Compare commits

...

32 Commits

Author SHA1 Message Date
hipudding 2a0dc97e56 CANN: Fix failed test cases (#12708)
* CANN: Fix memory waste in aclnn_tensor

* CANN: fix backend ops fail

* CANN: fix acl_tensor memory alloc.

* CANN: format

* CANN: remove trailing whitespace
2025-04-03 08:49:51 +08:00
lhez 97a20c012b opencl: use max_alloc_size in backend ctx instead of querying again (#12705) 2025-04-02 17:01:42 -07:00
Jeff Bolz f01bd02376 vulkan: Implement split_k for coopmat2 flash attention. (#12627)
When using group query attention, we have one workgroup per KV batch and this
can be very few workgroups (e.g. just 8 in some models). Enable split_k to
spread the work across SMs. This helps a lot when the KV cache is large.
2025-04-02 14:25:08 -05:00
bandoti 6f3bd38640 cmake: remove caching from vulkan coopmat checks (#12719) 2025-04-02 14:56:26 -03:00
Jeff Bolz be0a0f8cae vulkan: Implement grouped query attention in the coopmat2 FA shader (#12559)
When adjacent batches of Q share the same batches of K/V, batch them into
the same workgroup. For example, when:

dst(128,32,1,1) = FA(q(128,1,32,1), k(128,16640,8,1), v(128,16640,8,1))

previously we would run 32 workgroups computing 1 result each, now we will
run 8 workgroups computing 4 results each.

This doesn't directly translate to better performance (at least when you have
>=32 SMs), but in a subsequent change I'll enable split_k which will scale much
better with 4x fewer workgroups.
2025-04-02 19:40:32 +02:00
0cc4m 92e3006bb6 Vulkan: Fix mmq int dot float cache size (#12722) 2025-04-02 19:12:30 +02:00
Georgi Gerganov 833e2b7409 model : print tensor size during load (#12711)
* model : print tensor size during load

* cont : fix units MB -> MiB

Co-authored-by: Diego Devesa <slarengh@gmail.com>

---------

Co-authored-by: Diego Devesa <slarengh@gmail.com>
2025-04-02 16:38:54 +03:00
Diego Devesa e0e912f49b llama : add option to override model tensor buffers (#11397)
* llama : add option to override tensor buffers

* ggml : fix possible underflow in ggml_nbytes
2025-04-02 14:52:01 +02:00
Georgi Gerganov a10b36c91a llama : refactor kv cache guard (#12695)
* llama : refactor kv cache guard

ggml-ci

* cont : fix comment [no ci]

* llama : fix kv_cache restore logic

ggml-ci

* context : simplify kv cache updates

ggml-ci

* cont : better name [no ci]

* llama : fix llama_decode return code when could not find KV slot

ggml-ci

* context : change log err -> warn [no ci]

* kv-cache : add comment + warning
2025-04-02 14:32:59 +03:00
Sigbjørn Skjæret 83a88bd6af vocab : BailingMoE : change possessive quantifiers to greedy (#12677) 2025-04-02 11:21:48 +02:00
Xuan-Son Nguyen 42eb248f46 common : remove json.hpp from common.cpp (#12697)
* common : remove json.hpp from common.cpp

* fix comment
2025-04-02 09:58:34 +02:00
Chenguang Li 9bacd6b374 [CANN] get_rows and dup optimization (#12671)
* [CANN]get_rows and dup optimization.

Co-authored-by: hipudding <huafengchun@gmail.com>
Signed-off-by: noemotiovon <noemotiovon@gmail.com>

* [CANN]GET_ROWS and CPY/DUP optimization

Co-authored-by: hipudding <huafengchun@gmail.com>
Signed-off-by: noemotiovon <noemotiovon@gmail.com>

* [CANN]code style adjustment

Signed-off-by: noemotiovon <noemotiovon@gmail.com>

* [CANN]code style adjustment

Signed-off-by: noemotiovon <noemotiovon@gmail.com>

* [CANN]code style adjustment

Signed-off-by: noemotiovon <noemotiovon@gmail.com>

* [CANN]code style adjustment

Signed-off-by: noemotiovon <noemotiovon@gmail.com>

---------

Signed-off-by: noemotiovon <noemotiovon@gmail.com>
Co-authored-by: noemotiovon <noemotiovon@gmail.com>
Co-authored-by: hipudding <huafengchun@gmail.com>
2025-04-02 15:22:13 +08:00
Xuan-Son Nguyen 267c1399f1 common : refactor downloading system, handle mmproj with -hf option (#12694)
* (wip) refactor downloading system [no ci]

* fix all examples

* fix mmproj with -hf

* gemma3: update readme

* only handle mmproj in llava example

* fix multi-shard download

* windows: fix problem with std::min and std::max

* fix 2
2025-04-01 23:44:05 +02:00
Junil Kim f423981ac8 opencl : fix memory allocation size (#12649)
issue:
https://github.com/CodeLinaro/llama.cpp/pull/17#issuecomment-2760611283

This patch fixes the memory allocation size
not exceeding the maximum size of the OpenCL device.
2025-04-01 09:54:34 -07:00
jklincn e39e727e9a llama : use LLM_KV_GENERAL_FILE_TYPE instead of gguf_find_key (#12672) 2025-04-01 14:54:28 +02:00
Sigbjørn Skjæret 5936a616e4 convert : BailingMoE : fix qkv split when head_dim is 0 (#12687)
NOTE: Ling-lite-base is broken, see https://huggingface.co/inclusionAI/Ling-lite-base/discussions/2
2025-04-01 14:37:13 +02:00
Georgi Gerganov 3fd072a540 metal : use F32 prec in FA kernels (#12688)
* metal : use F32 prec in FA kernels

ggml-ci

* cont : fix FA vec kernel

ggml-ci
2025-04-01 14:57:19 +03:00
R0CKSTAR a6f32f0b34 Fix clang warning in gguf_check_reserved_keys (#12686)
* Fix clang warning in gguf_check_reserved_keys

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>

* Fix typo

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>

---------

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
2025-04-01 13:12:53 +02:00
Wagner Bruna 2bb3597e42 vulkan: fix build when glslc doesn't support coopmat (#12683) 2025-04-01 11:38:07 +02:00
Romain Biessy 8293970542 SYCL: Rename oneMKL to oneMath (#12192)
* Rename oneMKL Interface to oneMath

* Use oneMath for Intel vendor

* Rename occurences to mkl

* clang-format

* Silence verbose warnings

* Set oneMath HIP_TARGETS

* Fix silence warnings

* Remove step to build oneMath from build instructions

* Use fixed oneMath version

* Remove INTEL_CPU

* Fold CMake oneDNN conditions

* Use Intel oneMKL for Intel devices

* Improve CMake message

* Link against MKL::MKL_SYCL::BLAS only

* Move oneMath documentation to Nvidia and AMD sections
2025-04-01 16:24:29 +08:00
Akarshan Biswas 8bbf26083d SYCL: switch to SYCL namespace (#12674) 2025-04-01 10:11:39 +02:00
Sigbjørn Skjæret 35782aeedb convert : BailingMoE : avoid setting rope_dim to 0 (#12678) 2025-03-31 23:09:48 +02:00
Daniel Bevenius c80a7759da vocab : add special infill tokens for CodeLlama (#11850)
* vocab : add special infill tokens for CodeLlama

The commit adds the following special tokens for CodeLlama infill:
- `▁<PRE>`
- `▁<SUF>`
- `▁<MID>`

The motivation for this is that currently the infill example uses
CodeLlama as a suggested model. But when using this model the following
error is generated:
```console
/llama.cpp-debug/examples/infill/infill.cpp:165: GGML_ASSERT(llama_vocab_fim_pre(vocab) >= 0) failed

Could not attach to process.  If your uid matches the uid of the target
process, check the setting of /proc/sys/kernel/yama/ptrace_scope, or try
again as the root user.  For more details, see /etc/sysctl.d/10-ptrace.conf
ptrace: Operation not permitted.
No stack.
The program is not being run.
305251 Aborted                 (core dumped)
./build/bin/llama-infill -t 10 -ngl 0 -m models/codellama-13b.Q5_K_S.gguf \
  -c 4096 --temp 0.7 --repeat_penalty 1.1 -n 20 \
  --in-prefix "def helloworld():\n    print(\"hell" \
  --in-suffix "\n   print(\"goodbye world\")\n    "
```

* squash! vocab : add special infill tokens for CodeLlama

Add _<EOT> as well.
2025-03-31 18:40:56 +02:00
a3sh 250d7953e8 ggml : faster ssm scan (#10558)
* faster ssm_scan

* delete unused commnet

* clang format

* add space

* modify unnecessary calculations

* faster ssm conv implementatioin

* modify file name with dash
2025-03-31 18:05:13 +02:00
Sigbjørn Skjæret 403fbacbbc convert : Qwerky : use lora_rank_tokenshift and lora_rank_decay if present (#12667) 2025-03-31 16:36:25 +02:00
0cc4m a8a1f33567 Vulkan: Add DP4A MMQ and Q8_1 quantization shader (#12135)
* Vulkan: Add DP4A MMQ and Q8_1 quantization shader

* Add q4_0 x q8_1 matrix matrix multiplication support

* Vulkan: Add int8 coopmat MMQ support

* Vulkan: Add q4_1, q5_0 and q5_1 quants, improve integer dot code

* Add GL_EXT_integer_dot_product check

* Remove ggml changes, fix mmq pipeline picker

* Remove ggml changes, restore Intel coopmat behaviour

* Fix glsl compile attempt when integer vec dot is not supported

* Remove redundant code, use non-saturating integer dot, enable all matmul sizes for mmq

* Remove redundant comment

* Fix integer dot check

* Fix compile issue with unsupported int dot glslc

* Update Windows build Vulkan SDK version
2025-03-31 14:37:01 +02:00
Georgi Gerganov 1790e73157 cmake : fix whitespace (#0) 2025-03-31 15:07:32 +03:00
Georgi Gerganov 0114a32da0 sync : ggml
ggml-ci
2025-03-31 15:07:32 +03:00
Sandro Hanea a7724480fd cmake: improve Vulkan cooperative matrix support checks (whisper/2966)
Co-authored-by: Sandro Hanea <me@sandro.rocks>
2025-03-31 15:07:32 +03:00
Sigbjørn Skjæret 1a85949067 llava : proper description fix (#12668) 2025-03-31 11:28:30 +02:00
Akarshan Biswas 6c02a032fa SYCL: Remove misleading ggml_sycl_op_flatten function (#12387)
* SYCL: Remove misleading ggml_sycl_op_flatten function

