Zihao Chen
GitHub
commit
0f2350dcb6
6 changed files with 1 additions and 282 deletions
3
Makefile
3
Makefile
|
|
@ -1,7 +1,6 @@
|
||||||
# Define the default target now so that it is always the first target
|
# Define the default target now so that it is always the first target
|
||||||
BUILD_TARGETS = \
|
BUILD_TARGETS = \
|
||||||
libllava.a \
|
libllava.a \
|
||||||
llama-batched-bench \
|
|
||||||
llama-bench \
|
llama-bench \
|
||||||
llama-benchmark-matmult \
|
llama-benchmark-matmult \
|
||||||
llama-cli \
|
llama-cli \
|
||||||
|
|
@ -39,7 +38,7 @@ BUILD_TARGETS = \
|
||||||
|
|
||||||
# Legacy build targets that were renamed in #7809, but should still be removed when the project is cleaned
|
# Legacy build targets that were renamed in #7809, but should still be removed when the project is cleaned
|
||||||
LEGACY_TARGETS_CLEAN = main quantize quantize-stats perplexity imatrix embedding vdot q8dot convert-llama2c-to-ggml \
|
LEGACY_TARGETS_CLEAN = main quantize quantize-stats perplexity imatrix embedding vdot q8dot convert-llama2c-to-ggml \
|
||||||
simple batched-bench save-load-state server gguf gguf-split eval-callback llama-bench libllava.a llava-cli \
|
simple save-load-state server gguf gguf-split eval-callback llama-bench libllava.a llava-cli \
|
||||||
retrieval speculative infill tokenize benchmark-matmult parallel export-lora lookahead lookup passkey gritlm
|
retrieval speculative infill tokenize benchmark-matmult parallel export-lora lookahead lookup passkey gritlm
|
||||||
|
|
||||||
# Legacy build targets that were renamed in #7809, but we want to build binaries that for them that output a deprecation warning if people try to use them.
|
# Legacy build targets that were renamed in #7809, but we want to build binaries that for them that output a deprecation warning if people try to use them.
|
||||||
|
|
|
||||||
|
|
@ -13,7 +13,6 @@ include_directories(${CMAKE_CURRENT_SOURCE_DIR})
|
||||||
if (EMSCRIPTEN)
|
if (EMSCRIPTEN)
|
||||||
else()
|
else()
|
||||||
add_subdirectory(cvector-generator)
|
add_subdirectory(cvector-generator)
|
||||||
add_subdirectory(batched-bench)
|
|
||||||
add_subdirectory(benchmark)
|
add_subdirectory(benchmark)
|
||||||
add_subdirectory(convert-llama2c-to-ggml)
|
add_subdirectory(convert-llama2c-to-ggml)
|
||||||
add_subdirectory(embedding)
|
add_subdirectory(embedding)
|
||||||
|
|
|
||||||
|
|
@ -1,5 +0,0 @@
|
||||||
set(TARGET llama-batched-bench)
|
|
||||||
add_executable(${TARGET} batched-bench.cpp)
|
|
||||||
install(TARGETS ${TARGET} RUNTIME)
|
|
||||||
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
|
|
||||||
target_compile_features(${TARGET} PRIVATE cxx_std_11)
|
|
||||||
|
|
@ -1,51 +0,0 @@
|
||||||
# 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 |
|
|
||||||
|
|
@ -1,215 +0,0 @@
|
||||||
#include "common.h"
|
|
||||||
#include "llama.h"
|
|
||||||
|
|
||||||
#include <algorithm>
|
|
||||||
#include <cmath>
|
|
||||||
#include <cstdio>
|
|
||||||
#include <string>
|
|
||||||
#include <vector>
|
|
||||||
|
|
||||||
// mutates the input string
|
|
||||||
static std::vector<int> parse_list(char * p) {
|
|
||||||
std::vector<int> ret;
|
|
||||||
|
|
||||||
char * q = p;
|
|
||||||
|
|
||||||
while (*p) {
|
|
||||||
if (*p == ',') {
|
|
||||||
*p = '\0';
|
|
||||||
ret.push_back(std::atoi(q));
|
|
||||||
