mirror of
https://github.com/ggerganov/llama.cpp.git
synced 2024-09-22 21:16:20 +00:00
Compare commits
17 Commits
d1a7741e4f
...
80daabf452
Author | SHA1 | Date | |
---|---|---|---|
|
80daabf452 | ||
|
8db003a19d | ||
|
0996c5597f | ||
|
5bb2c5dbd2 | ||
|
67155ab7f5 | ||
|
5af118efda | ||
|
d2b496bff4 | ||
|
424e3a52fe | ||
|
56c5f988eb | ||
|
97efd5047a | ||
|
cc9514f941 | ||
|
f57f8cb3da | ||
|
3676778e82 | ||
|
d94ad56f87 | ||
|
74ba8516ce | ||
|
e914ac7c68 | ||
|
3666c861d4 |
@ -941,11 +941,37 @@ struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_p
|
||||
|
||||
#ifdef LLAMA_USE_CURL
|
||||
|
||||
#define CURL_MAX_RETRY 3
|
||||
#define CURL_RETRY_DELAY_SECONDS 2
|
||||
|
||||
|
||||
static bool starts_with(const std::string & str, const std::string & prefix) {
|
||||
// While we wait for C++20's std::string::starts_with...
|
||||
return str.rfind(prefix, 0) == 0;
|
||||
}
|
||||
|
||||
static bool curl_perform_with_retry(const std::string& url, CURL* curl, int max_attempts, int retry_delay_seconds) {
|
||||
int remaining_attempts = max_attempts;
|
||||
|
||||
while (remaining_attempts > 0) {
|
||||
fprintf(stderr, "%s: Trying to download from %s (attempt %d of %d)...\n", __func__ , url.c_str(), max_attempts - remaining_attempts + 1, max_attempts);
|
||||
|
||||
CURLcode res = curl_easy_perform(curl);
|
||||
if (res == CURLE_OK) {
|
||||
return true;
|
||||
}
|
||||
|
||||
int exponential_backoff_delay = std::pow(retry_delay_seconds, max_attempts - remaining_attempts) * 1000;
|
||||
fprintf(stderr, "%s: curl_easy_perform() failed: %s, retrying after %d milliseconds...\n", __func__, curl_easy_strerror(res), exponential_backoff_delay);
|
||||
|
||||
remaining_attempts--;
|
||||
std::this_thread::sleep_for(std::chrono::milliseconds(exponential_backoff_delay));
|
||||
}
|
||||
|
||||
fprintf(stderr, "%s: curl_easy_perform() failed after %d attempts\n", __func__, max_attempts);
|
||||
return false;
|
||||
}
|
||||
|
||||
static bool llama_download_file(const std::string & url, const std::string & path, const std::string & hf_token) {
|
||||
|
||||
// Initialize libcurl
|
||||
@ -1049,9 +1075,8 @@ static bool llama_download_file(const std::string & url, const std::string & pat
|
||||
curl_easy_setopt(curl.get(), CURLOPT_HEADERFUNCTION, static_cast<CURLOPT_HEADERFUNCTION_PTR>(header_callback));
|
||||
curl_easy_setopt(curl.get(), CURLOPT_HEADERDATA, &headers);
|
||||
|
||||
CURLcode res = curl_easy_perform(curl.get());
|
||||
if (res != CURLE_OK) {
|
||||
fprintf(stderr, "%s: curl_easy_perform() failed: %s\n", __func__, curl_easy_strerror(res));
|
||||
bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS);
|
||||
if (!was_perform_successful) {
|
||||
return false;
|
||||
}
|
||||
|
||||
@ -1126,11 +1151,10 @@ static bool llama_download_file(const std::string & url, const std::string & pat
|
||||
};
|
||||
|
||||
// start the download
|
||||
fprintf(stderr, "%s: downloading from %s to %s (server_etag:%s, server_last_modified:%s)...\n", __func__,
|
||||
llama_download_hide_password_in_url(url).c_str(), path.c_str(), headers.etag.c_str(), headers.last_modified.c_str());
|
||||
auto res = curl_easy_perform(curl.get());
|
||||
if (res != CURLE_OK) {
|
||||
fprintf(stderr, "%s: curl_easy_perform() failed: %s\n", __func__, curl_easy_strerror(res));
|
||||
fprintf(stderr, "%s: trying to download model from %s to %s (server_etag:%s, server_last_modified:%s)...\n", __func__,
|
||||
llama_download_hide_password_in_url(url).c_str(), path.c_str(), headers.etag.c_str(), headers.last_modified.c_str());
|
||||
bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS);
|
||||
if (!was_perform_successful) {
|
||||
return false;
|
||||
}
|
||||
|
||||
|
@ -31,6 +31,7 @@ import re
|
||||
import requests
|
||||
import sys
|
||||
import json
|
||||
import shutil
|
||||
|
||||
from hashlib import sha256
|
||||
from enum import IntEnum, auto
|
||||