* remove trailing whitespace

* Fix L2 norm from rebase

* remove try catch block from element_wise.cpp

* remove comment from common.hp

* ggml-sycl.cpp: Add try catch sycl::exception block in compute_forward

* norm.cpp: remove try catch exception block
2025-03-31 11:25:24 +02:00
Sigbjørn Skjæret f52d59d771 llava : fix clip loading GGUFs with missing description (#12660) 2025-03-31 11:07:07 +02:00
92 changed files with 3609 additions and 3872 deletions
+1 -1
View File
@@ -803,7 +803,7 @@ jobs:
env:
OPENBLAS_VERSION: 0.3.23
SDE_VERSION: 9.33.0-2024-01-07
VULKAN_VERSION: 1.4.304.1
VULKAN_VERSION: 1.4.309.0
strategy:
matrix:
+644 -71
View File
@@ -1,9 +1,20 @@
#include "gguf.h" // for reading GGUF splits
#include "arg.h"
#include "common.h"
#include "log.h"
#include "sampling.h"
#include "chat.h"
// fix problem with std::min and std::max
#if defined(_WIN32)
#define WIN32_LEAN_AND_MEAN
#ifndef NOMINMAX
# define NOMINMAX
#endif
#include <windows.h>
#endif
#include <algorithm>
#include <climits>
#include <cstdarg>
@@ -14,6 +25,14 @@
#include <thread>
#include <vector>
//#define LLAMA_USE_CURL
#if defined(LLAMA_USE_CURL)
#include <curl/curl.h>
#include <curl/easy.h>
#include <future>
#endif
#include "json-schema-to-grammar.h"
using json = nlohmann::ordered_json;
@@ -125,47 +144,549 @@ std::string common_arg::to_string() {
return ss.str();
}
//
// downloader
//
struct common_hf_file_res {
std::string repo; // repo name with ":tag" removed
std::string ggufFile;
std::string mmprojFile;
};
#ifdef LLAMA_USE_CURL
#ifdef __linux__
#include <linux/limits.h>
#elif defined(_WIN32)
# if !defined(PATH_MAX)
# define PATH_MAX MAX_PATH
# endif
#else
#include <sys/syslimits.h>
#endif
#define LLAMA_CURL_MAX_URL_LENGTH 2084 // Maximum URL Length in Chrome: 2083
//
// CURL utils
//
using curl_ptr = std::unique_ptr<CURL, decltype(&curl_easy_cleanup)>;
// cannot use unique_ptr for curl_slist, because we cannot update without destroying the old one
struct curl_slist_ptr {
struct curl_slist * ptr = nullptr;
~curl_slist_ptr() {
if (ptr) {
curl_slist_free_all(ptr);
}
}
};
#define CURL_MAX_RETRY 3
#define CURL_RETRY_DELAY_SECONDS 2
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) {
LOG_INF("%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;
LOG_WRN("%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));
}
LOG_ERR("%s: curl_easy_perform() failed after %d attempts\n", __func__, max_attempts);
return false;
}
// download one single file from remote URL to local path
static bool common_download_file_single(const std::string & url, const std::string & path, const std::string & bearer_token) {
// Initialize libcurl
curl_ptr curl(curl_easy_init(), &curl_easy_cleanup);
curl_slist_ptr http_headers;
if (!curl) {
LOG_ERR("%s: error initializing libcurl\n", __func__);
return false;
}
bool force_download = false;
// Set the URL, allow to follow http redirection
curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
curl_easy_setopt(curl.get(), CURLOPT_FOLLOWLOCATION, 1L);
// Check if hf-token or bearer-token was specified
if (!bearer_token.empty()) {
std::string auth_header = "Authorization: Bearer " + bearer_token;
http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str());
curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
}
#if defined(_WIN32)
// CURLSSLOPT_NATIVE_CA tells libcurl to use standard certificate store of
// operating system. Currently implemented under MS-Windows.
curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA);
#endif
// Check if the file already exists locally
auto file_exists = std::filesystem::exists(path);
// If the file exists, check its JSON metadata companion file.
std::string metadata_path = path + ".json";
nlohmann::json metadata;
std::string etag;
std::string last_modified;
if (file_exists) {
// Try and read the JSON metadata file (note: stream autoclosed upon exiting this block).
std::ifstream metadata_in(metadata_path);
if (metadata_in.good()) {
try {
metadata_in >> metadata;
LOG_INF("%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(), metadata.dump().c_str());
if (metadata.contains("url") && metadata.at("url").is_string()) {
auto previous_url = metadata.at("url").get<std::string>();
if (previous_url != url) {
LOG_ERR("%s: Model URL mismatch: %s != %s\n", __func__, url.c_str(), previous_url.c_str());
return false;
}
}
if (metadata.contains("etag") && metadata.at("etag").is_string()) {
etag = metadata.at("etag");
}
if (metadata.contains("lastModified") && metadata.at("lastModified").is_string()) {
last_modified = metadata.at("lastModified");
}
} catch (const nlohmann::json::exception & e) {
LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what());
return false;
}
}
} else {
LOG_INF("%s: no previous model file found %s\n", __func__, path.c_str());
}
// Send a HEAD request to retrieve the etag and last-modified headers
struct common_load_model_from_url_headers {
std::string etag;
std::string last_modified;
};
common_load_model_from_url_headers headers;
{
typedef size_t(*CURLOPT_HEADERFUNCTION_PTR)(char *, size_t, size_t, void *);
auto header_callback = [](char * buffer, size_t /*size*/, size_t n_items, void * userdata) -> size_t {
common_load_model_from_url_headers * headers = (common_load_model_from_url_headers *) userdata;
static std::regex header_regex("([^:]+): (.*)\r\n");
static std::regex etag_regex("ETag", std::regex_constants::icase);
static std::regex last_modified_regex("Last-Modified", std::regex_constants::icase);
std::string header(buffer, n_items);
std::smatch match;
if (std::regex_match(header, match, header_regex)) {
const std::string & key = match[1];
const std::string & value = match[2];
if (std::regex_match(key, match, etag_regex)) {
headers->etag = value;
} else if (std::regex_match(key, match, last_modified_regex)) {
headers->last_modified = value;
}
}
return n_items;
};
curl_easy_setopt(curl.get(), CURLOPT_NOBODY, 1L); // will trigger the HEAD verb
curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L); // hide head request progress
curl_easy_setopt(curl.get(), CURLOPT_HEADERFUNCTION, static_cast<CURLOPT_HEADERFUNCTION_PTR>(header_callback));
curl_easy_setopt(curl.get(), CURLOPT_HEADERDATA, &headers);
bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS);
if (!was_perform_successful) {
return false;
}
long http_code = 0;
curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &http_code);
if (http_code != 200) {
// HEAD not supported, we don't know if the file has changed
// force trigger downloading
force_download = true;
LOG_ERR("%s: HEAD invalid http status code received: %ld\n", __func__, http_code);
}
}
bool should_download = !file_exists || force_download;
if (!should_download) {
if (!etag.empty() && etag != headers.etag) {
LOG_WRN("%s: ETag header is different (%s != %s): triggering a new download\n", __func__, etag.c_str(), headers.etag.c_str());
should_download = true;
} else if (!last_modified.empty() && last_modified != headers.last_modified) {
LOG_WRN("%s: Last-Modified header is different (%s != %s): triggering a new download\n", __func__, last_modified.c_str(), headers.last_modified.c_str());
should_download = true;
}
}
if (should_download) {
std::string path_temporary = path + ".downloadInProgress";
if (file_exists) {
LOG_WRN("%s: deleting previous downloaded file: %s\n", __func__, path.c_str());
if (remove(path.c_str()) != 0) {
LOG_ERR("%s: unable to delete file: %s\n", __func__, path.c_str());
return false;
}
}
// Set the output file
struct FILE_deleter {
void operator()(FILE * f) const {
fclose(f);
}
};
std::unique_ptr<FILE, FILE_deleter> outfile(fopen(path_temporary.c_str(), "wb"));
if (!outfile) {
LOG_ERR("%s: error opening local file for writing: %s\n", __func__, path.c_str());
return false;
}
typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * data, size_t size, size_t nmemb, void * fd);
auto write_callback = [](void * data, size_t size, size_t nmemb, void * fd) -> size_t {
return fwrite(data, size, nmemb, (FILE *)fd);
};
curl_easy_setopt(curl.get(), CURLOPT_NOBODY, 0L);
curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast<CURLOPT_WRITEFUNCTION_PTR>(write_callback));
curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, outfile.get());
// display download progress
curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 0L);
// helper function to hide password in URL
auto llama_download_hide_password_in_url = [](const std::string & url) -> std::string {
std::size_t protocol_pos = url.find("://");
if (protocol_pos == std::string::npos) {
return url; // Malformed URL
}
std::size_t at_pos = url.find('@', protocol_pos + 3);
if (at_pos == std::string::npos) {
return url; // No password in URL
}
return url.substr(0, protocol_pos + 3) + "********" + url.substr(at_pos);
};
// start the download
LOG_INF("%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;
}
long http_code = 0;
curl_easy_getinfo (curl.get(), CURLINFO_RESPONSE_CODE, &http_code);
if (http_code < 200 || http_code >= 400) {
LOG_ERR("%s: invalid http status code received: %ld\n", __func__, http_code);
return false;
}
// Causes file to be closed explicitly here before we rename it.
outfile.reset();
// Write the updated JSON metadata file.
metadata.update({
{"url", url},
{"etag", headers.etag},
{"lastModified", headers.last_modified}
});
std::ofstream(metadata_path) << metadata.dump(4);
LOG_INF("%s: file metadata saved: %s\n", __func__, metadata_path.c_str());
if (rename(path_temporary.c_str(), path.c_str()) != 0) {
LOG_ERR("%s: unable to rename file: %s to %s\n", __func__, path_temporary.c_str(), path.c_str());
return false;
}
}
return true;
}
// download multiple files from remote URLs to local paths
// the input is a vector of pairs <url, path>
static bool common_download_file_multiple(const std::vector<std::pair<std::string, std::string>> & urls, const std::string & bearer_token) {
// Prepare download in parallel
std::vector<std::future<bool>> futures_download;
for (auto const & item : urls) {
futures_download.push_back(std::async(std::launch::async, [bearer_token](const std::pair<std::string, std::string> & it) -> bool {
return common_download_file_single(it.first, it.second, bearer_token);
}, item));
}
// Wait for all downloads to complete
for (auto & f : futures_download) {
if (!f.get()) {
return false;
}
}
return true;
}
static bool common_download_model(
const common_params_model & model,
const std::string & bearer_token) {
// Basic validation of the model.url
if (model.url.empty()) {
LOG_ERR("%s: invalid model url\n", __func__);
return false;
}
if (!common_download_file_single(model.url, model.path, bearer_token)) {
return false;
}
// check for additional GGUFs split to download
int n_split = 0;
{
struct gguf_init_params gguf_params = {
/*.no_alloc = */ true,
/*.ctx = */ NULL,
};
auto * ctx_gguf = gguf_init_from_file(model.path.c_str(), gguf_params);
if (!ctx_gguf) {
LOG_ERR("\n%s: failed to load input GGUF from %s\n", __func__, model.path.c_str());
return false;
}
auto key_n_split = gguf_find_key(ctx_gguf, LLM_KV_SPLIT_COUNT);
if (key_n_split >= 0) {
n_split = gguf_get_val_u16(ctx_gguf, key_n_split);
}
gguf_free(ctx_gguf);
}
if (n_split > 1) {
char split_prefix[PATH_MAX] = {0};
char split_url_prefix[LLAMA_CURL_MAX_URL_LENGTH] = {0};
// Verify the first split file format
// and extract split URL and PATH prefixes
{
if (!llama_split_prefix(split_prefix, sizeof(split_prefix), model.path.c_str(), 0, n_split)) {
LOG_ERR("\n%s: unexpected model file name: %s n_split=%d\n", __func__, model.path.c_str(), n_split);
return false;
}
if (!llama_split_prefix(split_url_prefix, sizeof(split_url_prefix), model.url.c_str(), 0, n_split)) {
LOG_ERR("\n%s: unexpected model url: %s n_split=%d\n", __func__, model.url.c_str(), n_split);
return false;
}
}
std::vector<std::pair<std::string, std::string>> urls;
for (int idx = 1; idx < n_split; idx++) {
char split_path[PATH_MAX] = {0};
llama_split_path(split_path, sizeof(split_path), split_prefix, idx, n_split);
char split_url[LLAMA_CURL_MAX_URL_LENGTH] = {0};
llama_split_path(split_url, sizeof(split_url), split_url_prefix, idx, n_split);
if (std::string(split_path) == model.path) {
continue; // skip the already downloaded file
}
urls.push_back({split_url, split_path});
}
// Download in parallel
common_download_file_multiple(urls, bearer_token);
}
return true;
}
/**
* Allow getting the HF file from the HF repo with tag (like ollama), for example:
* - bartowski/Llama-3.2-3B-Instruct-GGUF:q4
* - bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M
* - bartowski/Llama-3.2-3B-Instruct-GGUF:q5_k_s
* Tag is optional, default to "latest" (meaning it checks for Q4_K_M first, then Q4, then if not found, return the first GGUF file in repo)
*
* Return pair of <repo, file> (with "repo" already having tag removed)
*
* Note: we use the Ollama-compatible HF API, but not using the blobId. Instead, we use the special "ggufFile" field which returns the value for "hf_file". This is done to be backward-compatible with existing cache files.
*/
static struct common_hf_file_res common_get_hf_file(const std::string & hf_repo_with_tag, const std::string & bearer_token) {
auto parts = string_split<std::string>(hf_repo_with_tag, ':');
std::string tag = parts.size() > 1 ? parts.back() : "latest";
std::string hf_repo = parts[0];
if (string_split<std::string>(hf_repo, '/').size() != 2) {
throw std::invalid_argument("error: invalid HF repo format, expected <user>/<model>[:quant]\n");
}
// fetch model info from Hugging Face Hub API
curl_ptr curl(curl_easy_init(), &curl_easy_cleanup);
curl_slist_ptr http_headers;
std::string res_str;
std::string url = "https://huggingface.co/v2/" + hf_repo + "/manifests/" + tag;
curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L);
typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * ptr, size_t size, size_t nmemb, void * data);
auto write_callback = [](void * ptr, size_t size, size_t nmemb, void * data) -> size_t {
static_cast<std::string *>(data)->append((char * ) ptr, size * nmemb);
return size * nmemb;
};
curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast<CURLOPT_WRITEFUNCTION_PTR>(write_callback));
curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, &res_str);
#if defined(_WIN32)
curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA);
#endif
if (!bearer_token.empty()) {
std::string auth_header = "Authorization: Bearer " + bearer_token;
http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str());
}
// Important: the User-Agent must be "llama-cpp" to get the "ggufFile" field in the response
http_headers.ptr = curl_slist_append(http_headers.ptr, "User-Agent: llama-cpp");
http_headers.ptr = curl_slist_append(http_headers.ptr, "Accept: application/json");
curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
CURLcode res = curl_easy_perform(curl.get());
if (res != CURLE_OK) {
throw std::runtime_error("error: cannot make GET request to HF API");
}
long res_code;
std::string ggufFile = "";
std::string mmprojFile = "";
curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &res_code);
if (res_code == 200) {
// extract ggufFile.rfilename in json, using regex
{
std::regex pattern("\"ggufFile\"[\\s\\S]*?\"rfilename\"\\s*:\\s*\"([^\"]+)\"");
std::smatch match;
if (std::regex_search(res_str, match, pattern)) {
ggufFile = match[1].str();
}
}
// extract mmprojFile.rfilename in json, using regex
{
std::regex pattern("\"mmprojFile\"[\\s\\S]*?\"rfilename\"\\s*:\\s*\"([^\"]+)\"");
std::smatch match;
if (std::regex_search(res_str, match, pattern)) {
mmprojFile = match[1].str();
}
}
} else if (res_code == 401) {
throw std::runtime_error("error: model is private or does not exist; if you are accessing a gated model, please provide a valid HF token");
} else {
throw std::runtime_error(string_format("error from HF API, response code: %ld, data: %s", res_code, res_str.c_str()));
}
// check response
if (ggufFile.empty()) {
throw std::runtime_error("error: model does not have ggufFile");
}
return { hf_repo, ggufFile, mmprojFile };
}
#else
static bool common_download_file_single(const std::string &, const std::string &, const std::string &) {
LOG_ERR("error: built without CURL, cannot download model from internet\n");
return false;
}
static bool common_download_file_multiple(const std::vector<std::pair<std::string, std::string>> &, const std::string &) {
LOG_ERR("error: built without CURL, cannot download model from the internet\n");
return false;
}
static bool common_download_model(
const common_params_model &,
const std::string &) {
LOG_ERR("error: built without CURL, cannot download model from the internet\n");
return false;
}
static struct common_hf_file_res common_get_hf_file(const std::string &, const std::string &) {
LOG_ERR("error: built without CURL, cannot download model from the internet\n");
return {};
}
#endif // LLAMA_USE_CURL
//
// utils
//
static void common_params_handle_model_default(
std::string & model,
const std::string & model_url,
std::string & hf_repo,
std::string & hf_file,
const std::string & hf_token,
const std::string & model_default) {
if (!hf_repo.empty()) {
// short-hand to avoid specifying --hf-file -> default it to --model
if (hf_file.empty()) {
if (model.empty()) {
auto auto_detected = common_get_hf_file(hf_repo, hf_token);
if (auto_detected.first.empty() || auto_detected.second.empty()) {
exit(1); // built without CURL, error message already printed
static void common_params_handle_model(
struct common_params_model & model,
const std::string & bearer_token,
const std::string & model_path_default,
bool is_mmproj = false) { // TODO: move is_mmproj to an enum when we have more files?
// handle pre-fill default model path and url based on hf_repo and hf_file
{
if (!model.hf_repo.empty()) {
// short-hand to avoid specifying --hf-file -> default it to --model
if (model.hf_file.empty()) {
if (model.path.empty()) {
auto auto_detected = common_get_hf_file(model.hf_repo, bearer_token);
if (auto_detected.repo.empty() || auto_detected.ggufFile.empty()) {
exit(1); // built without CURL, error message already printed
}
model.hf_repo = auto_detected.repo;
model.hf_file = is_mmproj ? auto_detected.mmprojFile : auto_detected.ggufFile;
} else {
model.hf_file = model.path;
}
hf_repo = auto_detected.first;
hf_file = auto_detected.second;
} else {
hf_file = model;
}
// TODO: allow custom host
model.url = "https://huggingface.co/" + model.hf_repo + "/resolve/main/" + model.hf_file;
// make sure model path is present (for caching purposes)
if (model.path.empty()) {
// this is to avoid different repo having same file name, or same file name in different subdirs
std::string filename = model.hf_repo + "_" + model.hf_file;
// to make sure we don't have any slashes in the filename
string_replace_all(filename, "/", "_");
model.path = fs_get_cache_file(filename);
}
} else if (!model.url.empty()) {
if (model.path.empty()) {
auto f = string_split<std::string>(model.url, '#').front();
f = string_split<std::string>(f, '?').front();
model.path = fs_get_cache_file(string_split<std::string>(f, '/').back());
}
} else if (model.path.empty()) {
model.path = model_path_default;
}
// make sure model path is present (for caching purposes)
if (model.empty()) {
// this is to avoid different repo having same file name, or same file name in different subdirs
std::string filename = hf_repo + "_" + hf_file;
// to make sure we don't have any slashes in the filename
string_replace_all(filename, "/", "_");
model = fs_get_cache_file(filename);
}
// then, download it if needed
if (!model.url.empty()) {
bool ok = common_download_model(model, bearer_token);
if (!ok) {
LOG_ERR("error: failed to download model from %s\n", model.url.c_str());
exit(1);
}
} else if (!model_url.empty()) {
if (model.empty()) {
auto f = string_split<std::string>(model_url, '#').front();
f = string_split<std::string>(f, '?').front();
model = fs_get_cache_file(string_split<std::string>(f, '/').back());
}
} else if (model.empty()) {
model = model_default;
}
}
@@ -300,10 +821,16 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n");
}
// TODO: refactor model params in a common struct
common_params_handle_model_default(params.model, params.model_url, params.hf_repo, params.hf_file, params.hf_token, DEFAULT_MODEL_PATH);
common_params_handle_model_default(params.speculative.model, params.speculative.model_url, params.speculative.hf_repo, params.speculative.hf_file, params.hf_token, "");
common_params_handle_model_default(params.vocoder.model, params.vocoder.model_url, params.vocoder.hf_repo, params.vocoder.hf_file, params.hf_token, "");
common_params_handle_model(params.model, params.hf_token, DEFAULT_MODEL_PATH);
common_params_handle_model(params.speculative.model, params.hf_token, "");
common_params_handle_model(params.vocoder.model, params.hf_token, "");
// allow --mmproj to be set from -hf
// assuming that mmproj is always in the same repo as text model
if (!params.model.hf_repo.empty() && ctx_arg.ex == LLAMA_EXAMPLE_LLAVA) {
params.mmproj.hf_repo = params.model.hf_repo;
}
common_params_handle_model(params.mmproj, params.hf_token, "", true);
if (params.escape) {
string_process_escapes(params.prompt);
@@ -322,6 +849,10 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
params.kv_overrides.back().key[0] = 0;
}
if (!params.tensor_buft_overrides.empty()) {
params.tensor_buft_overrides.push_back({nullptr, nullptr});
}
if (params.reranking && params.embedding) {
throw std::invalid_argument("error: either --embedding or --reranking can be specified, but not both");
}
@@ -1561,7 +2092,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
{"--mmproj"}, "FILE",
"path to a multimodal projector file for LLaVA. see examples/llava/README.md",
[](common_params & params, const std::string & value) {
params.mmproj = value;
params.mmproj.path = value;
}
).set_examples({LLAMA_EXAMPLE_LLAVA}));
add_opt(common_arg(
{"--mmproj-url"}, "URL",
"URL to a multimodal projector file for LLaVA. see examples/llava/README.md",
[](common_params & params, const std::string & value) {
params.mmproj.url = value;
}
).set_examples({LLAMA_EXAMPLE_LLAVA}));
add_opt(common_arg(
@@ -1647,6 +2185,41 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
exit(0);
}
));
add_opt(common_arg(
{"--override-tensor", "-ot"}, "<tensor name pattern>=<buffer type>,...",
"override tensor buffer type", [](common_params & params, const std::string & value) {
/* static */ std::map<std::string, ggml_backend_buffer_type_t> buft_list;
if (buft_list.empty()) {
// enumerate all the devices and add their buffer types to the list
for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
auto * dev = ggml_backend_dev_get(i);
auto * buft = ggml_backend_dev_buffer_type(dev);
if (buft) {
buft_list[ggml_backend_buft_name(buft)] = buft;
}
}
}
for (const auto & override : string_split<std::string>(value, ',')) {
std::string::size_type pos = override.find('=');
if (pos == std::string::npos) {
throw std::invalid_argument("invalid value");
}
std::string tensor_name = override.substr(0, pos);
std::string buffer_type = override.substr(pos + 1);
if (buft_list.find(buffer_type) == buft_list.end()) {
printf("Available buffer types:\n");
for (const auto & it : buft_list) {
printf(" %s\n", ggml_backend_buft_name(it.second));
}
throw std::invalid_argument("unknown buffer type");
}
// FIXME: this leaks memory
params.tensor_buft_overrides.push_back({strdup(tensor_name.c_str()), buft_list.at(buffer_type)});
}
}
));
add_opt(common_arg(
{"-ngl", "--gpu-layers", "--n-gpu-layers"}, "N",
"number of layers to store in VRAM",
@@ -1790,14 +2363,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
"or `--model-url` if set, otherwise %s)", DEFAULT_MODEL_PATH
),
[](common_params & params, const std::string & value) {
params.model = value;
params.model.path = value;
}
).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_EXPORT_LORA}).set_env("LLAMA_ARG_MODEL"));
add_opt(common_arg(
{"-mu", "--model-url"}, "MODEL_URL",
"model download url (default: unused)",
[](common_params & params, const std::string & value) {
params.model_url = value;
params.model.url = value;
}
).set_env("LLAMA_ARG_MODEL_URL"));
add_opt(common_arg(
@@ -1806,35 +2379,35 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
"example: unsloth/phi-4-GGUF:q4_k_m\n"
"(default: unused)",
[](common_params & params, const std::string & value) {
params.hf_repo = value;
params.model.hf_repo = value;
}
).set_env("LLAMA_ARG_HF_REPO"));
add_opt(common_arg(
{"-hfd", "-hfrd", "--hf-repo-draft"}, "<user>/<model>[:quant]",
"Same as --hf-repo, but for the draft model (default: unused)",
[](common_params & params, const std::string & value) {
params.speculative.hf_repo = value;
params.speculative.model.hf_repo = value;
}
).set_env("LLAMA_ARG_HFD_REPO"));
add_opt(common_arg(
{"-hff", "--hf-file"}, "FILE",
"Hugging Face model file. If specified, it will override the quant in --hf-repo (default: unused)",
[](common_params & params, const std::string & value) {
params.hf_file = value;
params.model.hf_file = value;
}
).set_env("LLAMA_ARG_HF_FILE"));
add_opt(common_arg(
{"-hfv", "-hfrv", "--hf-repo-v"}, "<user>/<model>[:quant]",
"Hugging Face model repository for the vocoder model (default: unused)",
[](common_params & params, const std::string & value) {
params.vocoder.hf_repo = value;
params.vocoder.model.hf_repo = value;
}
).set_env("LLAMA_ARG_HF_REPO_V"));
add_opt(common_arg(
{"-hffv", "--hf-file-v"}, "FILE",
"Hugging Face model file for the vocoder model (default: unused)",
[](common_params & params, const std::string & value) {
params.vocoder.hf_file = value;
params.vocoder.model.hf_file = value;
}
).set_env("LLAMA_ARG_HF_FILE_V"));
add_opt(common_arg(
@@ -2454,7 +3027,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
{"-md", "--model-draft"}, "FNAME",
"draft model for speculative decoding (default: unused)",
[](common_params & params, const std::string & value) {
params.speculative.model = value;
params.speculative.model.path = value;
}
).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MODEL_DRAFT"));
@@ -2462,7 +3035,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
{"-mv", "--model-vocoder"}, "FNAME",
"vocoder model for audio generation (default: unused)",
[](common_params & params, const std::string & value) {
params.vocoder.model = value;
params.vocoder.model.path = value;
}
).set_examples({LLAMA_EXAMPLE_TTS, LLAMA_EXAMPLE_SERVER}));
add_opt(common_arg(
@@ -2485,10 +3058,10 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
{"--tts-oute-default"},
string_format("use default OuteTTS models (note: can download weights from the internet)"),
[](common_params & params) {
params.hf_repo = "OuteAI/OuteTTS-0.2-500M-GGUF";
params.hf_file = "OuteTTS-0.2-500M-Q8_0.gguf";
params.vocoder.hf_repo = "ggml-org/WavTokenizer";
params.vocoder.hf_file = "WavTokenizer-Large-75-F16.gguf";
params.model.hf_repo = "OuteAI/OuteTTS-0.2-500M-GGUF";
params.model.hf_file = "OuteTTS-0.2-500M-Q8_0.gguf";
params.vocoder.model.hf_repo = "ggml-org/WavTokenizer";
params.vocoder.model.hf_file = "WavTokenizer-Large-75-F16.gguf";
}
).set_examples({LLAMA_EXAMPLE_TTS}));
@@ -2496,8 +3069,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
{"--embd-bge-small-en-default"},
string_format("use default bge-small-en-v1.5 model (note: can download weights from the internet)"),
[](common_params & params) {
params.hf_repo = "ggml-org/bge-small-en-v1.5-Q8_0-GGUF";
params.hf_file = "bge-small-en-v1.5-q8_0.gguf";
params.model.hf_repo = "ggml-org/bge-small-en-v1.5-Q8_0-GGUF";
params.model.hf_file = "bge-small-en-v1.5-q8_0.gguf";
params.pooling_type = LLAMA_POOLING_TYPE_NONE;
params.embd_normalize = 2;
params.n_ctx = 512;
@@ -2510,8 +3083,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
{"--embd-e5-small-en-default"},
string_format("use default e5-small-v2 model (note: can download weights from the internet)"),
[](common_params & params) {
params.hf_repo = "ggml-org/e5-small-v2-Q8_0-GGUF";
params.hf_file = "e5-small-v2-q8_0.gguf";
params.model.hf_repo = "ggml-org/e5-small-v2-Q8_0-GGUF";
params.model.hf_file = "e5-small-v2-q8_0.gguf";
params.pooling_type = LLAMA_POOLING_TYPE_NONE;
params.embd_normalize = 2;
params.n_ctx = 512;
@@ -2524,8 +3097,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
{"--embd-gte-small-default"},
string_format("use default gte-small model (note: can download weights from the internet)"),
[](common_params & params) {
params.hf_repo = "ggml-org/gte-small-Q8_0-GGUF";
params.hf_file = "gte-small-q8_0.gguf";
params.model.hf_repo = "ggml-org/gte-small-Q8_0-GGUF";
params.model.hf_file = "gte-small-q8_0.gguf";
params.pooling_type = LLAMA_POOLING_TYPE_NONE;
params.embd_normalize = 2;
params.n_ctx = 512;
@@ -2538,8 +3111,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
{"--fim-qwen-1.5b-default"},
string_format("use default Qwen 2.5 Coder 1.5B (note: can download weights from the internet)"),
[](common_params & params) {
params.hf_repo = "ggml-org/Qwen2.5-Coder-1.5B-Q8_0-GGUF";
params.hf_file = "qwen2.5-coder-1.5b-q8_0.gguf";
params.model.hf_repo = "ggml-org/Qwen2.5-Coder-1.5B-Q8_0-GGUF";
params.model.hf_file = "qwen2.5-coder-1.5b-q8_0.gguf";
params.port = 8012;
params.n_gpu_layers = 99;
params.flash_attn = true;
@@ -2554,8 +3127,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
{"--fim-qwen-3b-default"},
string_format("use default Qwen 2.5 Coder 3B (note: can download weights from the internet)"),
[](common_params & params) {
params.hf_repo = "ggml-org/Qwen2.5-Coder-3B-Q8_0-GGUF";
params.hf_file = "qwen2.5-coder-3b-q8_0.gguf";
params.model.hf_repo = "ggml-org/Qwen2.5-Coder-3B-Q8_0-GGUF";
params.model.hf_file = "qwen2.5-coder-3b-q8_0.gguf";
params.port = 8012;
params.n_gpu_layers = 99;
params.flash_attn = true;
@@ -2570,8 +3143,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
{"--fim-qwen-7b-default"},
string_format("use default Qwen 2.5 Coder 7B (note: can download weights from the internet)"),
[](common_params & params) {
params.hf_repo = "ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF";
params.hf_file = "qwen2.5-coder-7b-q8_0.gguf";
params.model.hf_repo = "ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF";
params.model.hf_file = "qwen2.5-coder-7b-q8_0.gguf";
params.port = 8012;
params.n_gpu_layers = 99;
params.flash_attn = true;
@@ -2586,10 +3159,10 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
{"--fim-qwen-7b-spec"},
string_format("use Qwen 2.5 Coder 7B + 0.5B draft for speculative decoding (note: can download weights from the internet)"),
[](common_params & params) {
params.hf_repo = "ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF";
params.hf_file = "qwen2.5-coder-7b-q8_0.gguf";
params.speculative.hf_repo = "ggml-org/Qwen2.5-Coder-0.5B-Q8_0-GGUF";
params.speculative.hf_file = "qwen2.5-coder-0.5b-q8_0.gguf";
params.model.hf_repo = "ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF";
params.model.hf_file = "qwen2.5-coder-7b-q8_0.gguf";
params.speculative.model.hf_repo = "ggml-org/Qwen2.5-Coder-0.5B-Q8_0-GGUF";
params.speculative.model.hf_file = "qwen2.5-coder-0.5b-q8_0.gguf";
params.speculative.n_gpu_layers = 99;
params.port = 8012;
params.n_gpu_layers = 99;
@@ -2605,10 +3178,10 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
{"--fim-qwen-14b-spec"},
string_format("use Qwen 2.5 Coder 14B + 0.5B draft for speculative decoding (note: can download weights from the internet)"),
[](common_params & params) {
params.hf_repo = "ggml-org/Qwen2.5-Coder-14B-Q8_0-GGUF";
params.hf_file = "qwen2.5-coder-14b-q8_0.gguf";
params.speculative.hf_repo = "ggml-org/Qwen2.5-Coder-0.5B-Q8_0-GGUF";
params.speculative.hf_file = "qwen2.5-coder-0.5b-q8_0.gguf";
params.model.hf_repo = "ggml-org/Qwen2.5-Coder-14B-Q8_0-GGUF";
params.model.hf_file = "qwen2.5-coder-14b-q8_0.gguf";
params.speculative.model.hf_repo = "ggml-org/Qwen2.5-Coder-0.5B-Q8_0-GGUF";
params.speculative.model.hf_file = "qwen2.5-coder-0.5b-q8_0.gguf";
params.speculative.n_gpu_layers = 99;
params.port = 8012;
params.n_gpu_layers = 99;
+14 -519
View File
@@ -7,9 +7,6 @@
#include "common.h"
#include "log.h"
// Change JSON_ASSERT from assert() to GGML_ASSERT:
#define JSON_ASSERT GGML_ASSERT
#include "json.hpp"
#include "llama.h"
#include <algorithm>
@@ -51,47 +48,11 @@
#include <sys/stat.h>
#include <unistd.h>
#endif
#if defined(LLAMA_USE_CURL)
#include <curl/curl.h>
#include <curl/easy.h>
#include <future>
#endif
#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
#if defined(LLAMA_USE_CURL)
#ifdef __linux__
#include <linux/limits.h>
#elif defined(_WIN32)
# if !defined(PATH_MAX)
# define PATH_MAX MAX_PATH
# endif
#else
#include <sys/syslimits.h>
#endif
#define LLAMA_CURL_MAX_URL_LENGTH 2084 // Maximum URL Length in Chrome: 2083
//
// CURL utils
//
using curl_ptr = std::unique_ptr<CURL, decltype(&curl_easy_cleanup)>;
// cannot use unique_ptr for curl_slist, because we cannot update without destroying the old one
struct curl_slist_ptr {
struct curl_slist * ptr = nullptr;
~curl_slist_ptr() {
if (ptr) {
curl_slist_free_all(ptr);
}
}
};
#endif // LLAMA_USE_CURL
using json = nlohmann::ordered_json;
//
// CPU utils
//
@@ -900,22 +861,14 @@ std::string fs_get_cache_file(const std::string & filename) {