q = p + 1;
|
|
||||||
}
|
|
||||||
|
|
||||||
++p;
|
|
||||||
}
|
|
||||||
|
|
||||||
ret.push_back(std::atoi(q));
|
|
||||||
|
|
||||||
return ret;
|
|
||||||
}
|
|
||||||
|
|
||||||
static void print_usage(int argc, char ** argv, const gpt_params & params) {
|
|
||||||
gpt_params_print_usage(argc, argv, params);
|
|
||||||
|
|
||||||
LOG_TEE("\nexample usage:\n");
|
|
||||||
LOG_TEE("\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_TEE("\n");
|
|
||||||
}
|
|
||||||
|
|
||||||
int main(int argc, char ** argv) {
|
|
||||||
gpt_params params;
|
|
||||||
|
|
||||||
if (!gpt_params_parse(argc, argv, params)) {
|
|
||||||
print_usage(argc, argv, params);
|
|
||||||
return 1;
|
|
||||||
}
|
|
||||||
|
|
||||||
int is_pp_shared = params.is_pp_shared;
|
|
||||||
|
|
||||||
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 = llama_model_params_from_gpt_params(params);
|
|
||||||
|
|
||||||
llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
|
|
||||||
|
|
||||||
if (model == NULL) {
|
|
||||||
fprintf(stderr , "%s: error: unable to load model\n" , __func__);
|
|
||||||
return 1;
|
|
||||||
}
|
|
||||||
|
|
||||||
llama_context_params ctx_params = llama_context_params_from_gpt_params(params);
|
|
||||||
|
|
||||||
// ensure enough sequences are available
|
|
||||||
ctx_params.n_seq_max = *std::max_element(n_pl.begin(), n_pl.end());
|
|
||||||
|
|
||||||
llama_context * ctx = llama_new_context_with_model(model, ctx_params);
|
|
||||||
|
|
||||||
if (ctx == NULL) {
|
|
||||||
fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
|
|
||||||
return 1;
|
|
||||||
}
|
|
||||||
|
|
||||||
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) {
|
|
||||||
for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) {
|
|
||||||
const int32_t n_tokens = std::min(n_batch, (int32_t) (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,
|
|
||||||
0, 0, 0, // unused
|
|
||||||
};
|
|
||||||
|
|
||||||
const int ret = llama_decode(ctx, batch_view);
|
|
||||||
if (ret != 0) {
|
|
||||||
LOG_TEE("failed to decode the batch, n_batch = %d, ret = %d\n", n_batch, ret);
|
|
||||||
return false;
|
|
||||||
}
|
|
||||||
|
|
||||||
llama_synchronize(ctx);
|
|
||||||
}
|
|
||||||
|
|
||||||
return true;
|
|
||||||
};
|
|
||||||
|
|
||||||
// warm up
|
|
||||||
{
|
|
||||||
for (int i = 0; i < 16; ++i) {
|
|
||||||
llama_batch_add(batch, 0, i, { 0 }, false);
|
|
||||||
}
|
|
||||||
|
|
||||||
if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
|
|
||||||
LOG_TEE("%s: llama_decode() failed\n", __func__);
|
|
||||||
return 1;
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
LOG_TEE("\n");
|
|
||||||
LOG_TEE("%s: 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\n", __func__, n_kv_max, params.n_batch, params.n_ubatch, params.flash_attn, params.is_pp_shared, params.n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch);
|
|
||||||
LOG_TEE("\n");
|
|
||||||
|
|
||||||
LOG_TEE("|%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_TEE("|%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 ? pp + pl*tg : pl*(pp + tg);
|
|
||||||
|
|
||||||
if (n_ctx_req > n_kv_max) {
|
|
||||||
continue;
|
|
||||||
}
|
|
||||||
|
|
||||||
llama_batch_clear(batch);
|
|
||||||
|
|
||||||
for (int i = 0; i < pp; ++i) {
|
|
||||||
for (int j = 0; j < (is_pp_shared ? 1 : pl); ++j) {
|
|
||||||
