 1) / (n)) * (n))
+
+// forward declaration
+// TODO: improve this later
+struct clip_ctx;
+
+enum projector_type {
+    PROJECTOR_TYPE_MLP,
+    PROJECTOR_TYPE_MLP_NORM,
+    PROJECTOR_TYPE_LDP,
+    PROJECTOR_TYPE_LDPV2,
+    PROJECTOR_TYPE_MINICPMV,
+    PROJECTOR_TYPE_GLM_EDGE,
+    PROJECTOR_TYPE_QWEN2VL,
+    PROJECTOR_TYPE_QWEN3VL,
+    PROJECTOR_TYPE_GEMMA3,
+    PROJECTOR_TYPE_GEMMA3NV,
+    PROJECTOR_TYPE_GEMMA3NA,
+    PROJECTOR_TYPE_IDEFICS3,
+    PROJECTOR_TYPE_PIXTRAL,
+    PROJECTOR_TYPE_QWEN25VL,
+    PROJECTOR_TYPE_ULTRAVOX,
+    PROJECTOR_TYPE_INTERNVL,
+    PROJECTOR_TYPE_LLAMA4,
+    PROJECTOR_TYPE_QWEN2A,
+    PROJECTOR_TYPE_GLMA,
+    PROJECTOR_TYPE_QWEN25O, // will be replaced by QWEN2A or QWEN25VL depending on clip_ctx
+    PROJECTOR_TYPE_VOXTRAL,
+    PROJECTOR_TYPE_MUSIC_FLAMINGO,
+    PROJECTOR_TYPE_LFM2,
+    PROJECTOR_TYPE_KIMIVL,
+    PROJECTOR_TYPE_LIGHTONOCR,
+    PROJECTOR_TYPE_COGVLM,
+    PROJECTOR_TYPE_JANUS_PRO,
+    PROJECTOR_TYPE_LFM2A,
+    PROJECTOR_TYPE_GLM4V,
+    PROJECTOR_TYPE_YOUTUVL,
+    PROJECTOR_TYPE_UNKNOWN,
+};
+
+static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
+    { PROJECTOR_TYPE_MLP,       "mlp" },
+    { PROJECTOR_TYPE_LDP,       "ldp" },
+    { PROJECTOR_TYPE_LDPV2,     "ldpv2"},
+    { PROJECTOR_TYPE_MINICPMV,  "resampler"},
+    { PROJECTOR_TYPE_GLM_EDGE,  "adapter"},
+    { PROJECTOR_TYPE_QWEN2VL,   "qwen2vl_merger"},
+    { PROJECTOR_TYPE_QWEN25VL,  "qwen2.5vl_merger"},
+    { PROJECTOR_TYPE_QWEN3VL,   "qwen3vl_merger"},
+    { PROJECTOR_TYPE_GEMMA3,    "gemma3"},
+    { PROJECTOR_TYPE_GEMMA3NV,  "gemma3nv"},
+    { PROJECTOR_TYPE_GEMMA3NA,  "gemma3na"},
+    { PROJECTOR_TYPE_IDEFICS3,  "idefics3"},
+    { PROJECTOR_TYPE_PIXTRAL,   "pixtral"},
+    { PROJECTOR_TYPE_ULTRAVOX,  "ultravox"},
+    { PROJECTOR_TYPE_INTERNVL,  "internvl"},
+    { PROJECTOR_TYPE_LLAMA4,    "llama4"},
+    { PROJECTOR_TYPE_QWEN2A,    "qwen2a"},
+    { PROJECTOR_TYPE_GLMA,      "glma"},
+    { PROJECTOR_TYPE_QWEN25O,   "qwen2.5o"},
+    { PROJECTOR_TYPE_VOXTRAL,   "voxtral"},
+    { PROJECTOR_TYPE_MUSIC_FLAMINGO, "musicflamingo"},
+    { PROJECTOR_TYPE_LFM2,      "lfm2"},
+    { PROJECTOR_TYPE_KIMIVL,    "kimivl"},
+    { PROJECTOR_TYPE_LIGHTONOCR,"lightonocr"},
+    { PROJECTOR_TYPE_COGVLM,    "cogvlm"},
+    { PROJECTOR_TYPE_JANUS_PRO, "janus_pro"},
+    { PROJECTOR_TYPE_LFM2A,     "lfm2a"},
+    { PROJECTOR_TYPE_GLM4V,     "glm4v"},
+    { PROJECTOR_TYPE_YOUTUVL,   "youtuvl"},
+};
+
+static projector_type clip_projector_type_from_string(const std::string & str) {
+    for (const auto & pair : PROJECTOR_TYPE_NAMES) {
+        if (pair.second == str) {
+            return pair.first;
+        }
+    }
+    return PROJECTOR_TYPE_UNKNOWN;
+}
+
+// RGB uint8 image
+struct clip_image_u8 {
+    int nx;
+    int ny;
+
+    std::vector<uint8_t> buf;
+};
+
+// For images, buf.size() == nx*ny*3
+//     Memory layout: RGBRGBRGB...
+// For audio, only one channel is used, buf.size() == nx*ny
+//     nx will be n_frames and ny will be n_mel
+struct clip_image_f32 {
+    int nx;
+    int ny;
+
+    std::vector<float> buf;
+};
+
+//
+// logging
+//
+
+static void clip_log_callback_default(enum ggml_log_level level, const char * text, void * user_data) {
+    (void) level;
+    (void) user_data;
+    fputs(text, stderr);
+    fflush(stderr);
+}
+
+struct clip_logger_state {
+    ggml_log_callback log_callback;
+    void * log_callback_user_data;
+};
+
+extern struct clip_logger_state g_logger_state;
+
+static void clip_log_internal_v(enum ggml_log_level level, const char * format, va_list args) {
+    if (format == NULL) {
+        return;
+    }
+    va_list args_copy;
+    va_copy(args_copy, args);
+    char buffer[128];
+    int len = vsnprintf(buffer, 128, format, args);
+    if (len < 128) {
+        g_logger_state.log_callback(level, buffer, g_logger_state.log_callback_user_data);
+    } else {
+        char * buffer2 = (char *) calloc(len + 1, sizeof(char));
+        vsnprintf(buffer2, len + 1, format, args_copy);
+        buffer2[len] = 0;
+        g_logger_state.log_callback(level, buffer2, g_logger_state.log_callback_user_data);
+        free(buffer2);
+    }
+    va_end(args_copy);
+}
+
+static void clip_log_internal(enum ggml_log_level level, const char * format, ...) {
+    va_list args;
+    va_start(args, format);
+    clip_log_internal_v(level, format, args);
+    va_end(args);
+}
+
+#define LOG_INF(...) clip_log_internal(GGML_LOG_LEVEL_INFO,  __VA_ARGS__)
+#define LOG_WRN(...) clip_log_internal(GGML_LOG_LEVEL_WARN,  __VA_ARGS__)
+#define LOG_ERR(...) clip_log_internal(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
+#define LOG_DBG(...) clip_log_internal(GGML_LOG_LEVEL_DEBUG, __VA_ARGS__)
+#define LOG_CNT(...) clip_log_internal(GGML_LOG_LEVEL_CONT,  __VA_ARGS__)
+
+//
+// cpp wrappers
+//
+
+// wrapper for clip_image_size
+struct clip_image_size_deleter {
+    void operator()(clip_image_size * val) { clip_image_size_free(val); }
+};
+typedef std::unique_ptr<clip_image_size, clip_image_size_deleter> clip_image_size_ptr;
+
+// wrapper for clip_image_u8
+struct clip_image_u8_deleter {
+    void operator()(clip_image_u8 * val) { clip_image_u8_free(val); }
+};
+typedef std::unique_ptr<clip_image_u8, clip_image_u8_deleter> clip_image_u8_ptr;
+
+// wrapper for clip_image_f32
+struct clip_image_f32_deleter {
+    void operator()(clip_image_f32 * val) { clip_image_f32_free(val); }
+};
+typedef std::unique_ptr<clip_image_f32, clip_image_f32_deleter> clip_image_f32_ptr;
+
+struct clip_image_u8_batch {
+    std::vector<clip_image_u8_ptr> entries;
+};
+
+struct clip_image_f32_batch {
+    std::vector<clip_image_f32_ptr> entries;
+    bool is_audio = false;
+
+    // for llava-uhd style models, we need to know the grid size
+    // note: entries.size() == grid_x * grid_y + 1 (one overview image)
+    int grid_x = 0;
+    int grid_y = 0;
+
+    clip_image_f32_batch clone() const {
+        clip_image_f32_batch new_batch{
+            /* entries  */ {},
+            /* is_audio */ is_audio,
+            /* grid_x   */ grid_x,
+            /* grid_y   */ grid_y,
+        };
+        new_batch.entries.reserve(entries.size());
+        for (const auto & entry : entries) {
+            new_batch.entries.emplace_back(new clip_image_f32(*entry));
+        }
+        return new_batch;
+    }
+};
+
+//
+// common utils
+//
+
+static std::string string_format(const char * fmt, ...) {
+    va_list ap;
+    va_list ap2;
+    va_start(ap, fmt);
+    va_copy(ap2, ap);
+    int size = vsnprintf(NULL, 0, fmt, ap);
+    GGML_ASSERT(size >= 0 && size < INT_MAX); // NOLINT
+    std::vector<char> buf(size + 1);
+    int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
+    GGML_ASSERT(size2 == size);
+    va_end(ap2);
+    va_end(ap);
+    return std::string(buf.data(), buf.size());
+}
+
+static void string_replace_all(std::string & s, const std::string & search, const std::string & replace) {
+    if (search.empty()) {
+        return;
+    }
+    std::string builder;
+    builder.reserve(s.length());
+    size_t pos = 0;
+    size_t last_pos = 0;
+    while ((pos = s.find(search, last_pos)) != std::string::npos) {
+        builder.append(s, last_pos, pos - last_pos);
+        builder.append(replace);
+        last_pos = pos + search.length();
+    }
+    builder.append(s, last_pos, std::string::npos);
+    s = std::move(builder);
+}
+
+// split string by a `std::string delim` instead of `char delim`
+static std::vector<std::string> string_split_str(std::string s, const std::string & delimiter) {
+    std::vector<std::string> tokens;
+    size_t pos = 0;
+    std::string token;
+    while ((pos = s.find(delimiter)) != std::string::npos) {
+        token = s.substr(0, pos);
+        tokens.push_back(token);
+        s.erase(0, pos + delimiter.length());
+    }
+    tokens.push_back(s);
+    return tokens;
+}
+
+//
+// gguf utils
+//
+
+static std::string gguf_data_to_str(enum gguf_type type, const void * data, int i) {
+    switch (type) {
+        case GGUF_TYPE_UINT8:   return std::to_string(((const uint8_t  *)data)[i]);
+        case GGUF_TYPE_INT8:    return std::to_string(((const int8_t   *)data)[i]);
+        case GGUF_TYPE_UINT16:  return std::to_string(((const uint16_t *)data)[i]);
+        case GGUF_TYPE_INT16:   return std::to_string(((const int16_t  *)data)[i]);
+        case GGUF_TYPE_UINT32:  return std::to_string(((const uint32_t *)data)[i]);
+        case GGUF_TYPE_INT32:   return std::to_string(((const int32_t  *)data)[i]);
+        case GGUF_TYPE_UINT64:  return std::to_string(((const uint64_t *)data)[i]);
+        case GGUF_TYPE_INT64:   return std::to_string(((const int64_t  *)data)[i]);
+        case GGUF_TYPE_FLOAT32: return std::to_string(((const float    *)data)[i]);
+        case GGUF_TYPE_FLOAT64: return std::to_string(((const double   *)data)[i]);
+        case GGUF_TYPE_BOOL:    return ((const bool *)data)[i] ? "true" : "false";
+        default:                return string_format("unknown type %d", type);
+    }
+}
+
+static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {
+    const enum gguf_type type = gguf_get_kv_type(ctx_gguf, i);
+
+    switch (type) {
+        case GGUF_TYPE_STRING:
+            return gguf_get_val_str(ctx_gguf, i);
+        case GGUF_TYPE_ARRAY:
+            {
+                const enum gguf_type arr_type = gguf_get_arr_type(ctx_gguf, i);
+                int arr_n = gguf_get_arr_n(ctx_gguf, i);
+                const void * data = arr_type == GGUF_TYPE_STRING ? nullptr : gguf_get_arr_data(ctx_gguf, i);
+                std::stringstream ss;
+                ss << "[";
+                for (int j = 0; j < arr_n; j++) {
+                    if (arr_type == GGUF_TYPE_STRING) {
+                        std::string val = gguf_get_arr_str(ctx_gguf, i, j);
+                        // escape quotes
+                        string_replace_all(val, "\\", "\\\\");
+                        string_replace_all(val, "\"", "\\\"");
+                        ss << '"' << val << '"';
+                    } else if (arr_type == GGUF_TYPE_ARRAY) {
+                        ss << "???";
+                    } else {
+                        ss << gguf_data_to_str(arr_type, data, j);
+                    }
+                    if (j < arr_n - 1) {
+                        ss << ", ";
+                    }
+                }
+                ss << "]";
+                return ss.str();
+            }
+        default:
+            return gguf_data_to_str(type, gguf_get_val_data(ctx_gguf, i), 0);
+    }
+}
+
+//
+// debugging
+//
+
+static void print_tensor_shape(ggml_tensor * t) {
+    printf("%s.shape = [", t->name);
+    for (int i = 0; i < ggml_n_dims(t); ++i) {
+        printf("%" PRId64, t->ne[i]);
+        if (i < ggml_n_dims(t) - 1) {
+            printf(", ");
+        }
+    }
+    printf("]\n");
+}
+
+static void print_tensor_data(ggml_tensor * t, uint8_t * data, int64_t n) {
+    ggml_type type = t->type;
+    int64_t * ne = t->ne;
+    size_t * nb = t->nb;
+    for (int64_t i3 = 0; i3 < ne[3]; i3++) {
+        printf("%s.data: [\n", t->name);
+        for (int64_t i2 = 0; i2 < ne[2]; i2++) {
+            if (i2 == n && ne[2] > 2*n) {
+                printf("     ..., \n");
+                i2 = ne[2] - n;
+            }
+            printf("     [\n");
+            for (int64_t i1 = 0; i1 < ne[1]; i1++) {
+                if (i1 == n && ne[1] > 2*n) {
+                    printf("      ..., \n");
+                    i1 = ne[1] - n;
+                }
+                printf("      [");
+                for (int64_t i0 = 0; i0 < ne[0]; i0++) {
+                    if (i0 == n && ne[0] > 2*n) {
+                        printf("..., ");
+                        i0 = ne[0] - n;
+                    }
+                    size_t i = i3 * nb[3] + i2 * nb[2] + i1 * nb[1] + i0 * nb[0];
+                    float v;
+                    if (type == GGML_TYPE_F16) {
+                        v = ggml_fp16_to_fp32(*(ggml_fp16_t *) &data[i]);
+                    } else if (type == GGML_TYPE_F32) {
+                        v = *(float *) &data[i];
+                    } else if (type == GGML_TYPE_I32) {
+                        v = (float) *(int32_t *) &data[i];
+                    } else if (type == GGML_TYPE_I16) {
+                        v = (float) *(int16_t *) &data[i];
+                    } else if (type == GGML_TYPE_I8) {
+                        v = (float) *(int8_t *) &data[i];
+                    } else {
+                        GGML_ABORT("fatal error");
+                    }
+                    printf("%8.4f", v);
+                    if (i0 < ne[0] - 1) printf(", ");
+                }
+                printf("],\n");
+            }
+            printf("     ],\n");
+        }
+        printf("    ]\n");
+    }
+}
+
+void clip_debug_encode(clip_ctx * ctx, int h, int w, float fill_value);
+
+//
+// API used internally with mtmd
+//
+
+projector_type clip_get_projector_type(const struct clip_ctx * ctx);

tools/mtmd/clip-model.h

@@ -0,0 +1,389 @@
+#pragma once
+
+#include "ggml.h"
+#include "clip.h"
+#include "clip-impl.h"
+
+#include <array>
+#include <vector>
+#include <unordered_set>
+#include <cstdint>
+#include <cmath>
+
+enum ffn_op_type {
+    FFN_GELU,
+    FFN_GELU_ERF,
+    FFN_SILU,
+    FFN_GELU_QUICK,
+};
+
+enum norm_type {
+    NORM_TYPE_NORMAL,
+    NORM_TYPE_RMS,
+};
+
+enum patch_merge_type {
+    PATCH_MERGE_FLAT,
+    PATCH_MERGE_SPATIAL_UNPAD,
+};
+
+struct clip_hparams {
+    int32_t image_size = 0;
+    int32_t patch_size = 0;
+    int32_t n_embd = 0;
+    int32_t n_ff = 0;
+    int32_t projection_dim = 0;
+    int32_t n_head = 0;
+    int32_t n_layer = 0;
+    // idefics3
+    int32_t image_longest_edge = 0;
+    int32_t image_min_pixels = -1;
+    int32_t image_max_pixels = -1;
+    int32_t n_merge = 0; // number of patch merges **per-side**
+
+    float image_mean[3];
+    float image_std[3];
+
+    // for models using dynamic image size, we need to have a smaller image size to warmup
+    // otherwise, user will get OOM everytime they load the model
+    int32_t warmup_image_size = 0;
+    int32_t warmup_audio_size = 3000;
+
+    ffn_op_type ffn_op = FFN_GELU;
+
+    patch_merge_type mm_patch_merge_type = PATCH_MERGE_FLAT;
+
+    float eps = 1e-6;
+    float rope_theta = 0.0;
+
+    std::vector<clip_image_size> image_res_candidates; // for llava-uhd style models
+    int32_t image_crop_resolution;
+    std::unordered_set<int32_t> vision_feature_layer;
+    int32_t attn_window_size = 0;
+    int32_t n_wa_pattern = 0;
+    std::unordered_set<int32_t> wa_layer_indexes; // explicit layer indexes that use full attention (for irregular patterns like YoutuVL)
+
+    // audio
+    int32_t n_mel_bins = 0; // whisper preprocessor
+    int32_t proj_stack_factor = 0; // ultravox
+
+    // audio-to-mel preprocessor params
+    int32_t audio_chunk_len   = -1; // in seconds
+    int32_t audio_sample_rate = -1;
+    int32_t audio_n_fft       = -1;
+    int32_t audio_window_len  = -1;
+    int32_t audio_hop_len     = -1;
+
+    // legacy
+    bool has_llava_projector = false;
+    int minicpmv_version = 0;
+    int32_t minicpmv_query_num = 0;         // MiniCPM-V query number
+
+    // custom value provided by user, can be undefined if not set
+    int32_t custom_image_min_tokens = -1;
+    int32_t custom_image_max_tokens = -1;
+
+    void set_limit_image_tokens(int n_tokens_min, int n_tokens_max) {
+        const int cur_merge = n_merge == 0 ? 1 : n_merge;
+        const int patch_area = patch_size * patch_size * cur_merge * cur_merge;
+        image_min_pixels = (custom_image_min_tokens > 0 ? custom_image_min_tokens : n_tokens_min) * patch_area;
+        image_max_pixels = (custom_image_max_tokens > 0 ? custom_image_max_tokens : n_tokens_max) * patch_area;
+        warmup_image_size = static_cast<int>(std::sqrt(image_max_pixels));
+    }
+
+    void set_warmup_n_tokens(int n_tokens) {
+        int n_tok_per_side = static_cast<int>(std::sqrt(n_tokens));
+        GGML_ASSERT(n_tok_per_side * n_tok_per_side == n_tokens && "n_tokens must be n*n");
+        const int cur_merge = n_merge == 0 ? 1 : n_merge;
+        warmup_image_size = n_tok_per_side * patch_size * cur_merge;
+        // TODO: support warmup size for custom token numbers
+    }
+};
+
+struct clip_layer {
+    // attention
+    ggml_tensor * k_w = nullptr;
+    ggml_tensor * k_b = nullptr;
+    ggml_tensor * q_w = nullptr;
+    ggml_tensor * q_b = nullptr;
+    ggml_tensor * v_w = nullptr;
+    ggml_tensor * v_b = nullptr;
+    ggml_tensor * qkv_w = nullptr;
+    ggml_tensor * qkv_b = nullptr;
+
+    ggml_tensor * o_w = nullptr;
+    ggml_tensor * o_b = nullptr;
+
+    ggml_tensor * k_norm = nullptr;
+    ggml_tensor * q_norm = nullptr;
+
+    // layernorm 1
+    ggml_tensor * ln_1_w = nullptr;
+    ggml_tensor * ln_1_b = nullptr;
+
+    ggml_tensor * ff_up_w = nullptr;
+    ggml_tensor * ff_up_b = nullptr;
+    ggml_tensor * ff_gate_w = nullptr;
+    ggml_tensor * ff_gate_b = nullptr;
+    ggml_tensor * ff_down_w = nullptr;
+    ggml_tensor * ff_down_b = nullptr;
+
+    // layernorm 2
+    ggml_tensor * ln_2_w = nullptr;
+    ggml_tensor * ln_2_b = nullptr;
+
+    // layer scale (no bias)
+    ggml_tensor * ls_1_w = nullptr;
+    ggml_tensor * ls_2_w = nullptr;
+
+    // qwen3vl deepstack merger
+    ggml_tensor * deepstack_norm_w = nullptr;
+    ggml_tensor * deepstack_norm_b = nullptr;
+    ggml_tensor * deepstack_fc1_w = nullptr;
+    ggml_tensor * deepstack_fc1_b = nullptr;
+    ggml_tensor * deepstack_fc2_w = nullptr;
+    ggml_tensor * deepstack_fc2_b = nullptr;
+
+    // lfm2
+    ggml_tensor * ff_norm_w     = nullptr;
+    ggml_tensor * ff_norm_b     = nullptr;
+    ggml_tensor * ff_norm_1_w   = nullptr;
+    ggml_tensor * ff_norm_1_b   = nullptr;
+    ggml_tensor * ff_up_1_w     = nullptr;
+    ggml_tensor * ff_up_1_b     = nullptr;
+    ggml_tensor * ff_down_1_w   = nullptr;
+    ggml_tensor * ff_down_1_b   = nullptr;
+    ggml_tensor * pos_bias_u    = nullptr;
+    ggml_tensor * pos_bias_v    = nullptr;
+    ggml_tensor * norm_conv_w   = nullptr;
+    ggml_tensor * norm_conv_b   = nullptr;
+    ggml_tensor * linear_pos_w  = nullptr;
+
+    ggml_tensor * conv_norm_w   = nullptr;
+    ggml_tensor * conv_norm_b   = nullptr;
+    ggml_tensor * conv_dw_w     = nullptr;
+    ggml_tensor * conv_dw_b     = nullptr;
+    ggml_tensor * conv_pw1_w    = nullptr;
+    ggml_tensor * conv_pw1_b    = nullptr;
+    ggml_tensor * conv_pw2_w    = nullptr;
+    ggml_tensor * conv_pw2_b    = nullptr;
+
+    bool has_deepstack() const {
+        return deepstack_fc1_w != nullptr;
+    }
+};
+
+// Expanded MobileNetV5 block structure for Gemma3n vision encoder
+struct mobilenetv5_block {
+    // Stage 0 (Edge Residual)
+    ggml_tensor * s0_conv_exp_w = nullptr;
+    ggml_tensor * s0_bn1_w      = nullptr;
+    ggml_tensor * s0_conv_pwl_w = nullptr;
+    ggml_tensor * s0_bn2_w      = nullptr;
+
+    // Stage 1+ (Universal Inverted Residual)
+    ggml_tensor * dw_start_w    = nullptr;
+    ggml_tensor * dw_start_bn_w = nullptr;
+
+    ggml_tensor * pw_exp_w      = nullptr;
+    ggml_tensor * pw_exp_bn_w   = nullptr;
+
+    ggml_tensor * dw_mid_w      = nullptr;
+    ggml_tensor * dw_mid_bn_w   = nullptr;
+
+    ggml_tensor * pw_proj_w     = nullptr;
+    ggml_tensor * pw_proj_bn_w  = nullptr;
+
+    ggml_tensor * layer_scale_w = nullptr;
+
+    // Attention (MQA) components
+    ggml_tensor * attn_q_w = nullptr;
+    ggml_tensor * attn_k_w = nullptr;
+    ggml_tensor * attn_v_w = nullptr;
+    ggml_tensor * attn_o_w = nullptr;
+
+    // Optional downsampling/norm in attention
+    ggml_tensor * attn_k_dw_w   = nullptr;
+    ggml_tensor * attn_k_norm_w = nullptr;
+    ggml_tensor * attn_v_dw_w   = nullptr;
+    ggml_tensor * attn_v_norm_w = nullptr;
+
+    // Block norm (often present in attention blocks)
+    ggml_tensor * attn_norm_w   = nullptr;
+};
+
+struct clip_model {
+    clip_modality modality = CLIP_MODALITY_VISION;
+    projector_type proj_type = PROJECTOR_TYPE_MLP;
+    clip_hparams hparams;
+
+    // embeddings
+    ggml_tensor * class_embedding = nullptr;
+    ggml_tensor * patch_embeddings_0 = nullptr;
+    ggml_tensor * patch_embeddings_1 = nullptr;  // second Conv2D kernel when we decouple Conv3D along temproal dimension (Qwen2VL)
+    ggml_tensor * patch_bias = nullptr;
+    ggml_tensor * position_embeddings = nullptr;
+    ggml_tensor * norm_embd_w = nullptr;
+    ggml_tensor * norm_embd_b = nullptr;
+
+    ggml_tensor * pre_ln_w = nullptr;
+    ggml_tensor * pre_ln_b = nullptr;
+
+    std::vector<clip_layer> layers;
+
+    int32_t n_deepstack_layers = 0; // used by Qwen3-VL, calculated from clip_layer
+
+    ggml_tensor * post_ln_w;
+    ggml_tensor * post_ln_b;
+
+    ggml_tensor * projection; // TODO: rename it to fc (fully connected layer)
+    ggml_tensor * mm_fc_w;
+    ggml_tensor * mm_fc_b;
+    ggml_tensor * mm_ffn_up_w = nullptr;
+    ggml_tensor * mm_ffn_up_b = nullptr;
+    ggml_tensor * mm_ffn_gate_w = nullptr;
+    ggml_tensor * mm_ffn_gate_b = nullptr;
+    ggml_tensor * mm_ffn_down_w = nullptr;
+    ggml_tensor * mm_ffn_down_b = nullptr;
+    ggml_tensor * mm_post_norm_w = nullptr;
+    ggml_tensor * mm_post_norm_b = nullptr;
+
+    // LLaVA projection
+    ggml_tensor * mm_input_norm_w = nullptr;
+    ggml_tensor * mm_input_norm_b = nullptr;
+    ggml_tensor * mm_0_w = nullptr;
+    ggml_tensor * mm_0_b = nullptr;
+    ggml_tensor * mm_2_w = nullptr;
+    ggml_tensor * mm_2_b = nullptr;
+
+    ggml_tensor * image_newline = nullptr;
+
+    // Yi type models with mlp+normalization projection
+    ggml_tensor * mm_1_w = nullptr; // Yi type models have 0, 1, 3, 4
+    ggml_tensor * mm_1_b = nullptr;
+    ggml_tensor * mm_3_w = nullptr;
+    ggml_tensor * mm_3_b = nullptr;
+    ggml_tensor * mm_4_w = nullptr;
+    ggml_tensor * mm_4_b = nullptr;
+
+    // GLMV-Edge projection
+    ggml_tensor * mm_model_adapter_conv_w = nullptr;
+    ggml_tensor * mm_model_adapter_conv_b = nullptr;
+
+    // MobileVLM projection
+    ggml_tensor * mm_model_mlp_1_w = nullptr;
+    ggml_tensor * mm_model_mlp_1_b = nullptr;
+    ggml_tensor * mm_model_mlp_3_w = nullptr;
+    ggml_tensor * mm_model_mlp_3_b = nullptr;
+    ggml_tensor * mm_model_block_1_block_0_0_w = nullptr;
+    ggml_tensor * mm_model_block_1_block_0_1_w = nullptr;
+    ggml_tensor * mm_model_block_1_block_0_1_b = nullptr;
+    ggml_tensor * mm_model_block_1_block_1_fc1_w = nullptr;
+    ggml_tensor * mm_model_block_1_block_1_fc1_b = nullptr;
+    ggml_tensor * mm_model_block_1_block_1_fc2_w = nullptr;
+    ggml_tensor * mm_model_block_1_block_1_fc2_b = nullptr;
+    ggml_tensor * mm_model_block_1_block_2_0_w = nullptr;
+    ggml_tensor * mm_model_block_1_block_2_1_w = nullptr;
+    ggml_tensor * mm_model_block_1_block_2_1_b = nullptr;
+    ggml_tensor * mm_model_block_2_block_0_0_w = nullptr;
+    ggml_tensor * mm_model_block_2_block_0_1_w = nullptr;
+    ggml_tensor * mm_model_block_2_block_0_1_b = nullptr;
+    ggml_tensor * mm_model_block_2_block_1_fc1_w = nullptr;
+    ggml_tensor * mm_model_block_2_block_1_fc1_b = nullptr;
+    ggml_tensor * mm_model_block_2_block_1_fc2_w = nullptr;
+    ggml_tensor * mm_model_block_2_block_1_fc2_b = nullptr;
+    ggml_tensor * mm_model_block_2_block_2_0_w = nullptr;
+    ggml_tensor * mm_model_block_2_block_2_1_w = nullptr;
+    ggml_tensor * mm_model_block_2_block_2_1_b = nullptr;
+
+    // MobileVLM_V2 projection
+    ggml_tensor * mm_model_mlp_0_w = nullptr;
+    ggml_tensor * mm_model_mlp_0_b = nullptr;
+    ggml_tensor * mm_model_mlp_2_w = nullptr;
+    ggml_tensor * mm_model_mlp_2_b = nullptr;
+    ggml_tensor * mm_model_peg_0_w = nullptr;
+    ggml_tensor * mm_model_peg_0_b = nullptr;
+
+    // MINICPMV projection
+    ggml_tensor * mm_model_pos_embed_k = nullptr;
+    ggml_tensor * mm_model_query = nullptr;
+    ggml_tensor * mm_model_proj = nullptr;
+    ggml_tensor * mm_model_kv_proj = nullptr;
+    ggml_tensor * mm_model_attn_q_w = nullptr;
+    ggml_tensor * mm_model_attn_q_b = nullptr;
+    ggml_tensor * mm_model_attn_k_w = nullptr;
+    ggml_tensor * mm_model_attn_k_b = nullptr;
+    ggml_tensor * mm_model_attn_v_w = nullptr;
+    ggml_tensor * mm_model_attn_v_b = nullptr;
+    ggml_tensor * mm_model_attn_o_w = nullptr;
+    ggml_tensor * mm_model_attn_o_b = nullptr;
+    ggml_tensor * mm_model_ln_q_w = nullptr;
+    ggml_tensor * mm_model_ln_q_b = nullptr;
+    ggml_tensor * mm_model_ln_kv_w = nullptr;
+    ggml_tensor * mm_model_ln_kv_b = nullptr;
+    ggml_tensor * mm_model_ln_post_w = nullptr;
+    ggml_tensor * mm_model_ln_post_b = nullptr;
+
+    // gemma3
+    ggml_tensor * mm_input_proj_w = nullptr;
+    ggml_tensor * mm_soft_emb_norm_w = nullptr;
+
+    // mobilenetv5 for gemma3n
+    std::vector<mobilenetv5_block> mobilenet_blocks;
+    std::vector<int> mobilenet_stage_ends;
+    ggml_tensor * mobilenet_stem_conv_w = nullptr;
+    ggml_tensor * mobilenet_stem_conv_b = nullptr;
+    ggml_tensor * mobilenet_stem_norm_w = nullptr;
+    ggml_tensor * mm_post_proj_norm_w = nullptr;
+
+    // Multi-Scale Fusion Adapter (MSFA) components
+    ggml_tensor * msfa_concat_conv_w = nullptr;
+    ggml_tensor * msfa_concat_norm_w = nullptr;
+    ggml_tensor * msfa_ffn_expand_w = nullptr;
+    ggml_tensor * msfa_ffn_project_w = nullptr;
+    ggml_tensor * msfa_ffn_expand_bn = nullptr;
+    ggml_tensor * msfa_ffn_project_bn = nullptr;
+
+
+    // pixtral, glm4v
+    ggml_tensor * token_embd_img_break = nullptr;
+    ggml_tensor * mm_patch_merger_w = nullptr;
+    ggml_tensor * mm_patch_merger_b = nullptr;
+
+    // ultravox / whisper encoder
+    ggml_tensor * conv1d_1_w = nullptr;
+    ggml_tensor * conv1d_1_b = nullptr;
+    ggml_tensor * conv1d_2_w = nullptr;
+    ggml_tensor * conv1d_2_b = nullptr;
+    ggml_tensor * mm_norm_pre_w = nullptr;
+    ggml_tensor * mm_norm_pre_b = nullptr;
+    ggml_tensor * mm_norm_mid_w = nullptr;
+
+    // cogvlm
+    ggml_tensor * mm_post_fc_norm_w = nullptr;
+    ggml_tensor * mm_post_fc_norm_b = nullptr;
+    ggml_tensor * mm_h_to_4h_w = nullptr;
+    ggml_tensor * mm_gate_w = nullptr;
+    ggml_tensor * mm_4h_to_h_w = nullptr;
+    ggml_tensor * mm_boi = nullptr;
+    ggml_tensor * mm_eoi = nullptr;
+
+    // lfm2 audio
+    std::array<ggml_tensor *, 7> pre_encode_conv_X_w = {nullptr};
+    std::array<ggml_tensor *, 7> pre_encode_conv_X_b = {nullptr};
+    ggml_tensor * pre_encode_out_w = nullptr;
+    ggml_tensor * pre_encode_out_b = nullptr;
+
+    bool audio_has_avgpool() const {
+        return proj_type == PROJECTOR_TYPE_QWEN2A
+            || proj_type == PROJECTOR_TYPE_VOXTRAL
+            || proj_type == PROJECTOR_TYPE_MUSIC_FLAMINGO;
+    }
+
+    bool audio_has_stack_frames() const {
+        return proj_type == PROJECTOR_TYPE_ULTRAVOX
+            || proj_type == PROJECTOR_TYPE_VOXTRAL;
+    }
+};
+
+const clip_hparams * clip_get_hparams(const struct clip_ctx * ctx);

tools/mtmd/clip.h

@@ -0,0 +1,121 @@
+#pragma once
+
+#include "ggml.h"
+#include "mtmd.h"
+
+#include <stddef.h>
+#include <stdint.h>
+
+// !!! Internal header, to be used by mtmd only !!!
+
+#define MTMD_INTERNAL_HEADER
+
+struct clip_ctx;
+
+struct clip_image_size {
+    int width;
+    int height;
+};
+
+struct clip_image_f32;
+struct clip_image_u8_batch;
+struct clip_image_f32_batch;
+
+enum clip_modality {
+    CLIP_MODALITY_VISION,
+    CLIP_MODALITY_AUDIO,
+};
+
+enum clip_flash_attn_type {
+    CLIP_FLASH_ATTN_TYPE_AUTO     = -1,
+    CLIP_FLASH_ATTN_TYPE_DISABLED = 0,
+    CLIP_FLASH_ATTN_TYPE_ENABLED  = 1,
+};
+
+struct clip_context_params {
+    bool use_gpu;
+    enum clip_flash_attn_type flash_attn_type;
+    int image_min_tokens;
+    int image_max_tokens;
+    bool warmup;
+    ggml_backend_sched_eval_callback cb_eval;
+    void * cb_eval_user_data;
+};
+
+struct clip_init_result {
+    struct clip_ctx * ctx_v; // vision context
+    struct clip_ctx * ctx_a; // audio context
+};
+
+struct clip_init_result clip_init(const char * fname, struct clip_context_params ctx_params);
+
+void clip_free(struct clip_ctx * ctx);
+
+size_t clip_embd_nbytes(const struct clip_ctx * ctx);
+size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_w, int img_h);
+
+int32_t clip_get_image_size (const struct clip_ctx * ctx);
+int32_t clip_get_patch_size (const struct clip_ctx * ctx);
+int32_t clip_get_hidden_size(const struct clip_ctx * ctx);
+
+// TODO: should be enum, not string
+const char * clip_patch_merge_type(const struct clip_ctx * ctx);
+
+int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * img);
+
+// for M-RoPE, this will be the number of token positions in X and Y directions
+// for other models, X will be the total number of tokens and Y will be 1
+int clip_n_output_tokens_x(const struct clip_ctx * ctx, struct clip_image_f32 * img);
+int clip_n_output_tokens_y(const struct clip_ctx * ctx, struct clip_image_f32 * img);
+
+// this should be equal to the embedding dimension of the text model
+int clip_n_mmproj_embd(const struct clip_ctx * ctx);
+
+struct clip_image_size      * clip_image_size_init(void);
+struct clip_image_u8        * clip_image_u8_init (void);
+struct clip_image_f32       * clip_image_f32_init(void);
+struct clip_image_f32_batch * clip_image_f32_batch_init(void); // only used by libllava
+
+// nx, ny are the output image dimensions
+unsigned char * clip_image_u8_get_data(struct clip_image_u8 * img, uint32_t * nx, uint32_t * ny);
+
+void clip_image_size_free (struct clip_image_size * img_size);
+void clip_image_u8_free (struct clip_image_u8  * img);
+void clip_image_f32_free(struct clip_image_f32 * img);
+void clip_image_u8_batch_free (struct clip_image_u8_batch  * batch);
+void clip_image_f32_batch_free(struct clip_image_f32_batch * batch);
+
+// use for accessing underlay data of clip_image_f32_batch
+size_t clip_image_f32_batch_n_images(const struct clip_image_f32_batch * batch); // equivalent to batch->size()
+size_t clip_image_f32_batch_nx(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->nx
+size_t clip_image_f32_batch_ny(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->ny
+struct clip_image_f32 * clip_image_f32_get_img(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->data
+
+/**
+ * Build image from pixels decoded by other libraries instead of stb_image.h for better performance.
+ * The memory layout is RGBRGBRGB..., input buffer length must be 3*nx*ny bytes
+ */
+void clip_build_img_from_pixels(const unsigned char * rgb_pixels, int nx, int ny, struct clip_image_u8 * img);
+
+/** preprocess img and store the result in res_imgs, pad_to_square may be overridden to false depending on model configuration */
+bool clip_image_preprocess(struct clip_ctx * ctx, const struct clip_image_u8 * img, struct clip_image_f32_batch * res_imgs );
+
+struct ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx);
+
+bool clip_image_encode      (struct clip_ctx * ctx, int n_threads, struct clip_image_f32 * img, float * vec);
+bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, const struct clip_image_f32_batch * imgs, float * vec);
+
+int clip_is_minicpmv(const struct clip_ctx * ctx);
+bool clip_is_glm(const struct clip_ctx * ctx);
+bool clip_is_llava(const struct clip_ctx * ctx);
+// note for contributor: this clip_is_(model) pattern is deprecated
+//                       do NOT add new functions like this
+
+bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec);
+
+// use by audio input
+void clip_image_f32_batch_add_mel(struct clip_image_f32_batch * batch, int n_mel, int n_frames, float * mel);
+
+bool clip_has_vision_encoder(const struct clip_ctx * ctx);
+bool clip_has_audio_encoder(const struct clip_ctx * ctx);
+bool clip_has_whisper_encoder(const struct clip_ctx * ctx);

tools/mtmd/deprecation-warning.cpp

@@ -0,0 +1,22 @@
+#include <cstdio>
+#include <string>
+
+int main(int argc, char** argv) {
+    std::string filename = "main";
+    if (argc >= 1) {
+        filename = argv[0];
+    }
+
+    // Get only the program name from the full path
+    size_t pos = filename.find_last_of("/\\");
+    if (pos != std::string::npos) {
+        filename = filename.substr(pos+1);
+    }
+
+    fprintf(stdout, "\n");
+    fprintf(stdout, "WARNING: The binary '%s' is deprecated.\n", filename.c_str());
+    fprintf(stdout, "Please use 'llama-mtmd-cli' instead.\n");
+    fprintf(stdout, "\n");
+
+    return EXIT_FAILURE;
+}

tools/mtmd/legacy-models/convert_image_encoder_to_gguf.py

@@ -0,0 +1,412 @@
+import argparse
+import os
+import json
+import re
+
+import torch
+import numpy as np
+from gguf import *
+from transformers import CLIPModel, CLIPProcessor, CLIPVisionModel, SiglipVisionModel
+
+TEXT = "clip.text"
+VISION = "clip.vision"
+
+
+def k(raw_key: str, arch: str) -> str:
+    return raw_key.format(arch=arch)
+
+
+def should_skip_tensor(name: str, has_text: bool, has_vision: bool, has_llava: bool) -> bool:
+    if name in (
+        "logit_scale",
+        "text_model.embeddings.position_ids",
+        "vision_model.embeddings.position_ids",
+    ):
+        return True
+
+    if has_llava and name in ["visual_projection.weight", "vision_model.post_layernorm.weight", "vision_model.post_layernorm.bias"]:
+        return True
+
+    if name.startswith("v") and not has_vision:
+        return True
+
+    if name.startswith("t") and not has_text:
+        return True
+
+    return False
+
+
+def get_tensor_name(name: str) -> str:
+    # Standardize the transformers llava next keys for
+    # image newline / mm projector with the classes in haotian-liu LLaVA
+    if name == "image_newline":
+        return "model.image_newline"
+    if name.startswith("multi_modal_projector"):
+        name = name.replace("multi_modal_projector", "mm")
+        if "linear_1" in name:
+            name = name.replace("linear_1", "0")
+        if "linear_2" in name:
+            name = name.replace("linear_2", "2")
+        return name
+
+    if "projection" in name:
+        return name
+    if "mm_projector" in name:
+        name = name.replace("model.mm_projector", "mm")
+        name = re.sub(r'mm\.mlp\.mlp', 'mm.model.mlp', name, count=1)
+        name = re.sub(r'mm\.peg\.peg', 'mm.model.peg', name, count=1)
+        return name
+
+    return name.replace("text_model", "t").replace("vision_model", "v").replace("encoder.layers", "blk").replace("embeddings.", "").replace("_proj", "").replace("self_attn.", "attn_").replace("layer_norm", "ln").replace("layernorm", "ln").replace("mlp.fc1", "ffn_down").replace("mlp.fc2", "ffn_up").replace("embedding", "embd").replace("final", "post").replace("layrnorm", "ln")
+
+
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a significant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = (
+        list(range(ord("!"), ord("~") + 1))
+        + list(range(ord("¡"), ord("¬") + 1))
+        + list(range(ord("®"), ord("ÿ") + 1))
+    )
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8 + n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+
+ap = argparse.ArgumentParser()
+ap.add_argument("-m", "--model-dir", help="Path to model directory cloned from HF Hub", required=True)
+ap.add_argument("--use-f32", action="store_true", default=False, help="Use f32 instead of f16")
+ap.add_argument('--bigendian', action="store_true", default=False, help="Model is executed on big-endian machine")
+ap.add_argument("--text-only", action="store_true", required=False,
+                help="Save a text-only model. It can't be used to encode images")
+ap.add_argument("--vision-only", action="store_true", required=False,
+                help="Save a vision-only model. It can't be used to encode texts")
+ap.add_argument("--clip-model-is-vision", action="store_true", required=False,
+                help="The clip model is a pure vision model (ShareGPT4V vision extract for example)")
+
+# Selectable visual encoders that are compatible with this script
+encoder_group = ap.add_mutually_exclusive_group()
+encoder_group.add_argument("--clip-model-is-openclip", action="store_true", required=False,
+                help="The clip model is from openclip (for ViT-SO400M type))")
+encoder_group.add_argument("--clip-model-is-siglip", action="store_true", required=False,
+                help="the visual encoder is Siglip.")
+
+ap.add_argument("--llava-projector", help="Path to llava.projector file. If specified, save an image encoder for LLaVA models.")
+ap.add_argument("--projector-type", help="Type of projector. Possible values: mlp, ldp, ldpv2", choices=["mlp", "ldp", "ldpv2"], default="mlp")
+ap.add_argument("-o", "--output-dir", help="Directory to save GGUF files. Default is the original model directory", default=None)
+# Example --image_mean 0.48145466 0.4578275 0.40821073 --image_std 0.26862954 0.26130258 0.27577711
+# Example --image_mean 0.5 0.5 0.5 --image_std 0.5 0.5 0.5
+default_image_mean = [0.48145466, 0.4578275, 0.40821073]
+default_image_std = [0.26862954, 0.26130258, 0.27577711]
+ap.add_argument('--image-mean', type=float, nargs='+', help='Mean of the images for normalization (overrides processor) ', default=None)
+ap.add_argument('--image-std', type=float, nargs='+', help='Standard deviation of the images for normalization (overrides processor)', default=None)
+
+# with proper
+args = ap.parse_args()
+
+
+if args.text_only and args.vision_only:
+    print("--text-only and --image-only arguments cannot be specified at the same time.")
+    exit(1)
+
+if args.use_f32:
+    print("WARNING: Weights for the convolution op is always saved in f16, as the convolution op in GGML does not support 32-bit kernel weights yet.")
+
+# output in the same directory as the model if output_dir is None
+dir_model = args.model_dir
+
+if (
+    args.clip_model_is_vision or
+    not os.path.exists(dir_model + "/vocab.json") or
+    args.clip_model_is_openclip or
+    args.clip_model_is_siglip
+):
+    vocab = None
+    tokens = None
+else:
+    with open(dir_model + "/vocab.json", "r", encoding="utf-8") as f:
+        vocab = json.load(f)
+        tokens = [key for key in vocab]
+
+with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
+    config = json.load(f)
+    if args.clip_model_is_vision:
+        v_hparams = config
+        t_hparams = None
+    else:
+        v_hparams = config["vision_config"]
+        t_hparams = config["text_config"]
+
+# possible data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+#
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+ftype = 1
+if args.use_f32:
+    ftype = 0
+
+if args.clip_model_is_siglip:
+    model = SiglipVisionModel.from_pretrained(dir_model)
+    processor = None
+elif args.clip_model_is_vision or args.clip_model_is_openclip:
+    model = CLIPVisionModel.from_pretrained(dir_model)
+    processor = None
+else:
+    model = CLIPModel.from_pretrained(dir_model)
+    processor = CLIPProcessor.from_pretrained(dir_model)
+
+fname_middle = None
+has_text_encoder = True
+has_vision_encoder = True
+has_llava_projector = False
+if args.text_only:
+    fname_middle = "text-"
+    has_vision_encoder = False
+elif args.llava_projector is not None:
+    fname_middle = "mmproj-"
+    has_text_encoder = False
+    has_llava_projector = True
+elif args.vision_only:
+    fname_middle = "vision-"
+    has_text_encoder = False
+else:
+    fname_middle = ""
+
+output_dir = args.output_dir if args.output_dir is not None else dir_model
+os.makedirs(output_dir, exist_ok=True)
+output_prefix = os.path.basename(output_dir).replace("ggml_", "")
+fname_out = os.path.join(output_dir, f"{fname_middle}model-{ftype_str[ftype]}.gguf")
+fout = GGUFWriter(path=fname_out, arch="clip", endianess=GGUFEndian.LITTLE if not args.bigendian else GGUFEndian.BIG)
+
+fout.add_bool("clip.has_text_encoder", has_text_encoder)
+fout.add_bool("clip.has_vision_encoder", has_vision_encoder)
+fout.add_bool("clip.has_llava_projector", has_llava_projector)
+fout.add_file_type(ftype)
+model_name = config["_name_or_path"] if "_name_or_path" in config else os.path.basename(dir_model)
+fout.add_name(model_name)
+if args.text_only:
+    fout.add_description("text-only CLIP model")
+elif args.vision_only and not has_llava_projector:
+    fout.add_description("vision-only CLIP model")
+elif has_llava_projector:
+    fout.add_description("image encoder for LLaVA")
+    # add projector type
+    fout.add_string("clip.projector_type", args.projector_type)
+else:
+    fout.add_description("two-tower CLIP model")
+
+if has_text_encoder:
+    assert t_hparams is not None
+    assert tokens is not None
+    if args.clip_model_is_siglip:
+        text_projection_dim = 0
+    else:
+        text_projection_dim = t_hparams.get("projection_dim", config["projection_dim"])
+    # text_model hparams
+    fout.add_uint32(k(KEY_CONTEXT_LENGTH, TEXT), t_hparams["max_position_embeddings"])
+    fout.add_uint32(k(KEY_EMBEDDING_LENGTH, TEXT), t_hparams["hidden_size"])
+    fout.add_uint32(k(KEY_FEED_FORWARD_LENGTH, TEXT), t_hparams["intermediate_size"])
+    fout.add_uint32("clip.text.projection_dim", text_projection_dim)
+    fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, TEXT), t_hparams["num_attention_heads"])
+    fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, TEXT), t_hparams["layer_norm_eps"])
+    fout.add_uint32(k(KEY_BLOCK_COUNT, TEXT), t_hparams["num_hidden_layers"])
+    fout.add_token_list(tokens)
+
+
+
+def get_non_negative_vision_feature_layers(v_hparams):
+    """
+    Determine the vision feature layer(s) for the llava model, which are indices into the
+    hidden states of the visual encoder. Note that the hidden states array generally takes the
+    form:
+
+        [<emb input>, <output of enc block 0>, ... <output of enc block num_hidden_layers>]
+
+    so feature indices should be offset as n+1 to get the output of encoder block n.
+    We convert all vision feature layers to non-negative so that -1 can be used in
+    the model as an unset value. If no vision feature layer is found, we leave it unset.
+    """
+    num_hidden_layers = v_hparams["num_hidden_layers"]
+    to_non_negative = lambda layer_idx: layer_idx  if layer_idx >= 0 else num_hidden_layers + layer_idx + 1
+    feature_layers_key = None
+    # Key used for llava models in transformers
+    if "vision_feature_layer" in config:
+        feature_layers_key = "vision_feature_layer"
+    # Key used for llava models in the original format
+    elif "mm_vision_select_layer" in config:
+        feature_layers_key = "mm_vision_select_layer"
+    if feature_layers_key is not None:
+        feature_layers = config[feature_layers_key]
+        if isinstance(feature_layers, int):
+            feature_layers = [feature_layers]
+        return [to_non_negative(feature_layer) for feature_layer in feature_layers]
+
+# Determine if we have explicitly specified vision feature layers in our config
+feature_layers = get_non_negative_vision_feature_layers(v_hparams)
+
+if has_vision_encoder:
+    # Siglip does not have a visual projector; set projection dim to 0
+    if args.clip_model_is_siglip:
+        visual_projection_dim = 0
+    else:
+        visual_projection_dim = v_hparams.get("projection_dim", config["projection_dim"])
+
+    # set vision_model hparams
+    fout.add_uint32("clip.vision.image_size", v_hparams["image_size"])
+    fout.add_uint32("clip.vision.patch_size", v_hparams["patch_size"])
+    fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), v_hparams["hidden_size"])
+    fout.add_uint32(k(KEY_FEED_FORWARD_LENGTH, VISION), v_hparams["intermediate_size"])
+    fout.add_uint32("clip.vision.projection_dim", visual_projection_dim)
+    fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, VISION), v_hparams["num_attention_heads"])
+    fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, VISION), v_hparams["layer_norm_eps"])
+    if feature_layers:
+        block_count = max(feature_layers)
+    else:
+        block_count = v_hparams["num_hidden_layers"] - 1 if has_llava_projector else v_hparams["num_hidden_layers"]
+    fout.add_uint32(k(KEY_BLOCK_COUNT, VISION), block_count)
+                            #     /**
+                            #      "image_grid_pinpoints": [
+                            #         [
+                            #         336,
+                            #         672
+                            #         ],
+                            #         [
+                            #         672,
+                            #         336
+                            #         ],
+                            #         [
+                            #         672,
+                            #         672
+                            #         ],
+                            #         [
+                            #         1008,
+                            #         336
+                            #         ],
+                            #         [
+                            #         336,
+                            #         1008
+                            #         ]
+                            #     ],
+                            #     Flattened:
+                            #     [
+                            #         336, 672,
+                            #         672, 336,
+                            #         672, 672,
+                            #         1008, 336,
+                            #         336, 1008
+                            #     ]
+                            #  *
+                            #  */
+    if "image_grid_pinpoints" in v_hparams:
+        # flatten it
+        image_grid_pinpoints = []
+        for pinpoint in v_hparams["image_grid_pinpoints"]:
+            for p in pinpoint:
+                image_grid_pinpoints.append(p)
+        fout.add_array("clip.vision.image_grid_pinpoints", image_grid_pinpoints)
+    if "image_crop_resolution" in v_hparams:
+        fout.add_uint32("clip.vision.image_crop_resolution", v_hparams["image_crop_resolution"])
+    if "image_aspect_ratio" in v_hparams:
+        fout.add_string("clip.vision.image_aspect_ratio", v_hparams["image_aspect_ratio"])
+    if "image_split_resolution" in v_hparams:
+        fout.add_uint32("clip.vision.image_split_resolution", v_hparams["image_split_resolution"])
+    if "mm_patch_merge_type" in v_hparams:
+        fout.add_string("clip.vision.mm_patch_merge_type", v_hparams["mm_patch_merge_type"])
+    if "mm_projector_type" in v_hparams:
+        fout.add_string("clip.vision.mm_projector_type", v_hparams["mm_projector_type"])
+    if feature_layers:
+        fout.add_array("clip.vision.feature_layer", feature_layers)
+
+    if processor is not None:
+        image_mean = processor.image_processor.image_mean if args.image_mean is None or args.image_mean == default_image_mean else args.image_mean  # pyright: ignore[reportAttributeAccessIssue]
+        image_std = processor.image_processor.image_std if args.image_std is None or args.image_std == default_image_std else args.image_std  # pyright: ignore[reportAttributeAccessIssue]
+    else:
+        image_mean = args.image_mean if args.image_mean is not None else default_image_mean
+        image_std = args.image_std if args.image_std is not None else default_image_std
+    fout.add_array("clip.vision.image_mean", image_mean)
+    fout.add_array("clip.vision.image_std", image_std)
+
+use_gelu = v_hparams["hidden_act"] == "gelu"
+fout.add_bool("clip.use_gelu", use_gelu)
+
+
+if has_llava_projector:
+    # By default, we drop the last layer for llava projector
+    # models unless we have explicitly set vision feature layers
+    if feature_layers is None:
+        model.vision_model.encoder.layers.pop(-1)
+    else:
+        model.vision_model.encoder.layers = model.vision_model.encoder.layers[:max(feature_layers)]
+
+    projector = torch.load(args.llava_projector)
+    for name, data in projector.items():
+        name = get_tensor_name(name)
+        # pw and dw conv ndim==4
+        if data.ndim == 2 or data.ndim == 4:
+            data = data.squeeze().numpy().astype(np.float16)
+        else:
+            data = data.squeeze().numpy().astype(np.float32)
+
+        fout.add_tensor(name, data)
+
+    print("Projector tensors added\n")
+
+state_dict = model.state_dict()
+for name, data in state_dict.items():
+    if should_skip_tensor(name, has_text_encoder, has_vision_encoder, has_llava_projector):
+        # we don't need this
+        print(f"skipping parameter: {name}")
+        continue
+
+    name = get_tensor_name(name)
+    data = data.squeeze().numpy()
+
+    n_dims = len(data.shape)
+
+    # ftype == 0 -> float32, ftype == 1 -> float16
+    ftype_cur = 0
+    if n_dims == 4:
+        print(f"tensor {name} is always saved in f16")
+        data = data.astype(np.float16)
+        ftype_cur = 1
+    elif ftype == 1:
+        if name[-7:] == ".weight" and n_dims == 2:
+            print("  Converting to float16")
+            data = data.astype(np.float16)
+            ftype_cur = 1
+        else:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
+    else:
+        if data.dtype != np.float32:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
+
+    print(f"{name} - {ftype_str[ftype_cur]} - shape = {data.shape}")
+    fout.add_tensor(name, data)
+
+
+fout.write_header_to_file()
+fout.write_kv_data_to_file()
+fout.write_tensors_to_file()
+fout.close()
+
+print("Done. Output file: " + fname_out)

