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

Merge branch 'master' into gg/flash-attn

Browse source
This commit is contained in:
Georgi Gerganov 2024-01-31 18:49:43 +02:00
commit 2ddc9bbef1
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
30 changed files with 1758 additions and 1597 deletions
Download patch file Download diff file Expand all files Collapse all files

2
examples/batched-bench/batched-bench.cpp
View file

@ -88,7 +88,7 @@ int main(int argc, char ** argv) {
llama_model_params model_params = llama_model_default_params();
const std::vector<float> t_split (LLAMA_MAX_DEVICES, 0.0f);
const std::vector<float> t_split(llama_max_devices(), 0.0f);
model_params.n_gpu_layers = n_gpu_layers;
model_params.tensor_split = t_split.data();

31
examples/llama-bench/llama-bench.cpp
View file

@ -160,7 +160,7 @@ struct cmd_params {
std::vector<int> main_gpu;
std::vector<bool> no_kv_offload;
std::vector<bool> mul_mat_q;
std::vector<std::array<float, LLAMA_MAX_DEVICES>> tensor_split;
std::vector<std::vector<float>> tensor_split;
int reps;
bool verbose;
output_formats output_format;
@ -179,7 +179,7 @@ static const cmd_params cmd_params_defaults = {
/* main_gpu */ {0},
/* no_kv_offload */ {false},
/* mul_mat_q */ {true},
/* tensor_split */ {{}},
/* tensor_split */ {std::vector<float>(llama_max_devices(), 0.0f)},
/* reps */ 5,
/* verbose */ false,
/* output_format */ MARKDOWN
@ -380,10 +380,10 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
const std::regex regex{R"([;/]+)"};
std::sregex_token_iterator it{ts.begin(), ts.end(), regex, -1};
std::vector<std::string> split_arg{it, {}};
GGML_ASSERT(split_arg.size() <= LLAMA_MAX_DEVICES);
GGML_ASSERT(split_arg.size() <= llama_max_devices());
std::array<float, LLAMA_MAX_DEVICES> tensor_split;
for (size_t i = 0; i < LLAMA_MAX_DEVICES; ++i) {
std::vector<float> tensor_split(llama_max_devices());
for (size_t i = 0; i < llama_max_devices(); ++i) {
if (i < split_arg.size()) {
tensor_split[i] = std::stof(split_arg[i]);
} else {
@ -459,7 +459,7 @@ struct cmd_params_instance {
int main_gpu;
bool no_kv_offload;
bool mul_mat_q;
std::array<float, LLAMA_MAX_DEVICES> tensor_split;
std::vector<float> tensor_split;
llama_model_params to_llama_mparams() const {
llama_model_params mparams = llama_model_default_params();
@ -563,6 +563,7 @@ struct test {
static const bool cuda;
static const bool opencl;
static const bool vulkan;
static const bool kompute;
static const bool metal;
static const bool gpu_blas;
static const bool blas;
@ -581,7 +582,7 @@ struct test {
int main_gpu;
bool no_kv_offload;
bool mul_mat_q;
std::array<float, LLAMA_MAX_DEVICES> tensor_split;
std::vector<float> tensor_split;
int n_prompt;
int n_gen;
std::string test_time;
@ -647,6 +648,9 @@ struct test {
if (vulkan) {
return "Vulkan";
}
if (kompute) {
return "Kompute";
}
if (metal) {
return "Metal";
}
@ -662,7 +666,7 @@ struct test {
static const std::vector<std::string> & get_fields() {
static const std::vector<std::string> fields = {
"build_commit", "build_number",
"cuda", "opencl", "vulkan", "metal", "gpu_blas", "blas",
"cuda", "opencl", "vulkan", "kompute", "metal", "gpu_blas", "blas",
"cpu_info", "gpu_info",
"model_filename", "model_type", "model_size", "model_n_params",
"n_batch", "n_threads", "type_k", "type_v",
@ -686,8 +690,9 @@ struct test {
field == "avg_ns" || field == "stddev_ns") {
return INT;
}
if (field == "cuda" || field == "opencl" || field == "vulkan"|| field == "metal" || field == "gpu_blas" || field == "blas" ||
field == "f16_kv" || field == "no_kv_offload" || field == "mul_mat_q") {
if (field == "cuda" || field == "opencl" || field == "vulkan" || field == "kompute" || field == "metal" ||
field == "gpu_blas" || field == "blas" || field == "f16_kv" || field == "no_kv_offload" ||
field == "mul_mat_q") {
return BOOL;
}
if (field == "avg_ts" || field == "stddev_ts") {
@ -699,7 +704,7 @@ struct test {
std::vector<std::string> get_values() const {
std::string tensor_split_str;
int max_nonzero = 0;
for (int i = 0; i < LLAMA_MAX_DEVICES; i++) {
for (size_t i = 0; i < llama_max_devices(); i++) {
if (tensor_split[i] > 0) {
max_nonzero = i;
}
@ -714,7 +719,8 @@ struct test {
}
std::vector<std::string> values = {
build_commit, std::to_string(build_number),
std::to_string(cuda), std::to_string(opencl), std::to_string(vulkan), std::to_string(metal), std::to_string(gpu_blas), std::to_string(blas),
std::to_string(cuda), std::to_string(opencl), std::to_string(vulkan), std::to_string(vulkan),
std::to_string(metal), std::to_string(gpu_blas), std::to_string(blas),
cpu_info, gpu_info,
model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params),
std::to_string(n_batch), std::to_string(n_threads), ggml_type_name(type_k), ggml_type_name(type_v),
@ -743,6 +749,7 @@ const int test::build_number = LLAMA_BUILD_NUMBER;
const bool test::cuda = !!ggml_cpu_has_cublas();
const bool test::opencl = !!ggml_cpu_has_clblast();
const bool test::vulkan = !!ggml_cpu_has_vulkan();
const bool test::kompute = !!ggml_cpu_has_kompute();
const bool test::metal = !!ggml_cpu_has_metal();
const bool test::gpu_blas = !!ggml_cpu_has_gpublas();
const bool test::blas = !!ggml_cpu_has_blas();