@ -125,12 +126,27 @@ def download_model(model):
|
||||
if tokt == TOKENIZER_TYPE.UGM:
|
||||
files.append("spiece.model")
|
||||
|
||||
for file in files:
|
||||
save_path = f"models/tokenizers/{name}/{file}"
|
||||
if os.path.isfile(save_path):
|
||||
logger.info(f"{name}: File {save_path} already exists - skipping")
|
||||
continue
|
||||
download_file_with_auth(f"{repo}/resolve/main/{file}", token, save_path)
|
||||
if os.path.isdir(repo):
|
||||
# If repo is a path on the file system, copy the directory
|
||||
for file in files:
|
||||
src_path = os.path.join(repo, file)
|
||||
dst_path = f"models/tokenizers/{name}/{file}"
|
||||
if os.path.isfile(dst_path):
|
||||
logger.info(f"{name}: File {dst_path} already exists - skipping")
|
||||
continue
|
||||
if os.path.isfile(src_path):
|
||||
shutil.copy2(src_path, dst_path)
|
||||
logger.info(f"{name}: Copied {src_path} to {dst_path}")
|
||||
else:
|
||||
logger.warning(f"{name}: Source file {src_path} does not exist")
|
||||
else:
|
||||
# If repo is a URL, download the files
|
||||
for file in files:
|
||||
save_path = f"models/tokenizers/{name}/{file}"
|
||||
if os.path.isfile(save_path):
|
||||
logger.info(f"{name}: File {save_path} already exists - skipping")
|
||||
continue
|
||||
download_file_with_auth(f"{repo}/resolve/main/{file}", token, save_path)
|
||||
|
||||
|
||||
for model in models:
|
||||
|
@ -3,32 +3,10 @@
|
||||
#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, char ** argv) {
|
||||
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]);
|
||||
|
@ -18,8 +18,8 @@ struct llava_context {
|
||||
};
|
||||
|
||||
static void show_additional_info(int /*argc*/, char ** argv) {
|
||||
LOG_TEE("\n example usage: %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
|
||||
LOG_TEE(" note: a lower temperature value like 0.1 is recommended for better quality.\n");
|
||||
LOG_TEE("\nexample usage:\n\n%s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
|
||||
LOG_TEE("\nnote: a lower temperature value like 0.1 is recommended for better quality.\n");
|
||||
}
|
||||
|
||||
static void llama_log_callback_logTee(ggml_log_level level, const char * text, void * user_data) {
|
||||
@ -255,7 +255,7 @@ int main(int argc, char ** argv) {
|
||||
|
||||
gpt_params params;
|
||||
|
||||
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON, show_additional_info)) {
|
||||
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, show_additional_info)) {
|
||||
return 1;
|
||||
}
|
||||
|
||||
|
@ -11,6 +11,8 @@
|
||||
extern "C" {
|
||||
#endif
|
||||
|
||||
#define GGML_KOMPUTE_MAX_DEVICES 16
|
||||
|
||||
struct ggml_vk_device {
|
||||
int index;
|
||||
int type; // same as VkPhysicalDeviceType
|
||||
@ -23,10 +25,10 @@ struct ggml_vk_device {
|
||||
};
|
||||
|
||||
struct ggml_vk_device * ggml_vk_available_devices(size_t memoryRequired, size_t * count);
|
||||
int ggml_backend_kompute_get_device_count(void);
|
||||
void ggml_backend_kompute_get_device_memory(int device, size_t * free, size_t * total);
|
||||
bool ggml_vk_get_device(struct ggml_vk_device * device, size_t memoryRequired, const char * name);
|
||||
bool ggml_vk_has_vulkan(void);
|
||||
bool ggml_vk_has_device(void);
|
||||
struct ggml_vk_device ggml_vk_current_device(void);
|
||||
|
||||
//
|
||||
// backend API
|
||||
|
@ -26,7 +26,11 @@ void ggml_cuda_op_mul_mat_q(
|
||||
// nrows_dst == nrows of the matrix that the kernel writes into
|
||||
const int64_t nrows_dst = id == ctx.device ? ne0 : row_diff;
|
||||
|
||||
const mmq_args args = {src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stride00, src1_padded_row_size, src1_ncols, ne11, nrows_dst};
|
||||
// The stream-k decomposition is only faster for recent NVIDIA GPUs.
|
||||
// Also its fixup needs to allocate a temporary buffer in the memory pool.
|
||||
// There are multiple parallel CUDA streams for src1_ncols != ne11 which would introduce a race condition for this buffer.