tools/mtmd/legacy-models/glmedge-convert-image-encoder-to-gguf.py

@@ -0,0 +1,280 @@
+import argparse
+import os
+import json
+import re
+
+import torch
+import numpy as np
+from gguf import *
+
+TEXT = "clip.text"
+VISION = "clip.vision"
+from transformers import SiglipVisionModel, SiglipVisionConfig
+
+def k(raw_key: str, arch: str) -> str:
+    return raw_key.format(arch=arch)
+
+
+def should_skip_tensor(name: str, has_text: bool, has_vision: bool, has_llava: bool) -> bool:
+    if name in (
+        "logit_scale",
+        "text_model.embeddings.position_ids",
+        "vision_model.embeddings.position_ids",
+    ):
+        return True
+
+    if name in (
+        "vision_model.head.probe",
+        "vision_model.head.attention.in_proj_weight",
+        "vision_model.head.attention.in_proj_bias",
+        "vision_model.head.attention.out_proj.weight",
+        "vision_model.head.attention.out_proj.bias",
+        "vision_model.head.layernorm.weight",
+        "vision_model.head.layernorm.bias",
+        "vision_model.head.mlp.fc1.weight",
+        "vision_model.head.mlp.fc1.bias",
+        "vision_model.head.mlp.fc2.weight",
+        "vision_model.head.mlp.fc2.bias"
+    ):
+        return True
+
+    if name.startswith("v") and not has_vision:
+        return True
+
+    if name.startswith("t") and not has_text:
+        return True
+
+    return False
+
+
+def get_tensor_name(name: str) -> str:
+    if "projection" in name:
+        return name
+    if "mm_projector" in name:
+        name = name.replace("model.mm_projector", "mm")
+        name = re.sub(r'mm\.mlp\.mlp', 'mm.model.mlp', name, count=1)
+        name = re.sub(r'mm\.peg\.peg', 'mm.model.peg', name, count=1)
+        return name
+
+    return name.replace("text_model", "t").replace("vision_model", "v").replace("encoder.layers", "blk").replace("embeddings.", "").replace("_proj", "").replace("self_attn.", "attn_").replace("layer_norm", "ln").replace("layernorm", "ln").replace("mlp.fc1", "ffn_down").replace("mlp.fc2", "ffn_up").replace("embedding", "embd").replace("final", "post").replace("layrnorm", "ln")
+
+
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a significant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = (
+        list(range(ord("!"), ord("~") + 1))
+        + list(range(ord("¡"), ord("¬") + 1))
+        + list(range(ord("®"), ord("ÿ") + 1))
+    )
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8 + n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+
+ap = argparse.ArgumentParser()
+ap.add_argument("-m", "--model-dir", help="Path to model directory cloned from HF Hub", required=True)
+ap.add_argument("--use-f32", action="store_true", default=False, help="Use f32 instead of f16")
+ap.add_argument("--text-only", action="store_true", required=False,
+                help="Save a text-only model. It can't be used to encode images")
+ap.add_argument("--vision-only", action="store_true", required=False,
+                help="Save a vision-only model. It can't be used to encode texts")
+ap.add_argument("--clip-model-is-vision", action="store_true", required=False,
+                help="The clip model is a pure vision model (ShareGPT4V vision extract for example)")
+ap.add_argument("--clip-model-is-openclip", action="store_true", required=False,
+                help="The clip model is from openclip (for ViT-SO400M type))")
+ap.add_argument("--llava-projector", help="Path to llava.projector file. If specified, save an image encoder for LLaVA models.")
+ap.add_argument("--projector-type", help="Type of projector. Possible values: mlp, ldp, ldpv2", choices=["mlp", "ldp", "ldpv2","adapter"], default="adapter")
+ap.add_argument("-o", "--output-dir", help="Directory to save GGUF files. Default is the original model directory", default=None)
+# Example --image_mean 0.48145466 0.4578275 0.40821073 --image_std 0.26862954 0.26130258 0.27577711
+# Example --image_mean 0.5 0.5 0.5 --image_std 0.5 0.5 0.5
+default_image_mean = [0.5, 0.5, 0.5]
+default_image_std = [0.5, 0.5, 0.5]
+ap.add_argument('--image-mean', type=float, nargs='+', help='Mean of the images for normalization (overrides processor) ', default=None)
+ap.add_argument('--image-std', type=float, nargs='+', help='Standard deviation of the images for normalization (overrides processor)', default=None)
+
+# with proper
+args = ap.parse_args()
+
+
+if args.text_only and args.vision_only:
+    print("--text-only and --image-only arguments cannot be specified at the same time.")
+    exit(1)
+
+if args.use_f32:
+    print("WARNING: Weights for the convolution op is always saved in f16, as the convolution op in GGML does not support 32-bit kernel weights yet.")
+
+# output in the same directory as the model if output_dir is None
+dir_model = args.model_dir
+
+if args.clip_model_is_vision or not os.path.exists(dir_model + "/vocab.json") or args.clip_model_is_openclip:
+    vocab = None
+    tokens = None
+else:
+    with open(dir_model + "/vocab.json", "r", encoding="utf-8") as f:
+        vocab = json.load(f)
+        tokens = [key for key in vocab]
+
+with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
+    config = json.load(f)
+    if args.clip_model_is_vision:
+        v_hparams = config
+        t_hparams = None
+    else:
+        v_hparams = config["vision_config"]
+        t_hparams = None
+
+# possible data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+#
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+ftype = 1
+if args.use_f32:
+    ftype = 0
+
+vision_config = SiglipVisionConfig(**v_hparams)
+model = SiglipVisionModel(vision_config)
+model.load_state_dict(torch.load(os.path.join(dir_model, "glm.clip")))
+
+fname_middle = None
+has_text_encoder = False
+has_vision_encoder = True
+has_glm_projector = True
+if args.text_only:
+    fname_middle = "text-"
+    has_vision_encoder = False
+elif args.llava_projector is not None:
+    fname_middle = "mmproj-"
+    has_text_encoder = False
+    has_glm_projector = True
+elif args.vision_only:
+    fname_middle = "vision-"
+    has_text_encoder = False
+else:
+    fname_middle = ""
+
+output_dir = args.output_dir if args.output_dir is not None else dir_model
+os.makedirs(output_dir, exist_ok=True)
+output_prefix = os.path.basename(output_dir).replace("ggml_", "")
+fname_out = os.path.join(output_dir, f"{fname_middle}model-{ftype_str[ftype]}.gguf")
+fout = GGUFWriter(path=fname_out, arch="clip")
+
+fout.add_bool("clip.has_text_encoder", has_text_encoder)
+fout.add_bool("clip.has_vision_encoder", has_vision_encoder)
+fout.add_bool("clip.has_glm_projector", has_glm_projector)
+fout.add_file_type(ftype)
+model_name = config["_name_or_path"] if "_name_or_path" in config else os.path.basename(dir_model)
+fout.add_name(model_name)
+if has_glm_projector:
+    fout.add_description("image encoder for glm4v")
+    fout.add_string("clip.projector_type", "adapter")
+else:
+    fout.add_description("two-tower CLIP model")
+
+if has_text_encoder:
+    assert t_hparams is not None
+    assert tokens is not None
+    # text_model hparams
+    fout.add_uint32(k(KEY_CONTEXT_LENGTH, TEXT), t_hparams["max_position_embeddings"])
+    fout.add_uint32(k(KEY_EMBEDDING_LENGTH, TEXT), t_hparams["hidden_size"])
+    fout.add_uint32(k(KEY_FEED_FORWARD_LENGTH, TEXT), t_hparams["intermediate_size"])
+    fout.add_uint32("clip.text.projection_dim", t_hparams.get("projection_dim", config["projection_dim"]))
+    fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, TEXT), t_hparams["num_attention_heads"])
+    fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, TEXT), t_hparams["layer_norm_eps"])
+    fout.add_uint32(k(KEY_BLOCK_COUNT, TEXT), t_hparams["num_hidden_layers"])
+    fout.add_token_list(tokens)
+
+if has_vision_encoder:
+    # vision_model hparams
+    fout.add_uint32("clip.vision.image_size", v_hparams["image_size"])
+    fout.add_uint32("clip.vision.patch_size", v_hparams["patch_size"])
+    fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), v_hparams["hidden_size"])
+    fout.add_uint32(k(KEY_FEED_FORWARD_LENGTH, VISION), v_hparams["intermediate_size"])
+    fout.add_uint32("clip.vision.projection_dim", 0)
+    fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, VISION), v_hparams["num_attention_heads"])
+    fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, VISION), 1e-6)
+    fout.add_uint32(k(KEY_BLOCK_COUNT, VISION), v_hparams["num_hidden_layers"])
+
+    image_mean = args.image_mean if args.image_mean is not None else default_image_mean
+    image_std = args.image_std if args.image_std is not None else default_image_std
+    fout.add_array("clip.vision.image_mean", image_mean)
+    fout.add_array("clip.vision.image_std", image_std)
+
+fout.add_bool("clip.use_gelu", True)
+
+
+if has_glm_projector:
+    # model.vision_model.encoder.layers.pop(-1)  # pyright: ignore[reportAttributeAccessIssue]
+    projector = torch.load(args.llava_projector)
+    for name, data in projector.items():
+        name = get_tensor_name(name)
+        # pw and dw conv ndim==4
+        if data.ndim == 2 or data.ndim == 4:
+            data = data.squeeze().numpy().astype(np.float16)
+        else:
+            data = data.squeeze().numpy().astype(np.float32)
+        if name.startswith("vision."):
+            name=name.replace("vision.","")
+        fout.add_tensor(name, data)
+        print(f"Projector {name} - {data.dtype} - shape = {data.shape}")
+        # print(f"Projector {name} tensors added\n")
+
+state_dict = model.state_dict()  # pyright: ignore[reportAttributeAccessIssue]
+for name, data in state_dict.items():
+    if should_skip_tensor(name, has_text_encoder, has_vision_encoder, has_glm_projector):
+        # we don't need this
+        print(f"skipping parameter: {name}")
+        continue
+
+    name = get_tensor_name(name)
+    data = data.squeeze().numpy()
+
+    n_dims = len(data.shape)
+
+    # ftype == 0 -> float32, ftype == 1 -> float16
+    ftype_cur = 0
+    if n_dims == 4:
+        print(f"tensor {name} is always saved in f16")
+        data = data.astype(np.float16)
+        ftype_cur = 1
+    elif ftype == 1:
+        if name[-7:] == ".weight" and n_dims == 2:
+            # print("  Converting to float16")
+            data = data.astype(np.float16)
+            ftype_cur = 1
+        else:
+            # print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
+    else:
+        if data.dtype != np.float32:
+            # print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
+    print(f"siglip {name} - {data.dtype} - shape = {data.shape}")
+    # print(f"{name} - {ftype_str[ftype_cur]} - shape = {data.shape}")
+    fout.add_tensor(name, data)
+
+
+fout.write_header_to_file()
+fout.write_kv_data_to_file()
+fout.write_tensors_to_file()
+fout.close()
+
+print("Done. Output file: " + fname_out)

tools/mtmd/legacy-models/glmedge-surgery.py

@@ -0,0 +1,33 @@
+import argparse
+import os
+import torch
+from transformers import AutoModel
+
+ap = argparse.ArgumentParser()
+ap.add_argument("-m", "--model", help="Path to GLM model")
+args = ap.parse_args()
+
+# find the model part that includes the the multimodal projector weights
+model = AutoModel.from_pretrained(args.model, trust_remote_code=True, local_files_only=True)
+checkpoint = model.state_dict()
+
+# get a list of mm tensor names
+mm_tensors = [k for k, v in checkpoint.items() if k.startswith("vision.adapter.")]
+
+# store these tensors in a new dictionary and torch.save them
+projector = {name: checkpoint[name].float() for name in mm_tensors}
+torch.save(projector, f"{args.model}/glm.projector")
+
+clip_tensors = [k for k, v in checkpoint.items() if k.startswith("vision.vit.model.vision_model.")]
+if len(clip_tensors) > 0:
+    clip = {name.replace("vision.vit.model.", ""): checkpoint[name].float() for name in clip_tensors}
+    torch.save(clip, f"{args.model}/glm.clip")
+
+    # added tokens should be removed to be able to convert Mistral models
+    if os.path.exists(f"{args.model}/added_tokens.json"):
+        with open(f"{args.model}/added_tokens.json", "w") as f:
+            f.write("{}\n")
+
+print("Done!")
+print(f"Now you can convert {args.model} to a regular LLaMA GGUF file.")
+print(f"Also, use {args.model}glm.projector to prepare a glm-encoder.gguf file.")

tools/mtmd/legacy-models/llava_surgery.py

@@ -0,0 +1,38 @@
+import argparse
+import glob
+import os
+import torch
+
+
+ap = argparse.ArgumentParser()
+ap.add_argument("-m", "--model", help="Path to LLaVA v1.5 model")
+args = ap.parse_args()
+
+# find the model part that includes the the multimodal projector weights
+path = sorted(glob.glob(f"{args.model}/pytorch_model*.bin"))[-1]
+checkpoint = torch.load(path)
+
+# get a list of mm tensor names
+mm_tensors = [k for k, v in checkpoint.items() if k.startswith("model.mm_projector")]
+
+# store these tensors in a new dictionary and torch.save them
+projector = {name: checkpoint[name].float() for name in mm_tensors}
+torch.save(projector, f"{args.model}/llava.projector")
+
+# BakLLaVA models contain CLIP tensors in it
+clip_tensors = [k for k, v in checkpoint.items() if k.startswith("model.vision_tower")]
+if len(clip_tensors) > 0:
+    clip = {name.replace("vision_tower.vision_tower.", ""): checkpoint[name].float() for name in clip_tensors}
+    torch.save(clip, f"{args.model}/llava.clip")
+
+
+    # added tokens should be removed to be able to convert Mistral models
+    if os.path.exists(f"{args.model}/added_tokens.json"):
+        with open(f"{args.model}/added_tokens.json", "w") as f:
+            f.write("{}\n")
+
+
+
+print("Done!")
+print(f"Now you can convert {args.model} to a regular LLaMA GGUF file.")
+print(f"Also, use {args.model}/llava.projector to prepare a llava-encoder.gguf file.")

tools/mtmd/legacy-models/llava_surgery_v2.py

@@ -0,0 +1,180 @@
+import argparse
+import glob
+import os
+import torch
+from safetensors import safe_open
+from safetensors.torch import save_file
+from typing import Any, ContextManager, cast
+
+# Function to determine if file is a SafeTensor file
+def is_safetensor_file(file_path):
+    return file_path.endswith('.safetensors')
+
+
+# Unified loading function
+def load_model(file_path):
+    if is_safetensor_file(file_path):
+        tensors = {}
+        with cast(ContextManager[Any], safe_open(file_path, framework="pt", device="cpu")) as f:
+            for key in f.keys():
+                tensors[key] = f.get_tensor(key).clone()
+                # output shape
+                print(f"{key} : {tensors[key].shape}")
+        return tensors, 'safetensor'
+    else:
+        return torch.load(file_path, map_location=torch.device('cpu')), 'pytorch'
+
+
+# Unified saving function
+def save_model(model, file_path, file_type):
+    if file_type == 'safetensor':
+        # safe_save(model, file_path)
+        save_file(model, file_path)
+    else:
+        torch.save(model, file_path)
+
+# Helpers to match weight names from specific components or
+# determine if a saved shard contains that component
+def is_vision_tower(weight_name):
+    return (
+        weight_name.startswith("model.vision_tower") or
+        weight_name.startswith("vit.") or
+        weight_name.startswith("vision_tower")
+    )
+
+def is_newline(weight_name):
+    return (
+        weight_name.startswith("model.image_newline") or
+        weight_name.startswith("image_newline")
+    )
+
+def is_mm_projector(weight_name):
+    return (
+        weight_name.startswith("model.mm_projector") or
+        weight_name.startswith("vision_proj.") or
+        weight_name.startswith("multi_modal_projector")
+    )
+
+def newline_criteria(checkpoint):
+    return any(is_newline(k) for k in checkpoint.keys())
+
+def proj_criteria(checkpoint):
+    return any(is_mm_projector(k) for k in checkpoint.keys())
+
+# Adapted function to clean vision tower from checkpoint
+def clean_vision_tower_from_checkpoint(checkpoint_path):
+    checkpoint, file_type = load_model(checkpoint_path)
+    # file_type = 'pytorch'
+    model_path = os.path.dirname(checkpoint_path)
+    print(f"Searching for vision tower tensors in {checkpoint_path}")
+    clip_tensors = [k for k, v in checkpoint.items() if is_vision_tower(k)]
+
+    if len(clip_tensors) > 0:
+        print(f"Found {len(clip_tensors)} tensors to extract from {checkpoint_path}")
+        # Adapted for file type
+        clip_path = os.path.join(model_path, "llava.clip")
+
+        if os.path.exists(clip_path):
+            print(f"Loading existing llava.clip from {clip_path}")
+            existing_clip, _ = load_model(clip_path)
+        else:
+            print(f"Creating new llava.clip at {clip_path}")
+            existing_clip = {}
+        # Update existing_clip with new tensors, avoid duplicates
+        for name in clip_tensors:
+            simple_name = name[name.index('vision_model.'):] if 'vision_model.' in name else name
+            print(f"Adding {simple_name} to llava.clip")
+            if simple_name not in existing_clip:
+                existing_clip[simple_name] = checkpoint[name]
+
+        # Save the updated clip tensors back to llava.clip
+        save_model(existing_clip, clip_path, 'pytorch')
+
+        # Remove the tensors from the original checkpoint
+        for name in clip_tensors:
+            del checkpoint[name]
+
+        checkpoint_path = checkpoint_path
+        return True
+    return False
+
+def find_relevant_checkpoints(checkpoint_paths, newline_criteria, projector):
+    newline_checkpoint_path = None
+    projector_checkpoint_path = None
+
+    for path in checkpoint_paths:
+        checkpoint, _ = load_model(path)
+        if newline_criteria(checkpoint) and newline_checkpoint_path is None:
+            newline_checkpoint_path = path
+        if projector(checkpoint):
+            projector_checkpoint_path = path
+
+    return newline_checkpoint_path, projector_checkpoint_path
+
+
+# Command-line interface setup
+ap = argparse.ArgumentParser()
+ap.add_argument("-m", "--model", required=True, help="Path to LLaVA v1.5+ model")
+ap.add_argument("-C", "--clean-vision-tower", action="store_true", help="Remove any vision tower from the model files")
+args = ap.parse_args()
+
+if args.clean_vision_tower:
+    # Generalized to handle both PyTorch and SafeTensors models
+    model_files = sorted(glob.glob(f"{args.model}/*"), key=os.path.getmtime, reverse=True)
+    # checkpoint_paths = [path for path in model_files if (path.endswith('.bin') and path.startswith('pytorch')) or (path.endswith('.safetensors') and path.startswith('model'))]
+    checkpoint_paths = [path for path in model_files if (path.endswith('.bin') and 'pytorch' in path.split('/')[-1].split('\\')[-1]) or (path.endswith('.safetensors') and 'model' in path.split('/')[-1].split('\\')[-1])]
+    for projector_checkpoint_path in checkpoint_paths:
+        print(f"Cleaning {projector_checkpoint_path}")
+        if not clean_vision_tower_from_checkpoint(projector_checkpoint_path):
+            print(f"No vision tower found in {projector_checkpoint_path}")
+            # we break once none is found, so far all models append them at the end
+            # break
+    print("Done! All vision tower tensors are removed from the model files and stored in llava.clip file.")
+
+# Now we look for the projector in the last checkpoint
+model_files = sorted(glob.glob(f"{args.model}/*"), key=os.path.getmtime, reverse=True)
+checkpoint_paths = [path for path in model_files if (path.endswith('.bin') and 'pytorch' in path.split('/')[-1].split('\\')[-1]) or (path.endswith('.safetensors') and 'model' in path.split('/')[-1].split('\\')[-1])]
+# last_checkpoint_path = checkpoint_paths[0]
+# first_checkpoint_path = checkpoint_paths[-1]
+newline_checkpoint_path, projector_checkpoint_path = find_relevant_checkpoints(checkpoint_paths, newline_criteria, proj_criteria)
+
+print(f"Taking projector from {projector_checkpoint_path}")
+first_mm_tensors = []
+first_checkpoint = None
+if newline_checkpoint_path is not None:
+    print(f"Taking newline from {newline_checkpoint_path}")
+    first_checkpoint, file_type = load_model(newline_checkpoint_path)
+    first_mm_tensors = [k for k, v in first_checkpoint.items() if is_newline(k)]
+
+# Load the checkpoint
+mm_tensors = []
+last_checkpoint = None
+if projector_checkpoint_path is not None:
+    last_checkpoint, file_type = load_model(projector_checkpoint_path)
+    mm_tensors = [k for k, v in last_checkpoint.items() if is_mm_projector(k)]
+
+if len(mm_tensors) == 0:
+    if last_checkpoint is not None:
+        for k, v in last_checkpoint.items():
+            print(k)
+    print(f"Found {len(mm_tensors)} tensors to extract out of {len(last_checkpoint) if last_checkpoint is not None else 0} tensors.")
+    print("No tensors found. Is this a LLaVA model?")
+    exit()
+
+print(f"Found {len(mm_tensors)} tensors to extract.")
+print(f"Found additional {len(first_mm_tensors)} tensors to extract.")
+# projector = {name: checkpoint.[name].float() for name in mm_tensors}
+projector = {}
+for name in mm_tensors:
+    assert last_checkpoint is not None
+    projector[name] = last_checkpoint[name].float()
+for name in first_mm_tensors:
+    assert first_checkpoint is not None
+    projector[name] = first_checkpoint[name].float()
+
+if len(projector) > 0:
+    save_model(projector, f"{args.model}/llava.projector", 'pytorch')
+
+print("Done!")
+print(f"Now you can convert {args.model} to a regular LLaMA GGUF file.")
+print(f"Also, use {args.model}/llava.projector to prepare a llava-encoder.gguf file.")