//
// Model utils
//
struct common_init_result common_init_from_params(common_params & params) {
common_init_result iparams;
auto mparams = common_model_params_to_llama(params);
llama_model * model = nullptr;
if (!params.hf_repo.empty() && !params.hf_file.empty()) {
model = common_load_model_from_hf(params.hf_repo, params.hf_file, params.model, params.hf_token, mparams);
} else if (!params.model_url.empty()) {
model = common_load_model_from_url(params.model_url, params.model, params.hf_token, mparams);
} else {
model = llama_model_load_from_file(params.model.c_str(), mparams);
}
llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
if (model == NULL) {
LOG_ERR("%s: failed to load model '%s'\n", __func__, params.model.c_str());
LOG_ERR("%s: failed to load model '%s'\n", __func__, params.model.path.c_str());
return iparams;
}
@@ -950,7 +903,7 @@ struct common_init_result common_init_from_params(common_params & params) {
llama_context * lctx = llama_init_from_model(model, cparams);
if (lctx == NULL) {
LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.c_str());
LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.path.c_str());
llama_model_free(model);
return iparams;
}
@@ -1089,15 +1042,18 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
if (!params.devices.empty()) {
mparams.devices = params.devices.data();
}
if (params.n_gpu_layers != -1) {
mparams.n_gpu_layers = params.n_gpu_layers;
}
mparams.main_gpu = params.main_gpu;
mparams.split_mode = params.split_mode;
mparams.tensor_split = params.tensor_split;
mparams.use_mmap = params.use_mmap;
mparams.use_mlock = params.use_mlock;
mparams.check_tensors = params.check_tensors;
if (params.kv_overrides.empty()) {
mparams.kv_overrides = NULL;
} else {
@@ -1105,6 +1061,13 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
mparams.kv_overrides = params.kv_overrides.data();
}
if (params.tensor_buft_overrides.empty()) {
mparams.tensor_buft_overrides = NULL;
} else {
GGML_ASSERT(params.tensor_buft_overrides.back().pattern == nullptr && "Tensor buffer overrides not terminated with empty pattern");
mparams.tensor_buft_overrides = params.tensor_buft_overrides.data();
}
return mparams;
}
@@ -1164,451 +1127,6 @@ struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_p
return tpp;
}
#ifdef LLAMA_USE_CURL
#define CURL_MAX_RETRY 3
#define CURL_RETRY_DELAY_SECONDS 2
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) {
LOG_INF("%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;
LOG_WRN("%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));
}
LOG_ERR("%s: curl_easy_perform() failed after %d attempts\n", __func__, max_attempts);
return false;
}
static bool common_download_file(const std::string & url, const std::string & path, const std::string & hf_token) {
// Initialize libcurl
curl_ptr curl(curl_easy_init(), &curl_easy_cleanup);
curl_slist_ptr http_headers;
if (!curl) {
LOG_ERR("%s: error initializing libcurl\n", __func__);
return false;
}
bool force_download = false;
// Set the URL, allow to follow http redirection
curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
curl_easy_setopt(curl.get(), CURLOPT_FOLLOWLOCATION, 1L);
// Check if hf-token or bearer-token was specified
if (!hf_token.empty()) {
std::string auth_header = "Authorization: Bearer " + hf_token;
http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str());
curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
}
#if defined(_WIN32)
// CURLSSLOPT_NATIVE_CA tells libcurl to use standard certificate store of
// operating system. Currently implemented under MS-Windows.
curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA);
#endif
// Check if the file already exists locally
auto file_exists = std::filesystem::exists(path);
// If the file exists, check its JSON metadata companion file.
std::string metadata_path = path + ".json";
nlohmann::json metadata;
std::string etag;
std::string last_modified;
if (file_exists) {
// Try and read the JSON metadata file (note: stream autoclosed upon exiting this block).
std::ifstream metadata_in(metadata_path);
if (metadata_in.good()) {
try {
metadata_in >> metadata;
LOG_INF("%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(), metadata.dump().c_str());
if (metadata.contains("url") && metadata.at("url").is_string()) {
auto previous_url = metadata.at("url").get<std::string>();
if (previous_url != url) {
LOG_ERR("%s: Model URL mismatch: %s != %s\n", __func__, url.c_str(), previous_url.c_str());
return false;
}
}
if (metadata.contains("etag") && metadata.at("etag").is_string()) {
etag = metadata.at("etag");
}
if (metadata.contains("lastModified") && metadata.at("lastModified").is_string()) {
last_modified = metadata.at("lastModified");
}
} catch (const nlohmann::json::exception & e) {
LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what());
return false;
}
}
} else {
LOG_INF("%s: no previous model file found %s\n", __func__, path.c_str());
}
// Send a HEAD request to retrieve the etag and last-modified headers
struct common_load_model_from_url_headers {
std::string etag;
std::string last_modified;
};
common_load_model_from_url_headers headers;
{
typedef size_t(*CURLOPT_HEADERFUNCTION_PTR)(char *, size_t, size_t, void *);
auto header_callback = [](char * buffer, size_t /*size*/, size_t n_items, void * userdata) -> size_t {
common_load_model_from_url_headers * headers = (common_load_model_from_url_headers *) userdata;
static std::regex header_regex("([^:]+): (.*)\r\n");
static std::regex etag_regex("ETag", std::regex_constants::icase);
static std::regex last_modified_regex("Last-Modified", std::regex_constants::icase);
std::string header(buffer, n_items);
std::smatch match;
if (std::regex_match(header, match, header_regex)) {
const std::string & key = match[1];
const std::string & value = match[2];
if (std::regex_match(key, match, etag_regex)) {
headers->etag = value;
} else if (std::regex_match(key, match, last_modified_regex)) {
headers->last_modified = value;
}
}
return n_items;
};
curl_easy_setopt(curl.get(), CURLOPT_NOBODY, 1L); // will trigger the HEAD verb
curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L); // hide head request progress
curl_easy_setopt(curl.get(), CURLOPT_HEADERFUNCTION, static_cast<CURLOPT_HEADERFUNCTION_PTR>(header_callback));
curl_easy_setopt(curl.get(), CURLOPT_HEADERDATA, &headers);
bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS);
if (!was_perform_successful) {
return false;
}
long http_code = 0;
curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &http_code);
if (http_code != 200) {
// HEAD not supported, we don't know if the file has changed
// force trigger downloading
force_download = true;
LOG_ERR("%s: HEAD invalid http status code received: %ld\n", __func__, http_code);
}
}
bool should_download = !file_exists || force_download;
if (!should_download) {
if (!etag.empty() && etag != headers.etag) {
LOG_WRN("%s: ETag header is different (%s != %s): triggering a new download\n", __func__, etag.c_str(), headers.etag.c_str());
should_download = true;
} else if (!last_modified.empty() && last_modified != headers.last_modified) {
LOG_WRN("%s: Last-Modified header is different (%s != %s): triggering a new download\n", __func__, last_modified.c_str(), headers.last_modified.c_str());
should_download = true;
}
}
if (should_download) {
std::string path_temporary = path + ".downloadInProgress";
if (file_exists) {
LOG_WRN("%s: deleting previous downloaded file: %s\n", __func__, path.c_str());
if (remove(path.c_str()) != 0) {
LOG_ERR("%s: unable to delete file: %s\n", __func__, path.c_str());
return false;
}
}
// Set the output file
struct FILE_deleter {
void operator()(FILE * f) const {
fclose(f);
}
};
std::unique_ptr<FILE, FILE_deleter> outfile(fopen(path_temporary.c_str(), "wb"));
if (!outfile) {
LOG_ERR("%s: error opening local file for writing: %s\n", __func__, path.c_str());
return false;
}
typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * data, size_t size, size_t nmemb, void * fd);
auto write_callback = [](void * data, size_t size, size_t nmemb, void * fd) -> size_t {
return fwrite(data, size, nmemb, (FILE *)fd);
};
curl_easy_setopt(curl.get(), CURLOPT_NOBODY, 0L);
curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast<CURLOPT_WRITEFUNCTION_PTR>(write_callback));
curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, outfile.get());
// display download progress
curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 0L);
// helper function to hide password in URL
auto llama_download_hide_password_in_url = [](const std::string & url) -> std::string {
std::size_t protocol_pos = url.find("://");
if (protocol_pos == std::string::npos) {
return url; // Malformed URL
}
std::size_t at_pos = url.find('@', protocol_pos + 3);
if (at_pos == std::string::npos) {
return url; // No password in URL
}
return url.substr(0, protocol_pos + 3) + "********" + url.substr(at_pos);
};
// start the download
LOG_INF("%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;
}
long http_code = 0;
curl_easy_getinfo (curl.get(), CURLINFO_RESPONSE_CODE, &http_code);
if (http_code < 200 || http_code >= 400) {
LOG_ERR("%s: invalid http status code received: %ld\n", __func__, http_code);
return false;
}
// Causes file to be closed explicitly here before we rename it.
outfile.reset();
// Write the updated JSON metadata file.
metadata.update({
{"url", url},
{"etag", headers.etag},
{"lastModified", headers.last_modified}
});
std::ofstream(metadata_path) << metadata.dump(4);
LOG_INF("%s: file metadata saved: %s\n", __func__, metadata_path.c_str());
if (rename(path_temporary.c_str(), path.c_str()) != 0) {
LOG_ERR("%s: unable to rename file: %s to %s\n", __func__, path_temporary.c_str(), path.c_str());
return false;
}
}
return true;
}
struct llama_model * common_load_model_from_url(
const std::string & model_url,
const std::string & local_path,
const std::string & hf_token,
const struct llama_model_params & params) {
// Basic validation of the model_url
if (model_url.empty()) {
LOG_ERR("%s: invalid model_url\n", __func__);
return NULL;
}
if (!common_download_file(model_url, local_path, hf_token)) {
return NULL;
}
// check for additional GGUFs split to download
int n_split = 0;
{
struct gguf_init_params gguf_params = {
/*.no_alloc = */ true,
/*.ctx = */ NULL,
};
auto * ctx_gguf = gguf_init_from_file(local_path.c_str(), gguf_params);
if (!ctx_gguf) {
LOG_ERR("\n%s: failed to load input GGUF from %s\n", __func__, local_path.c_str());
return NULL;
}
auto key_n_split = gguf_find_key(ctx_gguf, LLM_KV_SPLIT_COUNT);
if (key_n_split >= 0) {
n_split = gguf_get_val_u16(ctx_gguf, key_n_split);
}
gguf_free(ctx_gguf);
}
if (n_split > 1) {
char split_prefix[PATH_MAX] = {0};
char split_url_prefix[LLAMA_CURL_MAX_URL_LENGTH] = {0};
// Verify the first split file format
// and extract split URL and PATH prefixes
{
if (!llama_split_prefix(split_prefix, sizeof(split_prefix), local_path.c_str(), 0, n_split)) {
LOG_ERR("\n%s: unexpected model file name: %s n_split=%d\n", __func__, local_path.c_str(), n_split);
return NULL;
}
if (!llama_split_prefix(split_url_prefix, sizeof(split_url_prefix), model_url.c_str(), 0, n_split)) {
LOG_ERR("\n%s: unexpected model url: %s n_split=%d\n", __func__, model_url.c_str(), n_split);
return NULL;
}
}
// Prepare download in parallel
std::vector<std::future<bool>> futures_download;
for (int idx = 1; idx < n_split; idx++) {
futures_download.push_back(std::async(std::launch::async, [&split_prefix, &split_url_prefix, &n_split, hf_token](int download_idx) -> bool {
char split_path[PATH_MAX] = {0};
llama_split_path(split_path, sizeof(split_path), split_prefix, download_idx, n_split);
char split_url[LLAMA_CURL_MAX_URL_LENGTH] = {0};
llama_split_path(split_url, sizeof(split_url), split_url_prefix, download_idx, n_split);
return common_download_file(split_url, split_path, hf_token);
}, idx));
}
// Wait for all downloads to complete
for (auto & f : futures_download) {
if (!f.get()) {
return NULL;
}
}
}
return llama_model_load_from_file(local_path.c_str(), params);
}
struct llama_model * common_load_model_from_hf(
const std::string & repo,
const std::string & remote_path,
const std::string & local_path,
const std::string & hf_token,
const struct llama_model_params & params) {
// construct hugging face model url:
//
// --repo ggml-org/models --file tinyllama-1.1b/ggml-model-f16.gguf
// https://huggingface.co/ggml-org/models/resolve/main/tinyllama-1.1b/ggml-model-f16.gguf
//
// --repo TheBloke/Mixtral-8x7B-v0.1-GGUF --file mixtral-8x7b-v0.1.Q4_K_M.gguf
// https://huggingface.co/TheBloke/Mixtral-8x7B-v0.1-GGUF/resolve/main/mixtral-8x7b-v0.1.Q4_K_M.gguf
//
std::string model_url = "https://huggingface.co/";
model_url += repo;
model_url += "/resolve/main/";
model_url += remote_path;
return common_load_model_from_url(model_url, local_path, hf_token, params);
}
/**
* Allow getting the HF file from the HF repo with tag (like ollama), for example:
* - bartowski/Llama-3.2-3B-Instruct-GGUF:q4
* - bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M
* - bartowski/Llama-3.2-3B-Instruct-GGUF:q5_k_s
* Tag is optional, default to "latest" (meaning it checks for Q4_K_M first, then Q4, then if not found, return the first GGUF file in repo)
*
* Return pair of <repo, file> (with "repo" already having tag removed)
*
* Note: we use the Ollama-compatible HF API, but not using the blobId. Instead, we use the special "ggufFile" field which returns the value for "hf_file". This is done to be backward-compatible with existing cache files.
*/
std::pair<std::string, std::string> common_get_hf_file(const std::string & hf_repo_with_tag, const std::string & hf_token) {
auto parts = string_split<std::string>(hf_repo_with_tag, ':');
std::string tag = parts.size() > 1 ? parts.back() : "latest";
std::string hf_repo = parts[0];
if (string_split<std::string>(hf_repo, '/').size() != 2) {
throw std::invalid_argument("error: invalid HF repo format, expected <user>/<model>[:quant]\n");
}
// fetch model info from Hugging Face Hub API
json model_info;
curl_ptr curl(curl_easy_init(), &curl_easy_cleanup);
curl_slist_ptr http_headers;
std::string res_str;
std::string url = "https://huggingface.co/v2/" + hf_repo + "/manifests/" + tag;
curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L);
typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * ptr, size_t size, size_t nmemb, void * data);
auto write_callback = [](void * ptr, size_t size, size_t nmemb, void * data) -> size_t {
static_cast<std::string *>(data)->append((char * ) ptr, size * nmemb);
return size * nmemb;
};
curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast<CURLOPT_WRITEFUNCTION_PTR>(write_callback));
curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, &res_str);
#if defined(_WIN32)
curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA);
#endif
if (!hf_token.empty()) {
std::string auth_header = "Authorization: Bearer " + hf_token;
http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str());
}
// Important: the User-Agent must be "llama-cpp" to get the "ggufFile" field in the response
http_headers.ptr = curl_slist_append(http_headers.ptr, "User-Agent: llama-cpp");
http_headers.ptr = curl_slist_append(http_headers.ptr, "Accept: application/json");
curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
CURLcode res = curl_easy_perform(curl.get());
if (res != CURLE_OK) {
throw std::runtime_error("error: cannot make GET request to HF API");
}
long res_code;
curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &res_code);
if (res_code == 200) {
model_info = json::parse(res_str);
} else if (res_code == 401) {
throw std::runtime_error("error: model is private or does not exist; if you are accessing a gated model, please provide a valid HF token");
} else {
throw std::runtime_error(string_format("error from HF API, response code: %ld, data: %s", res_code, res_str.c_str()));
}
// check response
if (!model_info.contains("ggufFile")) {
throw std::runtime_error("error: model does not have ggufFile");
}
json & gguf_file = model_info.at("ggufFile");
if (!gguf_file.contains("rfilename")) {
throw std::runtime_error("error: ggufFile does not have rfilename");
}
return std::make_pair(hf_repo, gguf_file.at("rfilename"));
}
#else
struct llama_model * common_load_model_from_url(
const std::string & /*model_url*/,
const std::string & /*local_path*/,
const std::string & /*hf_token*/,
const struct llama_model_params & /*params*/) {
LOG_WRN("%s: llama.cpp built without libcurl, downloading from an url not supported.\n", __func__);
return nullptr;
}
struct llama_model * common_load_model_from_hf(
const std::string & /*repo*/,
const std::string & /*remote_path*/,
const std::string & /*local_path*/,
const std::string & /*hf_token*/,
const struct llama_model_params & /*params*/) {
LOG_WRN("%s: llama.cpp built without libcurl, downloading from Hugging Face not supported.\n", __func__);
return nullptr;
}
std::pair<std::string, std::string> common_get_hf_file(const std::string &, const std::string &) {
LOG_WRN("%s: llama.cpp built without libcurl, downloading from Hugging Face not supported.\n", __func__);
return std::make_pair("", "");
}
#endif // LLAMA_USE_CURL
//
// Batch utils
//
@@ -2032,26 +1550,3 @@ common_control_vector_data common_control_vector_load(const std::vector<common_c
return result;
}
template <>
json common_grammar_trigger::to_json() const {
json out {
{"type", (int) type},
{"value", value},
};
if (type == COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN) {
out["token"] = (int) token;
}
return out;
}
template <>
common_grammar_trigger common_grammar_trigger::from_json(const json & in) {
common_grammar_trigger out;
out.type = (common_grammar_trigger_type) in.at("type").get<int>();
out.value = in.at("value").get<std::string>();
if (out.type == COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN) {
out.token = (llama_token) in.at("token").get<int>();
}
return out;
}
+13 -36
View File
@@ -121,10 +121,6 @@ struct common_grammar_trigger {
common_grammar_trigger_type type;
std::string value;
llama_token token = LLAMA_TOKEN_NULL;
// T can only be nlohmann::ordered_json
template <class T> T to_json() const;
template <class T> static common_grammar_trigger from_json(const T & in);
};
// sampling parameters
@@ -184,6 +180,13 @@ struct common_params_sampling {
std::string print() const;
};
struct common_params_model {
std::string path = ""; // model local path // NOLINT
std::string url = ""; // model url to download // NOLINT
std::string hf_repo = ""; // HF repo // NOLINT
std::string hf_file = ""; // HF file // NOLINT
};
struct common_params_speculative {
std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
@@ -197,19 +200,11 @@ struct common_params_speculative {
struct cpu_params cpuparams;
struct cpu_params cpuparams_batch;
std::string hf_repo = ""; // HF repo // NOLINT
std::string hf_file = ""; // HF file // NOLINT
std::string model = ""; // draft model for speculative decoding // NOLINT
std::string model_url = ""; // model url to download // NOLINT
struct common_params_model model;
};
struct common_params_vocoder {
std::string hf_repo = ""; // HF repo // NOLINT
std::string hf_file = ""; // HF file // NOLINT
std::string model = ""; // model path // NOLINT
std::string model_url = ""; // model url to download // NOLINT
struct common_params_model model;
std::string speaker_file = ""; // speaker file path // NOLINT
@@ -267,12 +262,10 @@ struct common_params {
struct common_params_speculative speculative;
struct common_params_vocoder vocoder;
std::string model = ""; // model path // NOLINT
struct common_params_model model;
std::string model_alias = ""; // model alias // NOLINT
std::string model_url = ""; // model url to download // NOLINT
std::string hf_token = ""; // HF token // NOLINT
std::string hf_repo = ""; // HF repo // NOLINT
std::string hf_file = ""; // HF file // NOLINT
std::string prompt = ""; // NOLINT
std::string system_prompt = ""; // NOLINT
std::string prompt_file = ""; // store the external prompt file name // NOLINT
@@ -286,6 +279,7 @@ struct common_params {
std::vector<std::string> in_files; // all input files
std::vector<std::string> antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts)
std::vector<llama_model_kv_override> kv_overrides;
std::vector<llama_model_tensor_buft_override> tensor_buft_overrides;
bool lora_init_without_apply = false; // only load lora to memory, but do not apply it to ctx (user can manually apply lora later using llama_adapter_lora_apply)
std::vector<common_adapter_lora_info> lora_adapters; // lora adapter path with user defined scale
@@ -347,7 +341,7 @@ struct common_params {
common_conversation_mode conversation_mode = COMMON_CONVERSATION_MODE_AUTO;
// multimodal models (see examples/llava)
std::string mmproj = ""; // path to multimodal projector // NOLINT
struct common_params_model mmproj;
std::vector<std::string> image; // path to image file(s)
// embedding
@@ -546,23 +540,6 @@ struct llama_model_params common_model_params_to_llama ( common_params
struct llama_context_params common_context_params_to_llama(const common_params & params);
struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params);
struct llama_model * common_load_model_from_url(
const std::string & model_url,
const std::string & local_path,
const std::string & hf_token,
const struct llama_model_params & params);
struct llama_model * common_load_model_from_hf(
const std::string & repo,
const std::string & remote_path,
const std::string & local_path,
const std::string & hf_token,
const struct llama_model_params & params);
std::pair<std::string, std::string> common_get_hf_file(
const std::string & hf_repo_with_tag,
const std::string & hf_token);
// clear LoRA adapters from context, then apply new list of adapters
void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora);
+2 -2
View File
@@ -240,7 +240,7 @@ public:
auto index = key.get<int>();
return array_->at(index < 0 ? array_->size() + index : index);
} else if (object_) {
if (!key.is_hashable()) throw std::runtime_error("Unashable type: " + dump());
if (!key.is_hashable()) throw std::runtime_error("Unhashable type: " + dump());
auto it = object_->find(key.primitive_);
if (it == object_->end()) return Value();
return it->second;
@@ -249,7 +249,7 @@ public:
}
void set(const Value& key, const Value& value) {
if (!object_) throw std::runtime_error("Value is not an object: " + dump());
if (!key.is_hashable()) throw std::runtime_error("Unashable type: " + dump());
if (!key.is_hashable()) throw std::runtime_error("Unhashable type: " + dump());
(*object_)[key.primitive_] = value;
}
Value call(const std::shared_ptr<Context> & context, ArgumentsValue & args) const {
+4 -7
View File
@@ -3557,8 +3557,8 @@ class RWKV6Qwen2Model(Rwkv6Model):
head_size = hidden_size // num_attention_heads
rms_norm_eps = self.hparams["rms_norm_eps"]
intermediate_size = self.hparams["intermediate_size"]
time_mix_extra_dim = 64 if hidden_size >= 4096 else 32
time_decay_extra_dim = 128 if hidden_size >= 4096 else 64
time_mix_extra_dim = self.hparams.get("lora_rank_tokenshift", 64 if hidden_size >= 4096 else 32)
time_decay_extra_dim = self.hparams.get("lora_rank_decay", 128 if hidden_size >= 4096 else 64)
# RWKV isn't context limited
self.gguf_writer.add_context_length(1048576)
@@ -5146,10 +5146,7 @@ class BailingMoeModel(Model):
def set_gguf_parameters(self):
super().set_gguf_parameters()
hparams = self.hparams
if "head_dim" in hparams:
rope_dim = hparams["head_dim"]
else:
rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
rope_dim = hparams.get("head_dim") or hparams["hidden_size"] // hparams["num_attention_heads"]
self.gguf_writer.add_rope_dimension_count(rope_dim)
self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
@@ -5175,7 +5172,7 @@ class BailingMoeModel(Model):
n_head = self.hparams["num_attention_heads"]
n_kv_head = self.hparams.get("num_key_value_heads")
n_embd = self.hparams["hidden_size"]
head_dim = self.hparams.get("head_dim", n_embd // n_head)
head_dim = self.hparams.get("head_dim") or n_embd // n_head
output_name = self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT)
+7 -32
View File
@@ -20,7 +20,7 @@
**oneAPI** is an open ecosystem and a standard-based specification, supporting multiple architectures including but not limited to intel CPUs, GPUs and FPGAs. The key components of the oneAPI ecosystem include:
- **DPCPP** *(Data Parallel C++)*: The primary oneAPI SYCL implementation, which includes the icpx/icx Compilers.
- **oneAPI Libraries**: A set of highly optimized libraries targeting multiple domains *(e.g. oneMKL and oneDNN)*.
- **oneAPI Libraries**: A set of highly optimized libraries targeting multiple domains *(e.g. Intel oneMKL, oneMath and oneDNN)*.
- **oneAPI LevelZero**: A high performance low level interface for fine-grained control over intel iGPUs and dGPUs.
- **Nvidia & AMD Plugins**: These are plugins extending oneAPI's DPCPP support to SYCL on Nvidia and AMD GPU targets.
@@ -227,16 +227,6 @@ Upon a successful installation, SYCL is enabled for the available intel devices,
**oneAPI Plugin**: In order to enable SYCL support on Nvidia GPUs, please install the [Codeplay oneAPI Plugin for Nvidia GPUs](https://developer.codeplay.com/products/oneapi/nvidia/download). User should also make sure the plugin version matches the installed base toolkit one *(previous step)* for a seamless "oneAPI on Nvidia GPU" setup.
**oneMKL for cuBlas**: The current oneMKL releases *(shipped with the oneAPI base-toolkit)* do not contain the cuBLAS backend. A build from source of the upstream [oneMKL](https://github.com/oneapi-src/oneMKL) with the *cuBLAS* backend enabled is thus required to run it on Nvidia GPUs.
```sh
git clone https://github.com/oneapi-src/oneMKL
cd oneMKL
cmake -B buildWithCublas -DCMAKE_CXX_COMPILER=icpx -DCMAKE_C_COMPILER=icx -DENABLE_MKLGPU_BACKEND=OFF -DENABLE_MKLCPU_BACKEND=OFF -DENABLE_CUBLAS_BACKEND=ON -DTARGET_DOMAINS=blas
cmake --build buildWithCublas --config Release
```
**oneDNN**: The current oneDNN releases *(shipped with the oneAPI base-toolkit)* do not include the NVIDIA backend. Therefore, oneDNN must be compiled from source to enable the NVIDIA target:
```sh
@@ -250,16 +240,6 @@ cmake --build build-nvidia --config Release
**oneAPI Plugin**: In order to enable SYCL support on AMD GPUs, please install the [Codeplay oneAPI Plugin for AMD GPUs](https://developer.codeplay.com/products/oneapi/amd/download). As with Nvidia GPUs, the user should also make sure the plugin version matches the installed base toolkit.
**oneMKL for rocBlas**: The current oneMKL releases *(shipped with the oneAPI base-toolkit)* doesn't contain the rocBLAS backend. A build from source of the upstream [oneMKL](https://github.com/oneapi-src/oneMKL) with the *rocBLAS* backend enabled is thus required to run it on AMD GPUs.
```sh
git clone https://github.com/oneapi-src/oneMKL
cd oneMKL
# Find your HIPTARGET with rocminfo, under the key 'Name:'
cmake -B buildWithrocBLAS -DCMAKE_CXX_COMPILER=icpx -DCMAKE_C_COMPILER=icx -DENABLE_MKLGPU_BACKEND=OFF -DENABLE_MKLCPU_BACKEND=OFF -DENABLE_ROCBLAS_BACKEND=ON -DHIPTARGETS=${HIPTARGET} -DTARGET_DOMAINS=blas
cmake --build buildWithrocBLAS --config Release
```
3. **Verify installation and environment**
In order to check the available SYCL devices on the machine, please use the `sycl-ls` command.
@@ -324,13 +304,10 @@ cmake --build build --config Release -j -v
#### Nvidia GPU
```sh
# Export relevant ENV variables
export LD_LIBRARY_PATH=/path/to/oneMKL/buildWithCublas/lib:$LD_LIBRARY_PATH
export LIBRARY_PATH=/path/to/oneMKL/buildWithCublas/lib:$LIBRARY_PATH
export CPLUS_INCLUDE_DIR=/path/to/oneMKL/buildWithCublas/include:$CPLUS_INCLUDE_DIR
export CPLUS_INCLUDE_DIR=/path/to/oneMKL/include:$CPLUS_INCLUDE_DIR
The SYCL backend depends on [oneMath](https://github.com/uxlfoundation/oneMath) for Nvidia and AMD devices.
By default it is automatically built along with the project. A specific build can be provided by setting the CMake flag `-DoneMath_DIR=/path/to/oneMath/install/lib/cmake/oneMath`.
```sh
# Build LLAMA with Nvidia BLAS acceleration through SYCL
# Setting GGML_SYCL_DEVICE_ARCH is optional but can improve performance
GGML_SYCL_DEVICE_ARCH=sm_80 # Example architecture
@@ -347,12 +324,10 @@ cmake --build build --config Release -j -v
#### AMD GPU
```sh
# Export relevant ENV variables
export LD_LIBRARY_PATH=/path/to/oneMKL/buildWithrocBLAS/lib:$LD_LIBRARY_PATH
export LIBRARY_PATH=/path/to/oneMKL/buildWithrocBLAS/lib:$LIBRARY_PATH
export CPLUS_INCLUDE_DIR=/path/to/oneMKL/buildWithrocBLAS/include:$CPLUS_INCLUDE_DIR
The SYCL backend depends on [oneMath](https://github.com/uxlfoundation/oneMath) for Nvidia and AMD devices.
By default it is automatically built along with the project. A specific build can be provided by setting the CMake flag `-DoneMath_DIR=/path/to/oneMath/install/lib/cmake/oneMath`.
```sh
# Build LLAMA with rocBLAS acceleration through SYCL
## AMD
+1 -1
View File
@@ -38,7 +38,7 @@ int main(int argc, char ** argv) {
llama_model_params model_params = common_model_params_to_llama(params);
llama_model * model = llama_model_load_from_file(params.model.c_str(), model_params);
llama_model * model = llama_model_load_from_file(params.model.path.c_str(), model_params);
if (model == NULL) {
fprintf(stderr , "%s: error: unable to load model\n" , __func__);
+1 -1
View File
@@ -41,7 +41,7 @@ int main(int argc, char ** argv) {
llama_model_params model_params = common_model_params_to_llama(params);
llama_model * model = llama_model_load_from_file(params.model.c_str(), model_params);
llama_model * model = llama_model_load_from_file(params.model.path.c_str(), model_params);
if (model == NULL) {
LOG_ERR("%s: error: unable to load model\n" , __func__);
+1 -1
View File
@@ -421,7 +421,7 @@ int main(int argc, char ** argv) {
g_verbose = (params.verbosity > 1);
try {
lora_merge_ctx ctx(params.model, params.lora_adapters, params.out_file, params.cpuparams.n_threads);
lora_merge_ctx ctx(params.model.path, params.lora_adapters, params.out_file, params.cpuparams.n_threads);
ctx.run_merge();
} catch (const std::exception & err) {
fprintf(stderr, "%s\n", err.what());
+1 -1
View File
@@ -168,7 +168,7 @@ int main(int argc, char * argv[]) {
llama_backend_init();
llama_model * model = llama_model_load_from_file(params.model.c_str(), mparams);
llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
// create generation context
llama_context * ctx = llama_init_from_model(model, cparams);
+20
View File
@@ -4,6 +4,26 @@
>
> This is very experimental, only used for demo purpose.
## Quick started
You can use pre-quantized model from [ggml-org](https://huggingface.co/ggml-org)'s Hugging Face account
```bash
# build
cmake -B build
cmake --build build --target llama-gemma3-cli
# alternatively, install from brew (MacOS)
brew install llama.cpp
# run it
llama-gemma3-cli -hf ggml-org/gemma-3-4b-it-GGUF
llama-gemma3-cli -hf ggml-org/gemma-3-12b-it-GGUF
llama-gemma3-cli -hf ggml-org/gemma-3-27b-it-GGUF
# note: 1B model does not support vision
```
## How to get mmproj.gguf?
```bash
+5 -3
View File
@@ -1396,14 +1396,16 @@ struct clip_ctx * clip_init(const char * fname, struct clip_context_params ctx_p
const int n_kv = gguf_get_n_kv(ctx);
const int ftype = get_u32(ctx, KEY_FTYPE);
const std::string ftype_str = get_ftype(ftype);
const int idx_desc = get_key_idx(ctx, KEY_DESCRIPTION);
const std::string description = gguf_get_val_str(ctx, idx_desc);
const int idx_name = gguf_find_key(ctx, KEY_NAME);
if (idx_name != -1) { // make name optional temporarily as some of the uploaded models missing it due to a bug
const std::string name = gguf_get_val_str(ctx, idx_name);
LOG_INF("%s: model name: %s\n", __func__, name.c_str());
}
LOG_INF("%s: description: %s\n", __func__, description.c_str());
const int idx_desc = gguf_find_key(ctx, KEY_DESCRIPTION);
if (idx_desc != -1) { // ditto
const std::string description = gguf_get_val_str(ctx, idx_desc);
LOG_INF("%s: description: %s\n", __func__, description.c_str());
}
LOG_INF("%s: GGUF version: %d\n", __func__, gguf_get_version(ctx));
LOG_INF("%s: alignment: %zu\n", __func__, gguf_get_alignment(ctx));
LOG_INF("%s: n_tensors: %d\n", __func__, n_tensors);
+3 -3
View File
@@ -78,7 +78,7 @@ struct gemma3_context {
}
void init_clip_model(common_params & params) {
const char * clip_path = params.mmproj.c_str();
const char * clip_path = params.mmproj.path.c_str();
ctx_clip = clip_model_load(clip_path, params.verbosity > 1);
}
@@ -232,13 +232,13 @@ int main(int argc, char ** argv) {
common_init();
if (params.mmproj.empty()) {
if (params.mmproj.path.empty()) {
show_additional_info(argc, argv);
return 1;
}
gemma3_context ctx(params);
printf("%s: %s\n", __func__, params.model.c_str());
printf("%s: %s\n", __func__, params.model.path.c_str());
bool is_single_turn = !params.prompt.empty() && !params.image.empty();
+3 -3
View File
@@ -225,7 +225,7 @@ static struct llama_model * llava_init(common_params * params) {
llama_model_params model_params = common_model_params_to_llama(*params);
llama_model * model = llama_model_load_from_file(params->model.c_str(), model_params);