58
examples/llava/MobileVLM-README.md
View file

@ -111,17 +111,71 @@ llama_print_timings: eval time = 1279.03 ms / 18 runs ( 71.06 m
llama_print_timings: total time = 34570.79 ms
```
## Orin compile and run
### compile
```sh
make LLAMA_CUBLAS=1 CUDA_DOCKER_ARCH=sm_87 LLAMA_CUDA_F16=1 -j 32
```
### run on Orin
### case 1
**input**
```sh
./llava-cli \
-m /data/local/tmp/ggml-model-q4_k.gguf \
--mmproj /data/local/tmp/mmproj-model-f16.gguf \
--image /data/local/tmp/demo.jpeg \
-p "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\nWho is the author of this book? \nAnswer the question using a single word or phrase. ASSISTANT:" \
--n-gpu-layers 999
```
**output**
```sh
encode_image_with_clip: image encoded in 296.62 ms by CLIP ( 2.06 ms per image patch)
Susan Wise Bauer
llama_print_timings: load time = 1067.64 ms
llama_print_timings: sample time = 1.53 ms / 6 runs ( 0.25 ms per token, 3934.43 tokens per second)
llama_print_timings: prompt eval time = 306.84 ms / 246 tokens ( 1.25 ms per token, 801.72 tokens per second)
llama_print_timings: eval time = 91.50 ms / 6 runs ( 15.25 ms per token, 65.58 tokens per second)
llama_print_timings: total time = 1352.63 ms / 252 tokens
```
### case 2
**input**
```sh
./llava-cli \
-m /data/local/tmp/ggml-model-q4_k.gguf \
--mmproj /data/local/tmp/mmproj-model-f16.gguf \
-p "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\nWhat is in the image? ASSISTANT:" \
--n-gpu-layers 999
```
**output**
```sh
encode_image_with_clip: image encoded in 302.15 ms by CLIP ( 2.10 ms per image patch)
The image features a cat lying in the grass.
llama_print_timings: load time = 1057.07 ms
llama_print_timings: sample time = 3.27 ms / 11 runs ( 0.30 ms per token, 3360.83 tokens per second)
llama_print_timings: prompt eval time = 213.60 ms / 232 tokens ( 0.92 ms per token, 1086.14 tokens per second)
llama_print_timings: eval time = 166.65 ms / 11 runs ( 15.15 ms per token, 66.01 tokens per second)
llama_print_timings: total time = 1365.47 ms / 243 tokens
```
## Minor shortcomings
The `n_patch` of output in `ldp` is 1/4 of the input. In order to implement quickly, we uniformly modified `clip_n_patches` function to a quarter. when counting the time consumption, the calculated time will be 4 times bigger than the real cost.
## TODO
- [ ] Support non-CPU backend for the new operators, such as `depthwise`, `hardswish`, `hardsigmoid`
- [x] Support non-CPU backend for the new operators, such as `depthwise`, `hardswish`, `hardsigmoid`
- [ ] Optimize LDP projector performance
- Optimize the structure definition to avoid unnecessary memory rearrangements, to reduce the use of `ggml_permute_cpy`;
- Optimize operator implementation (ARM CPU/NVIDIA GPU): such as depthwise conv, hardswish, hardsigmoid, etc.
- [ ] run MobileVLM on `Jetson Orin`
- [x] run MobileVLM on `Jetson Orin`
- [ ] Support more model variants, such as `MobileVLM-3B`.

44
examples/server/server.cpp
View file

@ -1789,28 +1789,28 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
printf(" -b N, --batch-size N batch size for prompt processing (default: %d)\n", params.n_batch);
printf(" --memory-f32 use f32 instead of f16 for memory key+value (default: disabled)\n");
printf(" not recommended: doubles context memory required and no measurable increase in quality\n");
if (llama_mlock_supported())
if (llama_supports_mlock())
{
printf(" --mlock force system to keep model in RAM rather than swapping or compressing\n");
}
if (llama_mmap_supported())
if (llama_supports_mmap())
{
printf(" --no-mmap do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
}
printf(" --numa attempt optimizations that help on some NUMA systems\n");
#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
printf(" -ngl N, --n-gpu-layers N\n");
printf(" number of layers to store in VRAM\n");