tools/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py

@@ -0,0 +1,885 @@
+# coding=utf-8
+# Copyright 2024 Google AI and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" PyTorch Siglip model. """
+# Copied from  HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit and add tgt_sizes
+
+
+import os
+import math
+import warnings
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import nn
+from torch.nn.init import _calculate_fan_in_and_fan_out
+
+from transformers.activations import ACT2FN
+from transformers.modeling_utils import PreTrainedModel
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import (
+    logging,
+)
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+class SiglipVisionConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`SiglipVisionModel`]. It is used to instantiate a
+    Siglip vision encoder according to the specified arguments, defining the model architecture. Instantiating a
+    configuration with the defaults will yield a similar configuration to that of the vision encoder of the Siglip
+    [google/siglip-base-patch16-224](https://huggingface.co/google/siglip-base-patch16-224) architecture.
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+    Args:
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        intermediate_size (`int`, *optional*, defaults to 3072):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        num_channels (`int`, *optional*, defaults to 3):
+            Number of channels in the input images.
+        image_size (`int`, *optional*, defaults to 224):
+            The size (resolution) of each image.
+        patch_size (`int`, *optional*, defaults to 16):
+            The size (resolution) of each patch.
+        hidden_act (`str` or `function`, *optional*, defaults to `"gelu_pytorch_tanh"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"selu"` and `"gelu_new"` ``"quick_gelu"` are supported.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-06):
+            The epsilon used by the layer normalization layers.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+    Example:
+    ```python
+    >>> from transformers import SiglipVisionConfig, SiglipVisionModel
+    >>> # Initializing a SiglipVisionConfig with google/siglip-base-patch16-224 style configuration
+    >>> configuration = SiglipVisionConfig()
+    >>> # Initializing a SiglipVisionModel (with random weights) from the google/siglip-base-patch16-224 style configuration
+    >>> model = SiglipVisionModel(configuration)
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "siglip_vision_model"
+
+    def __init__(
+        self,
+        hidden_size=768,
+        intermediate_size=3072,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        num_channels=3,
+        image_size=224,
+        patch_size=16,
+        hidden_act="gelu_pytorch_tanh",
+        layer_norm_eps=1e-6,
+        attention_dropout=0.0,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.num_channels = num_channels
+        self.patch_size = patch_size
+        self.image_size = image_size
+        self.attention_dropout = attention_dropout
+        self.layer_norm_eps = layer_norm_eps
+        self.hidden_act = hidden_act
+
+_CHECKPOINT_FOR_DOC = "google/siglip-base-patch16-224"
+
+SIGLIP_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "google/siglip-base-patch16-224",
+    # See all SigLIP models at https://huggingface.co/models?filter=siglip
+]
+
+# Copied from transformers.models.llama.modeling_llama._get_unpad_data
+def _get_unpad_data(attention_mask):
+    seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
+    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
+    max_seqlen_in_batch = seqlens_in_batch.max().item()
+    cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
+    return (
+        indices,
+        cu_seqlens,
+        max_seqlen_in_batch,
+    )
+
+
+def _trunc_normal_(tensor, mean, std, a, b):
+    # Cut & paste from PyTorch official master until it's in a few official releases - RW
+    # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
+    def norm_cdf(x):
+        # Computes standard normal cumulative distribution function
+        return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0
+
+    if (mean < a - 2 * std) or (mean > b + 2 * std):
+        warnings.warn(
+            "mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
+            "The distribution of values may be incorrect.",
+            stacklevel=2,
+        )
+
+    # Values are generated by using a truncated uniform distribution and
+    # then using the inverse CDF for the normal distribution.
+    # Get upper and lower cdf values
+    l = norm_cdf((a - mean) / std)
+    u = norm_cdf((b - mean) / std)
+
+    # Uniformly fill tensor with values from [l, u], then translate to
+    # [2l-1, 2u-1].
+    tensor.uniform_(2 * l - 1, 2 * u - 1)
+
+    # Use inverse cdf transform for normal distribution to get truncated
+    # standard normal
+    if tensor.dtype in [torch.float16, torch.bfloat16]:
+        # The `erfinv_` op is not (yet?) defined in float16+cpu, bfloat16+gpu
+        og_dtype = tensor.dtype
+        tensor = tensor.to(torch.float32)
+        tensor.erfinv_()
+        tensor = tensor.to(og_dtype)
+    else:
+        tensor.erfinv_()
+
+    # Transform to proper mean, std
+    tensor.mul_(std * math.sqrt(2.0))
+    tensor.add_(mean)
+
+    # Clamp to ensure it's in the proper range
+    if tensor.dtype == torch.float16:
+        # The `clamp_` op is not (yet?) defined in float16+cpu
+        tensor = tensor.to(torch.float32)
+        tensor.clamp_(min=a, max=b)
+        tensor = tensor.to(torch.float16)
+    else:
+        tensor.clamp_(min=a, max=b)
+
+
+def trunc_normal_tf_(
+    tensor: torch.Tensor, mean: float = 0.0, std: float = 1.0, a: float = -2.0, b: float = 2.0
+):
+    """Fills the input Tensor with values drawn from a truncated
+    normal distribution. The values are effectively drawn from the
+    normal distribution :math:`\\mathcal{N}(\text{mean}, \text{std}^2)`
+    with values outside :math:`[a, b]` redrawn until they are within
+    the bounds. The method used for generating the random values works
+    best when :math:`a \\leq \text{mean} \\leq b`.
+    NOTE: this 'tf' variant behaves closer to Tensorflow / JAX impl where the
+    bounds [a, b] are applied when sampling the normal distribution with mean=0, std=1.0
+    and the result is subsquently scaled and shifted by the mean and std args.
+    Args:
+        tensor: an n-dimensional `torch.Tensor`
+        mean: the mean of the normal distribution
+        std: the standard deviation of the normal distribution
+        a: the minimum cutoff value
+        b: the maximum cutoff value
+    """
+    with torch.no_grad():
+        _trunc_normal_(tensor, 0, 1.0, a, b)
+        tensor.mul_(std).add_(mean)
+
+
+def variance_scaling_(tensor, scale=1.0, mode="fan_in", distribution="normal"):
+    fan_in, fan_out = _calculate_fan_in_and_fan_out(tensor)
+    denom = fan_in
+    if mode == "fan_in":
+        denom = fan_in
+    elif mode == "fan_out":
+        denom = fan_out
+    elif mode == "fan_avg":
+        denom = (fan_in + fan_out) / 2
+
+    variance = scale / denom
+
+    if distribution == "truncated_normal":
+        # constant is stddev of standard normal truncated to (-2, 2)
+        trunc_normal_tf_(tensor, std=math.sqrt(variance) / 0.87962566103423978)
+    elif distribution == "normal":
+        with torch.no_grad():
+            tensor.normal_(std=math.sqrt(variance))
+    elif distribution == "uniform":
+        bound = math.sqrt(3 * variance)
+        with torch.no_grad():
+            tensor.uniform_(-bound, bound)
+    else:
+        raise ValueError(f"invalid distribution {distribution}")
+
+
+def lecun_normal_(tensor):
+    variance_scaling_(tensor, mode="fan_in", distribution="truncated_normal")
+
+
+def default_flax_embed_init(tensor):
+    variance_scaling_(tensor, mode="fan_in", distribution="normal")
+
+class SiglipVisionEmbeddings(nn.Module):
+    def __init__(self, config: SiglipVisionConfig):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.image_size = config.image_size
+        self.patch_size = config.patch_size
+
+        self.patch_embedding = nn.Conv2d(
+            in_channels=config.num_channels,
+            out_channels=self.embed_dim,
+            kernel_size=self.patch_size,
+            stride=self.patch_size,
+            padding="valid",
+        )
+
+        self.num_patches_per_side = self.image_size // self.patch_size
+        self.num_patches = self.num_patches_per_side**2
+        self.num_positions = self.num_patches
+        self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
+
+class SiglipAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    # Copied from transformers.models.clip.modeling_clip.CLIPAttention.__init__
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_heads
+        if self.head_dim * self.num_heads != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
+                f" {self.num_heads})."
+            )
+        self.scale = self.head_dim**-0.5
+        self.dropout = config.attention_dropout
+
+        self.k_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.v_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.q_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim)
+
+# Copied from transformers.models.clip.modeling_clip.CLIPMLP with CLIP->Siglip
+class SiglipMLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.activation_fn = ACT2FN[config.hidden_act]
+        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
+        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
+
+
+# Copied from transformers.models.clip.modeling_clip.CLIPEncoderLayer with CLIP->Siglip
+class SiglipEncoderLayer(nn.Module):
+    def __init__(self, config: SiglipVisionConfig):
+        super().__init__()
+        self.embed_dim = config.hidden_size
+        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
+        self.self_attn = (
+            SiglipAttention(config)
+        )
+        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+        self.mlp = SiglipMLP(config)
+        self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+
+class SiglipPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = SiglipVisionConfig
+    base_model_prefix = "siglip"
+    supports_gradient_checkpointing = True
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+
+        if isinstance(module, SiglipVisionEmbeddings):
+            width = self.config.hidden_size
+            nn.init.normal_(module.position_embedding.weight, std=1 / np.sqrt(width))
+        elif isinstance(module, nn.Embedding):
+            default_flax_embed_init(module.weight)
+        elif isinstance(module, SiglipAttention):
+            nn.init.normal_(module.q_proj.weight)
+            nn.init.normal_(module.k_proj.weight)
+            nn.init.normal_(module.v_proj.weight)
+            nn.init.normal_(module.out_proj.weight)
+            nn.init.zeros_(module.q_proj.bias)
+            nn.init.zeros_(module.k_proj.bias)
+            nn.init.zeros_(module.v_proj.bias)
+            nn.init.zeros_(module.out_proj.bias)
+        elif isinstance(module, SiglipMLP):
+            nn.init.normal_(module.fc1.weight)
+            nn.init.normal_(module.fc2.weight)
+            nn.init.normal_(module.fc1.bias, std=1e-6)
+            nn.init.normal_(module.fc2.bias, std=1e-6)
+        elif isinstance(module, (nn.Linear, nn.Conv2d)):
+            lecun_normal_(module.weight)
+            if module.bias is not None:
+                nn.init.zeros_(module.bias)
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+
+SIGLIP_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+    Parameters:
+        config ([`SiglipVisionConfig`]): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+
+SIGLIP_VISION_INPUTS_DOCSTRING = r"""
+    Args:
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using
+            [`AutoImageProcessor`]. See [`CLIPImageProcessor.__call__`] for details.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+# Copied from transformers.models.clip.modeling_clip.CLIPEncoder with CLIP->Siglip
+class SiglipEncoder(nn.Module):
+    """
+    Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
+    [`SiglipEncoderLayer`].
+    Args:
+        config: SiglipConfig
+    """
+
+    def __init__(self, config: SiglipVisionConfig):
+        super().__init__()
+        self.config = config
+        self.layers = nn.ModuleList([SiglipEncoderLayer(config) for _ in range(config.num_hidden_layers)])
+        self.gradient_checkpointing = False
+
+class SiglipVisionTransformer(SiglipPreTrainedModel):
+    config_class = SiglipVisionConfig
+    main_input_name = "pixel_values"
+    _supports_flash_attn_2 = True
+
+    def __init__(self, config: SiglipVisionConfig):
+        super().__init__(config)
+        self.config = config
+        embed_dim = config.hidden_size
+
+        self.embeddings = SiglipVisionEmbeddings(config)
+        self.encoder = SiglipEncoder(config)
+        self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
+        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.embeddings.patch_embedding
+
+import argparse
+import json
+import re
+
+import numpy as np
+from gguf import *
+from transformers.models.idefics2.modeling_idefics2 import Idefics2VisionTransformer
+from transformers.models.idefics2.configuration_idefics2 import Idefics2VisionConfig
+
+TEXT = "clip.text"
+VISION = "clip.vision"
+
+
+def add_key_str(raw_key: str, arch: str) -> str:
+    return raw_key.format(arch=arch)
+
+
+def should_skip_tensor(name: str, has_text: bool, has_vision: bool, has_minicpmv: bool) -> bool:
+    if name in (
+        "logit_scale",
+        "text_model.embeddings.position_ids",
+        "vision_model.embeddings.position_ids",
+    ):
+        return True
+
+    if has_minicpmv and name in ["visual_projection.weight"]:
+        return True
+
+    if name.startswith("v") and not has_vision:
+        return True
+
+    if name.startswith("t") and not has_text:
+        return True
+
+    return False
+
+
+def get_tensor_name(name: str) -> str:
+    if "projection" in name:
+        return name
+    if "mm_projector" in name:
+        name = name.replace("model.mm_projector", "mm")
+        name = re.sub(r'mm\.mlp\.mlp', 'mm.model.mlp', name, count=1)
+        name = re.sub(r'mm\.peg\.peg', 'mm.model.peg', name, count=1)
+        return name
+
+    return name.replace("text_model", "t").replace("vision_model", "v").replace("encoder.layers", "blk").replace("embeddings.", "").replace("_proj", "").replace("self_attn.", "attn_").replace("layer_norm", "ln").replace("layernorm", "ln").replace("mlp.fc1", "ffn_down").replace("mlp.fc2", "ffn_up").replace("embedding", "embd").replace("final", "post").replace("layrnorm", "ln")
+
+
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a significant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = (
+        list(range(ord("!"), ord("~") + 1))
+        + list(range(ord("¡"), ord("¬") + 1))
+        + list(range(ord("®"), ord("ÿ") + 1))
+    )
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8 + n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+
+ap = argparse.ArgumentParser()
+ap.add_argument("-m", "--model-dir", help="Path to model directory cloned from HF Hub", required=True)
+ap.add_argument("--use-f32", action="store_true", default=False, help="Use f32 instead of f16")
+ap.add_argument("--text-only", action="store_true", required=False,
+                help="Save a text-only model. It can't be used to encode images")
+ap.add_argument("--vision-only", action="store_true", required=False,
+                help="Save a vision-only model. It can't be used to encode texts")
+ap.add_argument("--clip-model-is-vision", action="store_true", required=False,
+                help="The clip model is a pure vision model (ShareGPT4V vision extract for example)")
+ap.add_argument("--clip-model-is-openclip", action="store_true", required=False,
+                help="The clip model is from openclip (for ViT-SO400M type))")
+ap.add_argument("--minicpmv-projector", help="Path to minicpmv.projector file. If specified, save an image encoder for MiniCPM-V models.")
+ap.add_argument("--projector-type", help="Type of projector. Possible values: mlp, ldp, ldpv2", choices=["mlp", "ldp", "ldpv2"], default="mlp")
+ap.add_argument("-o", "--output-dir", help="Directory to save GGUF files. Default is the original model directory", default=None)
+# Example --image_mean 0.48145466 0.4578275 0.40821073 --image_std 0.26862954 0.26130258 0.27577711
+# Example --image_mean 0.5 0.5 0.5 --image_std 0.5 0.5 0.5
+default_image_mean = [0.5, 0.5, 0.5]
+default_image_std = [0.5, 0.5, 0.5]
+ap.add_argument('--image-mean', type=float, nargs='+', help='Mean of the images for normalization (overrides processor) ', default=None)
+ap.add_argument('--image-std', type=float, nargs='+', help='Standard deviation of the images for normalization (overrides processor)', default=None)
+ap.add_argument('--minicpmv_version', type=int, help='minicpmv_version: MiniCPM-V-2 use 1; MiniCPM-V-2.5 use 2; MiniCPM-V-2.6 use 3; MiniCPM-o-2.6 use 4; MiniCPM-V 4.0 use 5; MiniCPM-o-4.0 use 6', default=2)
+
+# with proper
+args = ap.parse_args()
+
+
+if args.text_only and args.vision_only:
+    print("--text-only and --image-only arguments cannot be specified at the same time.")
+    exit(1)
+
+if args.use_f32:
+    print("WARNING: Weights for the convolution op is always saved in f16, as the convolution op in GGML does not support 32-bit kernel weights yet.")
+
+# output in the same directory as the model if output_dir is None
+dir_model = args.model_dir
+
+# Read config.json to get actual model configuration
+config_path = os.path.join(dir_model, "config.json")
+model_config = {}
+if os.path.isfile(config_path):
+    with open(config_path, "r", encoding="utf-8") as f:
+        model_config = json.load(f)
+    print(f"Loaded config from {config_path}")
+else:
+    print(f"Warning: config.json not found at {config_path}")
+
+# If minicpmv_projector is not specified but the default path exists, use the default path
+if args.minicpmv_projector is None:
+    default_projector_path = os.path.join(dir_model, "minicpmv.projector")
+    if os.path.isfile(default_projector_path):
+        args.minicpmv_projector = default_projector_path
+        print(f"Found default projector file: {default_projector_path}")
+
+# If output_dir is not specified, use model_dir as the default value
+if args.output_dir is None:
+    args.output_dir = dir_model
+
+if args.clip_model_is_vision or not os.path.exists(dir_model + "/vocab.json") or args.clip_model_is_openclip:
+    vocab = None
+    tokens = None
+else:
+    with open(dir_model + "/vocab.json", "r", encoding="utf-8") as f:
+        vocab = json.load(f)
+        tokens = [key for key in vocab]
+
+# possible data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+#
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+ftype = 1
+if args.use_f32:
+    ftype = 0
+
+# if args.clip_model_is_vision or args.clip_model_is_openclip:
+#     model = CLIPVisionModel.from_pretrained(dir_model)
+#     processor = None
+# else:
+#     model = CLIPModel.from_pretrained(dir_model)
+#     processor = CLIPProcessor.from_pretrained(dir_model)
+
+minicpmv_version = args.minicpmv_version
+
+# Use actual config values instead of hardcoded ones
+if model_config:
+    # For the projector/resampler, use the main model's hidden_size
+    emb_dim = model_config.get("hidden_size", 1536)
+
+    # For the vision model, use vision_config values
+    vision_config_dict = model_config.get("vision_config", {})
+    default_vision_config = {
+        "hidden_size": vision_config_dict.get("hidden_size", 1152),
+        "image_size": vision_config_dict.get("image_size", 980),
+        "intermediate_size": vision_config_dict.get("intermediate_size", 4304),
+        "model_type": vision_config_dict.get("model_type", "siglip"),
+        "num_attention_heads": vision_config_dict.get("num_attention_heads", 16),
+        "num_hidden_layers": vision_config_dict.get("num_hidden_layers", 27),
+        "patch_size": vision_config_dict.get("patch_size", 14),
+    }
+
+    # Use vision model's num_hidden_layers for block_count
+    block_count = vision_config_dict.get("num_hidden_layers", 27)
+
+    print(f"Using config values: emb_dim={emb_dim}, block_count={block_count}")
+    print(f"Vision config: {default_vision_config}")
+else:
+    # Fallback to original hardcoded logic if config.json not found
+    emb_dim = 4096
+    block_count = 26
+    if minicpmv_version == 1:
+        emb_dim = 2304
+        block_count = 26
+    elif minicpmv_version == 2:
+        emb_dim = 4096
+        block_count = 27
+    elif minicpmv_version == 3:
+        emb_dim = 3584
+        block_count = 27
+    elif minicpmv_version == 4:
+        emb_dim = 3584
+        block_count = 27
+    elif minicpmv_version == 5:
+        emb_dim = 2560
+        block_count = 27
+    elif minicpmv_version == 6:
+        emb_dim = 4096
+        block_count = 27
+
+    default_vision_config = {
+            "hidden_size": 1152,
+            "image_size": 980,
+            "intermediate_size": 4304,
+            "model_type": "idefics2",
+            "num_attention_heads": 16,
+            "num_hidden_layers": 27,
+            "patch_size": 14,
+        }
+
+vision_config = Idefics2VisionConfig(**default_vision_config)
+model = Idefics2VisionTransformer(vision_config)
+if minicpmv_version == 3 or (model_config and model_config.get("vision_config", {}).get("model_type") == "siglip"):
+    vision_config = SiglipVisionConfig(**default_vision_config)
+    model = SiglipVisionTransformer(vision_config)
+elif minicpmv_version == 4:
+    vision_config = SiglipVisionConfig(**default_vision_config)
+    model = SiglipVisionTransformer(vision_config)
+elif minicpmv_version == 5:
+    default_vision_config["model_type"] = "siglip_vision_model"
+    vision_config = SiglipVisionConfig(**default_vision_config)
+    model = SiglipVisionTransformer(vision_config)
+elif minicpmv_version == 6:
+    default_vision_config["model_type"] = "siglip_vision_model"
+    vision_config = SiglipVisionConfig(**default_vision_config)
+    model = SiglipVisionTransformer(vision_config)
+
+processor = None
+# if model.attn_pool is not None:
+#     model.attn_pool = torch.nn.Identity()
+
+# model.blocks = model.blocks[:-1]
+model.load_state_dict(torch.load(os.path.join(dir_model, "minicpmv.clip")))
+
+fname_middle = None
+has_text_encoder = True
+has_vision_encoder = True
+has_minicpmv_projector = False
+
+if args.text_only:
+    fname_middle = "text-"
+    has_vision_encoder = False
+elif args.minicpmv_projector is not None:
+    fname_middle = "mmproj-"
+    has_text_encoder = False
+    has_minicpmv_projector = True
+elif args.vision_only:
+    fname_middle = "vision-"
+    has_text_encoder = False
+else:
+    fname_middle = ""
+
+output_dir = args.output_dir
+os.makedirs(output_dir, exist_ok=True)
+output_prefix = os.path.basename(output_dir).replace("ggml_", "")
+fname_out = os.path.join(output_dir, f"{fname_middle}model-{ftype_str[ftype]}.gguf")
+fout = GGUFWriter(path=fname_out, arch="clip")
+
+fout.add_bool("clip.has_text_encoder", has_text_encoder)
+fout.add_bool("clip.has_vision_encoder", has_vision_encoder)
+fout.add_bool("clip.has_minicpmv_projector", has_minicpmv_projector)
+fout.add_file_type(ftype)
+if args.text_only:
+    fout.add_description("text-only CLIP model")
+elif args.vision_only and not has_minicpmv_projector:
+    fout.add_description("vision-only CLIP model")
+elif has_minicpmv_projector:
+    fout.add_description("image encoder for MiniCPM-V")
+    # add projector type
+    fout.add_string("clip.projector_type", "resampler")
+    fout.add_int32("clip.minicpmv_version", minicpmv_version)
+else:
+    fout.add_description("two-tower CLIP model")
+
+if has_vision_encoder:
+    # vision_model hparams - use actual config values
+    vision_image_size = model_config.get("image_size", 448) if model_config else 448
+    vision_patch_size = default_vision_config.get("patch_size", 14)
+    vision_hidden_size = default_vision_config.get("hidden_size", 1152)
+    vision_intermediate_size = default_vision_config.get("intermediate_size", 4304)
+    vision_attention_heads = default_vision_config.get("num_attention_heads", 16)
+
+    fout.add_uint32("clip.vision.image_size", vision_image_size)
+    fout.add_uint32("clip.vision.patch_size", vision_patch_size)
+    fout.add_uint32(add_key_str(KEY_EMBEDDING_LENGTH, VISION), vision_hidden_size)
+    fout.add_uint32(add_key_str(KEY_FEED_FORWARD_LENGTH, VISION), vision_intermediate_size)
+    fout.add_uint32("clip.vision.projection_dim", 0)
+    fout.add_uint32(add_key_str(KEY_ATTENTION_HEAD_COUNT, VISION), vision_attention_heads)
+    fout.add_float32(add_key_str(KEY_ATTENTION_LAYERNORM_EPS, VISION), 1e-6)
+    fout.add_uint32(add_key_str(KEY_BLOCK_COUNT, VISION), block_count)
+
+    # Add MiniCPM-V specific parameters
+    query_num = model_config.get("query_num", 0) if model_config else 0
+    resampler_emb_dim = model_config.get("hidden_size", 0) if model_config else 0
+    fout.add_uint32("clip.minicpmv_query_num", query_num)
+
+    if processor is not None:
+        image_mean = processor.image_processor.image_mean if args.image_mean is None or args.image_mean == default_image_mean else args.image_mean
+        image_std = processor.image_processor.image_std if args.image_std is None or args.image_std == default_image_std else args.image_std
+    else:
+        image_mean = args.image_mean if args.image_mean is not None else default_image_mean
+        image_std = args.image_std if args.image_std is not None else default_image_std
+    fout.add_array("clip.vision.image_mean", image_mean)
+    fout.add_array("clip.vision.image_std", image_std)
+
+use_gelu = True
+fout.add_bool("clip.use_gelu", use_gelu)
+
+def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
+    """
+    embed_dim: output dimension for each position
+    pos: a list of positions to be encoded: size (M,)
+    out: (M, D)
+    """
+    assert embed_dim % 2 == 0
+    omega = np.arange(embed_dim // 2, dtype=np.float32)
+    omega /= embed_dim / 2.
+    omega = 1. / 10000 ** omega  # (D/2,)
+
+    pos = pos.reshape(-1)  # (M,)
+    out = np.einsum('m,d->md', pos, omega)  # (M, D/2), outer product
+
+    emb_sin = np.sin(out)  # (M, D/2)
+    emb_cos = np.cos(out)  # (M, D/2)
+
+    emb = np.concatenate([emb_sin, emb_cos], axis=1)  # (M, D)
+    return emb
+
+def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
+    assert embed_dim % 2 == 0
+
+    # use half of dimensions to encode grid_h
+    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])  # (H*W, D/2)
+    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])  # (H*W, D/2)
+
+    emb = np.concatenate([emb_h, emb_w], axis=1)  # (H*W, D)
+    return emb
+
+
+# https://github.com/facebookresearch/mae/blob/efb2a8062c206524e35e47d04501ed4f544c0ae8/util/pos_embed.py#L20
+def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False):
+    """
+    grid_size: int of the grid height and width
+    return:
+    pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
+    """
+    if isinstance(grid_size, int):
+        grid_h_size, grid_w_size = grid_size, grid_size
+    else:
+        grid_h_size, grid_w_size = grid_size[0], grid_size[1]
+
+    grid_h = np.arange(grid_h_size, dtype=np.float32)
+    grid_w = np.arange(grid_w_size, dtype=np.float32)
+    grid = np.meshgrid(grid_w, grid_h)  # here w goes first
+    grid = np.stack(grid, axis=0)
+
+    grid = grid.reshape([2, 1, grid_h_size, grid_w_size])
+    pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
+    if cls_token:
+        pos_embed = np.concatenate([np.zeros([1, embed_dim]), pos_embed], axis=0)
+    return pos_embed
+
+def _replace_name_resampler(s, v):
+    if re.match("resampler.pos_embed", s):
+        return {
+            s: v,
+            re.sub("pos_embed", "pos_embed_k", s): torch.from_numpy(get_2d_sincos_pos_embed(emb_dim, (70, 70))),
+        }
+    if re.match("resampler.proj", s):
+        return {
+            re.sub("proj", "pos_embed_k", s): torch.from_numpy(get_2d_sincos_pos_embed(emb_dim, (70, 70))),
+            re.sub("proj", "proj.weight", s): v.transpose(-1, -2).contiguous(),
+        }
+    if re.match("resampler.attn.in_proj_.*", s):
+        return {
+            re.sub("attn.in_proj_", "attn.q.", s): v.chunk(3, dim=0)[0],
+            re.sub("attn.in_proj_", "attn.k.", s): v.chunk(3, dim=0)[1],
+            re.sub("attn.in_proj_", "attn.v.", s): v.chunk(3, dim=0)[2],
+        }
+    return {s: v}
+
+if has_minicpmv_projector:
+    projector = torch.load(args.minicpmv_projector)
+    new_state_dict = {}
+    for k, v in projector.items():
+        kvs = _replace_name_resampler(k, v)
+        for nk, nv in kvs.items():
+            new_state_dict[nk] = nv
+    projector = new_state_dict
+    ftype_cur = 0
+    for name, data in projector.items():
+        name = get_tensor_name(name)
+        data = data.squeeze().numpy()
+
+        n_dims = len(data.shape)
+        if ftype == 1:
+            if name[-7:] == ".weight" and n_dims == 2:
+                print("  Converting to float16")
+                data = data.astype(np.float16)
+                ftype_cur = 1
+            else:
+                print("  Converting to float32")
+                data = data.astype(np.float32)
+                ftype_cur = 0
+        else:
+            if data.dtype != np.float32:
+                print("  Converting to float32")
+                data = data.astype(np.float32)
+                ftype_cur = 0
+
+        fout.add_tensor(name, data)
+        print(f"{name} - {ftype_str[ftype_cur]} - shape = {data.shape}")
+
+    print("Projector tensors added\n")
+
+def _replace_name(s, v):
+    s = "vision_model." + s
+    if re.match("vision_model.embeddings.position_embedding", s):
+        v = v.unsqueeze(0)
+        return {s: v}
+
+    return {s: v}
+
+state_dict = model.state_dict()
+new_state_dict = {}
+for k, v in state_dict.items():
+    kvs = _replace_name(k, v)
+    for nk, nv in kvs.items():
+        new_state_dict[nk] = nv
+state_dict = new_state_dict
+for name, data in state_dict.items():
+    if should_skip_tensor(name, has_text_encoder, has_vision_encoder, has_minicpmv_projector):
+        # we don't need this
+        print(f"skipping parameter: {name}")
+        continue
+
+    name = get_tensor_name(name)
+    data = data.squeeze().numpy()
+
+    n_dims = len(data.shape)
+
+    # ftype == 0 -> float32, ftype == 1 -> float16
+    ftype_cur = 0
+    if n_dims == 4:
+        print(f"tensor {name} is always saved in f16")
+        data = data.astype(np.float16)
+        ftype_cur = 1
+    elif ftype == 1:
+        if name[-7:] == ".weight" and n_dims == 2:
+            print("  Converting to float16")
+            data = data.astype(np.float16)
+            ftype_cur = 1
+        else:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
+    else:
+        if data.dtype != np.float32:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
+
+    print(f"{name} - {ftype_str[ftype_cur]} - shape = {data.shape}")
+    fout.add_tensor(name, data)
+
+
+fout.write_header_to_file()
+fout.write_kv_data_to_file()
+fout.write_tensors_to_file()
+fout.close()
+
+print("Done. Output file: " + fname_out)

tools/mtmd/legacy-models/minicpmv-surgery.py

@@ -0,0 +1,47 @@
+import argparse
+import os
+import torch
+from transformers import AutoModel, AutoTokenizer
+
+ap = argparse.ArgumentParser()
+ap.add_argument("-m", "--model", help="Path to MiniCPM-V model")
+args = ap.parse_args()
+
+# find the model part that includes the the multimodal projector weights
+model = AutoModel.from_pretrained(args.model, trust_remote_code=True, local_files_only=True, torch_dtype=torch.bfloat16)
+checkpoint = model.state_dict()
+
+# get a list of mm tensor names
+mm_tensors = [k for k, v in checkpoint.items() if k.startswith("resampler")]
+
+# store these tensors in a new dictionary and torch.save them
+projector = {name: checkpoint[name].float() for name in mm_tensors}
+if 'resampler.proj' in projector.keys() and hasattr(model.llm.config,'scale_emb') is True:
+    projector['resampler.proj'] = projector['resampler.proj'] / model.llm.config.scale_emb
+torch.save(projector, f"{args.model}/minicpmv.projector")
+
+clip_tensors = [k for k, v in checkpoint.items() if k.startswith("vpm")]
+if len(clip_tensors) > 0:
+    clip = {name.replace("vpm.", ""): checkpoint[name].float() for name in clip_tensors}
+    torch.save(clip, f"{args.model}/minicpmv.clip")
+
+    # added tokens should be removed to be able to convert Mistral models
+    if os.path.exists(f"{args.model}/added_tokens.json"):
+        with open(f"{args.model}/added_tokens.json", "w") as f:
+            f.write("{}\n")
+
+config = model.llm.config
+config.auto_map = {
+    "AutoConfig": "configuration_minicpm.MiniCPMConfig",
+    "AutoModel": "modeling_minicpm.MiniCPMModel",
+    "AutoModelForCausalLM": "modeling_minicpm.MiniCPMForCausalLM",
+    "AutoModelForSeq2SeqLM": "modeling_minicpm.MiniCPMForCausalLM",
+    "AutoModelForSequenceClassification": "modeling_minicpm.MiniCPMForSequenceClassification"
+}
+model.llm.save_pretrained(f"{args.model}/model")
+tok = AutoTokenizer.from_pretrained(args.model, trust_remote_code=True)
+tok.save_pretrained(f"{args.model}/model")
+
+print("Done!")
+print(f"Now you can convert {args.model} to a regular LLaMA GGUF file.")
+print(f"Also, use {args.model}/minicpmv.projector to prepare a minicpmv-encoder.gguf file.")

tools/mtmd/models/cogvlm.cpp

@@ -0,0 +1,98 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_cogvlm::build() {
+    GGML_ASSERT(model.class_embedding != nullptr);
+    GGML_ASSERT(model.position_embeddings != nullptr);
+
+    const int n_pos = n_patches + 1; // +1 for [CLS]
+
+    // build input and concatenate class embedding
+    ggml_tensor * inp = build_inp();
+    inp = ggml_concat(ctx0, inp, model.class_embedding, 1);
+
+    inp = ggml_add(ctx0, inp, model.position_embeddings);
+    cb(inp, "inp_pos", -1);
+
+    ggml_tensor * inpL = inp;
+
+    for (int il = 0; il < n_layer; il++) {
+        auto & layer = model.layers[il];
+        ggml_tensor * cur = inpL;
+
+        cur = ggml_mul_mat(ctx0, layer.qkv_w, cur);
+
+        cur = ggml_add(ctx0, cur, layer.qkv_b);
+
+        ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, d_head, n_head, n_pos, d_head*sizeof(float),
+            cur->nb[1], 0);
+        ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, d_head, n_head, n_pos, d_head*sizeof(float),
+            cur->nb[1], n_embd * sizeof(float));
+        ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, d_head, n_head, n_pos, d_head*sizeof(float),
+            cur->nb[1], 2 * n_embd * sizeof(float));
+
+        cb(Qcur, "Qcur", il);
+        cb(Kcur, "Kcur", il);
+        cb(Vcur, "Vcur", il);
+
+        cur = build_attn(layer.o_w, layer.o_b,
+            Qcur, Kcur, Vcur, nullptr, kq_scale, il);
+        cb(cur, "attn_out", il);
+
+        cur = build_norm(cur, layer.ln_1_w, layer.ln_1_b, NORM_TYPE_NORMAL, eps, il);
+        cb(cur, "attn_post_norm", il);
+
+        cur = ggml_add(ctx0, cur, inpL);
+        inpL = cur;
+
+        cur = build_ffn(cur,
+            layer.ff_up_w, layer.ff_up_b,
+            layer.ff_gate_w, layer.ff_gate_b,
+            layer.ff_down_w, layer.ff_down_b,
+            hparams.ffn_op, il);
+
+        cb(cur, "ffn_out", il);
+
+        cur = build_norm(cur, layer.ln_2_w, layer.ln_2_b, NORM_TYPE_NORMAL, eps, il);
+        cb(cur, "ffn_post_norm", il);
+
+        cur = ggml_add(ctx0, cur, inpL);
+        cb(cur, "layer_out", il);
+        inpL = cur;
+
+    }
+
+    // remove CLS token (like build_llama4 does)
+    ggml_tensor * cur = ggml_view_2d(ctx0, inpL,
+        n_embd, n_patches,
+        ggml_row_size(inpL->type, n_embd), 0);
+
+    // Multiply with mm_model_proj
+    cur = ggml_mul_mat(ctx0, model.mm_model_proj, cur);
+
+    // Apply layernorm, weight, bias
+    cur = build_norm(cur, model.mm_post_fc_norm_w, model.mm_post_fc_norm_b, NORM_TYPE_NORMAL, 1e-5, -1);
+
+    // Apply GELU
+    cur = ggml_gelu_inplace(ctx0, cur);
+
+    // Branch 1: multiply with mm_h_to_4h_w
+    ggml_tensor * h_to_4h = ggml_mul_mat(ctx0, model.mm_h_to_4h_w, cur);
+
+    // Branch 2: multiply with mm_gate_w
+    ggml_tensor * gate = ggml_mul_mat(ctx0, model.mm_gate_w, cur);
+
+    // Apply silu
+    gate = ggml_swiglu_split(ctx0, gate, h_to_4h);
+
+    // Apply mm_4h_to_h_w
+    cur = ggml_mul_mat(ctx0, model.mm_4h_to_h_w, gate);
+
+    // Concatenate with boi and eoi
+    cur = ggml_concat(ctx0, model.mm_boi, cur, 1);
+    cur = ggml_concat(ctx0, cur, model.mm_eoi, 1);
+
+    // build the graph
+    ggml_build_forward_expand(gf, cur);
+
+    return gf;
+}

tools/mtmd/models/conformer.cpp

@@ -0,0 +1,217 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_conformer::build() {
+    const int n_frames   = img.nx;
+    const int n_pos      = n_frames / 2;
+    const int n_pos_embd = (((((n_frames + 1) / 2) + 1) / 2 + 1) / 2) * 2 - 1;
+    GGML_ASSERT(model.position_embeddings->ne[1] >= n_pos);
+
+    ggml_tensor * pos_emb = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, 512, n_pos_embd);
+    ggml_set_name(pos_emb, "pos_emb");
+    ggml_set_input(pos_emb);
+    ggml_build_forward_expand(gf, pos_emb);
+
+    ggml_tensor * inp = build_inp_raw(1);
+    cb(inp, "input", -1);
+
+    auto * cur = ggml_cont(ctx0, ggml_transpose(ctx0, inp));
+
+    // pre encode, conv subsampling
+    {
+        // layer.0 - conv2d
+        cur = ggml_conv_2d(ctx0, model.pre_encode_conv_X_w[0], cur, 2, 2, 1, 1, 1, 1);
+        cur = ggml_add(ctx0, cur, model.pre_encode_conv_X_b[0]);
+        cb(cur, "conformer.pre_encode.conv.{}", 0);
+
+        // layer.1 - relu
+        cur = ggml_relu_inplace(ctx0, cur);
+
+        // layer.2 conv2d dw
+        cur = ggml_conv_2d_dw_direct(ctx0, model.pre_encode_conv_X_w[2], cur, 2, 2, 1, 1, 1, 1);
+        cur = ggml_add(ctx0, cur, model.pre_encode_conv_X_b[2]);
+        cb(cur, "conformer.pre_encode.conv.{}", 2);
+
+        // layer.3 conv2d
+        cur = ggml_conv_2d_direct(ctx0, model.pre_encode_conv_X_w[3], cur, 1, 1, 0, 0, 1, 1);
+        cur = ggml_add(ctx0, cur, model.pre_encode_conv_X_b[3]);
+        cb(cur, "conformer.pre_encode.conv.{}", 3);
+
+        // layer.4 - relu
+        cur = ggml_relu_inplace(ctx0, cur);
+
+        // layer.5 conv2d dw
+        cur = ggml_conv_2d_dw_direct(ctx0, model.pre_encode_conv_X_w[5], cur, 2, 2, 1, 1, 1, 1);
+        cur = ggml_add(ctx0, cur, model.pre_encode_conv_X_b[5]);
+        cb(cur, "conformer.pre_encode.conv.{}", 5);
+
+        // layer.6 conv2d
+        cur = ggml_conv_2d_direct(ctx0, model.pre_encode_conv_X_w[6], cur, 1, 1, 0, 0, 1, 1);
+        cur = ggml_add(ctx0, cur, model.pre_encode_conv_X_b[6]);
+        cb(cur, "conformer.pre_encode.conv.{}", 6);
+
+        // layer.7 - relu
+        cur = ggml_relu_inplace(ctx0, cur);
+
+        // flatten channel and frequency axis
+        cur = ggml_cont(ctx0, ggml_permute(ctx0, cur, 0, 2, 1, 3));
+        cur = ggml_reshape_2d(ctx0, cur, cur->ne[0] * cur->ne[1], cur->ne[2]);
+
+        // calculate out
+        cur = ggml_mul_mat(ctx0, model.pre_encode_out_w, cur);
+        cur = ggml_add(ctx0, cur, model.pre_encode_out_b);
+        cb(cur, "conformer.pre_encode.out", -1);
+    }
+
+    // pos_emb
+    cb(pos_emb, "pos_emb", -1);
+
+    for (int il = 0; il < hparams.n_layer; il++) {
+        const auto & layer = model.layers[il];
+
+        auto * residual = cur;
+
+        cb(cur, "layer.in", il);
+
+        // feed_forward1
+        cur = build_norm(cur, layer.ff_norm_w, layer.ff_norm_b, NORM_TYPE_NORMAL, 1e-5, il);
+        cb(cur, "conformer.layers.{}.norm_feed_forward1", il);
+
+        cur = build_ffn(cur, layer.ff_up_w, layer.ff_up_b, nullptr, nullptr, layer.ff_down_w, layer.ff_down_b, FFN_SILU,
+                        il);
+        cb(cur, "conformer.layers.{}.feed_forward1.linear2", il);
+
+        const auto fc_factor = 0.5f;
+        residual             = ggml_add(ctx0, residual, ggml_scale(ctx0, cur, fc_factor));
+
+        // self-attention
+        {
+            cur = build_norm(residual, layer.ln_1_w, layer.ln_1_b, NORM_TYPE_NORMAL, 1e-5, il);
+            cb(cur, "conformer.layers.{}.norm_self_att", il);
+
+            ggml_tensor * Qcur     = ggml_mul_mat(ctx0, layer.q_w, cur);
+            Qcur                   = ggml_add(ctx0, Qcur, layer.q_b);
+            Qcur                   = ggml_reshape_3d(ctx0, Qcur, d_head, n_head, Qcur->ne[1]);
+            ggml_tensor * Q_bias_u = ggml_add(ctx0, Qcur, layer.pos_bias_u);
+            Q_bias_u               = ggml_permute(ctx0, Q_bias_u, 0, 2, 1, 3);
+            ggml_tensor * Q_bias_v = ggml_add(ctx0, Qcur, layer.pos_bias_v);
+            Q_bias_v               = ggml_permute(ctx0, Q_bias_v, 0, 2, 1, 3);
+
+            // TODO @ngxson : some cont can/should be removed when ggml_mul_mat support these cases
+            ggml_tensor * Kcur = ggml_mul_mat(ctx0, layer.k_w, cur);
+            Kcur               = ggml_add(ctx0, Kcur, layer.k_b);
+            Kcur               = ggml_reshape_3d(ctx0, Kcur, d_head, n_head, Kcur->ne[1]);
+            Kcur               = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3));
+
+            ggml_tensor * Vcur = ggml_mul_mat(ctx0, layer.v_w, cur);
+            Vcur               = ggml_add(ctx0, Vcur, layer.v_b);
+            Vcur               = ggml_reshape_3d(ctx0, Vcur, d_head, n_head, Vcur->ne[1]);
+            Vcur               = ggml_cont(ctx0, ggml_permute(ctx0, Vcur, 1, 2, 0, 3));
+
+            // build_attn won't fit due to matrix_ac and matrix_bd separation
+            ggml_tensor * matrix_ac = ggml_mul_mat(ctx0, Q_bias_u, Kcur);
+            matrix_ac               = ggml_cont(ctx0, ggml_permute(ctx0, matrix_ac, 1, 0, 2, 3));
+            cb(matrix_ac, "conformer.layers.{}.self_attn.id3", il);
+
+            auto * p = ggml_mul_mat(ctx0, layer.linear_pos_w, pos_emb);
+            cb(p, "conformer.layers.{}.self_attn.linear_pos", il);
+            p = ggml_reshape_3d(ctx0, p, d_head, n_head, p->ne[1]);
+            p = ggml_permute(ctx0, p, 0, 2, 1, 3);
+
+            auto * matrix_bd = ggml_mul_mat(ctx0, Q_bias_v, p);
+            matrix_bd        = ggml_cont(ctx0, ggml_permute(ctx0, matrix_bd, 1, 0, 2, 3));
+
+            // rel shift
+            {
+                const auto pos_len = matrix_bd->ne[0];
+                const auto q_len   = matrix_bd->ne[1];
+                const auto h       = matrix_bd->ne[2];
+                matrix_bd          = ggml_pad(ctx0, matrix_bd, 1, 0, 0, 0);
+                matrix_bd          = ggml_roll(ctx0, matrix_bd, 1, 0, 0, 0);
+                matrix_bd          = ggml_reshape_3d(ctx0, matrix_bd, q_len, pos_len + 1, h);
+                matrix_bd          = ggml_view_3d(ctx0, matrix_bd, q_len, pos_len, h, matrix_bd->nb[1],
+                                                        matrix_bd->nb[2], matrix_bd->nb[0] * q_len);
+                matrix_bd          = ggml_cont_3d(ctx0, matrix_bd, pos_len, q_len, h);
+            }
+
+            matrix_bd     = ggml_view_3d(ctx0, matrix_bd, matrix_ac->ne[0], matrix_bd->ne[1],
+                                               matrix_bd->ne[2], matrix_bd->nb[1], matrix_bd->nb[2], 0);
+            auto * scores = ggml_add(ctx0, matrix_ac, matrix_bd);
+            scores        = ggml_scale(ctx0, scores, 1.0f / std::sqrt(d_head));
+            cb(scores, "conformer.layers.{}.self_attn.id0", il);
+
+            ggml_tensor * attn = ggml_soft_max(ctx0, scores);
+            ggml_tensor * x    = ggml_mul_mat(ctx0, attn, Vcur);
+            x                  = ggml_permute(ctx0, x, 2, 0, 1, 3);
+            x                  = ggml_cont_2d(ctx0, x, x->ne[0] * x->ne[1], x->ne[2]);
+
+            ggml_tensor * out = ggml_mul_mat(ctx0, layer.o_w, x);
+            out               = ggml_add(ctx0, out, layer.o_b);
+            cb(out, "conformer.layers.{}.self_attn.linear_out", il);
+
+            cur = out;
+        }
+
+        residual = ggml_add(ctx0, residual, cur);
+        cur      = build_norm(residual, layer.norm_conv_w, layer.norm_conv_b, NORM_TYPE_NORMAL, 1e-5, il);
+        cb(cur, "conformer.layers.{}.norm_conv", il);
+
+        // conv
+        {
+            auto * x = cur;
+            x = ggml_mul_mat(ctx0, layer.conv_pw1_w, x);
+            x = ggml_add(ctx0, x, layer.conv_pw1_b);
+            cb(x, "conformer.layers.{}.conv.pointwise_conv1", il);
+
+            // ggml_glu doesn't support sigmoid
+            // TODO @ngxson : support this ops in ggml
+            {
+                int64_t       d    = x->ne[0] / 2;
+                ggml_tensor * gate = ggml_sigmoid(ctx0, ggml_view_2d(ctx0, x, d, x->ne[1], x->nb[1], d * x->nb[0]));
+                x                  = ggml_mul(ctx0, ggml_view_2d(ctx0, x, d, x->ne[1], x->nb[1], 0), gate);
+                x                  = ggml_cont(ctx0, ggml_transpose(ctx0, x));
+            }
+
+            // use ggml_ssm_conv for f32 precision
+            x = ggml_pad(ctx0, x, 4, 0, 0, 0);
+            x = ggml_roll(ctx0, x, 4, 0, 0, 0);
+            x = ggml_pad(ctx0, x, 4, 0, 0, 0);
+            x = ggml_ssm_conv(ctx0, x, layer.conv_dw_w);
+            x = ggml_add(ctx0, x, layer.conv_dw_b);
+
+            x = ggml_add(ctx0, ggml_mul(ctx0, x, layer.conv_norm_w), layer.conv_norm_b);
+            x = ggml_silu(ctx0, x);
+
+            // pointwise_conv2
+            x = ggml_mul_mat(ctx0, layer.conv_pw2_w, x);
+            x = ggml_add(ctx0, x, layer.conv_pw2_b);
+
+            cur = x;
+        }
+
+        residual = ggml_add(ctx0, residual, cur);
+
+        cur = build_norm(residual, layer.ff_norm_1_w, layer.ff_norm_1_b, NORM_TYPE_NORMAL, 1e-5, il);
+        cb(cur, "conformer.layers.{}.norm_feed_forward2", il);
+
+        cur = build_ffn(cur, layer.ff_up_1_w, layer.ff_up_1_b, nullptr, nullptr, layer.ff_down_1_w, layer.ff_down_1_b,
+                        FFN_SILU, il);  // TODO(tarek): read activation for ffn from hparams
+        cb(cur, "conformer.layers.{}.feed_forward2.linear2", il);
+
+        residual = ggml_add(ctx0, residual, ggml_scale(ctx0, cur, fc_factor));
+        cb(residual, "conformer.layers.{}.conv.id", il);
+
+        cur = build_norm(residual, layer.ln_2_w, layer.ln_2_b, NORM_TYPE_NORMAL, 1e-5, il);
+        cb(cur, "conformer.layers.{}.norm_out", il);
+    }
+
+    // audio adapter
+    cur = build_norm(cur, model.mm_0_w, model.mm_0_b, NORM_TYPE_NORMAL, 1e-5, -1);
+    cb(cur, "audio_adapter.model.{}", 0);
+    cur = build_ffn(cur, model.mm_1_w, model.mm_1_b, nullptr, nullptr, model.mm_3_w, model.mm_3_b, FFN_GELU_ERF, -1);
+
+    cb(cur, "projected", -1);
+
+    ggml_build_forward_expand(gf, cur);
+
+    return gf;
+}

tools/mtmd/models/glm4v.cpp

@@ -0,0 +1,120 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_glm4v::build() {
+    GGML_ASSERT(model.patch_bias != nullptr);
+    GGML_ASSERT(model.position_embeddings != nullptr);
+    GGML_ASSERT(model.class_embedding == nullptr);
+
+    const int batch_size = 1;
+
+    norm_type norm_t = NORM_TYPE_RMS;
+
+    ggml_tensor * inp_raw = build_inp_raw();
+    ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings_0, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
+
+    int mrope_sections[4] = {d_head/4, d_head/4, d_head/4, d_head/4};
+    ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches * 4);
+    ggml_set_name(positions, "positions");
+    ggml_set_input(positions);
+
+    GGML_ASSERT(img.nx % (patch_size * 2) == 0);
+    GGML_ASSERT(img.ny % (patch_size * 2) == 0);
+
+    // second conv dimension
+    {
+        auto inp_1 = ggml_conv_2d(ctx0, model.patch_embeddings_1, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
+        inp = ggml_add(ctx0, inp, inp_1);
+
+        inp = ggml_permute(ctx0, inp, 1, 2, 0, 3);  // [w, h, c, b] -> [c, w, h, b]
+        inp = ggml_cont_4d(
+            ctx0, inp,
+            n_embd * 2, n_patches_x / 2, n_patches_y, batch_size);
+        inp = ggml_reshape_4d(
+            ctx0, inp,
+            n_embd * 2, n_patches_x / 2, 2, batch_size * (n_patches_y / 2));
+        inp = ggml_permute(ctx0, inp, 0, 2, 1, 3);
+        inp = ggml_cont_3d(
+            ctx0, inp,
+            n_embd, n_patches_x * n_patches_y, batch_size);
+    }
+
+    // add patch bias
+    inp = ggml_add(ctx0, inp, model.patch_bias);
+    cb(inp, "patch_bias", -1);
+
+    // pos-conv norm
+    inp = build_norm(inp, model.norm_embd_w, model.norm_embd_b, norm_t, eps, -1);
+
+    // calculate absolute position embedding and apply
+    ggml_tensor * learned_pos_embd = resize_position_embeddings(GGML_SCALE_MODE_BICUBIC);
+    learned_pos_embd = ggml_cont_4d(
+        ctx0, learned_pos_embd,
+        n_embd * 2, n_patches_x / 2, n_patches_y, batch_size);
+    learned_pos_embd = ggml_reshape_4d(
+        ctx0, learned_pos_embd,
+        n_embd * 2, n_patches_x / 2, 2, batch_size * (n_patches_y / 2));
+    learned_pos_embd = ggml_permute(ctx0, learned_pos_embd, 0, 2, 1, 3);
+    learned_pos_embd = ggml_cont_3d(
+        ctx0, learned_pos_embd,
+        n_embd, n_patches_x * n_patches_y, batch_size);
+    cb(learned_pos_embd, "learned_pos_embd", -1);
+
+    auto add_pos = [&](ggml_tensor * cur, const clip_layer &) {
+        return ggml_rope_multi(
+                    ctx0, cur, positions, nullptr,
+                    d_head/2, mrope_sections, GGML_ROPE_TYPE_VISION,
+                    32768, hparams.rope_theta, 1, 0, 1, 32, 1);
+    };
+
+    ggml_tensor * cur = build_vit(
+                            inp, n_patches,
+                            norm_t,
+                            hparams.ffn_op,
+                            learned_pos_embd,
+                            add_pos);
+
+    cb(cur, "vit_out", -1);
+    // cb(ggml_sum(ctx0, cur), "vit_out_sum", -1);
+
+    // GLM4V projector
+    // ref: https://github.com/huggingface/transformers/blob/40dc11cd3eb4126652aa41ef8272525affd4a636/src/transformers/models/glm4v/modeling_glm4v.py#L116-L130
+
+    // patch merger (downsample)
+    {
+        int n_merge = hparams.n_merge;
+        GGML_ASSERT(n_merge > 0);
+
+        int n_token_out = n_patches / n_merge / n_merge;
+        cur = ggml_reshape_4d(ctx0, cur, n_embd, n_merge, n_merge, n_token_out);
+        cur = ggml_cont(ctx0, ggml_permute(ctx0, cur, 2, 0, 1, 3)); // [n_merge, n_merge, n_embd, n_token_out]
+        cur = ggml_conv_2d(ctx0, model.mm_patch_merger_w, cur, n_merge, n_merge, 0, 0, 1, 1);
+        cur = ggml_reshape_2d(ctx0, cur, cur->ne[2], n_token_out); // [n_embd_out, n_token_out]
+
+        cur = ggml_add(ctx0, cur, model.mm_patch_merger_b);
+    }
+
+    // FC projector
+    {
+        cur = ggml_mul_mat(ctx0, model.projection, cur);
+        // default LayerNorm (post_projection_norm)
+        cur = build_norm(cur, model.mm_post_norm_w, model.mm_post_norm_b, NORM_TYPE_NORMAL, 1e-5, -1);
+        cur = ggml_gelu_erf(ctx0, cur);
+        cb(cur, "after_fc_proj", -1);
+    }
+
+    // FFN projector
+    {
+        cur = build_ffn(cur,
+            model.mm_ffn_up_w, model.mm_ffn_up_b,
+            model.mm_ffn_gate_w, model.mm_ffn_gate_b,
+            model.mm_ffn_down_w, model.mm_ffn_down_b,
+            hparams.ffn_op, -1);
+        cb(cur, "after_ffn_proj", -1);
+        // cb(ggml_sum(ctx0, cur), "merged_sum", -1);
+    }
+
+    // build the graph
+    ggml_build_forward_expand(gf, cur);
+
+    return gf;
+}

tools/mtmd/models/internvl.cpp

@@ -0,0 +1,69 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_internvl::build() {
+    GGML_ASSERT(model.class_embedding != nullptr);
+    GGML_ASSERT(model.position_embeddings != nullptr);
+
+    const int n_pos = n_patches + 1;
+    ggml_tensor * inp = build_inp();
+
+    // add CLS token
+    inp = ggml_concat(ctx0, inp, model.class_embedding, 1);
+
+    // The larger models use a different ViT, which uses RMS norm instead of layer norm
+    // ref: https://github.com/ggml-org/llama.cpp/pull/13443#issuecomment-2869786188
+    norm_type norm_t = (hparams.n_embd == 3200 && hparams.n_layer == 45)
+        ? NORM_TYPE_RMS // 6B ViT (Used by InternVL 2.5/3 - 26B, 38B, 78B)
+        : NORM_TYPE_NORMAL; // 300M ViT (Used by all smaller InternVL models)
+
+    ggml_tensor * cur = build_vit(
+                            inp, n_pos,
+                            norm_t,
+                            hparams.ffn_op,
+                            model.position_embeddings,
+                            nullptr);
+
+    // remove CLS token
+    cur = ggml_view_2d(ctx0, cur,
+        n_embd, n_patches,
+        ggml_row_size(cur->type, n_embd), 0);
+
+    // pixel shuffle
+    {
+        const int scale_factor = model.hparams.n_merge;
+        const int bsz    = 1; // batch size, always 1 for now since we don't support batching
+        const int height = n_patches_y;
+        const int width  = n_patches_x;
+        GGML_ASSERT(scale_factor > 0);
+        cur = ggml_reshape_4d(ctx0, cur, n_embd * scale_factor, height / scale_factor, width, bsz);
+        cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
+        cur = ggml_cont_4d(ctx0, cur,
+            n_embd * scale_factor * scale_factor,
+            height / scale_factor,
+            width / scale_factor,
+            bsz);
+        cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
+        // flatten to 2D
+        cur = ggml_cont_2d(ctx0, cur,
+            n_embd * scale_factor * scale_factor,
+            cur->ne[1] * cur->ne[2]);
+    }
+
+    // projector (always using GELU activation)
+    {
+        // projector LayerNorm uses pytorch's default eps = 1e-5
+        // ref: https://huggingface.co/OpenGVLab/InternVL3-8B-Instruct/blob/a34d3e4e129a5856abfd6aa6de79776484caa14e/modeling_internvl_chat.py#L79
+        cur = build_norm(cur, model.mm_0_w, model.mm_0_b, NORM_TYPE_NORMAL, 1e-5, -1);
+        cur = build_ffn(cur,
+            model.mm_1_w, model.mm_1_b,
+            nullptr, nullptr,
+            model.mm_3_w, model.mm_3_b,
+            FFN_GELU,
+            -1);
+    }
+
+    // build the graph
+    ggml_build_forward_expand(gf, cur);
+
+    return gf;
+}

tools/mtmd/models/kimivl.cpp

@@ -0,0 +1,63 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_kimivl::build() {
+    // 2D input positions
+    ggml_tensor * pos_h = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches);
+    ggml_set_name(pos_h, "pos_h");
+    ggml_set_input(pos_h);
+
+    ggml_tensor * pos_w = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches);
+    ggml_set_name(pos_w, "pos_w");
+    ggml_set_input(pos_w);
+
+    ggml_tensor * learned_pos_embd = resize_position_embeddings();
+
+    // build ViT with 2D position embeddings
+    auto add_pos = [&](ggml_tensor * cur, const clip_layer &) {
+        // first half is X axis and second half is Y axis
+        return build_rope_2d(ctx0, cur, pos_w, pos_h, hparams.rope_theta, false);
+    };
+
+    ggml_tensor * inp = build_inp();
+    ggml_tensor * cur = build_vit(
+                            inp, n_patches,
+                            NORM_TYPE_NORMAL,
+                            hparams.ffn_op,
+                            learned_pos_embd,
+                            add_pos);
+
+    cb(cur, "vit_out", -1);
+
+    {
+        // patch_merger
+        const int scale_factor = model.hparams.n_merge;
+        cur = build_patch_merge_permute(cur, scale_factor);
+
+        // projection norm
+        int proj_inp_dim = cur->ne[0];
+        cur = ggml_view_2d(ctx0, cur,
+            n_embd, cur->ne[1] * scale_factor * scale_factor,
+            ggml_row_size(cur->type, n_embd), 0);
+        cur = ggml_norm(ctx0, cur, 1e-5); // default nn.LayerNorm
+        cur = ggml_mul(ctx0, cur, model.mm_input_norm_w);
+        cur = ggml_add(ctx0, cur, model.mm_input_norm_b);
+        cur = ggml_view_2d(ctx0, cur,
+            proj_inp_dim, cur->ne[1] / scale_factor / scale_factor,
+            ggml_row_size(cur->type, proj_inp_dim), 0);
+        cb(cur, "proj_inp_normed", -1);
+
+        // projection mlp
+        cur = build_ffn(cur,
+            model.mm_1_w, model.mm_1_b,
+            nullptr, nullptr,
+            model.mm_2_w, model.mm_2_b,
+            FFN_GELU,
+            -1);
+        cb(cur, "proj_out", -1);
+    }
+
+    // build the graph
+    ggml_build_forward_expand(gf, cur);
+
+    return gf;
+}

tools/mtmd/models/llama4.cpp

@@ -0,0 +1,96 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_llama4::build() {
+    GGML_ASSERT(model.class_embedding != nullptr);
+    GGML_ASSERT(model.position_embeddings != nullptr);
+
+    const int n_pos = n_patches + 1; // +1 for [CLS]
+
+    // 2D input positions
+    ggml_tensor * pos_h = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos);
+    ggml_set_name(pos_h, "pos_h");
+    ggml_set_input(pos_h);
+
+    ggml_tensor * pos_w = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos);
+    ggml_set_name(pos_w, "pos_w");
+    ggml_set_input(pos_w);
+
+    ggml_tensor * inp = build_inp_raw();
+
+    // Llama4UnfoldConvolution
+    {
+        ggml_tensor * kernel = ggml_reshape_4d(ctx0, model.patch_embeddings_0,
+                                                patch_size, patch_size, 3, n_embd);
+        inp = ggml_im2col(ctx0, kernel, inp, patch_size, patch_size, 0, 0, 1, 1, true, inp->type);
+        inp = ggml_mul_mat(ctx0, model.patch_embeddings_0, inp);
+        inp = ggml_reshape_2d(ctx0, inp, n_embd, n_patches);
+        cb(inp, "patch_conv", -1);
+    }
+
+    // add CLS token
+    inp = ggml_concat(ctx0, inp, model.class_embedding, 1);
+
+    // build ViT with 2D position embeddings
+    auto add_pos = [&](ggml_tensor * cur, const clip_layer &) {
+        // first half is X axis and second half is Y axis
+        // ref: https://github.com/huggingface/transformers/blob/40a493c7ed4f19f08eadb0639cf26d49bfa5e180/src/transformers/models/llama4/modeling_llama4.py#L1312
+        // ref: https://github.com/Blaizzy/mlx-vlm/blob/a57156aa87b33cca6e5ee6cfc14dd4ef8f611be6/mlx_vlm/models/llama4/vision.py#L441
+        return build_rope_2d(ctx0, cur, pos_w, pos_h, hparams.rope_theta, false);
+    };
+    ggml_tensor * cur = build_vit(
+                            inp, n_pos,
+                            NORM_TYPE_NORMAL,
+                            hparams.ffn_op,
+                            model.position_embeddings,
+                            add_pos);
+
+    // remove CLS token
+    cur = ggml_view_2d(ctx0, cur,
+        n_embd, n_patches,
+        ggml_row_size(cur->type, n_embd), 0);
+
+    // pixel shuffle
+    // based on Llama4VisionPixelShuffleMLP
+    // https://github.com/huggingface/transformers/blob/2932f318a20d9e54cc7aea052e040164d85de7d6/src/transformers/models/llama4/modeling_llama4.py#L1151
+    {
+        const int scale_factor = model.hparams.n_merge;
+        const int bsz = 1; // batch size, always 1 for now since we don't support batching
+        GGML_ASSERT(scale_factor > 0);
+        GGML_ASSERT(n_patches_x == n_patches_y); // llama4 only supports square images
+        cur = ggml_reshape_4d(ctx0, cur,
+            n_embd * scale_factor,
+            n_patches_x / scale_factor,
+            n_patches_y,
+            bsz);
+        cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
+        cur = ggml_cont_4d(ctx0, cur,
+            n_embd * scale_factor * scale_factor,
+            n_patches_x / scale_factor,
+            n_patches_y / scale_factor,