llama_model * model = llama_model_load_from_file(params->model.path.c_str(), model_params);
if (model == NULL) {
LOG_ERR("%s: unable to load model\n" , __func__);
return NULL;
@@ -234,7 +234,7 @@ static struct llama_model * llava_init(common_params * params) {
}
static struct llava_context * llava_init_context(common_params * params, llama_model * model) {
const char * clip_path = params->mmproj.c_str();
const char * clip_path = params->mmproj.path.c_str();
auto prompt = params->prompt;
if (prompt.empty()) {
@@ -283,7 +283,7 @@ int main(int argc, char ** argv) {
common_init();
if (params.mmproj.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
if (params.mmproj.path.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
print_usage(argc, argv);
return 1;
}
+3 -3
View File
@@ -31,7 +31,7 @@ static struct llama_model * llava_init(common_params * params) {
llama_model_params model_params = common_model_params_to_llama(*params);
llama_model * model = llama_model_load_from_file(params->model.c_str(), model_params);
llama_model * model = llama_model_load_from_file(params->model.path.c_str(), model_params);
if (model == NULL) {
LOG_ERR("%s: unable to load model\n" , __func__);
return NULL;
@@ -80,7 +80,7 @@ static void llava_free(struct llava_context * ctx_llava) {
}
static struct clip_ctx * clip_init_context(common_params * params) {
const char * clip_path = params->mmproj.c_str();
const char * clip_path = params->mmproj.path.c_str();
auto prompt = params->prompt;
if (prompt.empty()) {
@@ -290,7 +290,7 @@ int main(int argc, char ** argv) {
common_init();
if (params.mmproj.empty() || (params.image.empty())) {
if (params.mmproj.path.empty() || (params.image.empty())) {
show_additional_info(argc, argv);
return 1;
}
+3 -3
View File
@@ -314,7 +314,7 @@ static struct llama_model * llava_init(common_params * params) {
llama_model_params model_params = common_model_params_to_llama(*params);
llama_model * model = llama_model_load_from_file(params->model.c_str(), model_params);
llama_model * model = llama_model_load_from_file(params->model.path.c_str(), model_params);
if (model == NULL) {
LOG_ERR("%s: unable to load model\n" , __func__);
return NULL;
@@ -323,7 +323,7 @@ static struct llama_model * llava_init(common_params * params) {
}
static struct llava_context * llava_init_context(common_params * params, llama_model * model) {
const char * clip_path = params->mmproj.c_str();
const char * clip_path = params->mmproj.path.c_str();
auto prompt = params->prompt;
if (prompt.empty()) {
@@ -524,7 +524,7 @@ int main(int argc, char ** argv) {
common_init();
if (params.mmproj.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
if (params.mmproj.path.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
print_usage(argc, argv);
return 1;
}
+3 -1
View File
@@ -106,6 +106,8 @@ int main(int argc, char ** argv) {
common_params params;
params.n_predict = 128;
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_PARALLEL)) {
return 1;
}
@@ -405,7 +407,7 @@ int main(int argc, char ** argv) {
params.prompt_file = "used built-in defaults";
}
LOG_INF("External prompt file: \033[32m%s\033[0m\n", params.prompt_file.c_str());
LOG_INF("Model and path used: \033[32m%s\033[0m\n\n", params.model.c_str());
LOG_INF("Model and path used: \033[32m%s\033[0m\n\n", params.model.path.c_str());
LOG_INF("Total prompt tokens: %6d, speed: %5.2f t/s\n", n_total_prompt, (double) (n_total_prompt ) / (t_main_end - t_main_start) * 1e6);
LOG_INF("Total gen tokens: %6d, speed: %5.2f t/s\n", n_total_gen, (double) (n_total_gen ) / (t_main_end - t_main_start) * 1e6);
+1 -1
View File
@@ -64,7 +64,7 @@ int main(int argc, char ** argv) {
llama_model_params model_params = common_model_params_to_llama(params);
llama_model * model = llama_model_load_from_file(params.model.c_str(), model_params);
llama_model * model = llama_model_load_from_file(params.model.path.c_str(), model_params);
if (model == NULL) {
LOG_ERR("%s: unable to load model\n" , __func__);
+14 -16
View File
@@ -133,7 +133,8 @@ struct slot_params {
auto grammar_triggers = json::array();
for (const auto & trigger : sampling.grammar_triggers) {
grammar_triggers.push_back(trigger.to_json<json>());
server_grammar_trigger ct(std::move(trigger));
grammar_triggers.push_back(ct.to_json());
}
return json {
@@ -372,9 +373,9 @@ struct server_task {
const auto grammar_triggers = data.find("grammar_triggers");
if (grammar_triggers != data.end()) {
for (const auto & t : *grammar_triggers) {
auto ct = common_grammar_trigger::from_json(t);
if (ct.type == COMMON_GRAMMAR_TRIGGER_TYPE_WORD) {
const auto & word = ct.value;
server_grammar_trigger ct(t);
if (ct.value.type == COMMON_GRAMMAR_TRIGGER_TYPE_WORD) {
const auto & word = ct.value.value;
auto ids = common_tokenize(vocab, word, /* add_special= */ false, /* parse_special= */ true);
if (ids.size() == 1) {
auto token = ids[0];
@@ -392,7 +393,7 @@ struct server_task {
params.sampling.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, word});
}
} else {
params.sampling.grammar_triggers.push_back(ct);
params.sampling.grammar_triggers.push_back(std::move(ct.value));
}
}
}
@@ -1876,7 +1877,7 @@ struct server_context {
}
bool load_model(const common_params & params) {
SRV_INF("loading model '%s'\n", params.model.c_str());
SRV_INF("loading model '%s'\n", params.model.path.c_str());
params_base = params;
@@ -1886,7 +1887,7 @@ struct server_context {
ctx = llama_init.context.get();
if (model == nullptr) {
SRV_ERR("failed to load model, '%s'\n", params_base.model.c_str());
SRV_ERR("failed to load model, '%s'\n", params_base.model.path.c_str());
return false;
}
@@ -1897,16 +1898,13 @@ struct server_context {
add_bos_token = llama_vocab_get_add_bos(vocab);
has_eos_token = llama_vocab_eos(vocab) != LLAMA_TOKEN_NULL;
if (!params_base.speculative.model.empty() || !params_base.speculative.hf_repo.empty()) {
SRV_INF("loading draft model '%s'\n", params_base.speculative.model.c_str());
if (!params_base.speculative.model.path.empty() || !params_base.speculative.model.hf_repo.empty()) {
SRV_INF("loading draft model '%s'\n", params_base.speculative.model.path.c_str());
auto params_dft = params_base;
params_dft.devices = params_base.speculative.devices;
params_dft.hf_file = params_base.speculative.hf_file;
params_dft.hf_repo = params_base.speculative.hf_repo;
params_dft.model = params_base.speculative.model;
params_dft.model_url = params_base.speculative.model_url;
params_dft.n_ctx = params_base.speculative.n_ctx == 0 ? params_base.n_ctx / params_base.n_parallel : params_base.speculative.n_ctx;
params_dft.n_gpu_layers = params_base.speculative.n_gpu_layers;
params_dft.n_parallel = 1;
@@ -1920,12 +1918,12 @@ struct server_context {
model_dft = llama_init_dft.model.get();
if (model_dft == nullptr) {
SRV_ERR("failed to load draft model, '%s'\n", params_base.speculative.model.c_str());
SRV_ERR("failed to load draft model, '%s'\n", params_base.speculative.model.path.c_str());
return false;
}
if (!common_speculative_are_compatible(ctx, llama_init_dft.context.get())) {
SRV_ERR("the draft model '%s' is not compatible with the target model '%s'\n", params_base.speculative.model.c_str(), params_base.model.c_str());
SRV_ERR("the draft model '%s' is not compatible with the target model '%s'\n", params_base.speculative.model.path.c_str(), params_base.model.path.c_str());
return false;
}
@@ -3865,7 +3863,7 @@ int main(int argc, char ** argv) {
json data = {
{ "default_generation_settings", ctx_server.default_generation_settings_for_props },
{ "total_slots", ctx_server.params_base.n_parallel },
{ "model_path", ctx_server.params_base.model },
{ "model_path", ctx_server.params_base.model.path },
{ "chat_template", common_chat_templates_source(ctx_server.chat_templates.get()) },
{ "bos_token", common_token_to_piece(ctx_server.ctx, llama_vocab_bos(ctx_server.vocab), /* special= */ true)},
{ "eos_token", common_token_to_piece(ctx_server.ctx, llama_vocab_eos(ctx_server.vocab), /* special= */ true)},
@@ -4131,7 +4129,7 @@ int main(int argc, char ** argv) {
{"object", "list"},
{"data", {
{
{"id", params.model_alias.empty() ? params.model : params.model_alias},
{"id", params.model_alias.empty() ? params.model.path : params.model_alias},
{"object", "model"},
{"created", std::time(0)},
{"owned_by", "llamacpp"},
+28 -1
View File
@@ -58,6 +58,32 @@ static T json_value(const json & body, const std::string & key, const T & defaul
const static std::string build_info("b" + std::to_string(LLAMA_BUILD_NUMBER) + "-" + LLAMA_COMMIT);
// thin wrapper around common_grammar_trigger with (de)serialization functions
struct server_grammar_trigger {
common_grammar_trigger value;
server_grammar_trigger() = default;
server_grammar_trigger(const common_grammar_trigger & value) : value(value) {}
server_grammar_trigger(const json & in) {
value.type = (common_grammar_trigger_type) in.at("type").get<int>();
value.value = in.at("value").get<std::string>();
if (value.type == COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN) {
value.token = (llama_token) in.at("token").get<int>();
}
}
json to_json() const {
json out {
{"type", (int) value.type},
{"value", value.value},
};
if (value.type == COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN) {
out["token"] = (int) value.token;
}
return out;
}
};
//
// tokenizer and input processing utils
//
@@ -627,7 +653,8 @@ static json oaicompat_completion_params_parse(
llama_params["grammar_lazy"] = chat_params.grammar_lazy;
auto grammar_triggers = json::array();
for (const auto & trigger : chat_params.grammar_triggers) {
grammar_triggers.push_back(trigger.to_json<json>());
server_grammar_trigger ct(trigger);
grammar_triggers.push_back(ct.to_json());
}
llama_params["grammar_triggers"] = grammar_triggers;
llama_params["preserved_tokens"] = chat_params.preserved_tokens;
@@ -24,7 +24,7 @@ int main(int argc, char ** argv) {
common_init();
if (params.speculative.model.empty()) {
if (params.speculative.model.path.empty()) {
LOG_ERR("%s: --model-draft is required\n", __func__);
return 1;
}
+1 -1
View File
@@ -46,7 +46,7 @@ int main(int argc, char ** argv) {
common_init();
if (params.speculative.model.empty()) {
if (params.speculative.model.path.empty()) {
LOG_ERR("%s: --model-draft is required\n", __func__);
return 1;
}
+1 -6
View File
@@ -577,12 +577,7 @@ int main(int argc, char ** argv) {
const llama_vocab * vocab = llama_model_get_vocab(model_ttc);
// TODO: refactor in a common struct
params.model = params.vocoder.model;
params.model_url = params.vocoder.model_url;
params.hf_repo = params.vocoder.hf_repo;
params.hf_file = params.vocoder.hf_file;
params.model = params.vocoder.model;
params.embedding = true;
common_init_result llama_init_cts = common_init_from_params(params);
-2
View File
@@ -51,13 +51,11 @@ if (CANN_INSTALL_DIR)
${CANN_INSTALL_DIR}/acllib/include
)
add_subdirectory(kernels)
list(APPEND CANN_LIBRARIES
ascendcl
nnopbase
opapi
acl_op_compiler
ascendc_kernels
)
file(GLOB GGML_SOURCES_CANN "*.cpp")
+6 -4
View File
@@ -54,9 +54,7 @@ aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne,
// added.
int64_t acl_ne[GGML_MAX_DIMS * 2], acl_stride[GGML_MAX_DIMS * 2];
int64_t acl_storage_len = 0;
if (ne == nullptr) {
acl_storage_len = ggml_nbytes(tensor);
for (int i = 0; i < GGML_MAX_DIMS; i++) {
acl_ne[i] = tensor->ne[i];
// The step size of acl is in elements.
@@ -65,14 +63,18 @@ aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne,
} else {
// With bcast
for (int i = 0; i < dims; i++) {
acl_storage_len += (ne[i] - 1) * nb[i];
acl_ne[i] = ne[i];
acl_stride[i] = nb[i] / ggml_element_size(tensor);
}
}
// Reverse ne and stride.
int64_t final_dims = (dims == 0 ? GGML_MAX_DIMS : dims);
int64_t acl_storage_len = 1;
for (int i = 0; i < final_dims; i++) {
acl_storage_len += (acl_ne[i] - 1) * acl_stride[i];
}
// Reverse ne and stride.
std::reverse(acl_ne, acl_ne + final_dims);
std::reverse(acl_stride, acl_stride + final_dims);
+5 -5
View File
@@ -101,14 +101,14 @@ aclTensor* ggml_cann_create_tensor(void* data_ptr, aclDataType dtype,
tmp_stride[i] = nb[i] / type_size;
}
int64_t acl_storage_len = 1;
for (int i = 0; i < dims; i++) {
acl_storage_len += (tmp_ne[i] - 1) * tmp_stride[i];
}
std::reverse(tmp_ne, tmp_ne + dims);
std::reverse(tmp_stride, tmp_stride + dims);
int64_t acl_storage_len = 0;
for (int i = 0; i < dims; i++) {
acl_storage_len += (ne[i] - 1) * nb[i];
}
aclTensor* acl_tensor =
aclCreateTensor(tmp_ne, dims, dtype, tmp_stride, offset / type_size,
format, &acl_storage_len, 1, data_ptr);
+247 -236
View File
@@ -30,6 +30,7 @@
#include <aclnnop/aclnn_copy.h>
#include <aclnnop/aclnn_cos.h>
#include <aclnnop/aclnn_div.h>
#include <aclnnop/aclnn_embedding.h>
#include <aclnnop/aclnn_exp.h>
#include <aclnnop/aclnn_fill_scalar.h>
#include <aclnnop/aclnn_group_norm.h>
@@ -58,7 +59,6 @@
#include <vector>
#include "ggml-impl.h"
#include "kernels/ascendc_kernels.h"
#define GGML_COMMON_DECL_C
@@ -99,6 +99,35 @@ static void aclnn_repeat(ggml_backend_cann_context& ctx, aclTensor* acl_src,
ACL_CHECK(aclDestroyIntArray(repeats));
}
/**
* @brief Casts the elements of a tensor to a specified data type using the CANN backend.
*
* @details This function performs a type conversion on the elements of the input tensor `acl_src`
* and stores the results in the destination tensor `acl_dst`. The conversion type is
* determined based on the `dst` tensor's data type.
*
* @param ctx The context for the CANN backend operations.
* @param acl_src The source tensor whose elements will be cast.
* @param acl_dst The destination tensor that will store the casted elements.
* @param dst The ggml tensor specifying the target data type.
*/
static void aclnn_cast(ggml_backend_cann_context& ctx, aclTensor* acl_src,
aclTensor* acl_dst, ggml_tensor* dst) {
uint64_t workspaceSize = 0;
aclOpExecutor* executor;
void* workspaceAddr = nullptr;
ACL_CHECK(aclnnCastGetWorkspaceSize(acl_src,
ggml_cann_type_mapping(dst->type),
acl_dst, &workspaceSize, &executor));
if (workspaceSize > 0) {
ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
workspaceAddr = workspace_allocator.get();
}
ACL_CHECK(aclnnCast(workspaceAddr, workspaceSize, executor, ctx.stream()));
}
void ggml_cann_repeat(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
ggml_tensor* src = dst->src[0];
GGML_ASSERT(ggml_can_repeat(src, dst));
@@ -329,8 +358,6 @@ void ggml_cann_sqr(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
void ggml_cann_clamp(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
ggml_tensor* src = dst->src[0];
GGML_ASSERT(src->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
float min;
float max;
@@ -889,173 +916,76 @@ static void cann_copy(ggml_backend_cann_context& ctx, aclTensor* acl_src,
}
void ggml_cann_dup(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
ggml_tensor* src = dst->src[0];
ggml_tensor* src0 = dst->src[0];
aclTensor* acl_src = ggml_cann_create_tensor(src);
aclTensor* acl_src = ggml_cann_create_tensor(src0);
aclTensor* acl_dst = ggml_cann_create_tensor(dst);
ggml_cann_pool_alloc src_extra_allocator(ctx.pool(), sizeof(ggml_tensor));
ggml_cann_pool_alloc dst_extra_allocator(ctx.pool(), sizeof(ggml_tensor));
src->extra = src_extra_allocator.get();
dst->extra = dst_extra_allocator.get();
ACL_CHECK(aclrtMemcpyAsync(src->extra, sizeof(ggml_tensor), src,
sizeof(ggml_tensor), ACL_MEMCPY_HOST_TO_DEVICE,
ctx.stream()));
ACL_CHECK(aclrtMemcpyAsync(dst->extra, sizeof(ggml_tensor), dst,
sizeof(ggml_tensor), ACL_MEMCPY_HOST_TO_DEVICE,
ctx.stream()));
if ((dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32) &&
ggml_are_same_shape(src, dst)) {
cann_copy(ctx, acl_src, acl_dst);
ACL_CHECK(aclDestroyTensor(acl_src));
ACL_CHECK(aclDestroyTensor(acl_dst));
return;
}
// TODO: simplify
if (src->type == GGML_TYPE_F16) {
if (dst->type == GGML_TYPE_Q8_0) {
aclrtlaunch_ascendc_quantize_f16_q8_0(
24, ctx.stream(), src->data, dst->data,
((ggml_tensor*)src->extra)->ne, ((ggml_tensor*)src->extra)->nb,
((ggml_tensor*)dst->extra)->ne);
return;
}
if (dst->type == GGML_TYPE_Q4_0) {
aclrtlaunch_ascendc_quantize_f16_to_q4_0(
24, ctx.stream(), src->data, dst->data,
((ggml_tensor*)src->extra)->ne, ((ggml_tensor*)src->extra)->nb,
((ggml_tensor*)dst->extra)->ne);
return;
}
if (dst->type == GGML_TYPE_F16) {
if (ggml_are_same_shape(src, dst)) {
cann_copy(ctx, acl_src, acl_dst);
ACL_CHECK(aclDestroyTensor(acl_src));
ACL_CHECK(aclDestroyTensor(acl_dst));
return;
}
if (ggml_is_contiguous(dst)) {
const size_t src_type_size = ggml_type_size(src->type);
if (src->nb[0] == src_type_size) {
// src0 is contigous on first dimension, copy by rows
int64_t rows_num = ggml_nrows(src);
aclrtlaunch_ascendc_dup_by_rows_fp16(
rows_num, ctx.stream(), src->data, dst->data,
((ggml_tensor*)src->extra)->ne,
((ggml_tensor*)src->extra)->nb,
((ggml_tensor*)dst->extra)->ne,
((ggml_tensor*)dst->extra)->nb);
return;
}
GGML_ABORT("fatal error");
}
GGML_ABORT("fatal error");
}
if (dst->type == GGML_TYPE_F32) {
if (ggml_are_same_shape(src, dst)) {
cann_copy(ctx, acl_src, acl_dst);
ACL_CHECK(aclDestroyTensor(acl_src));
ACL_CHECK(aclDestroyTensor(acl_dst));
return;
}
if (ggml_is_contiguous(dst)) {
const size_t src_type_size = ggml_type_size(src->type);
if (src->nb[0] == src_type_size) {
// src0 is contigous on first dimension, copy by rows
int64_t rows_num = ggml_nrows(src);
aclrtlaunch_ascendc_dup_by_rows_fp16_to_fp32(
rows_num, ctx.stream(), src->data, dst->data,
((ggml_tensor*)src->extra)->ne,
((ggml_tensor*)src->extra)->nb,
((ggml_tensor*)dst->extra)->ne,
((ggml_tensor*)dst->extra)->nb);
return;
}
GGML_ABORT("fatal error");
}
GGML_ABORT("fatal error");
}
// TODO
GGML_ABORT("fatal error");
} else if (src->type == GGML_TYPE_F32) {
// TODO: if (src0->type == dst->type && ne00 == ne0 && nb00 == type_size
// && nb0 == type_size)
if (dst->type == GGML_TYPE_Q8_0) {
aclrtlaunch_ascendc_quantize_f32_q8_0(
24, ctx.stream(), src->data, dst->data,
((ggml_tensor*)src->extra)->ne, ((ggml_tensor*)src->extra)->nb,
((ggml_tensor*)dst->extra)->ne);
return;
}
if (dst->type == GGML_TYPE_Q4_0) {
aclrtlaunch_ascendc_quantize_f32_to_q4_0(
24, ctx.stream(), src->data, dst->data,
((ggml_tensor*)src->extra)->ne, ((ggml_tensor*)src->extra)->nb,
((ggml_tensor*)dst->extra)->ne);
return;
}
if (dst->type == GGML_TYPE_F32) {
if (ggml_are_same_shape(src, dst)) {
cann_copy(ctx, acl_src, acl_dst);
ACL_CHECK(aclDestroyTensor(acl_src));
ACL_CHECK(aclDestroyTensor(acl_dst));
return;
}
if (ggml_is_contiguous(dst)) {
const size_t src_type_size = ggml_type_size(src->type);
if (src->nb[0] == src_type_size) {
// src0 is contigous on first dimension, copy by rows
int64_t rows_num = ggml_nrows(src);
aclrtlaunch_ascendc_dup_by_rows_fp32(
rows_num, ctx.stream(), src->data, dst->data,
((ggml_tensor*)src->extra)->ne,
((ggml_tensor*)src->extra)->nb,
((ggml_tensor*)dst->extra)->ne,
((ggml_tensor*)dst->extra)->nb);
return;
}
GGML_ABORT("fatal error");
} else {
// TODO: dst not contiguous
GGML_ABORT("fatal error");
}
}
if (dst->type == GGML_TYPE_F16) {
if (ggml_are_same_shape(src, dst)) {
cann_copy(ctx, acl_src, acl_dst);
ACL_CHECK(aclDestroyTensor(acl_src));
ACL_CHECK(aclDestroyTensor(acl_dst));
return;
}
if (ggml_is_contiguous(dst)) {
const size_t src_type_size = ggml_type_size(src->type);
if (src->nb[0] == src_type_size) {
// src0 is contigous on first dimension, copy by rows
int64_t rows_num = ggml_nrows(src);
aclrtlaunch_ascendc_dup_by_rows_fp32_to_fp16(
rows_num, ctx.stream(), src->data, dst->data,
((ggml_tensor*)src->extra)->ne,
((ggml_tensor*)src->extra)->nb,
((ggml_tensor*)dst->extra)->ne,
((ggml_tensor*)dst->extra)->nb);
return;
}
GGML_ABORT("fatal error");
}
}
// TODO
GGML_ABORT("fatal error");
} else {
if (ggml_are_same_shape(src, dst)) {
if (ggml_are_same_shape(src0, dst)) {
if (dst->type == src0->type) {
cann_copy(ctx, acl_src, acl_dst);
ACL_CHECK(aclDestroyTensor(acl_src));
ACL_CHECK(aclDestroyTensor(acl_dst));
} else {
aclnn_cast(ctx, acl_src, acl_dst, dst);
}
} else {
if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst)) {
if (dst->type == src0->type) {
size_t cpy_size = ggml_nbytes(dst);
ACL_CHECK(aclrtMemcpyAsync(
dst->data, cpy_size, src0->data, cpy_size,
ACL_MEMCPY_DEVICE_TO_DEVICE, ctx.stream()));
return;
} else {
ggml_cann_pool_alloc src_buffer_allocator(
ctx.pool(),
ggml_nelements(dst) * ggml_type_size(dst->type));
void* src_trans_buffer = src_buffer_allocator.get();
size_t src_trans_nb[GGML_MAX_DIMS];
src_trans_nb[0] = ggml_type_size(dst->type);
for (int i = 1; i < GGML_MAX_DIMS; i++) {
src_trans_nb[i] = src_trans_nb[i - 1] * src0->ne[i - 1];
}
aclTensor* src_trans_tensor = ggml_cann_create_tensor(
src_trans_buffer, ggml_cann_type_mapping(dst->type),
ggml_type_size(dst->type), src0->ne, src_trans_nb,
GGML_MAX_DIMS);
aclnn_cast(ctx, acl_src, src_trans_tensor, dst);
size_t cpy_size = ggml_nbytes(dst);
ACL_CHECK(aclrtMemcpyAsync(
dst->data, cpy_size, src_trans_buffer, cpy_size,
ACL_MEMCPY_DEVICE_TO_DEVICE, ctx.stream()));
ACL_CHECK(aclDestroyTensor(src_trans_tensor));
return;
}
} else if (ggml_is_contiguous(dst)) {
ggml_cann_pool_alloc src_buffer_allocator(
ctx.pool(), ggml_nelements(dst) * ggml_type_size(dst->type));
void* src_trans_buffer = src_buffer_allocator.get();
size_t src_trans_nb[GGML_MAX_DIMS];
src_trans_nb[0] = ggml_type_size(dst->type);
for (int i = 1; i < GGML_MAX_DIMS; i++) {
src_trans_nb[i] = src_trans_nb[i - 1] * src0->ne[i - 1];
}
aclTensor* src_trans_tensor = ggml_cann_create_tensor(
src_trans_buffer, ggml_cann_type_mapping(dst->type),
ggml_type_size(dst->type), src0->ne, src_trans_nb,
GGML_MAX_DIMS);
aclnn_cast(ctx, acl_src, src_trans_tensor, dst);
size_t cpy_size = ggml_nbytes(dst);
ACL_CHECK(aclrtMemcpyAsync(dst->data, cpy_size, src_trans_buffer,
cpy_size, ACL_MEMCPY_DEVICE_TO_DEVICE,
ctx.stream()));
ACL_CHECK(aclDestroyTensor(src_trans_tensor));
return;
} else {
GGML_ABORT("Unsupport dst is not tontiguous.");
}
GGML_ABORT("fatal error");
}
ACL_CHECK(aclDestroyTensor(acl_src));
ACL_CHECK(aclDestroyTensor(acl_dst));
}
#ifdef __cplusplus
@@ -1158,8 +1088,6 @@ void ggml_cann_rms_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
float eps;
memcpy(&eps, dst->op_params, sizeof(float));
GGML_ASSERT(eps > 0.0f);
uint64_t workspaceSize = 0;
aclOpExecutor* executor;
void* workspaceAddr = nullptr;
@@ -2378,85 +2306,168 @@ void ggml_cann_softmax(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
ACL_CHECK(aclDestroyTensor(tmp_mask_tensor));
}
void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
ggml_tensor* src0 = dst->src[0];
ggml_tensor* src1 = dst->src[1];
/**
* @brief Performs embedding operation on a 4D tensor using the CANN backend.
*
* This function extracts slices from the source tensor (`src_buffer`),
* index tensor (`index`), and destination tensor (`dst`), and performs an
* embedding operation on them. The embedding operation is applied by iterating
* over the last two dimensions of the source tensor, creating the necessary
* tensors for the source, index, and output, and executing the embedding operation.
*
* @param ctx The context for CANN backend operations.
* @param src_buffer The source buffer holding the data for the source tensor.
* @param src_ne The dimensions of the source tensor.
* @param src_nb The strides (byte offsets) of the source tensor.
* @param index The index tensor used in the embedding operation.
* @param dst The destination tensor where the result will be stored.
*/
static void aclnn_embedding_4d(ggml_backend_cann_context& ctx, void* src_buffer,
int64_t* src_ne, size_t* src_nb, ggml_tensor* index,
ggml_tensor* dst) {
for (int64_t i = 0; i < src_ne[3]; i++) {
for (int64_t j = 0; j < src_ne[2]; j++) {
// src
int64_t acl_src_ne[2] = {src_ne[0], src_ne[1]};
size_t acl_src_nb[2] = {src_nb[0], src_nb[1]};
aclTensor* acl_src_tensor = ggml_cann_create_tensor(
(char*)src_buffer + i * src_nb[3] + j * src_nb[2],
ggml_cann_type_mapping(dst->type), ggml_element_size(dst),
acl_src_ne, acl_src_nb, 2);
ggml_cann_pool_alloc src0_extra_allocator(ctx.pool(), sizeof(ggml_tensor));
ggml_cann_pool_alloc src1_extra_allocator(ctx.pool(), sizeof(ggml_tensor));
ggml_cann_pool_alloc dst_extra_allocator(ctx.pool(), sizeof(ggml_tensor));
src0->extra = src0_extra_allocator.get();
src1->extra = src1_extra_allocator.get();
dst->extra = dst_extra_allocator.get();
ACL_CHECK(aclrtMemcpyAsync(src0->extra, sizeof(ggml_tensor), src0,
sizeof(ggml_tensor), ACL_MEMCPY_HOST_TO_DEVICE,
ctx.stream()));
ACL_CHECK(aclrtMemcpyAsync(src1->extra, sizeof(ggml_tensor), src1,
sizeof(ggml_tensor), ACL_MEMCPY_HOST_TO_DEVICE,
ctx.stream()));
ACL_CHECK(aclrtMemcpyAsync(dst->extra, sizeof(ggml_tensor), dst,
sizeof(ggml_tensor), ACL_MEMCPY_HOST_TO_DEVICE,
ctx.stream()));
// index
int64_t acl_index_ne[1] = {index->ne[0]};
size_t acl_index_nb[1] = {index->nb[0]};
aclTensor* acl_index = ggml_cann_create_tensor(
(char*)index->data + i * index->nb[2] + j * index->nb[1],
ggml_cann_type_mapping(index->type), ggml_element_size(index),
acl_index_ne, acl_index_nb, 1);
// out
int64_t acl_out_ne[2] = {dst->ne[0], dst->ne[1]};
size_t acl_out_nb[2] = {dst->nb[0], dst->nb[1]};
aclTensor* acl_out = ggml_cann_create_tensor(
(char*)dst->data + i * dst->nb[3] + j * dst->nb[2],
ggml_cann_type_mapping(dst->type), ggml_element_size(dst),
acl_out_ne, acl_out_nb, 2);
uint64_t workspaceSize = 0;
aclOpExecutor* executor;
void* workspaceAddr = nullptr;
ACL_CHECK(aclnnEmbeddingGetWorkspaceSize(
acl_src_tensor, acl_index, acl_out, &workspaceSize, &executor));
if (workspaceSize > 0) {
ggml_cann_pool_alloc workspace_allocator(ctx.pool(),
workspaceSize);
workspaceAddr = workspace_allocator.get();
}
ACL_CHECK(aclnnEmbedding(workspaceAddr, workspaceSize, executor,
ctx.stream()));
ACL_CHECK(aclDestroyTensor(acl_src_tensor));
ACL_CHECK(aclDestroyTensor(acl_index));
ACL_CHECK(aclDestroyTensor(acl_out));
}
}
}
void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
ggml_tensor* src0 = dst->src[0]; // src
ggml_tensor* src1 = dst->src[1]; // index
switch (src0->type) {
case GGML_TYPE_F32: {
#ifdef ASCEND_310P
// Special operation for get_row_f32 kernel of 310P: clear the
// content of dest data buffer when row is not aligned to 32 bytes
if ((src0->ne[0] % 8) != 0) {
size_t dst_len = src1->ne[0] * src1->ne[1] * src1->ne[2] *
src0->ne[0] * ggml_type_size(GGML_TYPE_F32);
ACL_CHECK(aclrtMemset((char*)dst->data, dst_len, 0, dst_len));
}
#endif
aclrtlaunch_ascendc_get_row_f32(
24, ctx.stream(), src0->data, src1->data, dst->data,
((ggml_tensor*)src0->extra)->ne,
((ggml_tensor*)src0->extra)->nb,
((ggml_tensor*)src1->extra)->ne,
((ggml_tensor*)src1->extra)->nb, ((ggml_tensor*)dst->extra)->ne,
((ggml_tensor*)dst->extra)->nb);
aclnn_embedding_4d(ctx, src0->data, src0->ne, src0->nb, src1,
dst);
break;
}
case GGML_TYPE_F16: {
#ifdef ASCEND_310P
// Special operation for get_row_f16 kernel of 310P: clear the
// content of dest data buffer when row is not aligned to 32 bytes
if ((src0->ne[0] % 16) != 0) {
size_t dst_len =
src1->ne[0] * src1->ne[1] * src1->ne[2] * src0->ne[0] *
ggml_type_size(
GGML_TYPE_F32); // out is also f32, even input is f16
ACL_CHECK(aclrtMemset((char*)dst->data, dst_len, 0, dst_len));
aclTensor* acl_src0 = ggml_cann_create_tensor(src0);
ggml_cann_pool_alloc src_buffer_allocator(
ctx.pool(), ggml_nelements(src0) * sizeof(float_t));
void* src_trans_buffer = src_buffer_allocator.get();
size_t src_trans_nb[GGML_MAX_DIMS];
src_trans_nb[0] = sizeof(float_t);
for (int i = 1; i < GGML_MAX_DIMS; i++) {
src_trans_nb[i] = src_trans_nb[i - 1] * src0->ne[i - 1];
}
#endif
aclrtlaunch_ascendc_get_row_f16(
24, ctx.stream(), src0->data, src1->data, dst->data,
((ggml_tensor*)src0->extra)->ne,
((ggml_tensor*)src0->extra)->nb,
((ggml_tensor*)src1->extra)->ne,
((ggml_tensor*)src1->extra)->nb, ((ggml_tensor*)dst->extra)->ne,
((ggml_tensor*)dst->extra)->nb);
aclTensor* src_trans_tensor = ggml_cann_create_tensor(
src_trans_buffer, ACL_FLOAT, ggml_type_size(dst->type),
src0->ne, src_trans_nb, GGML_MAX_DIMS);
aclnn_cast(ctx, acl_src0, src_trans_tensor, dst);
aclnn_embedding_4d(ctx, src_trans_buffer, src0->ne,
src_trans_nb, src1, dst);
ACL_CHECK(aclDestroyTensor(acl_src0));
ACL_CHECK(aclDestroyTensor(src_trans_tensor));
break;
}
case GGML_TYPE_Q4_0:
aclrtlaunch_ascendc_get_row_q4_0(
24, ctx.stream(), src0->data, src1->data, dst->data,
((ggml_tensor*)src0->extra)->ne,
((ggml_tensor*)src1->extra)->ne,
((ggml_tensor*)src1->extra)->nb, ((ggml_tensor*)dst->extra)->ne,
((ggml_tensor*)dst->extra)->nb);
break;
case GGML_TYPE_Q8_0:
aclrtlaunch_ascendc_get_row_q8_0(
24, ctx.stream(), src0->data, src1->data, dst->data,
((ggml_tensor*)src0->extra)->ne,
((ggml_tensor*)src1->extra)->ne,
((ggml_tensor*)src1->extra)->nb, ((ggml_tensor*)dst->extra)->ne,
((ggml_tensor*)dst->extra)->nb);
case GGML_TYPE_Q8_0: {
// add 1 dim for bcast mul.
size_t weight_nb[GGML_MAX_DIMS + 1], scale_nb[GGML_MAX_DIMS + 1],
dequant_nb[GGML_MAX_DIMS + 1];
int64_t weight_ne[GGML_MAX_DIMS + 1], scale_ne[GGML_MAX_DIMS + 1],
*dequant_ne;
int64_t scale_offset = 0;
// [3,4,5,64] -> [3,4,5,2,32]
weight_ne[0] = QK8_0;
weight_ne[1] = src0->ne[0] / QK8_0;
weight_nb[0] = sizeof(int8_t);
weight_nb[1] = weight_nb[0] * weight_ne[0];
for (int i = 2; i < GGML_MAX_DIMS + 1; i++) {
weight_ne[i] = src0->ne[i - 1];
weight_nb[i] = weight_nb[i - 1] * weight_ne[i - 1];
}
// [3,4,5,64] -> [3,4,5,2,1]
scale_ne[0] = 1;
scale_ne[1] = src0->ne[0] / QK8_0;
scale_nb[0] = sizeof(uint16_t);
scale_nb[1] = scale_nb[0] * scale_ne[0];
for (int i = 2; i < GGML_MAX_DIMS + 1; i++) {
scale_ne[i] = src0->ne[i - 1];
scale_nb[i] = scale_nb[i - 1] * scale_ne[i - 1];
}
// [3,4,5,64] -> [3,4,5,2,32]
dequant_ne = weight_ne;
dequant_nb[0] = sizeof(float_t);
for (int i = 1; i < GGML_MAX_DIMS + 1; i++) {
dequant_nb[i] = dequant_nb[i - 1] * dequant_ne[i - 1];
}
scale_offset = ggml_nelements(src0) * sizeof(int8_t);
ggml_cann_pool_alloc dequant_buffer_allocator(
ctx.pool(), ggml_nelements(src0) * sizeof(float_t));
aclTensor* acl_weight_tensor = ggml_cann_create_tensor(
src0->data, ACL_INT8, sizeof(int8_t), weight_ne, weight_nb,
GGML_MAX_DIMS + 1);
aclTensor* acl_scale_tensor = ggml_cann_create_tensor(
src0->data, ACL_FLOAT16, sizeof(float16_t), scale_ne, scale_nb,
GGML_MAX_DIMS + 1, ACL_FORMAT_ND, scale_offset);
aclTensor* dequant_tensor = ggml_cann_create_tensor(
dequant_buffer_allocator.get(), ACL_FLOAT, sizeof(float_t),
dequant_ne, dequant_nb, GGML_MAX_DIMS + 1);
aclnn_mul(ctx, acl_weight_tensor, acl_scale_tensor, dequant_tensor);
dequant_nb[0] = sizeof(float_t);
dequant_ne = src0->ne;
for (int i = 1; i < GGML_MAX_DIMS; i++) {
dequant_nb[i] = dequant_nb[i - 1] * src0->ne[i - 1];
}
aclnn_embedding_4d(ctx, dequant_buffer_allocator.get(),
dequant_ne, dequant_nb, src1, dst);
ACL_CHECK(aclDestroyTensor(dequant_tensor));
break;
}
default:
GGML_ABORT("fatal error");
GGML_ABORT("Unsupported tensor type for GGML_OP_GET_ROWS");
break;
}
}
@@ -2797,8 +2808,8 @@ static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
ACL_CHECK(aclnnWeightQuantBatchMatmulV2GetWorkspaceSize(
acl_input_tensor, acl_weight_tensor, acl_scale_tensor, nullptr,
nullptr, nullptr, nullptr, antiquantGroupSize, acl_output_tensor,
&workspaceSize, &executor));
nullptr, nullptr, nullptr, antiquantGroupSize,
acl_output_tensor, &workspaceSize, &executor));
if (workspaceAddr == nullptr) {
workspaceAddr = workspace_allocator.alloc(workspaceSize);
}
@@ -3137,7 +3148,7 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
// TODO: use ascendc
// Only test with LLAMA model.
ggml_tensor* src0 = dst->src[0]; // input
ggml_tensor* src2 = dst->src[2]; // freq_factors
// ggml_tensor* src2 = dst->src[2]; // freq_factors, not used now.
// param
float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
-6
View File
@@ -535,9 +535,6 @@ template <aclnnStatus getWorkspaceSize(const aclTensor*, aclTensor*, uint64_t*,
void ggml_cann_activation(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
ggml_tensor* src = dst->src[0];
GGML_ASSERT(src->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
aclTensor* acl_src = ggml_cann_create_tensor(src);
aclTensor* acl_dst = ggml_cann_create_tensor(dst);
@@ -566,9 +563,6 @@ template <aclnnStatus getWorkspaceSize(const aclTensor*, const aclTensor*,
void ggml_cann_activation(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
ggml_tensor* src = dst->src[0];
GGML_ASSERT(src->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
aclTensor* acl_src = ggml_cann_create_tensor(src);
aclTensor* acl_dst = ggml_cann_create_tensor(dst);
+33 -21
View File
@@ -1458,11 +1458,6 @@ static void ggml_backend_cann_free(ggml_backend_t backend) {
ACL_CHECK(aclrtSynchronizeDevice());
ACL_CHECK(aclrtResetDevice(cann_ctx->device));
// finalize when last backend freed.
if (cann_ctx->device == ggml_backend_cann_get_device_count() - 1) {
ACL_CHECK(aclFinalize());
}
delete cann_ctx;
delete backend;
}
@@ -1688,11 +1683,14 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
}
case GGML_OP_MUL_MAT: {
switch (op->src[0]->type) {
case GGML_TYPE_Q8_0:
case GGML_TYPE_F16:
case GGML_TYPE_F32:
case GGML_TYPE_Q4_0:
return true;
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q4_0:
// only support contiguous for quantized types.
return ggml_is_contiguous(op->src[0]) &&
