EngineX-Mthreads/enginex-mthreads-llama.cpp
4
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

Sync from upstream llama.cpp repository

Browse Source
This commit is contained in:
xiezhongtao
2026-01-16 10:43:34 +08:00
parent 3bc369a6f7
commit f4ae4cc7da
2053 changed files with 956010 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
examples/eval-callback/CMakeLists.txt Normal file
View File

@@ -0,0 +1,26 @@
set(TARGET llama-eval-callback)
add_executable(${TARGET} eval-callback.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_17)
if(LLAMA_BUILD_TESTS)
if(NOT ${CMAKE_SYSTEM_PROCESSOR} MATCHES "s390x")
set(MODEL_NAME "tinyllamas/stories15M-q4_0.gguf")
set(MODEL_HASH "SHA256=66967fbece6dbe97886593fdbb73589584927e29119ec31f08090732d1861739")
else()
set(MODEL_NAME "tinyllamas/stories15M-be.Q4_0.gguf")
set(MODEL_HASH "SHA256=9aec857937849d976f30397e97eb1cabb53eb9dcb1ce4611ba8247fb5f44c65d")
endif()
set(MODEL_DEST "${CMAKE_BINARY_DIR}/${MODEL_NAME}")
set(TEST_TARGET test-eval-callback)
add_test(NAME ${TEST_TARGET}-download-model COMMAND ${CMAKE_COMMAND}
-DDEST=${MODEL_DEST}
-DNAME=${MODEL_NAME}
-DHASH=${MODEL_HASH}
-P ${CMAKE_SOURCE_DIR}/cmake/download-models.cmake
)
set_tests_properties(${TEST_TARGET}-download-model PROPERTIES FIXTURES_SETUP ${TEST_TARGET}-download-model)
add_test(NAME ${TEST_TARGET} COMMAND llama-eval-callback -m "${MODEL_DEST}" --prompt hello --seed 42 -ngl 0)
set_tests_properties(${TEST_TARGET} PROPERTIES FIXTURES_REQUIRED ${TEST_TARGET}-download-model)
endif()

95
examples/eval-callback/README.md Normal file
View File

@@ -0,0 +1,95 @@
# llama.cpp/examples/eval-callback
A simple example which demonstrates how to use callback during the inference.
It simply prints to the console all operations and tensor data.
Usage:
```shell
llama-eval-callback \
--hf-repo ggml-org/models \
--hf-file phi-2/ggml-model-q4_0.gguf \
--model phi-2-q4_0.gguf \
--prompt hello \
--seed 42 \
-ngl 33
```
Will print:
```shell
llm_load_tensors: offloaded 33/33 layers to GPU
...
llama_new_context_with_model: n_ctx = 512
...
llama_new_context_with_model: CUDA0 compute buffer size = 105.00 MiB
llama_new_context_with_model: CUDA_Host compute buffer size = 6.01 MiB
llama_new_context_with_model: graph nodes = 1225
llama_new_context_with_model: graph splits = 2
ggml_debug: inp_embd = (f32) GET_ROWS(token_embd.weight{2560, 51200, 1, 1}, inp_tokens{1, 1, 1, 1}}) = {2560, 1, 1, 1}
[
[
[ -0.0181, 0.0272, 0.0272, ...],
],
]
ggml_debug: norm-0 = (f32) NORM(CUDA0#inp_embd#0{2560, 1, 1, 1}, }) = {2560, 1, 1, 1}
[
[
[ -0.6989, 1.0636, 1.0636, ...],
],
]
ggml_debug: norm_w-0 = (f32) MUL(norm-0{2560, 1, 1, 1}, blk.0.attn_norm.weight{2560, 1, 1, 1}}) = {2560, 1, 1, 1}
[
[
[ -0.1800, 0.2817, 0.2632, ...],
],
]
ggml_debug: attn_norm-0 = (f32) ADD(norm_w-0{2560, 1, 1, 1}, blk.0.attn_norm.bias{2560, 1, 1, 1}}) = {2560, 1, 1, 1}
[
[
[ -0.1863, 0.2970, 0.2604, ...],
],
]
ggml_debug: wqkv-0 = (f32) MUL_MAT(blk.0.attn_qkv.weight{2560, 7680, 1, 1}, attn_norm-0{2560, 1, 1, 1}}) = {7680, 1, 1, 1}
[
[
[ -1.1238, 1.2876, -1.8086, ...],
],
]
ggml_debug: bqkv-0 = (f32) ADD(wqkv-0{7680, 1, 1, 1}, blk.0.attn_qkv.bias{7680, 1, 1, 1}}) = {7680, 1, 1, 1}
[
[
[ -1.1135, 1.4604, -1.9226, ...],
],
]
ggml_debug: bqkv-0 (view) = (f32) VIEW(bqkv-0{7680, 1, 1, 1}, }) = {2560, 1, 1, 1}
[
[
[ -1.1135, 1.4604, -1.9226, ...],
],
]
ggml_debug: Qcur-0 = (f32) CONT(bqkv-0 (view){2560, 1, 1, 1}, }) = {2560, 1, 1, 1}
[
[
[ -1.1135, 1.4604, -1.9226, ...],
],
]
ggml_debug: Qcur-0 (reshaped) = (f32) RESHAPE(Qcur-0{2560, 1, 1, 1}, }) = {80, 32, 1, 1}
[
[
[ -1.1135, 1.4604, -1.9226, ...],
[ -0.3608, 0.5076, -1.8866, ...],
[ 1.7643, 0.0273, -2.1065, ...],
...
],
]
ggml_debug: Qcur-0 = (f32) ROPE(Qcur-0 (reshaped){80, 32, 1, 1}, CUDA0#inp_pos#0{1, 1, 1, 1}}) = {80, 32, 1, 1}
[
[
[ -1.1135, 1.4604, -1.9226, ...],
[ -0.3608, 0.5076, -1.8866, ...],
[ 1.7643, 0.0273, -2.1065, ...],
...
],
]
```

80
examples/eval-callback/eval-callback.cpp Normal file
View File

@@ -0,0 +1,80 @@
#include "arg.h"
#include "common.h"
#include "debug.h"
#include "log.h"
#include "llama.h"
#include "llama-cpp.h"
#include <string>
#include <vector>
static bool run(llama_context * ctx, const common_params & params) {
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);
std::vector<llama_token> tokens = common_tokenize(ctx, params.prompt, add_bos);
if (tokens.empty()) {
LOG_ERR("%s : there are not input tokens to process - (try to provide a prompt with '-p')\n", __func__);
return false;
}
if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size()))) {
LOG_ERR("%s : failed to eval\n", __func__);
return false;
}
return true;
}
int main(int argc, char ** argv) {
base_callback_data cb_data;
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);
// pass the callback to the backend scheduler
// it will be executed for each node during the graph computation
params.cb_eval = common_debug_cb_eval<false>;
params.cb_eval_user_data = &cb_data;
params.warmup = false;
// init
auto llama_init = common_init_from_params(params);
auto * model = llama_init->model();
auto * ctx = llama_init->context();
if (model == nullptr || ctx == nullptr) {
LOG_ERR("%s : failed to init\n", __func__);
return 1;
}
// print system information
{
LOG_INF("\n");
LOG_INF("%s\n", common_params_get_system_info(params).c_str());
LOG_INF("\n");
}
bool OK = run(ctx, params);
if (!OK) {
return 1;
}
LOG("\n");
llama_perf_context_print(ctx);
llama_backend_free();
return 0;
}