+            bsz);
+        //cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
+        // flatten to 2D
+        cur = ggml_cont_2d(ctx0, cur,
+            n_embd * scale_factor * scale_factor,
+            n_patches / scale_factor / scale_factor);
+        cb(cur, "pixel_shuffle", -1);
+    }
+
+    // based on Llama4VisionMLP2 (always uses GELU activation, no bias)
+    {
+        cur = ggml_mul_mat(ctx0, model.mm_model_mlp_1_w, cur);
+        cur = ggml_gelu(ctx0, cur);
+        cur = ggml_mul_mat(ctx0, model.mm_model_mlp_2_w, cur);
+        cur = ggml_gelu(ctx0, cur);
+        cb(cur, "adapter_mlp", -1);
+    }
+
+    // Llama4MultiModalProjector
+    cur = ggml_mul_mat(ctx0, model.mm_model_proj, cur);
+    cb(cur, "projected", -1);
+
+    // build the graph
+    ggml_build_forward_expand(gf, cur);
+
+    return gf;
+}

tools/mtmd/models/llava.cpp

@@ -0,0 +1,374 @@
+#include "models.h"
+
+// this graph is used by llava, granite and glm
+// due to having embedding_stack (used by granite), we cannot reuse build_vit
+ggml_cgraph * clip_graph_llava::build() {
+    const int batch_size = 1;
+    const int n_pos = n_patches + (model.class_embedding ? 1 : 0);
+
+    GGML_ASSERT(n_patches_x == n_patches_y && "only square images supported");
+
+    // Calculate the deepest feature layer based on hparams and projector type
+    int max_feature_layer = n_layer;
+    {
+        // Get the index of the second to last layer; this is the default for models that have a llava projector
+        int il_last = hparams.n_layer - 1;
+        int deepest_feature_layer = -1;
+
+        if (proj_type == PROJECTOR_TYPE_MINICPMV || proj_type == PROJECTOR_TYPE_GLM_EDGE) {
+            il_last += 1;
+        }
+
+        // If we set explicit vision feature layers, only go up to the deepest one
+        // NOTE: only used by granite-vision models for now
+        for (const auto & feature_layer : hparams.vision_feature_layer) {
+            if (feature_layer > deepest_feature_layer) {
+                deepest_feature_layer = feature_layer;
+            }
+        }
+        max_feature_layer = deepest_feature_layer < 0 ? il_last : deepest_feature_layer;
+    }
+
+    ggml_tensor * inp = build_inp();
+
+    // concat class_embeddings and patch_embeddings
+    if (model.class_embedding) {
+        inp = ggml_concat(ctx0, inp, model.class_embedding, 1);
+    }
+
+    ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos);
+    ggml_set_name(positions, "positions");
+    ggml_set_input(positions);
+
+    inp = ggml_add(ctx0, inp, ggml_get_rows(ctx0, model.position_embeddings, positions));
+
+    ggml_tensor * inpL = inp;
+
+    // pre-layernorm
+    if (model.pre_ln_w) {
+        inpL = build_norm(inpL, model.pre_ln_w, model.pre_ln_b, NORM_TYPE_NORMAL, eps, -1);
+        cb(inpL, "pre_ln", -1);
+    }
+
+    std::vector<ggml_tensor *> embedding_stack;
+    const auto & vision_feature_layer = hparams.vision_feature_layer;
+
+    // loop over layers
+    for (int il = 0; il < max_feature_layer; il++) {
+        auto & layer = model.layers[il];
+        ggml_tensor * cur = inpL; // inpL = residual, cur = hidden_states
+
+        // If this is an embedding feature layer, save the output.
+        // NOTE: 0 index here refers to the input to the encoder.
+        if (vision_feature_layer.find(il) != vision_feature_layer.end()) {
+            embedding_stack.push_back(cur);
+        }
+
+        // layernorm1
+        cur = build_norm(cur, layer.ln_1_w, layer.ln_1_b, NORM_TYPE_NORMAL, eps, il);
+        cb(cur, "layer_inp_normed", il);
+
+        // self-attention
+        {
+            ggml_tensor * Qcur = ggml_mul_mat(ctx0, layer.q_w, cur);
+            if (layer.q_b) {
+                Qcur = ggml_add(ctx0, Qcur, layer.q_b);
+            }
+
+            ggml_tensor * Kcur = ggml_mul_mat(ctx0, layer.k_w, cur);
+            if (layer.k_b) {
+                Kcur = ggml_add(ctx0, Kcur, layer.k_b);
+            }
+
+            ggml_tensor * Vcur = ggml_mul_mat(ctx0, layer.v_w, cur);
+            if (layer.v_b) {
+                Vcur = ggml_add(ctx0, Vcur, layer.v_b);
+            }
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, d_head, n_head, n_pos);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, d_head, n_head, n_pos);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, d_head, n_head, n_pos);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(layer.o_w, layer.o_b,
+                Qcur, Kcur, Vcur, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+        }
+
+        // re-add the layer input, e.g., residual
+        cur = ggml_add(ctx0, cur, inpL);
+
+        inpL = cur; // inpL = residual, cur = hidden_states
+
+        cb(cur, "ffn_inp", il);
+
+        // layernorm2
+        cur = build_norm(cur, layer.ln_2_w, layer.ln_2_b, NORM_TYPE_NORMAL, eps, il);
+        cb(cur, "ffn_inp_normed", il);
+
+        // ffn
+        cur = build_ffn(cur,
+            layer.ff_up_w, layer.ff_up_b,
+            layer.ff_gate_w, layer.ff_gate_b,
+            layer.ff_down_w, layer.ff_down_b,
+            hparams.ffn_op, il);
+
+        cb(cur, "ffn_out", il);
+
+        // residual 2
+        cur = ggml_add(ctx0, inpL, cur);
+        cb(cur, "layer_out", il);
+
+        inpL = cur;
+    }
+
+    // post-layernorm
+    if (model.post_ln_w) {
+        inpL = build_norm(inpL, model.post_ln_w, model.post_ln_b, NORM_TYPE_NORMAL, eps, -1);
+    }
+
+    ggml_tensor * embeddings = inpL;
+
+    // process vision feature layers (used by granite)
+    {
+        // final layer is a vision feature layer
+        if (vision_feature_layer.find(max_feature_layer) != vision_feature_layer.end()) {
+            embedding_stack.push_back(inpL);
+        }
+
+        // If feature layers are explicitly set, stack them (if we have multiple)
+        if (!embedding_stack.empty()) {
+            embeddings = embedding_stack[0];
+            for (size_t i = 1; i < embedding_stack.size(); i++) {
+                embeddings = ggml_concat(ctx0, embeddings, embedding_stack[i], 0);
+            }
+        }
+    }
+
+    // llava projector (also used by granite)
+    if (hparams.has_llava_projector) {
+        embeddings = ggml_reshape_2d(ctx0, embeddings, embeddings->ne[0], embeddings->ne[1]);
+
+        ggml_tensor * patches = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches);
+        ggml_set_name(patches, "patches");
+        ggml_set_input(patches);
+
+        // shape [1, 576, 1024]
+        // ne is whcn, ne = [1024, 576, 1, 1]
+        embeddings = ggml_get_rows(ctx0, embeddings, patches);
+
+        // print_tensor_info(embeddings, "embeddings");
+
+        // llava projector
+        if (proj_type == PROJECTOR_TYPE_MLP) {
+            embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings);
+            embeddings = ggml_add(ctx0, embeddings, model.mm_0_b);
+
+            embeddings = ggml_gelu(ctx0, embeddings);
+            if (model.mm_2_w) {
+                embeddings = ggml_mul_mat(ctx0, model.mm_2_w, embeddings);
+                embeddings = ggml_add(ctx0, embeddings, model.mm_2_b);
+            }
+        }
+        else if (proj_type == PROJECTOR_TYPE_MLP_NORM) {
+            embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings);
+            embeddings = ggml_add(ctx0, embeddings, model.mm_0_b);
+            // ggml_tensor_printf(embeddings, "mm_0_w",0,true,false);
+            // First LayerNorm
+            embeddings = ggml_norm(ctx0, embeddings, eps);
+            embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_1_w),
+                                model.mm_1_b);
+
+            // GELU activation
+            embeddings = ggml_gelu(ctx0, embeddings);
+
+            // Second linear layer
+            embeddings = ggml_mul_mat(ctx0, model.mm_3_w, embeddings);
+            embeddings = ggml_add(ctx0, embeddings, model.mm_3_b);
+
+            // Second LayerNorm
+            embeddings = ggml_norm(ctx0, embeddings, eps);
+            embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_4_w),
+                                model.mm_4_b);
+        }
+        else if (proj_type == PROJECTOR_TYPE_LDP) {
+            // MobileVLM projector
+            int n_patch = 24;
+            ggml_tensor * mlp_1 = ggml_mul_mat(ctx0, model.mm_model_mlp_1_w, embeddings);
+            mlp_1 = ggml_add(ctx0, mlp_1, model.mm_model_mlp_1_b);
+            mlp_1 = ggml_gelu(ctx0, mlp_1);
+            ggml_tensor * mlp_3 = ggml_mul_mat(ctx0, model.mm_model_mlp_3_w, mlp_1);
+            mlp_3 = ggml_add(ctx0, mlp_3, model.mm_model_mlp_3_b);
+            // mlp_3 shape = [1, 576, 2048], ne = [2048, 576, 1, 1]
+
+            // block 1
+            ggml_tensor * block_1 = nullptr;
+            {
+                // transpose from [1, 576, 2048] --> [1, 2048, 576] --> [1, 2048, 24, 24]
+                mlp_3 = ggml_permute(ctx0, mlp_3, 1, 0, 2, 3);
+                mlp_3 = ggml_cont_4d(ctx0, mlp_3, n_patch, n_patch, mlp_3->ne[1], mlp_3->ne[2]);
+                // stride = 1, padding = 1, bias is nullptr
+                block_1 = ggml_conv_2d_dw(ctx0, model.mm_model_block_1_block_0_0_w, mlp_3, 1, 1, 1, 1, 1, 1);
+
+                // layer norm
+                // // block_1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1]
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 2, 0, 3));
+                // block_1 shape = [1, 24, 24, 2048], ne = [2048, 24, 24, 1]
+                block_1 = ggml_norm(ctx0, block_1, eps);
+                block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_1_block_0_1_w), model.mm_model_block_1_block_0_1_b);
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 2, 0, 1, 3));
+
+                // block_1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1]
+                // hardswish
+                ggml_tensor * block_1_hw = ggml_hardswish(ctx0, block_1);
+
+                block_1 = ggml_pool_2d(ctx0, block_1_hw, GGML_OP_POOL_AVG, block_1_hw->ne[0], block_1_hw->ne[1], block_1_hw->ne[0], block_1_hw->ne[1], 0, 0);
+                // block_1 shape = [1, 2048, 1, 1], ne = [1, 1, 2048, 1]
+                // pointwise conv
+                block_1 = ggml_reshape_2d(ctx0, block_1, block_1->ne[0]*block_1->ne[1]*block_1->ne[2], block_1->ne[3]);
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_1_block_1_fc1_w, block_1);
+                block_1 = ggml_add(ctx0, block_1, model.mm_model_block_1_block_1_fc1_b);
+                block_1 = ggml_relu(ctx0, block_1);
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_1_block_1_fc2_w, block_1);
+                block_1 = ggml_add(ctx0, block_1, model.mm_model_block_1_block_1_fc2_b);
+                block_1 = ggml_hardsigmoid(ctx0, block_1);
+                // block_1_hw shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1], block_1 shape = [1, 2048], ne = [2048, 1, 1, 1]
+                block_1 = ggml_reshape_4d(ctx0, block_1, 1, 1, block_1->ne[0], block_1->ne[1]);
+                block_1 = ggml_mul(ctx0, block_1_hw, block_1);
+
+                int w = block_1->ne[0], h = block_1->ne[1];
+                block_1 = ggml_reshape_3d(ctx0, block_1, w*h, block_1->ne[2], block_1->ne[3]);
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 0, 2, 3));
+
+                // block_1 shape = [1, 24*24, 2048], ne = [24*24, 2048, 1]
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_1_block_2_0_w, block_1);
+                block_1 = ggml_reshape_4d(ctx0, block_1, block_1->ne[0], w, h, block_1->ne[3]);
+
+                // block_1 shape = [1, 24, 24, 2048], ne = [2048, 24, 24, 1]
+                block_1 = ggml_norm(ctx0, block_1, eps);
+                block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_1_block_2_1_w), model.mm_model_block_1_block_2_1_b);
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 2, 0, 1, 3));
+                // block1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1]
+                // residual
+                block_1 = ggml_add(ctx0, mlp_3, block_1);
+            }
+
+            // block_2
+            {
+                // stride = 2
+                block_1 = ggml_conv_2d_dw(ctx0, model.mm_model_block_2_block_0_0_w, block_1, 2, 2, 1, 1, 1, 1);
+
+                // block_1 shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1]
+                // layer norm
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 2, 0, 3));
+                // block_1 shape = [1, 12, 12, 2048], ne = [2048, 12, 12, 1]
+                block_1 = ggml_norm(ctx0, block_1, eps);
+                block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_2_block_0_1_w), model.mm_model_block_2_block_0_1_b);
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 2, 0, 1, 3));
+                // block_1 shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1]
+                // hardswish
+                ggml_tensor * block_1_hw = ggml_hardswish(ctx0, block_1);
+
+                // not sure the parameters is right for globalAvgPooling
+                block_1 = ggml_pool_2d(ctx0, block_1_hw, GGML_OP_POOL_AVG, block_1_hw->ne[0], block_1_hw->ne[1], block_1_hw->ne[0], block_1_hw->ne[1], 0, 0);
+                // block_1 shape = [1, 2048, 1, 1], ne = [1, 1, 2048, 1]
+                // pointwise conv
+                block_1 = ggml_reshape_2d(ctx0, block_1, block_1->ne[0]*block_1->ne[1]*block_1->ne[2], block_1->ne[3]);
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_2_block_1_fc1_w, block_1);
+                block_1 = ggml_add(ctx0, block_1, model.mm_model_block_2_block_1_fc1_b);
+                block_1 = ggml_relu(ctx0, block_1);
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_2_block_1_fc2_w, block_1);
+                block_1 = ggml_add(ctx0, block_1, model.mm_model_block_2_block_1_fc2_b);
+                block_1 = ggml_hardsigmoid(ctx0, block_1);
+
+                // block_1_hw shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1], block_1 shape = [1, 2048, 1, 1], ne = [1, 1, 2048, 1]
+                block_1 = ggml_reshape_4d(ctx0, block_1, 1, 1, block_1->ne[0], block_1->ne[1]);
+                block_1 = ggml_mul(ctx0, block_1_hw, block_1);
+
+                int w = block_1->ne[0], h = block_1->ne[1];
+                block_1 = ggml_reshape_3d(ctx0, block_1, w*h, block_1->ne[2], block_1->ne[3]);
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 0, 2, 3));
+                // block_1 shape = [1, 24*24, 2048], ne = [24*24, 2048, 1]
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_2_block_2_0_w, block_1);
+                block_1 = ggml_reshape_4d(ctx0, block_1, block_1->ne[0], w, h, block_1->ne[3]);
+
+
+                // block_1 shape = [1, 12, 12, 2048], ne = [2048, 12, 12, 1]
+                block_1 = ggml_norm(ctx0, block_1, eps);
+                block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_2_block_2_1_w), model.mm_model_block_2_block_2_1_b);
+                block_1 = ggml_reshape_3d(ctx0, block_1, block_1->ne[0], block_1->ne[1] * block_1->ne[2], block_1->ne[3]);
+                // block_1 shape = [1, 144, 2048], ne = [2048, 144, 1]
+            }
+            embeddings = block_1;
+        }
+        else if (proj_type == PROJECTOR_TYPE_LDPV2)
+        {
+            int n_patch = 24;
+            ggml_tensor * mlp_0 = ggml_mul_mat(ctx0, model.mm_model_mlp_0_w, embeddings);
+            mlp_0 = ggml_add(ctx0, mlp_0, model.mm_model_mlp_0_b);
+            mlp_0 = ggml_gelu(ctx0, mlp_0);
+            ggml_tensor * mlp_2 = ggml_mul_mat(ctx0, model.mm_model_mlp_2_w, mlp_0);
+            mlp_2 = ggml_add(ctx0, mlp_2, model.mm_model_mlp_2_b);
+            // mlp_2 ne = [2048, 576, 1, 1]
+            // // AVG Pool Layer 2*2, strides = 2
+            mlp_2 = ggml_permute(ctx0, mlp_2, 1, 0, 2, 3);
+            // mlp_2 ne = [576, 2048, 1, 1]
+            mlp_2 = ggml_cont_4d(ctx0, mlp_2, n_patch, n_patch, mlp_2->ne[1], mlp_2->ne[2]);
+            // mlp_2 ne [24, 24, 2048, 1]
+            mlp_2 = ggml_pool_2d(ctx0, mlp_2, GGML_OP_POOL_AVG, 2, 2, 2, 2, 0, 0);
+            // weight ne = [3, 3, 2048, 1]
+            ggml_tensor * peg_0 = ggml_conv_2d_dw(ctx0, model.mm_model_peg_0_w, mlp_2, 1, 1, 1, 1, 1, 1);
+            peg_0 = ggml_cont(ctx0, ggml_permute(ctx0, peg_0, 1, 2, 0, 3));
+            peg_0 = ggml_add(ctx0, peg_0, model.mm_model_peg_0_b);
+            mlp_2 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_2, 1, 2, 0, 3));
+            peg_0 = ggml_add(ctx0, peg_0, mlp_2);
+            peg_0 = ggml_reshape_3d(ctx0, peg_0, peg_0->ne[0], peg_0->ne[1] * peg_0->ne[2], peg_0->ne[3]);
+            embeddings = peg_0;
+        }
+        else {
+            GGML_ABORT("fatal error");
+        }
+    }
+
+    // glm projector
+    else if (proj_type == PROJECTOR_TYPE_GLM_EDGE) {
+        size_t gridsz = (size_t)sqrt(embeddings->ne[1]);
+        embeddings = ggml_permute(ctx0,embeddings,1,0,2,3);
+        embeddings = ggml_cont_3d(ctx0, embeddings, gridsz, gridsz, embeddings->ne[1]);
+        embeddings = ggml_conv_2d(ctx0, model.mm_model_adapter_conv_w, embeddings, 2, 2, 0, 0, 1, 1);
+        embeddings = ggml_reshape_3d(ctx0, embeddings,embeddings->ne[0]*embeddings->ne[1] , embeddings->ne[2], batch_size);
+        embeddings = ggml_cont(ctx0, ggml_permute(ctx0,embeddings, 1, 0, 2, 3));
+        embeddings = ggml_add(ctx0, embeddings, model.mm_model_adapter_conv_b);
+        // GLU
+        {
+            embeddings = ggml_mul_mat(ctx0, model.mm_model_mlp_0_w, embeddings);
+            embeddings = ggml_norm(ctx0, embeddings, eps);
+            embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_model_ln_q_w), model.mm_model_ln_q_b);
+            embeddings = ggml_gelu_inplace(ctx0, embeddings);
+            ggml_tensor * x = embeddings;
+            embeddings = ggml_mul_mat(ctx0, model.mm_model_mlp_2_w, embeddings);
+            x = ggml_mul_mat(ctx0, model.mm_model_mlp_1_w,x);
+            embeddings = ggml_swiglu_split(ctx0, embeddings, x);
+            embeddings = ggml_mul_mat(ctx0, model.mm_model_mlp_3_w, embeddings);
+        }
+        // arrangement of BOI/EOI token embeddings
+        // note: these embeddings are not present in text model, hence we cannot process them as text tokens
+        // see: https://huggingface.co/THUDM/glm-edge-v-2b/blob/main/siglip.py#L53
+        {
+            embeddings = ggml_concat(ctx0, model.mm_boi, embeddings, 1); // BOI
+            embeddings = ggml_concat(ctx0, embeddings, model.mm_eoi, 1); // EOI
+        }
+    }
+
+    else {
+        GGML_ABORT("llava: unknown projector type");
+    }
+
+    // build the graph
+    ggml_build_forward_expand(gf, embeddings);
+
+    return gf;
+}

tools/mtmd/models/minicpmv.cpp

@@ -0,0 +1,114 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_minicpmv::build() {
+    GGML_ASSERT(model.class_embedding == nullptr);
+    const int n_pos       = n_patches;
+    const int n_embd_proj = n_mmproj_embd;
+
+    // position embeddings for the projector (not for ViT)
+    // see: https://huggingface.co/openbmb/MiniCPM-o-2_6/blob/main/resampler.py#L70
+    // base frequency omega
+    ggml_tensor * omega = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, n_embd_proj / 4);
+    ggml_set_name(omega, "omega");
+    ggml_set_input(omega);
+
+    // 2D input positions (using float for sinusoidal embeddings)
+    ggml_tensor * pos_h = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, 1, n_pos);
+    ggml_set_name(pos_h, "pos_h");
+    ggml_set_input(pos_h);
+    ggml_tensor * pos_w = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, 1, n_pos);
+    ggml_set_name(pos_w, "pos_w");
+    ggml_set_input(pos_w);
+
+    // for selecting learned pos embd, used by ViT
+    struct ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos);
+    ggml_set_name(positions, "positions");
+    ggml_set_input(positions);
+
+    ggml_tensor * learned_pos_embd = ggml_get_rows(ctx0, model.position_embeddings, positions);
+
+    ggml_tensor * inp = build_inp();
+    ggml_tensor * embeddings = build_vit(
+                            inp, n_pos,
+                            NORM_TYPE_NORMAL,
+                            hparams.ffn_op,
+                            learned_pos_embd,
+                            nullptr);
+
+    // resampler projector (it is just another transformer)
+
+    ggml_tensor * q = model.mm_model_query;
+    ggml_tensor * v = ggml_mul_mat(ctx0, model.mm_model_kv_proj, embeddings);
+
+    // norm
+    q = build_norm(q, model.mm_model_ln_q_w,  model.mm_model_ln_q_b,  NORM_TYPE_NORMAL, eps, -1);
+    v = build_norm(v, model.mm_model_ln_kv_w, model.mm_model_ln_kv_b, NORM_TYPE_NORMAL, eps, -1);
+
+    // calculate sinusoidal pos embd
+    ggml_tensor * pos_embed = nullptr;
+    {
+        // outer product
+        ggml_tensor * omega_b = ggml_repeat_4d(ctx0, omega, omega->ne[0], n_pos, 1, 1); // n_pos rows
+        ggml_tensor * theta_x = ggml_mul(ctx0, omega_b, pos_w);
+        ggml_tensor * theta_y = ggml_mul(ctx0, omega_b, pos_h);
+        // sin and cos
+        ggml_tensor * pos_embd_x = ggml_concat(
+            ctx0,
+            ggml_sin(ctx0, theta_x),
+            ggml_cos(ctx0, theta_x),
+            0 // concat on first dim
+        );
+        ggml_tensor * pos_embd_y = ggml_concat(
+            ctx0,
+            ggml_sin(ctx0, theta_y),
+            ggml_cos(ctx0, theta_y),
+            0 // concat on first dim
+        );
+        pos_embed = ggml_concat(ctx0, pos_embd_x, pos_embd_y, 0);
+    }
+
+    // k = v + pos_embed
+    ggml_tensor * k = ggml_add(ctx0, v, pos_embed);
+
+    // attention
+    {
+        const int d_head = 128;
+        int n_head = n_embd_proj/d_head;
+        // Use actual config value if available, otherwise fall back to hardcoded values
+        int num_query = hparams.minicpmv_query_num;
+        ggml_tensor * Q = ggml_add(ctx0,
+            ggml_mul_mat(ctx0, model.mm_model_attn_q_w, q),
+            model.mm_model_attn_q_b);
+        ggml_tensor * K = ggml_add(ctx0,
+            ggml_mul_mat(ctx0, model.mm_model_attn_k_w, k),
+            model.mm_model_attn_k_b);
+        ggml_tensor * V = ggml_add(ctx0,
+            ggml_mul_mat(ctx0, model.mm_model_attn_v_w, v),
+            model.mm_model_attn_v_b);
+
+        Q = ggml_reshape_3d(ctx0, Q, d_head, n_head, num_query);
+        K = ggml_reshape_3d(ctx0, K, d_head, n_head, n_pos);
+        V = ggml_reshape_3d(ctx0, V, d_head, n_head, n_pos);
+
+        cb(Q, "resampler_Q", -1);
+        cb(K, "resampler_K", -1);
+        cb(V, "resampler_V", -1);
+
+        float resampler_kq_scale = 1.0f/ sqrtf(float(d_head));
+        embeddings = build_attn(
+            model.mm_model_attn_o_w,
+            model.mm_model_attn_o_b,
+            Q, K, V, nullptr, resampler_kq_scale, -1);
+        cb(embeddings, "resampler_attn_out", -1);
+    }
+    // layernorm
+    embeddings = build_norm(embeddings, model.mm_model_ln_post_w, model.mm_model_ln_post_b, NORM_TYPE_NORMAL, eps, -1);
+
+    // projection
+    embeddings = ggml_mul_mat(ctx0, model.mm_model_proj, embeddings);
+
+    // build the graph
+    ggml_build_forward_expand(gf, embeddings);
+
+    return gf;
+}

tools/mtmd/models/mobilenetv5.cpp

@@ -0,0 +1,451 @@
+#include "models.h"
+
+// Helpers for MobileNetV5 Blocks
+// RMS Norm 2D - normalizes over channels for each spatial position
+ggml_tensor * clip_graph_mobilenetv5::rms_norm_2d(ggml_tensor * inp, ggml_tensor * weight, float eps) {
+    // inp: [W, H, C, B]
+
+    ggml_tensor * cur = ggml_permute(ctx0, inp, 2, 1, 0, 3);
+    cur = ggml_cont(ctx0, cur);
+    cur = ggml_rms_norm(ctx0, cur, eps);
+
+    if (weight) {
+        cur = ggml_mul(ctx0, cur, weight);
+    }
+
+    cur = ggml_permute(ctx0, cur, 2, 1, 0, 3);
+    cur = ggml_cont(ctx0, cur);
+
+    return cur;
+}
+
+// Conv2dSame padding - asymmetric SAME padding like PyTorch/TF
+ggml_tensor* clip_graph_mobilenetv5::pad_same_2d(ggml_tensor* inp, int kernel_h, int kernel_w, int stride_h, int stride_w, int dilation_h, int dilation_w) {
+    const int64_t ih = inp->ne[1];  // height
+    const int64_t iw = inp->ne[0];  // width
+
+    // Calculate output size (ceil division)
+    const int64_t oh = (ih + stride_h - 1) / stride_h;
+    const int64_t ow = (iw + stride_w - 1) / stride_w;
+
+    // Calculate padding needed
+    const int64_t pad_h = std::max((int64_t)0, (oh - 1) * stride_h + (kernel_h - 1) * dilation_h + 1 - ih);
+    const int64_t pad_w = std::max((int64_t)0, (ow - 1) * stride_w + (kernel_w - 1) * dilation_w + 1 - iw);
+
+    // Split padding asymmetrically
+    const int pad_h_top = pad_h / 2;
+    const int pad_h_bottom = pad_h - pad_h_top;
+    const int pad_w_left = pad_w / 2;
+    const int pad_w_right = pad_w - pad_w_left;
+
+    // Apply padding if needed
+    // ggml_pad_ext: (ctx, tensor, lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3)
+    // For [W, H, C, B]: p0=width, p1=height, p2=channels, p3=batch
+    if (pad_h > 0 || pad_w > 0) {
+        inp = ggml_pad_ext(ctx0, inp,
+            pad_w_left, pad_w_right,     // width padding (dim 0)
+            pad_h_top, pad_h_bottom,      // height padding (dim 1)
+            0, 0,                         // no channel padding (dim 2)
+            0, 0);                        // no batch padding (dim 3)
+    }
+
+    return inp;
+}
+
+
+// Edge Residual Block (Stage 0)
+ggml_tensor * clip_graph_mobilenetv5::build_edge_residual(ggml_tensor * inp, const mobilenetv5_block & block, int stride) {
+    ggml_tensor * cur = inp;
+
+    // 1. Expansion Conv (3x3)
+    if (stride == 2) {
+        // Case: Downsampling (Block 0)
+        // Replicates Conv2dSame(kernel=3, stride=2)
+        cur = pad_same_2d(cur, 3, 3, stride, stride);
+        cur = ggml_conv_2d_direct(ctx0, block.s0_conv_exp_w, cur, stride, stride, 0, 0, 1, 1);
+    } else {
+        // Case: Normal 3x3 Block (Block 1, 2)
+        // Replicates Conv2d(kernel=3, stride=1, padding=1)
+        cur = ggml_conv_2d_direct(ctx0, block.s0_conv_exp_w, cur, stride, stride, 1, 1, 1, 1);
+    }
+
+    // BN + Activation
+    if (block.s0_bn1_w) cur = rms_norm_2d(cur, block.s0_bn1_w);
+    cur = ggml_gelu(ctx0, cur);
+
+    // 2. Pointwise Linear Conv (1x1)
+    // 1x1 Convs usually have padding=0 and stride=1
+    cur = ggml_conv_2d_direct(ctx0, block.s0_conv_pwl_w, cur, 1, 1, 0, 0, 1, 1);
+    if (block.s0_bn2_w) cur = rms_norm_2d(cur, block.s0_bn2_w);
+
+    // 3. Residual Connection
+    // Only apply residual if spatial dimensions and channels match (stride 1)
+    if (stride == 1 && inp->ne[2] == cur->ne[2] && inp->ne[0] == cur->ne[0]) {
+        cur = ggml_add(ctx0, cur, inp);
+    }
+
+    return cur;
+}
+
+// Universal Inverted Residual Block (Stage 1+)
+ggml_tensor * clip_graph_mobilenetv5::build_inverted_residual(ggml_tensor * inp, const mobilenetv5_block & block, int stride) {
+    ggml_tensor * cur = inp;
+
+    // 1. Depthwise Start (Optional)
+    // NOTE: dw_start always has stride=1 (no downsampling here)
+    if (block.dw_start_w) {
+        int k = block.dw_start_w->ne[0]; // 3 or 5
+        int p = k / 2;
+        cur = ggml_conv_2d_dw(ctx0, block.dw_start_w, cur, 1, 1, p, p, 1, 1);
+        if (block.dw_start_bn_w) cur = rms_norm_2d(cur, block.dw_start_bn_w);
+    }
+
+    // 2. Pointwise Expansion (1x1)
+    if (block.pw_exp_w) {
+        // Standard 1x1 conv, pad=0, stride=1
+        cur = ggml_conv_2d_direct(ctx0, block.pw_exp_w, cur, 1, 1, 0, 0, 1, 1);
+        if (block.pw_exp_bn_w) cur = rms_norm_2d(cur, block.pw_exp_bn_w);
+        cur = ggml_gelu(ctx0, cur);
+    }
+
+    // 3. Depthwise Mid (Optional)
+    // NOTE: dw_mid is where downsampling happens (stride=2 for first block of stage)
+    if (block.dw_mid_w) {
+        int k = block.dw_mid_w->ne[0]; // 3 or 5
+
+        if (stride > 1) {
+            // Case: Stride 2 (Downsample) -> Use Asymmetric "Same" Padding
+            cur = pad_same_2d(cur, k, k, stride, stride);
+            cur = ggml_conv_2d_dw(ctx0, block.dw_mid_w, cur, stride, stride, 0, 0, 1, 1); // pad=0
+        } else {
+            // Case: Stride 1 -> Use Standard Symmetric Padding
+            int p = k / 2;
+            cur = ggml_conv_2d_dw(ctx0, block.dw_mid_w, cur, stride, stride, p, p, 1, 1);
+        }
+
+        if (block.dw_mid_bn_w) cur = rms_norm_2d(cur, block.dw_mid_bn_w);
+        cur = ggml_gelu(ctx0, cur);
+    }
+
+    // 4. Pointwise Projection (1x1)
+    if (block.pw_proj_w) {
+        cur = ggml_conv_2d_direct(ctx0, block.pw_proj_w, cur, 1, 1, 0, 0, 1, 1);
+        if (block.pw_proj_bn_w) cur = rms_norm_2d(cur, block.pw_proj_bn_w);
+    }
+
+    // Apply Layer Scaling if present
+    if (block.layer_scale_w) {
+        cur = ggml_mul(ctx0, cur, block.layer_scale_w);
+    }
+
+    // 5. Residual Connection
+    bool same_spatial = (inp->ne[0] == cur->ne[0]) && (inp->ne[1] == cur->ne[1]);
+    bool same_channel = (inp->ne[2] == cur->ne[2]);
+    if (same_spatial && same_channel) {
+        cur = ggml_add(ctx0, cur, inp);
+    }
+
+    return cur;
+}
+
+// Attention Block (MQA)
+ggml_tensor * clip_graph_mobilenetv5::build_mobilenet_attn(ggml_tensor * inp, const mobilenetv5_block & block) {
+    ggml_tensor * cur = inp;
+
+    // Norm
+    if (block.attn_norm_w) {
+        cur = rms_norm_2d(cur, block.attn_norm_w, 1e-6f);
+    }
+
+    // 1. Q Calculation
+    ggml_tensor * q = ggml_conv_2d_direct(ctx0, block.attn_q_w, cur, 1, 1, 0, 0, 1, 1);
+
+    // 2. K Calculation (Downsampled)
+    // Uses Conv2dSame(640, 640, kernel_size=(3, 3), stride=(2, 2), groups=640)
+    ggml_tensor * k_inp = cur;
+    if (block.attn_k_dw_w) {
+        int k_size = block.attn_k_dw_w->ne[0];  // Usually 3
+        k_inp = pad_same_2d(cur, k_size, k_size, 2, 2);  // Apply SAME padding
+        k_inp = ggml_conv_2d_dw(ctx0, block.attn_k_dw_w, k_inp, 2, 2, 0, 0, 1, 1);  // padding=0
+        if (block.attn_k_norm_w) {
+            k_inp = rms_norm_2d(k_inp, block.attn_k_norm_w, 1e-6f);
+        }
+    }
+    ggml_tensor * k = ggml_conv_2d_direct(ctx0, block.attn_k_w, k_inp, 1, 1, 0, 0, 1, 1);
+
+    // 3. V Calculation (Downsampled)
+    // Uses Conv2dSame(640, 640, kernel_size=(3, 3), stride=(2, 2), groups=640)
+    ggml_tensor * v_inp = cur;
+    if (block.attn_v_dw_w) {
+        int v_size = block.attn_v_dw_w->ne[0];  // Usually 3
+        v_inp = pad_same_2d(cur, v_size, v_size, 2, 2);  // Apply SAME padding
+        v_inp = ggml_conv_2d_dw(ctx0, block.attn_v_dw_w, v_inp, 2, 2, 0, 0, 1, 1);  // padding=0
+        if (block.attn_v_norm_w) {
+            v_inp = rms_norm_2d(v_inp, block.attn_v_norm_w, 1e-6f);
+        }
+    }
+    ggml_tensor * v = ggml_conv_2d_direct(ctx0, block.attn_v_w, v_inp, 1, 1, 0, 0, 1, 1);
+
+    const int W = cur->ne[0]; const int H = cur->ne[1]; const int B = cur->ne[3];
+    const int D = k->ne[2]; // Head dimension
+    const int n_head = q->ne[2] / D;
+    const int N = W * H;
+
+    // Process Q: [W, H, D*n_head, B] -> [D, N, n_head, B]
+    q = ggml_reshape_3d(ctx0, q, N, D*n_head, B);
+    q = ggml_reshape_4d(ctx0, q, N, D, n_head, B);
+    q = ggml_permute(ctx0, q, 1, 0, 2, 3); // [D, N, n_head, B]
+    q = ggml_cont(ctx0, q);
+
+    const int Wk = k->ne[0]; const int Hk = k->ne[1];
+    const int M = Wk * Hk;
+
+    // Process K: [Wk, Hk, D, B] -> [D, M, 1, B]
+    k = ggml_reshape_3d(ctx0, k, M, D, B);
+    k = ggml_reshape_4d(ctx0, k, M, D, 1, B);
+    k = ggml_permute(ctx0, k, 1, 0, 2, 3); // [D, M, 1, B]
+    k = ggml_cont(ctx0, k);
+
+    // Process V: [Wk, Hk, D, B] -> [M, D, 1, B]
+    v = ggml_reshape_3d(ctx0, v, M, D, B);
+    v = ggml_reshape_4d(ctx0, v, M, D, 1, B);
+    v = ggml_cont(ctx0, v); // [M, D, 1, B]
+
+    // Multi-Query Attention
+    float scale = 1.0f / sqrtf((float)D);
+
+    // Step 1: Compute Q @ K.T
+    ggml_tensor * scores = ggml_mul_mat(ctx0, k, q);
+
+    scores = ggml_scale(ctx0, scores, scale);
+
+    scores = ggml_soft_max(ctx0, scores);
+
+    ggml_tensor * kqv = ggml_mul_mat(ctx0, v, scores);
+
+    kqv = ggml_permute(ctx0, kqv, 1, 0, 2, 3);
+    kqv = ggml_cont(ctx0, kqv);
+
+
+    kqv = ggml_reshape_3d(ctx0, kqv, N, D * n_head, B);
+    kqv = ggml_reshape_4d(ctx0, kqv, W, H, D * n_head, B);
+    kqv = ggml_cont(ctx0, kqv);
+
+    // Output projection
+    cur = ggml_conv_2d_direct(ctx0, block.attn_o_w, kqv, 1, 1, 0, 0, 1, 1);
+
+    // Residual & Layer Scale
+    if (inp->ne[0] == cur->ne[0] && inp->ne[2] == cur->ne[2]) {
+        if (block.layer_scale_w) {
+            cur = ggml_mul(ctx0, cur, block.layer_scale_w);
+        }
+        cur = ggml_add(ctx0, cur, inp);
+    }
+
+    return cur;
+}
+
+ggml_cgraph * clip_graph_mobilenetv5::build() {
+    ggml_tensor * inp = build_inp_raw();
+
+    // 1. Stem - Conv2dSame(3, 64, kernel_size=(3, 3), stride=(2, 2))
+    ggml_tensor * cur = pad_same_2d(inp, 3, 3, 2, 2);  // Apply SAME padding
+
+    cur = ggml_conv_2d_direct(ctx0, model.mobilenet_stem_conv_w, cur, 2, 2, 0, 0, 1, 1);  // padding=0
+    if (model.mobilenet_stem_conv_b) {
+        cur = ggml_add(ctx0, cur, model.mobilenet_stem_conv_b);
+    }
+    if (model.mobilenet_stem_norm_w) cur = rms_norm_2d(cur, model.mobilenet_stem_norm_w);
+    cur = ggml_gelu(ctx0, cur);
+
+
+    // 2. Blocks
+    std::vector<ggml_tensor*> intermediate_features;
+    const int total_blocks = model.mobilenet_blocks.size();
+
+    auto is_stage_start = [&](int i) {
+        if (i == 0) return true;
+        for (int end_idx : model.mobilenet_stage_ends) {
+            if (i == end_idx + 1) return true;
+        }
+        return false;
+    };
+
+    auto is_fusion_point = [&](int i) {
+        if (model.mobilenet_stage_ends.size() >= 4) {
+                if (i == model.mobilenet_stage_ends[2]) return true; // End of Stage 2
+                if (i == model.mobilenet_stage_ends[3]) return true; // End of Stage 3
+        } else {
+            if (i == total_blocks - 1) return true;
+        }
+        return false;
+    };
+
+    for (int i = 0; i < total_blocks; i++) {
+        const auto & block = model.mobilenet_blocks[i];
+        int stride = is_stage_start(i) ? 2 : 1;
+
+        if (block.s0_conv_exp_w)      cur = build_edge_residual(cur, block, stride);
+        else if (block.attn_q_w)      cur = build_mobilenet_attn(cur, block);
+        else                          cur = build_inverted_residual(cur, block, stride);
+
+        if (is_fusion_point(i)) {
+
+            intermediate_features.push_back(cur);
+        }
+    }
+
+    // 3. Multi-Scale Fusion Adapter (MSFA)
+    if (!intermediate_features.empty()) {
+
+        // A. Reference Resolution: PyTorch implementation uses inputs[0]
+        // We assume intermediate_features[0] is the "High Resolution" target.
+        // In MobileNet designs, this is typically the feature map with the smallest stride (e.g. 32x32).
+        ggml_tensor* target_feat = intermediate_features[0];
+        int high_res_w = target_feat->ne[0];
+        int high_res_h = target_feat->ne[1];
+
+        std::vector<ggml_tensor*> resized_feats;
+
+        // B. Resize inputs to match inputs[0] (High Resolution)
+        for (auto feat : intermediate_features) {
+            int feat_w = feat->ne[0];
+            int feat_h = feat->ne[1];
+
+            // PyTorch: if feat_size < high_resolution: interpolate
+            if (feat_w < high_res_w || feat_h < high_res_h) {
+                // Calculate scale factor.
+                // Note: PyTorch 'nearest' works on arbitrary float scales.
+                // ggml_upscale generally takes integer factors or target sizes depending on helper.
+                // Assuming standard power-of-2 scaling (e.g. 16 -> 32 means scale=2).
+                int scale_w = high_res_w / feat_w;
+                // int scale_h = high_res_h / feat_h;
+
+                // Safety check for non-integer scaling if strictly replicating
+                GGML_ASSERT(high_res_w % feat_w == 0);
+
+                // Upsample (Nearest Neighbor)
+                // 2 is the scale factor
+                feat = ggml_upscale(ctx0, feat, scale_w, ggml_scale_mode::GGML_SCALE_MODE_NEAREST);
+            }
+            resized_feats.push_back(feat);
+        }
+
+        // C. Concatenate at High Resolution (Channel Dim = 2 in ggml)
+        cur = resized_feats[0];
+        for (size_t k = 1; k < resized_feats.size(); ++k) {
+            cur = ggml_concat(ctx0, cur, resized_feats[k], 2);
+        }
+
+        // D. FFN (UniversalInvertedResidual)
+        // Structure: Expand Conv -> Norm -> GELU -> Project Conv -> Norm
+
+        // 1. Expansion
+        if (model.msfa_ffn_expand_w) {
+            // 1x1 Conv
+            cur = ggml_conv_2d_direct(ctx0, model.msfa_ffn_expand_w, cur, 1, 1, 0, 0, 1, 1);
+
+            if (model.msfa_ffn_expand_bn) {
+                cur = rms_norm_2d(cur, model.msfa_ffn_expand_bn);
+            }
+
+            cur = ggml_gelu(ctx0, cur);
+
+        }
+
+        // 2. Projection (No DW because kernel_size=0)
+        if (model.msfa_ffn_project_w) {
+            // 1x1 Conv
+            cur = ggml_conv_2d_direct(ctx0, model.msfa_ffn_project_w, cur, 1, 1, 0, 0, 1, 1);
+
+            // UniversalInvertedResidual typically has a norm after projection
+            if (model.msfa_ffn_project_bn) {
+                cur = rms_norm_2d(cur, model.msfa_ffn_project_bn);
+            }
+
+        }
+
+        // E. Final Downsample to Target Resolution (Output Resolution)
+        // PyTorch: matches self.output_resolution (e.g. 16x16)
+        const int target_out_res = 16;
+        int current_w = cur->ne[0];
+
+        if (current_w > target_out_res) {
+            int s = current_w / target_out_res;
+
+            GGML_ASSERT(current_w % target_out_res == 0);
+
+            // Avg Pool: Kernel=s, Stride=s
+            cur = ggml_pool_2d(ctx0, cur, GGML_OP_POOL_AVG, s, s, s, s, 0, 0);
+
+        }
+
+        // F. Final Norm
+        if (model.msfa_concat_norm_w) {
+            cur = rms_norm_2d(cur, model.msfa_concat_norm_w);
+
+        }
+    }
+
+    // 4. Gemma 3n Multimodal Projection (Embedder)
+    // Input: 'cur' is [Width, Height, Channels, Batch]
+    int W = cur->ne[0];
+    int H = cur->ne[1];
+    int C = cur->ne[2];
+    int B = cur->ne[3];
+
+    GGML_ASSERT(C == hparams.n_embd);
+
+    // 1. Permute and Flatten to [Channels, Tokens, Batch]
+    // PyTorch expects (Batch, Seq, Hidden), GGML usually processes (Hidden, Seq, Batch)
+    cur = ggml_permute(ctx0, cur, 2, 1, 0, 3); // -> [C, H, W, B]
+    cur = ggml_permute(ctx0, cur, 0, 2, 1, 3); // -> [C, W, H, B]
+    cur = ggml_cont(ctx0, cur);
+    cur = ggml_reshape_3d(ctx0, cur, C, W*H, B);
+    cur = ggml_cont(ctx0, cur);
+
+
+    // 2. FEATURE SCALING
+    // PyTorch: vision_outputs *= self.config.vision_config.hidden_size**0.5
+    const float scale_factor = sqrtf((float)C);
+    cur = ggml_scale(ctx0, cur, scale_factor);
+
+
+    // 3. SOFT EMBEDDING NORM
+    // PyTorch: self._norm(x) * self.weight
+    // We must normalize regardless, then multiply if weight exists.
+    {
+        const float eps = 1e-6f; // Gemma3n uses 1e-6
+        cur = ggml_rms_norm(ctx0, cur, eps);
+
+        if (model.mm_soft_emb_norm_w) {
+            // Weight shape is (2048,) -> Element-wise broadcast multiply
+            cur = ggml_mul(ctx0, cur, model.mm_soft_emb_norm_w);
+        }
+
+    }
+
+    // 4. PROJECTION
+    // PyTorch: embedding_projection = nn.Linear(vision_hidden, text_hidden, bias=False)
+    // Weight stored as [out_features, in_features] = [text_hidden_size, vision_hidden_size]
+    if (model.mm_input_proj_w) {
+        cur = ggml_mul_mat(ctx0, model.mm_input_proj_w, cur);
+    }
+
+    // 5. POST PROJECTION NORM
+    // PyTorch: embedding_post_projection_norm = Gemma3nRMSNorm(..., with_scale=False)
+    // with_scale=False means weight is registered as buffer with value 1.0
+    // So output = rms_norm(x) * 1.0 = rms_norm(x), magnitude ~1
+    {
+        const float eps = 1e-6f;
+        cur = ggml_rms_norm(ctx0, cur, eps);
+
+        if (model.mm_post_proj_norm_w) {
+            // If weight is loaded, multiply (should be ~1.0 anyway)
+            cur = ggml_mul(ctx0, cur, model.mm_post_proj_norm_w);
+        }
+    }
+
+    ggml_build_forward_expand(gf, cur);
+    return gf;
+}

tools/mtmd/models/models.h

@@ -0,0 +1,111 @@
+#pragma once
+
+#include "../clip-graph.h"
+
+/*
+ * IMPORTANT: The mtmd module does NOT accept pull requests that are fully or predominantly AI-generated.
+ * We encourage human contributors to ensure the quality and reliability of the codebase.
+ */
+
+struct clip_graph_siglip : clip_graph {
+    clip_graph_siglip(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+};
+
+struct clip_graph_pixtral : clip_graph {
+    clip_graph_pixtral(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+};
+
+struct clip_graph_qwen2vl : clip_graph {
+    clip_graph_qwen2vl(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+};
+
+struct clip_graph_qwen3vl : clip_graph {
+    clip_graph_qwen3vl(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+};
+
+struct clip_graph_youtuvl : clip_graph {
+    clip_graph_youtuvl(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+};
+
+struct clip_graph_minicpmv : clip_graph {
+    clip_graph_minicpmv(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+};
+
+struct clip_graph_internvl : clip_graph {
+    clip_graph_internvl(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+};
+
+struct clip_graph_llama4 : clip_graph {
+    clip_graph_llama4(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+};
+
+struct clip_graph_kimivl : clip_graph {
+    clip_graph_kimivl(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+};
+
+struct clip_graph_cogvlm : clip_graph {
+    clip_graph_cogvlm(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+};
+
+struct clip_graph_llava : clip_graph {
+    clip_graph_llava(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+};
+
+struct clip_graph_whisper_enc : clip_graph {
+    clip_graph_whisper_enc(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+};
+
+struct clip_graph_conformer : clip_graph {
+    clip_graph_conformer(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+};
+
+struct clip_graph_glm4v : clip_graph {
+    clip_graph_glm4v(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+};
+
+struct clip_graph_mobilenetv5 : clip_graph {
+    clip_graph_mobilenetv5(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+
+    ggml_tensor * rms_norm_2d(
+        ggml_tensor * inp,
+        ggml_tensor * weight,
+        float eps = 1e-6f);
+
+    ggml_tensor* pad_same_2d(
+        ggml_tensor* inp,
+        int kernel_h,
+        int kernel_w,
+        int stride_h,
+        int stride_w,
+        int dilation_h = 1,
+        int dilation_w = 1);
+
+    ggml_tensor * build_edge_residual(
+        ggml_tensor * inp,
+        const mobilenetv5_block & block,
+        int stride);
+
+    ggml_tensor * build_inverted_residual(
+        ggml_tensor * inp,
+        const mobilenetv5_block & block,
+        int stride);
+
+    ggml_tensor * build_mobilenet_attn(
+        ggml_tensor * inp,
+        const mobilenetv5_block & block);
+};

tools/mtmd/models/pixtral.cpp

@@ -0,0 +1,86 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_pixtral::build() {
+    const int n_merge = hparams.n_merge;
+
+    // 2D input positions
+    ggml_tensor * pos_h = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches);
+    ggml_set_name(pos_h, "pos_h");
+    ggml_set_input(pos_h);
+
+    ggml_tensor * pos_w = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches);
+    ggml_set_name(pos_w, "pos_w");
+    ggml_set_input(pos_w);
+
+    auto add_pos = [&](ggml_tensor * cur, const clip_layer &) {
+        return build_rope_2d(ctx0, cur, pos_h, pos_w, hparams.rope_theta, true);
+    };
+
+    ggml_tensor * inp = build_inp();
+    ggml_tensor * cur = build_vit(
+                            inp, n_patches,
+                            NORM_TYPE_RMS,
+                            hparams.ffn_op,
+                            nullptr, // no learned pos embd
+                            add_pos);
+
+    // mistral small 3.1 patch merger
+    // ref: https://github.com/huggingface/transformers/blob/7a3e208892c06a5e278144eaf38c8599a42f53e7/src/transformers/models/mistral3/modeling_mistral3.py#L67
+    if (model.mm_patch_merger_w) {
+        GGML_ASSERT(hparams.n_merge > 0);
+
+        cur = ggml_mul(ctx0, ggml_rms_norm(ctx0, cur, eps), model.mm_input_norm_w);
+
+        // reshape image tokens to 2D grid
+        cur = ggml_reshape_3d(ctx0, cur, n_embd, n_patches_x, n_patches_y);
+        cur = ggml_permute(ctx0, cur, 2, 0, 1, 3); // [x, y, n_embd]
+        cur = ggml_cont(ctx0, cur);
+
+        // torch.nn.functional.unfold is just an im2col under the hood
+        // we just need a dummy kernel to make it work
+        ggml_tensor * kernel = ggml_view_3d(ctx0, cur, n_merge, n_merge, cur->ne[2], 0, 0, 0);
+        cur = ggml_im2col(ctx0, kernel, cur, n_merge, n_merge, 0, 0, 1, 1, true, inp->type);
+
+        // project to n_embd
+        cur = ggml_reshape_2d(ctx0, cur, cur->ne[0], cur->ne[1] * cur->ne[2]);
+        cur = ggml_mul_mat(ctx0, model.mm_patch_merger_w, cur);
+    }
+
+    // LlavaMultiModalProjector (always using GELU activation)
+    {
+        cur = build_ffn(cur,
+            model.mm_1_w, model.mm_1_b,
+            nullptr, nullptr,
+            model.mm_2_w, model.mm_2_b,
+            FFN_GELU,
+            -1);
+    }
+
+    // arrangement of the [IMG_BREAK] token
+    if (model.token_embd_img_break) {
+        // not efficient, but works
+        // the trick is to view the embeddings as a 3D tensor with shape [n_embd, n_patches_per_row, n_rows]
+        // and then concatenate the [IMG_BREAK] token to the end of each row, aka n_patches_per_row dimension
+        // after the concatenation, we have a tensor with shape [n_embd, n_patches_per_row + 1, n_rows]
+
+        const int p_y             = n_merge > 0 ? n_patches_y / n_merge : n_patches_y;
+        const int p_x             = n_merge > 0 ? n_patches_x / n_merge : n_patches_x;
+        const int p_total         = p_x * p_y;
+        const int n_embd_text     = cur->ne[0];
+        const int n_tokens_output = p_total + p_y - 1; // one [IMG_BREAK] per row, except the last row
+
+        ggml_tensor * tmp = ggml_reshape_3d(ctx0, cur, n_embd_text, p_x, p_y);
+        ggml_tensor * tok = ggml_new_tensor_3d(ctx0, tmp->type, n_embd_text, 1, p_y);
+        tok = ggml_scale(ctx0, tok, 0.0); // clear the tensor
+        tok = ggml_add(ctx0, tok, model.token_embd_img_break);
+        tmp = ggml_concat(ctx0, tmp, tok, 1);
+        cur = ggml_view_2d(ctx0, tmp,
+            n_embd_text, n_tokens_output,
+            ggml_row_size(tmp->type, n_embd_text), 0);
+    }
+
+    // build the graph
+    ggml_build_forward_expand(gf, cur);
+
+    return gf;
+}

tools/mtmd/models/qwen2vl.cpp

@@ -0,0 +1,183 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_qwen2vl::build() {
+    GGML_ASSERT(model.patch_bias == nullptr);
+    GGML_ASSERT(model.class_embedding == nullptr);
+
+    const int batch_size       = 1;
+    const bool use_window_attn = hparams.n_wa_pattern > 0;
+    const int n_wa_pattern     = hparams.n_wa_pattern;
+    const int n_pos            = n_patches;
+    const int num_position_ids = n_pos * 4; // m-rope requires 4 dim per position
+
+    norm_type norm_t = proj_type == PROJECTOR_TYPE_QWEN25VL
+        ? NORM_TYPE_RMS // qwen 2.5 vl
+        : NORM_TYPE_NORMAL; // qwen 2 vl
+
+    int mrope_sections[4] = {d_head/4, d_head/4, d_head/4, d_head/4};
+
+    ggml_tensor * inp_raw = build_inp_raw();
+    ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings_0, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
+
+    GGML_ASSERT(img.nx % (patch_size * 2) == 0);
+    GGML_ASSERT(img.ny % (patch_size * 2) == 0);
+
+    // second conv dimension
+    {
+        auto inp_1 = ggml_conv_2d(ctx0, model.patch_embeddings_1, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
+        inp = ggml_add(ctx0, inp, inp_1);
+
+        inp = ggml_permute(ctx0, inp, 1, 2, 0, 3);  // [w, h, c, b] -> [c, w, h, b]
+        inp = ggml_cont_4d(
+            ctx0, inp,
+            n_embd * 2, n_patches_x / 2, n_patches_y, batch_size);
+        inp = ggml_reshape_4d(
+            ctx0, inp,
+            n_embd * 2, n_patches_x / 2, 2, batch_size * (n_patches_y / 2));
+        inp = ggml_permute(ctx0, inp, 0, 2, 1, 3);
+        inp = ggml_cont_3d(
+            ctx0, inp,
+            n_embd, n_patches_x * n_patches_y, batch_size);
+    }
+
+    ggml_tensor * inpL           = inp;
+    ggml_tensor * window_mask    = nullptr;
+    ggml_tensor * window_idx     = nullptr;
+    ggml_tensor * inv_window_idx = nullptr;
+
+    ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_position_ids);
+    ggml_set_name(positions, "positions");
+    ggml_set_input(positions);
+
+    // pre-layernorm
+    if (model.pre_ln_w) {
+        inpL = build_norm(inpL, model.pre_ln_w, model.pre_ln_b, norm_t, eps, -1);
+    }
+
+    if (use_window_attn) {
+        // handle window attention inputs
+        inv_window_idx = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos / 4);
+        ggml_set_name(inv_window_idx, "inv_window_idx");
+        ggml_set_input(inv_window_idx);
+        // mask for window attention
+        window_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_pos, n_pos);
+        ggml_set_name(window_mask, "window_mask");
+        ggml_set_input(window_mask);
+
+        // if flash attn is used, we need to pad the mask and cast to f16
+        if (flash_attn_type == CLIP_FLASH_ATTN_TYPE_ENABLED) {
+            window_mask = ggml_cast(ctx0, window_mask, GGML_TYPE_F16);
+        }
+
+        // inpL shape: [n_embd, n_patches_x * n_patches_y, batch_size]
+        GGML_ASSERT(batch_size == 1);
+        inpL = ggml_reshape_2d(ctx0, inpL, n_embd * 4, n_patches_x * n_patches_y * batch_size / 4);
+        inpL = ggml_get_rows(ctx0, inpL, inv_window_idx);
+        inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_patches_x * n_patches_y, batch_size);
+    }
+
+    // loop over layers
+    for (int il = 0; il < n_layer; il++) {
+        const auto & layer = model.layers[il];
+        const bool full_attn = use_window_attn ? (il + 1) % n_wa_pattern == 0 : true;
+
+        ggml_tensor * cur = inpL; // inpL = residual, cur = hidden_states
+
+        // layernorm1
+        cur = build_norm(cur, layer.ln_1_w, layer.ln_1_b, norm_t, eps, il);
+        cb(cur, "ln1", il);
+
+        // self-attention
+        {
+            ggml_tensor * Qcur = ggml_add(ctx0,
+                ggml_mul_mat(ctx0, layer.q_w, cur), layer.q_b);
+            ggml_tensor * Kcur = ggml_add(ctx0,
+                ggml_mul_mat(ctx0, layer.k_w, cur), layer.k_b);
+            ggml_tensor * Vcur = ggml_add(ctx0,
+                ggml_mul_mat(ctx0, layer.v_w, cur), layer.v_b);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, d_head, n_head, n_patches);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, d_head, n_head, n_patches);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, d_head, n_head, n_patches);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            // apply M-RoPE
+            Qcur = ggml_rope_multi(
+                ctx0, Qcur, positions, nullptr,
+                d_head/2, mrope_sections, GGML_ROPE_TYPE_VISION, 32768, 10000, 1, 0, 1, 32, 1);
+            Kcur = ggml_rope_multi(
+                ctx0, Kcur, positions, nullptr,
+                d_head/2, mrope_sections, GGML_ROPE_TYPE_VISION, 32768, 10000, 1, 0, 1, 32, 1);
+
+            cb(Qcur, "Qcur_rope", il);
+            cb(Kcur, "Kcur_rope", il);
+
+            ggml_tensor * attn_mask = full_attn ? nullptr : window_mask;
+
+            cur = build_attn(layer.o_w, layer.o_b,
+                Qcur, Kcur, Vcur, attn_mask, kq_scale, il);
+            cb(cur, "attn_out", il);
+        }
+
+        // re-add the layer input, e.g., residual
+        cur = ggml_add(ctx0, cur, inpL);
+
+        inpL = cur; // inpL = residual, cur = hidden_states
+
+        cb(cur, "ffn_inp", il);
+
+        // layernorm2
+        cur = build_norm(cur, layer.ln_2_w, layer.ln_2_b, norm_t, eps, il);
+        cb(cur, "ffn_inp_normed", il);
+
+        // ffn
+        cur = build_ffn(cur,
+            layer.ff_up_w, layer.ff_up_b,
+            layer.ff_gate_w, layer.ff_gate_b,
+            layer.ff_down_w, layer.ff_down_b,
+            hparams.ffn_op, il);
+
+        cb(cur, "ffn_out", il);
+
+        // residual 2
+        cur = ggml_add(ctx0, inpL, cur);
+        cb(cur, "layer_out", il);
+
+        inpL = cur;
+    }
+
+    // post-layernorm
+    if (model.post_ln_w) {
+        inpL = build_norm(inpL, model.post_ln_w, model.post_ln_b, norm_t, eps, n_layer);
+    }
+
+    // multimodal projection
+    ggml_tensor * embeddings = inpL;
+    embeddings = ggml_reshape_3d(ctx0, embeddings, n_embd * 4, n_pos / 4, batch_size);
+    embeddings = build_ffn(embeddings,
+                        model.mm_0_w, model.mm_0_b,
+                        nullptr, nullptr,
+                        model.mm_1_w, model.mm_1_b,
+                        FFN_GELU,
+                        -1);
+
+    if (use_window_attn) {
+        window_idx = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos / 4);
+        ggml_set_name(window_idx, "window_idx");
+        ggml_set_input(window_idx);
+
+        // embeddings shape: [n_embd, n_patches_x * n_patches_y, batch_size]
+        GGML_ASSERT(batch_size == 1);
+        embeddings = ggml_reshape_2d(ctx0, embeddings, hparams.projection_dim, n_patches_x * n_patches_y / 4);
+        embeddings = ggml_get_rows(ctx0, embeddings, window_idx);
+        embeddings = ggml_reshape_3d(ctx0, embeddings, hparams.projection_dim, n_patches_x * n_patches_y / 4, batch_size);
+    }
+
+    // build the graph
+    ggml_build_forward_expand(gf, embeddings);
+
+    return gf;
+}

tools/mtmd/models/qwen3vl.cpp

@@ -0,0 +1,191 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_qwen3vl::build() {
+    GGML_ASSERT(model.patch_bias != nullptr);
+    GGML_ASSERT(model.position_embeddings != nullptr);
+    GGML_ASSERT(model.class_embedding == nullptr);
+
+    const int batch_size       = 1;
+    const int n_pos            = n_patches;
+    const int num_position_ids = n_pos * 4; // m-rope requires 4 dim per position
+
+    norm_type norm_t = NORM_TYPE_NORMAL;
+
+    int mrope_sections[4] = {d_head/4, d_head/4, d_head/4, d_head/4};
+
+    ggml_tensor * inp_raw = build_inp_raw();
+    ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings_0, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
+
+    GGML_ASSERT(img.nx % (patch_size * 2) == 0);
+    GGML_ASSERT(img.ny % (patch_size * 2) == 0);
+
+    // second conv dimension
+    {
+        auto inp_1 = ggml_conv_2d(ctx0, model.patch_embeddings_1, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
+        inp = ggml_add(ctx0, inp, inp_1);
+
+        inp = ggml_permute(ctx0, inp, 1, 2, 0, 3);  // [w, h, c, b] -> [c, w, h, b]
+        inp = ggml_cont_4d(
+            ctx0, inp,
+            n_embd * 2, n_patches_x / 2, n_patches_y, batch_size);
+        inp = ggml_reshape_4d(
+            ctx0, inp,
+            n_embd * 2, n_patches_x / 2, 2, batch_size * (n_patches_y / 2));
+        inp = ggml_permute(ctx0, inp, 0, 2, 1, 3);
+        inp = ggml_cont_3d(
+            ctx0, inp,
+            n_embd, n_patches_x * n_patches_y, batch_size);
+    }
+
+    // add patch bias
+    if (model.patch_bias != nullptr) {
+        inp = ggml_add(ctx0, inp, model.patch_bias);
+        cb(inp, "patch_bias", -1);
+    }
+
+    // calculate absolute position embedding and apply
+    ggml_tensor * learned_pos_embd = resize_position_embeddings();
+    learned_pos_embd = ggml_cont_4d(
+        ctx0, learned_pos_embd,
+        n_embd * 2, n_patches_x / 2, n_patches_y, batch_size);
+    learned_pos_embd = ggml_reshape_4d(
+        ctx0, learned_pos_embd,
+        n_embd * 2, n_patches_x / 2, 2, batch_size * (n_patches_y / 2));
+    learned_pos_embd = ggml_permute(ctx0, learned_pos_embd, 0, 2, 1, 3);
+    learned_pos_embd = ggml_cont_3d(
+        ctx0, learned_pos_embd,
+        n_embd, n_patches_x * n_patches_y, batch_size);
+    inp = ggml_add(ctx0, inp, learned_pos_embd);
+    cb(inp, "inp_pos_emb", -1);
+
+    ggml_tensor * inpL = inp;
+
+    ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_position_ids);
+    ggml_set_name(positions, "positions");
+    ggml_set_input(positions);
+
+    // pre-layernorm
+    if (model.pre_ln_w) {
+        inpL = build_norm(inpL, model.pre_ln_w, model.pre_ln_b, norm_t, eps, -1);
+    }
+
+    // deepstack features (stack along the feature dimension), [n_embd * len(deepstack_layers), n_patches_x * n_patches_y, batch_size]
+    ggml_tensor * deepstack_features = nullptr;
+    const int merge_factor = hparams.n_merge > 0 ? hparams.n_merge * hparams.n_merge : 4; // default 2x2=4 for qwen3vl
+
+    // loop over layers
+    for (int il = 0; il < n_layer; il++) {
+        auto & layer = model.layers[il];
+
+        ggml_tensor * cur = inpL; // inpL = residual, cur = hidden_states
+
+        // layernorm1
+        cur = build_norm(cur, layer.ln_1_w, layer.ln_1_b, norm_t, eps, il);
+        cb(cur, "ln1", il);
+
+        // self-attention
+        {
+            cur = ggml_mul_mat(ctx0, layer.qkv_w, cur);
+            cur = ggml_add(ctx0, cur, layer.qkv_b);
+
+            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, d_head, n_head, n_pos,
+                    /* nb1    */ ggml_row_size(cur->type, d_head),
+                    /* nb2    */ cur->nb[1],
+                    /* offset */ 0);
+
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, d_head, n_head, n_pos,
+                    /* nb1    */ ggml_row_size(cur->type, d_head),
+                    /* nb2    */ cur->nb[1],
+                    /* offset */ ggml_row_size(cur->type, n_embd));
+
+            ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, d_head, n_head, n_pos,
+                    /* nb1    */ ggml_row_size(cur->type, d_head),
+                    /* nb2    */ cur->nb[1],
+                    /* offset */ ggml_row_size(cur->type, 2 * n_embd));
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            // apply M-RoPE
+            Qcur = ggml_rope_multi(
+                ctx0, Qcur, positions, nullptr,
+                d_head/2, mrope_sections, GGML_ROPE_TYPE_VISION, 32768, 10000, 1, 0, 1, 32, 1);
+            Kcur = ggml_rope_multi(
+                ctx0, Kcur, positions, nullptr,
+                d_head/2, mrope_sections, GGML_ROPE_TYPE_VISION, 32768, 10000, 1, 0, 1, 32, 1);
+
+            cb(Qcur, "Qcur_rope", il);
+            cb(Kcur, "Kcur_rope", il);
+
+            cur = build_attn(layer.o_w, layer.o_b,
+                Qcur, Kcur, Vcur, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+        }
+
+        // re-add the layer input, e.g., residual
+        cur = ggml_add(ctx0, cur, inpL);
+
+        inpL = cur; // inpL = residual, cur = hidden_states
+
+        cb(cur, "ffn_inp", il);
+
+        // layernorm2
+        cur = build_norm(cur, layer.ln_2_w, layer.ln_2_b, norm_t, eps, il);
+        cb(cur, "ffn_inp_normed", il);
+
+        // ffn
+        cur = build_ffn(cur,
+            layer.ff_up_w, layer.ff_up_b,
+            layer.ff_gate_w, layer.ff_gate_b,
+            layer.ff_down_w, layer.ff_down_b,
+            hparams.ffn_op, il);
+
+        cb(cur, "ffn_out", il);
+
+        // residual 2
+        cur = ggml_add(ctx0, inpL, cur);
+        cb(cur, "layer_out", il);
+
+        if (layer.has_deepstack()) {
+            ggml_tensor * feat = ggml_reshape_3d(ctx0, cur, n_embd * merge_factor, n_pos / merge_factor, batch_size);
