vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

eval-callback: Example how to use eval callback for debugging (#6576)

Browse source 
* gguf-debug: Example how to use ggml callback for debugging

* gguf-debug: no mutex, verify type, fix stride.

* llama: cv eval: move cb eval field in common gpt_params

* ggml_debug: use common gpt_params to pass cb eval.
Fix get tensor SIGV random.

* ggml_debug: ci: add tests

* ggml_debug: EOL in CMakeLists.txt

* ggml_debug: Remove unused param n_batch, no batching here

* ggml_debug: fix trailing spaces

* ggml_debug: fix trailing spaces

* common: fix cb_eval and user data not initialized

* ci: build revert label

* ggml_debug: add main test label

* doc: add a model: add a link to ggml-debug

* ggml-debug: add to make toolchain

* ggml-debug: tests add the main label

* ggml-debug: ci add test curl label

* common: allow the warmup to be disabled in llama_init_from_gpt_params

* ci: add curl test

* ggml-debug: better tensor type support

* gitignore : ggml-debug

* ggml-debug: printing also the sum of each tensor

* ggml-debug: remove block size

* eval-callback: renamed from ggml-debug

* eval-callback: fix make toolchain

---------

Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Pierrick Hymbert 2024-04-11 14:51:07 +02:00 • committed by GitHub
parent 8228b66dbc
commit b804b1ef77
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
12 changed files with 319 additions and 22 deletions
Download patch file Download diff file Expand all files Collapse all files

9
examples/eval-callback/CMakeLists.txt Normal file
View file

@ -0,0 +1,9 @@
set(TARGET 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_11)
set(TEST_TARGET test-eval-callback)
add_test(NAME ${TEST_TARGET} COMMAND eval-callback --hf-repo ggml-org/models --hf-file tinyllamas/stories260K.gguf --model stories260K.gguf --prompt hello --seed 42)
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
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, ...],
...
],
]
```

185
examples/eval-callback/eval-callback.cpp Normal file
View file

@ -0,0 +1,185 @@
#include "common.h"
#include "llama.h"
#include "ggml.h"
#include <cstdio>
#include <random>
#include <string>
#include <tuple>
#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 void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne, const size_t * nb, int64_t n) {
float sum = 0;
for (int64_t i3 = 0; i3 < ne[3]; i3++) {
printf(" [\n");
for (int64_t i2 = 0; i2 < ne[2] && i2 < n; i2++) {
printf(" [\n");
for (int64_t i1 = 0; i1 < ne[1] && i1 < n; i1++) {
printf(" [");
for (int64_t i0 = 0; i0 < ne[0] && i0 < n; i0++) {
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_ASSERT(false);
}
printf("%8.4f", v);
sum += v;
if (i0 < ne[0] - 1 && i0 < n - 1) printf(", ");
}
if (ne[0] > n) printf(", ...");
printf("],\n");
}
if (ne[1] > n) printf(" ...\n");
printf(" ],\n");
}
if (ne[2] > n) printf(" ...\n");
printf(" ]\n");
printf(" sum = %f\n", sum);
}
}
/**
* 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) {
sprintf(src1_str, "%s{%s}", src1->name, ggml_ne_string(src1).c_str());
}
printf("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__,
t->name, ggml_type_name(t->type), ggml_op_name(t->op),
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 gpt_params & params) {
const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), 0, 0))) {
fprintf(stderr, "%s : failed to eval\n", __func__);
return false;
}
return true;
}
int main(int argc, char ** argv) {
callback_data cb_data;
gpt_params params;
if (!gpt_params_parse(argc, argv, params)) {
return 1;
}
print_build_info();
std::mt19937 rng(params.seed);
if (params.random_prompt) {
params.prompt = gpt_random_prompt(rng);
}
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
llama_model * model;
llama_context * ctx;
std::tie(model, ctx) = llama_init_from_gpt_params(params);
if (model == nullptr || ctx == nullptr) {
fprintf(stderr, "%s : failed to init\n", __func__);
return 1;
}
// print system information
{
fprintf(stderr, "\n");
fprintf(stderr, "%s\n", get_system_info(params).c_str());
}
bool OK = run(ctx, params);
if (!OK) {
return 1;
}
llama_print_timings(ctx);
llama_free(ctx);
llama_free_model(model);
llama_backend_free();
return 0;
}