ggml_is_contiguous(op->src[1]);
default:
return false;
}
@@ -1704,7 +1702,6 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
switch (op->src[0]->type) {
case GGML_TYPE_F32:
case GGML_TYPE_F16:
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q8_0:
return true;
default:
@@ -1712,16 +1709,21 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
}
} break;
case GGML_OP_CPY: {
switch (op->type) {
case GGML_TYPE_F32:
case GGML_TYPE_F16:
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q4_0:
return true;
default:
return false;
ggml_tensor *src = op->src[0];
if ((op->type != GGML_TYPE_F32 && op->type != GGML_TYPE_F16) ||
(src->type != GGML_TYPE_F32 &&
src->type != GGML_TYPE_F16)) {
// only support F32 and F16.
return false;
}
}
if (!ggml_are_same_shape(op, src) && !ggml_is_contiguous(op)) {
// unsupport dst is not contiguous.
return false;
}
return true;
} break;
case GGML_OP_CONT: {
// TODO: support GGML_TYPE_BF16
switch (op->src[0]->type) {
@@ -1734,13 +1736,14 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
}
case GGML_OP_ROPE: {
// TODO: with ops-test v == 1
float * ext_factor = (float*)((int32_t*)op->op_params + 7);
float ext_factor = 0.0f;
memcpy(&ext_factor, (const float *) op->op_params + 7, sizeof(float));
// TODO: n_dims <= ne0
if (op->src[0]->ne[0] != op->op_params[1]) {
return false;
}
// TODO: ext_factor != 0
if (*ext_factor != 0) {
if (ext_factor != 0) {
return false;
}
@@ -1762,9 +1765,19 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
}
return true;
}
case GGML_OP_POOL_2D: {
const int32_t * opts = (const int32_t *) op->op_params;
const int k0 = opts[1];
const int k1 = opts[2];
const int p0 = opts[5];
const int p1 = opts[6];
// value of paddingH should be at most half of kernelH
// value of paddingW should be at most half of kernelW
return (p0 <= (k0 / 2)) && (p1 <= (k1 / 2));
}
case GGML_OP_DUP:
case GGML_OP_IM2COL:
case GGML_OP_CONCAT:
case GGML_OP_DUP:
case GGML_OP_REPEAT:
case GGML_OP_NONE:
case GGML_OP_RESHAPE:
@@ -1781,7 +1794,6 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
case GGML_OP_CLAMP:
case GGML_OP_DIAG_MASK_INF:
case GGML_OP_SOFT_MAX:
case GGML_OP_POOL_2D:
case GGML_OP_SUM_ROWS:
case GGML_OP_ARGSORT:
case GGML_OP_ACC:
-30
View File
@@ -1,30 +0,0 @@
file(GLOB SRC_FILES
get_row_f32.cpp
get_row_f16.cpp
get_row_q4_0.cpp
get_row_q8_0.cpp
quantize_f32_q8_0.cpp
quantize_f16_q8_0.cpp
quantize_float_to_q4_0.cpp
dup.cpp
)
set(ASCEND_CANN_PACKAGE_PATH ${CANN_INSTALL_DIR})
set(RUN_MODE "npu" CACHE STRING "run mode: npu/sim")
if(EXISTS ${ASCEND_CANN_PACKAGE_PATH}/compiler/tikcpp/ascendc_kernel_cmake)
set(ASCENDC_CMAKE_DIR ${ASCEND_CANN_PACKAGE_PATH}/compiler/tikcpp/ascendc_kernel_cmake)
elseif(EXISTS ${ASCEND_CANN_PACKAGE_PATH}/ascendc_devkit/tikcpp/samples/cmake)
set(ASCENDC_CMAKE_DIR ${ASCEND_CANN_PACKAGE_PATH}/ascendc_devkit/tikcpp/samples/cmake)
else()
message(FATAL_ERROR "ascendc_kernel_cmake does not exist, please check whether the compiler package is installed.")
endif()
include(${ASCENDC_CMAKE_DIR}/ascendc.cmake)
ascendc_library(ascendc_kernels STATIC
${SRC_FILES}
)
message(STATUS "CANN: compile ascend kernels witch SOC_TYPE:${SOC_TYPE}, SOC_VERSION:${SOC_VERSION}, compile macro:-D${SOC_TYPE_COMPILE_OPTION}.")
ascendc_compile_definitions(ascendc_kernels PRIVATE "-D${SOC_TYPE_COMPILE_OPTION}")
# ascendc_compile_definitions(ascendc_kernels PRIVATE -DASCENDC_DUMP)
@@ -1,19 +0,0 @@
#ifndef ASCENDC_KERNELS_H
#define ASCENDC_KERNELS_H
#include "aclrtlaunch_ascendc_get_row_f32.h"
#include "aclrtlaunch_ascendc_get_row_f16.h"
#include "aclrtlaunch_ascendc_get_row_q8_0.h"
#include "aclrtlaunch_ascendc_get_row_q4_0.h"
#include "aclrtlaunch_ascendc_quantize_f32_q8_0.h"
#include "aclrtlaunch_ascendc_quantize_f16_q8_0.h"
#include "aclrtlaunch_ascendc_quantize_f16_to_q4_0.h"
#include "aclrtlaunch_ascendc_quantize_f32_to_q4_0.h"
#include "aclrtlaunch_ascendc_dup_by_rows_fp16.h"
#include "aclrtlaunch_ascendc_dup_by_rows_fp32.h"
#include "aclrtlaunch_ascendc_dup_by_rows_fp32_to_fp16.h"
#include "aclrtlaunch_ascendc_dup_by_rows_fp16_to_fp32.h"
#endif // ASCENDC_KERNELS_H
-234
View File
@@ -1,234 +0,0 @@
#include "kernel_operator.h"
using namespace AscendC;
#define BUFFER_NUM 2
const int64_t SUPPORTED_MAX_DIM = 65535; // currently the limit of max block dim supportted by dup kernel is 65535template <typename SRC_T, typename DST_T>
template <typename SRC_T, typename DST_T>
class DupByRows {
public:
__aicore__ inline DupByRows() {}
__aicore__ inline void init(GM_ADDR src, GM_ADDR dst, int64_t *input_ne_ub,
size_t *input_nb_ub) {
/* Dup by rows when src is contigous on first dimension and dst is
contiguous, each kernel process one row.
*/
// Input has four dims.
int64_t op_block_num = GetBlockNum();
int64_t op_block_idx = GetBlockIdx();
// param
num_rows = input_ne_ub[1] * input_ne_ub[2] * input_ne_ub[3];
num_elem = input_ne_ub[0];
// index for (ne[1], ne[2], ne[3]): (idx_ne1, idx_ne2, idx_ne3)
idx_ne3 = op_block_idx / (input_ne_ub[1] * input_ne_ub[2]);
idx_ne2 = (op_block_idx - idx_ne3 * (input_ne_ub[1] * input_ne_ub[2]))
/ (input_ne_ub[1]);
idx_ne1 = op_block_idx - idx_ne3 * (input_ne_ub[1] * input_ne_ub[2])
- idx_ne2 * input_ne_ub[1];
// src may not contiguous in dim [1,2,3], so stride decited by ne&nb
src_stride = input_nb_ub[3] * idx_ne3 + input_nb_ub[2] * idx_ne2
+ input_nb_ub[1] * idx_ne1;
// dst is contiguous
dst_stride = op_block_idx * (input_ne_ub[0] * sizeof(DST_T));
src_gm.SetGlobalBuffer(reinterpret_cast<__gm__ SRC_T *>(src +
src_stride));
dst_gm.SetGlobalBuffer(reinterpret_cast<__gm__ DST_T *>(dst +
dst_stride));
pipe.InitBuffer(src_queue, BUFFER_NUM, (sizeof(SRC_T) * num_elem +
32 - 1) / 32 * 32);
pipe.InitBuffer(dst_queue, BUFFER_NUM, (sizeof(DST_T) * num_elem +
32 - 1) / 32 * 32);
}
__aicore__ inline void copy_in() {
LocalTensor<SRC_T> src_local = src_queue.AllocTensor<SRC_T>();
const size_t elem_per_block = 32 / sizeof(SRC_T);
size_t tail = num_elem % elem_per_block;
size_t cpy_elements_len = tail > 0 ? num_elem + 1 : num_elem;
DataCopy(src_local, src_gm, cpy_elements_len);
src_queue.EnQue(src_local);
}
__aicore__ inline void copy_out() {
LocalTensor<DST_T> dst_local = dst_queue.DeQue<DST_T>();
#ifdef ASCEND_310P
const size_t elem_per_block = 32 / sizeof(DST_T);
size_t tail = num_elem % elem_per_block;
size_t len = num_elem & ~(elem_per_block - 1);
if (len > 0) {
DataCopy(dst_gm, dst_local, len);
}
if(tail != 0) {
for (size_t i = tail; i < elem_per_block; i++) {
dst_local[len + i].SetValue(0, 0);
}
SetAtomicAdd<float>();
DataCopy(dst_gm[len], dst_local[len], elem_per_block);
SetAtomicNone();
}
#else
DataCopyExtParams dataCopyParams;
dataCopyParams.blockCount = 1;
dataCopyParams.blockLen = num_elem * sizeof(DST_T);
DataCopyPad(dst_gm, dst_local, dataCopyParams);
#endif
dst_queue.FreeTensor(dst_local);
}
__aicore__ inline void dup() {
// main process, copy one row data from src to dst.
copy_in();
LocalTensor<SRC_T> src_local = src_queue.DeQue<SRC_T>();
LocalTensor<DST_T> dst_local = dst_queue.AllocTensor<DST_T>();
int32_t BLOCK_NUM = 32 / sizeof(DST_T);
DataCopy(dst_local, src_local, (num_elem + BLOCK_NUM - 1)
/ BLOCK_NUM * BLOCK_NUM);
dst_queue.EnQue<DST_T>(dst_local);
src_queue.FreeTensor(src_local);
copy_out();
}
__aicore__ inline void dup_with_cast() {
// main process, copy one row data from src to dst.
// cast dtype from src to dst.
copy_in();
LocalTensor<SRC_T> src_local = src_queue.DeQue<SRC_T>();
LocalTensor<DST_T> dst_local = dst_queue.AllocTensor<DST_T>();
Cast(dst_local, src_local, RoundMode::CAST_NONE, num_elem);
dst_queue.EnQue<DST_T>(dst_local);
src_queue.FreeTensor(src_local);
copy_out();
}
private:
TPipe pipe;
GlobalTensor<SRC_T> src_gm;
GlobalTensor<DST_T> dst_gm;
int64_t num_rows;
int64_t num_elem;
int64_t idx_ne3;
int64_t idx_ne2;
int64_t idx_ne1;
int64_t src_stride;
int64_t dst_stride;
TQue<QuePosition::VECIN, BUFFER_NUM> src_queue;
TQue<QuePosition::VECOUT, BUFFER_NUM> dst_queue;
};
template <typename T>
__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
auto gm_ptr = (__gm__ uint8_t *)gm;
auto ub_ptr = (uint8_t *)(ub);
for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
*ub_ptr = *gm_ptr;
}
}
extern "C" __global__ __aicore__ void ascendc_dup_by_rows_fp16(
GM_ADDR src_gm,
GM_ADDR dst_gm,
GM_ADDR input_ne_gm,
GM_ADDR input_nb_gm,
GM_ADDR output_ne_gm,
GM_ADDR output_nb_gm) {
int64_t input_ne_ub[4];
size_t input_nb_ub[4];
int64_t output_ne_ub[4];
size_t output_nb_ub[4];
copy_to_ub(input_ne_gm, input_ne_ub, 32);
copy_to_ub(input_nb_gm, input_nb_ub, 32);
copy_to_ub(output_ne_gm, output_ne_ub, 32);
copy_to_ub(output_nb_gm, output_nb_ub, 32);
DupByRows<half, half> op;
op.init(src_gm, dst_gm, input_ne_ub, input_nb_ub);
op.dup();
}
extern "C" __global__ __aicore__ void ascendc_dup_by_rows_fp32(
GM_ADDR src_gm,
GM_ADDR dst_gm,
GM_ADDR input_ne_gm,
GM_ADDR input_nb_gm,
GM_ADDR output_ne_gm,
GM_ADDR output_nb_gm) {
int64_t input_ne_ub[4];
size_t input_nb_ub[4];
int64_t output_ne_ub[4];
size_t output_nb_ub[4];
copy_to_ub(input_ne_gm, input_ne_ub, 32);
copy_to_ub(input_nb_gm, input_nb_ub, 32);
copy_to_ub(output_ne_gm, output_ne_ub, 32);
copy_to_ub(output_nb_gm, output_nb_ub, 32);
DupByRows<float, float> op;
op.init(src_gm, dst_gm, input_ne_ub, input_nb_ub);
op.dup();
}
extern "C" __global__ __aicore__ void ascendc_dup_by_rows_fp32_to_fp16(
GM_ADDR src_gm,
GM_ADDR dst_gm,
GM_ADDR input_ne_gm,
GM_ADDR input_nb_gm,
GM_ADDR output_ne_gm,
GM_ADDR output_nb_gm) {
int64_t input_ne_ub[4];
size_t input_nb_ub[4];
int64_t output_ne_ub[4];
size_t output_nb_ub[4];
copy_to_ub(input_ne_gm, input_ne_ub, 32);
copy_to_ub(input_nb_gm, input_nb_ub, 32);
copy_to_ub(output_ne_gm, output_ne_ub, 32);
copy_to_ub(output_nb_gm, output_nb_ub, 32);
DupByRows<float, half> op;
op.init(src_gm, dst_gm, input_ne_ub, input_nb_ub);
op.dup_with_cast();
}
extern "C" __global__ __aicore__ void ascendc_dup_by_rows_fp16_to_fp32(
GM_ADDR src_gm,
GM_ADDR dst_gm,
GM_ADDR input_ne_gm,
GM_ADDR input_nb_gm,
GM_ADDR output_ne_gm,
GM_ADDR output_nb_gm) {
// copy params from gm to ub.
int64_t input_ne_ub[4];
size_t input_nb_ub[4];
int64_t output_ne_ub[4];
size_t output_nb_ub[4];
copy_to_ub(input_ne_gm, input_ne_ub, 32);
copy_to_ub(input_nb_gm, input_nb_ub, 32);
copy_to_ub(output_ne_gm, output_ne_ub, 32);
copy_to_ub(output_nb_gm, output_nb_ub, 32);
DupByRows<half, float> op;
op.init(src_gm, dst_gm, input_ne_ub, input_nb_ub);
op.dup_with_cast();
}
-197
View File
@@ -1,197 +0,0 @@
#include "kernel_operator.h"
// optimize me. Use template to avoid copy code.
using namespace AscendC;
#define BUFFER_NUM 2
class GET_ROW_F16 {
public:
__aicore__ inline GET_ROW_F16() {}
__aicore__ inline void init(GM_ADDR input, GM_ADDR indices, GM_ADDR output,
int64_t *input_ne_ub, size_t *input_nb_ub,
int64_t *indices_ne_ub, size_t *indices_nb_ub,
int64_t *output_ne_ub, size_t *output_nb_ub) {
// TODO, use template for F16/f32
int64_t op_block_num = GetBlockNum();
op_block_idx = GetBlockIdx();
for (int i = 0; i < 4; i++) {
input_ne[i] = input_ne_ub[i];
input_stride[i] = input_nb_ub[i] / input_nb_ub[0];
indices_ne[i] = indices_ne_ub[i];
indices_stride[i] = indices_nb_ub[i] / indices_nb_ub[0];
output_ne[i] = output_ne_ub[i];
output_stride[i] = output_nb_ub[i] / output_nb_ub[0];
}
// Indices has two dims. n_elements = all rows should get.
// dr, all rows should this thread get.
uint64_t n_elements =
indices_ne[0] * indices_ne[1] * indices_ne[2] * indices_ne[3];
dr = n_elements / op_block_num;
uint64_t tails = n_elements % op_block_num;
if (op_block_idx < tails) {
dr += 1;
ir = dr * op_block_idx;
} else {
ir = dr * op_block_idx + tails;
}
input_gm.SetGlobalBuffer((__gm__ half *)input);
indices_gm.SetGlobalBuffer((__gm__ int32_t *)indices);
output_gm.SetGlobalBuffer((__gm__ float *)output);
uint64_t input_local_buffer_size = ((input_ne[0] * sizeof(half) + 31)
& ~31);
uint64_t output_local_buffer_size = ((input_ne[0] * sizeof(float) + 31)
& ~31);
local_buffer_elems = input_local_buffer_size / sizeof(half);
// TODO, consider long row that can't put in UB.
// All data should asign to 32. It's ok because all data is align to 32.
pipe.InitBuffer(input_queue, BUFFER_NUM, input_local_buffer_size);
pipe.InitBuffer(output_queue, BUFFER_NUM, output_local_buffer_size);
}
__aicore__ inline void copy_in(uint32_t offset, size_t len) {
size_t origin_len = len;
LocalTensor<half> input_local = input_queue.AllocTensor<half>();
const size_t elem_per_block = 32 / sizeof(half);
size_t tail = len % elem_per_block;
len = len & ~(elem_per_block - 1);
if(tail != 0) {
len += elem_per_block;
}
DataCopy(input_local, input_gm[offset], len);
input_queue.EnQue(input_local);
}
__aicore__ inline void copy_out(uint32_t offset, size_t len) {
LocalTensor<float> output_local = output_queue.DeQue<float>();
const size_t elem_per_block = 32 / sizeof(float);
size_t tail = len % elem_per_block;
len = len & ~(elem_per_block - 1);
if (len > 0) {
DataCopy(output_gm[offset], output_local, len);
}
if(tail != 0) {
#ifdef ASCEND_310P
for (size_t i = tail; i < elem_per_block; i++) {
output_local[len + i].SetValue(0, 0);
}
SetAtomicAdd<float>();
DataCopy(output_gm[offset + len], output_local[len], elem_per_block);
SetAtomicNone();
#else
DataCopyExtParams dataCopyParams;
dataCopyParams.blockCount = 1;
dataCopyParams.blockLen = tail * sizeof(float);
DataCopyPad(output_gm[offset + len], output_local[len],
dataCopyParams);
#endif
}
output_queue.FreeTensor(output_local);
}
__aicore__ inline void calculate_row(int64_t idx) {
const int64_t indices_ne2_idx = idx / (indices_ne[0] * indices_ne[1]);
const int64_t indices_ne1_idx =
(idx - indices_ne2_idx * indices_ne[0] * indices_ne[1]) /
indices_ne[0];
const int64_t indices_ne0_idx =
(idx - indices_ne2_idx * indices_ne[0] * indices_ne[1] -
indices_ne1_idx * indices_ne[0]);
const int64_t indices_offset = indices_ne0_idx * indices_stride[0] +
indices_ne1_idx * indices_stride[1] +
indices_ne2_idx * indices_stride[2];
const int32_t selected_row_idx = indices_gm.GetValue(indices_offset);
const int64_t input_offset = selected_row_idx * input_stride[1] +
indices_ne1_idx * input_stride[2] +
indices_ne2_idx * input_stride[3];
const int64_t output_offset = indices_ne0_idx * output_stride[1] +
indices_ne1_idx * output_stride[2] +
indices_ne2_idx * output_stride[3];
copy_in(input_offset, input_ne[0]);
LocalTensor<half> input_local = input_queue.DeQue<half>();
LocalTensor<float> output_local = output_queue.AllocTensor<float>();
Cast(output_local, input_local, RoundMode::CAST_NONE,
local_buffer_elems);
output_queue.EnQue(output_local);
copy_out(output_offset, input_ne[0]);
input_queue.FreeTensor(input_local);
}
__aicore__ inline void calculate() {
for (int64_t i = ir; i < ir + dr; i++) {
calculate_row(i);
}
}
private:
int64_t input_ne[4];
size_t input_stride[4];
int64_t indices_ne[4];
size_t indices_stride[4];
int64_t output_ne[4];
size_t output_stride[4];
size_t local_buffer_elems;
int64_t ir;
int64_t dr;
TPipe pipe;
GlobalTensor<half> input_gm;
GlobalTensor<int32_t> indices_gm;
GlobalTensor<float> output_gm;
TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
int64_t op_block_idx;
};
template <typename T>
__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
auto gm_ptr = (__gm__ uint8_t *)gm;
auto ub_ptr = (uint8_t *)(ub);
for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
*ub_ptr = *gm_ptr;
}
}
extern "C" __global__ __aicore__ void ascendc_get_row_f16(
GM_ADDR input_gm, GM_ADDR indices_gm, GM_ADDR output_gm,
GM_ADDR input_ne_gm, GM_ADDR input_nb_gm, GM_ADDR indices_ne_gm,
GM_ADDR indices_nb_gm, GM_ADDR output_ne_gm, GM_ADDR output_nb_gm) {
int64_t input_ne_ub[4];
size_t input_nb_ub[4];
int64_t indices_ne_ub[4];
size_t indices_nb_ub[4];
int64_t output_ne_ub[4];
size_t output_nb_ub[4];
copy_to_ub(input_ne_gm, input_ne_ub, 32);
copy_to_ub(input_nb_gm, input_nb_ub, 32);
copy_to_ub(indices_ne_gm, indices_ne_ub, 32);
copy_to_ub(indices_nb_gm, indices_nb_ub, 32);
copy_to_ub(output_ne_gm, output_ne_ub, 32);
copy_to_ub(output_nb_gm, output_nb_ub, 32);
GET_ROW_F16 op;
op.init(input_gm, indices_gm, output_gm, input_ne_ub, input_nb_ub,
indices_ne_ub, indices_nb_ub, output_ne_ub, output_nb_ub);
op.calculate();
}
-190
View File
@@ -1,190 +0,0 @@
#include "kernel_operator.h"
// optimize me. Use template to avoid copy code.
using namespace AscendC;
#define BUFFER_NUM 2
class GET_ROW_F32 {
public:
__aicore__ inline GET_ROW_F32() {}
__aicore__ inline void init(GM_ADDR input, GM_ADDR indices, GM_ADDR output,
int64_t *input_ne_ub, size_t *input_nb_ub,
int64_t *indices_ne_ub, size_t *indices_nb_ub,
int64_t *output_ne_ub, size_t *output_nb_ub) {
int64_t op_block_num = GetBlockNum();
op_block_idx = GetBlockIdx();
for (int i = 0; i < 4; i++) {
input_ne[i] = input_ne_ub[i];
input_stride[i] = input_nb_ub[i] / input_nb_ub[0];
indices_ne[i] = indices_ne_ub[i];
indices_stride[i] = indices_nb_ub[i] / indices_nb_ub[0];
output_ne[i] = output_ne_ub[i];
output_stride[i] = output_nb_ub[i] / output_nb_ub[0];
}
// Indices has two dims. n_elements = all rows should get.
// dr, all rows should this thread get.
uint64_t n_elements =
indices_ne[0] * indices_ne[1] * indices_ne[2] * indices_ne[3];
dr = n_elements / op_block_num;
uint64_t tails = n_elements % op_block_num;
if (op_block_idx < tails) {
dr += 1;
ir = dr * op_block_idx;
} else {
ir = dr * op_block_idx + tails;
}
input_gm.SetGlobalBuffer((__gm__ float *)input);
indices_gm.SetGlobalBuffer((__gm__ int32_t *)indices);
output_gm.SetGlobalBuffer((__gm__ float *)output);
uint64_t local_buffer_size = ((input_ne[0] * sizeof(float) + 31) & ~31);
local_buffer_elems = local_buffer_size / sizeof(float);
// TODO, consider long row that can't put in UB.
// All data should asign to 32. It's ok because all data is align to 32.
pipe.InitBuffer(input_queue, BUFFER_NUM, local_buffer_size);
pipe.InitBuffer(output_queue, BUFFER_NUM, local_buffer_size);
}
__aicore__ inline void copy_in(uint32_t offset, size_t len) {
LocalTensor<float> input_local = input_queue.AllocTensor<float>();
const size_t elem_per_block = 32 / sizeof(float);
size_t tail = len % elem_per_block;
len = len & ~(elem_per_block - 1);
if(tail != 0) {
len += elem_per_block;
}
DataCopy(input_local, input_gm[offset], len);
input_queue.EnQue(input_local);
}
__aicore__ inline void copy_out(uint32_t offset, size_t len) {
LocalTensor<float> output_local = output_queue.DeQue<float>();
const size_t elem_per_block = 32 / sizeof(float);
size_t tail = len % elem_per_block;
len = len & ~(elem_per_block - 1);
if (len > 0) {
DataCopy(output_gm[offset], output_local, len);
}
if(tail != 0) {
#ifdef ASCEND_310P
for (size_t i = tail; i < elem_per_block; i++) {
output_local[len + i].SetValue(0, 0);
}
SetAtomicAdd<float>();
DataCopy(output_gm[offset + len], output_local[len], elem_per_block);
SetAtomicNone();
#else
DataCopyExtParams dataCopyParams;
dataCopyParams.blockCount = 1;
dataCopyParams.blockLen = tail * sizeof(float);
DataCopyPad(output_gm[offset + len], output_local[len],
dataCopyParams);
#endif
}
output_queue.FreeTensor(output_local);
}
__aicore__ inline void calculate_row(int64_t idx) {
const int64_t indices_ne2_idx = idx / (indices_ne[0] * indices_ne[1]);
const int64_t indices_ne1_idx =
(idx - indices_ne2_idx * indices_ne[0] * indices_ne[1]) /
indices_ne[0];
const int64_t indices_ne0_idx =
(idx - indices_ne2_idx * indices_ne[0] * indices_ne[1] -
indices_ne1_idx * indices_ne[0]);
const int64_t indices_offset = indices_ne0_idx * indices_stride[0] +
indices_ne1_idx * indices_stride[1] +
indices_ne2_idx * indices_stride[2];
const int32_t selected_row_idx = indices_gm.GetValue(indices_offset);
const int64_t input_offset = selected_row_idx * input_stride[1] +
indices_ne1_idx * input_stride[2] +
indices_ne2_idx * input_stride[3];
const int64_t output_offset = indices_ne0_idx * output_stride[1] +
indices_ne1_idx * output_stride[2] +
indices_ne2_idx * output_stride[3];
copy_in(input_offset, input_ne[0]);
LocalTensor<float> input_local = input_queue.DeQue<float>();
LocalTensor<float> output_local = output_queue.AllocTensor<float>();
DataCopy(output_local, input_local, local_buffer_elems);
output_queue.EnQue(output_local);
copy_out(output_offset, input_ne[0]);
input_queue.FreeTensor(input_local);
}
__aicore__ inline void calculate() {
for (int64_t i = ir; i < ir + dr; i++) {
calculate_row(i);
}
}
private:
int64_t input_ne[4];
size_t input_stride[4];
int64_t indices_ne[4];
size_t indices_stride[4];
int64_t output_ne[4];
size_t output_stride[4];
size_t local_buffer_elems;
int64_t ir;
int64_t dr;
TPipe pipe;
GlobalTensor<float> input_gm;
GlobalTensor<int32_t> indices_gm;
GlobalTensor<float> output_gm;
TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
int64_t op_block_idx;
};
template <typename T>
__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
auto gm_ptr = (__gm__ uint8_t *)gm;
auto ub_ptr = (uint8_t *)(ub);
for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
*ub_ptr = *gm_ptr;
}
}
extern "C" __global__ __aicore__ void ascendc_get_row_f32(
GM_ADDR input_gm, GM_ADDR indices_gm, GM_ADDR output_gm,
GM_ADDR input_ne_gm, GM_ADDR input_nb_gm, GM_ADDR indices_ne_gm,
GM_ADDR indices_nb_gm, GM_ADDR output_ne_gm, GM_ADDR output_nb_gm) {
int64_t input_ne_ub[4];
size_t input_nb_ub[4];
int64_t indices_ne_ub[4];
size_t indices_nb_ub[4];
int64_t output_ne_ub[4];
size_t output_nb_ub[4];
copy_to_ub(input_ne_gm, input_ne_ub, 32);
copy_to_ub(input_nb_gm, input_nb_ub, 32);
copy_to_ub(indices_ne_gm, indices_ne_ub, 32);
copy_to_ub(indices_nb_gm, indices_nb_ub, 32);
copy_to_ub(output_ne_gm, output_ne_ub, 32);
copy_to_ub(output_nb_gm, output_nb_ub, 32);
GET_ROW_F32 op;
op.init(input_gm, indices_gm, output_gm, input_ne_ub, input_nb_ub,
indices_ne_ub, indices_nb_ub, output_ne_ub, output_nb_ub);
op.calculate();
}
-204
View File
@@ -1,204 +0,0 @@
#include "kernel_operator.h"
// optimize me. Use template to avoid copy code.
using namespace AscendC;
#ifdef ASCEND_310P // 310P not support 4bit get row
extern "C" __global__ __aicore__ void ascendc_get_row_q4_0(
GM_ADDR input_gm, GM_ADDR indices_gm, GM_ADDR output_gm,
GM_ADDR input_ne_gm, GM_ADDR indices_ne_gm, GM_ADDR indices_nb_gm,
GM_ADDR output_ne_gm, GM_ADDR output_nb_gm) {
// let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed.
printf("Ascend310P not support 4bit get row.\n");
}
#else
#define BUFFER_NUM 2
#define QK4_0 32
class GET_ROW_Q4_0 {
public:
__aicore__ inline GET_ROW_Q4_0() {}
__aicore__ inline void init(GM_ADDR input, GM_ADDR indices, GM_ADDR output,
int64_t *input_ne_ub, int64_t *indices_ne_ub,
size_t *indices_nb_ub, int64_t *output_ne_ub,
size_t *output_nb_ub) {
int64_t op_block_num = GetBlockNum();
int64_t op_block_idx = GetBlockIdx();
for (int i = 0; i < 4; i++) {
input_ne[i] = input_ne_ub[i];
indices_ne[i] = indices_ne_ub[i];
indices_stride[i] = indices_nb_ub[i] / indices_nb_ub[0];
scale_ne[i] = input_ne_ub[i];
output_ne[i] = output_ne_ub[i];
output_stride[i] = output_nb_ub[i] / output_nb_ub[0];
}
// one scale for a group.
scale_ne[0] /= QK4_0;
input_stride[0] = 1;
scale_stride[0] = 1;
output_stride[0] = 1;
for (int i = 1; i < 4; i++) {
input_stride[i] = input_stride[i - 1] * input_ne[i - 1];
scale_stride[i] = scale_stride[i - 1] * scale_ne[i - 1];
}
group_size_in_row = input_ne[0] / QK4_0;
int64_t scale_offset = input_ne[0] * input_ne[1] * input_ne[2] *
input_ne[3] / 2;
// Indices has two dims. n_elements = all rows should get.
// dr, all rows should this thread get.
uint64_t n_elements =
indices_ne[0] * indices_ne[1] * indices_ne[2] * indices_ne[3];
dr = n_elements / op_block_num;
uint64_t tails = n_elements % op_block_num;
if (op_block_idx < tails) {
dr += 1;
ir = dr * op_block_idx;
} else {
ir = dr * op_block_idx + tails;
}
input_gm.SetGlobalBuffer((__gm__ int4b_t *)input);
scale_gm.SetGlobalBuffer((__gm__ half *)(input + scale_offset));
indices_gm.SetGlobalBuffer((__gm__ int32_t *)indices);
output_gm.SetGlobalBuffer((__gm__ float *)output);
pipe.InitBuffer(input_queue, BUFFER_NUM, QK4_0 * sizeof(int4b_t));
pipe.InitBuffer(cast_queue, BUFFER_NUM, QK4_0 * sizeof(half));
pipe.InitBuffer(output_queue, BUFFER_NUM, QK4_0 * sizeof(float));
}
__aicore__ inline void copy_in(uint32_t offset) {
LocalTensor<int4b_t> input_local = input_queue.AllocTensor<int4b_t>();
// 32 * sizeof(int4b_t) = 16, which is not aligned to 32, why no error?
DataCopy(input_local, input_gm[offset], QK4_0);
input_queue.EnQue(input_local);
}
__aicore__ inline void copy_out(uint32_t offset) {
LocalTensor<float> output_local = output_queue.DeQue<float>();
DataCopy(output_gm[offset], output_local, QK4_0);
output_queue.FreeTensor(output_local);
}
__aicore__ inline void calculate_group(int64_t idx, int64_t group) {
const int64_t indices_ne2_idx = idx / (indices_ne[0] * indices_ne[1]);
const int64_t indices_ne1_idx =
(idx - indices_ne2_idx * indices_ne[0] * indices_ne[1]) /
indices_ne[0];
const int64_t indices_ne0_idx =
(idx - indices_ne2_idx * indices_ne[0] * indices_ne[1] -
indices_ne1_idx * indices_ne[0]);
const int64_t indices_offset = indices_ne0_idx * indices_stride[0] +
indices_ne1_idx * indices_stride[1] +
indices_ne2_idx * indices_stride[2];
const int32_t selected_row_idx = indices_gm.GetValue(indices_offset);
const int64_t input_offset = selected_row_idx * input_stride[1] +
indices_ne1_idx * input_stride[2] +
indices_ne2_idx * input_stride[3] +
group * QK4_0;
const int64_t scale_offset = selected_row_idx * scale_stride[1] +
indices_ne1_idx * scale_stride[2] +
indices_ne2_idx * scale_stride[3] + group;
const int64_t output_offset = indices_ne0_idx * output_stride[1] +
indices_ne1_idx * output_stride[2] +
indices_ne2_idx * output_stride[3] +
group * QK4_0;
copy_in(input_offset);
LocalTensor<int4b_t> input_local = input_queue.DeQue<int4b_t>();
LocalTensor<half> cast_local = cast_queue.AllocTensor<half>();
LocalTensor<float> output_local = output_queue.AllocTensor<float>();
// TODO: cast more data to speed up.
Cast(cast_local, input_local, RoundMode::CAST_NONE, QK4_0);
Cast(output_local, cast_local, RoundMode::CAST_NONE, QK4_0);
// Only mul need compile by group.
half scale = scale_gm.GetValue(scale_offset);
Muls(output_local, output_local, (float)scale, QK4_0);
input_queue.FreeTensor(input_local);
cast_queue.FreeTensor(cast_local);
output_queue.EnQue(output_local);
copy_out(output_offset);
}
__aicore__ inline void calculate() {
for (int64_t i = ir; i < ir + dr; i++) {
for (int64_t j = 0; j < group_size_in_row; j++) {
calculate_group(i, j);
}
}
}
private:
int64_t input_ne[4];
size_t input_stride[4];
int64_t scale_ne[4];
size_t scale_stride[4];
int64_t indices_ne[4];
size_t indices_stride[4];
int64_t output_ne[4];
size_t output_stride[4];
int64_t ir;
int64_t dr;
int64_t group_size_in_row;
TPipe pipe;
GlobalTensor<int4b_t> input_gm;
GlobalTensor<half> scale_gm;
GlobalTensor<int32_t> indices_gm;
GlobalTensor<float> output_gm;
TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
TQue<QuePosition::VECIN, BUFFER_NUM> cast_queue;
};
template <typename T>
__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
auto gm_ptr = (__gm__ uint8_t *)gm;
auto ub_ptr = (uint8_t *)(ub);
for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
*ub_ptr = *gm_ptr;
}
}
extern "C" __global__ __aicore__ void ascendc_get_row_q4_0(
GM_ADDR input_gm, GM_ADDR indices_gm, GM_ADDR output_gm,
GM_ADDR input_ne_gm, GM_ADDR indices_ne_gm, GM_ADDR indices_nb_gm,
GM_ADDR output_ne_gm, GM_ADDR output_nb_gm) {
int64_t input_ne_ub[4];
int64_t indices_ne_ub[4];
size_t indices_nb_ub[4];
int64_t output_ne_ub[4];
size_t output_nb_ub[4];
copy_to_ub(input_ne_gm, input_ne_ub, 32);
copy_to_ub(indices_ne_gm, indices_ne_ub, 32);
copy_to_ub(indices_nb_gm, indices_nb_ub, 32);
copy_to_ub(output_ne_gm, output_ne_ub, 32);
copy_to_ub(output_nb_gm, output_nb_ub, 32);
GET_ROW_Q4_0 op;
op.init(input_gm, indices_gm, output_gm, input_ne_ub, indices_ne_ub,
indices_nb_ub, output_ne_ub, output_nb_ub);
op.calculate();
}
#endif // #ifdef ASCEND_310P
-191
View File
@@ -1,191 +0,0 @@
#include "kernel_operator.h"
// optimize me. Use template to avoid copy code.
using namespace AscendC;
#define BUFFER_NUM 2
#define QK8_0 32
class GET_ROW_Q8_0 {
public:
__aicore__ inline GET_ROW_Q8_0() {}
__aicore__ inline void init(GM_ADDR input, GM_ADDR indices, GM_ADDR output,
int64_t *input_ne_ub, int64_t *indices_ne_ub,
size_t *indices_nb_ub, int64_t *output_ne_ub,
size_t *output_nb_ub) {
int64_t op_block_num = GetBlockNum();
int64_t op_block_idx = GetBlockIdx();
for (int i = 0; i < 4; i++) {
input_ne[i] = input_ne_ub[i];
indices_ne[i] = indices_ne_ub[i];
indices_stride[i] = indices_nb_ub[i] / indices_nb_ub[0];
scale_ne[i] = input_ne_ub[i];
output_ne[i] = output_ne_ub[i];
output_stride[i] = output_nb_ub[i] / output_nb_ub[0];
}
// one scale for a group.
scale_ne[0] /= QK8_0;
input_stride[0] = 1;
scale_stride[0] = 1;
output_stride[0] = 1;
for (int i = 1; i < 4; i++) {
input_stride[i] = input_stride[i - 1] * input_ne[i - 1];
scale_stride[i] = scale_stride[i - 1] * scale_ne[i - 1];
}
group_size_in_row = input_ne[0] / QK8_0;
int64_t scale_offset = input_ne[0] * input_ne[1] * input_ne[2] *
input_ne[3] * sizeof(int8_t);
// Indices has two dims. n_elements = all rows should get.
// dr, all rows should this thread get.
uint64_t n_elements =
indices_ne[0] * indices_ne[1] * indices_ne[2] * indices_ne[3];
dr = n_elements / op_block_num;
uint64_t tails = n_elements % op_block_num;
if (op_block_idx < tails) {
dr += 1;
ir = dr * op_block_idx;
} else {
ir = dr * op_block_idx + tails;
}
input_gm.SetGlobalBuffer((__gm__ int8_t *)input);
scale_gm.SetGlobalBuffer((__gm__ half *)(input + scale_offset));
indices_gm.SetGlobalBuffer((__gm__ int32_t *)indices);
output_gm.SetGlobalBuffer((__gm__ float *)output);
pipe.InitBuffer(input_queue, BUFFER_NUM, QK8_0 * sizeof(int8_t));
pipe.InitBuffer(cast_queue, BUFFER_NUM, QK8_0 * sizeof(half));
pipe.InitBuffer(output_queue, BUFFER_NUM, QK8_0 * sizeof(float));
}
__aicore__ inline void copy_in(uint32_t offset) {
LocalTensor<int8_t> input_local = input_queue.AllocTensor<int8_t>();
DataCopy(input_local, input_gm[offset], QK8_0);
input_queue.EnQue(input_local);
}
__aicore__ inline void copy_out(uint32_t offset) {
LocalTensor<float> output_local = output_queue.DeQue<float>();
DataCopy(output_gm[offset], output_local, QK8_0);
output_queue.FreeTensor(output_local);
}
__aicore__ inline void calculate_group(int64_t idx, int64_t group) {
const int64_t indices_ne2_idx = idx / (indices_ne[0] * indices_ne[1]);
const int64_t indices_ne1_idx =
(idx - indices_ne2_idx * indices_ne[0] * indices_ne[1]) /
indices_ne[0];
const int64_t indices_ne0_idx =
(idx - indices_ne2_idx * indices_ne[0] * indices_ne[1] -
indices_ne1_idx * indices_ne[0]);
const int64_t indices_offset = indices_ne0_idx * indices_stride[0] +
indices_ne1_idx * indices_stride[1] +
indices_ne2_idx * indices_stride[2];
const int32_t selected_row_idx = indices_gm.GetValue(indices_offset);
const int64_t input_offset = selected_row_idx * input_stride[1] +
indices_ne1_idx * input_stride[2] +
indices_ne2_idx * input_stride[3] +
group * QK8_0;
const int64_t scale_offset = selected_row_idx * scale_stride[1] +
indices_ne1_idx * scale_stride[2] +
indices_ne2_idx * scale_stride[3] + group;
const int64_t output_offset = indices_ne0_idx * output_stride[1] +
indices_ne1_idx * output_stride[2] +
indices_ne2_idx * output_stride[3] +
group * QK8_0;
copy_in(input_offset);
LocalTensor<int8_t> input_local = input_queue.DeQue<int8_t>();
LocalTensor<half> cast_local = cast_queue.AllocTensor<half>();
LocalTensor<float> output_local = output_queue.AllocTensor<float>();
// TODO: cast more data to speed up.
Cast(cast_local, input_local, RoundMode::CAST_NONE, QK8_0);
Cast(output_local, cast_local, RoundMode::CAST_NONE, QK8_0);
// Only mul need compile by group.
half scale = scale_gm.GetValue(scale_offset);
Muls(output_local, output_local, (float)scale, QK8_0);
input_queue.FreeTensor(input_local);
cast_queue.FreeTensor(cast_local);
output_queue.EnQue(output_local);
copy_out(output_offset);
}
__aicore__ inline void calculate() {
for (int64_t i = ir; i < ir + dr; i++) {
for (int64_t j = 0; j < group_size_in_row; j++) {
calculate_group(i, j);
}
}
}
private:
int64_t input_ne[4];
size_t input_stride[4];
int64_t scale_ne[4];
size_t scale_stride[4];
int64_t indices_ne[4];
size_t indices_stride[4];
int64_t output_ne[4];
size_t output_stride[4];
int64_t ir;
int64_t dr;
int64_t group_size_in_row;
TPipe pipe;
GlobalTensor<int8_t> input_gm;
GlobalTensor<half> scale_gm;
GlobalTensor<int32_t> indices_gm;
GlobalTensor<float> output_gm;
TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
TQue<QuePosition::VECIN, BUFFER_NUM> cast_queue;
};
template <typename T>
__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
auto gm_ptr = (__gm__ uint8_t *)gm;
auto ub_ptr = (uint8_t *)(ub);
for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
*ub_ptr = *gm_ptr;
}
}
extern "C" __global__ __aicore__ void ascendc_get_row_q8_0(
GM_ADDR input_gm, GM_ADDR indices_gm, GM_ADDR output_gm,
GM_ADDR input_ne_gm, GM_ADDR indices_ne_gm, GM_ADDR indices_nb_gm,
GM_ADDR output_ne_gm, GM_ADDR output_nb_gm) {
int64_t input_ne_ub[4];
int64_t indices_ne_ub[4];
size_t indices_nb_ub[4];
int64_t output_ne_ub[4];
size_t output_nb_ub[4];