+            feat = build_norm(feat, layer.deepstack_norm_w, layer.deepstack_norm_b, norm_t, eps, il);
+            feat = build_ffn(feat,
+                layer.deepstack_fc1_w, layer.deepstack_fc1_b,
+                nullptr, nullptr,
+                layer.deepstack_fc2_w, layer.deepstack_fc2_b,
+                ffn_op_type::FFN_GELU, il);
+
+            if(!deepstack_features) {
+                deepstack_features = feat;
+            } else {
+                // concat along the feature dimension
+                deepstack_features = ggml_concat(ctx0, deepstack_features, feat, 0);
+            }
+        }
+
+        inpL = cur;
+    }
+
+    // post-layernorm
+    if (model.post_ln_w) {
+        inpL = build_norm(inpL, model.post_ln_w, model.post_ln_b, norm_t, eps, n_layer);
+    }
+
+    // multimodal projection
+    ggml_tensor * embeddings = inpL;
+    embeddings = ggml_reshape_3d(ctx0, embeddings, n_embd * 4, n_pos / 4, batch_size);
+
+    embeddings = build_ffn(embeddings,
+        model.mm_0_w, model.mm_0_b,
+        nullptr, nullptr,
+        model.mm_1_w, model.mm_1_b,
+        ffn_op_type::FFN_GELU, -1);
+
+    embeddings = ggml_concat(ctx0, embeddings, deepstack_features, 0); // concat along the feature dimension
+
+    // build the graph
+    ggml_build_forward_expand(gf, embeddings);
+
+    return gf;
+}

tools/mtmd/models/siglip.cpp

@@ -0,0 +1,86 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_siglip::build() {
+    ggml_tensor * inp = build_inp();
+
+    ggml_tensor * learned_pos_embd = model.position_embeddings;
+    if (proj_type == PROJECTOR_TYPE_LFM2) {
+        learned_pos_embd = resize_position_embeddings();
+    }
+
+    ggml_tensor * cur = build_vit(
+                            inp, n_patches,
+                            NORM_TYPE_NORMAL,
+                            hparams.ffn_op,
+                            learned_pos_embd,
+                            nullptr);
+
+    if (proj_type == PROJECTOR_TYPE_GEMMA3) {
+        const int batch_size = 1;
+        GGML_ASSERT(n_patches_x == n_patches_y);
+        const int patches_per_image = n_patches_x;
+        const int kernel_size = hparams.n_merge;
+
+        cur = ggml_transpose(ctx0, cur);
+        cur = ggml_cont_4d(ctx0, cur, patches_per_image, patches_per_image, n_embd, batch_size);
+
+        // doing a pool2d to reduce the number of output tokens
+        cur = ggml_pool_2d(ctx0, cur, GGML_OP_POOL_AVG, kernel_size, kernel_size, kernel_size, kernel_size, 0, 0);
+        cur = ggml_reshape_3d(ctx0, cur, cur->ne[0] * cur->ne[0], n_embd, batch_size);
+        cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
+
+        // apply norm before projection
+        cur = ggml_rms_norm(ctx0, cur, eps);
+        cur = ggml_mul(ctx0, cur, model.mm_soft_emb_norm_w);
+
+        // apply projection
+        cur = ggml_mul_mat(ctx0,
+            ggml_cont(ctx0, ggml_transpose(ctx0, model.mm_input_proj_w)),
+            cur);
+
+    } else if (proj_type == PROJECTOR_TYPE_IDEFICS3) {
+        // pixel_shuffle
+        // https://github.com/huggingface/transformers/blob/0a950e0bbe1ed58d5401a6b547af19f15f0c195e/src/transformers/models/idefics3/modeling_idefics3.py#L578
+        const int scale_factor = model.hparams.n_merge;
+        cur = build_patch_merge_permute(cur, scale_factor);
+        cur = ggml_mul_mat(ctx0, model.projection, cur);
+
+    } else if (proj_type == PROJECTOR_TYPE_LFM2) {
+        // pixel unshuffle block
+        const int scale_factor = model.hparams.n_merge;
+        cur = build_patch_merge_permute(cur, scale_factor);
+
+        // projection, in LFM2-VL input norm is optional
+        if (model.mm_input_norm_w) {
+            cur = ggml_norm(ctx0, cur, 1e-5); // default nn.LayerNorm
+            cur = ggml_mul(ctx0, cur, model.mm_input_norm_w);
+        }
+
+        if (model.mm_input_norm_b) {
+            cur = ggml_add(ctx0, cur, model.mm_input_norm_b);
+        }
+
+        cur = build_ffn(cur,
+            model.mm_1_w, model.mm_1_b,
+            nullptr, nullptr,
+            model.mm_2_w, model.mm_2_b,
+            FFN_GELU,
+            -1);
+
+    } else if (proj_type == PROJECTOR_TYPE_JANUS_PRO) {
+        cur = build_ffn(cur,
+            model.mm_0_w, model.mm_0_b,
+            nullptr, nullptr,
+            model.mm_1_w, model.mm_1_b,
+            hparams.ffn_op,
+            -1);
+
+    } else {
+        GGML_ABORT("SigLIP: Unsupported projector type");
+    }
+
+    // build the graph
+    ggml_build_forward_expand(gf, cur);
+
+    return gf;
+}

tools/mtmd/models/whisper-enc.cpp

@@ -0,0 +1,115 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_whisper_enc::build() {
+    const int n_frames = img.nx;
+    const int n_pos    = n_frames / 2;
+    GGML_ASSERT(model.position_embeddings->ne[1] >= n_pos);
+
+    ggml_tensor * inp = build_inp_raw(1);
+
+    // conv1d block
+    {
+        // convolution + gelu
+        ggml_tensor * cur = ggml_conv_1d_ph(ctx0, model.conv1d_1_w, inp, 1, 1);
+        cur = ggml_add(ctx0, cur, model.conv1d_1_b);
+
+        cur = ggml_gelu_erf(ctx0, cur);
+
+        cur = ggml_conv_1d_ph(ctx0, model.conv1d_2_w, cur, 2, 1);
+        cur = ggml_add(ctx0, cur, model.conv1d_2_b);
+
+        cur = ggml_gelu_erf(ctx0, cur);
+        // transpose
+        inp = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
+        cb(inp, "after_conv1d", -1);
+    }
+
+    // sanity check (only check one layer, but it should be the same for all)
+    GGML_ASSERT(model.layers[0].ln_1_w && model.layers[0].ln_1_b);
+    GGML_ASSERT(model.layers[0].ln_2_w && model.layers[0].ln_2_b);
+    GGML_ASSERT(model.layers[0].q_b);
+    GGML_ASSERT(model.layers[0].v_b);
+    GGML_ASSERT(!model.layers[0].k_b); // no bias for k
+
+    ggml_tensor * pos_embd_selected = ggml_view_2d(
+        ctx0, model.position_embeddings,
+        model.position_embeddings->ne[0], n_pos,
+        model.position_embeddings->nb[1], 0
+    );
+    ggml_tensor * cur = build_vit(
+                            inp, n_pos,
+                            NORM_TYPE_NORMAL,
+                            hparams.ffn_op,
+                            pos_embd_selected,
+                            nullptr);
+
+    cb(cur, "after_transformer", -1);
+
+    if (model.audio_has_stack_frames()) {
+        // StackAudioFrames
+        // https://huggingface.co/fixie-ai/ultravox-v0_5-llama-3_2-1b/blob/main/ultravox_model.py
+        cur = build_stack(cur, hparams.proj_stack_factor, n_embd);
+        cb(cur, "after_stacked", -1);
+    }
+
+    if (proj_type == PROJECTOR_TYPE_ULTRAVOX) {
+        // UltravoxProjector
+        // pre-norm
+        cur = ggml_rms_norm(ctx0, cur, 1e-6);
+        cur = ggml_mul(ctx0, cur, model.mm_norm_pre_w);
+
+        // ffn in
+        cur = ggml_mul_mat(ctx0, model.mm_1_w, cur);
+
+        // swiglu
+        // see SwiGLU in ultravox_model.py, the second half passed through is silu, not the first half
+        cur = ggml_swiglu_swapped(ctx0, cur);
+
+        // mid-norm
+        cur = ggml_rms_norm(ctx0, cur, 1e-6);
+        cur = ggml_mul(ctx0, cur, model.mm_norm_mid_w);
+
+        // ffn out
+        cur = ggml_mul_mat(ctx0, model.mm_2_w, cur);
+
+    } else if (proj_type == PROJECTOR_TYPE_QWEN2A) {
+        // projector
+        cur = ggml_mul_mat(ctx0, model.mm_fc_w, cur);
+        cur = ggml_add(ctx0, cur, model.mm_fc_b);
+
+    } else if (proj_type == PROJECTOR_TYPE_VOXTRAL) {
+        // projector
+        cur = build_ffn(cur,
+            model.mm_1_w, model.mm_1_b,
+            nullptr, nullptr,
+            model.mm_2_w, model.mm_2_b,
+            FFN_GELU_ERF,
+            -1);
+
+    } else if (proj_type == PROJECTOR_TYPE_MUSIC_FLAMINGO) {
+        // projector
+        cur = build_ffn(cur,
+            model.mm_1_w, model.mm_1_b,
+            nullptr, nullptr,
+            model.mm_2_w, model.mm_2_b,
+            FFN_GELU_ERF,
+            -1);
+
+    } else if (proj_type == PROJECTOR_TYPE_GLMA) {
+            cur = ggml_norm(ctx0, cur, hparams.eps);
+            cur = ggml_mul(ctx0, cur, model.mm_norm_pre_w);
+            cur = ggml_add(ctx0, cur, model.mm_norm_pre_b);
+            cur = build_stack(cur, hparams.proj_stack_factor, n_embd);
+            cur = build_ffn(cur, model.mm_1_w, model.mm_1_b, nullptr, nullptr, model.mm_2_w, model.mm_2_b, hparams.ffn_op, 0);
+            cur = ggml_concat(ctx0, model.mm_boi, cur, 1);
+            cur = ggml_concat(ctx0, cur, model.mm_eoi, 1);
+    } else {
+        GGML_ABORT("%s: unknown projector type", __func__);
+    }
+
+    cb(cur, "projected", -1);
+
+    ggml_build_forward_expand(gf, cur);
+
+    return gf;
+}

tools/mtmd/models/youtuvl.cpp

@@ -0,0 +1,179 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_youtuvl::build() {
+    GGML_ASSERT(model.class_embedding == nullptr);
+    const int batch_size       = 1;
+    const bool use_window_attn = !hparams.wa_layer_indexes.empty();
+    const int n_pos            = n_patches;
+    const int num_position_ids = n_pos * 4;
+    const int m = 2;
+    const int Wp = n_patches_x;
+    const int Hp = n_patches_y;
+    const int Hm = Hp / m;
+    const int Wm = Wp / m;
+    norm_type norm_t = NORM_TYPE_NORMAL;
+
+    int mrope_sections[4] = {d_head/4, d_head/4, d_head/4, d_head/4};
+
+    ggml_tensor * inp = build_inp_raw();
+
+    // change conv3d to linear
+    // reshape and permute to get patches, permute from (patch_size, m, Wm, patch_size, m, Hm, C) to (C, patch_size, patch_size, m, m, Wm, Hm)
+    {
+        inp = ggml_reshape_4d(
+            ctx0, inp,
+            Wm * m * patch_size, m * patch_size, Hm, 3);
+        inp = ggml_permute(ctx0, inp, 1, 2, 3, 0);
+        inp = ggml_cont_4d(
+            ctx0, inp,
+            m * patch_size * 3, Wm, m * patch_size, Hm);
+
+        inp = ggml_permute(ctx0, inp, 0, 2, 1, 3);
+        inp = ggml_cont_4d(
+            ctx0, inp,
+            m * patch_size * 3, patch_size, m, Hm * Wm);
+
+        inp = ggml_permute(ctx0, inp, 1, 0, 2, 3);
+        inp = ggml_cont_4d(
+            ctx0, inp,
+            patch_size, 3, patch_size, Hm * Wm * m * m);
+
+        inp = ggml_permute(ctx0, inp, 2, 0, 1, 3);
+        inp = ggml_cont_3d(
+            ctx0, inp,
+            3*patch_size* patch_size,  Hm * Wm * m * m, 1);
+    }
+    inp = ggml_mul_mat(ctx0, model.patch_embeddings_0, inp);
+
+    if (model.patch_bias) {
+        inp = ggml_add(ctx0, inp, model.patch_bias);
+    }
+
+    inp = ggml_reshape_2d(ctx0, inp, n_embd, n_patches);
+
+    ggml_tensor * inpL           = inp;
+    ggml_tensor * window_mask    = nullptr;
+    ggml_tensor * window_idx     = nullptr;
+    ggml_tensor * inv_window_idx = nullptr;
+
+    ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_position_ids);
+    ggml_set_name(positions, "positions");
+    ggml_set_input(positions);
+
+    // pre-layernorm
+    if (model.pre_ln_w) {
+        inpL = build_norm(inpL, model.pre_ln_w, model.pre_ln_b, norm_t, eps, -1);
+    }
+    if (use_window_attn) {
+        inv_window_idx = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos / 4);
+        ggml_set_name(inv_window_idx, "inv_window_idx");
+        ggml_set_input(inv_window_idx);
+        // mask for window attention
+        window_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_pos, n_pos);
+        ggml_set_name(window_mask, "window_mask");
+        ggml_set_input(window_mask);
+
+        // if flash attn is used, we need to pad the mask and cast to f16
+        if (flash_attn_type == CLIP_FLASH_ATTN_TYPE_ENABLED) {
+            window_mask = ggml_cast(ctx0, window_mask, GGML_TYPE_F16);
+        }
+
+        // inpL shape: [n_embd, n_patches_x * n_patches_y, batch_size]
+        GGML_ASSERT(batch_size == 1);
+        inpL = ggml_reshape_2d(ctx0, inpL, n_embd * 4, n_patches_x * n_patches_y * batch_size / 4);
+        inpL = ggml_get_rows(ctx0, inpL, inv_window_idx);
+        inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_patches_x * n_patches_y, batch_size);
+    }
+
+    // loop over layers
+    for (int il = 0; il < n_layer; il++) {
+        const auto & layer = model.layers[il];
+        const bool full_attn = use_window_attn ? hparams.wa_layer_indexes.count(il) > 0 : true;
+
+        ggml_tensor * cur = inpL; // inpL = residual, cur = hidden_states
+
+        // layernorm1
+        cur = build_norm(cur, layer.ln_1_w, layer.ln_1_b, norm_t, eps, il);
+        // self-attention
+        {
+            ggml_tensor * Qcur = ggml_add(ctx0,
+                ggml_mul_mat(ctx0, layer.q_w, cur), layer.q_b);
+            ggml_tensor * Kcur = ggml_add(ctx0,
+                ggml_mul_mat(ctx0, layer.k_w, cur), layer.k_b);
+            ggml_tensor * Vcur = ggml_add(ctx0,
+                ggml_mul_mat(ctx0, layer.v_w, cur), layer.v_b);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, d_head, n_head, n_patches);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, d_head, n_head, n_patches);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, d_head, n_head, n_patches);
+
+            Qcur = ggml_rope_multi(
+                ctx0, Qcur, positions, nullptr,
+                d_head/2, mrope_sections, GGML_ROPE_TYPE_VISION, 32768, 10000, 1, 0, 1, 32, 1);
+            Kcur = ggml_rope_multi(
+                ctx0, Kcur, positions, nullptr,
+                d_head/2, mrope_sections, GGML_ROPE_TYPE_VISION, 32768, 10000, 1, 0, 1, 32, 1);
+
+            ggml_tensor * attn_mask = full_attn ? nullptr : window_mask;
+
+            cur = build_attn(layer.o_w, layer.o_b,
+                Qcur, Kcur, Vcur, attn_mask, kq_scale, il);
+        }
+        // re-add the layer input, e.g., residual
+        cur = ggml_add(ctx0, cur, inpL);
+
+        inpL = cur; // inpL = residual, cur = hidden_states
+
+        // layernorm2
+        cur = build_norm(cur, layer.ln_2_w, layer.ln_2_b, norm_t, eps, il);
+
+        // ffn
+        cur = build_ffn(cur,
+            layer.ff_up_w, layer.ff_up_b,
+            nullptr, nullptr,
+            layer.ff_down_w, layer.ff_down_b,
+            hparams.ffn_op, il);
+
+        // residual 2
+        cur = ggml_add(ctx0, inpL, cur);
+
+        inpL = cur;
+    }
+
+    ggml_tensor * embeddings = inpL;
+    if (use_window_attn) {
+        const int spatial_merge_unit = 4;
+        window_idx = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos / spatial_merge_unit);
+        ggml_set_name(window_idx, "window_idx");
+        ggml_set_input(window_idx);
+        GGML_ASSERT(batch_size == 1);
+        embeddings = ggml_reshape_2d(ctx0, embeddings, n_embd * spatial_merge_unit, n_patches / spatial_merge_unit);
+        embeddings = ggml_get_rows(ctx0, embeddings, window_idx);
+        embeddings = ggml_reshape_3d(ctx0, embeddings, n_embd, n_patches, batch_size);
+        cb(embeddings, "window_order_restored", -1);
+    }
+
+    // post-layernorm (part of Siglip2VisionTransformer, applied after encoder)
+    if (model.post_ln_w) {
+        embeddings = build_norm(embeddings, model.post_ln_w, model.post_ln_b, norm_t, eps, n_layer);
+    }
+
+    // Now apply merger (VLPatchMerger):
+    // 1. Apply RMS norm (ln_q in VLPatchMerger)
+    embeddings = build_norm(embeddings, model.mm_input_norm_w, nullptr, NORM_TYPE_RMS, 1e-6, -1);
+    cb(embeddings, "merger_normed", -1);
+
+    // 2. First reshape for spatial merge (merge 2x2 patches)
+    embeddings = ggml_reshape_3d(ctx0, embeddings, n_embd * 4, n_pos / 4, batch_size);
+    cb(embeddings, "merger_reshaped", -1);
+
+    embeddings = build_ffn(embeddings,
+                    model.mm_0_w, model.mm_0_b,
+                    nullptr, nullptr,
+                    model.mm_1_w, model.mm_1_b,
+                    FFN_GELU,
+                    -1);
+    ggml_build_forward_expand(gf, embeddings);
+
+    return gf;
+}

tools/mtmd/mtmd-audio.cpp

@@ -0,0 +1,730 @@
+#include "mtmd-audio.h"
+
+#define _USE_MATH_DEFINES // for M_PI
+#include <cmath>
+#include <cstdint>
+#include <cstring>
+#include <thread>
+#include <vector>
+#include <fstream>
+#include <algorithm>
+
+// some of the code here is copied from whisper.cpp
+
+constexpr bool DEBUG = false;
+
+void mtmd_audio_cache::fill_sin_cos_table(int n) {
+    sin_vals.resize(n);
+    cos_vals.resize(n);
+    for (int i = 0; i < n; i++) {
+        double theta = (2 * M_PI * i) / n;
+        sin_vals[i]  = sinf(theta);
+        cos_vals[i]  = cosf(theta);
+    }
+}
+
+void mtmd_audio_cache::fill_hann_window(int length, bool periodic) {
+    hann_window.resize(length);
+    int offset = -1;
+    if (periodic) {
+        offset = 0;
+    }
+    for (int i = 0; i < length; i++) {
+        hann_window[i] = 0.5 * (1.0 - cosf((2.0 * M_PI * i) / (length + offset)));
+    }
+}
+
+void mtmd_audio_cache::fill_mel_filterbank_matrix(int   n_mel,
+                                                  int   n_fft,
+                                                  int   sample_rate,
+                                                  float fmin,
+                                                  float fmax,
+                                                  bool  slaney_area_norm,
+                                                  float scale) {
+    GGML_ASSERT(n_mel > 0 && n_fft > 1);
+    if (fmax <= 0.0f) {
+        fmax = 0.5f * sample_rate;
+    }
+
+    // Slaney scale (matches librosa default)
+    const double min_log_hz  = 1000.0;
+    const double lin_slope   = 3 / 200.;
+    const double min_log_mel = min_log_hz * lin_slope;
+    const double log_step    = log(6.4) / 27.0;
+    auto         hz_to_mel   = [min_log_hz, lin_slope, log_step, min_log_mel](const double f_hz) -> double {
+        return (f_hz < min_log_hz) ? f_hz * lin_slope : min_log_mel + log(f_hz / min_log_hz) / log_step;
+    };
+    auto mel_to_hz = [min_log_hz, lin_slope, log_step, min_log_mel](const double m) -> double {
+        return (m < min_log_mel) ? m / lin_slope : min_log_hz * exp((m - min_log_mel) * log_step);
+    };
+
+    // infer N_fft from n_fft_bins
+    const double bin_hz_step = double(sample_rate) / double(n_fft);
+
+    // mel grid: n_mel + 2 edges
+    const double        m_lo = hz_to_mel(fmin);
+    const double        m_hi = hz_to_mel(fmax);
+    std::vector<double> mel_pts(n_mel + 2);
+    for (int i = 0; i < n_mel + 2; ++i) {
+        mel_pts[i] = m_lo + (m_hi - m_lo) * (double(i) / (n_mel + 1));
+    }
+
+    // convert to Hz
+    std::vector<double> hz_pts(n_mel + 2);
+    for (int i = 0; i < n_mel + 2; ++i) {
+        hz_pts[i] = mel_to_hz(mel_pts[i]);
+    }
+
+    const int n_fft_bins = n_fft / 2 + 1;
+
+    // filterbank
+    std::vector<float> out(n_mel * n_fft_bins, 0);
+    for (int m = 0; m < n_mel; ++m) {
+        const double f_left   = hz_pts[m];
+        const double f_center = hz_pts[m + 1];
+        const double f_right  = hz_pts[m + 2];
+
+        const double denom_l = std::max(1e-30, f_center - f_left);
+        const double denom_r = std::max(1e-30, f_right - f_center);
+        const double enorm   = slaney_area_norm ? (2.0 / std::max(1e-30, f_right - f_left)) : 1.0;
+
+        for (int k = 0; k < n_fft_bins; ++k) {
+            const double f = k * bin_hz_step;
+            double       w = 0.0;
+            if (f >= f_left && f <= f_center) {
+                w = (f - f_left) / denom_l;
+            } else if (f > f_center && f <= f_right) {
+                w = (f_right - f) / denom_r;
+            }
+            out[size_t(m) * size_t(n_fft_bins) + size_t(k)] = float(w * enorm * scale);
+        }
+    }
+
+    filters.n_mel = n_mel;
+    filters.n_fft = n_fft;
+    filters.data  = std::move(out);
+
+    if (DEBUG) {  // debug
+        for (size_t i = 0; i < filters.data.size(); ++i) {
+            if (filters.data[i] != 0.0f) {
+                printf("filters[%zu] = %f\n", i, filters.data[i] * 1000.0f);
+            }
+        }
+    }
+}
+
+// Unified DFT implementation for both forward and inverse transforms
+// Template parameters:
+//   Inverse: false = DFT with exp(-2πi·k·n/N), no scaling
+//            true  = IDFT with exp(+2πi·k·n/N), scales by 1/N
+//   RealInput: true = input is real-valued (stride 1), avoids imaginary computations
+//              false = input is complex-valued (interleaved real/imag, stride 2)
+template <bool Inverse, bool RealInput>
+static void dft_impl(const mtmd_audio_cache & cache, const float * in, int N, float * out) {
+    const int n_sin_cos_vals = cache.sin_vals.size();
+    const int sin_cos_step   = n_sin_cos_vals / N;
+
+    constexpr float sign  = Inverse ? 1.0f : -1.0f;
+    const float     scale = Inverse ? (1.0f / N) : 1.0f;
+
+    for (int k = 0; k < N; k++) {
+        float re = 0;
+        float im = 0;
+
+        for (int n = 0; n < N; n++) {
+            int   idx     = (k * n * sin_cos_step) % n_sin_cos_vals;
+            float cos_val = cache.cos_vals[idx];
+            float sin_val = cache.sin_vals[idx];
+
+            if constexpr (RealInput) {
+                // Real input: in_im = 0, simplifies to:
+                // re += in_re * cos_val
+                // im += sign * in_re * sin_val
+                float in_re = in[n];
+                re += in_re * cos_val;
+                im += sign * in_re * sin_val;
+            } else {
+                float in_re = in[n * 2 + 0];
+                float in_im = in[n * 2 + 1];
+                // (a + bi) * (cos + sign*i*sin) = (a*cos - sign*b*sin) + (sign*a*sin + b*cos)i
+                re += in_re * cos_val - sign * in_im * sin_val;
+                im += sign * in_re * sin_val + in_im * cos_val;
+            }
+        }
+
+        out[k * 2 + 0] = re * scale;
+        out[k * 2 + 1] = im * scale;
+    }
+}
+
+// Cooley-Tukey FFT/IFFT unified implementation
+// Template parameters:
+//   Inverse: false = FFT with exp(-2πi·k/N), no scaling
+//            true  = IFFT with exp(+2πi·k/N), scales by 0.5 at each level
+//   RealInput: true = input is real-valued (stride 1)
+//              false = input is complex-valued (interleaved real/imag, stride 2)
+template <bool Inverse, bool RealInput>
+static void fft_impl(const mtmd_audio_cache & cache, float * in, int N, float * out) {
+    const int n_sin_cos_vals = cache.sin_vals.size();
+
+    if (N == 1) {
+        out[0] = in[0];
+        if constexpr (RealInput) {
+            out[1] = 0.0f;
+        } else {
+            out[1] = in[1];
+        }
+        return;
+    }
+
+    const int half_N = N / 2;
+    if (N - half_N * 2 == 1) {
+        // Odd N: fall back to DFT
+        dft_impl<Inverse, RealInput>(cache, in, N, out);
+        return;
+    }
+
+    // Split into even and odd
+    if constexpr (RealInput) {
+        // Real input: stride is 1, copy only real values
+        float * even = in + N;
+        for (int i = 0; i < half_N; ++i) {
+            even[i] = in[2 * i];
+        }
+        float * even_fft = out + 2 * N;
+        fft_impl<Inverse, true>(cache, even, half_N, even_fft);
+
+        float * odd = even;
+        for (int i = 0; i < half_N; ++i) {
+            odd[i] = in[2 * i + 1];
+        }
+        float * odd_fft = even_fft + N;
+        fft_impl<Inverse, true>(cache, odd, half_N, odd_fft);
+    } else {
+        // Complex input: stride is 2, copy complex pairs
+        float * even = in + N * 2;
+        for (int i = 0; i < half_N; ++i) {
+            even[i * 2 + 0] = in[2 * i * 2 + 0];
+            even[i * 2 + 1] = in[2 * i * 2 + 1];
+        }
+        float * even_fft = out + 2 * N;
+        fft_impl<Inverse, false>(cache, even, half_N, even_fft);
+
+        float * odd = even;
+        for (int i = 0; i < half_N; ++i) {
+            odd[i * 2 + 0] = in[(2 * i + 1) * 2 + 0];
+            odd[i * 2 + 1] = in[(2 * i + 1) * 2 + 1];
+        }
+        float * odd_fft = even_fft + N;
+        fft_impl<Inverse, false>(cache, odd, half_N, odd_fft);
+    }
+
+    float * even_fft = out + 2 * N;
+    float * odd_fft  = even_fft + N;
+
+    const int sin_cos_step = n_sin_cos_vals / N;
+
+    constexpr float sign  = Inverse ? 1.0f : -1.0f;
+    constexpr float scale = Inverse ? 0.5f : 1.0f;
+
+    for (int k = 0; k < half_N; k++) {
+        int   idx = k * sin_cos_step;  // t = 2*M_PI*k/N
+        float re  = cache.cos_vals[idx];
+        float im  = sign * cache.sin_vals[idx];
+
+        float re_odd = odd_fft[2 * k + 0];
+        float im_odd = odd_fft[2 * k + 1];
+
+        out[2 * k + 0] = scale * (even_fft[2 * k + 0] + re * re_odd - im * im_odd);
+        out[2 * k + 1] = scale * (even_fft[2 * k + 1] + re * im_odd + im * re_odd);
+
+        out[2 * (k + half_N) + 0] = scale * (even_fft[2 * k + 0] - re * re_odd + im * im_odd);
+        out[2 * (k + half_N) + 1] = scale * (even_fft[2 * k + 1] - re * im_odd - im * re_odd);
+    }
+}
+
+// Forward FFT for real input (used by mel spectrogram)
+static void fft(const mtmd_audio_cache & cache, float * in, int N, float * out) {
+    fft_impl<false, true>(cache, in, N, out);
+}
+
+// Inverse FFT for complex input
+static void ifft(const mtmd_audio_cache & cache, float * in, int N, float * out) {
+    fft_impl<true, false>(cache, in, N, out);
+}
+
+struct filter_params {
+    int32_t n_mel;
+    int32_t n_fft_bins;
+    int32_t hann_window_size;
+    int32_t hop_length;
+    int32_t sample_rate;
+    bool    center_padding = false;
+    float   preemph = 0.f;
+    bool    use_natural_log = false;
+    bool    norm_per_feature = false;
+};
+
+static void log_mel_spectrogram_worker_thread(int                        ith,
+                                              const float *              hann,
+                                              const std::vector<float> & samples,
+                                              int                        n_samples,
+                                              int                        frame_size,
+                                              int                        frame_step,
+                                              int                        n_threads,
+                                              const filter_params &      params,
+                                              const mtmd_audio_cache &   cache,
+                                              mtmd_audio_mel &           out) {
+    std::vector<float> fft_in(frame_size * 2, 0.0);
+    std::vector<float> fft_out(frame_size * 2 * 2 * 2);
+
+    int n_fft_bins = params.n_fft_bins;
+    int i = ith;
+
+    const auto & filters = cache.filters;
+
+    // make sure n_fft == 1 + (WHISPER_N_FFT / 2), bin_0 to bin_nyquist
+    GGML_ASSERT(n_fft_bins == 1 + (frame_size / 2));
+    GGML_ASSERT(cache.sin_vals.size() == cache.cos_vals.size());
+    // calculate FFT only when fft_in are not all zero
+    for (; i < std::min(n_samples / frame_step + 1, out.n_len); i += n_threads) {
+        const int offset = i * frame_step;
+
+        // apply Hann window (~10% faster)
+        for (int j = 0; j < std::min(frame_size, n_samples - offset); j++) {
+            fft_in[j] = hann[j] * samples[offset + j];
+        }
+
+        // fill the rest with zeros
+        if (n_samples - offset < frame_size) {
+            std::fill(fft_in.begin() + (n_samples - offset), fft_in.end(), 0.0);
+        }
+
+        // FFT
+        fft(cache, fft_in.data(), frame_size, fft_out.data());
+
+        // Calculate modulus^2 of complex numbers
+        // Use pow(fft_out[2 * j + 0], 2) + pow(fft_out[2 * j + 1], 2) causes inference quality problem? Interesting.
+        for (int j = 0; j < n_fft_bins; j++) {
+            fft_out[j] = (fft_out[2 * j + 0] * fft_out[2 * j + 0] + fft_out[2 * j + 1] * fft_out[2 * j + 1]);
+        }
+
+        // mel spectrogram
+        for (int j = 0; j < out.n_mel; j++) {
+            double sum = 0.0;
+            // unroll loop (suggested by GH user @lunixbochs)
+            int k = 0;
+            for (k = 0; k < n_fft_bins - 3; k += 4) {
+                size_t idx = size_t(j) * size_t(n_fft_bins) + size_t(k);
+                sum +=
+                        fft_out[k + 0] * filters.data[idx + 0] +
+                        fft_out[k + 1] * filters.data[idx + 1] +
+                        fft_out[k + 2] * filters.data[idx + 2] +
+                        fft_out[k + 3] * filters.data[idx + 3];
+            }
+            // handle n_fft remainder
+            for (; k < n_fft_bins; k++) {
+                sum += fft_out[k] * filters.data[j * n_fft_bins + k];
+            }
+            sum = params.use_natural_log
+                ? log(sum + 5.960464477539063e-08)
+                : log10(std::max(sum, 1e-10));
+            out.data[j * out.n_len + i] = sum;
+        }
+    }
+
+    // Otherwise fft_out are all zero
+    double sum = params.use_natural_log ? log(1e-10) : log10(1e-10);
+    for (; i < out.n_len; i += n_threads) {
+        for (int j = 0; j < out.n_mel; j++) {
+            out.data[j * out.n_len + i] = sum;
+        }
+    }
+}
+
+// ref: https://github.com/openai/whisper/blob/main/whisper/audio.py#L110-L157
+static bool log_mel_spectrogram(
+        const float * samples,
+        const int     n_samples_in,
+        const int     n_threads,
+        const filter_params & params,
+        const mtmd_audio_cache & cache,
+        mtmd_audio_mel & out) {
+    //const int64_t t_start_us = ggml_time_us();
+
+    out.n_len_org = n_samples_in;
+    int n_samples = n_samples_in;
+
+    // Hann window
+    const float * hann       = cache.hann_window.data();
+    const int     frame_size = (params.n_fft_bins - 1) * 2;
+    const int     frame_step = params.hop_length;
+
+    // Padding
+    std::vector<float> samples_padded;
+    if (params.center_padding) {
+        const auto pad_amount = frame_size / 2;
+        samples_padded = std::vector<float>(n_samples + 2 * pad_amount, 0);
+        std::copy(samples, samples + n_samples, samples_padded.data() + pad_amount);
+        samples = samples_padded.data();
+        n_samples = samples_padded.size();
+    } else {
+        // existing padding logic
+        int64_t stage_1_pad = params.sample_rate * 30;
+        int64_t stage_2_pad = frame_size / 2;
+        samples_padded.resize(n_samples + stage_1_pad + stage_2_pad * 2);
+        std::copy(samples, samples + n_samples, samples_padded.begin() + stage_2_pad);
+        // pad 30 seconds of zeros at the end of audio (480,000 samples) + reflective pad 200 samples at the end of audio
+        std::fill(samples_padded.begin() + n_samples + stage_2_pad, samples_padded.begin() + n_samples + stage_1_pad + 2 * stage_2_pad, 0);
+        // reflective pad 200 samples at the beginning of audio
+        if (n_samples < stage_2_pad + 1) {
+            // TODO: Handle short audio differently or return error
+            return false;
+        }
+        std::reverse_copy(samples + 1, samples + 1 + stage_2_pad, samples_padded.begin());
+    }
+
+    // preemphasis
+    if (params.preemph) {
+        const int   pad_amount = frame_size / 2;
+        const float preemph = 0.97f;
+        float       prev = samples_padded[pad_amount];
+        for (int i = pad_amount + 1; i + pad_amount < n_samples; ++i) {
+            float cur = samples_padded[i];
+            samples_padded[i] = cur - preemph * prev;
+            prev = cur;
+        }
+    }
+
+    // pad hann window if it's smaller than frame_size
+    // TODO: probably unnecessary here? (or better doing it in g_cache?)
+    std::vector<float> hann_window_padded;
+    if (params.hann_window_size < frame_size) {
+        hann_window_padded.resize(frame_size);
+        const int padding = (frame_size - params.hann_window_size) / 2;
+        std::copy(hann, hann + params.hann_window_size, &hann_window_padded[padding]);
+        hann = hann_window_padded.data();
+    }
+
+
+    out.n_mel = params.n_mel;
+    out.n_len = (n_samples - frame_size) / frame_step + 1;
+    // TODO: handle these checks better
+    if (out.n_mel > 0 && (unsigned long)out.n_len > SIZE_MAX / out.n_mel) {
+        LOG_ERR("%s: size overflow\n", __func__);
+        return false;
+    }
+    if (n_samples < frame_size) {
+        LOG_ERR("%s: not enough samples after padding\n", __func__);
+        return false;
+    }
+    out.data.resize(out.n_mel * out.n_len);
+
+    {
+        std::vector<std::thread> workers(n_threads - 1);
+        for (int iw = 0; iw < n_threads - 1; ++iw) {
+            workers[iw] =
+                std::thread(log_mel_spectrogram_worker_thread, iw + 1, hann, std::cref(samples_padded), n_samples,
+                            frame_size, frame_step, n_threads, std::cref(params), std::cref(cache), std::ref(out));
+        }
+
+        // main thread
+        log_mel_spectrogram_worker_thread(0, hann, samples_padded, n_samples, frame_size, frame_step, n_threads, params,
+                                          cache, out);
+        for (int iw = 0; iw < n_threads - 1; ++iw) {
+            workers[iw].join();
+        }
+    }
+
+    const int effective_n_len = n_samples_in / frame_step;
+    if (params.norm_per_feature) {
+        for (int i = 0; i < out.n_mel; i++) {
+            double mean = 0;
+            for (int j = 0; j < effective_n_len; ++j) {
+                mean += out.data[i * out.n_len + j];
+            }
+            mean /= effective_n_len;
+
+            double var = 0.0;
+            for (int j = 0; j < effective_n_len; ++j) {
+                const double value = out.data[i * out.n_len + j] - mean;
+                var += value * value;
+            }
+            var /= effective_n_len - 1;  // unbiased
+            const double mstd = std::sqrt(var + 1e-5);
+
+            for (int j = 0; j < effective_n_len; ++j) {
+                auto &value = out.data[i * out.n_len + j];
+                value        = (value - mean) / mstd;
+            }
+
+            // pad the rest with zeros
+            for (int j = effective_n_len; j < out.n_len; ++j) {
+                out.data[i * out.n_len + j] = 0.0;
+            }
+        }
+    } else {
+        // clamping and normalization
+        double mmax = -1e20;
+        for (int i = 0; i < out.n_mel*out.n_len; i++) {
+            if (out.data[i] > mmax) {
+                mmax = out.data[i];
+            }
+        }
+
+        mmax -= 8.0;
+
+        for (int i = 0; i < out.n_mel*out.n_len; i++) {
+            if (out.data[i] < mmax) {
+                out.data[i] = mmax;
+            }
+            out.data[i] = (out.data[i] + 4.0)/4.0;
+        }
+    }
+
+    // Dump log_mel_spectrogram
+    if (DEBUG) {
+        std::ofstream outFile("log_mel_spectrogram.json");
+        outFile << "[";
+        for (uint64_t i = 0; i < out.data.size() - 1; i++) {
+            outFile << out.data[i] << ", ";
+        }
+        outFile << out.data[out.data.size() - 1] << "]";
+        outFile.close();
+    }
+
+    return true;
+}
+
+//
+// mtmd_audio_preprocessor_whisper
+//
+
+void mtmd_audio_preprocessor_whisper::initialize() {
+    cache.fill_sin_cos_table(hparams.audio_n_fft);
+    cache.fill_hann_window(hparams.audio_window_len, true);
+    cache.fill_mel_filterbank_matrix(hparams.n_mel_bins, hparams.audio_n_fft, hparams.audio_sample_rate);
+}
+
+bool mtmd_audio_preprocessor_whisper::preprocess(const float *                 samples,
+                                                 size_t                        n_samples,
+                                                 std::vector<mtmd_audio_mel> & output) {
+    if (n_samples == 0) {
+        // empty audio
+        return false;
+    }
+
+    std::vector<float> smpl;
+    // if input is too short, pad with zeros
+    // this is to avoid potential issues with stage1/2 padding in log_mel_spectrogram
+    // TODO: maybe handle this better
+    size_t min_samples = (size_t) hparams.audio_sample_rate * (hparams.audio_chunk_len + 1);  // +1 second margin
+    if (n_samples < min_samples) {
+        smpl.resize(min_samples, 0.0f);
+        std::memcpy(smpl.data(), samples, n_samples * sizeof(float));
+        samples   = smpl.data();
+        n_samples = smpl.size();
+    }
+
+    filter_params params;
+    params.n_mel            = hparams.n_mel_bins;
+    params.n_fft_bins       = 1 + (hparams.audio_n_fft / 2);
+    params.hann_window_size = hparams.audio_window_len;
+    params.hop_length       = hparams.audio_hop_len;
+    params.sample_rate      = hparams.audio_sample_rate;
+    params.center_padding   = false;
+    params.preemph          = 0.0f;  // disabled
+    params.use_natural_log  = false;
+    params.norm_per_feature = false;
+
+    // make sure the cache is initialized
+    GGML_ASSERT(!cache.sin_vals.empty());
+    GGML_ASSERT(!cache.cos_vals.empty());
+    GGML_ASSERT(!cache.filters.data.empty());
+
+    mtmd_audio_mel out_full;
+    bool           ok = log_mel_spectrogram(samples, n_samples,
+                                            4,  // n_threads
+                                            params, cache, out_full);
+    if (!ok) {
+        return false;
+    }
+
+    // because the cgraph in clip.cpp only accepts 3000 frames each, we need to split the mel
+    // we always expect the mel to have 3000 silent frames at the end
+    if (DEBUG) {
+        printf("output: n_mel = %d, n_len = %d\n", out_full.n_mel, out_full.n_len);
+    }
+    const size_t frames_per_chunk = 3000;
+    GGML_ASSERT((size_t) out_full.n_len > frames_per_chunk);
+    for (size_t off = 0; off < (size_t) out_full.n_len; off += frames_per_chunk) {
+        int n_len = std::min(frames_per_chunk, (size_t) out_full.n_len - off);
+        if ((size_t) n_len < frames_per_chunk) {
+            break;  // last uncomplete chunk will always be a padded chunk, safe to ignore
+        }
+
+        mtmd_audio_mel out_chunk;
+        out_chunk.n_len     = n_len;
+        out_chunk.n_mel     = out_full.n_mel;
+        out_chunk.n_len_org = out_full.n_mel;  // unused
+        out_chunk.data.reserve(out_chunk.n_mel * out_chunk.n_len);
+
+        for (int i = 0; i < out_full.n_mel; i++) {
+            auto src = out_full.data.begin() + i * out_full.n_len + off;
+            out_chunk.data.insert(out_chunk.data.end(), src, src + frames_per_chunk);
+        }
+
+        output.push_back(std::move(out_chunk));
+    }
+
+    return true;
+}
+
+//
+// mtmd_audio_preprocessor_conformer
+//
+
+void mtmd_audio_preprocessor_conformer::initialize() {
+    cache.fill_sin_cos_table(hparams.audio_n_fft);
+    cache.fill_hann_window(hparams.audio_window_len, true);
+    cache.fill_mel_filterbank_matrix(hparams.n_mel_bins, hparams.audio_n_fft, hparams.audio_sample_rate);
+}
+
+bool mtmd_audio_preprocessor_conformer::preprocess(const float *                 samples,
+                                                   size_t                        n_samples,
+                                                   std::vector<mtmd_audio_mel> & output) {
+    // empty audio
+    if (n_samples == 0) {
+        return false;
+    }
+
+    filter_params params;
+    params.n_mel            = hparams.n_mel_bins;
+    params.n_fft_bins       = 1 + (hparams.audio_n_fft / 2);
+    params.hann_window_size = hparams.audio_window_len;
+    params.hop_length       = hparams.audio_hop_len;
+    params.sample_rate      = hparams.audio_sample_rate;
+    params.center_padding   = true;
+    params.preemph          = 0.97f;
+    params.use_natural_log  = true;
+    params.norm_per_feature = true;
+
+    // make sure the cache is initialized
+    GGML_ASSERT(!cache.sin_vals.empty());
+    GGML_ASSERT(!cache.cos_vals.empty());
+    GGML_ASSERT(!cache.filters.data.empty());
+
+    mtmd_audio_mel out_full;
+    bool           ok = log_mel_spectrogram(samples, n_samples,
+                                            4,  // n_threads
+                                            params, cache, out_full);
+    if (!ok) {
+        return false;
+    }
+
+    output.push_back(std::move(out_full));
+    return true;
+}
+
+//
+// mtmd_audio_streaming_istft implementation
+//
+
+mtmd_audio_streaming_istft::mtmd_audio_streaming_istft(int n_fft, int hop_length) :
+    n_fft(n_fft),
+    hop_length(hop_length),
+    n_fft_bins(n_fft / 2 + 1),
+    overlap_buffer(n_fft, 0.0f),
+    window_sum_buffer(n_fft, 0.0f),
+    padding_to_remove((n_fft - hop_length) / 2),
+    ifft_in(n_fft * 2 * 4, 0.0f),  // extra space for recursive IFFT
+    ifft_out(n_fft * 2 * 4, 0.0f) {
+    cache.fill_sin_cos_table(n_fft);
+    cache.fill_hann_window(n_fft, true);
+}
+
+void mtmd_audio_streaming_istft::reset() {
+    std::fill(overlap_buffer.begin(), overlap_buffer.end(), 0.0f);
+    std::fill(window_sum_buffer.begin(), window_sum_buffer.end(), 0.0f);
+    padding_to_remove = (n_fft - hop_length) / 2;
+}
+
+std::vector<float> mtmd_audio_streaming_istft::process_frame(const float * frame_spectrum) {
+    std::vector<float> output(hop_length);
+
+    // copy frequencies
+    for (int j = 0; j < n_fft_bins; j++) {
+        ifft_in[j * 2 + 0] = frame_spectrum[j * 2 + 0];
+        ifft_in[j * 2 + 1] = frame_spectrum[j * 2 + 1];
+    }
+
+    // mirror negative frequencies
+    for (int j = 1; j < n_fft_bins - 1; j++) {
+        int mirror_idx              = n_fft - j;
+        ifft_in[mirror_idx * 2 + 0] = ifft_in[j * 2 + 0];
+        ifft_in[mirror_idx * 2 + 1] = -ifft_in[j * 2 + 1];  // conjugate
+    }
+
+    ifft(cache, ifft_in.data(), n_fft, ifft_out.data());
+
+    // update window sum and overlap buffer
+    for (int j = 0; j < n_fft; j++) {
+        window_sum_buffer[j] += cache.hann_window[j] * cache.hann_window[j];
+        overlap_buffer[j] += ifft_out[j * 2] * cache.hann_window[j];
+    }
+
+    // extract hop_length samples with normalization
+    for (int i = 0; i < hop_length; i++) {
+        if (window_sum_buffer[i] > 1e-8f) {
+            output[i] = overlap_buffer[i] / window_sum_buffer[i];
+        } else {
+            output[i] = overlap_buffer[i];
+        }
+    }
+
+    // shift buffers left by hop_length
+    std::copy(overlap_buffer.begin() + hop_length, overlap_buffer.end(), overlap_buffer.begin());
+    std::fill(overlap_buffer.end() - hop_length, overlap_buffer.end(), 0.0f);
+
+    std::copy(window_sum_buffer.begin() + hop_length, window_sum_buffer.end(), window_sum_buffer.begin());
+    std::fill(window_sum_buffer.end() - hop_length, window_sum_buffer.end(), 0.0f);
+
+    // Remove padding if needed
+    int to_remove = std::min(padding_to_remove, (int) output.size());
+    padding_to_remove -= to_remove;
+    output.erase(output.begin(), output.begin() + to_remove);
+
+    return output;
+}
+
+std::vector<float> mtmd_audio_streaming_istft::flush() {
+    std::vector<float> output;
+
+    // Extract remaining samples from overlap buffer
+    // Continue until we've extracted all meaningful samples
+    int remaining = n_fft - hop_length;
+    while (remaining > 0) {
+        int chunk_size = std::min(remaining, hop_length);
+
+        for (int i = 0; i < chunk_size; i++) {
+            float sample;
+            if (window_sum_buffer[i] > 1e-8f) {
+                sample = overlap_buffer[i] / window_sum_buffer[i];
+            } else {
+                sample = overlap_buffer[i];
+            }
+            output.push_back(sample);
+        }
+
+        // Shift buffers
+        std::copy(overlap_buffer.begin() + chunk_size, overlap_buffer.end(), overlap_buffer.begin());
+        std::fill(overlap_buffer.end() - chunk_size, overlap_buffer.end(), 0.0f);
+
+        std::copy(window_sum_buffer.begin() + chunk_size, window_sum_buffer.end(), window_sum_buffer.begin());
+        std::fill(window_sum_buffer.end() - chunk_size, window_sum_buffer.end(), 0.0f);
+
+        remaining -= chunk_size;
+    }
+
+    return output;
+}

tools/mtmd/mtmd-audio.h

@@ -0,0 +1,113 @@
+#pragma once
+
+#include "ggml.h"
+#include "clip-model.h"
+
+#include <cstdint>
+#include <vector>
+#include <string>
+
+#define MTMD_INTERNAL_HEADER
+
+struct mtmd_audio_mel {
+    int n_len;
+    int n_len_org;
+    int n_mel;
+
+    std::vector<float> data;
+};
+
+struct mtmd_audio_mel_filters {
+    int32_t n_mel;
+    int32_t n_fft;
+
+    std::vector<float> data;
+};
+
+// cache for audio processing, each processor instance owns its own cache
+struct mtmd_audio_cache {
+    std::vector<float> sin_vals;
+    std::vector<float> cos_vals;
+
+    std::vector<float> hann_window;
+
+    mtmd_audio_mel_filters filters;
+
+    void fill_sin_cos_table(int n);
+
+    void fill_hann_window(int length, bool periodic);
+
+    // Build mel filterbank matrix [n_mel × n_fft_bins] at runtime.
+    // n_fft_bins must be (N_fft / 2 + 1). Example: if N_fft=512 -> n_fft_bins=257.
+    void fill_mel_filterbank_matrix(int   n_mel,
+                                    int   n_fft,
+                                    int   sample_rate,               // e.g. 16000
+                                    float fmin             = 0.0f,   // e.g. 0.0
+                                    float fmax             = -1.0f,  // e.g. sr/2; pass -1 for auto
+                                    bool  slaney_area_norm = true,
+                                    float scale = 1.0f  // optional extra scaling
+    );
+};
+
+struct mtmd_audio_preprocessor {
+    const clip_hparams & hparams;
+
+    mtmd_audio_preprocessor(const clip_ctx * ctx): hparams(*clip_get_hparams(ctx)) {}
+
+    virtual ~mtmd_audio_preprocessor() = default;
+    virtual void initialize() = 0; // NOT thread-safe
+    virtual bool preprocess(const float * samples, size_t n_samples, std::vector<mtmd_audio_mel> & output) = 0;
+};
+
+struct mtmd_audio_preprocessor_whisper : mtmd_audio_preprocessor {
+    mtmd_audio_preprocessor_whisper(const clip_ctx * ctx) : mtmd_audio_preprocessor(ctx) {}
+    void initialize() override;
+    bool preprocess(const float * samples, size_t n_samples, std::vector<mtmd_audio_mel> & output) override;
+
+  private:
+    mtmd_audio_cache cache;
+};
+
+struct mtmd_audio_preprocessor_conformer : mtmd_audio_preprocessor {
+    mtmd_audio_preprocessor_conformer(const clip_ctx * ctx) : mtmd_audio_preprocessor(ctx) {}
+    void initialize() override;
+    bool preprocess(const float * samples, size_t n_samples, std::vector<mtmd_audio_mel> & output) override;
+
+  private:
+    mtmd_audio_cache cache;
+};
+
+//
+// streaming ISTFT - converts spectrogram frames back to audio one frame at a time
+//
+struct mtmd_audio_streaming_istft {
+    mtmd_audio_streaming_istft(int n_fft, int hop_length);
+
+    // reset streaming state
+    void reset();
+
+    // process a single STFT frame (streaming)
+    // frame_spectrum: [n_fft_bins x 2] interleaved real/imag
+    // returns: up to hop_length samples
+    std::vector<float> process_frame(const float * frame_spectrum);
+
+    // flush remaining samples at end of stream
+    std::vector<float> flush();
+
+  private:
+    int n_fft;
+    int hop_length;
+    int n_fft_bins;
+
+    // Own cache for output processing
+    mtmd_audio_cache cache;
+
+    // Streaming state
+    std::vector<float> overlap_buffer;
+    std::vector<float> window_sum_buffer;
+    int                padding_to_remove;
+
+    // Working buffers for IFFT
+    std::vector<float> ifft_in;
+    std::vector<float> ifft_out;
+};

tools/mtmd/mtmd-cli.cpp

@@ -0,0 +1,437 @@
+#include "arg.h"
+#include "debug.h"
+#include "log.h"
+#include "common.h"
+#include "sampling.h"
+#include "llama.h"
+#include "ggml.h"
+#include "console.h"
+#include "chat.h"
+#include "mtmd.h"
+#include "mtmd-helper.h"
+
+#include <vector>
+#include <limits.h>
+#include <cinttypes>
+
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
+#include <signal.h>
+#include <unistd.h>
+#elif defined (_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#define NOMINMAX
+#endif
+#include <windows.h>
+#include <signal.h>
+#endif
+
+// volatile, because of signal being an interrupt
+static volatile bool g_is_generating = false;
+static volatile bool g_is_interrupted = false;
+
+/**
+ * Please note that this is NOT a production-ready stuff.
+ * It is a playground for trying multimodal support in llama.cpp.
+ * For contributors: please keep this code simple and easy to understand.
+ */
+
+static void show_additional_info(int /*argc*/, char ** argv) {
+    LOG(
+        "Experimental CLI for multimodal\n\n"
+        "Usage: %s [options] -m <model> --mmproj <mmproj> --image <image> --audio <audio> -p <prompt>\n\n"
+        "  -m and --mmproj are required\n"
+        "  -hf user/repo can replace both -m and --mmproj in most cases\n"
+        "  --image, --audio and -p are optional, if NOT provided, the CLI will run in chat mode\n"
+        "  to disable using GPU for mmproj model, add --no-mmproj-offload\n",
+        argv[0]
+    );
+}
+
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
+static void sigint_handler(int signo) {
+    if (signo == SIGINT) {
+        if (g_is_generating) {
+            g_is_generating = false;
+        } else {
+            console::cleanup();
+            if (g_is_interrupted) {
+                _exit(1);
+            }
+            g_is_interrupted = true;
+        }
+    }
+}
+#endif
+
+struct mtmd_cli_context {
+    mtmd::context_ptr ctx_vision;
+    common_init_result_ptr llama_init;
+
+    llama_model       * model;
+    llama_context     * lctx;
+    const llama_vocab * vocab;
+    common_sampler    * smpl;
+    llama_batch         batch;
+    int                 n_batch;
+
+    mtmd::bitmaps bitmaps;
+
+    // chat template
+    common_chat_templates_ptr tmpls;
+    std::vector<common_chat_msg> chat_history;
+    bool use_jinja = false;
+    // TODO: support for --system-prompt with /clear command
+
+    // support for legacy templates (models not having EOT token)
+    llama_tokens antiprompt_tokens;
+
+    int n_threads    = 1;
+    llama_pos n_past = 0;
+
+    base_callback_data cb_data;
+
+    mtmd_cli_context(common_params & params) : llama_init(common_init_from_params(params)) {
+        model = llama_init->model();
+        lctx = llama_init->context();
+        vocab = llama_model_get_vocab(model);
+        smpl = common_sampler_init(model, params.sampling);
+        n_threads = params.cpuparams.n_threads;
+        batch = llama_batch_init(1, 0, 1); // batch for next token generation
+        n_batch = params.n_batch;
+
+        if (!model || !lctx) {
+            exit(1);
+        }
+
+        if (!llama_model_chat_template(model, nullptr) && params.chat_template.empty()) {
+            LOG_ERR("Model does not have chat template.\n");
+            LOG_ERR("  For old llava models, you may need to use '--chat-template vicuna'\n");
+            LOG_ERR("  For MobileVLM models, use '--chat-template deepseek'\n");
+            LOG_ERR("  For Mistral Small 3.1, use '--chat-template mistral-v7'\n");
+            exit(1);
+        }
+
+        tmpls = common_chat_templates_init(model, params.chat_template);
+        use_jinja = params.use_jinja;
+        chat_history.clear();
+        LOG_INF("%s: chat template example:\n%s\n", __func__, common_chat_format_example(tmpls.get(), params.use_jinja, params.default_template_kwargs).c_str());
+
+        init_vision_context(params);
+
+        // load antiprompt tokens for legacy templates
+        if (params.chat_template == "vicuna") {
+            antiprompt_tokens = common_tokenize(lctx, "ASSISTANT:", false, true);
+        } else if (params.chat_template == "deepseek") {
+            antiprompt_tokens = common_tokenize(lctx, "###", false, true);
+        }
+    }
+
+    ~mtmd_cli_context() {
+        llama_batch_free(batch);
+        common_sampler_free(smpl);
+    }
+
+    void init_vision_context(common_params & params) {
+        const char * clip_path = params.mmproj.path.c_str();
+        mtmd_context_params mparams = mtmd_context_params_default();
+        mparams.use_gpu          = params.mmproj_use_gpu;
+        mparams.print_timings    = true;
+        mparams.n_threads        = params.cpuparams.n_threads;
+        mparams.flash_attn_type  = params.flash_attn_type;
+        mparams.warmup           = params.warmup;
+        mparams.image_min_tokens = params.image_min_tokens;
+        mparams.image_max_tokens = params.image_max_tokens;
+        if (std::getenv("MTMD_DEBUG_GRAPH") != nullptr) {
+            mparams.cb_eval_user_data = &cb_data;
+            mparams.cb_eval = common_debug_cb_eval<false>;
+        }
+        ctx_vision.reset(mtmd_init_from_file(clip_path, model, mparams));
+        if (!ctx_vision.get()) {
+            LOG_ERR("Failed to load vision model from %s\n", clip_path);
+            exit(1);
+        }
+    }
+
+    bool check_antiprompt(const llama_tokens & generated_tokens) {
+        if (antiprompt_tokens.empty() || generated_tokens.size() < antiprompt_tokens.size()) {
+            return false;
+        }
+        return std::equal(
+            generated_tokens.end() - antiprompt_tokens.size(),
+            generated_tokens.end(),
+            antiprompt_tokens.begin()
+        );
+    }
+
+    bool load_media(const std::string & fname) {
+        mtmd::bitmap bmp(mtmd_helper_bitmap_init_from_file(ctx_vision.get(), fname.c_str()));
+        if (!bmp.ptr) {
+            return false;
+        }
+        bitmaps.entries.push_back(std::move(bmp));
+        return true;
+    }
+};
+
+static int generate_response(mtmd_cli_context & ctx, int n_predict) {
+    llama_tokens generated_tokens;
+    for (int i = 0; i < n_predict; i++) {
+        if (i > n_predict || !g_is_generating || g_is_interrupted) {
+            LOG("\n");
+            break;
+        }
+
+        llama_token token_id = common_sampler_sample(ctx.smpl, ctx.lctx, -1);
+        generated_tokens.push_back(token_id);
+        common_sampler_accept(ctx.smpl, token_id, true);
+
+        if (llama_vocab_is_eog(ctx.vocab, token_id) || ctx.check_antiprompt(generated_tokens)) {
+            LOG("\n");
+            break; // end of generation
+        }
+
+        LOG("%s", common_token_to_piece(ctx.lctx, token_id).c_str());
+        fflush(stdout);
+
+        if (g_is_interrupted) {
+            LOG("\n");
+            break;
+        }
+
+        // eval the token
+        common_batch_clear(ctx.batch);
+        common_batch_add(ctx.batch, token_id, ctx.n_past++, {0}, true);
+        if (llama_decode(ctx.lctx, ctx.batch)) {
+            LOG_ERR("failed to decode token\n");
+            return 1;
+        }
+    }
+
+    std::string generated_text = common_detokenize(ctx.lctx, generated_tokens);
+    common_chat_msg msg;
+    msg.role    = "assistant";
+    msg.content = generated_text;
+    ctx.chat_history.push_back(std::move(msg));
+
+    return 0;
+}
+
+static std::string chat_add_and_format(mtmd_cli_context & ctx, common_chat_msg & new_msg) {
+    LOG_DBG("chat_add_and_format: new_msg.role='%s', new_msg.content='%s'\n",
+        new_msg.role.c_str(), new_msg.content.c_str());
+    auto formatted = common_chat_format_single(ctx.tmpls.get(), ctx.chat_history,
+        new_msg, new_msg.role == "user",
+        ctx.use_jinja);
+    ctx.chat_history.push_back(new_msg);
+    return formatted;
+}
+
+static int eval_message(mtmd_cli_context & ctx, common_chat_msg & msg) {
+    bool add_bos = ctx.chat_history.empty();
+    auto formatted_chat = chat_add_and_format(ctx, msg);
+    LOG_DBG("formatted_chat.prompt: %s\n", formatted_chat.c_str());
+
+    mtmd_input_text text;
+    text.text          = formatted_chat.c_str();
+    text.add_special   = add_bos;
+    text.parse_special = true;
+
+    if (g_is_interrupted) return 0;
+
+    mtmd::input_chunks chunks(mtmd_input_chunks_init());
+    auto bitmaps_c_ptr = ctx.bitmaps.c_ptr();
+    int32_t res = mtmd_tokenize(ctx.ctx_vision.get(),
+                        chunks.ptr.get(), // output
+                        &text, // text
+                        bitmaps_c_ptr.data(),
+                        bitmaps_c_ptr.size());
+    if (res != 0) {
+        LOG_ERR("Unable to tokenize prompt, res = %d\n", res);
+        return 1;
+    }
+
+    ctx.bitmaps.entries.clear();
+
+    llama_pos new_n_past;
+    if (mtmd_helper_eval_chunks(ctx.ctx_vision.get(),
+                ctx.lctx, // lctx
+                chunks.ptr.get(), // chunks
+                ctx.n_past, // n_past
+                0, // seq_id
+                ctx.n_batch, // n_batch
+                true, // logits_last
+                &new_n_past)) {
+        LOG_ERR("Unable to eval prompt\n");
+        return 1;
+    }
+
+    ctx.n_past = new_n_past;
+
+    LOG("\n");
+
+    return 0;
+}
+
+int main(int argc, char ** argv) {
+    ggml_time_init();
+
+    common_params params;
+
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_MTMD, show_additional_info)) {
+        return 1;
+    }
+
+    common_init();
+    mtmd_helper_log_set(common_log_default_callback, nullptr);
+
+    if (params.mmproj.path.empty()) {
+        show_additional_info(argc, argv);
+        LOG_ERR("ERR: Missing --mmproj argument\n");
+        return 1;
+    }
+
+    mtmd_cli_context ctx(params);
+    LOG_INF("%s: loading model: %s\n", __func__, params.model.path.c_str());
+
+    bool is_single_turn = !params.prompt.empty() && !params.image.empty();
+
+    int n_predict = params.n_predict < 0 ? INT_MAX : params.n_predict;
+
+    // Ctrl+C handling
+    {
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
+        struct sigaction sigint_action;
+        sigint_action.sa_handler = sigint_handler;
+        sigemptyset (&sigint_action.sa_mask);
+        sigint_action.sa_flags = 0;
+        sigaction(SIGINT, &sigint_action, NULL);
+#elif defined (_WIN32)
+        auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
+            return (ctrl_type == CTRL_C_EVENT) ? (sigint_handler(SIGINT), true) : false;
+        };
+        SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
+#endif
+    }
+
+    if (g_is_interrupted) return 130;
+
+    auto eval_system_prompt_if_present = [&] {
+        if (params.system_prompt.empty()) {
+            return 0;
+        }
+
+        common_chat_msg msg;
+        msg.role = "system";
+        msg.content = params.system_prompt;
+        return eval_message(ctx, msg);
+    };
+
+    LOG_WRN("WARN: This is an experimental CLI for testing multimodal capability.\n");
+    LOG_WRN("      For normal use cases, please use the standard llama-cli\n");
+
+    if (eval_system_prompt_if_present()) {
+        return 1;
+    }
+
+    if (is_single_turn) {
+        g_is_generating = true;
+        if (params.prompt.find(mtmd_default_marker()) == std::string::npos) {
+            for (size_t i = 0; i < params.image.size(); i++) {
+                // most models require the marker before each image
+                // ref: https://github.com/ggml-org/llama.cpp/pull/17616
+                params.prompt = mtmd_default_marker() + params.prompt;
+            }
+        }
+
+        common_chat_msg msg;
+        msg.role = "user";
+        msg.content = params.prompt;
+        for (const auto & image : params.image) {
+            if (!ctx.load_media(image)) {
+                return 1; // error is already printed by libmtmd
+            }
+        }
+        if (eval_message(ctx, msg)) {
+            return 1;
+        }
+        if (!g_is_interrupted && generate_response(ctx, n_predict)) {
+            return 1;
+        }
+
+    } else {
+        LOG("\n Running in chat mode, available commands:");
+        if (mtmd_support_vision(ctx.ctx_vision.get())) {
+            LOG("\n   /image <path>    load an image");
+        }
+        if (mtmd_support_audio(ctx.ctx_vision.get())) {
+            LOG("\n   /audio <path>    load an audio");
+        }
+        LOG("\n   /clear           clear the chat history");
+        LOG("\n   /quit or /exit   exit the program");
+        LOG("\n");
+
+        std::string content;
+
+        while (!g_is_interrupted) {
+            g_is_generating = false;
+            LOG("\n> ");
+            console::set_display(DISPLAY_TYPE_USER_INPUT);
