EngineX-Iluvatar/enginex-bi_150-llama.cpp
15
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

同步 b7516

Browse Source
This commit is contained in:
xiezhongtao
2026-01-23 11:34:20 +08:00
parent d3752412b3
commit b1cf23ae3e
2015 changed files with 935048 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
examples/eval-callback/CMakeLists.txt Normal file
View File

@@ -0,0 +1,15 @@
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)
set(TEST_TARGET test-eval-callback)
if(NOT ${CMAKE_SYSTEM_PROCESSOR} MATCHES "s390x")
add_test(NAME ${TEST_TARGET}
COMMAND llama-eval-callback --hf-repo ggml-org/models --hf-file tinyllamas/stories260K.gguf --model stories260K.gguf --prompt hello --seed 42 -ngl 0)
else()
add_test(NAME ${TEST_TARGET}
COMMAND llama-eval-callback --hf-repo ggml-org/models --hf-file tinyllamas/stories260K-be.gguf --model stories260K-be.gguf --prompt hello --seed 42 -ngl 0)
endif()
set_property(TEST ${TEST_TARGET} PROPERTY LABELS eval-callback curl)

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, ...],
...
],
]
```

233
examples/eval-callback/eval-callback.cpp Normal file
View File

@@ -0,0 +1,233 @@
#include "arg.h"
#include "common.h"
#include "log.h"
#include "llama.h"
#include "ggml.h"
#include <cmath>
#include <cstdio>
#include <string>
#include <vector>
/**
* This the arbitrary data which will be passed to each callback.
* Later on we can for example add operation or tensor name filter from the CLI arg, or a file descriptor to dump the tensor.
*/
struct callback_data {
std::vector<uint8_t> data;
};
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 inline float ggml_compute_bf16_to_fp32(ggml_bf16_t h) {
union {
float f;
uint32_t i;
} u;
u.i = (uint32_t)h.bits << 16;
return u.f;
}
static float ggml_get_float_value(const uint8_t * data, ggml_type type, const size_t * nb, size_t i0, size_t i1, size_t i2, size_t i3) {
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(*(const ggml_fp16_t *) &data[i]);
} else if (type == GGML_TYPE_F32) {
v = *(const float *) &data[i];
} else if (type == GGML_TYPE_I64) {
v = (float) *(const int64_t *) &data[i];
} else if (type == GGML_TYPE_I32) {
v = (float) *(const int32_t *) &data[i];
} else if (type == GGML_TYPE_I16) {
v = (float) *(const int16_t *) &data[i];
} else if (type == GGML_TYPE_I8) {
v = (float) *(const int8_t *) &data[i];
} else if (type == GGML_TYPE_BF16) {
v = ggml_compute_bf16_to_fp32(*(const ggml_bf16_t *) &data[i]);
} else {
GGML_ABORT("fatal error");
}
return v;
}
static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne, const size_t * nb, int64_t n) {
GGML_ASSERT(n > 0);
float sum = 0;
for (int64_t i3 = 0; i3 < ne[3]; i3++) {
for (int64_t i2 = 0; i2 < ne[2]; i2++) {
for (int64_t i1 = 0; i1 < ne[1]; i1++) {
for (int64_t i0 = 0; i0 < ne[0]; i0++) {
const float v = ggml_get_float_value(data, type, nb, i0, i1, i2, i3);
sum += v;
}
}
}
}
for (int64_t i3 = 0; i3 < ne[3]; i3++) {
LOG(" [\n");
for (int64_t i2 = 0; i2 < ne[2]; i2++) {
if (i2 == n && ne[2] > 2*n) {
LOG(" ..., \n");
i2 = ne[2] - n;
}
LOG(" [\n");
for (int64_t i1 = 0; i1 < ne[1]; i1++) {
if (i1 == n && ne[1] > 2*n) {
LOG(" ..., \n");
i1 = ne[1] - n;
}
LOG(" [");
for (int64_t i0 = 0; i0 < ne[0]; i0++) {
if (i0 == n && ne[0] > 2*n) {
LOG("..., ");
i0 = ne[0] - n;
}
const float v = ggml_get_float_value(data, type, nb, i0, i1, i2, i3);
LOG("%12.4f", v);
if (i0 < ne[0] - 1) LOG(", ");
}
LOG("],\n");
}
LOG(" ],\n");
}
LOG(" ]\n");
LOG(" sum = %f\n", sum);
}
// TODO: make this abort configurable/optional?
if (std::isnan(sum)) {
LOG_ERR("encountered NaN - aborting\n");
exit(0);
}
}
/**
* GGML operations callback during the graph execution.
*
* @param t current tensor
* @param ask when ask is true, the scheduler wants to know if we are interested in data from this tensor
* if we return true, a follow-up call will be made with ask=false in which we can do the actual collection.
* see ggml_backend_sched_eval_callback
* @param user_data user data to pass at each call back
* @return true to receive data or continue the graph, false otherwise
*/
static bool ggml_debug(struct ggml_tensor * t, bool ask, void * user_data) {
auto * cb_data = (callback_data *) user_data;
const struct ggml_tensor * src0 = t->src[0];
const struct ggml_tensor * src1 = t->src[1];
if (ask) {
return true; // Always retrieve data
}
char src1_str[128] = {0};
if (src1) {
snprintf(src1_str, sizeof(src1_str), "%s{%s}", src1->name, ggml_ne_string(src1).c_str());
}
LOG("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__,
t->name, ggml_type_name(t->type), ggml_op_desc(t),
src0->name, ggml_ne_string(src0).c_str(),
src1 ? src1_str : "",
ggml_ne_string(t).c_str());
// copy the data from the GPU memory if needed
const bool is_host = ggml_backend_buffer_is_host(t->buffer);
if (!is_host) {
auto n_bytes = ggml_nbytes(t);
cb_data->data.resize(n_bytes);
ggml_backend_tensor_get(t, cb_data->data.data(), 0, n_bytes);
}
if (!ggml_is_quantized(t->type)) {
uint8_t * data = is_host ? (uint8_t *) t->data : cb_data->data.data();
ggml_print_tensor(data, t->type, t->ne, t->nb, 3);
}
return true;
}
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) {
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 = ggml_debug;
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;
}