copy_to_ub(input_ne_gm, input_ne_ub, 32);
copy_to_ub(indices_ne_gm, indices_ne_ub, 32);
copy_to_ub(indices_nb_gm, indices_nb_ub, 32);
copy_to_ub(output_ne_gm, output_ne_ub, 32);
copy_to_ub(output_nb_gm, output_nb_ub, 32);
GET_ROW_Q8_0 op;
op.init(input_gm, indices_gm, output_gm, input_ne_ub, indices_ne_ub,
indices_nb_ub, output_ne_ub, output_nb_ub);
op.calculate();
}
@@ -1,218 +0,0 @@
#include "kernel_operator.h"
using namespace AscendC;
#ifdef ASCEND_310P
extern "C" __global__ __aicore__ void ascendc_quantize_f16_q8_0(
GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
// let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed.
printf("Ascend310P not support f16->8bit quantization.\n");
}
#else
#define BUFFER_NUM 2
#define QK8_0 32
class QUANTIZE_F16_Q8_0 {
public:
__aicore__ inline QUANTIZE_F16_Q8_0() {}
__aicore__ inline void init(GM_ADDR input, GM_ADDR output,
int64_t *input_ne_ub, size_t *input_nb_ub,
int64_t *output_ne_ub) {
int64_t op_block_num = GetBlockNum();
int64_t op_block_idx = GetBlockIdx();
for (int i = 0; i < 4; i++) {
input_ne[i] = input_ne_ub[i];
input_stride[i] = input_nb_ub[i] / input_nb_ub[0];
output_ne[i] = output_ne_ub[i];
}
output_stride[0] = 1;
for (int i = 1; i < 4; i++) {
output_stride[i] = output_stride[i - 1] * output_ne[i - 1];
}
scale_ne = input_ne;
scale_stride[0] = 1;
scale_stride[1] = input_ne[0] / QK8_0;
for (int i = 2; i < 4; i++) {
scale_stride[i] = scale_stride[i - 1] * scale_ne[i - 1];
}
// split input tensor by rows.
uint64_t nr = input_ne[1] * input_ne[2] * input_ne[3];
dr = nr / op_block_num;
uint64_t tails = nr % op_block_num;
if (op_block_idx < tails) {
dr += 1;
ir = dr * op_block_idx;
} else {
ir = dr * op_block_idx + tails;
}
group_size_in_row = scale_stride[1];
int64_t output_size = output_ne[0] * output_ne[1] * output_ne[2] *
output_ne[3] * sizeof(uint8_t);
input_gm.SetGlobalBuffer((__gm__ half *)input);
output_gm.SetGlobalBuffer((__gm__ int8_t *)output);
scale_gm.SetGlobalBuffer((__gm__ half *)(output + output_size + ir *
group_size_in_row *
sizeof(half)));
pipe.InitBuffer(input_queue, BUFFER_NUM, QK8_0 * sizeof(half));
pipe.InitBuffer(output_queue, BUFFER_NUM, QK8_0 * sizeof(int8_t));
pipe.InitBuffer(work_queue, 1, 32);
pipe.InitBuffer(max_queue, 1, 32);
pipe.InitBuffer(abs_queue, 1, QK8_0 * sizeof(float));
pipe.InitBuffer(scale_queue, 1, 32);
pipe.InitBuffer(cast_queue ,1 ,QK8_0 * sizeof(float));
}
__aicore__ inline void copy_in(uint32_t offset) {
LocalTensor<half> input_local = input_queue.AllocTensor<half>();
DataCopy(input_local, input_gm[offset], QK8_0);
input_queue.EnQue(input_local);
}
__aicore__ inline void copy_out(uint32_t offset) {
LocalTensor<int8_t> output_local = output_queue.DeQue<int8_t>();
DataCopy(output_gm[offset], output_local, QK8_0);
output_queue.FreeTensor(output_local);
}
__aicore__ inline half calculate_group(int64_t row, int64_t group) {
const int64_t i3 = row / (input_ne[1] * input_ne[2]);
const int64_t i2 = (row - i3 * input_ne[1] * input_ne[2]) / input_ne[1];
const int64_t i1 =
row - i3 * input_ne[1] * input_ne[2] - i2 * input_ne[1];
const int64_t input_offset = i1 * input_stride[1] +
i2 * input_stride[2] +
i3 * input_stride[3] + QK8_0 * group;
const int64_t output_offset = i1 * output_stride[1] +
i2 * output_stride[2] +
i3 * output_stride[3] + QK8_0 * group;
copy_in(input_offset);
LocalTensor<half> input_local = input_queue.DeQue<half>();
LocalTensor<int8_t> output_local = output_queue.AllocTensor<int8_t>();
LocalTensor<float> work_local = work_queue.AllocTensor<float>();
LocalTensor<float> abs_local = abs_queue.AllocTensor<float>();
LocalTensor<float> max_local = max_queue.AllocTensor<float>();
LocalTensor<float> cast_local = cast_queue.AllocTensor<float>();
Cast(cast_local, input_local, RoundMode::CAST_NONE, QK8_0);
Abs(abs_local, cast_local, QK8_0);
ReduceMax(max_local, abs_local, work_local, QK8_0);
pipe_barrier(PIPE_ALL);
float d = max_local.GetValue(0);
d = d / ((1 << 7) - 1);
if (d != 0) {
Muls(cast_local, cast_local, 1.0f / d, QK8_0);
}
Cast(cast_local, cast_local, RoundMode::CAST_ROUND, QK8_0);
Cast(input_local, cast_local, RoundMode::CAST_ROUND, QK8_0);
Cast(output_local, input_local, RoundMode::CAST_ROUND, QK8_0);
output_queue.EnQue(output_local);
copy_out(output_offset);
input_queue.FreeTensor(input_local);
work_queue.FreeTensor(work_local);
abs_queue.FreeTensor(abs_local);
max_queue.FreeTensor(max_local);
cast_queue.FreeTensor(cast_local);
return (half)d;
}
__aicore__ inline void calculate() {
LocalTensor<half> scale_local = scale_queue.AllocTensor<half>();
uint32_t scale_local_offset = 0;
uint32_t scale_global_offset = 0;
for (int64_t i = ir; i < ir + dr; i++) {
for (int64_t j = 0; j < group_size_in_row; j++) {
half scale = calculate_group(i, j);
scale_local.SetValue(scale_local_offset++, scale);
if (scale_local_offset == 16) {
scale_local_offset = 0;
// TODO: OPTIMIZE ME
pipe_barrier(PIPE_ALL);
DataCopy(scale_gm[scale_global_offset], scale_local, 16);
pipe_barrier(PIPE_ALL);
scale_global_offset += 16;
}
}
}
if (scale_local_offset != 0) {
pipe_barrier(PIPE_ALL);
DataCopyExtParams dataCopyParams;
dataCopyParams.blockCount = 1;
dataCopyParams.blockLen = scale_local_offset * sizeof(half);
DataCopyPad(scale_gm[scale_global_offset], scale_local,
dataCopyParams);
pipe_barrier(PIPE_ALL);
}
}
private:
int64_t input_ne[4];
size_t input_stride[4];
int64_t *scale_ne;
size_t scale_stride[4];
int64_t output_ne[4];
size_t output_stride[4];
int64_t group_size_in_row;
int64_t ir;
int64_t dr;
TPipe pipe;
GlobalTensor<half> input_gm;
GlobalTensor<half> scale_gm;
GlobalTensor<int8_t> output_gm;
TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
TQue<QuePosition::VECIN, 1> work_queue;
TQue<QuePosition::VECOUT, 1> max_queue;
TQue<QuePosition::VECIN, 1> abs_queue;
TQue<QuePosition::VECOUT, 1> scale_queue;
TQue<QuePosition::VECOUT, 1> cast_queue;
};
template <typename T>
__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
auto gm_ptr = (__gm__ uint8_t *)gm;
auto ub_ptr = (uint8_t *)(ub);
for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
*ub_ptr = *gm_ptr;
}
}
extern "C" __global__ __aicore__ void ascendc_quantize_f16_q8_0(
GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
int64_t input_ne_ub[4];
size_t input_nb_ub[4];
int64_t output_ne_ub[4];
copy_to_ub(input_ne_gm, input_ne_ub, 32);
copy_to_ub(input_nb_gm, input_nb_ub, 32);
copy_to_ub(output_ne_gm, output_ne_ub, 32);
QUANTIZE_F16_Q8_0 op;
op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
op.calculate();
}
#endif // #ifdef ASCEND_310P
@@ -1,216 +0,0 @@
#include "kernel_operator.h"
using namespace AscendC;
#ifdef ASCEND_310P // 310P not support f32->8bit quantization
extern "C" __global__ __aicore__ void ascendc_quantize_f32_q8_0(
GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
// let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed.
printf("Ascend310P not support f32->8bit quantization.\n");
}
#else
#define BUFFER_NUM 2
#define QK8_0 32
class QUANTIZE_F32_Q8_0 {
public:
__aicore__ inline QUANTIZE_F32_Q8_0() {}
__aicore__ inline void init(GM_ADDR input, GM_ADDR output,
int64_t *input_ne_ub, size_t *input_nb_ub,
int64_t *output_ne_ub) {
int64_t op_block_num = GetBlockNum();
int64_t op_block_idx = GetBlockIdx();
for (int i = 0; i < 4; i++) {
input_ne[i] = input_ne_ub[i];
input_stride[i] = input_nb_ub[i] / input_nb_ub[0];
output_ne[i] = output_ne_ub[i];
}
output_stride[0] = 1;
for (int i = 1; i < 4; i++) {
output_stride[i] = output_stride[i - 1] * output_ne[i - 1];
}
scale_ne = input_ne;
scale_stride[0] = 1;
scale_stride[1] = input_ne[0] / QK8_0;
for (int i = 2; i < 4; i++) {
scale_stride[i] = scale_stride[i - 1] * scale_ne[i - 1];
}
// split input tensor by rows.
uint64_t nr = input_ne[1] * input_ne[2] * input_ne[3];
dr = nr / op_block_num;
uint64_t tails = nr % op_block_num;
if (op_block_idx < tails) {
dr += 1;
ir = dr * op_block_idx;
} else {
ir = dr * op_block_idx + tails;
}
group_size_in_row = scale_stride[1];
int64_t output_size = output_ne[0] * output_ne[1] * output_ne[2] *
output_ne[3] * sizeof(uint8_t);
input_gm.SetGlobalBuffer((__gm__ float *)input);
output_gm.SetGlobalBuffer((__gm__ int8_t *)output);
scale_gm.SetGlobalBuffer((__gm__ half *)(output + output_size +
ir * group_size_in_row *
sizeof(half)));
pipe.InitBuffer(input_queue, BUFFER_NUM, QK8_0 * sizeof(float));
pipe.InitBuffer(output_queue, BUFFER_NUM, QK8_0 * sizeof(int8_t));
pipe.InitBuffer(work_queue, 1, 32);
pipe.InitBuffer(max_queue, 1, 32);
pipe.InitBuffer(abs_queue, 1, QK8_0 * sizeof(float));
pipe.InitBuffer(cast_queue, 1, QK8_0 * sizeof(half));
pipe.InitBuffer(scale_queue, 1, 32);
}
__aicore__ inline void copy_in(uint32_t offset) {
LocalTensor<float> input_local = input_queue.AllocTensor<float>();
DataCopy(input_local, input_gm[offset], QK8_0);
input_queue.EnQue(input_local);
}
__aicore__ inline void copy_out(uint32_t offset) {
LocalTensor<int8_t> output_local = output_queue.DeQue<int8_t>();
DataCopy(output_gm[offset], output_local, QK8_0);
output_queue.FreeTensor(output_local);
}
__aicore__ inline half calculate_group(int64_t row, int64_t group) {
const int64_t i3 = row / (input_ne[1] * input_ne[2]);
const int64_t i2 = (row - i3 * input_ne[1] * input_ne[2]) / input_ne[1];
const int64_t i1 =
row - i3 * input_ne[1] * input_ne[2] - i2 * input_ne[1];
const int64_t input_offset = i1 * input_stride[1] +
i2 * input_stride[2] +
i3 * input_stride[3] + QK8_0 * group;
const int64_t output_offset = i1 * output_stride[1] +
i2 * output_stride[2] +
i3 * output_stride[3] + QK8_0 * group;
copy_in(input_offset);
LocalTensor<float> input_local = input_queue.DeQue<float>();
LocalTensor<int8_t> output_local = output_queue.AllocTensor<int8_t>();
LocalTensor<float> work_local = work_queue.AllocTensor<float>();
LocalTensor<float> abs_local = abs_queue.AllocTensor<float>();
LocalTensor<float> max_local = max_queue.AllocTensor<float>();
LocalTensor<half> cast_local = cast_queue.AllocTensor<half>();
Abs(abs_local, input_local, QK8_0);
ReduceMax(max_local, abs_local, work_local, QK8_0);
pipe_barrier(PIPE_ALL);
float d = max_local.GetValue(0);
d = d / ((1 << 7) - 1);
if (d != 0) {
Muls(input_local, input_local, 1.0f / d, QK8_0);
}
Cast(input_local, input_local, RoundMode::CAST_ROUND, QK8_0);
Cast(cast_local, input_local, RoundMode::CAST_ROUND, QK8_0);
Cast(output_local, cast_local, RoundMode::CAST_ROUND, QK8_0);
output_queue.EnQue(output_local);
copy_out(output_offset);
input_queue.FreeTensor(input_local);
work_queue.FreeTensor(work_local);
abs_queue.FreeTensor(abs_local);
max_queue.FreeTensor(max_local);
cast_queue.FreeTensor(cast_local);
return (half)d;
}
__aicore__ inline void calculate() {
LocalTensor<half> scale_local = scale_queue.AllocTensor<half>();
uint32_t scale_local_offset = 0;
uint32_t scale_global_offset = 0;
for (int64_t i = ir; i < ir + dr; i++) {
for (int64_t j = 0; j < group_size_in_row; j++) {
half scale = calculate_group(i, j);
scale_local.SetValue(scale_local_offset++, scale);
if (scale_local_offset == 16) {
scale_local_offset = 0;
// TODO: OPTIMIZE ME
pipe_barrier(PIPE_ALL);
DataCopy(scale_gm[scale_global_offset], scale_local, 16);
pipe_barrier(PIPE_ALL);
scale_global_offset += 16;
}
}
}
if (scale_local_offset != 0) {
pipe_barrier(PIPE_ALL);
DataCopyExtParams dataCopyParams;
dataCopyParams.blockCount = 1;
dataCopyParams.blockLen = scale_local_offset * sizeof(half);
DataCopyPad(scale_gm[scale_global_offset], scale_local,
dataCopyParams);
pipe_barrier(PIPE_ALL);
}
}
private:
int64_t input_ne[4];
size_t input_stride[4];
int64_t *scale_ne;
size_t scale_stride[4];
int64_t output_ne[4];
size_t output_stride[4];
int64_t group_size_in_row;
int64_t ir;
int64_t dr;
TPipe pipe;
GlobalTensor<float> input_gm;
GlobalTensor<half> scale_gm;
GlobalTensor<int8_t> output_gm;
TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
TQue<QuePosition::VECIN, 1> work_queue;
TQue<QuePosition::VECOUT, 1> max_queue;
TQue<QuePosition::VECIN, 1> abs_queue;
TQue<QuePosition::VECIN, 1> cast_queue;
TQue<QuePosition::VECOUT, 1> scale_queue;
};
template <typename T>
__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
auto gm_ptr = (__gm__ uint8_t *)gm;
auto ub_ptr = (uint8_t *)(ub);
for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
*ub_ptr = *gm_ptr;
}
}
extern "C" __global__ __aicore__ void ascendc_quantize_f32_q8_0(
GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
int64_t input_ne_ub[4];
size_t input_nb_ub[4];
int64_t output_ne_ub[4];
copy_to_ub(input_ne_gm, input_ne_ub, 32);
copy_to_ub(input_nb_gm, input_nb_ub, 32);
copy_to_ub(output_ne_gm, output_ne_ub, 32);
QUANTIZE_F32_Q8_0 op;
op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
op.calculate();
}
#endif // #ifdef ASCEND_310P
@@ -1,295 +0,0 @@
#include "kernel_operator.h"
using namespace AscendC;
#ifdef ASCEND_310P // 310P not support float->4bit quantization
extern "C" __global__ __aicore__ void ascendc_quantize_f32_to_q4_0(
GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
// let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed.
printf("Ascend310P not support f32->4bit quantization.\n");
}
extern "C" __global__ __aicore__ void ascendc_quantize_f16_to_q4_0(
GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
// let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed.
printf("Ascend310P not support f16->4bit quantization.\n");
}
#else
#define BUFFER_NUM 2
#define Group_Size 32
template <typename SRC_T>
class QUANTIZE_FLOAT_TO_Q4_0 {
public:
__aicore__ inline QUANTIZE_FLOAT_TO_Q4_0() {}
__aicore__ inline void init(GM_ADDR input, GM_ADDR output,
int64_t *input_ne_ub, size_t *input_nb_ub,
int64_t *output_ne_ub) {
// TODO: fix test_case CPY(type_src=f16,type_dst=q4_0,ne=[256,4,4,4],
// permute=[0,0,0,0]):
// [CPY] NMSE = 0.000008343 > 0.000001000 FAIL
int64_t op_block_num = GetBlockNum();
int64_t op_block_idx = GetBlockIdx();
// input stride of data elements
for (int i = 0; i < 4; i++) {
input_ne[i] = input_ne_ub[i];
input_stride[i] = input_nb_ub[i] / input_nb_ub[0];
output_ne[i] = output_ne_ub[i];
}
// output stride of data elements
output_stride[0] = 1;
for (int i = 1; i < 4; i++) {
output_stride[i] = output_stride[i - 1] * output_ne[i - 1];
}
// scale saved one by one after data:. [group1_scale, group2_scale, ...]
scale_ne = input_ne;
scale_stride[0] = 1;
scale_stride[1] = input_ne[0] / Group_Size;
for (int i = 2; i < 4; i++) {
scale_stride[i] = scale_stride[i - 1] * scale_ne[i - 1];
}
// split input tensor by rows.
uint64_t nr = input_ne[1] * input_ne[2] * input_ne[3];
dr = nr / op_block_num;
uint64_t tails = nr % op_block_num;
if (op_block_idx < tails) {
dr += 1;
ir = dr * op_block_idx;
} else {
ir = dr * op_block_idx + tails;
}
group_size_in_row = scale_stride[1];
int64_t scale_offset = output_ne[0] * output_ne[1] * output_ne[2] *
output_ne[3] * sizeof(uint8_t) / 2;
input_gm.SetGlobalBuffer((__gm__ SRC_T *)input);
output_gm.SetGlobalBuffer((__gm__ int8_t *)output);
scale_gm.SetGlobalBuffer((__gm__ half *)(output + scale_offset + ir *
group_size_in_row *
sizeof(half)));
pipe.InitBuffer(input_queue, BUFFER_NUM, Group_Size * sizeof(SRC_T));
pipe.InitBuffer(output_queue, BUFFER_NUM,
Group_Size * sizeof(int8_t) / 2);
pipe.InitBuffer(cast_queue , 1, Group_Size * sizeof(float));
pipe.InitBuffer(work_queue, 1, Group_Size * sizeof(float));
pipe.InitBuffer(max_queue, 1, Group_Size * sizeof(float));
pipe.InitBuffer(min_queue, 1, Group_Size * sizeof(float));
pipe.InitBuffer(scale_queue, 1, Group_Size / 2 * sizeof(half));
pipe.InitBuffer(int8_queue, 1, Group_Size * sizeof(int8_t));
pipe.InitBuffer(half_queue, 1, Group_Size * sizeof(half));
}
__aicore__ inline void copy_in(uint32_t offset) {
LocalTensor<SRC_T> input_local = input_queue.AllocTensor<SRC_T>();
DataCopy(input_local, input_gm[offset], Group_Size);
input_queue.EnQue(input_local);
}
__aicore__ inline void copy_out(uint32_t offset) {
// reinterpretcast Group_Size(32) * int4b_t to Group_Size / 2 * int8_t,
// and using DataCopyPad to avoid 32 bits align.
LocalTensor<int4b_t> output_local = output_queue.DeQue<int4b_t>();
LocalTensor<int8_t> output_int8_local =
output_local.ReinterpretCast<int8_t>();
DataCopyExtParams dataCopyParams;
dataCopyParams.blockCount = 1;
dataCopyParams.blockLen = Group_Size / 2 * sizeof(int8_t);
DataCopyPad(output_gm[offset], output_int8_local, dataCopyParams);
output_queue.FreeTensor(output_local);
}
__aicore__ inline void input_to_cast(LocalTensor<float> cast_local,
LocalTensor<float> input_local) {
DataCopy(cast_local, input_local, Group_Size);
}
__aicore__ inline void input_to_cast(LocalTensor<float> cast_local,
LocalTensor<half> input_local) {
Cast(cast_local, input_local, RoundMode::CAST_NONE, Group_Size);
}
__aicore__ inline half calculate_group(int64_t row, int64_t group) {
const int64_t i3 = row / (input_ne[1] * input_ne[2]);
const int64_t i2 = (row - i3 * input_ne[1] * input_ne[2]) / input_ne[1];
const int64_t i1 =
row - i3 * input_ne[1] * input_ne[2] - i2 * input_ne[1];
const int64_t input_offset = i1 * input_stride[1] +
i2 * input_stride[2] +
i3 * input_stride[3] + Group_Size * group;
// output_offset is stride for output_gm which datatype is int8_t and
// divided by 2 is needed for int4b_t.
const int64_t output_offset = (i1 * output_stride[1] +
i2 * output_stride[2] +
i3 * output_stride[3] +
Group_Size * group) / 2;
copy_in(input_offset);
LocalTensor<SRC_T> input_local = input_queue.DeQue<SRC_T>();
LocalTensor<int4b_t> output_local = output_queue.AllocTensor<int4b_t>();
LocalTensor<float> cast_local = cast_queue.AllocTensor<float>();
LocalTensor<float> work_local = work_queue.AllocTensor<float>();
LocalTensor<float> max_local = max_queue.AllocTensor<float>();
LocalTensor<float> min_local = min_queue.AllocTensor<float>();
LocalTensor<int8_t> int8_local = int8_queue.AllocTensor<int8_t>();
LocalTensor<half> half_local = half_queue.AllocTensor<half>();
input_to_cast(cast_local, input_local);
ReduceMax(max_local, cast_local, work_local, Group_Size);
ReduceMin(min_local, cast_local, work_local, Group_Size);
const float max_value = max_local.GetValue(0);
const float min_value = min_local.GetValue(0);
float d = max_value;
if (min_value < 0 && (-1 * min_value) > max_value) {
d = min_value;
}
d = d / (-8);
if (d != 0) {
Muls(cast_local, cast_local, 1.0f / d, Group_Size);
}
// range: [-8,8] -> [0.5,16.5] -> [0,16] -> [0,15] -> [-8,7]
float scalar = 8.5f;
Adds(cast_local, cast_local, scalar, Group_Size);
Cast(cast_local, cast_local, RoundMode::CAST_FLOOR, Group_Size);
scalar = 15.0f;
Mins(cast_local, cast_local, scalar, Group_Size);
scalar = -8.0f;
Adds(cast_local, cast_local, scalar, Group_Size);
// float->half->int4b
Cast(half_local, cast_local, RoundMode::CAST_NONE, Group_Size);
Cast(output_local, half_local, RoundMode::CAST_NONE, Group_Size);
output_queue.EnQue(output_local);
copy_out(output_offset);
input_queue.FreeTensor(input_local);
work_queue.FreeTensor(work_local);
max_queue.FreeTensor(max_local);
min_queue.FreeTensor(min_local);
int8_queue.FreeTensor(int8_local);
half_queue.FreeTensor(half_local);
cast_queue.FreeTensor(cast_local);
return (half)d;
}
__aicore__ inline void calculate() {
LocalTensor<half> scale_local = scale_queue.AllocTensor<half>();
uint32_t scale_local_offset = 0;
uint32_t scale_global_offset = 0;
for (int64_t i = ir; i < ir + dr; i++) {
for (int64_t j = 0; j < group_size_in_row; j++) {
half scale = calculate_group(i, j);
scale_local.SetValue(scale_local_offset++, scale);
// Copy Group_Size/2 length data each time.
if (scale_local_offset == Group_Size / 2) {
scale_local_offset = 0;
// TODO: OPTIMIZE ME
pipe_barrier(PIPE_ALL);
DataCopy(scale_gm[scale_global_offset], scale_local,
Group_Size / 2);
pipe_barrier(PIPE_ALL);
scale_global_offset += Group_Size / 2;
}
}
}
if (scale_local_offset != 0) {
pipe_barrier(PIPE_ALL);
DataCopyExtParams dataCopyParams;
dataCopyParams.blockCount = 1;
dataCopyParams.blockLen = scale_local_offset * sizeof(half);
DataCopyPad(scale_gm[scale_global_offset], scale_local,
dataCopyParams);
pipe_barrier(PIPE_ALL);
}
scale_queue.FreeTensor(scale_local);
}
private:
int64_t input_ne[4];
size_t input_stride[4];
int64_t *scale_ne;
size_t scale_stride[4];
int64_t output_ne[4];
size_t output_stride[4];
int64_t group_size_in_row;
int64_t ir;
int64_t dr;
TPipe pipe;
GlobalTensor<SRC_T> input_gm;
GlobalTensor<half> scale_gm;
GlobalTensor<int8_t> output_gm;
TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
TQue<QuePosition::VECIN, BUFFER_NUM> work_queue;
TQue<QuePosition::VECOUT, BUFFER_NUM> max_queue;
TQue<QuePosition::VECOUT, BUFFER_NUM> min_queue;
TQue<QuePosition::VECOUT, BUFFER_NUM> scale_queue;
TQue<QuePosition::VECOUT, BUFFER_NUM> cast_queue;
TQue<QuePosition::VECOUT, BUFFER_NUM> int8_queue;
TQue<QuePosition::VECOUT, BUFFER_NUM> half_queue;
};
template <typename T>
__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
auto gm_ptr = (__gm__ uint8_t *)gm;
auto ub_ptr = (uint8_t *)(ub);
for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
*ub_ptr = *gm_ptr;
}
}
extern "C" __global__ __aicore__ void ascendc_quantize_f16_to_q4_0(
GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
int64_t input_ne_ub[4];
size_t input_nb_ub[4];
int64_t output_ne_ub[4];
copy_to_ub(input_ne_gm, input_ne_ub, 32);
copy_to_ub(input_nb_gm, input_nb_ub, 32);
copy_to_ub(output_ne_gm, output_ne_ub, 32);
QUANTIZE_FLOAT_TO_Q4_0<half> op;
op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
op.calculate();
}
extern "C" __global__ __aicore__ void ascendc_quantize_f32_to_q4_0(
GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
int64_t input_ne_ub[4];
size_t input_nb_ub[4];
int64_t output_ne_ub[4];
copy_to_ub(input_ne_gm, input_ne_ub, 32);
copy_to_ub(input_nb_gm, input_nb_ub, 32);
copy_to_ub(output_ne_gm, output_ne_ub, 32);
QUANTIZE_FLOAT_TO_Q4_0<float> op;
op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
op.calculate();
}
#endif // #ifdef ASCEND_310P
+10
View File
@@ -31,6 +31,8 @@
#include "ggml-cuda/rope.cuh"
#include "ggml-cuda/scale.cuh"
#include "ggml-cuda/softmax.cuh"
#include "ggml-cuda/ssm-conv.cuh"
#include "ggml-cuda/ssm-scan.cuh"
#include "ggml-cuda/sum.cuh"
#include "ggml-cuda/sumrows.cuh"
#include "ggml-cuda/tsembd.cuh"
@@ -2296,6 +2298,12 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
case GGML_OP_SUM_ROWS:
ggml_cuda_op_sum_rows(ctx, dst);
break;
case GGML_OP_SSM_CONV:
ggml_cuda_op_ssm_conv(ctx, dst);
break;
case GGML_OP_SSM_SCAN:
ggml_cuda_op_ssm_scan(ctx, dst);
break;
case GGML_OP_ARGSORT:
ggml_cuda_op_argsort(ctx, dst);
break;
@@ -3193,6 +3201,8 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
case GGML_OP_COS:
case GGML_OP_CLAMP:
case GGML_OP_LOG:
case GGML_OP_SSM_SCAN:
case GGML_OP_SSM_CONV:
return true;
case GGML_OP_CONT:
return op->src[0]->type != GGML_TYPE_BF16;
+151
View File
@@ -0,0 +1,151 @@
#include "ssm-conv.cuh"
template <size_t split_d_inner, size_t d_conv>
static __global__ void ssm_conv_f32(const float * __restrict__ src0, const float * __restrict__ src1,
const int src0_nb0, const int src0_nb1, const int src0_nb2, const int src1_nb1,
float * __restrict__ dst, const int dst_nb0, const int dst_nb1, const int dst_nb2,
const int nc, const int ncs, const int nr, const int n_t, const int n_s) {
const int tid = threadIdx.x;
const int bidx = blockIdx.x;
const int bidy = blockIdx.y;
const float * x_block = (const float *) ((char *) src0 + bidx * src0_nb2 + bidy * split_d_inner * src0_nb1);
const float * w_block = (const float *) ((char *) src1 + bidy * split_d_inner * src1_nb1);
float * y_block = (float *) ((char *) dst + bidx * dst_nb2 + bidy * split_d_inner * dst_nb0);
const int stride_x = src0_nb1 / sizeof(float);
const int stride_w = src1_nb1 / sizeof(float);
const int stride_y = dst_nb1 / sizeof(float);
float x[d_conv] = { 0.0f };
float w[d_conv] = { 0.0f };
#pragma unroll
for (int j = 0; j < d_conv; j++) {
w[j] = w_block[tid * stride_w + j];
}
for (int i = 0; i < n_t; i++) {
float sumf = 0.0f;
if (i == 0) {
for (int j = 0; j < d_conv; j++) {
x[j] = x_block[tid * stride_x + j];
}
} else {
x[(i - 1) % d_conv] = x_block[tid * stride_x + i + d_conv - 1];
}
#pragma unroll
for (int j = 0; j < d_conv; j++) {
sumf += x[(i + j) % d_conv] * w[j];
}
y_block[i * stride_y + tid] = sumf;
}
}
template <size_t split_d_inner, size_t d_conv, size_t split_n_t>
static __global__ void ssm_conv_long_token_f32(const float * __restrict__ src0, const float * __restrict__ src1,
const int src0_nb0, const int src0_nb1, const int src0_nb2,
const int src1_nb1, float * __restrict__ dst, const int dst_nb0,
const int dst_nb1, const int dst_nb2, const int nc, const int ncs,
const int nr, const int n_t, const int n_s) {
const int tid = threadIdx.x;
const int bidx = blockIdx.x;
const int bidy = blockIdx.y;
const int bidz = blockIdx.z;
const float * x_block = (const float *) ((char *) src0 + bidx * src0_nb2 + bidy * split_d_inner * src0_nb1 +
bidz * split_n_t * src0_nb0);
const float * w_block = (const float *) ((char *) src1 + bidy * split_d_inner * src1_nb1);
float * y_block =
(float *) ((char *) dst + bidx * dst_nb2 + bidz * split_n_t * dst_nb1 + bidy * split_d_inner * dst_nb0);
const int stride_x = src0_nb1 / sizeof(float);
const int stride_w = src1_nb1 / sizeof(float);
const int stride_y = dst_nb1 / sizeof(float);
float x[d_conv] = { 0.0f };
float w[d_conv] = { 0.0f };
#pragma unroll
for (int j = 0; j < d_conv; j++) {
w[j] = w_block[tid * stride_w + j];
}
#pragma unroll
for (int i = 0; i < split_n_t; i++) {
if (bidz * split_n_t + i < n_t) {
float sumf = 0.0f;
if (i == 0) {
for (int j = 0; j < d_conv; j++) {
x[j] = x_block[tid * stride_x + j];
}
} else {
x[(i - 1) % d_conv] = x_block[tid * stride_x + i + d_conv - 1];
}
#pragma unroll
for (int j = 0; j < d_conv; j++) {
sumf += x[(i + j) % d_conv] * w[j];
}
y_block[i * stride_y + tid] = sumf;
}
}
}
static void ssm_conv_f32_cuda(const float * src0, const float * src1, const int src0_nb0, const int src0_nb1,
const int src0_nb2, const int src1_nb1, float * dst, const int dst_nb0, const int dst_nb1,
const int dst_nb2, const int nc, const int ncs, const int nr, const int n_t,
const int n_s, cudaStream_t stream) {
const int threads = 128;
GGML_ASSERT(nr % threads == 0);
if (n_t <= 32) {
const dim3 blocks(n_s, (nr + threads - 1) / threads, 1);
if (nc == 4) {
ssm_conv_f32<threads, 4><<<blocks, threads, 0, stream>>>(src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1,
dst, dst_nb0, dst_nb1, dst_nb2, nc, ncs, nr, n_t,
n_s);
} else {
GGML_ABORT("Only support kernel size = 4 now.");
}
} else {
if (nc == 4) {
const int split_n_t = 32;
dim3 blocks(n_s, (nr + threads - 1) / threads, (n_t + split_n_t - 1) / split_n_t);
ssm_conv_long_token_f32<threads, 4, split_n_t>
<<<blocks, threads, 0, stream>>>(src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1, dst, dst_nb0,
dst_nb1, dst_nb2, nc, ncs, nr, n_t, n_s);
} else {
GGML_ABORT("Only support kernel size = 4 right now.");
}
}
}
void ggml_cuda_op_ssm_conv(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const struct ggml_tensor * src0 = dst->src[0]; // conv_x
const struct ggml_tensor * src1 = dst->src[1]; // conv1d.weight
const int nc = src1->ne[0]; // d_conv
const int ncs = src0->ne[0]; // d_conv - 1 + n_t
const int nr = src0->ne[1]; // d_inner
const int n_t = dst->ne[1]; // tokens per sequence
const int n_s = dst->ne[2]; // number of sequences in the batch
GGML_ASSERT(dst->ne[0] == nr);
GGML_ASSERT(src0->nb[0] == sizeof(float));
GGML_ASSERT(src1->nb[0] == sizeof(float));
GGML_ASSERT(src0->nb[1] == src0->ne[0] * sizeof(float));
const float * src0_d = (const float *) src0->data;
const float * src1_d = (const float *) src1->data;
float * dst_d = (float *) dst->data;
cudaStream_t stream = ctx.stream();
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
ssm_conv_f32_cuda(src0_d, src1_d, src0->nb[0], src0->nb[1], src0->nb[2], src1->nb[1], dst_d, dst->nb[0], dst->nb[1],
dst->nb[2], nc, ncs, nr, n_t, n_s, stream);
}
+3
View File
@@ -0,0 +1,3 @@
#include "common.cuh"
void ggml_cuda_op_ssm_conv(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+155
View File
@@ -0,0 +1,155 @@
#include "ssm-scan.cuh"
// #include <cuda_runtime.h>
// static __device__ void global_to_shared(const float *src, float *dst) {
// asm volatile("cp.async.");
// }
template <size_t splitD, size_t N>
__global__ void __launch_bounds__(splitD, 2)
ssm_scan_f32(const float * __restrict__ src0, const float * __restrict__ src1, const float * __restrict__ src2,
const float * __restrict__ src3, const float * __restrict__ src4, const float * __restrict__ src5,
const int src0_nb1, const int src0_nb2, const int src1_nb0, const int src1_nb1, const int src1_nb2,
const int src1_nb3, const int src2_nb0, const int src2_nb1, const int src2_nb2, const int src3_nb1,
const int src4_nb1, const int src4_nb2, const int src5_nb1, const int src5_nb2,
float * __restrict__ dst, const int D, const int L, const int B) {
const int bidx = blockIdx.x; // split along B
const int bidy = blockIdx.y; // split along D
const int tid = threadIdx.x;
const int wid = tid / 32;
const int wtid = tid % 32;
extern __shared__ float smem[];
const int stride_sA = N + 1;
const int stride_ss0 = N + 1;
float * smem_A = smem;
float * smem_s0 = smem_A + splitD * stride_sA;
const float * s0_block = (const float *) ((char *) src0 + bidx * src0_nb2 + bidy * splitD * src0_nb1);
const float * x_block = (const float *) ((char *) src1 + (bidx * src1_nb2) + bidy * splitD * sizeof(float));
const float * dt_block = (const float *) ((char *) src2 + (bidx * src2_nb2) + bidy * splitD * sizeof(float));
const float * A_block = (const float *) ((char *) src3 + bidy * splitD * src3_nb1);
const float * B_block = (const float *) ((char *) src4 + (bidx * src4_nb2));
const float * C_block = (const float *) ((char *) src5 + (bidx * src5_nb2));
float * y_block = (float *) ((char *) dst + (bidx * src1_nb2) + bidy * splitD * sizeof(float));
float * s_block = (float *) ((char *) dst + src1_nb3 + bidx * src0_nb2 + bidy * splitD * src0_nb1);
const int stride_s0 = src0_nb1 / sizeof(float);
const int stride_x = src1_nb1 / sizeof(float);
const int stride_dt = src2_nb1 / sizeof(float);
const int stride_A = src3_nb1 / sizeof(float);
const int stride_B = src4_nb1 / sizeof(float);
const int stride_C = src5_nb1 / sizeof(float);
const int stride_s = stride_s0;
const int stride_y = stride_x;
// can N not be 16? for example 32?
if (N == 16) {
#pragma unroll
for (int i = 0; i < splitD / 4; i += 2) {
float value = A_block[(wid * warpSize + i) * stride_A + wtid];
// todo: bank conflict
// I am always confused with how to use the swizzling method to solve
// bank conflit. Hoping somebody can tell me.
smem_A[(wid * warpSize + i) * stride_sA + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value;
}
#pragma unroll
for (int i = 0; i < splitD / 4; i += 2) {
float value = s0_block[(wid * warpSize + i) * stride_s0 + wtid];
smem_s0[(wid * warpSize + i) * stride_ss0 + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value;
}
}
__syncthreads();
for (int i = 0; i < L; i++) {
float dt_soft_plus = dt_block[i * stride_dt + tid];
if (dt_soft_plus <= 20.0f) {
dt_soft_plus = log1pf(exp(dt_soft_plus));
}
float x_dt = x_block[i * stride_x + tid] * dt_soft_plus;
float sumf = 0.0f;
#pragma unroll
for (int j = 0; j < N; j++) {
float state = (smem_s0[tid * stride_ss0 + j] * expf(dt_soft_plus * smem_A[tid * stride_sA + j])) +
(B_block[i * stride_B + j] * x_dt);
sumf += state * C_block[i * stride_C + j];
if (i == L - 1) {
s_block[tid * stride_s + j] = state;
} else {
smem_s0[tid * stride_ss0 + j] = state;
}
}
__syncthreads();
y_block[i * stride_y + tid] = sumf;
}
}
static void ssm_scan_f32_cuda(const float * src0, const float * src1, const float * src2, const float * src3,
const float * src4, const float * src5, const int src0_nb1, const int src0_nb2,
const int src1_nb0, const int src1_nb1, const int src1_nb2, const int src1_nb3,
const int src2_nb0, const int src2_nb1, const int src2_nb2, const int src3_nb1,
const int src4_nb1, const int src4_nb2, const int src5_nb1, const int src5_nb2,
float * dst, const int N, const int D, const int L, const int B, cudaStream_t stream) {
const int threads = 128;
// todo: consider D cannot be divided,does this situation exist?
GGML_ASSERT(D % threads == 0);
const dim3 blocks(B, (D + threads - 1) / threads, 1);
const int smem_size = (threads * (N + 1) * 2) * sizeof(float);
if (N == 16) {
ssm_scan_f32<128, 16><<<blocks, threads, smem_size, stream>>>(
src0, src1, src2, src3, src4, src5, src0_nb1, src0_nb2, src1_nb0, src1_nb1, src1_nb2, src1_nb3, src2_nb0,
src2_nb1, src2_nb2, src3_nb1, src4_nb1, src4_nb2, src5_nb1, src5_nb2, dst, D, L, B);
} else {
GGML_ABORT("doesn't support N!=16.");
}
}
void ggml_cuda_op_ssm_scan(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const struct ggml_tensor * src0 = dst->src[0]; // s
const struct ggml_tensor * src1 = dst->src[1]; // x
const struct ggml_tensor * src2 = dst->src[2]; // dt
const struct ggml_tensor * src3 = dst->src[3]; // A
const struct ggml_tensor * src4 = dst->src[4]; // B
const struct ggml_tensor * src5 = dst->src[5]; // C
// const int64_t d_state = src0->ne[0];
// const int64_t d_inner = src0->ne[1];
// const int64_t l = src1->ne[1];