+            std::string line;
+            console::readline(line, false);
+            if (g_is_interrupted) break;
+            console::set_display(DISPLAY_TYPE_RESET);
+            line = string_strip(line);
+            if (line.empty()) {
+                continue;
+            }
+            if (line == "/quit" || line == "/exit") {
+                break;
+            }
+            if (line == "/clear") {
+                ctx.n_past = 0;
+                ctx.chat_history.clear();
+                llama_memory_clear(llama_get_memory(ctx.lctx), true);
+                if (eval_system_prompt_if_present()) {
+                    return 1;
+                }
+                LOG("Chat history cleared\n\n");
+                continue;
+            }
+            g_is_generating = true;
+            bool is_image = line == "/image" || line.find("/image ") == 0;
+            bool is_audio = line == "/audio" || line.find("/audio ") == 0;
+            if (is_image || is_audio) {
+                if (line.size() < 8) {
+                    LOG_ERR("ERR: Missing media filename\n");
+                    continue;
+                }
+                std::string media_path = line.substr(7);
+                if (ctx.load_media(media_path)) {
+                    LOG("%s %s loaded\n", media_path.c_str(), is_image ? "image" : "audio");
+                    content += mtmd_default_marker();
+                }
+                // else, error is already printed by libmtmd
+                continue;
+            } else {
+                content += line;
+            }
+            common_chat_msg msg;
+            msg.role = "user";
+            msg.content = content;
+            int ret = eval_message(ctx, msg);
+            if (ret) {
+                return 1;
+            }
+            if (g_is_interrupted) break;
+            if (generate_response(ctx, n_predict)) {
+                return 1;
+            }
+            content.clear();
+        }
+    }
+    if (g_is_interrupted) LOG("\nInterrupted by user\n");
+    LOG("\n\n");
+    llama_perf_context_print(ctx.lctx);
+    return g_is_interrupted ? 130 : 0;
+}

tools/mtmd/mtmd-helper.cpp

@@ -0,0 +1,521 @@
+// fix problem with std::min and std::max
+#if defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#   define NOMINMAX
+#endif
+#include <windows.h>
+#endif
+
+#include "mtmd.h"
+#include "mtmd-helper.h"
+#include "llama.h"
+
+#include <algorithm>
+#include <cinttypes>
+#include <vector>
+
+//#define MTMD_AUDIO_DEBUG
+
+#define MINIAUDIO_IMPLEMENTATION
+#ifndef MTMD_AUDIO_DEBUG
+#   define MA_NO_ENCODING
+#endif
+#define MA_NO_DEVICE_IO
+#define MA_NO_RESOURCE_MANAGER
+#define MA_NO_NODE_GRAPH
+#define MA_NO_ENGINE
+#define MA_NO_GENERATION
+#define MA_API static
+#include "miniaudio/miniaudio.h"
+
+#define STB_IMAGE_IMPLEMENTATION
+#include "stb/stb_image.h"
+
+#ifdef MTMD_INTERNAL_HEADER
+#error "mtmd-helper is a public library outside of mtmd. it must not include internal headers"
+#endif
+
+//
+// internal logging functions
+//
+
+struct mtmd_helper_logger {
+    ggml_log_callback default_callback = [](ggml_log_level level, const char * text, void * user_data) {
+        (void) level;
+        (void) user_data;
+        fputs(text, stderr);
+        fflush(stderr);
+    };
+
+    ggml_log_callback log_callback = default_callback;
+    void * log_callback_user_data;
+
+    void log_v(enum ggml_log_level level, const char * format, va_list args) {
+        if (format == NULL) {
+            return;
+        }
+        va_list args_copy;
+        va_copy(args_copy, args);
+        char buffer[128];
+        int len = vsnprintf(buffer, 128, format, args);
+        if (len < 128) {
+            log_callback(level, buffer, log_callback_user_data);
+        } else {
+            char * buffer2 = (char *) calloc(len + 1, sizeof(char));
+            vsnprintf(buffer2, len + 1, format, args_copy);
+            buffer2[len] = 0;
+            log_callback(level, buffer2, log_callback_user_data);
+            free(buffer2);
+        }
+        va_end(args_copy);
+    }
+
+    void log(enum ggml_log_level level, const char * format, ...) {
+        va_list args;
+        va_start(args, format);
+        log_v(level, format, args);
+        va_end(args);
+    }
+} g_logger;
+
+#define LOG_INF(...) g_logger.log(GGML_LOG_LEVEL_INFO,  __VA_ARGS__)
+#define LOG_WRN(...) g_logger.log(GGML_LOG_LEVEL_WARN,  __VA_ARGS__)
+#define LOG_ERR(...) g_logger.log(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
+
+void mtmd_helper_log_set(ggml_log_callback log_callback, void * user_data) {
+    if (log_callback == nullptr) {
+        log_callback = g_logger.default_callback;
+    }
+    g_logger.log_callback = log_callback;
+    g_logger.log_callback_user_data = user_data;
+    mtmd_log_set(log_callback, user_data);
+}
+
+//
+// helper functions
+//
+
+size_t mtmd_helper_get_n_tokens(const mtmd_input_chunks * chunks) {
+    size_t n_tokens = 0;
+    for (size_t i = 0; i < mtmd_input_chunks_size(chunks); i++) {
+        auto chunk = mtmd_input_chunks_get(chunks, i);
+        n_tokens += mtmd_input_chunk_get_n_tokens(chunk);
+    }
+    return n_tokens;
+}
+
+llama_pos mtmd_helper_get_n_pos(const mtmd_input_chunks * chunks) {
+    llama_pos n_pos = 0;
+    for (size_t i = 0; i < mtmd_input_chunks_size(chunks); i++) {
+        auto chunk = mtmd_input_chunks_get(chunks, i);
+        n_pos += mtmd_input_chunk_get_n_pos(chunk);
+    }
+    return n_pos;
+}
+
+// helper struct to make working with embd batch easier
+// note: this will be removed after llama_batch_ext refactoring
+struct decode_embd_batch {
+    int n_pos_per_embd;
+    int n_mmproj_embd;
+    std::vector<llama_pos>      pos;
+    std::vector<llama_pos>      pos_view; // used by mrope
+    std::vector<int32_t>        n_seq_id;
+    std::vector<llama_seq_id>   seq_id_0;
+    std::vector<llama_seq_id *> seq_ids;
+    std::vector<int8_t>         logits;
+    llama_batch batch;
+    decode_embd_batch(float * embd, int32_t n_tokens, int n_pos_per_embd, int n_mmproj_embd) : n_pos_per_embd(n_pos_per_embd), n_mmproj_embd(n_mmproj_embd) {
+        pos     .resize(n_tokens * n_pos_per_embd);
+        n_seq_id.resize(n_tokens);
+        seq_ids .resize(n_tokens + 1);
+        logits  .resize(n_tokens);
+        seq_id_0.resize(1);
+        seq_ids [n_tokens] = nullptr;
+        batch = {
+            /*n_tokens       =*/ n_tokens,
+            /*tokens         =*/ nullptr,
+            /*embd           =*/ embd,
+            /*pos            =*/ pos.data(),
+            /*n_seq_id       =*/ n_seq_id.data(),
+            /*seq_id         =*/ seq_ids.data(),
+            /*logits         =*/ logits.data(),
+        };
+    }
+
+    void set_position_normal(llama_pos pos_0, llama_seq_id seq_id) {
+        seq_id_0[0] = seq_id;
+        for (int i = 0; i < batch.n_tokens; i++) {
+            batch.pos     [i] = pos_0 + i;
+            batch.n_seq_id[i] = 1;
+            batch.seq_id  [i] = seq_id_0.data();
+            batch.logits  [i] = false;
+        }
+    }
+
+    // M-RoPE for image
+    void set_position_mrope_2d(llama_pos pos_0, int nx, int ny, llama_seq_id seq_id) {
+        GGML_ASSERT(n_pos_per_embd == 4);
+        seq_id_0[0] = seq_id;
+        for (int y = 0; y < ny; y++) {
+            for (int x = 0; x < nx; x++) {
+                int i = y * nx + x;
+                pos[i                     ] = pos_0;
+                pos[i + batch.n_tokens    ] = pos_0 + y;
+                pos[i + batch.n_tokens * 2] = pos_0 + x;
+                pos[i + batch.n_tokens * 3] = 0; // last pos dim is unused
+            }
+        }
+        for (int i = 0; i < batch.n_tokens; i++) {
+            batch.n_seq_id[i] = 1;
+            batch.seq_id  [i] = seq_id_0.data();
+            batch.logits  [i] = false;
+        }
+    }
+
+    // M-RoPE for audio
+    void set_position_mrope_1d(llama_pos pos_0, llama_seq_id seq_id) {
+        GGML_ASSERT(n_pos_per_embd == 4);
+        seq_id_0[0] = seq_id;
+        for (int i = 0; i < batch.n_tokens; i++) {
+            pos[i                     ] = pos_0 + i;
+            pos[i + batch.n_tokens    ] = pos_0 + i;
+            pos[i + batch.n_tokens * 2] = pos_0 + i;
+            pos[i + batch.n_tokens * 3] = 0; // last pos dim is unused
+        }
+        for (int i = 0; i < batch.n_tokens; i++) {
+            batch.n_seq_id[i] = 1;
+            batch.seq_id  [i] = seq_id_0.data();
+            batch.logits  [i] = false;
+        }
+    }
+
+    llama_batch get_view(int offset, int n_tokens) {
+        llama_pos * pos_ptr;
+        pos_view.clear();
+        pos_view.reserve(n_tokens * n_pos_per_embd);
+        if (n_pos_per_embd > 1) {
+            // mrope
+            // for example, with layout of src: 1234...1234...1234...1234...
+            //       offset 2 will give us dst: 34...34...34...34...
+            for (int i = 0; i < n_pos_per_embd; i++) {
+                // assume n_tokens is less than or equal to batch.n_tokens
+                // batch.n_tokens is number of **total** tokens
+                // n_tokens is number of viewed token
+                size_t src_idx = i * batch.n_tokens + offset;
+                pos_view.insert(pos_view.end(),
+                    pos.data() + src_idx,
+                    pos.data() + src_idx + n_tokens);
+            }
+            pos_ptr = pos_view.data();
+        } else {
+            // normal
+            pos_ptr = pos.data() + offset;
+        }
+        return {
+            /*n_tokens       =*/ n_tokens,
+            /*tokens         =*/ nullptr,
+            /*embd           =*/ batch.embd     + offset * n_mmproj_embd,
+            /*pos            =*/ pos_ptr,
+            /*n_seq_id       =*/ batch.n_seq_id + offset,
+            /*seq_id         =*/ batch.seq_id   + offset,
+            /*logits         =*/ batch.logits   + offset,
+        };
+    }
+};
+
+// Helper function for decoding an image whose embeddings have already been calculated
+int32_t mtmd_helper_decode_image_chunk(
+        mtmd_context * ctx,
+        struct llama_context * lctx,
+        const mtmd_input_chunk * chunk,
+        float * encoded_embd,
+        llama_pos n_past,
+        llama_seq_id seq_id,
+        int32_t n_batch,
+        llama_pos * new_n_past) {
+    auto chunk_type = mtmd_input_chunk_get_type(chunk);
+    const char * name = chunk_type == MTMD_INPUT_CHUNK_TYPE_IMAGE ? "image" : "audio";
+    if (chunk_type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
+        LOG_ERR("failed to decode chunk: input chunk not of image/audio type\n");
+        return -1;
+    }
+
+    const llama_model * model = llama_get_model(lctx);
+    int n_mmproj_embd = llama_model_n_embd_inp(model);
+    int n_pos_per_embd = mtmd_decode_use_mrope(ctx) ? 4 : 1;
+
+    int32_t n_tokens = mtmd_input_chunk_get_n_tokens(chunk);
+    int32_t i_batch = 0;
+    int32_t n_img_batches = GGML_PAD(n_tokens, n_batch) / n_batch;
+    decode_embd_batch batch_embd(encoded_embd, n_tokens, n_pos_per_embd, n_mmproj_embd);
+
+    if (mtmd_decode_use_mrope(ctx)) {
+        if (chunk_type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
+            const auto image_tokens = mtmd_input_chunk_get_tokens_image(chunk);
+            if (!image_tokens) {
+                LOG_ERR("failed to decode chunk: image tokens are null\n");
+                return -1;
+            }
+            const int nx = mtmd_image_tokens_get_nx(image_tokens);
+            const int ny = mtmd_image_tokens_get_ny(image_tokens);
+            batch_embd.set_position_mrope_2d(n_past, nx, ny, seq_id);
+        } else if (chunk_type == MTMD_INPUT_CHUNK_TYPE_AUDIO) {
+            batch_embd.set_position_mrope_1d(n_past, seq_id);
+        } else {
+            GGML_ABORT("invalid chunk type for M-RoPE");
+        }
+    } else {
+        batch_embd.set_position_normal(n_past, seq_id);
+    }
+
+    if (mtmd_decode_use_non_causal(ctx)) {
+        llama_set_causal_attn(lctx, false);
+        // TODO @ngxson : need to make sure only one image is processed at a time, and n_ubatch must be enough to hold the image
+    }
+
+    while (i_batch < n_img_batches) { // split into batches
+        int pos_offset = i_batch*n_batch;
+        int n_tokens_batch = std::min(n_batch, n_tokens - pos_offset);
+        llama_batch batch_embd_view = batch_embd.get_view(pos_offset, n_tokens_batch);
+
+        LOG_INF("decoding %s batch %d/%d, n_tokens_batch = %d\n", name, i_batch+1, n_img_batches, n_tokens_batch);
+
+        int64_t t1 = ggml_time_ms();
+        int32_t ret = llama_decode(lctx, batch_embd_view);
+        if (ret != 0) {
+            LOG_ERR("failed to decode %s\n", name);
+            llama_set_causal_attn(lctx, true); // restore causal attn
+            return ret;
+        }
+
+        LOG_INF("%s decoded (batch %d/%d) in %" PRId64 " ms\n", name, i_batch+1, n_img_batches, ggml_time_ms() - t1);
+
+        i_batch++;
+    }
+
+    n_past += mtmd_input_chunk_get_n_pos(chunk);
+    *new_n_past = n_past;
+
+    if (mtmd_decode_use_non_causal(ctx)) {
+        llama_set_causal_attn(lctx, true);
+    }
+    return 0;
+}
+
+int32_t mtmd_helper_eval_chunk_single(mtmd_context * ctx,
+        struct llama_context * lctx,
+        const mtmd_input_chunk * chunk,
+        llama_pos n_past,
+        llama_seq_id seq_id,
+        int32_t n_batch,
+        bool logits_last,
+        llama_pos * new_n_past) {
+    int32_t ret;
+    llama_batch text_batch = llama_batch_init(n_batch, 0, 1);
+    auto chunk_type = mtmd_input_chunk_get_type(chunk);
+
+    if (chunk_type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
+        size_t n_tokens;
+        const auto tokens = mtmd_input_chunk_get_tokens_text(chunk, &n_tokens);
+        // LOG_INF("decoding text chunk, n_tokens = %zu\n", n_tokens);
+        size_t i = 0;
+        while (i < n_tokens) { // split into batches
+            text_batch.n_tokens = 0; // clear the batch
+            for (; i < n_tokens && text_batch.n_tokens < n_batch; i++) {
+                int32_t j = text_batch.n_tokens;
+                text_batch.token   [j]    = tokens[i];
+                text_batch.pos     [j]    = n_past++;
+                text_batch.n_seq_id[j]    = 1;
+                text_batch.seq_id  [j][0] = seq_id;
+                text_batch.logits  [j]    = false;
+
+                text_batch.n_tokens++;
+            }
+            bool is_last_token = (i == n_tokens);
+            if (logits_last && is_last_token) {
+                text_batch.logits[text_batch.n_tokens - 1] = true;
+            }
+            ret = llama_decode(lctx, text_batch);
+            if (ret != 0) {
+                LOG_ERR("failed to decode text\n");
+                llama_batch_free(text_batch);
+                return ret;
+            }
+            *new_n_past += text_batch.n_tokens;
+        }
+
+    } else if (chunk_type == MTMD_INPUT_CHUNK_TYPE_IMAGE || chunk_type == MTMD_INPUT_CHUNK_TYPE_AUDIO) {
+        const char * name = chunk_type == MTMD_INPUT_CHUNK_TYPE_IMAGE ? "image" : "audio";
+        int64_t t0 = ggml_time_ms();
+
+        LOG_INF("encoding %s slice...\n", name);
+
+        ret = mtmd_encode_chunk(ctx, chunk);
+        if (ret != 0) {
+            LOG_ERR("failed to encode %s slice\n", name);
+            llama_batch_free(text_batch);
+            return ret;
+        }
+
+        LOG_INF("%s slice encoded in %" PRId64 " ms\n", name, ggml_time_ms() - t0);
+
+        float * embd = mtmd_get_output_embd(ctx);
+        ret = mtmd_helper_decode_image_chunk(ctx, lctx, chunk, embd, n_past, seq_id, n_batch, new_n_past);
+        if (ret != 0) {
+            LOG_ERR("failed to decode %s\n", name);
+            llama_batch_free(text_batch);
+            return ret;
+        }
+    } else {
+        GGML_ABORT("chunk type not supported");
+    }
+
+    llama_batch_free(text_batch);
+    return 0;
+}
+
+int32_t mtmd_helper_eval_chunks(mtmd_context * ctx,
+                                struct llama_context * lctx,
+                                const mtmd_input_chunks * chunks,
+                                llama_pos n_past,
+                                llama_seq_id seq_id,
+                                int32_t n_batch,
+                                bool logits_last,
+                                llama_pos * new_n_past) {
+    size_t n_chunks = mtmd_input_chunks_size(chunks);
+    if (n_chunks == 0) {
+        LOG_WRN("no chunks to eval\n");
+        return 0;
+    }
+
+    for (size_t i = 0; i < n_chunks; i++) {
+        bool chunk_logits_last = (i == n_chunks - 1) && logits_last;
+        auto chunk = mtmd_input_chunks_get(chunks, i);
+
+        int32_t res = mtmd_helper_eval_chunk_single(ctx, lctx, chunk, n_past, seq_id, n_batch, chunk_logits_last, &n_past);
+        if (res != 0) {
+            LOG_ERR("failed to eval chunk %zu\n", i);
+            return res;
+        }
+        *new_n_past = n_past;
+    }
+
+    return 0;
+}
+
+namespace audio_helpers {
+
+static bool is_audio_file(const char * buf, size_t len) {
+    if (len < 12) {
+        return false;
+    }
+
+    // RIFF ref: https://en.wikipedia.org/wiki/Resource_Interchange_File_Format
+    // WAV ref: https://www.mmsp.ece.mcgill.ca/Documents/AudioFormats/WAVE/WAVE.html
+    bool is_wav = memcmp(buf, "RIFF", 4) == 0 && memcmp(buf + 8, "WAVE", 4) == 0;
+    bool is_mp3 = len >= 3 && (
+        memcmp(buf, "ID3", 3) == 0 ||
+        // Check for MPEG sync word (simplified check)
+        ((unsigned char)buf[0] == 0xFF && ((unsigned char)buf[1] & 0xE0) == 0xE0)
+    );
+    bool is_flac = memcmp(buf, "fLaC", 4) == 0;
+
+    return is_wav || is_mp3 || is_flac;
+}
+
+// returns true if the buffer is a valid audio file
+static bool decode_audio_from_buf(const unsigned char * buf_in, size_t len, int target_sampler_rate, std::vector<float> & pcmf32_mono) {
+    ma_result result;
+    const int channels = 1;
+    ma_decoder_config decoder_config = ma_decoder_config_init(ma_format_f32, channels, target_sampler_rate);
+    ma_decoder decoder;
+
+    result = ma_decoder_init_memory(buf_in, len, &decoder_config, &decoder);
+    if (result != MA_SUCCESS) {
+        return false;
+    }
+
+    ma_uint64 frame_count;
+    ma_uint64 frames_read;
+    result = ma_decoder_get_length_in_pcm_frames(&decoder, &frame_count);
+    if (result != MA_SUCCESS) {
+        ma_decoder_uninit(&decoder);
+        return false;
+    }
+
+    pcmf32_mono.resize(frame_count);
+    result = ma_decoder_read_pcm_frames(&decoder, pcmf32_mono.data(), frame_count, &frames_read);
+    if (result != MA_SUCCESS) {
+        ma_decoder_uninit(&decoder);
+        return false;
+    }
+
+#ifdef MTMD_AUDIO_DEBUG
+    // save audio to wav file
+    ma_encoder_config config = ma_encoder_config_init(ma_encoding_format_wav, ma_format_f32, 1, target_sampler_rate);
+    ma_encoder encoder;
+    ma_encoder_init_file("output.wav", &config, &encoder);
+    ma_encoder_write_pcm_frames(&encoder, pcmf32_mono.data(), pcmf32_mono.size(), &frames_read);
+    ma_encoder_uninit(&encoder);
+#endif
+
+    ma_decoder_uninit(&decoder);
+    return true;
+}
+
+} // namespace audio_helpers
+
+mtmd_bitmap * mtmd_helper_bitmap_init_from_buf(mtmd_context * ctx, const unsigned char * buf, size_t len) {
+    if (audio_helpers::is_audio_file((const char *)buf, len)) {
+        std::vector<float> pcmf32;
+        int bitrate = mtmd_get_audio_bitrate(ctx);
+        if (bitrate < 0) {
+            LOG_ERR("This model does not support audio input\n");
+            return nullptr;
+        }
+        if (!audio_helpers::decode_audio_from_buf(buf, len, bitrate, pcmf32)) {
+            LOG_ERR("Unable to read WAV audio file from buffer\n");
+            return nullptr;
+        }
+        return mtmd_bitmap_init_from_audio(pcmf32.size(), pcmf32.data());
+    }
+
+    // otherwise, we assume it's an image
+    mtmd_bitmap * result = nullptr;
+    {
+        int nx, ny, nc;
+        auto * data = stbi_load_from_memory(buf, len, &nx, &ny, &nc, 3);
+        if (!data) {
+            LOG_ERR("%s: failed to decode image bytes\n", __func__);
+            return nullptr;
+        }
+        result = mtmd_bitmap_init(nx, ny, data);
+        stbi_image_free(data);
+    }
+    return result;
+}
+
+mtmd_bitmap * mtmd_helper_bitmap_init_from_file(mtmd_context * ctx, const char * fname) {
+    std::vector<unsigned char> buf;
+    FILE * f = fopen(fname, "rb");
+    if (!f) {
+        LOG_ERR("Unable to open file %s: %s\n", fname, strerror(errno));
+        return nullptr;
+    }
+
+    fseek(f, 0, SEEK_END);
+    long file_size = ftell(f);
+    fseek(f, 0, SEEK_SET);
+    buf.resize(file_size);
+
+    size_t n_read = fread(buf.data(), 1, file_size, f);
+    fclose(f);
+    if (n_read != (size_t)file_size) {
+        LOG_ERR("Failed to read entire file %s", fname);
+        return nullptr;
+    }
+
+    return mtmd_helper_bitmap_init_from_buf(ctx, buf.data(), buf.size());
+}

tools/mtmd/mtmd-helper.h

@@ -0,0 +1,96 @@
+#ifndef MTMD_HELPER_H
+#define MTMD_HELPER_H
+
+#include "ggml.h"
+#include "llama.h"
+#include "mtmd.h"
+
+#include <stddef.h>
+#include <stdint.h>
+#include <stdbool.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+//
+// libmtmd helper functions
+//
+// Please note that these helpers are not guaranteed to be stable.
+// BREAKING CHANGES are expected.
+//
+
+// Set callback for all future logging events.
+// If this is not called, or NULL is supplied, everything is output on stderr.
+// Note: this also call mtmd_log_set() internally
+MTMD_API void mtmd_helper_log_set(ggml_log_callback log_callback, void * user_data);
+
+// helper function to construct a mtmd_bitmap from a file
+// it calls mtmd_helper_bitmap_init_from_buf() internally
+// returns nullptr on failure
+// this function is thread-safe
+MTMD_API mtmd_bitmap * mtmd_helper_bitmap_init_from_file(mtmd_context * ctx, const char * fname);
+
+// helper function to construct a mtmd_bitmap from a buffer containing a file
+// supported formats:
+//     image: formats supported by stb_image: jpg, png, bmp, gif, etc.
+//     audio: formats supported by miniaudio: wav, mp3, flac
+// note: audio files will be auto-detected based on magic bytes
+// returns nullptr on failure
+// this function is thread-safe
+MTMD_API mtmd_bitmap * mtmd_helper_bitmap_init_from_buf(mtmd_context * ctx, const unsigned char * buf, size_t len);
+
+// helper to count the total number of tokens from a list of chunks, useful to keep track of KV cache
+MTMD_API size_t mtmd_helper_get_n_tokens(const mtmd_input_chunks * chunks);
+
+// helper to count the total position of tokens from a list of chunks, useful to keep track of n_past
+// normally, n_pos is equal to n_tokens, but for M-RoPE it is different
+MTMD_API llama_pos mtmd_helper_get_n_pos(const mtmd_input_chunks * chunks);
+
+// helper function that automatically:
+// 1. run llama_decode() on text chunks
+// 2. run mtmd_encode() on image chunks, then mtmd_get_output_embd() and then llama_decode()
+// if any of the mtmd_encode() or llama_decode() calls return non-zero, stop and forward the error
+// otherwise, returns 0 on success
+// this function is NOT thread-safe
+MTMD_API int32_t mtmd_helper_eval_chunks(mtmd_context * ctx,
+                                         struct llama_context * lctx,
+                                         const mtmd_input_chunks * chunks,
+                                         llama_pos n_past,
+                                         llama_seq_id seq_id,
+                                         int32_t n_batch,
+                                         bool logits_last,
+                                         llama_pos * new_n_past);
+
+// works like mtmd_helper_eval_chunks(), but only for a single chunk
+// this function is NOT thread-safe
+MTMD_API int32_t mtmd_helper_eval_chunk_single(mtmd_context * ctx,
+                                               struct llama_context * lctx,
+                                               const mtmd_input_chunk * chunk,
+                                               llama_pos n_past,
+                                               llama_seq_id seq_id,
+                                               int32_t n_batch,
+                                               bool logits_last,
+                                               llama_pos * new_n_past);
+
+// helper function to decode an image whose embeddings have already been calculated
+// this helper will handle batching and pre/post decoding setup (for ex. gemma 3 requires non-causal attention)
+// ret 0 on success, -1 on chunk not being a valid image chunk, 1 on decode failure
+MTMD_API int32_t mtmd_helper_decode_image_chunk(mtmd_context * ctx,
+                                                struct llama_context * lctx,
+                                                const mtmd_input_chunk * chunk,
+                                                float * encoded_embd,
+                                                llama_pos n_past,
+                                                llama_seq_id seq_id,
+                                                int32_t n_batch,
+                                                llama_pos * new_n_past);
+
+#ifdef __cplusplus
+} // extern "C"
+#endif
+
+//
+// C++ wrappers
+//
+
+#endif

tools/mtmd/mtmd.h

@@ -0,0 +1,319 @@
+#ifndef MTMD_H
+#define MTMD_H
+
+#include "ggml.h"
+#include "llama.h"
+
+#include <stddef.h>
+#include <stdint.h>
+#include <stdbool.h>
+
+#ifdef __cplusplus
+#include <string>
+#include <vector>
+#include <cinttypes>
+#include <memory>
+#endif
+
+/**
+ * libmtmd: A library for multimodal support in llama.cpp.
+ *
+ * WARNING: This API is experimental and subject to many BREAKING CHANGES.
+ *          Issues related to API usage may receive lower priority support.
+ *
+ * For the usage, see an example in mtmd-cli.cpp
+ *
+ * For contributors:
+ * - Make sure the C API is aligned with the libllama C API (as in llama.h)
+ * - Do not include model name (e.g., qwen, gemma) in the API, use generic terms instead
+ * - Keep the API minimal, do not expose internal details unless necessary
+ *
+ * IMPORTANT: The mtmd module does NOT accept pull requests that are fully or predominantly AI-generated.
+ * We encourage human contributors to ensure the quality and reliability of the codebase.
+ */
+
+#ifdef LLAMA_SHARED
+#    if defined(_WIN32) && !defined(__MINGW32__)
+#        ifdef LLAMA_BUILD
+#            define MTMD_API __declspec(dllexport)
+#        else
+#            define MTMD_API __declspec(dllimport)
+#        endif
+#    else
+#        define MTMD_API __attribute__ ((visibility ("default")))
+#    endif
+#else
+#    define MTMD_API
+#endif
+
+// deprecated marker, use mtmd_default_marker() instead
+#define MTMD_DEFAULT_IMAGE_MARKER "<__image__>"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+enum mtmd_input_chunk_type {
+    MTMD_INPUT_CHUNK_TYPE_TEXT,
+    MTMD_INPUT_CHUNK_TYPE_IMAGE,
+    MTMD_INPUT_CHUNK_TYPE_AUDIO,
+};
+
+// opaque types
+struct mtmd_context;
+struct mtmd_bitmap;
+struct mtmd_image_tokens;
+struct mtmd_input_chunk;
+struct mtmd_input_chunks;
+
+struct mtmd_input_text {
+    const char * text;
+    bool add_special;
+    bool parse_special;
+};
+
+//
+// C API
+//
+
+typedef struct mtmd_context      mtmd_context;
+typedef struct mtmd_bitmap       mtmd_bitmap;
+typedef struct mtmd_image_tokens mtmd_image_tokens;
+typedef struct mtmd_input_chunk  mtmd_input_chunk;
+typedef struct mtmd_input_chunks mtmd_input_chunks;
+typedef struct mtmd_input_text   mtmd_input_text;
+
+struct mtmd_context_params {
+    bool use_gpu;
+    bool print_timings;
+    int n_threads;
+    const char * image_marker; // deprecated, use media_marker instead
+    const char * media_marker;
+    enum llama_flash_attn_type flash_attn_type;
+    bool warmup; // whether to run a warmup encode pass after initialization
+
+    // limit number of image tokens, only for vision models with dynamic resolution
+    int image_min_tokens; // minimum number of tokens for image input (default: read from metadata)
+    int image_max_tokens; // maximum number of tokens for image input (default: read from metadata)
+
+    // callback function passed over to mtmd proper
+    ggml_backend_sched_eval_callback cb_eval;
+    void * cb_eval_user_data;
+};
+
+MTMD_API const char * mtmd_default_marker(void);
+
+MTMD_API struct mtmd_context_params mtmd_context_params_default(void);
+
+// initialize the mtmd context
+// return nullptr on failure
+MTMD_API mtmd_context * mtmd_init_from_file(const char * mmproj_fname,
+                                            const struct llama_model * text_model,
+                                            const struct mtmd_context_params ctx_params);
+
+MTMD_API void mtmd_free(mtmd_context * ctx);
+
+// whether we need to set non-causal mask before llama_decode
+MTMD_API bool mtmd_decode_use_non_causal(mtmd_context * ctx);
+
+// whether the current model use M-RoPE for llama_decode
+MTMD_API bool mtmd_decode_use_mrope(mtmd_context * ctx);
+
+// whether the current model supports vision input
+MTMD_API bool mtmd_support_vision(mtmd_context * ctx);
+
+// whether the current model supports audio input
+MTMD_API bool mtmd_support_audio(mtmd_context * ctx);
+
+// get audio bitrate in Hz, for example 16000 for Whisper
+// return -1 if audio is not supported
+MTMD_API int mtmd_get_audio_bitrate(mtmd_context * ctx);
+
+// mtmd_bitmap
+//
+// if bitmap is image:
+//     length of data must be nx * ny * 3
+//     the data is in RGBRGBRGB... format
+// if bitmap is audio:
+//     length of data must be n_samples * sizeof(float)
+//     the data is in float format (PCM F32)
+MTMD_API mtmd_bitmap *         mtmd_bitmap_init           (uint32_t nx, uint32_t ny, const unsigned char * data);
+MTMD_API mtmd_bitmap *         mtmd_bitmap_init_from_audio(size_t n_samples,         const float         * data);
+MTMD_API uint32_t              mtmd_bitmap_get_nx     (const mtmd_bitmap * bitmap);
+MTMD_API uint32_t              mtmd_bitmap_get_ny     (const mtmd_bitmap * bitmap);
+MTMD_API const unsigned char * mtmd_bitmap_get_data   (const mtmd_bitmap * bitmap);
+MTMD_API size_t                mtmd_bitmap_get_n_bytes(const mtmd_bitmap * bitmap);
+MTMD_API bool                  mtmd_bitmap_is_audio   (const mtmd_bitmap * bitmap);
+MTMD_API void                  mtmd_bitmap_free       (mtmd_bitmap * bitmap);
+// bitmap ID is optional, but useful for KV cache tracking
+// these getters/setters are dedicated functions, so you can for example calculate the hash of the image based on mtmd_bitmap_get_data()
+MTMD_API const char * mtmd_bitmap_get_id(const mtmd_bitmap * bitmap);
+MTMD_API void         mtmd_bitmap_set_id(mtmd_bitmap * bitmap, const char * id);
+
+
+// mtmd_input_chunks
+//
+// this is simply a list of mtmd_input_chunk
+// the elements can only be populated via mtmd_tokenize()
+MTMD_API mtmd_input_chunks *      mtmd_input_chunks_init(void);
+MTMD_API size_t                   mtmd_input_chunks_size(const mtmd_input_chunks * chunks);
+MTMD_API const mtmd_input_chunk * mtmd_input_chunks_get (const mtmd_input_chunks * chunks, size_t idx);
+MTMD_API void                     mtmd_input_chunks_free(mtmd_input_chunks * chunks);
+
+// mtmd_input_chunk
+//
+// the instance will be constructed via mtmd_tokenize()
+// it will be freed along with mtmd_input_chunks
+MTMD_API enum mtmd_input_chunk_type mtmd_input_chunk_get_type        (const mtmd_input_chunk * chunk);
+MTMD_API const llama_token *        mtmd_input_chunk_get_tokens_text (const mtmd_input_chunk * chunk, size_t * n_tokens_output);
+MTMD_API const mtmd_image_tokens *  mtmd_input_chunk_get_tokens_image(const mtmd_input_chunk * chunk);
+MTMD_API size_t                     mtmd_input_chunk_get_n_tokens    (const mtmd_input_chunk * chunk);
+// returns nullptr for ID on text chunk
+MTMD_API const char *               mtmd_input_chunk_get_id          (const mtmd_input_chunk * chunk);
+// number of temporal positions (equals to max(t,h,w) for M-RoPE; equals to n_tokens otherwise)
+MTMD_API llama_pos                  mtmd_input_chunk_get_n_pos       (const mtmd_input_chunk * chunk);
+
+// in case you want to use custom logic to handle the chunk (i.e. KV cache management)
+// you can move the chunk ownership to your own code by copying it
+// remember to free the chunk when you are done with it
+MTMD_API mtmd_input_chunk * mtmd_input_chunk_copy(const mtmd_input_chunk * chunk);
+MTMD_API void               mtmd_input_chunk_free(mtmd_input_chunk * chunk);
+
+
+// mtmd_image_tokens
+//
+// the instance will be constructed via mtmd_tokenize()
+// it will be freed along with mtmd_input_chunk
+MTMD_API size_t       mtmd_image_tokens_get_n_tokens(const mtmd_image_tokens * image_tokens); // TODO: deprecate
+MTMD_API size_t       mtmd_image_tokens_get_nx      (const mtmd_image_tokens * image_tokens);
+MTMD_API size_t       mtmd_image_tokens_get_ny      (const mtmd_image_tokens * image_tokens);
+MTMD_API const char * mtmd_image_tokens_get_id      (const mtmd_image_tokens * image_tokens); // TODO: deprecate
+// number of temporal positions (equals to max(t,h,w) for M-RoPE; equals to n_tokens otherwise)
+MTMD_API llama_pos    mtmd_image_tokens_get_n_pos   (const mtmd_image_tokens * image_tokens); // TODO: deprecate
+
+// tokenize an input text prompt and a list of bitmaps (images/audio)
+// the prompt must have the input image marker (default: "<__media__>") in it
+// the default marker is defined by mtmd_default_marker()
+// the marker will be replaced with the image/audio chunk
+// for example:
+//   "here is an image: <__media__>\ndescribe it in detail."
+//   this will gives 3 chunks:
+//   1. "here is an image: <start_of_image>"
+//   2. (image/audio tokens)
+//   3. "<end_of_image>\ndescribe it in detail."
+// number of bitmaps must be equal to the number of markers in the prompt
+// this function is thread-safe (shared ctx)
+// return values:
+//   0 on success
+//   1 on number of bitmaps not matching the number of markers
+//   2 on image preprocessing error
+MTMD_API int32_t mtmd_tokenize(mtmd_context * ctx,
+                               mtmd_input_chunks * output,
+                               const mtmd_input_text * text,
+                               const mtmd_bitmap ** bitmaps,
+                               size_t n_bitmaps);
+
+// returns 0 on success
+// TODO: deprecate
+MTMD_API int32_t mtmd_encode(mtmd_context * ctx,
+                             const mtmd_image_tokens * image_tokens);
+
+// returns 0 on success
+MTMD_API int32_t mtmd_encode_chunk(mtmd_context * ctx,
+                                   const mtmd_input_chunk * chunk);
+
+// get output embeddings from the last encode pass
+// the reading size (in bytes) is equal to:
+// llama_model_n_embd(model) * mtmd_input_chunk_get_n_tokens(chunk) * sizeof(float)
+MTMD_API float * mtmd_get_output_embd(mtmd_context * ctx);
+
+// Set callback for all future logging events.
+// If this is not called, or NULL is supplied, everything is output on stderr.
+MTMD_API void mtmd_log_set(ggml_log_callback log_callback, void * user_data);
+
+/////////////////////////////////////////
+
+// test function, to be used in test-mtmd-c-api.c
+MTMD_API mtmd_input_chunks * mtmd_test_create_input_chunks(void);
+
+#ifdef __cplusplus
+} // extern "C"
+#endif
+
+//
+// C++ wrappers
+//
+
+#ifdef __cplusplus
+
+namespace mtmd {
+
+struct mtmd_context_deleter {
+    void operator()(mtmd_context * val) { mtmd_free(val); }
+};
+using context_ptr = std::unique_ptr<mtmd_context, mtmd_context_deleter>;
+
+struct mtmd_bitmap_deleter {
+    void operator()(mtmd_bitmap * val) { mtmd_bitmap_free(val); }
+};
+using bitmap_ptr = std::unique_ptr<mtmd_bitmap, mtmd_bitmap_deleter>;
+
+struct mtmd_input_chunks_deleter {
+    void operator()(mtmd_input_chunks * val) { mtmd_input_chunks_free(val); }
+};
+using input_chunks_ptr = std::unique_ptr<mtmd_input_chunks, mtmd_input_chunks_deleter>;
+
+struct mtmd_input_chunk_deleter {
+    void operator()(mtmd_input_chunk * val) { mtmd_input_chunk_free(val); }
+};
+using input_chunk_ptr = std::unique_ptr<mtmd_input_chunk, mtmd_input_chunk_deleter>;
+
+struct bitmap {
+    bitmap_ptr ptr;
+    bitmap() : ptr(nullptr) {}
+    bitmap(mtmd_bitmap * bitmap) : ptr(bitmap) {}
+    bitmap(bitmap && other) noexcept : ptr(std::move(other.ptr)) {}
+    bitmap(uint32_t nx, uint32_t ny, const unsigned char * data) {
+        ptr.reset(mtmd_bitmap_init(nx, ny, data));
+    }
+    ~bitmap() = default;
+    uint32_t nx() const { return mtmd_bitmap_get_nx(ptr.get()); }
+    uint32_t ny() const { return mtmd_bitmap_get_ny(ptr.get()); }
+    const unsigned char * data() const { return mtmd_bitmap_get_data(ptr.get()); }
+    size_t n_bytes() const { return mtmd_bitmap_get_n_bytes(ptr.get()); }
+    std::string id() const { return mtmd_bitmap_get_id(ptr.get()); }
+    void set_id(const char * id) const { mtmd_bitmap_set_id(ptr.get(), id); }
+};
+
+struct bitmaps {
+    std::vector<bitmap> entries;
+    ~bitmaps() = default;
+    // return list of pointers to mtmd_bitmap
+    // example:
+    //   auto bitmaps_c_ptr = bitmaps.c_ptr();
+    //   int32_t res = mtmd_tokenize(... bitmaps_c_ptr.data(), bitmaps_c_ptr.size());
+    std::vector<const mtmd_bitmap *> c_ptr() {
+        std::vector<const mtmd_bitmap *> res(entries.size());
+        for (size_t i = 0; i < entries.size(); i++) {
+            res[i] = entries[i].ptr.get();
+        }
+        return res;
+    }
+};
+
+struct input_chunks {
+    input_chunks_ptr ptr;
+    input_chunks() = default;
+    input_chunks(mtmd_input_chunks * chunks) : ptr(chunks) {}
+    ~input_chunks() = default;
+    size_t size() const { return mtmd_input_chunks_size(ptr.get()); }
+    const mtmd_input_chunk * operator[](size_t idx) const {
+        return mtmd_input_chunks_get(ptr.get(), idx);
+    }
+};
+
+} // namespace mtmd
+
+#endif
+
+#endif

tools/mtmd/requirements.txt

@@ -0,0 +1,5 @@
+-r ../../requirements/requirements-convert_legacy_llama.txt
+--extra-index-url https://download.pytorch.org/whl/cpu
+pillow~=11.3.0
+torch~=2.6.0
+torchvision~=0.21.0

tools/mtmd/tests.sh

@@ -0,0 +1,183 @@
+#!/usr/bin/env bash
+
+# make sure we are in the right directory
+SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )
+cd $SCRIPT_DIR
+
+#export LLAMA_CACHE="$SCRIPT_DIR/tmp"
+
+set -eux
+
+mkdir -p $SCRIPT_DIR/output
+
+PROJ_ROOT="$SCRIPT_DIR/../.."
+cd $PROJ_ROOT
+
+# Check if the first argument is "big", then run test with big models
+# This is useful if we're running the script on a larger machine, so we can test the big models
+RUN_BIG_TESTS=false
+if [ "${1:-}" = "big" ]; then
+    RUN_BIG_TESTS=true
+    echo "Include BIG models..."
+fi
+
+RUN_HUGE_TESTS=false
+if [ "${1:-}" = "huge" ]; then
+    RUN_HUGE_TESTS=true
+    RUN_BIG_TESTS=true
+    echo "Include BIG and HUGE models..."
+fi
+
+###############
+
+arr_prefix=()
+arr_hf=()
+arr_extra_args=()
+arr_file=()
+
+add_test_vision() {
+    local hf=$1
+    shift
+    local extra_args=""
+    if [ $# -gt 0 ]; then
+        extra_args=$(printf " %q" "$@")
+    fi
+    arr_prefix+=("[vision]")
+    arr_hf+=("$hf")
+    arr_extra_args+=("$extra_args")
+    arr_file+=("test-1.jpeg")
+}
+
+add_test_audio() {
+    local hf=$1
+    shift
+    local extra_args=""
+    if [ $# -gt 0 ]; then
+        extra_args=$(printf " %q" "$@")
+    fi
+    arr_prefix+=("[audio] ")
+    arr_hf+=("$hf")
+    arr_extra_args+=("$extra_args")
+    arr_file+=("test-2.mp3")
+}
+
+add_test_vision "ggml-org/SmolVLM-500M-Instruct-GGUF:Q8_0"
+add_test_vision "ggml-org/SmolVLM2-2.2B-Instruct-GGUF:Q4_K_M"
+add_test_vision "ggml-org/SmolVLM2-500M-Video-Instruct-GGUF:Q8_0"
+add_test_vision "ggml-org/gemma-3-4b-it-GGUF:Q4_K_M"
+add_test_vision "THUDM/glm-edge-v-5b-gguf:Q4_K_M" -p "name of the newspaper?<__media__>"
+add_test_vision "second-state/Llava-v1.5-7B-GGUF:Q2_K" --chat-template vicuna
+add_test_vision "cjpais/llava-1.6-mistral-7b-gguf:Q3_K_M" --chat-template vicuna
+add_test_vision "ibm-research/granite-vision-3.2-2b-GGUF:Q4_K_M"
+add_test_vision "second-state/MiniCPM-Llama3-V-2_5-GGUF:Q2_K"  # model from openbmb is corrupted
+add_test_vision "openbmb/MiniCPM-V-2_6-gguf:Q2_K"
+add_test_vision "openbmb/MiniCPM-o-2_6-gguf:Q4_0"
+add_test_vision "bartowski/Qwen2-VL-2B-Instruct-GGUF:Q4_K_M"
+add_test_vision "ggml-org/Qwen2.5-VL-3B-Instruct-GGUF:Q4_K_M"
+add_test_vision "ggml-org/InternVL2_5-1B-GGUF:Q8_0"
+add_test_vision "ggml-org/InternVL3-1B-Instruct-GGUF:Q8_0"
+add_test_vision "ggml-org/Qwen2.5-Omni-3B-GGUF:Q4_K_M"
+add_test_vision "ggml-org/LFM2-VL-450M-GGUF:Q8_0"
+add_test_vision "ggml-org/granite-docling-258M-GGUF:Q8_0"
+add_test_vision "ggml-org/LightOnOCR-1B-1025-GGUF:Q8_0"
+
+add_test_audio  "ggml-org/ultravox-v0_5-llama-3_2-1b-GGUF:Q8_0"
+add_test_audio  "ggml-org/Qwen2.5-Omni-3B-GGUF:Q4_K_M"
+add_test_audio  "ggml-org/Voxtral-Mini-3B-2507-GGUF:Q4_K_M"
+add_test_audio  "ggml-org/LFM2-Audio-1.5B-GGUF:Q8_0"
+
+# to test the big models, run: ./tests.sh big
+if [ "$RUN_BIG_TESTS" = true ]; then
+    add_test_vision "ggml-org/pixtral-12b-GGUF:Q4_K_M"
+    add_test_vision "ggml-org/Mistral-Small-3.1-24B-Instruct-2503-GGUF" --chat-template mistral-v7
+    add_test_vision "ggml-org/Qwen2-VL-2B-Instruct-GGUF:Q4_K_M"
+    add_test_vision "ggml-org/Qwen2-VL-7B-Instruct-GGUF:Q4_K_M"
+    add_test_vision "ggml-org/Qwen2.5-VL-3B-Instruct-GGUF:Q4_K_M"
+    add_test_vision "ggml-org/Qwen2.5-VL-7B-Instruct-GGUF:Q4_K_M"
+    add_test_vision "ggml-org/Qwen3-VL-2B-Instruct-GGUF:Q8_0"
+    add_test_vision "ggml-org/InternVL3-8B-Instruct-GGUF:Q4_K_M"
+    add_test_vision "ggml-org/InternVL3-14B-Instruct-GGUF:Q4_K_M"
+    add_test_vision "ggml-org/Qwen2.5-Omni-7B-GGUF:Q4_K_M"
+    # add_test_vision "ggml-org/Qwen2.5-VL-32B-Instruct-GGUF:Q4_K_M" # does not work on my mac M3 Ultra
+    # add_test_vision "ggml-org/Kimi-VL-A3B-Thinking-2506-GGUF:Q4_K_M" # not always working
+
+    add_test_audio  "ggml-org/ultravox-v0_5-llama-3_1-8b-GGUF:Q4_K_M"
+    add_test_audio  "ggml-org/Qwen2.5-Omni-7B-GGUF:Q4_K_M"
+fi
+
+# to test the huge models, run: ./tests.sh huge
+# this will run both the big and huge models
+# huge models are > 32B parameters
+if [ "$RUN_HUGE_TESTS" = true ]; then
+    add_test_vision "ggml-org/Qwen2.5-VL-72B-Instruct-GGUF:Q4_K_M"
+    add_test_vision "ggml-org/Llama-4-Scout-17B-16E-Instruct-GGUF:IQ1_S"
+fi
+
+# these models always give the wrong answer, not sure why
+# add_test_vision "ggml-org/SmolVLM-Instruct-GGUF:Q4_K_M"
+# add_test_vision "ggml-org/SmolVLM-256M-Instruct-GGUF:Q8_0"
+# add_test_vision "ggml-org/SmolVLM2-256M-Video-Instruct-GGUF:Q8_0"
+
+# this model has broken chat template, not usable
+# add_test_vision "cmp-nct/Yi-VL-6B-GGUF:Q5_K"
+# add_test_vision "guinmoon/MobileVLM-3B-GGUF:Q4_K_M" "deepseek"
+
+###############
+
+cmake --build build -j --target llama-mtmd-cli
+
+arr_res=()
+
+for i in "${!arr_hf[@]}"; do
+    bin="llama-mtmd-cli"
+    prefix="${arr_prefix[$i]}"
+    hf="${arr_hf[$i]}"
+    extra_args="${arr_extra_args[$i]}"
+    inp_file="${arr_file[$i]}"
+
+    echo "Running test with binary: $bin and HF model: $hf"
+    echo ""
+    echo ""
+
+    cmd="$(printf %q "$PROJ_ROOT/build/bin/$bin") \
+        -hf $(printf %q "$hf") \
+        --image $(printf %q "$SCRIPT_DIR/$inp_file") \
+        --temp 0 -n 128 \
+        ${extra_args}"
+
+    # if extra_args does not contain -p, we add a default prompt
+    if ! [[ "$extra_args" =~ "-p" ]]; then
+        cmd+=" -p \"what is the publisher name of the newspaper?\""
+    fi
+
+    output=$(eval "$cmd" 2>&1 | tee /dev/tty)
+
+    echo "$output" > $SCRIPT_DIR/output/$bin-$(echo "$hf" | tr '/' '-').log
+
+    # either contains "new york" or both "men" and "walk"
+    if echo "$output" | grep -iq "new york" \
+            || (echo "$output" | grep -iq "men" && echo "$output" | grep -iq "walk")
+    then
+        result="$prefix \033[32mOK\033[0m:   $hf"
+    else
+        result="$prefix \033[31mFAIL\033[0m: $hf"
+    fi
+    echo -e "$result"
+    arr_res+=("$result")
+
+    echo ""
+    echo ""
+    echo ""
+    echo "#################################################"
+    echo "#################################################"
+    echo ""
+    echo ""
+done
+
+set +x
+
+for i in "${!arr_res[@]}"; do
+    echo -e "${arr_res[$i]}"
+done
+echo ""
+echo "Output logs are saved in $SCRIPT_DIR/output"

tools/perplexity/CMakeLists.txt

@@ -0,0 +1,8 @@
+set(TARGET llama-perplexity)
+add_executable(${TARGET} perplexity.cpp)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_17)
+
+if(LLAMA_TOOLS_INSTALL)
+    install(TARGETS ${TARGET} RUNTIME)
+endif()

tools/perplexity/README.md

@@ -0,0 +1,193 @@
+# Perplexity
+
+The `perplexity` example can be used to calculate the so-called perplexity value of a language model over a given text corpus.
+Perplexity measures how well the model can predict the next token with lower values being better.
+Note that perplexity is **not** directly comparable between models, especially if they use different tokenizers.
+Also note that finetunes typically result in a higher perplexity value even though the human-rated quality of outputs increases.
+
+Within llama.cpp the perplexity of base models is used primarily to judge the quality loss from e.g. quantized models vs. FP16.
+The convention among contributors is to use the Wikitext-2 test set for testing unless noted otherwise (can be obtained with `scripts/get-wikitext-2.sh`).
+When numbers are listed all command line arguments and compilation options are left at their defaults unless noted otherwise.
+llama.cpp numbers are **not** directly comparable to those of other projects because the exact values depend strongly on the implementation details.
+
+By default only the mean perplexity value and the corresponding uncertainty is calculated.
+The uncertainty is determined empirically by assuming a Gaussian distribution of the "correct" logits per and then applying error propagation.
+
+More statistics can be obtained by recording the logits from the FP16 version of a model.
+To do this, supply `perplexity` with `--kl-divergence-base path/to/logit/binary/file.kld`.
+The program will then record all logits and save them to the provided path in binary format.
+**The logit file will be very large, 11 GiB for LLaMA 2 or 37 GiB for LLaMA 3 when using the Wikitext-2 test set.**
+Once you have the file, supply `perplexity` with the quantized model, the logits file via `--kl-divergence-base`,
+and finally the `--kl-divergence` argument to indicate that the program should calculate the so-called Kullback-Leibler divergence.
+This is a measure of how similar the FP16 and the quantized logit distributions are with a value of 0 indicating that the distribution are the same.
+The uncertainty on the mean KL divergence is calculated by assuming the KL divergence per token follows a Gaussian distribution.
+
+In addition to the KL divergence the following statistics are calculated with `--kl-divergence`:
+
+* Ratio of mean FP16 PPL and quantized PPL. Uncertainty is estimated on logits, then propagated. The logarithm of this metric is also calculated and printed, it is 0 if the logit distributions are the same.
+* Difference of mean FP16 PPL and quantized PPL. Uncertainty is estimated on logits, then propagated.
+* Mean change in "correct" token probability. Positive values mean the model gets better at prediction, negative values mean it gets worse.
+* Pearson correlation coefficient of the "correct" token probabilites between models.
+* Percentiles of change in "correct" token probability. Positive values mean the model gets better at prediction, negative values mean it gets worse. Can be used to judge noise vs. quality loss from quantization. If the percentiles are symmetric then the quantization is essentially just adding noise. If the negative values are significantly larger than the positive values then this indicates that the model is actually becoming worse from the quantization.
+* The root mean square of the change in token probabilities. If you were to assume that the quantization simply causes Gaussian noise on the token probabilities then this would be the standard deviation of said noise. The uncertainty on the value is calculated that the change in token probabilities follows a Gaussian distribution. Related discussion: https://github.com/ggerganov/llama.cpp/discussions/2875 .
+* Same top p: Percentage of how often the token was assigned the highest probabilites by both models. The uncertainty is calculated from the Gaussian approximation of the binomial distribution.
+
+## LLaMA 3 8b Scoreboard
+
+| Revision | f364eb6f           |
+|:---------|:-------------------|
+| Backend  | CUDA               |
+| CPU      | AMD Epyc 7742      |
+| GPU      | 1x NVIDIA RTX 4090 |
+
+Results were generated using the CUDA backend and are sorted by Kullback-Leibler divergence relative to FP16.
+The "WT" importance matrices were created using varying numbers of Wikitext tokens and can be found [here](https://huggingface.co/JohannesGaessler/llama.cpp_importance_matrices/blob/main/imatrix-llama_3-8b-f16-2.7m_tokens.dat).
+Note: the FP16 logits used for the calculation of all metrics other than perplexity are stored in a binary file between runs.
+In order to save space this file does **not** contain the exact same FP32 logits but instead casts them to 16 bit unsigned integers (with some scaling).
+So the "f16" results are to be understood as the difference resulting only from this downcast.
+
+| Quantization | imatrix | Model size [GiB] | PPL                    | ΔPPL                   | KLD                   | Mean Δp           | RMS Δp           |
+|--------------|---------|------------------|------------------------|------------------------|-----------------------|-------------------|------------------|
+| f16          | None    |            14.97 | 6.233160 ±   0.037828  | 0.001524 ±   0.000755  | 0.000551 ±   0.000002 |  0.001 ± 0.002 %  | 0.787 ± 0.004 %  |
+| q8_0         | None    |             7.96 | 6.234284 ±   0.037878  | 0.002650 ±   0.001006  | 0.001355 ±   0.000006 | -0.019 ± 0.003 %  | 1.198 ± 0.007 %  |
+| q6_K         | None    |             6.14 | 6.253382 ±   0.038078  | 0.021748 ±   0.001852  | 0.005452 ±   0.000035 | -0.007 ± 0.006 %  | 2.295 ± 0.019 %  |
+| q5_K_M       | None    |             5.33 | 6.288607 ±   0.038338  | 0.056974 ±   0.002598  | 0.010762 ±   0.000079 | -0.114 ± 0.008 %  | 3.160 ± 0.031 %  |
+| q5_K_S       | None    |             5.21 | 6.336598 ±   0.038755  | 0.104964 ±   0.003331  | 0.016595 ±   0.000122 | -0.223 ± 0.010 %  | 3.918 ± 0.036 %  |
+| q5_1         | None    |             5.65 | 6.337857 ±   0.038677  | 0.106223 ±   0.003476  | 0.018045 ±   0.000139 | -0.287 ± 0.011 %  | 4.123 ± 0.039 %  |
+| q5_0         | None    |             5.21 | 6.363224 ±   0.038861  | 0.131591 ±   0.003894  | 0.022239 ±   0.000166 | -0.416 ± 0.012 %  | 4.634 ± 0.043 %  |
+| q4_K_M       | WT 10m  |             4.58 | 6.382937 ±   0.039055  | 0.151303 ±   0.004429  | 0.028152 ±   0.000240 | -0.389 ± 0.014 %  | 5.251 ± 0.049 %  |
+| q4_K_M       | None    |             4.58 | 6.407115 ±   0.039119  | 0.175482 ±   0.004620  | 0.031273 ±   0.000238 | -0.596 ± 0.014 %  | 5.519 ± 0.050 %  |
+| q4_K_S       | WT 10m  |             4.37 | 6.409697 ±   0.039189  | 0.178064 ±   0.004744  | 0.031951 ±   0.000259 | -0.531 ± 0.015 %  | 5.645 ± 0.051 %  |
+| iq4_NL       | WT 10m  |             4.35 | 6.455593 ±   0.039630  | 0.223959 ±   0.005201  | 0.035742 ±   0.000288 | -0.590 ± 0.016 %  | 5.998 ± 0.054 %  |
+| iq4_XS       | WT 10m  |             4.14 | 6.459705 ±   0.039595  | 0.228071 ±   0.005207  | 0.036334 ±   0.000284 | -0.668 ± 0.016 %  | 6.044 ± 0.054 %  |
+| q4_K_S       | None    |             4.37 | 6.500529 ±   0.039778  | 0.268895 ±   0.005638  | 0.043136 ±   0.000314 | -0.927 ± 0.017 %  | 6.562 ± 0.055 %  |
+| q4_1         | None    |             4.78 | 6.682737 ±   0.041285  | 0.451103 ±   0.008030  | 0.071683 ±   0.000505 | -0.927 ± 0.017 %  | 8.512 ± 0.063 %  |
+| q4_0         | None    |             4.34 | 6.700147 ±   0.041226  | 0.468514 ±   0.007951  | 0.071940 ±   0.000491 | -1.588 ± 0.022 %  | 8.434 ± 0.061 %  |
+| q3_K_L       | WT 10m  |             4.03 | 6.671223 ±   0.041427  | 0.439590 ±   0.008154  | 0.073077 ±   0.000529 | -0.940 ± 0.023 %  | 8.662 ± 0.064 %  |
+| q3_K_M       | WT 10m  |             3.74 | 6.734255 ±   0.041838  | 0.502622 ±   0.008901  | 0.084358 ±   0.000588 | -1.198 ± 0.024 %  | 9.292 ± 0.065 %  |
+| q3_K_L       | None    |             4.03 | 6.787876 ±   0.042104  | 0.556242 ±   0.009171  | 0.087176 ±   0.000614 | -1.532 ± 0.025 %  | 9.432 ± 0.067 %  |
+| q3_K_M       | None    |             3.74 | 6.888498 ±   0.042669  | 0.656864 ±   0.010071  | 0.101913 ±   0.000677 | -1.990 ± 0.026 %  | 10.203 ± 0.068 % |
+| iq3_M        | WT 10m  |             3.53 | 6.898327 ±   0.041643  | 0.666694 ±   0.009449  | 0.102534 ±   0.000663 | -3.178 ± 0.026 %  | 10.513 ± 0.066 % |
+| iq3_S        | WT 10m  |             3.42 | 6.965501 ±   0.042406  | 0.733867 ±   0.010245  | 0.111278 ±   0.000710 | -3.066 ± 0.027 %  | 10.845 ± 0.068 % |
+| iq3_XS       | WT 10m  |             3.28 | 7.163043 ±   0.043772  | 0.931409 ±   0.012084  | 0.138693 ±   0.000857 | -3.667 ± 0.031 %  | 12.148 ± 0.070 % |
+| iq3_XXS      | WT 10m  |             3.05 | 7.458436 ±   0.046404  | 1.226803 ±   0.015234  | 0.183625 ±   0.001042 | -3.918 ± 0.035 %  | 13.836 ± 0.074 % |
+| q3_K_S       | WT 10m  |             3.41 | 7.602878 ±   0.046848  | 1.371244 ±   0.015688  | 0.199821 ±   0.001008 | -5.046 ± 0.037 %  | 14.980 ± 0.070 % |
+| q3_K_S       | None    |             3.41 | 7.863786 ±   0.048885  | 1.632152 ±   0.017733  | 0.228217 ±   0.001079 | -5.604 ± 0.038 %  | 15.541 ± 0.070 % |
+| iq2_M        | WT 10m  |             2.74 | 8.600799 ±   0.055124  | 2.369166 ±   0.025244  | 0.325989 ±   0.00160  | -6.463 ± 0.046 %  | 18.519 ± 0.080 % |
+| q2_K         | WT 10k  |             2.96 | 8.652290 ±   0.055572  | 2.420657 ±   0.025587  | 0.331393 ±   0.001562 | -6.606 ± 0.046 %  | 18.790 ± 0.078 % |
+| q2_K         | WT 100k |             2.96 | 8.641993 ±   0.055406  | 2.410359 ±   0.025495  | 0.331672 ±   0.001569 | -6.628 ± 0.047 %  | 18.856 ± 0.078 % |
+| q2_K         | WT 10m  |             2.96 | 8.647825 ±   0.055610  | 2.416191 ±   0.025683  | 0.332223 ±   0.001572 | -6.500 ± 0.047 %  | 18.881 ± 0.078 % |
+| q2_K         | WT 1m   |             2.96 | 8.674365 ±   0.055743  | 2.442732 ±   0.025843  | 0.335308 ±   0.001576 | -6.634 ± 0.047 %  | 19.009 ± 0.079 % |
+| q2_K         | WT 1k   |             2.96 | 8.682605 ±   0.055916  | 2.450972 ±   0.026069  | 0.337093 ±   0.001596 | -6.596 ± 0.047 %  | 18.977 ± 0.079 % |
+| q2_K_S       | WT 10m  |             2.96 | 9.323778 ±   0.061551  | 3.092145 ±   0.031914  | 0.403360 ±   0.001787 | -7.131 ± 0.049 %  | 20.050 ± 0.081 % |
+| q2_K_S       | WT 1m   |             2.96 | 9.329321 ±   0.061378  | 3.097688 ±   0.031816  | 0.403590 ±   0.001797 | -7.289 ± 0.049 %  | 20.123 ± 0.081 % |
+| q2_K_S       | WT 100k |             2.96 | 9.362973 ±   0.061740  | 3.131339 ±   0.032169  | 0.408367 ±   0.001802 | -7.198 ± 0.050 %  | 20.132 ± 0.081 % |
+| q2_K_S       | WT 10k  |             2.96 | 9.376479 ±   0.062045  | 3.144846 ±   0.032464  | 0.408662 ±   0.001819 | -7.141 ± 0.050 %  | 20.120 ± 0.081 % |
+| q2_K_S       | WT 1k   |             2.96 | 9.415200 ±   0.062475  | 3.183567 ±   0.032993  | 0.415865 ±   0.001846 | -7.153 ± 0.050 %  | 20.311 ± 0.082 % |
+| iq2_S        | WT 10m  |             2.56 | 9.650781 ±   0.063209  | 3.419148 ±   0.034017  | 0.439197 ±   0.001976 | -8.319 ± 0.052 %  | 21.491 ± 0.083 % |
+| q2_K         | None    |             2.96 | 9.751568 ±   0.063312  | 3.519934 ±   0.033863  | 0.445132 ±   0.001835 | -9.123 ± 0.051 %  | 21.421 ± 0.079 % |
+| iq2_XS       | WT 10m  |             2.43 | 10.761424 ±   0.071056 | 4.529791 ±   0.042229  | 0.546290 ±   0.002133 | -10.576 ± 0.056 % | 23.872 ± 0.082 % |
+| iq2_XXS      | WT 10m  |             2.24 | 14.091782 ±   0.098396 | 7.860148 ±   0.070752  | 0.812022 ±   0.002741 | -14.363 ± 0.065 % | 28.576 ± 0.084 % |
+| iq1_M        | WT 10m  |             2.01 | 25.493722 ±   0.177903 | 19.262089 ±   0.152396 | 1.393084 ±   0.003529 | -24.672 ± 0.077 % | 38.287 ± 0.084 % |
+| iq1_S        | WT 1m   |             1.88 | 58.097760 ±   0.438604 | 51.866126 ±   0.416604 | 2.211278 ±   0.004688 | -32.471 ± 0.087 % | 46.418 ± 0.085 % |
+| iq1_S        | WT 1k   |             1.88 | 58.267851 ±   0.446208 | 52.036218 ±   0.424373 | 2.214858 ±   0.004778 | -31.880 ± 0.089 % | 46.330 ± 0.086 % |
+| iq1_S        | WT 100k |             1.88 | 58.581498 ±   0.453145 | 52.349864 ±   0.431360 | 2.220834 ±   0.004818 | -32.261 ± 0.089 % | 46.002 ± 0.086 % |
+| iq1_S        | WT 10m  |             1.88 | 60.694593 ±   0.471290 | 54.462959 ±   0.449644 | 2.254554 ±   0.004868 | -31.973 ± 0.088 % | 46.271 ± 0.086 % |
+| iq1_S        | WT 10k  |             1.88 | 63.221324 ±   0.493077 | 56.989691 ±   0.471423 | 2.293527 ±   0.004885 | -32.261 ± 0.089 % | 46.562 ± 0.086 % |
+
+There seems to be no consistent improvement from using more Wikitext tokens for the importance matrix.
+K-quants score better on mean Δp than the legacy quants than e.g. KL divergence would suggest.
+
+## LLaMA 2 vs. LLaMA 3 Quantization comparison
+
+| Revision | f364eb6f           |
+|:---------|:-------------------|
+| Backend  | CUDA               |
+| CPU      | AMD Epyc 7742      |
+| GPU      | 1x NVIDIA RTX 4090 |
+
+| Metric          |          L2 7b q2_K |          L3 8b q2_K |        L2 7b q4_K_M |        L3 8b q4_K_M |          L2 7b q6_K |          L3 8b q6_K |          L2 7b q8_0 |          L3 8b q8_0 |
+|-----------------|---------------------|---------------------|---------------------|---------------------|---------------------|---------------------|---------------------|---------------------|
+| Mean PPL        | 5.794552 ± 0.032298 | 9.751568 ± 0.063312 | 5.877078 ± 0.032781 | 6.407115 ± 0.039119 | 5.808494 ± 0.032425 | 6.253382 ± 0.038078 | 5.798542 ± 0.032366 | 6.234284 ± 0.037878 |
+| Mean PPL ratio  | 1.107955 ± 0.001427 | 1.564849 ± 0.004525 | 1.014242 ± 0.000432 | 1.028160 ± 0.000723 | 1.002406 ± 0.000191 | 1.003490 ± 0.000296 | 1.000689 ± 0.000107 | 1.000425 ± 0.000161 |
+| Mean ΔPPL       | 0.625552 ± 0.008725 | 3.519934 ± 0.033863 | 0.082526 ± 0.002530 | 0.175482 ± 0.004620 | 0.013941 ± 0.001110 | 0.021748 ± 0.001852 | 0.003990 ± 0.000624 | 0.002650 ± 0.001006 |
+| PPL correlation |              97.36% |              89.62% |              99.71% |              99.34% |              99.94% |              99.88% |              99.98% |              99.96% |
+| Mean KLD        | 0.108903 ± 0.000645 | 0.445132 ± 0.001835 | 0.012686 ± 0.000079 | 0.031273 ± 0.000238 | 0.002098 ± 0.000014 | 0.005452 ± 0.000035 | 0.000369 ± 0.000007 | 0.001355 ± 0.000006 |
+| Mean Δp         |    -2.710 ± 0.023 % |    -9.123 ± 0.051 % |    -0.416 ± 0.008 % |    -0.596 ± 0.014 % |    -0.035 ± 0.003 % |    -0.007 ± 0.006 % |    -0.005 ± 0.002 % |    -0.019 ± 0.003 % |
+| Maximum Δp      |             85.136% |             94.268% |             45.209% |             95.054% |             23.593% |             53.601% |             43.925% |             28.734% |
+| 99.9% Δp        |             37.184% |             50.003% |             17.461% |             27.084% |              7.798% |             13.613% |              3.387% |              6.402% |