printf(" -sm SPLIT_MODE, --split-mode SPLIT_MODE\n");
printf(" how to split the model across multiple GPUs, one of:\n");
printf(" - none: use one GPU only\n");
printf(" - layer (default): split layers and KV across GPUs\n");
printf(" - row: split rows across GPUs\n");
printf(" -ts SPLIT --tensor-split SPLIT\n");
printf(" fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1\n");
printf(" -mg i, --main-gpu i the GPU to use for the model (with split-mode = none),\n");
printf(" or for intermediate results and KV (with split-mode = row)\n");
#endif
if (llama_supports_gpu_offload()) {
printf(" -ngl N, --n-gpu-layers N\n");
printf(" number of layers to store in VRAM\n");
printf(" -sm SPLIT_MODE, --split-mode SPLIT_MODE\n");
printf(" how to split the model across multiple GPUs, one of:\n");
printf(" - none: use one GPU only\n");
printf(" - layer (default): split layers and KV across GPUs\n");
printf(" - row: split rows across GPUs\n");
printf(" -ts SPLIT --tensor-split SPLIT\n");
printf(" fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1\n");
printf(" -mg i, --main-gpu i the GPU to use for the model (with split-mode = none),\n");
printf(" or for intermediate results and KV (with split-mode = row)\n");
}
printf(" -m FNAME, --model FNAME\n");
printf(" model path (default: %s)\n", params.model.c_str());
printf(" -a ALIAS, --alias ALIAS\n");
@ -2066,13 +2066,13 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
invalid_param = true;
break;
}
#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
params.n_gpu_layers = std::stoi(argv[i]);
#else
LOG_WARNING("Not compiled with GPU offload support, --n-gpu-layers option will be ignored. "
if (llama_supports_gpu_offload()) {
params.n_gpu_layers = std::stoi(argv[i]);
} else {
LOG_WARNING("Not compiled with GPU offload support, --n-gpu-layers option will be ignored. "
"See main README.md for information on enabling GPU BLAS support",
{{"n_gpu_layers", params.n_gpu_layers}});
#endif
}
}
else if (arg == "--split-mode" || arg == "-sm")
{
@ -2115,9 +2115,9 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
const std::regex regex{R"([,/]+)"};
std::sregex_token_iterator it{arg_next.begin(), arg_next.end(), regex, -1};
std::vector<std::string> split_arg{it, {}};
GGML_ASSERT(split_arg.size() <= LLAMA_MAX_DEVICES);
GGML_ASSERT(split_arg.size() <= llama_max_devices());
for (size_t i_device = 0; i_device < LLAMA_MAX_DEVICES; ++i_device)
for (size_t i_device = 0; i_device < llama_max_devices(); ++i_device)
{
if (i_device < split_arg.size())
{

10
examples/sycl/ls-sycl-device.cpp
View file

@ -1,7 +1,9 @@
/*MIT license
Copyright (C) 2024 Intel Corporation
SPDX-License-Identifier: MIT
*/
//
// MIT license
// Copyright (C) 2024 Intel Corporation
// SPDX-License-Identifier: MIT
//
#include "ggml-sycl.h"

23
examples/sycl/win-build-sycl.bat Normal file
View file

@ -0,0 +1,23 @@
:: MIT license
:: Copyright (C) 2024 Intel Corporation
:: SPDX-License-Identifier: MIT
mkdir -p build
cd build
@call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
:: for FP16
:: faster for long-prompt inference
:: cmake -G "MinGW Makefiles" .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx -DCMAKE_BUILD_TYPE=Release -DLLAMA_SYCL_F16=ON
:: for FP32
cmake -G "MinGW Makefiles" .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx -DCMAKE_BUILD_TYPE=Release
:: build example/main only
:: make main
:: build all binary
make -j
cd ..

13
examples/sycl/win-run-llama2.bat Normal file
View file

@ -0,0 +1,13 @@
:: MIT license
:: Copyright (C) 2024 Intel Corporation
:: SPDX-License-Identifier: MIT
INPUT2="Building a website can be done in 10 simple steps:\nStep 1:"
@call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
set GGML_SYCL_DEVICE=0
rem set GGML_SYCL_DEBUG=1
.\build\bin\main.exe -m models\llama-2-7b.Q4_0.gguf -p %INPUT2% -n 400 -e -ngl 33 -s 0