// const int64_t b = src0->ne[2];
const int64_t nc = src0->ne[0]; // d_state
const int64_t nr = src0->ne[1]; // d_inner
const int64_t n_t = src1->ne[1]; // number of tokens per sequence
const int64_t n_s = src0->ne[2]; // number of sequences in the batch
GGML_ASSERT(ggml_nelements(src1) + ggml_nelements(src0) == ggml_nelements(dst));
GGML_ASSERT(src0->nb[0] == sizeof(float));
GGML_ASSERT(src1->nb[0] == sizeof(float));
GGML_ASSERT(src2->nb[0] == sizeof(float));
GGML_ASSERT(src3->nb[0] == sizeof(float));
GGML_ASSERT(src4->nb[0] == sizeof(float));
GGML_ASSERT(src5->nb[0] == sizeof(float));
// required for the dot product between s and C
GGML_ASSERT(src0->nb[1] == src0->ne[0] * sizeof(float));
// required for per-sequence offsets for states
GGML_ASSERT(src0->nb[2] == src0->ne[0] * src0->ne[1] * sizeof(float));
// required to get correct offset for state destination (i.e. src1->nb[3])
GGML_ASSERT(src1->nb[3] == src1->ne[0] * src1->ne[1] * src1->ne[2] * sizeof(float));
const float * src0_d = (const float *) src0->data;
const float * src1_d = (const float *) src1->data;
const float * src2_d = (const float *) src2->data;
const float * src3_d = (const float *) src3->data;
const float * src4_d = (const float *) src4->data;
const float * src5_d = (const float *) src5->data;
float * dst_d = (float *) dst->data;
cudaStream_t stream = ctx.stream();
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
ssm_scan_f32_cuda(src0_d, src1_d, src2_d, src3_d, src4_d, src5_d, src0->nb[1], src0->nb[2], src1->nb[0],
src1->nb[1], src1->nb[2], src1->nb[3], src2->nb[0], src2->nb[1], src2->nb[2], src3->nb[1],
src4->nb[1], src4->nb[2], src5->nb[1], src5->nb[2], dst_d, nc, nr, n_t, n_s, stream);
}
+3
View File
@@ -0,0 +1,3 @@
#include "common.cuh"
void ggml_cuda_op_ssm_scan(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+1 -1
View File
@@ -4179,7 +4179,7 @@ static void ggml_metal_encode_node(
// ne00*(nsg)
// each simdgroup has a full f16 head vector in shared mem to accumulate results
//
#define FATTN_SMEM(nsg) (GGML_PAD((nqptg*(GGML_PAD(ne00, 128) + 2*ncpsg*(nsg)) + ne20*(nsg))*(sizeof(float)/2), 16))
#define FATTN_SMEM(nsg) (GGML_PAD((nqptg*(GGML_PAD(ne00, 128) + 4*ncpsg*(nsg)) + ne20*(nsg))*(sizeof(float)/2), 16))
int64_t nsgmax = 2;
while (true) {
+47 -47
View File
@@ -3184,8 +3184,8 @@ kernel void kernel_flash_attn_ext(
threadgroup_barrier(mem_flags::mem_threadgroup);
{
half S[Q] = { [0 ... Q-1] = 0.0f };
half M[Q] = { [0 ... Q-1] = -__FLT16_MAX__/2 };
float S[Q] = { [0 ... Q-1] = 0.0f };
float M[Q] = { [0 ... Q-1] = -__FLT16_MAX__/2 };
// thread indices inside the simdgroup
// TODO: see if we can utilize quad-group functions for better performance
@@ -3202,13 +3202,13 @@ kernel void kernel_flash_attn_ext(
const bool has_mask = mask != q;
half slope = 1.0f;
float slope = 1.0f;
// ALiBi
if (args.max_bias > 0.0f) {
const short h = iq2;
const half base = h < args.n_head_log2 ? args.m0 : args.m1;
const float base = h < args.n_head_log2 ? args.m0 : args.m1;
const short exph = h < args.n_head_log2 ? h + 1 : 2*(h - args.n_head_log2) + 1;
slope = pow(base, exph);
@@ -3224,14 +3224,14 @@ kernel void kernel_flash_attn_ext(
if (has_mask) {
// used to detect blocks full of -INF
half smax = -INFINITY;
float smax = -INFINITY;
// load the mask in shared memory
#pragma unroll(Q)
for (short j = 0; j < Q; ++j) {
device const half * pm = (device const half *) ((device const char *) mask + (iq1 + j)*args.nb31);
const half m = pm[ic + tiisg];
const float m = pm[ic + tiisg];
ss[j*TS + C + tiisg] = m;
smax = max(smax, m);
@@ -3327,10 +3327,10 @@ kernel void kernel_flash_attn_ext(
// online softmax
{
for (ushort j = 0; j < Q; ++j) {
const half m = M[j];
const float m = M[j];
// scale and apply the logitcap / mask
half s = ss[j*TS + tiisg]*args.scale;
float s = ss[j*TS + tiisg]*args.scale;
if (args.logit_softcap != 0.0f) {
s = args.logit_softcap*precise::tanh(s);
@@ -3341,8 +3341,8 @@ kernel void kernel_flash_attn_ext(
M[j] = simd_max(max(M[j], s));
const half ms = exp(m - M[j]);
const half vs = exp(s - M[j]);
const float ms = exp(m - M[j]);
const float vs = exp(s - M[j]);
S[j] = S[j]*ms + simd_sum(vs);
@@ -3444,8 +3444,8 @@ kernel void kernel_flash_attn_ext(
// reduce the warps sequentially
for (ushort sg = 1; sg < nsg; ++sg) {
half S = { 0.0f };
half M = { -__FLT16_MAX__/2 };
float S = { 0.0f };
float M = { -__FLT16_MAX__/2 };
threadgroup_barrier(mem_flags::mem_threadgroup);
@@ -3461,16 +3461,16 @@ kernel void kernel_flash_attn_ext(
// the first simdgroup accumulates the results from the other simdgroups
if (sgitg == 0) {
for (short j = 0; j < Q; ++j) {
const half S0 = ss[j*TS + 0];
const half S1 = ss[j*TS + sg*SH + 0];
const float S0 = ss[j*TS + 0];
const float S1 = ss[j*TS + sg*SH + 0];
const half M0 = ss[j*TS + 1];
const half M1 = ss[j*TS + sg*SH + 1];
const float M0 = ss[j*TS + 1];
const float M1 = ss[j*TS + sg*SH + 1];
M = max(M0, M1);
const half ms0 = exp(M0 - M);
const half ms1 = exp(M1 - M);
const float ms0 = exp(M0 - M);
const float ms1 = exp(M1 - M);
S = S0*ms0 + S1*ms1;
@@ -3646,16 +3646,16 @@ kernel void kernel_flash_attn_ext_vec(
constexpr short DV4 = DV/4;
constexpr short NW = N_SIMDWIDTH;
constexpr short NL = NW/NE; // note: this can be adjusted to support different head sizes and simdgroup work loads
constexpr short SH = 2*C; // shared memory per simdgroup
constexpr short SH = 4*C; // shared memory per simdgroup
const short T = DK + nsg*SH; // shared memory size per query in (half)
//threadgroup q_t * sq = (threadgroup q_t *) (shmem_f16 + 0*DK); // holds the query data
threadgroup q4_t * sq4 = (threadgroup q4_t *) (shmem_f16 + 0*DK); // same as above but in q4_t
threadgroup s_t * ss = (threadgroup s_t *) (shmem_f16 + sgitg*SH + Q*DK); // scratch buffer for attention
threadgroup s4_t * ss4 = (threadgroup s4_t *) (shmem_f16 + sgitg*SH + Q*DK); // same as above but in s4_t
threadgroup half * sm = (threadgroup half *) (shmem_f16 + sgitg*SH + C + Q*DK); // scratch buffer for mask
threadgroup o4_t * sr4 = (threadgroup o4_t *) (shmem_f16 + sgitg*DV + Q*T); // scratch buffer for the results
//threadgroup q_t * sq = (threadgroup q_t *) (shmem_f16 + 0*DK); // holds the query data
threadgroup q4_t * sq4 = (threadgroup q4_t *) (shmem_f16 + 0*DK); // same as above but in q4_t
threadgroup s_t * ss = (threadgroup s_t *) (shmem_f16 + sgitg*SH + Q*DK); // scratch buffer for attention
threadgroup s4_t * ss4 = (threadgroup s4_t *) (shmem_f16 + sgitg*SH + Q*DK); // same as above but in s4_t
threadgroup float * sm = (threadgroup float *) (shmem_f16 + sgitg*SH + 2*C + Q*DK); // scratch buffer for mask
threadgroup o4_t * sr4 = (threadgroup o4_t *) (shmem_f16 + sgitg*DV + Q*T); // scratch buffer for the results
// store the result for all queries in local memory (the O matrix from the paper)
o4_t lo[DV4/NL];
@@ -3684,8 +3684,8 @@ kernel void kernel_flash_attn_ext_vec(
threadgroup_barrier(mem_flags::mem_threadgroup);
{
half S = 0.0f;
half M = -__FLT16_MAX__/2;
float S = 0.0f;
float M = -__FLT16_MAX__/2;
// thread indices inside the simdgroup
const short tx = tiisg%NL;
@@ -3703,13 +3703,13 @@ kernel void kernel_flash_attn_ext_vec(
// pointer to the mask
device const half * pm = (device const half *) (mask + iq1*args.nb31);
half slope = 1.0f;
float slope = 1.0f;
// ALiBi
if (args.max_bias > 0.0f) {
const short h = iq2;
const half base = h < args.n_head_log2 ? args.m0 : args.m1;
const float base = h < args.n_head_log2 ? args.m0 : args.m1;
const short exph = h < args.n_head_log2 ? h + 1 : 2*(h - args.n_head_log2) + 1;
slope = pow(base, exph);
@@ -3799,13 +3799,13 @@ kernel void kernel_flash_attn_ext_vec(
// online softmax
{
const half m = M;
const half s = ss[tiisg];
const float m = M;
const float s = ss[tiisg];
M = simd_max(max(M, s));
const half ms = exp(m - M);
const half vs = exp(s - M);
const float ms = exp(m - M);
const float vs = exp(s - M);
S = S*ms + simd_sum(vs);
@@ -3836,7 +3836,7 @@ kernel void kernel_flash_attn_ext_vec(
v4_t mv;
deq_v_t4(pv4 + i/nl_v, i%nl_v, mv);
lo[ii/NL] += mv*ms;
lo[ii/NL] += o4_t(float4(mv)*float4(ms));
}
}
}
@@ -3907,18 +3907,18 @@ kernel void kernel_flash_attn_ext_vec(
// parallel reduce
for (short r = nsg/2; r > 0; r >>= 1) {
if (sgitg < r) {
const half S0 = ss[ 0];
const half S1 = ss[r*SH + 0];
const float S0 = ss[ 0];
const float S1 = ss[r*(SH/2) + 0];
const half M0 = ss[ 1];
const half M1 = ss[r*SH + 1];
const float M0 = ss[ 1];
const float M1 = ss[r*(SH/2) + 1];
const half M = max(M0, M1);
const float M = max(M0, M1);
const half ms0 = exp(M0 - M);
const half ms1 = exp(M1 - M);
const float ms0 = exp(M0 - M);
const float ms1 = exp(M1 - M);
const half S = S0*ms0 + S1*ms1;
const float S = S0*ms0 + S1*ms1;
if (tiisg == 0) {
ss[0] = S;
@@ -3950,11 +3950,11 @@ kernel void kernel_flash_attn_ext_vec(
// in the other (non-vec) kernel, we need s_t to also be float because we scale during the soft_max
//
#define FA_TYPES \
half4, \
half4, \
half4, \
float, \
half, half4, \
half4, \
half4, \
half4, \
float, \
float, float4, \
half4
typedef decltype(kernel_flash_attn_ext_vec<FA_TYPES, half4, 1, dequantize_f16_t4, half4, 1, dequantize_f16_t4, 128, 128, 4>) flash_attn_ext_vec_t;
+23 -3
View File
@@ -921,10 +921,30 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
backend_ctx->program_CL_gemm = build_program_from_source(context, device, kernel_src_CL_gemm.c_str(), compile_opts);
CL_CHECK((backend_ctx->CL_mul_mat_Ab_Bi_8x4 = clCreateKernel(backend_ctx->program_CL_gemm, "kernel_mul_mat_Ab_Bi_8x4", &err), err));
// TODO: fixme: these sizes are hardcoded for now.
// they should be allocated based on the model's size
// and the device's max alloc size
// Allocate intermediate buffers and images
size_t max_A_q_d_bytes = 311164928;
size_t max_A_s_d_bytes = 38895616;
size_t max_B_d_bytes = 45088768;
size_t required_A_q_d_bytes = 311164928;
size_t required_A_s_d_bytes = 38895616;
size_t required_B_d_bytes = 45088768;
// Ensure buffer sizes do not exceed the maximum allocation size
size_t max_A_q_d_bytes = MIN(required_A_q_d_bytes, backend_ctx->max_alloc_size);
size_t max_A_s_d_bytes = MIN(required_A_s_d_bytes, backend_ctx->max_alloc_size);
size_t max_B_d_bytes = MIN(required_B_d_bytes, backend_ctx->max_alloc_size);
if (required_A_q_d_bytes > backend_ctx->max_alloc_size) {
GGML_LOG_WARN("ggml_opencl: A_q_d buffer size reduced from %zu to %zu due to device limitations.\n",
required_A_q_d_bytes, max_A_q_d_bytes);
}
if (required_A_s_d_bytes > backend_ctx->max_alloc_size) {
GGML_LOG_WARN("ggml_opencl: A_s_d buffer size reduced from %zu to %zu due to device limitations.\n",
required_A_s_d_bytes, max_A_s_d_bytes);
}
if (required_B_d_bytes > backend_ctx->max_alloc_size) {
GGML_LOG_WARN("ggml_opencl: B_d buffer size reduced from %zu to %zu due to device limitations.\n",
required_B_d_bytes, max_B_d_bytes);
}
CL_CHECK((backend_ctx->A_q_d_max = clCreateBuffer(context, 0, max_A_q_d_bytes, NULL, &err), err));
CL_CHECK((backend_ctx->A_s_d_max = clCreateBuffer(context, 0, max_A_s_d_bytes, NULL, &err), err));
+87 -22
View File
@@ -23,6 +23,23 @@ ggml_add_backend_library(ggml-sycl
../../include/ggml-sycl.h
)
file(GLOB GGML_HEADERS_SYCL "*.hpp")
file(GLOB GGML_SOURCES_SYCL "*.cpp")
target_sources(ggml-sycl PRIVATE ${GGML_HEADERS_SYCL} ${GGML_SOURCES_SYCL})
find_package(IntelSYCL)
if (IntelSYCL_FOUND)
# Use oneAPI CMake when possible
target_link_libraries(ggml-sycl PRIVATE IntelSYCL::SYCL_CXX)
else()
# Fallback to the simplest way of enabling SYCL when using intel/llvm nightly for instance
target_compile_options(ggml-sycl PRIVATE "-fsycl")
target_link_options(ggml-sycl PRIVATE "-fsycl")
endif()
target_compile_options(ggml-sycl PRIVATE "-Wno-narrowing")
# Link against oneDNN
find_package(DNNL)
set(GGML_SYCL_DNNL 0)
if(DNNL_FOUND)
@@ -62,8 +79,6 @@ if (GGML_SYCL_F16)
add_compile_definitions(GGML_SYCL_F16)
endif()
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing -fsycl")
if (GGML_SYCL_TARGET STREQUAL "NVIDIA")
add_compile_definitions(GGML_SYCL_WARP_SIZE=32)
elseif (GGML_SYCL_TARGET STREQUAL "AMD")
@@ -76,34 +91,84 @@ else()
add_compile_definitions(GGML_SYCL_WARP_SIZE=16)
endif()
file(GLOB GGML_HEADERS_SYCL "*.hpp")
file(GLOB GGML_SOURCES_SYCL "*.cpp")
target_sources(ggml-sycl PRIVATE ${GGML_HEADERS_SYCL} ${GGML_SOURCES_SYCL})
if (GGML_SYCL_GRAPH)
target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_GRAPH)
endif()
if (WIN32)
find_package(IntelSYCL REQUIRED)
# Link against Intel oneMKL or oneMath
if (GGML_SYCL_TARGET STREQUAL "INTEL")
# Intel devices use Intel oneMKL directly instead of oneMath to avoid the limitation of linking Intel oneMKL statically
# See https://github.com/uxlfoundation/oneMath/issues/654
find_package(MKL REQUIRED)
target_link_libraries(ggml-sycl PRIVATE IntelSYCL::SYCL_CXX MKL::MKL MKL::MKL_SYCL)
target_link_libraries(ggml-sycl PRIVATE MKL::MKL_SYCL::BLAS)
target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_USE_INTEL_ONEMKL)
else()
if (GGML_SYCL_GRAPH)
add_compile_definitions(GGML_SYCL_GRAPH)
find_package(oneMath QUIET)
if (NOT oneMath_FOUND)
message(STATUS "oneMath not found: oneMath will be automatically downloaded")
# Use FetchContent to automatically pull and build oneMath
include(FetchContent)
set(BUILD_FUNCTIONAL_TESTS False)
set(BUILD_EXAMPLES False)
set(TARGET_DOMAINS blas)
if (GGML_SYCL_TARGET STREQUAL "NVIDIA")
set(ENABLE_MKLCPU_BACKEND False)
set(ENABLE_MKLGPU_BACKEND False)
set(ENABLE_CUBLAS_BACKEND True)
elseif (GGML_SYCL_TARGET STREQUAL "AMD")
set(ENABLE_MKLCPU_BACKEND False)
set(ENABLE_MKLGPU_BACKEND False)
set(ENABLE_ROCBLAS_BACKEND True)
# Ensure setting a string variable here is not overriden by oneMath CACHE variables
cmake_policy(SET CMP0126 NEW)
# Setting the device architecture is only needed and useful for AMD devices in oneMath
set(HIP_TARGETS ${GGML_SYCL_DEVICE_ARCH} CACHE STRING "oneMath HIP target" FORCE)
endif()
FetchContent_Declare(
ONEMATH
GIT_REPOSITORY https://github.com/uxlfoundation/oneMath.git
GIT_TAG c255b1b4c41e2ee3059455c1f96a965d6a62568a
)
FetchContent_MakeAvailable(ONEMATH)
# Create alias to match with find_package targets name
function(onemath_alias target)
if (TARGET ${target}_obj)
# Silence verbose warnings from external libraries
target_compile_options(${target}_obj PRIVATE -w)
endif()
if (TARGET ${target})
add_library(ONEMATH::${target} ALIAS ${target})
endif()
endfunction()
onemath_alias(onemath)
onemath_alias(onemath_blas_mklcpu)
onemath_alias(onemath_blas_mklgpu)
onemath_alias(onemath_blas_cublas)
onemath_alias(onemath_blas_rocblas)
endif()
if (GGML_SYCL_TARGET STREQUAL "INTEL")
target_link_libraries(ggml-sycl PRIVATE sycl OpenCL mkl_core pthread m dl mkl_sycl_blas mkl_intel_ilp64 mkl_tbb_thread)
elseif (GGML_SYCL_TARGET STREQUAL "NVIDIA")
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=nvptx64-nvidia-cuda")
add_compile_definitions(GGML_SYCL_NVIDIA)
target_link_libraries(ggml-sycl PRIVATE sycl pthread m dl onemkl_blas_cublas)
# Below oneMath compile-time dispatching is used for better performance
if (GGML_SYCL_TARGET STREQUAL "NVIDIA")
target_link_libraries(ggml-sycl PRIVATE ONEMATH::onemath_blas_cublas)
target_compile_options(ggml-sycl PRIVATE "-fsycl-targets=nvptx64-nvidia-cuda")
target_link_options(ggml-sycl PRIVATE "-fsycl-targets=nvptx64-nvidia-cuda")
target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_NVIDIA)
elseif (GGML_SYCL_TARGET STREQUAL "AMD")
if (NOT GGML_SYCL_DEVICE_ARCH)
message(ERROR "Can't enable SYCL hip backend, GGML_SYCL_DEVICE_ARCH has not been set.")
endif()
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=amdgcn-amd-amdhsa")
target_link_libraries(ggml-sycl PRIVATE sycl pthread m dl onemkl)
target_link_libraries(ggml-sycl PRIVATE ONEMATH::onemath_blas_rocblas)
target_compile_options(ggml-sycl PRIVATE "-fsycl-targets=amdgcn-amd-amdhsa")
target_link_options(ggml-sycl PRIVATE "-fsycl-targets=amdgcn-amd-amdhsa")
target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_AMD)
else()
# Fallback to oneMath runtime dispatcher
target_link_libraries(ggml-sycl PRIVATE ONEMATH::onemath)
target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_GENERIC)
endif()
endif()
if (GGML_SYCL_DEVICE_ARCH)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Xsycl-target-backend --offload-arch=${GGML_SYCL_DEVICE_ARCH}")
endif()
if (GGML_SYCL_DEVICE_ARCH)
target_compile_options(ggml-sycl PRIVATE -Xsycl-target-backend --offload-arch=${GGML_SYCL_DEVICE_ARCH})
target_link_options(ggml-sycl PRIVATE -Xsycl-target-backend --offload-arch=${GGML_SYCL_DEVICE_ARCH})
endif()
-35
View File
@@ -66,41 +66,6 @@ int64_t downsample_sycl_global_range(int64_t accumulate_block_num, int64_t block
return sycl_down_blk_size;
}
void ggml_sycl_op_flatten(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1, ggml_tensor *dst,
const ggml_sycl_op_flatten_t op) try {
const bool use_src1 = src1 != nullptr;
if(use_src1)
GGML_ASSERT(strcmp(src1->buffer->buft->iface.get_name(src1->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
GGML_ASSERT(strcmp(dst->buffer->buft->iface.get_name(dst->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
// dd = data device
float * src0_ddf = (float *) src0->data;
float * src1_ddf = use_src1 ? (float *) src1->data : nullptr;
float * dst_ddf = (float *) dst->data;
ggml_sycl_pool_alloc<float> src0_f(ctx.pool());
ggml_sycl_pool_alloc<float> src1_f(ctx.pool());
ggml_sycl_pool_alloc<float> dst_f(ctx.pool());
ggml_sycl_set_device(ctx.device);
queue_ptr main_stream = ctx.stream();
// GGML_SYCL_DEBUG("ctx.device=%d, main_stream=%p src0_on_device=%d, src1_on_device=%d, dst_on_device=%d\n",
// ctx.device, main_stream, src0_on_device, src1_on_device, dst_on_device);
// do the computation
op(ctx, src0, src1, dst, src0_ddf, src1_ddf, dst_ddf, main_stream);
// print_ggml_tensor("tensor", dst);
}
catch (sycl::exception const &exc) {
std::cerr << exc.what() << "Exception caught at file:" << __FILE__
<< ", line:" << __LINE__ << std::endl;
std::exit(1);
}
void release_extra_gpu(ggml_tensor_extra_gpu * extra, std::vector<queue_ptr> streams) {
for (int i = 0; i < ggml_sycl_info().device_count; ++i) {
for (int64_t is = 0; is < GGML_SYCL_MAX_STREAMS; ++is) {
+8 -20
View File
@@ -494,12 +494,6 @@ static __dpct_inline__ Tp* get_pointer(sycl::local_accessor<Tp, dim> acc) {
int64_t downsample_sycl_global_range(int64_t accumulate_block_num, int64_t block_size);
typedef void (*ggml_sycl_op_flatten_t)(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1,
ggml_tensor *dst, const float *src0_dd,
const float *src1_dd, float *dst_dd,
const queue_ptr &main_stream);
template<float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
static void k_bin_bcast(const src0_t * src0, const src1_t * src1, dst_t * dst,
int ne0, int ne1, int ne2, int ne3,
@@ -757,24 +751,22 @@ struct bin_bcast_sycl {
template <class op>
inline void ggml_sycl_op_bin_bcast(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1, ggml_tensor *dst,
const float *src0_dd, const float *src1_dd,
float *dst_dd,
const queue_ptr &main_stream) {
const ggml_tensor *src1, ggml_tensor *dst) {
dpct::queue_ptr main_stream = ctx.stream();
if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
op()(ctx, src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream);
op()(ctx, src0, src1, dst, (const float *)src0->data, (const float *)src1->data, (float *)dst->data, main_stream);
} else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
op()(ctx, src0, src1, dst, (const sycl::half *)src0_dd, src1_dd,
(sycl::half *)dst_dd, main_stream);
op()(ctx, src0, src1, dst, (const sycl::half *)src0->data, (const float *)src1->data,
(sycl::half *)dst->data, main_stream);
} else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) {
op()(ctx, src0, src1, dst, (const sycl::half *)src0_dd, src1_dd, dst_dd,
op()(ctx, src0, src1, dst, (const sycl::half *)src0->data, (const float *)src1->data, (float *)dst->data,
main_stream);
} else if (src0->type == GGML_TYPE_I32 && dst->type == GGML_TYPE_I32) {
op()(ctx, src0, src1, dst, (const int32_t *)src0_dd, (const int32_t *)src1_dd, (int32_t *)dst_dd,
op()(ctx, src0, src1, dst, (const int32_t *)src0->data, (const int32_t *)src1->data, (int32_t *)dst->data,
main_stream);
} else if (src0->type == GGML_TYPE_I16 && dst->type == GGML_TYPE_I16) {
op()(ctx, src0, src1, dst, (const int16_t *)src0_dd, (const int16_t *)src1_dd, (int16_t *)dst_dd,
op()(ctx, src0, src1, dst, (const int16_t *)src0->data, (const int16_t *)src1->data, (int16_t *)dst->data,
main_stream);
} else {
fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s, src1: %s\n", __func__,
@@ -784,8 +776,4 @@ inline void ggml_sycl_op_bin_bcast(ggml_backend_sycl_context & ctx, const ggml_t
}
bool gpu_has_xmx(sycl::device &dev);
void ggml_sycl_op_flatten(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1, ggml_tensor *dst,
const ggml_sycl_op_flatten_t op);
#endif // GGML_SYCL_COMMON_HPP
+79 -90
View File
@@ -16,9 +16,18 @@
#include <sycl/sycl.hpp>
#include <sycl/half_type.hpp>
#include <syclcompat/math.hpp>
#include <oneapi/mkl.hpp>
#include <map>
#ifdef GGML_SYCL_USE_INTEL_ONEMKL
#include <oneapi/mkl.hpp>
// Allow to use the same namespace for Intel oneMKL and oneMath
namespace oneapi {
namespace math = mkl;
}
#else
#include <oneapi/math.hpp>
#endif
#include "ggml.h"
#if defined(__linux__)
@@ -83,13 +92,32 @@ inline std::string get_device_backend_and_type(const sycl::device &device) {
}
template <typename Ts> struct matrix_info_t {
oneapi::mkl::transpose transpose_info[2];
oneapi::math::transpose transpose_info[2];
Ts value_info[2];
std::int64_t size_info[3];
std::int64_t ld_info[3];
std::int64_t groupsize_info;
};
inline auto get_onemath_backend(sycl::queue& queue)
#if defined(GGML_SYCL_GENERIC) || defined(GGML_SYCL_USE_INTEL_ONEMKL)
-> sycl::queue&
#endif
{
// If the backend is known at compile-time, use oneMath backend_selector to use
// compile-time dispatching and avoid the need to dlopen libraries. Otherwise
// fallback to runtime dispatching.
#if defined(GGML_SYCL_NVIDIA)
return oneapi::math::backend_selector<oneapi::math::backend::cublas>{ queue };
#elif defined(GGML_SYCL_AMD)
return oneapi::math::backend_selector<oneapi::math::backend::rocblas>{ queue };
#elif defined(GGML_SYCL_GENERIC) || defined(GGML_SYCL_USE_INTEL_ONEMKL)
return queue;
#else
static_assert(false, "Unsupported backend");
#endif
}
namespace dpct
{
typedef sycl::queue *queue_ptr;
@@ -1686,26 +1714,18 @@ namespace dpct
namespace detail
{
template <class Ta, class Tb, class Tc, class Ts>
inline void gemm_impl(sycl::queue &q, oneapi::mkl::transpose a_trans,
oneapi::mkl::transpose b_trans, int m, int n, int k,
const void *alpha, const void *a, int lda, const void *b,
int ldb, const void *beta, void *c, int ldc)
{
Ts alpha_value = dpct::get_value(reinterpret_cast<const Ts *>(alpha), q);
Ts beta_value = dpct::get_value(reinterpret_cast<const Ts *>(beta), q);
auto data_a = get_memory<const Ta>(a);
auto data_b = get_memory<const Tb>(b);
auto data_c = get_memory<Tc>(c);
#ifdef GGML_SYCL_NVIDIA
oneapi::mkl::blas::column_major::gemm(oneapi::mkl::backend_selector<oneapi::mkl::backend::cublas>{ q },
a_trans, b_trans, m, n, k, alpha_value, data_a, lda, data_b, ldb,
beta_value, data_c, ldc);
#else
oneapi::mkl::blas::column_major::gemm(q, a_trans, b_trans, m, n, k, alpha_value, data_a, lda, data_b, ldb,
beta_value, data_c, ldc);
#endif
}
template <class Ta, class Tb, class Tc, class Ts>
inline void gemm_impl(sycl::queue & q, oneapi::math::transpose a_trans, oneapi::math::transpose b_trans, int m,
int n, int k, const void * alpha, const void * a, int lda, const void * b, int ldb,
const void * beta, void * c, int ldc) {
Ts alpha_value = dpct::get_value(reinterpret_cast<const Ts *>(alpha), q);
Ts beta_value = dpct::get_value(reinterpret_cast<const Ts *>(beta), q);
auto data_a = get_memory<const Ta>(a);
auto data_b = get_memory<const Tb>(b);
auto data_c = get_memory<Tc>(c);
oneapi::math::blas::column_major::gemm(get_onemath_backend(q), a_trans, b_trans, m, n, k, alpha_value, data_a,
lda, data_b, ldb, beta_value, data_c, ldc);
}
template <typename VecT, class BinaryOperation, class = void>
class vectorized_binary
@@ -1735,7 +1755,7 @@ namespace dpct
};
template <class Ta, class Tb, class Tc, class Ts>
inline void gemm_batch_impl(sycl::queue & q, oneapi::mkl::transpose a_trans, oneapi::mkl::transpose b_trans,
inline void gemm_batch_impl(sycl::queue & q, oneapi::math::transpose a_trans, oneapi::math::transpose b_trans,
int m, int n, int k, const void * alpha, const void ** a, int lda, const void ** b,
int ldb, const void * beta, void ** c, int ldc, int batch_size,
matrix_info_t<float> * matrix_info) {
@@ -1754,48 +1774,28 @@ namespace dpct
matrix_info->ld_info[2] = ldc;
matrix_info->groupsize_info = batch_size;
#ifdef GGML_SYCL_NVIDIA
sycl::event e = oneapi::mkl::blas::column_major::gemm_batch(
oneapi::mkl::backend_selector<oneapi::mkl::backend::cublas>{ q }, matrix_info->transpose_info,
matrix_info->transpose_info + 1, matrix_info->size_info, matrix_info->size_info + 1,
matrix_info->size_info + 2, reinterpret_cast<Ts *>(matrix_info->value_info),
reinterpret_cast<const Ta **>(a), matrix_info->ld_info, reinterpret_cast<const Tb **>(b),
matrix_info->ld_info + 1, reinterpret_cast<Ts *>(matrix_info->value_info + 1),
reinterpret_cast<Tc **>(c), matrix_info->ld_info + 2, 1, &(matrix_info->groupsize_info));
#else
sycl::event e = oneapi::mkl::blas::column_major::gemm_batch(
q, matrix_info->transpose_info, matrix_info->transpose_info + 1, matrix_info->size_info,
matrix_info->size_info + 1, matrix_info->size_info + 2, reinterpret_cast<Ts *>(matrix_info->value_info),
reinterpret_cast<const Ta **>(a), matrix_info->ld_info, reinterpret_cast<const Tb **>(b),
matrix_info->ld_info + 1, reinterpret_cast<Ts *>(matrix_info->value_info + 1),
reinterpret_cast<Tc **>(c), matrix_info->ld_info + 2, 1, &(matrix_info->groupsize_info));
#endif
sycl::event e = oneapi::math::blas::column_major::gemm_batch(
get_onemath_backend(q), matrix_info->transpose_info, matrix_info->transpose_info + 1,
matrix_info->size_info, matrix_info->size_info + 1, matrix_info->size_info + 2,
reinterpret_cast<Ts *>(matrix_info->value_info), reinterpret_cast<const Ta **>(a), matrix_info->ld_info,
reinterpret_cast<const Tb **>(b), matrix_info->ld_info + 1,
reinterpret_cast<Ts *>(matrix_info->value_info + 1), reinterpret_cast<Tc **>(c),
matrix_info->ld_info + 2, 1, &(matrix_info->groupsize_info));
}
template <class Ta, class Tb, class Tc, class Ts>
inline void
gemm_batch_impl(sycl::queue &q, oneapi::mkl::transpose a_trans,
oneapi::mkl::transpose b_trans, int m, int n,
int k, const void *alpha, const void *a, int lda,
long long int stride_a, const void *b, int ldb,
long long int stride_b, const void *beta, void *c,
int ldc, long long int stride_c, int batch_size)
{
inline void gemm_batch_impl(sycl::queue & q, oneapi::math::transpose a_trans, oneapi::math::transpose b_trans,
int m, int n, int k, const void * alpha, const void * a, int lda,
long long int stride_a, const void * b, int ldb, long long int stride_b,
const void * beta, void * c, int ldc, long long int stride_c, int batch_size) {
Ts alpha_value = dpct::get_value(reinterpret_cast<const Ts *>(alpha), q);
Ts beta_value = dpct::get_value(reinterpret_cast<const Ts *>(beta), q);
auto data_a = get_memory<const Ta>(a);
auto data_b = get_memory<const Tb>(b);
auto data_c = get_memory<Tc>(c);
#ifdef GGML_SYCL_NVIDIA
oneapi::mkl::blas::column_major::gemm_batch(
oneapi::mkl::backend_selector<oneapi::mkl::backend::cublas>{ q }, a_trans, b_trans, m, n, k,
alpha_value, data_a, lda, stride_a, data_b, ldb, stride_b, beta_value, data_c, ldc, stride_c,
batch_size);
#else
oneapi::mkl::blas::column_major::gemm_batch(q, a_trans, b_trans, m, n, k, alpha_value, data_a, lda,
stride_a, data_b, ldb, stride_b, beta_value, data_c, ldc,
stride_c, batch_size);
#endif
oneapi::math::blas::column_major::gemm_batch(get_onemath_backend(q), a_trans, b_trans, m, n, k, alpha_value,
data_a, lda, stride_a, data_b, ldb, stride_b, beta_value,
data_c, ldc, stride_c, batch_size);
}
} // namespace detail
@@ -2259,13 +2259,10 @@ namespace dpct
sycl::range<3>(x, y, 1), direction);
}
inline void gemm(sycl::queue &q, oneapi::mkl::transpose a_trans,
oneapi::mkl::transpose b_trans, int m, int n, int k,
const void *alpha, const void *a, library_data_t a_type,
int lda, const void *b, library_data_t b_type, int ldb,
const void *beta, void *c, library_data_t c_type, int ldc,
library_data_t scaling_type)
{
inline void gemm(sycl::queue & q, oneapi::math::transpose a_trans, oneapi::math::transpose b_trans, int m, int n,
int k, const void * alpha, const void * a, library_data_t a_type, int lda, const void * b,
library_data_t b_type, int ldb, const void * beta, void * c, library_data_t c_type, int ldc,
library_data_t scaling_type) {
if (scaling_type == library_data_t::real_float &&
c_type == library_data_t::complex_float)
{
@@ -2329,9 +2326,8 @@ namespace dpct
library_data_t::real_bfloat16, library_data_t::real_bfloat16,
library_data_t::real_float, library_data_t::real_float):
{
detail::gemm_impl<oneapi::mkl::bfloat16, oneapi::mkl::bfloat16, float,
float>(q, a_trans, b_trans, m, n, k, alpha, a, lda, b,
ldb, beta, c, ldc);
detail::gemm_impl<oneapi::math::bfloat16, oneapi::math::bfloat16, float, float>(
q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
break;
}
case detail::get_type_combination_id(
@@ -2369,8 +2365,7 @@ namespace dpct
library_data_t::real_bfloat16, library_data_t::real_bfloat16,
library_data_t::real_bfloat16, library_data_t::real_float):
{
detail::gemm_impl<oneapi::mkl::bfloat16, oneapi::mkl::bfloat16,
oneapi::mkl::bfloat16, float>(
detail::gemm_impl<oneapi::math::bfloat16, oneapi::math::bfloat16, oneapi::math::bfloat16, float>(
q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
break;
}
@@ -2390,7 +2385,7 @@ namespace dpct
default:
throw std::runtime_error("the combination of data type is unsupported");
}
} // gemm()
} // gemm()
/// Computes a batch of matrix-matrix product with general matrices.
/// \param [in] q The queue where the routine should be executed.
@@ -2412,7 +2407,7 @@ namespace dpct
/// \param [in] ldc Leading dimension of C.
/// \param [in] batch_size Specifies the number of matrix multiply operations to perform.
/// \param [in] scaling_type Data type of the scaling factors.
inline void gemm_batch(sycl::queue & q, oneapi::mkl::transpose a_trans, oneapi::mkl::transpose b_trans, int m,
inline void gemm_batch(sycl::queue & q, oneapi::math::transpose a_trans, oneapi::math::transpose b_trans, int m,
int n, int k, const void * alpha, const void * a[], library_data_t a_type, int lda,
const void * b[], library_data_t b_type, int ldb, const void * beta, void * c[],
library_data_t c_type, int ldc, int batch_size, library_data_t scaling_type,
@@ -2450,7 +2445,7 @@ namespace dpct
library_data_t::real_bfloat16, library_data_t::real_bfloat16,
library_data_t::real_bfloat16, library_data_t::real_float):
{
detail::gemm_batch_impl<oneapi::mkl::bfloat16, oneapi::mkl::bfloat16, oneapi::mkl::bfloat16, float>(
detail::gemm_batch_impl<oneapi::math::bfloat16, oneapi::math::bfloat16, oneapi::math::bfloat16, float>(
q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc, batch_size, matrix_info);
break;
}
@@ -2458,7 +2453,7 @@ namespace dpct
library_data_t::real_bfloat16, library_data_t::real_bfloat16,
library_data_t::real_float, library_data_t::real_float):
{
detail::gemm_batch_impl<oneapi::mkl::bfloat16, oneapi::mkl::bfloat16, float, float>(
detail::gemm_batch_impl<oneapi::math::bfloat16, oneapi::math::bfloat16, float, float>(
q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc, batch_size, matrix_info);
break;
}
@@ -2534,15 +2529,11 @@ namespace dpct
/// \param [in] stride_c Stride between the different C matrices.
/// \param [in] batch_size Specifies the number of matrix multiply operations to perform.
/// \param [in] scaling_type Data type of the scaling factors.
inline void gemm_batch(sycl::queue &q, oneapi::mkl::transpose a_trans,
oneapi::mkl::transpose b_trans, int m, int n, int k,
const void *alpha, const void *a, library_data_t a_type,
int lda, long long int stride_a, const void *b,
library_data_t b_type, int ldb, long long int stride_b,
const void *beta, void *c, library_data_t c_type,
int ldc, long long int stride_c, int batch_size,
library_data_t scaling_type)
{
inline void gemm_batch(sycl::queue & q, oneapi::math::transpose a_trans, oneapi::math::transpose b_trans, int m,
int n, int k, const void * alpha, const void * a, library_data_t a_type, int lda,
long long int stride_a, const void * b, library_data_t b_type, int ldb,
long long int stride_b, const void * beta, void * c, library_data_t c_type, int ldc,
long long int stride_c, int batch_size, library_data_t scaling_type) {
if (scaling_type == library_data_t::real_float &&
c_type == library_data_t::complex_float)
{
@@ -2611,20 +2602,18 @@ namespace dpct
library_data_t::real_bfloat16, library_data_t::real_bfloat16,
library_data_t::real_bfloat16, library_data_t::real_float):
{
detail::gemm_batch_impl<oneapi::mkl::bfloat16, oneapi::mkl::bfloat16,
oneapi::mkl::bfloat16, float>(
q, a_trans, b_trans, m, n, k, alpha, a, lda, stride_a, b, ldb, stride_b,