+| 99.0% Δp        |             18.131% |             25.875% |              7.798% |             12.084% |              3.838% |              6.407% |              1.867% |              3.544% |
+| Median Δp       |             -0.391% |             -2.476% |             -0.026% |             -0.024% |             -0.001% |              0.000% |             -0.000% |             -0.000% |
+| 1.0% Δp         |            -39.762% |            -87.173% |            -11.433% |            -19.567% |             -4.222% |             -6.767% |             -1.862% |             -3.698% |
+| 0.1% Δp         |            -79.002% |            -98.897% |            -26.433% |            -56.054% |             -9.091% |            -16.584% |             -3.252% |             -6.579% |
+| Minimum Δp      |            -99.915% |            -99.965% |            -83.383% |            -98.699% |            -43.142% |            -68.487% |             -9.343% |            -24.301% |
+| RMS Δp          |     9.762 ± 0.053 % |    21.421 ± 0.079 % |     3.252 ± 0.024 % |     5.519 ± 0.050 % |     1.339 ± 0.010 % |     2.295 ± 0.019 % |     0.618 ± 0.011 % |     1.198 ± 0.007 % |
+| Same top p      |    85.584 ± 0.086 % |    71.138 ± 0.119 % |    94.665 ± 0.055 % |    91.901 ± 0.072 % |    97.520 ± 0.038 % |    96.031 ± 0.051 % |    98.846 ± 0.026 % |    97.674 ± 0.040 % |
+
+## LLaMA 3 BF16 vs. FP16 comparison
+
+| Revision | 83330d8c      |
+|:---------|:--------------|
+| Backend  | CPU           |
+| CPU      | AMD Epyc 7742 |
+| GPU      | N/A           |
+
+Results were calculated with LLaMA 3 8b BF16 as `--kl-divergence-base` and LLaMA 3 8b FP16 as the `--model` for comparison.
+
+| Metric                         |                    Value |
+|--------------------------------|--------------------------|
+| Mean PPL(Q)                    |      6.227711 ± 0.037833 |
+| Mean PPL(base)                 |      6.225194 ± 0.037771 |
+| Cor(ln(PPL(Q)), ln(PPL(base))) |                  99.990% |
+| Mean ln(PPL(Q)/PPL(base))      |      0.000404 ± 0.000086 |
+| Mean PPL(Q)/PPL(base)          |      1.000404 ± 0.000086 |
+| Mean PPL(Q)-PPL(base)          |      0.002517 ± 0.000536 |
+| Mean    KLD                    |  0.00002515 ± 0.00000020 |
+| Maximum KLD                    |                 0.012206 |
+| 99.9%   KLD                    |                 0.000799 |
+| 99.0%   KLD                    |                 0.000222 |
+| 99.0%   KLD                    |                 0.000222 |
+| Median  KLD                    |                 0.000013 |
+| 10.0%   KLD                    |                -0.000002 |
+| 5.0%   KLD                     |                -0.000008 |
+| 1.0%   KLD                     |                -0.000023 |
+| Minimum KLD                    |                -0.000059 |
+| Mean    Δp                     | -0.0000745 ± 0.0003952 % |
+| Maximum Δp                     |                   4.186% |
+| 99.9%   Δp                     |                   1.049% |
+| 99.0%   Δp                     |                   0.439% |
+| 95.0%   Δp                     |                   0.207% |
+| 90.0%   Δp                     |                   0.125% |
+| 75.0%   Δp                     |                   0.029% |
+| Median  Δp                     |                   0.000% |
+| 25.0%   Δp                     |                  -0.030% |
+| 10.0%   Δp                     |                  -0.126% |
+| 5.0%   Δp                      |                  -0.207% |
+| 1.0%   Δp                      |                  -0.434% |
+| 0.1%   Δp                      |                  -1.016% |
+| Minimum Δp                     |                  -4.672% |
+| RMS Δp                         |          0.150 ± 0.001 % |
+| Same top p                     |         99.739 ± 0.013 % |
+
+## Old Numbers
+
+<details>
+<summary>Llama 2 70B Scoreboard</summary>
+
+| Quantization | Model size (GiB) | Perplexity | Delta to fp16 |
+|--------------|------------------|------------|---------------|
+| Q4_0         | 36.20            | 3.5550     | 3.61%         |
+| Q4_1         | 40.20            | 3.5125     | 2.37%         |
+| Q5_0         | 44.20            | 3.4744     | 1.26%         |
+| Q2_K         | 27.27            | 3.7339     | 8.82%         |
+| Q3_K_S       | 27.86            | 3.7019     | 7.89%         |
+| Q3_K_M       | 30.83            | 3.5932     | 4.72%         |
+| Q3_K_L       | 33.67            | 3.5617     | 3.80%         |
+| Q4_K_S       | 36.39            | 3.4852     | 1.57%         |
+| Q4_K_M       | 38.54            | 3.4725     | 1.20%         |
+| Q5_K_S       | 44.20            | 3.4483     | 0.50%         |
+| Q5_K_M       | 45.41            | 3.4451     | 0.40%         |
+| Q6_K         | 52.70            | 3.4367     | 0.16%         |
+| fp16         | 128.5            | 3.4313     | -             |
+
+</details>

tools/quantize/CMakeLists.txt

@@ -0,0 +1,9 @@
+set(TARGET llama-quantize)
+add_executable(${TARGET} quantize.cpp)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_include_directories(${TARGET} PRIVATE ../../common)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)
+
+if(LLAMA_TOOLS_INSTALL)
+    install(TARGETS ${TARGET} RUNTIME)
+endif()

tools/quantize/README.md

@@ -0,0 +1,171 @@
+# quantize
+
+This tool takes a GGUF input model file, typically in a high-precision format like F32 or BF16, and converts it to a quantized format.
+Quantization reduces the precision of model weights (e.g., from 32-bit floats to 4-bit integers), which shrinks the model's size and can speed up inference.
+This process however, may introduce some accuracy loss which is usually measured in [Perplexity](https://huggingface.co/docs/transformers/en/perplexity) (ppl) and/or [Kullback–Leibler Divergence](https://en.wikipedia.org/wiki/Kullback%E2%80%93Leibler_divergence) (kld).
+This can be minimized by using a suitable imatrix file.
+
+You can also use the [GGUF-my-repo](https://huggingface.co/spaces/ggml-org/gguf-my-repo) space on Hugging Face to build your own quants without any setup.
+
+Note: It is synced from llama.cpp `main` every 6 hours.
+
+Example usage:
+
+```./llama-quantize [options] input-model-f32.gguf [output-model-quant.gguf] type [threads]```
+
+```bash
+# from Hugginface, obtain the official meta-llama/Llama-3.1-8B model weights and place them in ./models
+ls ./models
+config.json             model-00001-of-00004.safetensors  model-00004-of-00004.safetensors  README.md                tokenizer.json
+generation_config.json  model-00002-of-00004.safetensors  model.safetensors.index.json      special_tokens_map.json  USE_POLICY.md
+LICENSE                 model-00003-of-00004.safetensors  original                          tokenizer_config.json
+
+# [Optional] for PyTorch .bin models like Mistral-7B
+ls ./models
+<folder containing weights and tokenizer json>
+
+# install Python dependencies
+python3 -m pip install -r requirements.txt
+
+# convert the model to ggml FP16 format
+python3 convert_hf_to_gguf.py ./models/mymodel/
+
+# quantize the model to 4-bits (using Q4_K_M method)
+./llama-quantize ./models/mymodel/ggml-model-f16.gguf ./models/mymodel/ggml-model-Q4_K_M.gguf Q4_K_M
+
+# update the gguf filetype to current version if older version is now unsupported
+./llama-quantize ./models/mymodel/ggml-model-Q4_K_M.gguf ./models/mymodel/ggml-model-Q4_K_M-v2.gguf COPY
+```
+
+Run the quantized model:
+
+```bash
+# start inference on a gguf model
+./llama-cli -m ./models/mymodel/ggml-model-Q4_K_M.gguf -cnv -p "You are a helpful assistant"
+```
+
+Options:
+* `--allow-requantize` allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit
+* `--leave-output-tensor` will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing
+* `--pure` disables k-quant mixtures and quantizes all tensors to the same type
+* `--imatrix` uses data in file generated by `llama-imatrix` as importance matrix for quant optimizations (highly recommended)
+* `--include-weights` use an importance matrix for tensor(s) in the list. Cannot be used with `--exclude-weights`
+* `--exclude-weights` use an importance matrix for tensor(s) in the list. Cannot be used with `--include-weights`
+* `--output-tensor-type` use a specific quant type for the output.weight tensor
+* `--token-embedding-type` use a specific quant type for the token embeddings tensor
+* `--keep-split` will generate the quantized model in the same shards as the input file otherwise it will produce a single quantized file
+
+Advanced options:
+* `--tensor-type` quantize specific tensor(s) to specific quant types. Supports regex syntax. May be specified multiple times.
+* `--prune-layers` prune (remove) the layers in the list
+* `--override-kv` option to override model metadata by key in the quantized model. May be specified multiple times
+
+Examples:
+
+```bash
+# naive Q4_K_M quantization using default settings and 8 CPU threads. Output will be "ggml-model-Q4_K_M.gguf"
+./llama-quantize input-model-f32.gguf q4_k_m 8
+```
+
+```bash
+#  quantize model enabling re-quantization, leaving the output tensor unquantized and all others quantized at the same level (Q4_K)
+./llama-quantize --allow-requantize --leave-output-tensor --pure input-model-f32.gguf q4_k_m 8
+```
+
+```bash
+# quantize model using an importance matrix for specified tensors only (attn_v and ffn_down)
+./llama-quantize --imatrix imatrix.gguf --include-weights attn_v --include-weights ffn_down input-model-f32.gguf q4_k_m 8
+```
+
+```bash
+# quantize model setting output tensor to Q5_K_M, token embeddings to Q3_K_M, and keeping the input file's shards
+./llama-quantize --imatrix imatrix.gguf --output-tensor-type q5_k --token-embedding-type q3_k --keep-split input-model-f32.gguf q4_k_m 8
+```
+
+```bash
+# quantize model using a regex to quantize attn_k tensors in odd layers to Q5_K_M and attn_q tensors in even layers to Q3_K_M
+./llama-quantize --imatrix imatrix.gguf --tensor-type "\.(\d*[13579])\.attn_k=q5_k" --tensor-type "\.(\d*[02468])\.attn_q=q3_k" input-model-f32.gguf q4_k_m 8
+```
+
+```bash
+# quantize model setting tensors attn_v and ffn_down to Q5_K_M and pruning layers 20, 21, and 22
+./llama-quantize --imatrix imatrix.gguf --tensor-type attn_v=q5_k --tensor-type ffn_down=q5_k --prune-layers 20,21,22 input-model-f32.gguf q4_k_m 8
+```
+
+```bash
+# override expert used count metadata to 16, prune layers 20, 21, and 22 without quantizing the model (copy tensors) and use specified name for the output file
+./llama-quantize --imatrix imatrix.gguf --override-kv qwen3moe.expert_used_count=int:16 --prune-layers 20,21,22 input-model-f32.gguf pruned-model-f32.gguf copy 8
+```
+
+## Memory/Disk Requirements
+
+When running the larger models, make sure you have enough disk space to store all the intermediate files.
+As the models are currently fully loaded into memory, you will need adequate disk space to save them and sufficient RAM to load them. At the moment, memory and disk requirements are the same. For exmaple (Llama 3.1):
+
+| Model | Original size | Quantized size (Q4_K_M) |
+| ----: | ------------: | ----------------------: |
+|    8B |       32.1 GB |                  4.9 GB |
+|   70B |      280.9 GB |                 43.1 GB |
+|  405B |    1,625.1 GB |                249.1 GB |
+
+
+## Quantization
+
+Several quantization methods are supported. They differ in the resulting model disk size and inference speed. For example,
+
+### [meta-llama/Llama-3.1-8B](https://huggingface.co/meta-llama/Llama-3.1-8B)
+
+| Measure                     | IQ1_S        | IQ1_M        | IQ2_XXS      | IQ2_XS        | IQ2_S         | IQ2_M        |
+| --------------------------- | ------------ | ------------ | ------------ | ------------- | ------------- | ------------ |
+| bits/weight                 |       2.0042 |       2.1460 |       2.3824 |        2.5882 |        2.7403 |       2.9294 |
+| size (GiB)                  |       1.87   |       2.01   |       2.23   |        2.42   |        2.56   |       2.74   |
+| prompt processing t/s @ 512 | 858.88 ±1.22 | 847.99 ±0.47 | 852.39 ±0.85 | 826.99 ±12.51 | 783.55 ±13.73 | 787.68 ±7.00 |
+| text generation t/s @ 128   |  79.73 ±0.79 |  72.92 ±0.14 |  79.86 ±0.22 |  78.04 ±0.46  |  77.30 ±2.47  |  74.44 ±0.15 |
+
+| Measure                     | IQ3_XXS      | IQ3_XS       | IQ3_S        | IQ3_M         | IQ4_XS        | IQ4_NL       |
+| --------------------------- | ------------ | ------------ | ------------ | ------------- | ------------- | ------------ |
+| bits/weight                 |       3.2548 |       3.4977 |       3.6606 |        3.7628 |        4.4597 |       4.6818 |
+| size (GiB)                  |       3.04   |       3.27   |       3.42   |        3.52   |        4.17   |       4.38   |
+| prompt processing t/s @ 512 | 813.88 ±6.53 | 708.71 ±1.26 | 798.78 ±8.81 | 768.70 ±13.73 | 771.80 ±11.38 | 806.03 ±7.07 |
+| text generation t/s @ 128   |  73.95 ±0.20 |  71.67 ±0.54 |  69.31 ±0.63 |  70.15 ±0.33  |  77.51 ±0.20  |  76.63 ±0.28 |
+
+
+| Measure                     | Q2_K_S       | Q2_K         | Q3_K_S       | Q3_K_M       | Q3_K_L       | Q4_K_S       |
+| --------------------------- | ------------ | ------------ | ------------ | ------------ | ------------ | ------------ |
+| bits/weight                 |       2.9697 |       3.1593 |       3.6429 |       3.9960 |       4.2979 |       4.6672 |
+| size (GiB)                  |       2.78   |       2.95   |       3.41   |       3.74   |       4.02   |       4.36   |
+| prompt processing t/s @ 512 | 798.91 ±6.40 | 784.45 ±7.85 | 752.17 ±7.94 | 783.44 ±9.92 | 761.17 ±7.55 | 818.55 ±9.58 |
+| text generation t/s @ 128   |  90.01 ±0.12 |  79.85 ±0.20 |  69.84 ±0.18 |  71.68 ±0.22 |  69.38 ±0.49 |  76.71 ±0.20 |
+
+| Measure                     | Q4_K_S       | Q4_K_M        | Q5_K_S       | Q5_K_M       | Q6_K          | Q8_0         |
+| --------------------------- | ------------ | ------------- | ------------ | ------------ | ------------- | ------------ |
+| bits/weight                 |       4.6672 |        4.8944 |       5.5704 |       5.7036 |        6.5633 |       8.5008 |
+| size (GiB)                  |       4.36   |        4.58   |       5.21   |       5.33   |        6.14   |       7.95   |
+| prompt processing t/s @ 512 | 818.55 ±9.58 | 821.81 ±21.44 | 752.52 ±0.99 | 758.69 ±7.43 | 812.01 ±10.82 | 865.09 ±8.30 |
+| text generation t/s @ 128   |  76.71 ±0.20 |  71.93 ±1.52  |  69.53 ±0.18 |  67.23 ±1.08 |  58.67 ±3.13  |  50.93 ±0.08 |
+
+| Measure                     | F16          |
+| --------------------------- | ------------ |
+| bits/weight                 |      16.0005 |
+| size (GiB)                  |      14.96   |
+| prompt processing t/s @ 512 | 923.49 ±0.53 |
+| text generation t/s @ 128   |  29.17 ±0.04 |
+
+## Background information on llama-quantize
+
+- [k-quants](https://github.com/ggml-org/llama.cpp/pull/1684)
+- k-quants improvements and i-quants
+  - [#2707](https://github.com/ggml-org/llama.cpp/pull/2707)
+  - [#2807](https://github.com/ggml-org/llama.cpp/pull/2807)
+  - [#4773 - 2-bit i-quants (inference)](https://github.com/ggml-org/llama.cpp/pull/4773)
+  - [#4856 - 2-bit i-quants (inference)](https://github.com/ggml-org/llama.cpp/pull/4856)
+  - [#4861 - importance matrix](https://github.com/ggml-org/llama.cpp/pull/4861)
+  - [#4872 - MoE models](https://github.com/ggml-org/llama.cpp/pull/4872)
+  - [#4897 - 2-bit quantization](https://github.com/ggml-org/llama.cpp/pull/4897)
+  - [#4930 - imatrix for all k-quants](https://github.com/ggml-org/llama.cpp/pull/4930)
+  - [#4951 - imatrix on the GPU](https://github.com/ggml-org/llama.cpp/pull/4957)
+  - [#4969 - imatrix for legacy quants](https://github.com/ggml-org/llama.cpp/pull/4969)
+  - [#4996 - k-quants tuning](https://github.com/ggml-org/llama.cpp/pull/4996)
+  - [#5060 - Q3_K_XS](https://github.com/ggml-org/llama.cpp/pull/5060)
+  - [#5196 - 3-bit i-quants](https://github.com/ggml-org/llama.cpp/pull/5196)
+  - [quantization tuning](https://github.com/ggml-org/llama.cpp/pull/5320), [another one](https://github.com/ggml-org/llama.cpp/pull/5334), and [another one](https://github.com/ggml-org/llama.cpp/pull/5361)

tools/quantize/quantize.cpp

@@ -0,0 +1,688 @@
+#include "common.h"
+#include "llama.h"
+#include "gguf.h"
+
+#include <cstdio>
+#include <cstring>
+#include <vector>
+#include <string>
+#include <unordered_map>
+#include <map>
+#include <fstream>
+#include <cmath>
+#include <cctype>
+#include <algorithm>
+#include <filesystem>
+
+struct quant_option {
+    std::string name;
+    llama_ftype ftype;
+    std::string desc;
+};
+
+static const std::vector<quant_option> QUANT_OPTIONS = {
+    { "Q4_0",     LLAMA_FTYPE_MOSTLY_Q4_0,     " 4.34G, +0.4685 ppl @ Llama-3-8B",  },
+    { "Q4_1",     LLAMA_FTYPE_MOSTLY_Q4_1,     " 4.78G, +0.4511 ppl @ Llama-3-8B",  },
+    { "MXFP4_MOE",LLAMA_FTYPE_MOSTLY_MXFP4_MOE," MXFP4 MoE",  },
+    { "Q5_0",     LLAMA_FTYPE_MOSTLY_Q5_0,     " 5.21G, +0.1316 ppl @ Llama-3-8B",  },
+    { "Q5_1",     LLAMA_FTYPE_MOSTLY_Q5_1,     " 5.65G, +0.1062 ppl @ Llama-3-8B",  },
+    { "IQ2_XXS",  LLAMA_FTYPE_MOSTLY_IQ2_XXS,  " 2.06 bpw quantization",            },
+    { "IQ2_XS",   LLAMA_FTYPE_MOSTLY_IQ2_XS,   " 2.31 bpw quantization",            },
+    { "IQ2_S",    LLAMA_FTYPE_MOSTLY_IQ2_S,    " 2.5  bpw quantization",            },
+    { "IQ2_M",    LLAMA_FTYPE_MOSTLY_IQ2_M,    " 2.7  bpw quantization",            },
+    { "IQ1_S",    LLAMA_FTYPE_MOSTLY_IQ1_S,    " 1.56 bpw quantization",            },
+    { "IQ1_M",    LLAMA_FTYPE_MOSTLY_IQ1_M,    " 1.75 bpw quantization",            },
+    { "TQ1_0",    LLAMA_FTYPE_MOSTLY_TQ1_0,    " 1.69 bpw ternarization",           },
+    { "TQ2_0",    LLAMA_FTYPE_MOSTLY_TQ2_0,    " 2.06 bpw ternarization",           },
+    { "Q2_K",     LLAMA_FTYPE_MOSTLY_Q2_K,     " 2.96G, +3.5199 ppl @ Llama-3-8B",  },
+    { "Q2_K_S",   LLAMA_FTYPE_MOSTLY_Q2_K_S,   " 2.96G, +3.1836 ppl @ Llama-3-8B",  },
+    { "IQ3_XXS",  LLAMA_FTYPE_MOSTLY_IQ3_XXS,  " 3.06 bpw quantization",            },
+    { "IQ3_S",    LLAMA_FTYPE_MOSTLY_IQ3_S,    " 3.44 bpw quantization",            },
+    { "IQ3_M",    LLAMA_FTYPE_MOSTLY_IQ3_M,    " 3.66 bpw quantization mix",        },
+    { "Q3_K",     LLAMA_FTYPE_MOSTLY_Q3_K_M,   "alias for Q3_K_M"                   },
+    { "IQ3_XS",   LLAMA_FTYPE_MOSTLY_IQ3_XS,   " 3.3 bpw quantization",             },
+    { "Q3_K_S",   LLAMA_FTYPE_MOSTLY_Q3_K_S,   " 3.41G, +1.6321 ppl @ Llama-3-8B",  },
+    { "Q3_K_M",   LLAMA_FTYPE_MOSTLY_Q3_K_M,   " 3.74G, +0.6569 ppl @ Llama-3-8B",  },
+    { "Q3_K_L",   LLAMA_FTYPE_MOSTLY_Q3_K_L,   " 4.03G, +0.5562 ppl @ Llama-3-8B",  },
+    { "IQ4_NL",   LLAMA_FTYPE_MOSTLY_IQ4_NL,   " 4.50 bpw non-linear quantization", },
+    { "IQ4_XS",   LLAMA_FTYPE_MOSTLY_IQ4_XS,   " 4.25 bpw non-linear quantization", },
+    { "Q4_K",     LLAMA_FTYPE_MOSTLY_Q4_K_M,   "alias for Q4_K_M",                  },
+    { "Q4_K_S",   LLAMA_FTYPE_MOSTLY_Q4_K_S,   " 4.37G, +0.2689 ppl @ Llama-3-8B",  },
+    { "Q4_K_M",   LLAMA_FTYPE_MOSTLY_Q4_K_M,   " 4.58G, +0.1754 ppl @ Llama-3-8B",  },
+    { "Q5_K",     LLAMA_FTYPE_MOSTLY_Q5_K_M,   "alias for Q5_K_M",                  },
+    { "Q5_K_S",   LLAMA_FTYPE_MOSTLY_Q5_K_S,   " 5.21G, +0.1049 ppl @ Llama-3-8B",  },
+    { "Q5_K_M",   LLAMA_FTYPE_MOSTLY_Q5_K_M,   " 5.33G, +0.0569 ppl @ Llama-3-8B",  },
+    { "Q6_K",     LLAMA_FTYPE_MOSTLY_Q6_K,     " 6.14G, +0.0217 ppl @ Llama-3-8B",  },
+    { "Q8_0",     LLAMA_FTYPE_MOSTLY_Q8_0,     " 7.96G, +0.0026 ppl @ Llama-3-8B",  },
+    { "F16",      LLAMA_FTYPE_MOSTLY_F16,      "14.00G, +0.0020 ppl @ Mistral-7B",  },
+    { "BF16",     LLAMA_FTYPE_MOSTLY_BF16,     "14.00G, -0.0050 ppl @ Mistral-7B",  },
+    { "F32",      LLAMA_FTYPE_ALL_F32,         "26.00G              @ 7B",          },
+    // Note: Ensure COPY comes after F32 to avoid ftype 0 from matching.
+    { "COPY",     LLAMA_FTYPE_ALL_F32,         "only copy tensors, no quantizing",  },
+};
+
+// Quantization types. Changes to this struct must be replicated in llama-quantize.cpp
+struct tensor_quantization {
+    std::string name;
+    ggml_type quant = GGML_TYPE_COUNT;
+};
+
+static const char * const LLM_KV_QUANTIZE_IMATRIX_FILE       = "quantize.imatrix.file";
+static const char * const LLM_KV_QUANTIZE_IMATRIX_DATASET    = "quantize.imatrix.dataset";
+static const char * const LLM_KV_QUANTIZE_IMATRIX_N_ENTRIES  = "quantize.imatrix.entries_count";
+static const char * const LLM_KV_QUANTIZE_IMATRIX_N_CHUNKS   = "quantize.imatrix.chunks_count";
+
+// TODO: share with imatrix.cpp
+static const char * const LLM_KV_IMATRIX_DATASETS    = "imatrix.datasets";
+static const char * const LLM_KV_IMATRIX_CHUNK_COUNT = "imatrix.chunk_count";
+static const char * const LLM_KV_IMATRIX_CHUNK_SIZE  = "imatrix.chunk_size";
+
+static bool striequals(const char * a, const char * b) {
+    while (*a && *b) {
+        if (std::tolower(*a) != std::tolower(*b)) {
+            return false;
+        }
+        a++; b++;
+    }
+    return *a == *b;
+}
+
+static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftype, std::string & ftype_str_out) {
+    std::string ftype_str;
+
+    for (auto ch : ftype_str_in) {
+        ftype_str.push_back(std::toupper(ch));
+    }
+    for (const auto & it : QUANT_OPTIONS) {
+        if (striequals(it.name.c_str(), ftype_str.c_str())) {
+            ftype = it.ftype;
+            ftype_str_out = it.name;
+            return true;
+        }
+    }
+    try {
+        int ftype_int = std::stoi(ftype_str);
+        for (const auto & it : QUANT_OPTIONS) {
+            if (it.ftype == ftype_int) {
+                ftype = it.ftype;
+                ftype_str_out = it.name;
+                return true;
+            }
+        }
+    }
+    catch (...) {
+        // stoi failed
+    }
+    return false;
+}
+
+[[noreturn]]
+static void usage(const char * executable) {
+    printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] [--imatrix] [--include-weights]\n", executable);
+    printf("       [--exclude-weights] [--output-tensor-type] [--token-embedding-type] [--tensor-type] [--prune-layers] [--keep-split] [--override-kv]\n");
+    printf("       model-f32.gguf [model-quant.gguf] type [nthreads]\n\n");
+    printf("  --allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n");
+    printf("  --leave-output-tensor: Will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing\n");
+    printf("  --pure: Disable k-quant mixtures and quantize all tensors to the same type\n");
+    printf("  --imatrix file_name: use data in file_name as importance matrix for quant optimizations\n");
+    printf("  --include-weights tensor_name: use importance matrix for this/these tensor(s)\n");
+    printf("  --exclude-weights tensor_name: use importance matrix for this/these tensor(s)\n");
+    printf("  --output-tensor-type ggml_type: use this ggml_type for the output.weight tensor\n");
+    printf("  --token-embedding-type ggml_type: use this ggml_type for the token embeddings tensor\n");
+    printf("  --tensor-type TENSOR=TYPE: quantize this tensor to this ggml_type. example: --tensor-type attn_q=q8_0\n");
+    printf("      Advanced option to selectively quantize tensors. May be specified multiple times.\n");
+    printf("  --prune-layers L0,L1,L2...comma-separated list of layer numbers to prune from the model\n");
+    printf("      Advanced option to remove all tensors from the given layers\n");
+    printf("  --keep-split: will generate quantized model in the same shards as input\n");
+    printf("  --override-kv KEY=TYPE:VALUE\n");
+    printf("      Advanced option to override model metadata by key in the quantized model. May be specified multiple times.\n");
+    printf("Note: --include-weights and --exclude-weights cannot be used together\n");
+    printf("\nAllowed quantization types:\n");
+    for (const auto & it : QUANT_OPTIONS) {
+        if (it.name != "COPY") {
+            printf("  %2d  or  ", it.ftype);
+        } else {
+            printf("          ");
+        }
+        printf("%-7s : %s\n", it.name.c_str(), it.desc.c_str());
+    }
+    exit(1);
+}
+
+static int load_legacy_imatrix(const std::string & imatrix_file, std::vector<std::string> & imatrix_datasets, std::unordered_map<std::string, std::vector<float>> & imatrix_data) {
+    std::ifstream in(imatrix_file.c_str(), std::ios::binary);
+    if (!in) {
+        printf("%s: failed to open %s\n",__func__, imatrix_file.c_str());
+        exit(1);
+    }
+    int n_entries;
+    in.read((char *)&n_entries, sizeof(n_entries));
+    if (in.fail() || n_entries < 1) {
+        printf("%s: no data in file %s\n", __func__, imatrix_file.c_str());
+        exit(1);
+    }
+    for (int i = 0; i < n_entries; ++i) {
+        int len; in.read((char *)&len, sizeof(len));
+        std::vector<char> name_as_vec(len+1);
+        in.read((char *)name_as_vec.data(), len);
+        if (in.fail()) {
+            printf("%s: failed reading name for entry %d from %s\n", __func__, i+1, imatrix_file.c_str());
+            exit(1);
+        }
+        name_as_vec[len] = 0;
+        std::string name{name_as_vec.data()};
+        auto & e = imatrix_data[name];
+        int ncall;
+        in.read((char *)&ncall, sizeof(ncall));
+        int nval;
+        in.read((char *)&nval, sizeof(nval));
+        if (in.fail() || nval < 1) {
+            printf("%s: failed reading number of values for entry %d\n", __func__, i);
+            imatrix_data = {};
+            exit(1);
+        }
+        e.resize(nval);
+        in.read((char *)e.data(), nval*sizeof(float));
+        if (in.fail()) {
+            printf("%s: failed reading data for entry %d\n", __func__, i);
+            imatrix_data = {};
+            exit(1);
+        }
+        if (ncall > 0) {
+            for (auto & v : e) {
+                v /= ncall;
+            }
+        }
+
+        if (getenv("LLAMA_TRACE")) {
+            printf("%s: loaded data (size = %6d, ncall = %6d) for '%s'\n", __func__, int(e.size()), ncall, name.c_str());
+        }
+    }
+
+    // latest legacy imatrix version contains the dataset filename at the end of the file
+    int m_last_call = 0;
+    if (in.peek() != EOF) {
+        in.read((char *)&m_last_call, sizeof(m_last_call));
+        int dataset_len;
+        in.read((char *)&dataset_len, sizeof(dataset_len));
+        std::vector<char> dataset_as_vec(dataset_len);
+        in.read(dataset_as_vec.data(), dataset_len);
+        imatrix_datasets.resize(1);
+        imatrix_datasets[0].assign(dataset_as_vec.begin(), dataset_as_vec.end());
+        printf("%s: imatrix dataset='%s'\n", __func__, imatrix_datasets[0].c_str());
+    }
+    printf("%s: loaded %d importance matrix entries from %s computed on %d chunks\n", __func__, int(imatrix_data.size()), imatrix_file.c_str(), m_last_call);
+    return m_last_call;
+}
+
+static int load_imatrix(const std::string & imatrix_file, std::vector<std::string> & imatrix_datasets, std::unordered_map<std::string, std::vector<float>> & imatrix_data) {
+
+    struct ggml_context * ctx = nullptr;
+    struct gguf_init_params meta_gguf_params = {
+        /* .no_alloc = */ false, // the data is needed
+        /* .ctx      = */ &ctx,
+    };
+    struct gguf_context * ctx_gguf = gguf_init_from_file(imatrix_file.c_str(), meta_gguf_params);
+    if (!ctx_gguf) {
+        fprintf(stderr, "%s: imatrix file '%s' is using old format\n", __func__, imatrix_file.c_str());
+        return load_legacy_imatrix(imatrix_file, imatrix_datasets, imatrix_data);
+    }
+    const int32_t n_entries = gguf_get_n_tensors(ctx_gguf);
+    if (n_entries < 1) {
+        fprintf(stderr, "%s: no data in file %s\n", __func__, imatrix_file.c_str());
+        gguf_free(ctx_gguf);
+        ggml_free(ctx);
+        exit(1);
+    }
+
+    const int dataset_idx     = gguf_find_key(ctx_gguf, LLM_KV_IMATRIX_DATASETS);
+    const int chunk_count_idx = gguf_find_key(ctx_gguf, LLM_KV_IMATRIX_CHUNK_COUNT);
+    const int chunk_size_idx  = gguf_find_key(ctx_gguf, LLM_KV_IMATRIX_CHUNK_SIZE);
+    if (dataset_idx < 0 || chunk_count_idx < 0 || chunk_size_idx < 0) {
+        fprintf(stderr, "%s: missing imatrix metadata in file %s\n", __func__, imatrix_file.c_str());
+        gguf_free(ctx_gguf);
+        ggml_free(ctx);
+        exit(1);
+    }
+
+    const uint32_t chunk_size = gguf_get_val_u32(ctx_gguf, chunk_size_idx);
+
+    const std::string sums_suffix{ ".in_sum2" };
+    const std::string counts_suffix{ ".counts" };
+
+    // Using an ordered map to get a deterministic iteration order.
+    std::map<std::string, std::pair<struct ggml_tensor *, struct ggml_tensor *>> sums_counts_for;
+
+    for (struct ggml_tensor * cur = ggml_get_first_tensor(ctx); cur; cur = ggml_get_next_tensor(ctx, cur)) {
+        std::string name = cur->name;
+
+        if (name.empty()) { continue; }
+
+        if (string_remove_suffix(name, sums_suffix)) {
+            // in_sum2
+            sums_counts_for[std::move(name)].first = cur;
+        } else if (string_remove_suffix(name, counts_suffix)) {
+            // counts
+            sums_counts_for[std::move(name)].second = cur;
+        } else {
+            // ignore other tensors
+        }
+    }
+
+    for (const auto & sc : sums_counts_for) {
+        const        std::string & name   = sc.first;
+        const struct ggml_tensor * sums   = sc.second.first;
+        const struct ggml_tensor * counts = sc.second.second;
+
+        if (!sums || !counts) {
+            fprintf(stderr, "%s: mismatched sums and counts for %s\n", __func__, name.c_str());
+            gguf_free(ctx_gguf);
+            ggml_free(ctx);
+            exit(1);
+        }
+
+        const int64_t ne0 = sums->ne[0];
+        const int64_t ne1 = sums->ne[1];
+
+        auto & e = imatrix_data[name];
+        e.resize(ggml_nelements(sums));
+        float max_count = 0.0f;
+        for (int64_t j = 0; j < ne1; ++j) {
+            const float count = ((const float *) counts->data)[j];
+            if (count > 0.0f) {
+                for (int64_t i = 0; i < ne0; ++i) {
+                    e[j*ne0 + i] = ((const float *) sums->data)[j*ne0 + i] / count;
+                }
+            } else {
+                // Partial imatrix data, this tensor never got any input during calibration
+                for (int64_t i = 0; i < ne0; ++i) {
+                    e[j*ne0 + i] = 1;
+                }
+            }
+            if (count > max_count) {
+                max_count = count;
+            }
+        }
+        if (getenv("LLAMA_TRACE")) {
+            printf("%s: loaded data (size = %6d, n_tokens = %6d, n_chunks = %6d) for '%s'\n", __func__, int(e.size()), int(max_count), int(max_count / chunk_size), name.c_str());
+        }
+    }
+
+    int m_last_chunk = gguf_get_val_u32(ctx_gguf, chunk_count_idx);
+
+    int64_t n_datasets = gguf_get_arr_n(ctx_gguf, dataset_idx);
+    imatrix_datasets.reserve(n_datasets);
+    for (int64_t i = 0; i < n_datasets; ++i) {
+        imatrix_datasets.push_back(gguf_get_arr_str(ctx_gguf, dataset_idx, i));
+    }
+    printf("%s: imatrix datasets=['%s'", __func__, imatrix_datasets[0].c_str());
+    for (size_t i = 1; i < imatrix_datasets.size(); ++i) {
+        printf(", '%s'", imatrix_datasets[i].c_str());
+    }
+    printf("]\n");
+
+    printf("%s: loaded %d importance matrix entries from %s computed on %d chunks\n", __func__, int(imatrix_data.size()), imatrix_file.c_str(), m_last_chunk);
+
+    gguf_free(ctx_gguf);
+    ggml_free(ctx);
+
+    return m_last_chunk;
+}
+
+static int prepare_imatrix(const std::string & imatrix_file,
+        std::vector<std::string> & imatrix_dataset,
+        const std::vector<std::string> & included_weights,
+        const std::vector<std::string> & excluded_weights,
+        std::unordered_map<std::string, std::vector<float>> & imatrix_data) {
+    int m_last_call = -1;
+    if (!imatrix_file.empty()) {
+        m_last_call = load_imatrix(imatrix_file, imatrix_dataset, imatrix_data);
+    }
+    if (imatrix_data.empty()) {
+        return m_last_call;
+    }
+    if (!excluded_weights.empty()) {
+        for (const auto & name : excluded_weights) {
+            for (auto it = imatrix_data.begin(); it != imatrix_data.end();) {
+                auto pos = it->first.find(name);
+                if (pos != std::string::npos) {
+                    it = imatrix_data.erase(it);
+                } else {
+                    ++it;
+                }
+            }
+        }
+    }
+    if (!included_weights.empty()) {
+        std::unordered_map<std::string, std::vector<float>> tmp;
+        for (const auto & name : included_weights) {
+            for (auto & e : imatrix_data) {
+                auto pos = e.first.find(name);
+                if (pos != std::string::npos) {
+                    tmp.emplace(std::move(e));
+                }
+            }
+        }
+        imatrix_data = std::move(tmp);
+    }
+    if (!imatrix_data.empty()) {
+        printf("%s: have %d importance matrix entries\n", __func__, int(imatrix_data.size()));
+    }
+    return m_last_call;
+}
+
+static ggml_type parse_ggml_type(const char * arg) {
+    for (int i = 0; i < GGML_TYPE_COUNT; ++i) {
+        auto type = (ggml_type)i;
+        const auto * name = ggml_type_name(type);
+        if (name && striequals(name, arg)) {
+            return type;
+        }
+    }
+    fprintf(stderr, "\n%s: invalid ggml_type '%s'\n\n", __func__, arg);
+    return GGML_TYPE_COUNT;
+}
+
+static bool parse_tensor_type(const char * data, std::vector<tensor_quantization> & tensor_type) {
+    const char * sep = strchr(data, '=');
+    if (sep == nullptr) {
+        printf("\n%s: malformed tensor type '%s'\n\n", __func__, data);
+        return false;
+    }
+
+    const size_t tn_len = sep - data;
+    if (tn_len == 0) {
+        printf("\n%s: missing tensor name\n\n", __func__);
+        return false;
+    }
+    if (const size_t qt_len = strlen(sep); qt_len == 1) {
+        printf("\n%s: missing quantization type\n\n", __func__);
+        return false;
+    }
+
+    std::string tn(data, tn_len);
+    std::transform(tn.begin(), tn.end(), tn.begin(), tolower);
+    sep++;
+    tensor_quantization tqz;
+    tqz.name = tn;
+    tqz.quant = parse_ggml_type(sep);
+    tensor_type.emplace_back(std::move(tqz));
+    if (tqz.quant == GGML_TYPE_COUNT) {
+        printf("\n%s: invalid quantization type '%s'\n\n", __func__, sep);
+        return false;
+    }
+
+    return true;
+}
+
+static bool parse_layer_prune(const char * data, std::vector<int> & prune_layers) {
+    if (!data) {
+        printf("\n%s: no layer pruning ids provided\n\n", __func__);
+        return false;
+    }
+
+    const auto block_ids = string_split<std::string>(data, ',');
+    for (const auto & block_id : block_ids) {
+        int id;
+        try {
+            id = std::stoi(block_id);
+        } catch (...) {
+            id = -1;
+        }
+        if (id < 0) {
+            printf("\n%s: invalid layer id '%s'\n\n", __func__, block_id.c_str());
+            return false;
+        }
+        prune_layers.emplace_back(id);
+    }
+
+    sort(prune_layers.begin(), prune_layers.end());
+    prune_layers.erase(std::unique(prune_layers.begin(), prune_layers.end()), prune_layers.end());
+    return true;
+}
+
+int main(int argc, char ** argv) {
+    if (argc < 3) {
+        usage(argv[0]);
+    }
+
+    llama_model_quantize_params params = llama_model_quantize_default_params();
+
+    int arg_idx = 1;
+    std::string imatrix_file;
+    std::vector<std::string> included_weights, excluded_weights;
+    std::vector<llama_model_kv_override> kv_overrides;
+    std::vector<tensor_quantization> tensor_types;
+    std::vector<int> prune_layers;
+
+    for (; arg_idx < argc && strncmp(argv[arg_idx], "--", 2) == 0; arg_idx++) {
+        if (strcmp(argv[arg_idx], "--leave-output-tensor") == 0) {
+            params.quantize_output_tensor = false;
+        } else if (strcmp(argv[arg_idx], "--output-tensor-type") == 0) {
+            if (arg_idx < argc-1) {
+                params.output_tensor_type = parse_ggml_type(argv[++arg_idx]);
+                if (params.output_tensor_type == GGML_TYPE_COUNT) {
+                    usage(argv[0]);
+                }
+            } else {
+                usage(argv[0]);
+            }
+        } else if (strcmp(argv[arg_idx], "--token-embedding-type") == 0) {
+            if (arg_idx < argc-1) {
+                params.token_embedding_type = parse_ggml_type(argv[++arg_idx]);
+                if (params.token_embedding_type == GGML_TYPE_COUNT) {
+                    usage(argv[0]);
+                }
+            } else {
+                usage(argv[0]);
+            }
+        } else if (strcmp(argv[arg_idx], "--tensor-type") == 0) {
+            if (arg_idx == argc-1 || !parse_tensor_type(argv[++arg_idx], tensor_types)) {
+                usage(argv[0]);
+            }
+        } else if (strcmp(argv[arg_idx], "--prune-layers") == 0) {
+            if (arg_idx == argc-1 || !parse_layer_prune(argv[++arg_idx], prune_layers)) {
+                usage(argv[0]);
+            }
+        } else if (strcmp(argv[arg_idx], "--override-kv") == 0) {
+            if (arg_idx == argc-1 || !string_parse_kv_override(argv[++arg_idx], kv_overrides)) {
+                usage(argv[0]);
+            }
+        } else if (strcmp(argv[arg_idx], "--allow-requantize") == 0) {
+            params.allow_requantize = true;
+        } else if (strcmp(argv[arg_idx], "--pure") == 0) {
+            params.pure = true;
+        } else if (strcmp(argv[arg_idx], "--imatrix") == 0) {
+            if (arg_idx < argc-1) {
+                imatrix_file = argv[++arg_idx];
+            } else {
+                usage(argv[0]);
+            }
+        } else if (strcmp(argv[arg_idx], "--include-weights") == 0) {
+            if (arg_idx < argc-1) {
+                included_weights.emplace_back(argv[++arg_idx]);
+            } else {
+                usage(argv[0]);
+            }
+        } else if (strcmp(argv[arg_idx], "--exclude-weights") == 0) {
+            if (arg_idx < argc-1) {
+                excluded_weights.emplace_back(argv[++arg_idx]);
+            } else {
+                usage(argv[0]);
+            }
+        } else if (strcmp(argv[arg_idx], "--keep-split") == 0) {
+            params.keep_split = true;
+        } else {
+            usage(argv[0]);
+        }
+    }
+
+    if (argc - arg_idx < 2) {
+        printf("%s: bad arguments\n", argv[0]);
+        usage(argv[0]);
+    }
+    if (!included_weights.empty() && !excluded_weights.empty()) {
+        usage(argv[0]);
+    }
+
+    std::vector<std::string> imatrix_datasets;
+    std::unordered_map<std::string, std::vector<float>> imatrix_data;
+    int m_last_call = prepare_imatrix(imatrix_file, imatrix_datasets, included_weights, excluded_weights, imatrix_data);
+    if (!imatrix_data.empty()) {
+        params.imatrix = &imatrix_data;
+        {
+            llama_model_kv_override kvo;
+            std::strcpy(kvo.key, LLM_KV_QUANTIZE_IMATRIX_FILE);
+            kvo.tag = LLAMA_KV_OVERRIDE_TYPE_STR;
+            strncpy(kvo.val_str, imatrix_file.c_str(), 127);
+            kvo.val_str[127] = '\0';
+            kv_overrides.emplace_back(std::move(kvo));
+        }
+        if (!imatrix_datasets.empty()) {
+            llama_model_kv_override kvo;
+            // TODO: list multiple datasets when there are more than one
+            std::strcpy(kvo.key, LLM_KV_QUANTIZE_IMATRIX_DATASET);
+            kvo.tag = LLAMA_KV_OVERRIDE_TYPE_STR;
+            strncpy(kvo.val_str, imatrix_datasets[0].c_str(), 127);
+            kvo.val_str[127] = '\0';
+            kv_overrides.emplace_back(std::move(kvo));
+        }
+
+        {
+            llama_model_kv_override kvo;
+            std::strcpy(kvo.key, LLM_KV_QUANTIZE_IMATRIX_N_ENTRIES);
+            kvo.tag = LLAMA_KV_OVERRIDE_TYPE_INT;
+            kvo.val_i64 = imatrix_data.size();
+            kv_overrides.emplace_back(std::move(kvo));
+        }
+
+        if (m_last_call > 0) {
+            llama_model_kv_override kvo;
+            std::strcpy(kvo.key, LLM_KV_QUANTIZE_IMATRIX_N_CHUNKS);
+            kvo.tag = LLAMA_KV_OVERRIDE_TYPE_INT;
+            kvo.val_i64 = m_last_call;
+            kv_overrides.emplace_back(std::move(kvo));
+        }
+    }
+    if (!kv_overrides.empty()) {
+        kv_overrides.emplace_back();
+        kv_overrides.back().key[0] = 0;
+        params.kv_overrides = &kv_overrides;
+    }
+    if (!tensor_types.empty()) {
+        params.tensor_types = &tensor_types;
+    }
+    if (!prune_layers.empty()) {
+        params.prune_layers = &prune_layers;
+    }
+
+    llama_backend_init();
+
+    // parse command line arguments
+    const std::string fname_inp = argv[arg_idx];
+    arg_idx++;
+    std::string fname_out;
+
+    std::string ftype_str;
+    std::string suffix = ".gguf";
+    if (try_parse_ftype(argv[arg_idx], params.ftype, ftype_str)) {
+        std::string fpath;
+        const size_t pos = fname_inp.find_last_of("/\\");
+        if (pos != std::string::npos) {
+            fpath = fname_inp.substr(0, pos + 1);
+        }
+
+        // export as [inp path]/ggml-model-[ftype]. Only add extension if there is no splitting
+        fname_out = fpath + "ggml-model-" + ftype_str;
+        if (!params.keep_split) {
+            fname_out += suffix;
+        }
+        arg_idx++;
+        if (ftype_str == "COPY") {
+            params.only_copy = true;
+        }
+    } else {
+        fname_out = argv[arg_idx];
+        if (params.keep_split && fname_out.find(suffix) != std::string::npos) {
+            fname_out = fname_out.substr(0, fname_out.length() - suffix.length());
+        }
+        arg_idx++;
+
+        if (argc <= arg_idx) {
+            fprintf(stderr, "%s: missing ftype\n", __func__);
+            return 1;
+        }
+        if (!try_parse_ftype(argv[arg_idx], params.ftype, ftype_str)) {
+            fprintf(stderr, "%s: invalid ftype '%s'\n", __func__, argv[arg_idx]);
+            return 1;
+        }
+        if (ftype_str == "COPY") {
+           params.only_copy = true;
+        }
+        arg_idx++;
+    }
+
+    // parse nthreads
+    if (argc > arg_idx) {
+        try {
+            params.nthread = std::stoi(argv[arg_idx]);
+        }
+        catch (const std::exception & e) {
+            fprintf(stderr, "%s: invalid nthread '%s' (%s)\n", __func__, argv[arg_idx], e.what());
+            return 1;
+        }
+    }
+
+    if ((params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS ||
+         params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_S  ||
+         params.ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S ||
+         params.ftype == LLAMA_FTYPE_MOSTLY_IQ1_S  ||
+         params.ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) && imatrix_data.empty()) {
+        fprintf(stderr, "\n==========================================================================================================\n");
+        fprintf(stderr, "Please do not use IQ1_S, IQ1_M, IQ2_S, IQ2_XXS, IQ2_XS or Q2_K_S quantization without an importance matrix\n");
+        fprintf(stderr, "==========================================================================================================\n\n\n");
+        return 1;
+    }
+
+    if (std::error_code ec; std::filesystem::equivalent(fname_inp, fname_out, ec)) {
+        fprintf(stderr, "%s: error: input and output files are the same: '%s'\n", __func__, fname_inp.c_str());
+        return 1;
+    }
+
+    print_build_info();
+
+    fprintf(stderr, "%s: quantizing '%s' to '%s' as %s", __func__, fname_inp.c_str(), fname_out.c_str(), ftype_str.c_str());
+    if (params.nthread > 0) {
+        fprintf(stderr, " using %d threads", params.nthread);
+    }
+    fprintf(stderr, "\n");
+
+    const int64_t t_main_start_us = llama_time_us();
+
+    int64_t t_quantize_us = 0;
+
+    // load the model
+    {
+        const int64_t t_start_us = llama_time_us();
+
+        if (llama_model_quantize(fname_inp.c_str(), fname_out.c_str(), &params)) {
+            fprintf(stderr, "%s: failed to quantize model from '%s'\n", __func__, fname_inp.c_str());
+            return 1;
+        }
+
+        t_quantize_us = llama_time_us() - t_start_us;
+    }
+
+    // report timing
+    {
+        const int64_t t_main_end_us = llama_time_us();
+
+        printf("\n");
+        printf("%s: quantize time = %8.2f ms\n", __func__, t_quantize_us/1000.0);
+        printf("%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0);
+    }
+
+    llama_backend_free();
+
+    return 0;
+}

tools/quantize/tests.sh

@@ -0,0 +1,65 @@
+#!/usr/bin/env bash
+
+set -eu
+
+if [ $# -lt 1 ]
+then
+    echo "usage:   $0 path_to_build_binary [path_to_temp_folder]"
+    echo "example: $0 ../../build/bin ../../tmp"
+    exit 1
+fi
+
+if [ $# -gt 1 ]
+then
+    TMP_DIR=$2
+else
+    TMP_DIR=/tmp
+fi
+
+set -x
+
+SPLIT=$1/llama-gguf-split
+QUANTIZE=$1/llama-quantize
+MAIN=$1/llama-completion
+WORK_PATH=$TMP_DIR/quantize
+ROOT_DIR=$(realpath $(dirname $0)/../../)
+
+mkdir -p "$WORK_PATH"
+
+# Clean up in case of previously failed test
+rm -f $WORK_PATH/ggml-model-split*.gguf $WORK_PATH/ggml-model-requant*.gguf
+
+# 1. Get a model
+(
+cd $WORK_PATH
+"$ROOT_DIR"/scripts/hf.sh --repo ggml-org/Qwen3-0.6B-GGUF --file Qwen3-0.6B-Q8_0.gguf
+)
+echo PASS
+
+# 2. Split model
+$SPLIT --split-max-tensors 28  $WORK_PATH/Qwen3-0.6B-Q8_0.gguf $WORK_PATH/ggml-model-split
+echo PASS
+echo
+
+# 3. Requant model with '--keep-split'
+$QUANTIZE --allow-requantize --keep-split $WORK_PATH/ggml-model-split-00001-of-00012.gguf $WORK_PATH/ggml-model-requant.gguf Q4_K
+echo PASS
+echo
+
+# 3a. Test the requanted model is loading properly
+$MAIN -no-cnv --model $WORK_PATH/ggml-model-requant-00001-of-00012.gguf -p "I believe the meaning of life is" --n-predict 32
+echo PASS
+echo
+
+# 4. Requant mode without '--keep-split'
+$QUANTIZE --allow-requantize $WORK_PATH/ggml-model-split-00001-of-00012.gguf $WORK_PATH/ggml-model-requant-merge.gguf Q4_K
+echo PASS
+echo
+
+# 4b. Test the requanted model is loading properly
+$MAIN -no-cnv --model $WORK_PATH/ggml-model-requant-merge.gguf -p "I believe the meaning of life is" --n-predict 32
+echo PASS
+echo
+
+# Clean up
+rm -f $WORK_PATH/ggml-model-split*.gguf $WORK_PATH/ggml-model-requant*.gguf

tools/rpc/CMakeLists.txt

@@ -0,0 +1,8 @@
+set(TARGET rpc-server)
+add_executable(${TARGET} rpc-server.cpp)
+target_link_libraries(${TARGET} PRIVATE ggml)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)
+
+if(LLAMA_TOOLS_INSTALL)
+    install(TARGETS ${TARGET} RUNTIME)
+endif()

tools/rpc/README.md

@@ -0,0 +1,104 @@
+## Overview
+
+> [!IMPORTANT]
+> This example and the RPC backend are currently in a proof-of-concept development stage. As such, the functionality is fragile and
+> insecure. **Never run the RPC server on an open network or in a sensitive environment!**
+
+The `rpc-server` allows exposing `ggml` devices on a remote host.
+The RPC backend communicates with one or several instances of `rpc-server` and offloads computations to them.
+This can be used for distributed LLM inference with `llama.cpp` in the following way:
+
+```mermaid
+flowchart TD
+    rpcb<-->|TCP|srva
+    rpcb<-->|TCP|srvb
+    rpcb<-.->|TCP|srvn
+    subgraph hostn[Host N]
+    srvn[rpc-server]<-.->dev4["CUDA0"]
+    srvn[rpc-server]<-.->dev5["CPU"]
+    end
+    subgraph hostb[Host B]
+    srvb[rpc-server]<-->dev3["Metal"]
+    end
+    subgraph hosta[Host A]
+    srva[rpc-server]<-->dev["CUDA0"]
+    srva[rpc-server]<-->dev2["CUDA1"]
+    end
+    subgraph host[Main Host]
+    local["Local devices"]<-->ggml[llama-cli]
+    ggml[llama-cli]<-->rpcb[RPC backend]
+    end
+    style hostn stroke:#66,stroke-width:2px,stroke-dasharray: 5 5
+    classDef devcls fill:#5B9BD5
+    class local,dev,dev2,dev3,dev4,dev5 devcls
+```
+
+By default, `rpc-server` exposes all available accelerator devices on the host.
+If there are no accelerators, it exposes a single `CPU` device.
+
+## Usage
+
+### Remote hosts
+
+On each remote host, build the backends for each accelerator by adding `-DGGML_RPC=ON` to the build options.
+For example, to build the `rpc-server` with support for CUDA accelerators:
+
+```bash
+mkdir build-rpc-cuda
+cd build-rpc-cuda
+cmake .. -DGGML_CUDA=ON -DGGML_RPC=ON
+cmake --build . --config Release
+```
+
+When started, the `rpc-server` will detect and expose all available `CUDA` devices:
+
+```bash
+$ bin/rpc-server
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 CUDA devices:
+  Device 0: NVIDIA GeForce RTX 5090, compute capability 12.0, VMM: yes
+Starting RPC server v3.0.0
+  endpoint       : 127.0.0.1:50052
+  local cache    : n/a
+Devices:
+  CUDA0: NVIDIA GeForce RTX 5090 (32109 MiB, 31588 MiB free)
+```
+
+You can control the set of exposed CUDA devices with the `CUDA_VISIBLE_DEVICES` environment variable or the `--device` command line option. The following two commands have the same effect:
+```bash
+$ CUDA_VISIBLE_DEVICES=0 bin/rpc-server -p 50052
+$ bin/rpc-server --device CUDA0 -p 50052
+```
+
+### Main host
+
+On the main host build `llama.cpp` with the backends for the local devices and add `-DGGML_RPC=ON` to the build options.
+Finally, when running `llama-cli` or `llama-server`, use the `--rpc` option to specify the host and port of each `rpc-server`:
+
+```bash
+$ llama-cli -hf ggml-org/gemma-3-1b-it-GGUF -ngl 99 --rpc 192.168.88.10:50052,192.168.88.11:50052
+```
+
+By default, llama.cpp distributes model weights and the KV cache across all available devices -- both local and remote -- in proportion to each device's available memory.
+You can override this behavior with the `--tensor-split` option and set custom proportions when splitting tensor data across devices.
+
+### Local cache
+
+The RPC server can use a local cache to store large tensors and avoid transferring them over the network.
+This can speed up model loading significantly, especially when using large models.
+To enable the cache, use the `-c` option:
+
+```bash
+$ bin/rpc-server -c
+```
+
+By default, the cache is stored in the `$HOME/.cache/llama.cpp/rpc` directory and can be controlled via the `LLAMA_CACHE` environment variable.
+
+### Troubleshooting
+
+Use the `GGML_RPC_DEBUG` environment variable to enable debug messages from `rpc-server`:
+```bash
+$ GGML_RPC_DEBUG=1 bin/rpc-server
+```
+

tools/rpc/rpc-server.cpp

@@ -0,0 +1,302 @@
+#if defined(_MSC_VER)
+#define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING
+#endif
+
+#include "ggml-rpc.h"
+#ifdef _WIN32
+#  define NOMINMAX
+#  define DIRECTORY_SEPARATOR '\\'
+#  include <locale>
+#  include <windows.h>
+#  include <fcntl.h>
+#  include <io.h>
+#else
+#  define DIRECTORY_SEPARATOR '/'
+#  include <unistd.h>
+#  include <sys/stat.h>
+#endif
+#include <codecvt>
+#include <string>
+#include <stdio.h>
+#include <vector>
+#include <filesystem>
+#include <algorithm>
+#include <thread>
+#include <regex>
+
+namespace fs = std::filesystem;
+
+// NOTE: this is copied from common.cpp to avoid linking with libcommon
+// returns true if successful, false otherwise
+static bool fs_create_directory_with_parents(const std::string & path) {
+#ifdef _WIN32
+    std::wstring_convert<std::codecvt_utf8<wchar_t>> converter;
+    std::wstring wpath = converter.from_bytes(path);
+
+    // if the path already exists, check whether it's a directory
+    const DWORD attributes = GetFileAttributesW(wpath.c_str());
+    if ((attributes != INVALID_FILE_ATTRIBUTES) && (attributes & FILE_ATTRIBUTE_DIRECTORY)) {
+        return true;
+    }
+
+    size_t pos_slash = 0;
+
+    // process path from front to back, procedurally creating directories
+    while ((pos_slash = path.find('\\', pos_slash)) != std::string::npos) {
+        const std::wstring subpath = wpath.substr(0, pos_slash);
+        const wchar_t * test = subpath.c_str();
+
+        const bool success = CreateDirectoryW(test, NULL);
+        if (!success) {
+            const DWORD error = GetLastError();
+
+            // if the path already exists, ensure that it's a directory
+            if (error == ERROR_ALREADY_EXISTS) {
+                const DWORD attributes = GetFileAttributesW(subpath.c_str());
+                if (attributes == INVALID_FILE_ATTRIBUTES || !(attributes & FILE_ATTRIBUTE_DIRECTORY)) {
+                    return false;
+                }
+            } else {
+                return false;
+            }
+        }
+
+        pos_slash += 1;
+    }
+
+    return true;
+#else
+    // if the path already exists, check whether it's a directory
+    struct stat info;
+    if (stat(path.c_str(), &info) == 0) {
+        return S_ISDIR(info.st_mode);
+    }
+
+    size_t pos_slash = 1; // skip leading slashes for directory creation
+
+    // process path from front to back, procedurally creating directories
+    while ((pos_slash = path.find('/', pos_slash)) != std::string::npos) {
+        const std::string subpath = path.substr(0, pos_slash);
+        struct stat info;
+
+        // if the path already exists, ensure that it's a directory
+        if (stat(subpath.c_str(), &info) == 0) {
+            if (!S_ISDIR(info.st_mode)) {
+                return false;
+            }
+        } else {
+            // create parent directories
+            const int ret = mkdir(subpath.c_str(), 0755);
+            if (ret != 0) {
+                return false;
+            }
+        }
+
+        pos_slash += 1;
+    }
+
+    return true;
+#endif // _WIN32
+}
+
+// NOTE: this is copied from common.cpp to avoid linking with libcommon
+static std::string fs_get_cache_directory() {
+    std::string cache_directory = "";
+    auto ensure_trailing_slash = [](std::string p) {
+        // Make sure to add trailing slash
+        if (p.back() != DIRECTORY_SEPARATOR) {
+            p += DIRECTORY_SEPARATOR;
+        }
+        return p;
+    };
+    if (getenv("LLAMA_CACHE")) {
+        cache_directory = std::getenv("LLAMA_CACHE");
+    } else {
+#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX) || defined(__OpenBSD__)
+        if (std::getenv("XDG_CACHE_HOME")) {
+            cache_directory = std::getenv("XDG_CACHE_HOME");
+        } else {
+            cache_directory = std::getenv("HOME") + std::string("/.cache/");
+        }
+#elif defined(__APPLE__)
+        cache_directory = std::getenv("HOME") + std::string("/Library/Caches/");
+#elif defined(_WIN32)
+        cache_directory = std::getenv("LOCALAPPDATA");
+#else
+#  error Unknown architecture
+#endif
+        cache_directory = ensure_trailing_slash(cache_directory);
+        cache_directory += "llama.cpp";
+    }
+    return ensure_trailing_slash(cache_directory);
+}
+
+struct rpc_server_params {
+    std::string              host        = "127.0.0.1";
+    int                      port        = 50052;
+    bool                     use_cache   = false;
+    int                      n_threads   = std::max(1U, std::thread::hardware_concurrency()/2);
+    std::vector<std::string> devices;
+};
+
+static void print_usage(int /*argc*/, char ** argv, rpc_server_params params) {
+    fprintf(stderr, "Usage: %s [options]\n\n", argv[0]);
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h, --help                       show this help message and exit\n");
+    fprintf(stderr, "  -t, --threads N                  number of threads for the CPU device (default: %d)\n", params.n_threads);
+    fprintf(stderr, "  -d, --device <dev1,dev2,...>     comma-separated list of devices\n");
+    fprintf(stderr, "  -H, --host HOST                  host to bind to (default: %s)\n", params.host.c_str());
+    fprintf(stderr, "  -p, --port PORT                  port to bind to (default: %d)\n", params.port);
+    fprintf(stderr, "  -c, --cache                      enable local file cache\n");
+    fprintf(stderr, "\n");
+}
+
+static bool rpc_server_params_parse(int argc, char ** argv, rpc_server_params & params) {
+    std::string arg;
+    for (int i = 1; i < argc; i++) {
+        arg = argv[i];
+        if (arg == "-H" || arg == "--host") {
+            if (++i >= argc) {
+                return false;
+            }
+            params.host = argv[i];
+        } else if (arg == "-t" || arg == "--threads") {
+            if (++i >= argc) {
+                return false;
+            }
+            params.n_threads = std::stoi(argv[i]);
+            if (params.n_threads <= 0) {
+                fprintf(stderr, "error: invalid number of threads: %d\n", params.n_threads);
+                return false;
+            }
+        } else if (arg == "-d" || arg == "--device") {
+            if (++i >= argc) {
+                return false;
+            }
+            const std::regex regex{ R"([,/]+)" };
+            std::string dev_str = argv[i];
+            std::sregex_token_iterator iter(dev_str.begin(), dev_str.end(), regex, -1);
+            std::sregex_token_iterator end;
+            for ( ; iter != end; ++iter) {
+                try {
+                    params.devices.push_back(*iter);
+                } catch (const std::exception & ) {
+                    fprintf(stderr, "error: invalid device: %s\n", iter->str().c_str());
+                    return false;
+                }
+            }
+        } else if (arg == "-p" || arg == "--port") {
+            if (++i >= argc) {
+                return false;
+            }
+            params.port = std::stoi(argv[i]);
+            if (params.port <= 0 || params.port > 65535) {
+                return false;
+            }
+        } else if (arg == "-c" || arg == "--cache") {
+            params.use_cache = true;
+        } else if (arg == "-h" || arg == "--help") {
+            print_usage(argc, argv, params);
+            exit(0);
+        } else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            print_usage(argc, argv, params);
+            exit(0);
+        }
+    }
+    return true;
+}
+
+static std::vector<ggml_backend_dev_t> get_devices(const rpc_server_params & params) {
+    std::vector<ggml_backend_dev_t> devices;
+    if (!params.devices.empty()) {
+        for (auto device : params.devices) {
+            ggml_backend_dev_t dev = ggml_backend_dev_by_name(device.c_str());
+            if (dev) {
+                devices.push_back(dev);
+            } else {
+                fprintf(stderr, "error: unknown device: %s\n", device.c_str());
+                fprintf(stderr, "available devices:\n");
+                for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
+                    auto * dev = ggml_backend_dev_get(i);
+                    size_t free, total;
+                    ggml_backend_dev_memory(dev, &free, &total);
+                    printf("  %s: %s (%zu MiB, %zu MiB free)\n", ggml_backend_dev_name(dev), ggml_backend_dev_description(dev), total / 1024 / 1024, free / 1024 / 1024);
+                }
+                return {};
+            }
+        }
+    }
+
+    // Try non-CPU devices first
+    if (devices.empty()) {
+        for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
+            ggml_backend_dev_t dev = ggml_backend_dev_get(i);
+            if (ggml_backend_dev_type(dev) != GGML_BACKEND_DEVICE_TYPE_CPU) {
+                devices.push_back(dev);
+            }
+        }
+    }
+
+    // If there are no accelerators, fallback to CPU device
+    if (devices.empty()) {
+        ggml_backend_dev_t dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+        if (dev) {
+            devices.push_back(dev);
+        }
+    }
+
+    return devices;
+}
+
+int main(int argc, char * argv[]) {
+    ggml_backend_load_all();
+
+    rpc_server_params params;
+    if (!rpc_server_params_parse(argc, argv, params)) {