llama_batch_add(batch, 0, i, { j }, false);
|
|
||||||
}
|
|
||||||
}
|
|
||||||
batch.logits[batch.n_tokens - 1] = true;
|
|
||||||
|
|
||||||
const auto t_pp_start = ggml_time_us();
|
|
||||||
|
|
||||||
llama_kv_cache_clear(ctx);
|
|
||||||
|
|
||||||
if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
|
|
||||||
LOG_TEE("%s: llama_decode() failed\n", __func__);
|
|
||||||
return 1;
|
|
||||||
}
|
|
||||||
|
|
||||||
if (is_pp_shared) {
|
|
||||||
for (int32_t i = 1; i < pl; ++i) {
|
|
||||||
llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
const auto t_pp_end = ggml_time_us();
|
|
||||||
|
|
||||||
const auto t_tg_start = ggml_time_us();
|
|
||||||
|
|
||||||
for (int i = 0; i < tg; ++i) {
|
|
||||||
llama_batch_clear(batch);
|
|
||||||
|
|
||||||
for (int j = 0; j < pl; ++j) {
|
|
||||||
llama_batch_add(batch, 0, pp + i, { j }, true);
|
|
||||||
}
|
|
||||||
|
|
||||||
if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
|
|
||||||
LOG_TEE("%s: llama_decode() failed\n", __func__);
|
|
||||||
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 = n_kv / t;
|
|
||||||
|
|
||||||
LOG_TEE("|%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);
|
|
||||||
}
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
llama_print_timings(ctx);
|
|
||||||
|
|
||||||
llama_batch_free(batch);
|
|
||||||
|
|
||||||
llama_free(ctx);
|
|
||||||
llama_free_model(model);
|
|
||||||
|
|
||||||
llama_backend_free();
|
|
||||||
|
|
||||||
fprintf(stderr, "\n\n");
|
|
||||||
|
|
||||||
return 0;
|
|
||||||
}
|
|
||||||
|
|
@ -183,11 +183,6 @@ if [ "$1" -eq "1" ]; then
|
||||||
|
|
||||||
make -j && ./bin/llama-perplexity -m ../models/tinyllama-1b/ggml-model-f16.gguf -f ./wikitext-2-raw/wiki.test.raw -ngl 100 --chunks 32
|
make -j && ./bin/llama-perplexity -m ../models/tinyllama-1b/ggml-model-f16.gguf -f ./wikitext-2-raw/wiki.test.raw -ngl 100 --chunks 32
|
||||||
|
|
||||||
# batched-bench
|
|
||||||
cd /workspace/llama.cpp
|
|
||||||
|
|
||||||
GGML_CUDA=1 make -j && ./llama-batched-bench ./models/tinyllama-1b/ggml-model-f16.gguf 4608 1 99 0 512 128 1,2,3,4,5,6,7,8,16,32
|
|
||||||
|
|
||||||
# parallel
|
# parallel
|
||||||
cd /workspace/llama.cpp
|
cd /workspace/llama.cpp
|
||||||
|
|
||||||
|
|
@ -202,6 +197,3 @@ fi
|
||||||
# GGML_CUDA=1 make -j && ./llama-speculative -m ./models/codellama-34b-instruct/ggml-model-f16.gguf -md ./models/codellama-7b-instruct/ggml-model-q4_0.gguf -p "# Dijkstra's shortest path algorithm in Python (4 spaces indentation) + complexity analysis:\n\n" -e -ngl 999 -ngld 999 -t 4 -n 512 -c 4096 -s 21 --draft 16 -np 1 --temp 0.0
|
# GGML_CUDA=1 make -j && ./llama-speculative -m ./models/codellama-34b-instruct/ggml-model-f16.gguf -md ./models/codellama-7b-instruct/ggml-model-q4_0.gguf -p "# Dijkstra's shortest path algorithm in Python (4 spaces indentation) + complexity analysis:\n\n" -e -ngl 999 -ngld 999 -t 4 -n 512 -c 4096 -s 21 --draft 16 -np 1 --temp 0.0
|
||||||
#fi
|
#fi
|
||||||
|
|
||||||
# more benches
|
|
||||||
#GGML_CUDA=1 make -j && ./llama-batched-bench ./models/codellama-7b/ggml-model-q4_k.gguf 4096 1 99 1 512,3200 128,128,800 1
|
|
||||||
#GGML_CUDA=1 make -j && ./llama-batched-bench ./models/codellama-13b/ggml-model-q4_k.gguf 4096 1 99 1 512,3200 128,128,800 1
|
|
||||||
|
|
|
||||||
Loading…
Add table
Add a link
Reference in a new issue