|
||||
const bool use_stream_k = compute_capability >= CC_VOLTA && compute_capability < CC_OFFSET_AMD && src1_ncols == ne11;
|
||||
const mmq_args args = {src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stride00, src1_padded_row_size, src1_ncols, ne11, nrows_dst, use_stream_k};
|
||||
|
||||
switch (src0->type) {
|
||||
case GGML_TYPE_Q4_0:
|
||||
|
@ -2742,6 +2742,7 @@ struct mmq_args {
|
||||
int64_t ne00; int64_t ne01; int64_t stride01;
|
||||
int64_t ne10; int64_t ne11; int64_t stride11;
|
||||
int64_t ne0;
|
||||
bool use_stream_k;
|
||||
};
|
||||
|
||||
template<ggml_type type>
|
||||
@ -2777,8 +2778,7 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
|
||||
const int ntx = (args.ne11 + mmq_x - 1) / mmq_x;
|
||||
const dim3 block_nums_xy_tiling(nty, ntx, 1);
|
||||
|
||||
const bool use_stream_k = cc >= CC_VOLTA && cc < CC_OFFSET_AMD;
|
||||
if (!use_stream_k) {
|
||||
if (!args.use_stream_k) {
|
||||
if (args.ne01 % mmq_y == 0) {
|
||||
constexpr bool need_check = false;
|
||||
mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, shmem, stream>>>
|
||||
|
File diff suppressed because it is too large
Load Diff
@ -3330,6 +3330,8 @@ static size_t llama_get_device_count(const llama_model & model) {
|
||||
count = ggml_backend_sycl_get_device_count();
|
||||
#elif defined(GGML_USE_VULKAN)
|
||||
count = ggml_backend_vk_get_device_count();
|
||||
#elif defined(GGML_USE_KOMPUTE)
|
||||
count = ggml_backend_kompute_get_device_count();
|
||||
#elif defined(GGML_USE_CANN)
|
||||
return ggml_backend_cann_get_device_count();
|
||||
#endif
|
||||
@ -3434,6 +3436,11 @@ static size_t llama_get_device_memory(const llama_model & model, int device) {
|
||||
size_t free;
|
||||
ggml_backend_vk_get_device_memory(local_device, &free, &total);
|
||||
return free;
|
||||
#elif defined(GGML_USE_KOMPUTE)
|
||||
size_t total;
|
||||
size_t free;
|
||||
ggml_backend_kompute_get_device_memory(device, &free, &total);
|
||||
return free;
|
||||
#elif defined(GGML_USE_CANN)
|
||||
size_t total;
|
||||
size_t free;
|
||||
@ -17985,6 +17992,8 @@ size_t llama_max_devices(void) {
|
||||
return GGML_SYCL_MAX_DEVICES;
|
||||
#elif defined(GGML_USE_VULKAN)
|
||||
return GGML_VK_MAX_DEVICES;
|
||||
#elif defined(GGML_USE_KOMPUTE)
|
||||
return GGML_KOMPUTE_MAX_DEVICES;
|
||||
#elif defined(GGML_USE_CANN)
|
||||
return GGML_CANN_MAX_DEVICES;
|
||||
#else
|
||||
@ -18337,13 +18346,35 @@ struct llama_context * llama_new_context_with_model(
|
||||
}
|
||||
#elif defined(GGML_USE_KOMPUTE)
|
||||
if (model->n_gpu_layers > 0) {
|
||||
auto * backend = ggml_backend_kompute_init(model->main_gpu);
|
||||
if (backend == nullptr) {
|
||||
LLAMA_LOG_ERROR("%s: failed to initialize Kompute backend\n", __func__);
|
||||
if (model->split_mode == LLAMA_SPLIT_MODE_NONE) {
|
||||
auto * backend = ggml_backend_kompute_init(model->main_gpu);
|
||||
if (!backend) {
|
||||
LLAMA_LOG_ERROR("%s: failed to initialize Kompute backend\n", __func__);
|
||||
llama_free(ctx);
|
||||
return nullptr;
|
||||
}
|
||||
ctx->backends.push_back(backend);
|
||||
} else if (model->split_mode == LLAMA_SPLIT_MODE_LAYER) {
|
||||
size_t count = 0;
|
||||
auto * devices =ggml_vk_available_devices(0, &count);
|
||||
for (size_t i = 0; i < count; i++) {
|
||||
LLAMA_LOG_INFO("Kompute: Found device #%d, %s, %s, max-alloc %ld, heap-size %lu\n",
|
||||
devices[i].index, devices[i].vendor, devices[i].name,
|
||||
devices[i].maxAlloc, devices[i].heapSize);
|
||||
auto * backend = ggml_backend_kompute_init(devices[i].index);
|
||||
if (!backend) {
|
||||
LLAMA_LOG_ERROR("%s: failed to initialize Kompute backend\n", __func__);
|
||||
llama_free(ctx);
|
||||
return nullptr;
|
||||
}
|
||||
ctx->backends.push_back(backend);
|
||||
}
|
||||
free(devices);
|
||||
} else {
|
||||
LLAMA_LOG_ERROR("%s: Failed to init Kompute backend: split mode %d not supported\n", __func__, model->split_mode);
|
||||
llama_free(ctx);
|
||||
return nullptr;
|
||||
}
|
||||
ctx->backends.push_back(backend);
|
||||
}
|
||||
#elif defined(GGML_USE_CANN)
|
||||
// with split_mode LLAMA_SPLIT_MODE_NONE or LLAMA_SPLIT_MODE_ROW, only the main GPU backend is used
|
||||
|
Loading…
Reference in New Issue
Block a user