+        fprintf(stderr, "Invalid parameters\n");
+        return 1;
+    }
+
+    if (params.host != "127.0.0.1") {
+        fprintf(stderr, "\n");
+        fprintf(stderr, "!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!\n");
+        fprintf(stderr, "WARNING: Host ('%s') is != '127.0.0.1'\n", params.host.c_str());
+        fprintf(stderr, "         Never expose the RPC server to an open network!\n");
+        fprintf(stderr, "         This is an experimental feature and is not secure!\n");
+        fprintf(stderr, "!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!\n");
+        fprintf(stderr, "\n");
+    }
+
+    auto devices = get_devices(params);
+    if (devices.empty()) {
+        fprintf(stderr, "No devices found\n");
+        return 1;
+    }
+    std::string endpoint = params.host + ":" + std::to_string(params.port);
+    const char * cache_dir = nullptr;
+    std::string cache_dir_str;
+    if (params.use_cache) {
+        cache_dir_str = fs_get_cache_directory() + "rpc/";
+        if (!fs_create_directory_with_parents(cache_dir_str)) {
+            fprintf(stderr, "Failed to create cache directory: %s\n", cache_dir_str.c_str());
+            return 1;
+        }
+        cache_dir = cache_dir_str.c_str();
+    }
+
+    ggml_backend_reg_t reg = ggml_backend_reg_by_name("RPC");
+    if (!reg) {
+        fprintf(stderr, "Failed to find RPC backend\n");
+        return 1;
+    }
+
+    auto start_server_fn = (decltype(ggml_backend_rpc_start_server)*) ggml_backend_reg_get_proc_address(reg, "ggml_backend_rpc_start_server");
+    if (!start_server_fn) {
+        fprintf(stderr, "Failed to obtain RPC backend start server function\n");
+        return 1;
+    }
+
+    start_server_fn(endpoint.c_str(), cache_dir, params.n_threads, devices.size(), devices.data());
+    return 0;
+}

tools/server/CMakeLists.txt

@@ -0,0 +1,70 @@
+include_directories(${CMAKE_CURRENT_SOURCE_DIR} ${CMAKE_CURRENT_BINARY_DIR})
+
+# server-context containing the core server logic, used by llama-server and CLI
+
+set(TARGET server-context)
+
+add_library(${TARGET} STATIC
+    server-task.cpp
+    server-task.h
+    server-queue.cpp
+    server-queue.h
+    server-common.cpp
+    server-common.h
+    server-context.cpp
+    server-context.h
+)
+
+if (BUILD_SHARED_LIBS)
+    set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+endif()
+
+target_include_directories(${TARGET} PRIVATE ../mtmd)
+target_include_directories(${TARGET} PRIVATE ${CMAKE_SOURCE_DIR})
+target_link_libraries(${TARGET} PUBLIC common mtmd ${CMAKE_THREAD_LIBS_INIT})
+
+
+# llama-server executable
+
+set(TARGET llama-server)
+
+if (NOT LLAMA_HTTPLIB)
+    message(FATAL_ERROR "LLAMA_HTTPLIB is OFF, cannot build llama-server. Hint: to skip building server, set -DLLAMA_BUILD_SERVER=OFF")
+endif()
+
+set(TARGET_SRCS
+    server.cpp
+    server-http.cpp
+    server-http.h
+    server-models.cpp
+    server-models.h
+)
+set(PUBLIC_ASSETS
+    index.html.gz
+    loading.html
+)
+
+foreach(asset ${PUBLIC_ASSETS})
+    set(input "${CMAKE_CURRENT_SOURCE_DIR}/public/${asset}")
+    set(output "${CMAKE_CURRENT_BINARY_DIR}/${asset}.hpp")
+    list(APPEND TARGET_SRCS ${output})
+    add_custom_command(
+        DEPENDS "${input}"
+        OUTPUT "${output}"
+        COMMAND "${CMAKE_COMMAND}" "-DINPUT=${input}" "-DOUTPUT=${output}" -P "${PROJECT_SOURCE_DIR}/scripts/xxd.cmake"
+    )
+    set_source_files_properties(${output} PROPERTIES GENERATED TRUE)
+endforeach()
+
+add_executable(${TARGET} ${TARGET_SRCS})
+install(TARGETS ${TARGET} RUNTIME)
+
+target_include_directories(${TARGET} PRIVATE ../mtmd)
+target_include_directories(${TARGET} PRIVATE ${CMAKE_SOURCE_DIR})
+target_link_libraries(${TARGET} PRIVATE server-context PUBLIC common cpp-httplib ${CMAKE_THREAD_LIBS_INIT})
+
+if (WIN32)
+    TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
+endif()
+
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

tools/server/README-dev.md

@@ -0,0 +1,179 @@
+# llama-server Development Documentation
+
+This document provides an in-depth technical overview of `llama-server`, intended for maintainers and contributors.
+
+If you are an end user consuming `llama-server` as a product, please refer to the main [README](./README.md) instead.
+
+## Backend
+
+### Overview
+
+The server supports two primary operating modes:
+
+- **Inference mode**: The default mode for performing inference with a single loaded GGUF model.
+- **Router mode**: Enables management of multiple inference server instances behind a single API endpoint. Requests are automatically routed to the appropriate backend instance based on the requested model.
+
+The core architecture consists of the following components:
+
+- `server_context`: Holds the primary inference state, including the main `llama_context` and all active slots.
+- `server_slot`: An abstraction over a single “sequence” in llama.cpp, responsible for managing individual parallel inference requests.
+- `server_routes`: Middleware layer between `server_context` and the HTTP interface; handles JSON parsing/formatting and request routing logic.
+- `server_http_context`: Implements the HTTP server using `cpp-httplib`.
+- `server_queue`: Thread-safe queue used by HTTP workers to submit new tasks to `server_context`.
+- `server_response`: Thread-safe queue used by `server_context` to return results to HTTP workers.
+- `server_response_reader`: Higher-level wrapper around the two queues above for cleaner code.
+- `server_task`: Unit of work pushed into `server_queue`.
+- `server_task_result`: Unit of result pushed into `server_response`.
+- `server_tokens`: Unified representation of token sequences (supports both text and multimodal tokens); used by `server_task` and `server_slot`.
+- `server_prompt_checkpoint`: For recurrent (e.g., RWKV) and SWA models, stores snapshots of KV cache state. Enables reuse when subsequent requests share the same prompt prefix, saving redundant computation.
+- `server_models`: Standalone component for managing multiple backend instances (used in router mode). It is completely independent of `server_context`.
+
+```mermaid
+graph TD
+    API_User <--> server_http_context
+    server_http_context <-- router mode --> server_models
+    server_http_context <-- inference mode --> server_routes
+    server_routes -- server_task --> server_queue
+    subgraph server_context
+        server_queue --> server_slot
+        server_slot -- server_task_result --> server_response
+        server_slot[multiple server_slot]
+    end
+    server_response --> server_routes
+```
+
+### Batching
+
+The server context maintains a single batch shared across all slots. When `update_slots()` is invoked, the system iterates through all active slots to populate this batch. For each slot, either a generated token from the previous decoding step or available prompt tokens are added to the batch.
+
+Batching constraints apply: slots can only be batched together if they share compatible configurations. For instance, slots using a specific LoRA adapter can be batched with each other, but not with slots using a different LoRA adapter or no adapter at all.
+
+Once the batch reaches capacity or all slots have been processed, `llama_decode` is called to execute the inference. This operation represents the primary computational bottleneck in `update_slots()`.
+
+Following decoding, the system either retrieves embeddings or samples the next token using `common_sampler_sample`. If a slot has remaining prompt tokens to process, it yields until the next `update_slots()` iteration.
+
+### Thread Management
+
+`server_context` runs on a dedicated single thread. Because it is single-threaded, heavy post-processing (especially after token generation) should be avoided, as it directly impacts multi-sequence throughput.
+
+Each incoming HTTP request is handled by its own thread managed by the HTTP library. The following operations are performed in HTTP worker threads:
+
+- JSON request parsing
+- Chat template application
+- Tokenization
+- Conversion of `server_task_result` into final JSON response
+- Error formatting into JSON
+- Tracking of partial/incremental responses (e.g., streaming tool calls or reasoning steps)
+
+**Best practices to follow:**
+
+- All JSON formatting and chat template logic must stay in the HTTP layer.
+- Avoid passing raw JSON between the HTTP layer and `server_slot`. Instead, parse everything into native C++ types as early as possible.
+
+### Example trace of a request
+
+Here is an example trace of an API request for text completion:
+
+- A request arrives at the HTTP layer.
+- The request is routed to the corresponding handler inside `server_routes`. In this case, `handle_completions_impl` is invoked.
+- The handler parses the input request, constructs a new `server_task`, and passes it to `server_res_generator`.
+- `server_res_generator` creates a new `task_result_state` for each task:
+    - `task_result_state` stays in the HTTP layer, responsible for keeping track of the current state of the response (e.g., parsing tool calls or thinking messages).
+    - `server_task` is moved into `server_queue` inside `server_context`.
+- `server_context` launches the task by moving it into an available slot (see `launch_slot_with_task()`).
+- `update_slot()` processes the task as described in the "Batching" section above.
+- Results may be sent using `send_partial_response` or `send_final_response`, which creates a new `server_task_result` and pushes it to the response queue.
+- At the same time, `server_res_generator` listens to the response queue and retrieves this response.
+- As the response is stateless, `server_res_generator` calls `response->update()` to update the response with the current state.
+- `server_res_generator` then calls `response->to_json()` and passes the response to the HTTP layer.
+
+### Testing
+
+`llama-server` includes an automated test suite based on `pytest`.
+
+The framework automatically starts a `llama-server` instance, sends requests, and validates responses.
+
+For detailed instructions, see the [test documentation](./tests/README.md).
+
+### Notable Related PRs
+
+- Initial server implementation: https://github.com/ggml-org/llama.cpp/pull/1443
+- Parallel decoding support: https://github.com/ggml-org/llama.cpp/pull/3228
+- Refactor introducing `server_queue` and `server_response`: https://github.com/ggml-org/llama.cpp/pull/5065
+- Reranking endpoint: https://github.com/ggml-org/llama.cpp/pull/9510
+- Multimodal model support (`libmtmd`): https://github.com/ggml-org/llama.cpp/pull/12898
+- Unified KV cache handling: https://github.com/ggml-org/llama.cpp/pull/16736
+- Separation of HTTP logic into dedicated files: https://github.com/ggml-org/llama.cpp/pull/17216
+- Large-scale code base split into smaller files: https://github.com/ggml-org/llama.cpp/pull/17362
+- Introduction of router mode: https://github.com/ggml-org/llama.cpp/pull/17470
+- Speculative decoding: https://github.com/ggml-org/llama.cpp/pull/17808 and rework in https://github.com/ggml-org/llama.cpp/pull/17808
+- INI presets: https://github.com/ggml-org/llama.cpp/pull/17859 (+ refactoring: https://github.com/ggml-org/llama.cpp/pull/18169)
+- Sleeping mode: https://github.com/ggml-org/llama.cpp/pull/18228
+
+
+
+
+## Web UI
+
+The project includes a web-based user interface for interacting with `llama-server`. It supports both single-model (`MODEL` mode) and multi-model (`ROUTER` mode) operation.
+
+The SvelteKit-based Web UI is introduced in this PR: https://github.com/ggml-org/llama.cpp/pull/14839
+
+### Features
+
+-   **Chat interface** with streaming responses
+-   **Multi-model support** (ROUTER mode) - switch between models, auto-load on selection
+-   **Modality validation** - ensures selected model supports conversation's attachments (images, audio)
+-   **Conversation management** - branching, regeneration, editing with history preservation
+-   **Attachment support** - images, audio, PDFs (with vision/text fallback)
+-   **Configurable parameters** - temperature, top_p, etc. synced with server defaults
+-   **Dark/light theme**
+
+### Tech Stack
+
+-   **SvelteKit** - frontend framework with Svelte 5 runes for reactive state
+-   **TailwindCSS** + **shadcn-svelte** - styling and UI components
+-   **Vite** - build tooling
+-   **IndexedDB** (Dexie) - local storage for conversations
+-   **LocalStorage** - user settings persistence
+
+### Architecture
+
+The WebUI follows a layered architecture:
+
+```
+Routes → Components → Hooks → Stores → Services → Storage/API
+```
+
+-   **Stores** - reactive state management (`chatStore`, `conversationsStore`, `modelsStore`, `serverStore`, `settingsStore`)
+-   **Services** - stateless API/database communication (`ChatService`, `ModelsService`, `PropsService`, `DatabaseService`)
+-   **Hooks** - reusable logic (`useModelChangeValidation`, `useProcessingState`)
+
+For detailed architecture diagrams, see [`tools/server/webui/docs/`](webui/docs/):
+
+-   `high-level-architecture.mmd` - full architecture with all modules
+-   `high-level-architecture-simplified.mmd` - simplified overview
+-   `data-flow-simplified-model-mode.mmd` - data flow for single-model mode
+-   `data-flow-simplified-router-mode.mmd` - data flow for multi-model mode
+-   `flows/*.mmd` - detailed per-domain flows (chat, conversations, models, etc.)
+
+### Development
+
+```sh
+# make sure you have Node.js installed
+cd tools/server/webui
+npm i
+
+# run dev server (with hot reload)
+npm run dev
+
+# run tests
+npm run test
+
+# build production bundle
+npm run build
+```
+
+After `public/index.html.gz` has been generated, rebuild `llama-server` as described in the [build](#build) section to include the updated UI.
+
+**Note:** The Vite dev server automatically proxies API requests to `http://localhost:8080`. Make sure `llama-server` is running on that port during development.

tools/server/bench/README.md

@@ -0,0 +1,119 @@
+### Server benchmark tools
+
+Benchmark is using [k6](https://k6.io/).
+
+##### Install k6 and sse extension
+
+SSE is not supported by default in k6, you have to build k6 with the [xk6-sse](https://github.com/phymbert/xk6-sse) extension.
+
+Example (assuming golang >= 1.21 is installed):
+```shell
+go install go.k6.io/xk6/cmd/xk6@latest
+$GOPATH/bin/xk6 build master \
+--with github.com/phymbert/xk6-sse
+```
+
+#### Download a dataset
+
+This dataset was originally proposed in [vLLM benchmarks](https://github.com/vllm-project/vllm/blob/main/benchmarks/README.md).
+
+```shell
+wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
+```
+
+#### Download a model
+Example for PHI-2
+
+```shell
+../../../scripts/hf.sh --repo ggml-org/models --file phi-2/ggml-model-q4_0.gguf
+```
+
+#### Start the server
+The server must answer OAI Chat completion requests on `http://localhost:8080/v1` or according to the environment variable `SERVER_BENCH_URL`.
+
+Example:
+```shell
+llama-server --host localhost --port 8080 \
+  --model ggml-model-q4_0.gguf \
+  --cont-batching \
+  --metrics \
+  --parallel 8 \
+  --batch-size 512 \
+  --ctx-size 4096 \
+  -ngl 33
+```
+
+#### Run the benchmark
+
+For 500 chat completions request with 8 concurrent users during maximum 10 minutes, run:
+```shell
+./k6 run script.js --duration 10m --iterations 500 --vus 8
+```
+
+The benchmark values can be overridden with:
+- `SERVER_BENCH_URL` server url prefix for chat completions, default `http://localhost:8080/v1`
+- `SERVER_BENCH_N_PROMPTS` total prompts to randomly select in the benchmark, default `480`
+- `SERVER_BENCH_MODEL_ALIAS` model alias to pass in the completion request, default `my-model`
+- `SERVER_BENCH_MAX_TOKENS` max tokens to predict, default: `512`
+- `SERVER_BENCH_DATASET` path to the benchmark dataset file
+- `SERVER_BENCH_MAX_PROMPT_TOKENS` maximum prompt tokens to filter out in the dataset: default `1024`
+- `SERVER_BENCH_MAX_CONTEXT` maximum context size of the completions request to filter out in the dataset: prompt + predicted tokens, default `2048`
+
+Note: the local tokenizer is just a string space split, real number of tokens will differ.
+
+Or with [k6 options](https://k6.io/docs/using-k6/k6-options/reference/):
+
+```shell
+SERVER_BENCH_N_PROMPTS=500 k6 run script.js --duration 10m --iterations 500 --vus 8
+```
+
+To [debug http request](https://k6.io/docs/using-k6/http-debugging/) use `--http-debug="full"`.
+
+#### Metrics
+
+Following metrics are available computed from the OAI chat completions response `usage`:
+- `llamacpp_tokens_second` Trend of `usage.total_tokens / request duration`
+- `llamacpp_prompt_tokens` Trend of `usage.prompt_tokens`
+- `llamacpp_prompt_tokens_total_counter` Counter of `usage.prompt_tokens`
+- `llamacpp_completion_tokens` Trend of `usage.completion_tokens`
+- `llamacpp_completion_tokens_total_counter` Counter of `usage.completion_tokens`
+- `llamacpp_completions_truncated_rate` Rate of completions truncated, i.e. if `finish_reason === 'length'`
+- `llamacpp_completions_stop_rate` Rate of completions stopped by the model, i.e. if `finish_reason === 'stop'`
+
+The script will fail if too many completions are truncated, see `llamacpp_completions_truncated_rate`.
+
+K6 metrics might be compared against [server metrics](../README.md), with:
+
+```shell
+curl http://localhost:8080/metrics
+```
+
+### Using the CI python script
+The `bench.py` script does several steps:
+- start the server
+- define good variable for k6
+- run k6 script
+- extract metrics from prometheus
+
+It aims to be used in the CI, but you can run it manually:
+
+```shell
+LLAMA_SERVER_BIN_PATH=../../../cmake-build-release/bin/llama-server python bench.py \
+              --runner-label local \
+              --name local \
+              --branch `git rev-parse --abbrev-ref HEAD` \
+              --commit `git rev-parse HEAD` \
+              --scenario script.js \
+              --duration 5m \
+              --hf-repo ggml-org/models	 \
+              --hf-file phi-2/ggml-model-q4_0.gguf \
+              --model-path-prefix models \
+              --parallel 4 \
+              -ngl 33 \
+              --batch-size 2048 \
+              --ubatch-size	256 \
+              --ctx-size 4096 \
+              --n-prompts 200 \
+              --max-prompt-tokens 256 \
+              --max-tokens 256
+```

tools/server/bench/bench.py

@@ -0,0 +1,322 @@
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import re
+import signal
+import socket
+import subprocess
+import sys
+import threading
+import time
+import traceback
+from contextlib import closing
+from datetime import datetime
+
+import matplotlib
+import matplotlib.dates
+import matplotlib.pyplot as plt
+import requests
+from statistics import mean
+
+
+def main(args_in: list[str] | None = None) -> None:
+    parser = argparse.ArgumentParser(description="Start server benchmark scenario")
+    parser.add_argument("--name", type=str, help="Bench name", required=True)
+    parser.add_argument("--runner-label", type=str, help="Runner label", required=True)
+    parser.add_argument("--branch", type=str, help="Branch name", default="detached")
+    parser.add_argument("--commit", type=str, help="Commit name", default="dirty")
+    parser.add_argument("--host", type=str, help="Server listen host", default="0.0.0.0")
+    parser.add_argument("--port", type=int, help="Server listen host", default="8080")
+    parser.add_argument("--model-path-prefix", type=str, help="Prefix where to store the model files", default="models")
+    parser.add_argument("--n-prompts", type=int,
+                        help="SERVER_BENCH_N_PROMPTS: total prompts to randomly select in the benchmark", required=True)
+    parser.add_argument("--max-prompt-tokens", type=int,
+                        help="SERVER_BENCH_MAX_PROMPT_TOKENS: maximum prompt tokens to filter out in the dataset",
+                        required=True)
+    parser.add_argument("--max-tokens", type=int,
+                        help="SERVER_BENCH_MAX_CONTEXT: maximum context size of the completions request to filter out in the dataset: prompt + predicted tokens",
+                        required=True)
+    parser.add_argument("--hf-repo", type=str, help="Hugging Face model repository", required=True)
+    parser.add_argument("--hf-file", type=str, help="Hugging Face model file", required=True)
+    parser.add_argument("-ngl", "--n-gpu-layers", type=int, help="layers to the GPU for computation", required=True)
+    parser.add_argument("--ctx-size", type=int, help="Set the size of the prompt context", required=True)
+    parser.add_argument("--parallel", type=int, help="Set the number of slots for process requests", required=True)
+    parser.add_argument("--batch-size", type=int, help="Set the batch size for prompt processing", required=True)
+    parser.add_argument("--ubatch-size", type=int, help="physical maximum batch size", required=True)
+    parser.add_argument("--scenario", type=str, help="Scenario to run", required=True)
+    parser.add_argument("--duration", type=str, help="Bench scenario", required=True)
+
+    args = parser.parse_args(args_in)
+
+    start_time = time.time()
+
+    # Start the server and performance scenario
+    try:
+        server_process = start_server(args)
+    except Exception:
+        print("bench: server start error :")
+        traceback.print_exc(file=sys.stdout)
+        sys.exit(1)
+
+    # start the benchmark
+    iterations = 0
+    data = {}
+    try:
+        start_benchmark(args)
+
+        with open("results.github.env", 'w') as github_env:
+            # parse output
+            with open('k6-results.json', 'r') as bench_results:
+                # Load JSON data from file
+                data = json.load(bench_results)
+                for metric_name in data['metrics']:
+                    for metric_metric in data['metrics'][metric_name]:
+                        value = data['metrics'][metric_name][metric_metric]
+                        if isinstance(value, float) or isinstance(value, int):
+                            value = round(value, 2)
+                            data['metrics'][metric_name][metric_metric]=value
+                            github_env.write(
+                                f"{escape_metric_name(metric_name)}_{escape_metric_name(metric_metric)}={value}\n")
+                iterations = data['root_group']['checks']['success completion']['passes']
+
+    except Exception:
+        print("bench: error :")
+        traceback.print_exc(file=sys.stdout)
+
+    # Stop the server
+    if server_process:
+        try:
+            print(f"bench: shutting down server pid={server_process.pid} ...")
+            if os.name == 'nt':
+                interrupt = signal.CTRL_C_EVENT
+            else:
+                interrupt = signal.SIGINT
+            server_process.send_signal(interrupt)
+            server_process.wait(0.5)
+
+        except subprocess.TimeoutExpired:
+            print(f"server still alive after 500ms, force-killing pid={server_process.pid} ...")
+            server_process.kill()  # SIGKILL
+            server_process.wait()
+
+        while is_server_listening(args.host, args.port):
+            time.sleep(0.1)
+
+    title = (f"llama.cpp {args.name} on {args.runner_label}\n "
+             f"duration={args.duration} {iterations} iterations")
+    xlabel = (f"{args.hf_repo}/{args.hf_file}\n"
+              f"parallel={args.parallel} ctx-size={args.ctx_size} ngl={args.n_gpu_layers} batch-size={args.batch_size} ubatch-size={args.ubatch_size} pp={args.max_prompt_tokens} pp+tg={args.max_tokens}\n"
+              f"branch={args.branch} commit={args.commit}")
+
+    # Prometheus
+    end_time = time.time()
+    prometheus_metrics = {}
+    if is_server_listening("0.0.0.0", 9090):
+        metrics = ['prompt_tokens_seconds', 'predicted_tokens_seconds',
+                   'kv_cache_usage_ratio', 'requests_processing', 'requests_deferred']
+
+        for metric in metrics:
+            resp = requests.get(f"http://localhost:9090/api/v1/query_range",
+                                params={'query': 'llamacpp:' + metric, 'start': start_time, 'end': end_time, 'step': 2})
+
+            with open(f"{metric}.json", 'w') as metric_json:
+                metric_json.write(resp.text)
+
+            if resp.status_code != 200:
+                print(f"bench: unable to extract prometheus metric {metric}: {resp.text}")
+            else:
+                metric_data = resp.json()
+                values = metric_data['data']['result'][0]['values']
+                timestamps, metric_values = zip(*values)
+                metric_values = [float(value) for value in metric_values]
+                prometheus_metrics[metric] = metric_values
+                timestamps_dt = [str(datetime.fromtimestamp(int(ts))) for ts in timestamps]
+                plt.figure(figsize=(16, 10), dpi=80)
+                plt.plot(timestamps_dt, metric_values, label=metric)
+                plt.xticks(rotation=0, fontsize=14, horizontalalignment='center', alpha=.7)
+                plt.yticks(fontsize=12, alpha=.7)
+
+                ylabel = f"llamacpp:{metric}"
+                plt.title(title,
+                          fontsize=14, wrap=True)
+                plt.grid(axis='both', alpha=.3)
+                plt.ylabel(ylabel, fontsize=22)
+                plt.xlabel(xlabel, fontsize=14, wrap=True)
+                plt.gca().xaxis.set_major_locator(matplotlib.dates.MinuteLocator())
+                plt.gca().xaxis.set_major_formatter(matplotlib.dates.DateFormatter("%Y-%m-%d %H:%M:%S"))
+                plt.gcf().autofmt_xdate()
+
+                # Remove borders
+                plt.gca().spines["top"].set_alpha(0.0)
+                plt.gca().spines["bottom"].set_alpha(0.3)
+                plt.gca().spines["right"].set_alpha(0.0)
+                plt.gca().spines["left"].set_alpha(0.3)
+
+                # Save the plot as a jpg image
+                plt.savefig(f'{metric}.jpg', dpi=60)
+                plt.close()
+
+                # Mermaid format in case images upload failed
+                with open(f"{metric}.mermaid", 'w') as mermaid_f:
+                    mermaid = (
+                    f"""---
+config:
+    xyChart:
+        titleFontSize: 12
+        width: 900
+        height: 600
+    themeVariables:
+        xyChart:
+            titleColor: "#000000"
+---
+xychart-beta
+    title "{title}"
+    y-axis "llamacpp:{metric}"
+    x-axis "llamacpp:{metric}" {int(min(timestamps))} --> {int(max(timestamps))}
+    line [{', '.join([str(round(float(value), 2)) for value in metric_values])}]
+                    """)
+                    mermaid_f.write(mermaid)
+
+    # 140 chars max for commit status description
+    bench_results = {
+        "i": iterations,
+        "req": {
+            "p95": round(data['metrics']["http_req_duration"]["p(95)"], 2),
+            "avg": round(data['metrics']["http_req_duration"]["avg"], 2),
+        },
+        "pp": {
+            "p95": round(data['metrics']["llamacpp_prompt_processing_second"]["p(95)"], 2),
+            "avg": round(data['metrics']["llamacpp_prompt_processing_second"]["avg"], 2),
+            "0": round(mean(prometheus_metrics['prompt_tokens_seconds']), 2) if 'prompt_tokens_seconds' in prometheus_metrics else 0,
+        },
+        "tg": {
+            "p95": round(data['metrics']["llamacpp_tokens_second"]["p(95)"], 2),
+            "avg": round(data['metrics']["llamacpp_tokens_second"]["avg"], 2),
+            "0": round(mean(prometheus_metrics['predicted_tokens_seconds']), 2) if 'predicted_tokens_seconds' in prometheus_metrics else 0,
+        },
+    }
+    with open("results.github.env", 'a') as github_env:
+        github_env.write(f"BENCH_RESULTS={json.dumps(bench_results, indent=None, separators=(',', ':') )}\n")
+        github_env.write(f"BENCH_ITERATIONS={iterations}\n")
+
+        title = title.replace('\n', ' ')
+        xlabel = xlabel.replace('\n', ' ')
+        github_env.write(f"BENCH_GRAPH_TITLE={title}\n")
+        github_env.write(f"BENCH_GRAPH_XLABEL={xlabel}\n")
+
+
+def start_benchmark(args):
+    k6_path = './k6'
+    if 'BENCH_K6_BIN_PATH' in os.environ:
+        k6_path = os.environ['BENCH_K6_BIN_PATH']
+    k6_args = [
+        'run', args.scenario,
+        '--no-color',
+        '--no-connection-reuse',
+        '--no-vu-connection-reuse',
+    ]
+    k6_args.extend(['--duration', args.duration])
+    k6_args.extend(['--iterations', args.n_prompts])
+    k6_args.extend(['--vus', args.parallel])
+    k6_args.extend(['--summary-export', 'k6-results.json'])
+    k6_args.extend(['--out', 'csv=k6-results.csv'])
+    args = f"SERVER_BENCH_N_PROMPTS={args.n_prompts} SERVER_BENCH_MAX_PROMPT_TOKENS={args.max_prompt_tokens} SERVER_BENCH_MAX_CONTEXT={args.max_tokens} "
+    args = args + ' '.join([str(arg) for arg in [k6_path, *k6_args]])
+    print(f"bench: starting k6 with: {args}")
+    k6_completed = subprocess.run(args, shell=True, stdout=sys.stdout, stderr=sys.stderr)
+    if k6_completed.returncode != 0:
+        raise Exception("bench: unable to run k6")
+
+
+def start_server(args):
+    server_process = start_server_background(args)
+
+    attempts = 0
+    max_attempts = 600
+    if 'GITHUB_ACTIONS' in os.environ:
+        max_attempts *= 2
+
+    while not is_server_listening(args.host, args.port):
+        attempts += 1
+        if attempts > max_attempts:
+            assert False, "server not started"
+        print(f"bench:     waiting for server to start ...")
+        time.sleep(0.5)
+
+    attempts = 0
+    while not is_server_ready(args.host, args.port):
+        attempts += 1
+        if attempts > max_attempts:
+            assert False, "server not ready"
+        print(f"bench:     waiting for server to be ready ...")
+        time.sleep(0.5)
+
+    print("bench: server started and ready.")
+    return server_process
+
+
+def start_server_background(args):
+    # Start the server
+    server_path = '../../../build/bin/llama-server'
+    if 'LLAMA_SERVER_BIN_PATH' in os.environ:
+        server_path = os.environ['LLAMA_SERVER_BIN_PATH']
+    server_args = [
+        '--host', args.host,
+        '--port', args.port,
+    ]
+    server_args.extend(['--hf-repo', args.hf_repo])
+    server_args.extend(['--hf-file', args.hf_file])
+    server_args.extend(['--n-gpu-layers', args.n_gpu_layers])
+    server_args.extend(['--ctx-size', args.ctx_size])
+    server_args.extend(['--parallel', args.parallel])
+    server_args.extend(['--batch-size', args.batch_size])
+    server_args.extend(['--ubatch-size', args.ubatch_size])
+    server_args.extend(['--n-predict', args.max_tokens * 2])
+    server_args.append('--cont-batching')
+    server_args.append('--metrics')
+    server_args.append('--flash-attn')
+    args = [str(arg) for arg in [server_path, *server_args]]
+    print(f"bench: starting server with: {' '.join(args)}")
+    pkwargs = {
+        'stdout': subprocess.PIPE,
+        'stderr': subprocess.PIPE
+    }
+    server_process = subprocess.Popen(
+        args,
+        **pkwargs)  # pyright: ignore[reportArgumentType, reportCallIssue]
+
+    def server_log(in_stream, out_stream):
+        for line in iter(in_stream.readline, b''):
+            print(line.decode('utf-8'), end='', file=out_stream)
+
+    thread_stdout = threading.Thread(target=server_log, args=(server_process.stdout, sys.stdout))
+    thread_stdout.start()
+    thread_stderr = threading.Thread(target=server_log, args=(server_process.stderr, sys.stderr))
+    thread_stderr.start()
+
+    return server_process
+
+
+def is_server_listening(server_fqdn, server_port):
+    with closing(socket.socket(socket.AF_INET, socket.SOCK_STREAM)) as sock:
+        result = sock.connect_ex((server_fqdn, server_port))
+        _is_server_listening = result == 0
+        if _is_server_listening:
+            print(f"server is listening on {server_fqdn}:{server_port}...")
+        return _is_server_listening
+
+
+def is_server_ready(server_fqdn, server_port):
+    url = f"http://{server_fqdn}:{server_port}/health"
+    response = requests.get(url)
+    return response.status_code == 200
+
+
+def escape_metric_name(metric_name):
+    return re.sub('[^A-Z0-9]', '_', metric_name.upper())
+
+
+if __name__ == '__main__':
+    main()

tools/server/bench/prometheus.yml

@@ -0,0 +1,9 @@
+global:
+  scrape_interval:     10s
+  external_labels:
+    llamacpp: 'server'
+
+scrape_configs:
+  - job_name: 'llama.cpp server'
+    static_configs:
+      - targets: ['localhost:8080']

tools/server/bench/requirements.txt

@@ -0,0 +1,2 @@
+matplotlib
+requests

tools/server/bench/script.js

@@ -0,0 +1,162 @@
+import sse from 'k6/x/sse'
+import {check, sleep} from 'k6'
+import {SharedArray} from 'k6/data'
+import {Counter, Rate, Trend} from 'k6/metrics'
+import exec from 'k6/execution';
+
+// Server chat completions prefix
+const server_url = __ENV.SERVER_BENCH_URL ? __ENV.SERVER_BENCH_URL : 'http://localhost:8080/v1'
+
+// Number of total prompts in the dataset - default 10m / 10 seconds/request * number of users
+const n_prompt = __ENV.SERVER_BENCH_N_PROMPTS ? parseInt(__ENV.SERVER_BENCH_N_PROMPTS) : 600 / 10 * 8
+
+// Model name to request
+const model = __ENV.SERVER_BENCH_MODEL_ALIAS ? __ENV.SERVER_BENCH_MODEL_ALIAS : 'my-model'
+
+// Dataset path
+const dataset_path = __ENV.SERVER_BENCH_DATASET ? __ENV.SERVER_BENCH_DATASET : './ShareGPT_V3_unfiltered_cleaned_split.json'
+
+// Max tokens to predict
+const max_tokens = __ENV.SERVER_BENCH_MAX_TOKENS ? parseInt(__ENV.SERVER_BENCH_MAX_TOKENS) : 512
+
+// Max prompt tokens
+const n_prompt_tokens = __ENV.SERVER_BENCH_MAX_PROMPT_TOKENS ? parseInt(__ENV.SERVER_BENCH_MAX_PROMPT_TOKENS) : 1024
+
+// Max slot context
+const n_ctx_slot = __ENV.SERVER_BENCH_MAX_CONTEXT ? parseInt(__ENV.SERVER_BENCH_MAX_CONTEXT) : 2048
+
+export function setup() {
+    console.info(`Benchmark config: server_url=${server_url} n_prompt=${n_prompt} model=${model} dataset_path=${dataset_path} max_tokens=${max_tokens}`)
+}
+
+const data = new SharedArray('conversations', function () {
+    const tokenizer = (message) => message.split(/[\s,'".?]/)
+
+    return JSON.parse(open(dataset_path))
+        // Filter out the conversations with less than 2 turns.
+        .filter(data => data["conversations"].length >= 2)
+        .filter(data => data["conversations"][0]["from"] === "human")
+        .map(data => {
+            return {
+                prompt: data["conversations"][0]["value"],
+                n_prompt_tokens: tokenizer(data["conversations"][0]["value"]).length,
+                n_completion_tokens: tokenizer(data["conversations"][1]["value"]).length,
+            }
+        })
+        // Filter out too short sequences
+        .filter(conv => conv.n_prompt_tokens >= 4 && conv.n_completion_tokens >= 4)
+        // Filter out too long sequences.
+        .filter(conv => conv.n_prompt_tokens <= n_prompt_tokens && conv.n_prompt_tokens + conv.n_completion_tokens <= n_ctx_slot)
+        // Keep only first n prompts
+        .slice(0, n_prompt)
+})
+
+const llamacpp_prompt_tokens = new Trend('llamacpp_prompt_tokens')
+const llamacpp_completion_tokens = new Trend('llamacpp_completion_tokens')
+
+const llamacpp_tokens_second = new Trend('llamacpp_tokens_second')
+const llamacpp_prompt_processing_second = new Trend('llamacpp_prompt_processing_second')
+const llamacpp_emit_first_token_second = new Trend('llamacpp_emit_first_token_second')
+
+const llamacpp_prompt_tokens_total_counter = new Counter('llamacpp_prompt_tokens_total_counter')
+const llamacpp_completion_tokens_total_counter = new Counter('llamacpp_completion_tokens_total_counter')
+
+const llamacpp_completions_truncated_rate = new Rate('llamacpp_completions_truncated_rate')
+const llamacpp_completions_stop_rate = new Rate('llamacpp_completions_stop_rate')
+
+export const options = {
+    thresholds: {
+        llamacpp_completions_truncated_rate: [
+            // more than 80% of truncated input will abort the test
+            {threshold: 'rate < 0.8', abortOnFail: true, delayAbortEval: '1m'},
+        ],
+    },
+    duration: '10m',
+    vus: 8,
+}
+
+export default function () {
+    const conversation = data[exec.scenario.iterationInInstance % data.length]
+    const payload = {
+        "messages": [
+            {
+                "role": "system",
+                "content": "You are ChatGPT, an AI assistant.",
+            },
+            {
+                "role": "user",
+                "content": conversation.prompt,
+            }
+        ],
+        "model": model,
+        "stream": true,
+        "stream_options": {
+          "include_usage": true, // False to be supported in llama.cpp server
+        },
+        "seed": 42,
+        "max_tokens": max_tokens,
+        "stop": ["<|im_end|>"] // This is temporary for phi-2 base (i.e. not instructed) since the server expects that the model always to emit BOS
+    }
+
+    const params = {method: 'POST', body: JSON.stringify(payload)};
+
+    const startTime = new Date()
+    let promptEvalEndTime = null
+    let prompt_tokens = 0
+    let completions_tokens = 0
+    let finish_reason = null
+    const res = sse.open(`${server_url}/chat/completions`, params, function (client) {
+        client.on('event', function (event) {
+            if (promptEvalEndTime == null) {
+                promptEvalEndTime = new Date()
+                llamacpp_emit_first_token_second.add((promptEvalEndTime - startTime) / 1.e3)
+            }
+
+            if (event.data === '[DONE]' || event.data === '') {
+                return
+            }
+
+            let chunk = JSON.parse(event.data)
+
+            if (chunk.choices && chunk.choices.length > 0) {
+                let choice = chunk.choices[0]
+                if (choice.finish_reason) {
+                    finish_reason = choice.finish_reason
+                }
+            }
+
+            if (chunk.usage) {
+                prompt_tokens = chunk.usage.prompt_tokens
+                llamacpp_prompt_tokens.add(prompt_tokens)
+                llamacpp_prompt_tokens_total_counter.add(prompt_tokens)
+
+                completions_tokens = chunk.usage.completion_tokens
+                llamacpp_completion_tokens.add(completions_tokens)
+                llamacpp_completion_tokens_total_counter.add(completions_tokens)
+            }
+        })
+
+        client.on('error', function (e) {
+            console.log('An unexpected error occurred: ', e.error());
+            throw e;
+        })
+    })
+
+    check(res, {'success completion': (r) => r.status === 200})
+
+    const endTime = new Date()
+
+    const promptEvalTime = promptEvalEndTime - startTime
+    if (promptEvalTime > 0) {
+        llamacpp_prompt_processing_second.add(prompt_tokens / (promptEvalEndTime - startTime) * 1.e3)
+    }
+
+    const completion_time = endTime - promptEvalEndTime
+    if (completions_tokens > 0 && completion_time > 0) {
+        llamacpp_tokens_second.add(completions_tokens / completion_time * 1.e3)
+    }
+    llamacpp_completions_truncated_rate.add(finish_reason === 'length')
+    llamacpp_completions_stop_rate.add(finish_reason === 'stop')
+
+    sleep(0.3)
+}

tools/server/chat-llama2.sh

@@ -0,0 +1,109 @@
+#!/usr/bin/env bash
+
+API_URL="${API_URL:-http://127.0.0.1:8080}"
+
+CHAT=(
+    "Hello, Assistant."
+    "Hello. How may I help you today?"
+)
+
+INSTRUCTION="A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions."
+
+trim() {
+    shopt -s extglob
+    set -- "${1##+([[:space:]])}"
+    printf "%s" "${1%%+([[:space:]])}"
+}
+
+trim_trailing() {
+    shopt -s extglob
+    printf "%s" "${1%%+([[:space:]])}"
+}
+
+format_prompt() {
+    if [[ "${#CHAT[@]}" -eq 0 ]]; then
+        echo -n "[INST] <<SYS>>\n${INSTRUCTION}\n<</SYS>>"
+    else
+        LAST_INDEX=$(( ${#CHAT[@]} - 1 ))
+        echo -n "${CHAT[$LAST_INDEX]}\n[INST] $1 [/INST]"
+    fi
+}
+
+tokenize() {
+    curl \
+        --silent \
+        --request POST \
+        --url "${API_URL}/tokenize" \
+        --header "Content-Type: application/json" \
+        --data-raw "$(jq -ns --arg content "$1" '{content:$content}')" \
+    | jq '.tokens[]'
+}
+
+N_KEEP=$(tokenize "[INST] <<SYS>>\n${INSTRUCTION}\n<</SYS>>" | wc -l)
+
+chat_completion() {
+    PROMPT="$(trim_trailing "$(format_prompt "$1")")"
+    DATA="$(echo -n "$PROMPT" | jq -Rs --argjson n_keep $N_KEEP '{
+        prompt: .,
+        temperature: 0.2,
+        top_k: 40,
+        top_p: 0.9,
+        n_keep: $n_keep,
+        n_predict: 1024,
+        stop: ["[INST]"],
+        stream: true
+    }')"
+
+    # Create a temporary file to hold the Python output
+    TEMPFILE=$(mktemp)
+
+    exec 3< <(curl \
+        --silent \
+        --no-buffer \
+        --request POST \
+        --url "${API_URL}/completion" \
+        --header "Content-Type: application/json" \
+        --data-raw "${DATA}")
+
+    python -c "
+import json
+import sys
+
+answer = ''
+while True:
+    line = sys.stdin.readline()
+    if not line:
+        break
+    if line.startswith('data: '):
+        json_content = line[6:].strip()
+        content = json.loads(json_content)['content']
+        sys.stdout.write(content)
+        sys.stdout.flush()
+        answer += content
+
+answer = answer.rstrip('\n')
+
+# Write the answer to the temporary file
+with open('$TEMPFILE', 'w') as f:
+    f.write(answer)
+    " <&3
+
+    exec 3<&-
+
+    # Read the answer from the temporary file
+    ANSWER=$(cat $TEMPFILE)
+
+    # Clean up the temporary file
+    rm $TEMPFILE
+
+    printf "\n"
+
+    CHAT+=("$1" "$(trim "$ANSWER")")
+}
+
+while true; do
+    echo -en "\033[0;32m"  # Green color
+    read -r -e -p "> " QUESTION
+    echo -en "\033[0m"  # Reset color
+    chat_completion "${QUESTION}"
+done

tools/server/chat.mjs

@@ -0,0 +1,131 @@
+import * as readline from 'node:readline'
+import { stdin, stdout } from 'node:process'
+import { readFileSync } from 'node:fs'
+import { SchemaConverter }  from './public_legacy/json-schema-to-grammar.mjs'
+
+const args = process.argv.slice(2);
+const grammarJsonSchemaFile = args.find(
+    (_, index) => args[index - 1] === "--grammar-json-schema"
+);
+
+const no_cached_prompt = args.find(
+    (_, index) => args[index - 1] === "--no-cache-prompt"
+) ?? "false";
+
+const grammarFile = args.find((_, index) => args[index - 1] === "--grammar");
+
+// Example usage: function,arguments
+const grammarJsonSchemaPropOrder = args.find(
+    (_, index) => args[index - 1] === "--grammar-json-schema-prop-order"
+);
+const propOrder = grammarJsonSchemaPropOrder
+    ? grammarJsonSchemaPropOrder
+          .split(",")
+          .reduce((acc, cur, index) => ({ ...acc, [cur]: index }), {})
+    : {};
+
+let grammar = null
+if (grammarJsonSchemaFile) {
+    let schema = JSON.parse(readFileSync(grammarJsonSchemaFile, 'utf-8'))
+    const converter = new SchemaConverter({prop_order: propOrder, allow_fetch: true})
+    schema = await converter.resolveRefs(schema, grammarJsonSchemaFile)
+    converter.visit(schema, '')
+    grammar = converter.formatGrammar()
+}
+if (grammarFile) {
+    grammar = readFileSync(grammarFile, 'utf-8')
+}
+
+// for cached prompt
+let slot_id = -1;
+
+const API_URL = 'http://127.0.0.1:8080'
+
+const chat = [
+    {
+        human: "Hello, Assistant.",
+        assistant: "Hello. How may I help you today?"
+    },
+    {
+        human: "Please tell me the largest city in Europe.",
+        assistant: "Sure. The largest city in Europe is Moscow, the capital of Russia."
+    },
+]
+
+const instruction = `A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.`
+
+function format_prompt(question) {
+    return `${instruction}\n${
+        chat.map(m =>`### Human: ${m.human}\n### Assistant: ${m.assistant}`).join("\n")
+    }\n### Human: ${question}\n### Assistant:`
+}
+
+async function tokenize(content) {
+    const result = await fetch(`${API_URL}/tokenize`, {
+        method: 'POST',
+        body: JSON.stringify({ content })
+    })
+
+    if (!result.ok) {
+        return []
+    }
+
+    return await result.json().tokens
+}
+
+const n_keep = await tokenize(instruction).length
+
+async function chat_completion(question) {
+    const result = await fetch(`${API_URL}/completion`, {
+        method: 'POST',
+        body: JSON.stringify({
+            prompt: format_prompt(question),
+            temperature: 0.2,
+            top_k: 40,
+            top_p: 0.9,
+            n_keep: n_keep,
+            n_predict: 256,
+            cache_prompt: no_cached_prompt === "false",
+            slot_id: slot_id,
+            stop: ["\n### Human:"], // stop completion after generating this
+            grammar,
+            stream: true,
+        })
+    })
+
+    if (!result.ok) {
+        return
+    }
+
+    let answer = ''
+
+    for await (var chunk of result.body) {
+        const t = Buffer.from(chunk).toString('utf8')
+        if (t.startsWith('data: ')) {
+            const message = JSON.parse(t.substring(6))
+            slot_id = message.slot_id
+            answer += message.content
+            process.stdout.write(message.content)
+            if (message.stop) {
+                if (message.truncated) {
+                    chat.shift()
+                }
+                break
+            }
+        }
+    }
+
+    process.stdout.write('\n')
+    chat.push({ human: question, assistant: answer.trimStart() })
+}
+
+const rl = readline.createInterface({ input: stdin, output: stdout });
+
+const readlineQuestion = (rl, query, options) => new Promise((resolve, reject) => {
+    rl.question(query, options, resolve)
+});
+
+while(true) {
+    const question = await readlineQuestion(rl, '> ')
+    await chat_completion(question)
+}

tools/server/chat.sh

@@ -0,0 +1,80 @@
+#!/usr/bin/env bash
+
+API_URL="${API_URL:-http://127.0.0.1:8080}"
+
+CHAT=(
+    "Hello, Assistant."
+    "Hello. How may I help you today?"
+    "Please tell me the largest city in Europe."
+    "Sure. The largest city in Europe is Moscow, the capital of Russia."
+)
+
+INSTRUCTION="A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions."
+
+trim() {
+    shopt -s extglob
+    set -- "${1##+([[:space:]])}"
+    printf "%s" "${1%%+([[:space:]])}"
+}
+
+trim_trailing() {
+    shopt -s extglob
+    printf "%s" "${1%%+([[:space:]])}"
+}
+
+format_prompt() {
+    echo -n "${INSTRUCTION}"
+    printf "\n### Human: %s\n### Assistant: %s" "${CHAT[@]}" "$1"
+}
+
+tokenize() {
+    curl \
+        --silent \
+        --request POST \
+        --url "${API_URL}/tokenize" \
+        --header "Content-Type: application/json" \
+        --data-raw "$(jq -ns --arg content "$1" '{content:$content}')" \
+    | jq '.tokens[]'
+}
+
+N_KEEP=$(tokenize "${INSTRUCTION}" | wc -l)
+
+chat_completion() {
+    PROMPT="$(trim_trailing "$(format_prompt "$1")")"
+    DATA="$(echo -n "$PROMPT" | jq -Rs --argjson n_keep $N_KEEP '{
+        prompt: .,
+        temperature: 0.2,
+        top_k: 40,
+        top_p: 0.9,
+        n_keep: $n_keep,
+        n_predict: 256,
+        cache_prompt: true,
+        stop: ["\n### Human:"],
+        stream: true
+    }')"
+
+    ANSWER=''
+
+    while IFS= read -r LINE; do
+        if [[ $LINE = data:* ]]; then
+            CONTENT="$(echo "${LINE:5}" | jq -r '.content')"
+            printf "%s" "${CONTENT}"
+            ANSWER+="${CONTENT}"
+        fi
+    done < <(curl \
+        --silent \
+        --no-buffer \
+        --request POST \
+        --url "${API_URL}/completion" \
+        --header "Content-Type: application/json" \
+        --data-raw "${DATA}")
+
+    printf "\n"
+
+    CHAT+=("$1" "$(trim "$ANSWER")")
+}
+
+while true; do
+    read -r -e -p "> " QUESTION
+    chat_completion "${QUESTION}"
+done

tools/server/public/loading.html

@@ -0,0 +1,12 @@
+<!DOCTYPE html>
+<html>
+    <head>
+        <meta http-equiv="refresh" content="5">
+    </head>
+    <body>
+        <div id="loading">
+            The model is loading. Please wait.<br/>
+            The user interface will appear soon.
+        </div>
+    </body>
+</html>

tools/server/public_legacy/colorthemes.css

@@ -0,0 +1,402 @@
+@import url("theme-snowstorm.css");
+@import url("theme-polarnight.css");
+@import url("theme-ketivah.css");
+@import url("theme-mangotango.css");
+@import url("theme-playground.css");
+@import url("theme-beeninorder.css");
+
+:root {
+/* ---------- PRIMARY COLORS ----------------- */
+--primary-color-1: hsl(217.5, 26.7%, 94.1%);
+    --primary-color-1-hue:             217.5;
+    --primary-color-1-saturation:      26.7%;
+    --primary-color-1-lightness:       94.1%;
+
+--primary-color-2: hsl(218.2, 26.8%, 92.0%);
+    --primary-color-2-hue:             218.2;
+    --primary-color-2-saturation:      26.8%;
+    --primary-color-2-lightness:       92.0%;
+
+--primary-color-3: hsl(218.8, 27.9%, 88.0%);
+    --primary-color-3-hue:             218.8;
+    --primary-color-3-saturation:      27.9%;
+    --primary-color-3-lightness:       88.0%;
+
+--primary-color-4: hsl(218.8, 18.3%, 81.8%);
+    --primary-color-4-hue:             218.8;
+    --primary-color-4-saturation:      18.3%;
+    --primary-color-4-lightness:       81.8%;
+
+
+/* ---------- SECONDARY COLORS --------------- */
+--secondary-color-1: hsl(220.0, 16.4%, 21.6%);
+    --secondary-color-1-hue:             220.0;
+    --secondary-color-1-saturation:      16.4%;
+    --secondary-color-1-lightness:       21.6%;
+
+--secondary-color-2: hsl(221.7, 16.3%, 27.6%);
+    --secondary-color-2-hue:             221.7;
+    --secondary-color-2-saturation:      16.3%;
+    --secondary-color-2-lightness:       27.6%;
+
+--secondary-color-3: hsl(220.0, 16.8%, 31.6%);
+    --secondary-color-3-hue:             220.0;
+    --secondary-color-3-saturation:      16.8%;
+    --secondary-color-3-lightness:       31.6%;
+
+--secondary-color-4: hsl(220.0, 16.5%, 35.7%);
+    --secondary-color-4-hue:             220.0;
+    --secondary-color-4-saturation:      16.5%;
+    --secondary-color-4-lightness:       35.7%;
+
+
+
+/* ----------- NUANCES COLORS ---------------- */
+--theme-nuance-color-1: hsl(178.7, 25.1%, 64.9%);
+    --theme-nuance-color-1-hue:             178.7;
+    --theme-nuance-color-1-saturation:      25.1%;
+    --theme-nuance-color-1-lightness:       64.9%;
+
+--theme-nuance-color-2: hsl(193.3, 43.4%, 67.5%);
+    --theme-nuance-color-2-hue:             193.3;
+    --theme-nuance-color-2-saturation:      43.4%;
+    --theme-nuance-color-2-lightness:       67.5%;
+
+--theme-nuance-color-3: hsl(210.0, 34.0%, 63.1%);
+    --theme-nuance-color-3-hue:             210.0;
+    --theme-nuance-color-3-saturation:      34.0%;
+    --theme-nuance-color-3-lightness:       63.1%;
+
+--theme-nuance-color-4: hsl(213.1, 32.0%, 52.2%);
+    --theme-nuance-color-4-hue:             213.1;
+    --theme-nuance-color-4-saturation:      32.0%;
+    --theme-nuance-color-4-lightness:       52.2%;
+
+
+
+/* ----------- ROYGP COLORS ------------------ */
+--theme-red-color:    hsl(32.5, 80%, 50%);
+--theme-orange-color: hsl(32.5, 70%, 45%);
+--theme-yellow-color: hsl(40.0,   0.6%, 73.3%);
+--theme-green-color:  hsl(92.4,  27.8%, 64.7%);
+--theme-purple-color: hsl(311.1, 20.2%, 63.1%);
+
+
+
+/* ------------------------------------------- */
+--background-color-1:    var(--primary-color-1);
+--background-color-2:    var(--primary-color-2);
+--background-color-3:    var(--primary-color-3);
+--background-color-4:    var(--primary-color-4);
+
+--border-color-1:        var(--primary-color-2);
+--border-color-2:        var(--primary-color-3);
+--border-color-3:        var(--primary-color-4);
+
+--border-focus-color:    var(--theme-nuance-color-2);
+--border-focus-shadow:   var(--theme-nuance-color-1);
+
+--text-color-plain:      var(--secondary-color-1);
+--text-color-subtile-1:  var(--secondary-color-2);
+--text-color-subtile-2:  var(--secondary-color-3);
+
+--code-background-color: var(--secondary-color-2);
+--code-text-color:       var(--primary-color-2);
+
+--ui-range-thumb-color:  var(--theme-nuance-color-3);
+--ui-range-thumb-border: var(--ui-ranger-thumb-color);
+
+--textarea-border-color: var(--secondary-color-4);
+
+--chat-id-color:         var(--theme-nuance-color-4);
+
+
+
+/* ------------------------------------------- */
+--button-alert-text-hover:       var(--primary-color-1);
+--button-alert-color-hover:      var(--theme-orange-color);
+--button-alert-border-hover:     var(--theme-orange-color);
+
+--button-alert-text-active:      var(--primary-color-1);
+--button-alert-color-active:     var(--theme-red-color);
+--button-alert-border-active:    var(--theme-red-color);
+
+
+
+/* ----------- PRIMARY BUTTONS --------------- */
+/* - button should immediately catch the eye - */
+--button-primary-text:   var(--secondary-color-1);
+--button-primary-color:  var(--theme-nuance-color-3);
+--button-primary-border: var(--theme-nuance-color-3);
+
+
+/* ---------hover---------- */
+--button-primary-text-hover:
+    hsl(217.5,
+    calc(var(--secondary-color-1-saturation) + 35%),
+    calc(var(--secondary-color-1-lightness)  - 30%));
+
+--button-primary-color-hover:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) -  2%),
+    calc(var(--theme-nuance-color-3-lightness)  - 10%));
+
+--button-primary-border-hover:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) -  2%),
+    calc(var(--theme-nuance-color-3-lightness)  - 10%));
+
+
+/* ---------active--------- */
+--button-primary-text-active:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 20%),
+    calc(var(--theme-nuance-color-3-lightness)  + 35%));
+
+--button-primary-color-active:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 10%),
+    calc(var(--theme-nuance-color-3-lightness)  - 25%));
+
+--button-primary-border-active:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 10%),
+    calc(var(--theme-nuance-color-3-lightness)  - 25%));
+
+
+
+/* ---------- SECONDARY BUTTONS -------------- */
+/* these should NOT immediately catch the eye  */
+--button-secondary-text:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 20%),
+    calc(var(--theme-nuance-color-3-lightness)  - 50%));
+
+--button-secondary-color:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 20%),
+    calc(var(--theme-nuance-color-3-lightness)  + 10%));
+
+--button-secondary-border:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 20%),
+    calc(var(--theme-nuance-color-3-lightness)  + 10%));
+
+
+/* ---------hover---------- */
+--button-secondary-text-hover:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 20%),
+    calc(var(--theme-nuance-color-3-lightness)  - 80%));
+
+--button-secondary-color-hover:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 22%),
+    calc(var(--theme-nuance-color-3-lightness)  +  1%));
+
+--button-secondary-border-hover:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 22%),
+    calc(var(--theme-nuance-color-3-lightness)  +  1%));
+
+
+/* ---------active--------- */
+--button-secondary-text-active:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) + 40%),
+    calc(var(--theme-nuance-color-3-lightness)  - 55%));
+
+--button-secondary-color-active:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 30%),
+    calc(var(--theme-nuance-color-3-lightness)  -  5%));
+
+--button-secondary-border-active:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 30%),
+    calc(var(--theme-nuance-color-3-lightness)  -  5%));
+
+
+
+/* ---------- TERTIARY BUTTONS --------------- */
+/* ---------- disabled buttons --------------- */
+--button-tertiary-text:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 40%),
+    calc(var(--theme-nuance-color-3-lightness)  -  5%));
+
+--button-tertiary-color:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 40%),
+    calc(var(--theme-nuance-color-3-lightness)  + 20%));
+
+--button-tertiary-border:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 40%),
+    calc(var(--theme-nuance-color-3-lightness)  + 20%));
+
+/* ---------hover---------- */
+--button-tertiary-text-hover:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 40%),
+    calc(var(--theme-nuance-color-3-lightness)  -  5%));
+
+--button-tertiary-color-hover:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 40%),
+    calc(var(--theme-nuance-color-3-lightness)  + 20%));
+
+--button-tertiary-border-hover:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 40%),
+    calc(var(--theme-nuance-color-3-lightness)  + 20%));
+}
+
+/*
+
+.theme-template {
+
+
+    If light theme: should go from bright to darker
+    If dark theme: should go from dark to brighter
+    ideally this should not be anything but steps of
+    gray or slightly variants from it
+
+    --primary-color-1: #2E3440;
+    --primary-color-2: #3B4252;
+    --primary-color-3: #434C5E;
+    --primary-color-4: #4C566A;
+
+
+
+    If light theme: should go from dark to brighter
+    If dark theme: should go from bright to darker
+    ideally this should not be anything but steps of
+    gray or slightly variants from it
+
+    --secondary-color-1: #ECEFF4;
+    --secondary-color-2: #E5E9F0;
+    --secondary-color-3: #D8DEE9;
+    --secondary-color-4: #C8CED9;
+
+
+
+    Choose wisely nuance colors. It is not easy to find
+    4 harmonizing nuance colors. But keep in mind, that
+    only one accent color could work too.
+
+    --theme-nuance-color-1: #8FBCBB;
+    --theme-nuance-color-2: #88C0D0;
+    --theme-nuance-color-3: #81A1C1;
+    --theme-nuance-color-4: #5E81AC;
+
+
+
+    adapt the color red, orange, yellow, green,
+    purple to the 'mood' of your overall design
+    e.g is it low-contrast? vibrant? dynamic? etc
+
+    --theme-red-color:    #BF616A;
+    --theme-orange-color: #D08770;
+    --theme-yellow-color: #EBCB8B;
+    --theme-green-color:  #A3BE8C;
+    --theme-purple-color: #B48EAD;
+
+
+
+NOTE: comment all those line `--- ...` out
+------------------------------------------------
+--background-color-1:
+--background-color-2:
+--background-color-3:
+--background-color-4:
+
+--border-color-1:
+--border-color-2:
+--border-color-3:
+
+--border-focus-color:
+--border-focus-shadow:
+
+--text-color-plain:
+--text-color-subtile-1:
+--text-color-subtile-2:
+
+--code-background-color:
+--code-text-color:
+
+--ui-range-thumb-color:
+--ui-range-thumb-border:
+
+--textarea-border-color:
+
+
+
+-------------------------------------------
+--button-alert-text-hover:
+--button-alert-color-hover:
+--button-alert-border-hover:
+
+--button-alert-text-active:
+--button-alert-color-active:
+--button-alert-border-active:
+
+
+
+----------- PRIMARY -----------------------
+--button should immediately catch the eye--
+
+--button-primary-text:
+--button-primary-color:
+--button-primary-border:
+
+
+---------hover----------
+--button-primary-text-hover:
+--button-primary-color-hover:
+--button-primary-border-hover:
+
+
+---------active---------
+--button-primary-text-active:
+--button-primary-color-active:
+--button-primary-border-active:
+
+
+
+------------ SECONDARY ------------------------
+--button should NOT immediately catch the eye--
+
+--button-secondary-text:
+--button-secondary-color:
+--button-secondary-border:
+
+
+---------hover----------
+--button-secondary-text-hover:
+--button-secondary-color-hover:
+--button-secondary-border-hover:
+
+
+---------active---------
+--button-secondary-text-active:
+--button-secondary-color-active:
+--button-secondary-border-active:
+
+
+
+---------- TERTIARY -----------------------
+---------- disabled buttons ---------------
+--button-tertiary-text:
+--button-tertiary-color:
+--button-tertiary-border:
+
+
+---------hover----------
+--button-tertiary-text:
+--button-tertiary-color:
+--button-tertiary-border:
+
+}
+
+*/