beta, c, ldc, stride_c, batch_size);
detail::gemm_batch_impl<oneapi::math::bfloat16, oneapi::math::bfloat16, oneapi::math::bfloat16, float>(
q, a_trans, b_trans, m, n, k, alpha, a, lda, stride_a, b, ldb, stride_b, beta, c, ldc, stride_c,
batch_size);
break;
}
case detail::get_type_combination_id(
library_data_t::real_bfloat16, library_data_t::real_bfloat16,
library_data_t::real_float, library_data_t::real_float):
{
detail::gemm_batch_impl<oneapi::mkl::bfloat16, oneapi::mkl::bfloat16, float,
float>(q, a_trans, b_trans, m, n, k, alpha, a, lda,
stride_a, b, ldb, stride_b, beta, c, ldc,
stride_c, batch_size);
detail::gemm_batch_impl<oneapi::math::bfloat16, oneapi::math::bfloat16, float, float>(
q, a_trans, b_trans, m, n, k, alpha, a, lda, stride_a, b, ldb, stride_b, beta, c, ldc, stride_c,
batch_size);
break;
}
#endif
+185 -273
View File
@@ -509,497 +509,409 @@ static void pad_f32_sycl(const float *x, float *dst, const int ne00,
});
}
inline void ggml_sycl_op_silu(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
ggml_tensor *dst, const float *src0_dd,
const float *src1_dd, float *dst_dd,
const queue_ptr &main_stream) {
inline void ggml_sycl_op_silu(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
silu_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
silu_f32_sycl(src0_dd, dst_dd, ggml_nelements(dst->src[0]), main_stream);
}
inline void ggml_sycl_op_gelu(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
ggml_tensor *dst, const float *src0_dd,
const float *src1_dd, float *dst_dd,
const queue_ptr &main_stream) {
inline void ggml_sycl_op_gelu(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
gelu_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
}
inline void ggml_sycl_op_gelu_quick(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1, ggml_tensor *dst,
const float *src0_dd, const float *src1_dd,
float *dst_dd,
const queue_ptr &main_stream) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
gelu_quick_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
gelu_f32_sycl(src0_dd, dst_dd, ggml_nelements(dst->src[0]), main_stream);
}
inline void ggml_sycl_op_tanh(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
ggml_tensor *dst, const float *src0_dd,
const float *src1_dd, float *dst_dd,
const queue_ptr &main_stream) {
inline void ggml_sycl_op_gelu_quick(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
tanh_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
gelu_quick_f32_sycl(src0_dd, dst_dd, ggml_nelements(dst->src[0]), main_stream);
}
inline void ggml_sycl_op_relu(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
ggml_tensor *dst, const float *src0_dd,
const float *src1_dd, float *dst_dd,
const queue_ptr &main_stream) {
inline void ggml_sycl_op_tanh(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
relu_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
tanh_f32_sycl(src0_dd, dst_dd, ggml_nelements(dst->src[0]), main_stream);
}
inline void ggml_sycl_op_hardsigmoid(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1, ggml_tensor *dst,
const float *src0_dd, const float *src1_dd,
float *dst_dd,
const queue_ptr &main_stream) {
inline void ggml_sycl_op_relu(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
hardsigmoid_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
relu_f32_sycl(src0_dd, dst_dd, ggml_nelements(dst->src[0]), main_stream);
}
inline void ggml_sycl_op_hardswish(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1, ggml_tensor *dst,
const float *src0_dd, const float *src1_dd,
float *dst_dd, const queue_ptr &main_stream) {
inline void ggml_sycl_op_hardsigmoid(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
hardswish_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
hardsigmoid_f32_sycl(src0_dd, dst_dd, ggml_nelements(dst->src[0]), main_stream);
}
inline void ggml_sycl_op_exp(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1, ggml_tensor *dst,
const float *src0_dd, const float *src1_dd,
float *dst_dd, const queue_ptr &main_stream) {
inline void ggml_sycl_op_hardswish(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
exp_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
hardswish_f32_sycl(src0_dd, dst_dd, ggml_nelements(dst->src[0]), main_stream);
}
inline void ggml_sycl_op_log(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1, ggml_tensor *dst,
const float *src0_dd, const float *src1_dd,
float *dst_dd, const queue_ptr &main_stream) {
inline void ggml_sycl_op_exp(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
log_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
exp_f32_sycl(src0_dd, dst_dd, ggml_nelements(dst->src[0]), main_stream);
}
inline void ggml_sycl_op_sigmoid(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1, ggml_tensor *dst,
const float *src0_dd, const float *src1_dd,
float *dst_dd, const queue_ptr &main_stream) {
inline void ggml_sycl_op_log(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
sigmoid_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
log_f32_sycl(src0_dd, dst_dd, ggml_nelements(dst->src[0]), main_stream);
}
inline void ggml_sycl_op_sqrt(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1, ggml_tensor *dst,
const float *src0_dd, const float *src1_dd,
float *dst_dd, const queue_ptr &main_stream) {
inline void ggml_sycl_op_sigmoid(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
sqrt_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
sigmoid_f32_sycl(src0_dd, dst_dd, ggml_nelements(dst->src[0]), main_stream);
}
inline void ggml_sycl_op_sin(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1, ggml_tensor *dst,
const float *src0_dd, const float *src1_dd,
float *dst_dd, const queue_ptr &main_stream) {
inline void ggml_sycl_op_sqrt(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
sin_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
sqrt_f32_sycl(src0_dd, dst_dd, ggml_nelements(dst->src[0]), main_stream);
}
inline void ggml_sycl_op_cos(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1, ggml_tensor *dst,
const float *src0_dd, const float *src1_dd,
float *dst_dd, const queue_ptr &main_stream) {
inline void ggml_sycl_op_sin(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
cos_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
sin_f32_sycl(src0_dd, dst_dd, ggml_nelements(dst->src[0]), main_stream);
}
inline void ggml_sycl_op_step(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1, ggml_tensor *dst,
const float *src0_dd, const float *src1_dd,
float *dst_dd, const queue_ptr &main_stream) {
inline void ggml_sycl_op_cos(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
step_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
cos_f32_sycl(src0_dd, dst_dd, ggml_nelements(dst->src[0]), main_stream);
}
inline void ggml_sycl_op_neg(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1, ggml_tensor *dst,
const float *src0_dd, const float *src1_dd,
float *dst_dd, const queue_ptr &main_stream) {
inline void ggml_sycl_op_step(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
neg_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
step_f32_sycl(src0_dd, dst_dd, ggml_nelements(dst->src[0]), main_stream);
}
inline void ggml_sycl_op_leaky_relu(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1, ggml_tensor *dst,
const float *src0_dd, const float *src1_dd,
float *dst_dd,
const queue_ptr &main_stream) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
inline void ggml_sycl_op_neg(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
neg_f32_sycl(src0_dd, dst_dd, ggml_nelements(dst->src[0]), main_stream);
}
inline void ggml_sycl_op_leaky_relu(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
float negative_slope;
memcpy(&negative_slope, dst->op_params, sizeof(float));
leaky_relu_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), negative_slope, main_stream);
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
leaky_relu_f32_sycl(src0_dd, dst_dd, ggml_nelements(dst->src[0]), negative_slope, main_stream);
}
inline void ggml_sycl_op_sqr(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
ggml_tensor *dst, const float *src0_dd,
const float *src1_dd, float *dst_dd,
const queue_ptr &main_stream) {
inline void ggml_sycl_op_sqr(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
sqr_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
sqr_f32_sycl(src0_dd, dst_dd, ggml_nelements(dst->src[0]), main_stream);
}
inline void ggml_sycl_op_upscale(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1, ggml_tensor *dst,
const float *src0_dd, const float *src1_dd,
float *dst_dd,
const queue_ptr &main_stream) {
inline void ggml_sycl_op_upscale(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
const float sf0 = (float)dst->ne[0]/src0->ne[0];
const float sf1 = (float)dst->ne[1]/src0->ne[1];
const float sf2 = (float)dst->ne[2]/src0->ne[2];
const float sf3 = (float)dst->ne[3]/src0->ne[3];
const float sf0 = (float)dst->ne[0]/dst->src[0]->ne[0];
const float sf1 = (float)dst->ne[1]/dst->src[0]->ne[1];
const float sf2 = (float)dst->ne[2]/dst->src[0]->ne[2];
const float sf3 = (float)dst->ne[3]/dst->src[0]->ne[3];
upscale_f32_sycl(src0_dd, dst_dd, src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
upscale_f32_sycl(src0_dd, dst_dd, dst->src[0]->nb[0], dst->src[0]->nb[1], dst->src[0]->nb[2], dst->src[0]->nb[3],
dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], sf0, sf1, sf2, sf3,
main_stream);
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
}
inline void ggml_sycl_op_pad(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
ggml_tensor *dst, const float *src0_dd,
const float *src1_dd, float *dst_dd,
const queue_ptr &main_stream) {
inline void ggml_sycl_op_pad(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors
GGML_ASSERT(dst->src[0]->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
pad_f32_sycl(src0_dd, dst_dd,
src0->ne[0], src0->ne[1], src0->ne[2],
dst->src[0]->ne[0], dst->src[0]->ne[1], dst->src[0]->ne[2],
dst->ne[0], dst->ne[1], dst->ne[2], main_stream);
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
}
inline void ggml_sycl_op_acc(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
ggml_tensor *dst, const float *src0_dd,
const float *src1_dd, float *dst_dd,
const queue_ptr &main_stream) {
inline void ggml_sycl_op_acc(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[1]->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
GGML_ASSERT(dst->ne[3] == 1); // just 3D tensors supported
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
const float * src1_dd = static_cast<const float*>(dst->src[1]->data);
float * dst_dd = static_cast<float *>(dst->data);
int nb1 = dst->op_params[0] / 4; // 4 bytes of float32
int nb2 = dst->op_params[1] / 4; // 4 bytes of float32
// int nb3 = dst->op_params[2] / 4; // 4 bytes of float32 - unused
int offset = dst->op_params[3] / 4; // offset in bytes
acc_f32_sycl(src0_dd, src1_dd, dst_dd, ggml_nelements(dst), src1->ne[0], src1->ne[1], src1->ne[2], nb1, nb2, offset, main_stream);
GGML_UNUSED(dst);
GGML_UNUSED(ctx);
acc_f32_sycl(src0_dd, src1_dd, dst_dd, ggml_nelements(dst), dst->src[1]->ne[0], dst->src[1]->ne[1], dst->src[1]->ne[2], nb1, nb2, offset, main_stream);
}
inline void ggml_sycl_op_add(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
ggml_tensor *dst, const float *src0_dd,
const float *src1_dd, float *dst_dd,
const queue_ptr &main_stream) {
inline void ggml_sycl_op_add(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_add>>(ctx, src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream);
ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_add>>(ctx, dst->src[0], dst->src[1], dst);
}
inline void ggml_sycl_op_sub(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
ggml_tensor *dst, const float *src0_dd,
const float *src1_dd, float *dst_dd,
const queue_ptr &main_stream) {
inline void ggml_sycl_op_sub(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_sub>>(ctx, src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream);
ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_sub>>(ctx, dst->src[0], dst->src[1], dst);
}
inline void ggml_sycl_op_mul(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
ggml_tensor *dst, const float *src0_dd,
const float *src1_dd, float *dst_dd,
const queue_ptr &main_stream) {
inline void ggml_sycl_op_mul(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_mul>>(ctx, src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream);
ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_mul>>(ctx, dst->src[0], dst->src[1], dst);
}
inline void ggml_sycl_op_div(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
ggml_tensor *dst, const float *src0_dd,
const float *src1_dd, float *dst_dd,
const queue_ptr &main_stream) {
inline void ggml_sycl_op_div(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_div>>(ctx, src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream);
ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_div>>(ctx, dst->src[0], dst->src[1], dst);
}
void ggml_sycl_sqrt(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_sqrt);
ggml_sycl_op_sqrt(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
void ggml_sycl_sin(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_sin);
ggml_sycl_op_sin(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
void ggml_sycl_cos(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_cos);
ggml_sycl_op_cos(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
void ggml_sycl_acc(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_acc);
ggml_sycl_op_acc(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
void ggml_sycl_gelu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_gelu);
ggml_sycl_op_gelu(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
void ggml_sycl_silu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_silu);
ggml_sycl_op_silu(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
void ggml_sycl_gelu_quick(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_gelu_quick);
ggml_sycl_op_gelu_quick(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
void ggml_sycl_tanh(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_tanh);
ggml_sycl_op_tanh(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
void ggml_sycl_relu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_relu);
ggml_sycl_op_relu(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
void ggml_sycl_sigmoid(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_sigmoid);
ggml_sycl_op_sigmoid(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
void ggml_sycl_hardsigmoid(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_hardsigmoid);
ggml_sycl_op_hardsigmoid(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
void ggml_sycl_hardswish(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_hardswish);
ggml_sycl_op_hardswish(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
void ggml_sycl_exp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_exp);
ggml_sycl_op_exp(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
void ggml_sycl_log(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_log);
ggml_sycl_op_log(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
void ggml_sycl_neg(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_neg);
ggml_sycl_op_neg(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
void ggml_sycl_step(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_step);
ggml_sycl_op_step(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
void ggml_sycl_leaky_relu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_leaky_relu);
ggml_sycl_op_leaky_relu(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
void ggml_sycl_sqr(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_sqr);
ggml_sycl_op_sqr(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
void ggml_sycl_upscale(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_upscale);
ggml_sycl_op_upscale(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
void ggml_sycl_pad(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_pad);
ggml_sycl_op_pad(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
@@ -1007,24 +919,24 @@ void ggml_sycl_pad(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
void ggml_sycl_add(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_add);
ggml_sycl_op_add(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
void ggml_sycl_sub(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_sub);
ggml_sycl_op_sub(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
void ggml_sycl_mul(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_mul);
ggml_sycl_op_mul(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
void ggml_sycl_div(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_div);
ggml_sycl_op_div(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
+24 -20
View File
@@ -257,50 +257,54 @@ static void get_rows_sycl_float(ggml_backend_sycl_context & ctx, const ggml_tens
GGML_UNUSED(ctx);
}
void ggml_sycl_op_get_rows(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1, ggml_tensor *dst,
const float *src0_d, const float *src1_d,
float *dst_d, const queue_ptr &stream) {
void ggml_sycl_op_get_rows(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
GGML_ASSERT(src1->type == GGML_TYPE_I32);
GGML_ASSERT(dst->src[1]->type == GGML_TYPE_I32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
GGML_ASSERT(src1->nb[0] == ggml_type_size(src1->type));
GGML_ASSERT(dst->src[0]->nb[0] == ggml_type_size(dst->src[0]->type));
GGML_ASSERT(dst->src[1]->nb[0] == ggml_type_size(dst->src[1]->type));
GGML_ASSERT(dst->nb[0] == ggml_type_size(dst->type));
const int32_t * src1_i32 = (const int32_t *) src1_d;
switch (src0->type) {
const int32_t * src1_i32 = (const int32_t *) dst->src[1]->data;
/* TODO: Refactor and remove duplicates */
switch (dst->src[0]->type) {
case GGML_TYPE_F16:
get_rows_sycl_float(ctx, src0, src1, dst, (const sycl::half *)src0_d,
src1_i32, dst_d, stream);
get_rows_sycl_float(ctx, dst->src[0], dst->src[1], dst, (const sycl::half *)dst->src[0]->data,
src1_i32, (float *)dst->data, ctx.stream());
break;
case GGML_TYPE_F32:
get_rows_sycl_float(ctx, src0, src1, dst, src0_d, src1_i32, dst_d, stream);
get_rows_sycl_float(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data,
src1_i32, (float *)dst->data, ctx.stream());
break;
case GGML_TYPE_Q4_0:
if (ctx.opt_feature.reorder && dst->op == GGML_OP_MUL_MAT) {
get_rows_sycl_reorder<QK4_0, QR4_0, dequantize_q4_0_reorder>(ctx, src0, src1, dst, src0_d, src1_i32, dst_d, stream);
get_rows_sycl_reorder<QK4_0, QR4_0, dequantize_q4_0_reorder>(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data,
src1_i32, (float *)dst->data, ctx.stream());
} else {
get_rows_sycl<QK4_0, QR4_0, dequantize_q4_0>(ctx, src0, src1, dst, src0_d, src1_i32, dst_d, stream);
get_rows_sycl<QK4_0, QR4_0, dequantize_q4_0>(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data,
src1_i32, (float *)dst->data, ctx.stream());
}
break;
case GGML_TYPE_Q4_1:
get_rows_sycl<QK4_1, QR4_1, dequantize_q4_1>(ctx, src0, src1, dst, src0_d, src1_i32, dst_d, stream);
get_rows_sycl<QK4_1, QR4_1, dequantize_q4_1>(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data,
src1_i32, (float *)dst->data, ctx.stream());
break;
case GGML_TYPE_Q5_0:
get_rows_sycl<QK5_0, QR5_0, dequantize_q5_0>(ctx, src0, src1, dst, src0_d, src1_i32, dst_d, stream);
get_rows_sycl<QK5_0, QR5_0, dequantize_q5_0>(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data,
src1_i32, (float *)dst->data, ctx.stream());
break;
case GGML_TYPE_Q5_1:
get_rows_sycl<QK5_1, QR5_1, dequantize_q5_1>(ctx, src0, src1, dst, src0_d, src1_i32, dst_d, stream);
get_rows_sycl<QK5_1, QR5_1, dequantize_q5_1>(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data,
src1_i32, (float *)dst->data, ctx.stream());
break;
case GGML_TYPE_Q8_0:
get_rows_sycl<QK8_0, QR8_0, dequantize_q8_0>(ctx, src0, src1, dst, src0_d, src1_i32, dst_d, stream);
get_rows_sycl<QK8_0, QR8_0, dequantize_q8_0>(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data,
src1_i32, (float *)dst->data, ctx.stream());
break;
default:
// TODO: k-quants
GGML_LOG_ERROR("%s: unsupported type: %s\n", __func__, ggml_type_name(src0->type));
GGML_LOG_ERROR("%s: unsupported type: %s\n", __func__, ggml_type_name(dst->src[0]->type));
GGML_ABORT("fatal error");
}
}
+1 -4
View File
@@ -15,9 +15,6 @@
#include "common.hpp"
void ggml_sycl_op_get_rows(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1, ggml_tensor *dst,
const float *src0_d, const float *src1_d,
float *dst_d, const queue_ptr &stream);
void ggml_sycl_op_get_rows(ggml_backend_sycl_context & ctx, ggml_tensor *dst);
#endif // GGML_SYCL_GETROWS_HPP
+96 -149
View File
@@ -1988,16 +1988,8 @@ catch (sycl::exception const &exc) {
std::exit(1);
}
static void ggml_sycl_op_repeat(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1, ggml_tensor *dst,
const float *src0_d, const float *src1_d,
float *dst_d,
const queue_ptr &main_stream) {
ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_repeat>>(ctx, dst, src0, dst, nullptr, src0_d, dst_d, main_stream);
GGML_UNUSED(src1);
GGML_UNUSED(src1_d);
static void ggml_sycl_op_repeat(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_repeat>>(ctx, dst, dst->src[0], dst);
}
@@ -2067,8 +2059,8 @@ inline void ggml_sycl_op_mul_mat_sycl(
const sycl::half alpha_f16 = 1.0f;
const sycl::half beta_f16 = 0.0f;
SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm(
*stream, oneapi::mkl::transpose::trans,
oneapi::mkl::transpose::nontrans, row_diff, src1_ncols, ne10,
*stream, oneapi::math::transpose::trans,
oneapi::math::transpose::nontrans, row_diff, src1_ncols, ne10,
&alpha_f16, src0_ptr, dpct::library_data_t::real_half, ne00,
src1_ptr, dpct::library_data_t::real_half, ne10, &beta_f16,
dst_f16.get(), dpct::library_data_t::real_half, ldc,
@@ -2105,17 +2097,10 @@ inline void ggml_sycl_op_mul_mat_sycl(
#if !GGML_SYCL_DNNL
const float alpha = 1.0f;
const float beta = 0.0f;
# ifdef GGML_SYCL_NVIDIA
SYCL_CHECK(CHECK_TRY_ERROR(oneapi::mkl::blas::column_major::gemm(
oneapi::mkl::backend_selector<oneapi::mkl::backend::cublas>{ *stream }, oneapi::mkl::transpose::trans,
oneapi::mkl::transpose::nontrans, row_diff, src1_ncols, ne10, dpct::get_value(&alpha, *stream), src0_ddf_i,
ne00, src1_ddf1_i, ne10, dpct::get_value(&beta, *stream), dst_dd_i, ldc)));
# else
SYCL_CHECK(CHECK_TRY_ERROR(oneapi::mkl::blas::column_major::gemm(
*stream, oneapi::mkl::transpose::trans, oneapi::mkl::transpose::nontrans, row_diff, src1_ncols, ne10,
dpct::get_value(&alpha, *stream), src0_ddf_i, ne00, src1_ddf1_i, ne10, dpct::get_value(&beta, *stream),
dst_dd_i, ldc)));
# endif
SYCL_CHECK(CHECK_TRY_ERROR(oneapi::math::blas::column_major::gemm(
get_onemath_backend(*stream), oneapi::math::transpose::trans, oneapi::math::transpose::nontrans, row_diff,
src1_ncols, ne10, dpct::get_value(&alpha, *stream), src0_ddf_i, ne00, src1_ddf1_i, ne10,
dpct::get_value(&beta, *stream), dst_dd_i, ldc)));
#else
DnnlGemmWrapper::row_gemm(ctx, false, true, src1_ncols, row_diff, ne10, src1_ddf1_i,
DnnlGemmWrapper::to_dt<float>(), src0_ddf_i, DnnlGemmWrapper::to_dt<float>(),
@@ -2132,13 +2117,14 @@ catch (sycl::exception const &exc) {
std::exit(1);
}
static void ggml_sycl_op_pool2d(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1, ggml_tensor *dst,
const float *src0_dd, const float *src1_dd,
float *dst_dd, const queue_ptr &main_stream) {
static void ggml_sycl_op_pool2d(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
const int32_t * opts = (const int32_t *)dst->op_params;
enum ggml_op_pool op = static_cast<ggml_op_pool>(opts[0]);
@@ -2149,8 +2135,8 @@ static void ggml_sycl_op_pool2d(ggml_backend_sycl_context & ctx, const ggml_tens
const int p0 = opts[5];
const int p1 = opts[6];
const int64_t IH = src0->ne[1];
const int64_t IW = src0->ne[0];
const int64_t IH = dst->src[0]->ne[1];
const int64_t IW = dst->src[0]->ne[0];
const int64_t N = dst->ne[3];
const int64_t OC = dst->ne[2];
@@ -2169,163 +2155,125 @@ static void ggml_sycl_op_pool2d(ggml_backend_sycl_context & ctx, const ggml_tens
parallel_elements, src0_dd, dst_dd, op,
item_ct1);
});
GGML_UNUSED(src1);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
}
inline void ggml_sycl_op_sum(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1, ggml_tensor *dst,
const float *src0_dd, const float *src1_dd,
float *dst_dd,
const queue_ptr &main_stream) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
inline void ggml_sycl_op_sum(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
const int64_t ne = ggml_nelements(src0);
const int64_t ne = ggml_nelements(dst->src[0]);
sum_rows_f32_sycl(src0_dd, dst_dd, ne, 1, main_stream);
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
}
inline void ggml_sycl_op_sum_rows(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1, ggml_tensor *dst,
const float *src0_dd, const float *src1_dd,
float *dst_dd,
const queue_ptr &main_stream) {
inline void ggml_sycl_op_sum_rows(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
const int64_t ncols = src0->ne[0];
const int64_t nrows = ggml_nrows(src0);
const int64_t ncols = dst->src[0]->ne[0];
const int64_t nrows = ggml_nrows(dst->src[0]);
sum_rows_f32_sycl(src0_dd, dst_dd, ncols, nrows, main_stream);
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
}
inline void ggml_sycl_op_argsort(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1, ggml_tensor *dst,
const float *src0_dd, const float *src1_dd,
float *dst_dd,
const queue_ptr &main_stream) {
inline void ggml_sycl_op_argsort(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_I32);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
int32_t * dst_dd = static_cast<int32_t *>(dst->data);
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_I32);
const int64_t ncols = src0->ne[0];
const int64_t nrows = ggml_nrows(src0);
const int64_t ncols = dst->src[0]->ne[0];
const int64_t nrows = ggml_nrows(dst->src[0]);
enum ggml_sort_order order = (enum ggml_sort_order) dst->op_params[0];
argsort_f32_i32_sycl(src0_dd, (int *)dst_dd, ncols, nrows, order, main_stream);
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
argsort_f32_i32_sycl(src0_dd, (int *) dst_dd, ncols, nrows, order, main_stream);
}
inline void ggml_sycl_op_argmax(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1, ggml_tensor *dst,
const float *src0_dd, const float *src1_dd,
float *dst_dd,
const queue_ptr &main_stream) {
inline void ggml_sycl_op_argmax(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_I32);
const int64_t ncols = src0->ne[0];
const int64_t nrows = ggml_nrows(src0);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
int32_t * dst_dd = static_cast<int32_t *>(dst->data);
argmax_f32_i32_sycl(src0_dd, (int *)dst_dd, ncols, nrows, main_stream);
const int64_t ncols = dst->src[0]->ne[0];
const int64_t nrows = ggml_nrows(dst->src[0]);
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
argmax_f32_i32_sycl(src0_dd, dst_dd, ncols, nrows, main_stream);
}
inline void ggml_sycl_op_diag_mask_inf(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1,
ggml_tensor *dst, const float *src0_dd,
const float *src1_dd, float *dst_dd,
const queue_ptr &main_stream) {
inline void ggml_sycl_op_diag_mask_inf(ggml_backend_sycl_context & ctx,ggml_tensor *dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
const int64_t ne00 = src0->ne[0];
const int64_t ne01 = src0->ne[1];
const int nrows0 = ggml_nrows(src0);
const int64_t ne00 = dst->src[0]->ne[0];
const int64_t ne01 = dst->src[0]->ne[1];
const int nrows0 = ggml_nrows(dst->src[0]);
const int n_past = ((int32_t *) dst->op_params)[0];
diag_mask_inf_f32_sycl(src0_dd, dst_dd, ne00, nrows0, ne01, n_past, main_stream);
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
}
inline void ggml_sycl_op_scale(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
ggml_tensor *dst, const float *src0_dd,
const float *src1_dd, float *dst_dd,
const queue_ptr &main_stream) {
inline void ggml_sycl_op_scale(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
float scale;
memcpy(&scale, dst->op_params, sizeof(float));
scale_f32_sycl(src0_dd, dst_dd, scale, ggml_nelements(src0), main_stream);
scale_f32_sycl(src0_dd, dst_dd, scale, ggml_nelements(dst->src[0]), main_stream);
/*
DPCT1010:87: SYCL uses exceptions to report errors and does not use the
error codes. The call was replaced with 0. You need to rewrite this code.
*/
SYCL_CHECK(0);
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
}
inline void ggml_sycl_op_clamp(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
ggml_tensor *dst, const float *src0_dd,
const float *src1_dd, float *dst_dd,
const queue_ptr &main_stream) {
inline void ggml_sycl_op_clamp(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
float min;
float max;
memcpy(&min, dst->op_params, sizeof(float));
memcpy(&max, (float *) dst->op_params + 1, sizeof(float));
clamp_f32_sycl(src0_dd, dst_dd, min, max, ggml_nelements(src0), main_stream);
clamp_f32_sycl(src0_dd, dst_dd, min, max, ggml_nelements(dst->src[0]), ctx.stream());
/*
DPCT1010:88: SYCL uses exceptions to report errors and does not use the
error codes. The call was replaced with 0. You need to rewrite this code.
*/
SYCL_CHECK(0);
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
}
static void ggml_sycl_set_peer_access(const int n_tokens, int main_device) {
@@ -2695,37 +2643,37 @@ catch (sycl::exception const &exc) {
static void ggml_sycl_repeat(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_repeat);
ggml_sycl_op_repeat(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
static void ggml_sycl_get_rows(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_get_rows);
ggml_sycl_op_get_rows(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
static void ggml_sycl_norm(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_norm);
ggml_sycl_op_norm(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
static void ggml_sycl_rms_norm(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_rms_norm);
ggml_sycl_op_rms_norm(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
static void ggml_sycl_l2_norm(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_l2_norm);
ggml_sycl_op_l2_norm(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
static void ggml_sycl_group_norm(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_group_norm);
ggml_sycl_op_group_norm(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
@@ -2881,14 +2829,10 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx,
if (r2 == 1 && r3 == 1 && ggml_is_contiguous_2(src0) && ggml_is_contiguous_2(src1)) {
// there is no broadcast and src0, src1 are contiguous across dims 2, 3
SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch(
*main_stream, oneapi::mkl::transpose::trans,
oneapi::mkl::transpose::nontrans, ne01, ne11, ne10, alpha,
(const char *)src0_as_f16, dpct::library_data_t::real_half,
nb01 / nb00, nb02 / nb00,
(const char *)src1_f16, dpct::library_data_t::real_half,
nb11 / nb10, nb12 / nb10, beta,
(char *)dst_t, cu_data_type, ne01, nb2 / nb0,
ne12 * ne13, cu_compute_type)));
*main_stream, oneapi::math::transpose::trans, oneapi::math::transpose::nontrans, ne01, ne11, ne10, alpha,
(const char *) src0_as_f16, dpct::library_data_t::real_half, nb01 / nb00, nb02 / nb00,
(const char *) src1_f16, dpct::library_data_t::real_half, nb11 / nb10, nb12 / nb10, beta, (char *) dst_t,
cu_data_type, ne01, nb2 / nb0, ne12 * ne13, cu_compute_type)));
} else {
const int ne23 = ne12*ne13;
@@ -2923,7 +2867,7 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx,
});
}
SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch(
*main_stream, oneapi::mkl::transpose::trans, oneapi::mkl::transpose::nontrans, ne01, ne11, ne10, alpha,
*main_stream, oneapi::math::transpose::trans, oneapi::math::transpose::nontrans, ne01, ne11, ne10, alpha,
(const void **) (ptrs_src.get() + 0 * ne23), dpct::library_data_t::real_half, nb01 / nb00,
(const void **) (ptrs_src.get() + 1 * ne23), dpct::library_data_t::real_half, nb11 / nb10, beta,
(void **) (ptrs_dst.get() + 0 * ne23), cu_data_type, ne01, ne23, cu_compute_type, matrix_info.get())));
@@ -3269,48 +3213,48 @@ catch (sycl::exception const &exc) {
}
static void ggml_sycl_scale(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_scale);
ggml_sycl_op_scale(ctx, dst);
}
static void ggml_sycl_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_clamp);
ggml_sycl_op_clamp(ctx, dst);
}
static void ggml_sycl_diag_mask_inf(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_diag_mask_inf);
ggml_sycl_op_diag_mask_inf(ctx, dst);
}
static void ggml_sycl_rope(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_ASSERT(ggml_is_contiguous(dst->src[0])); // TODO: this restriction is temporary until non-cont support is implemented
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_rope);
ggml_sycl_op_rope(ctx, dst);
}
static void ggml_sycl_pool2d(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_pool2d);
ggml_sycl_op_pool2d(ctx, dst);
}
static void ggml_sycl_im2col(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_im2col);
ggml_sycl_op_im2col(ctx, dst);
}
static void ggml_sycl_sum(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_ASSERT(ggml_is_contiguous(dst->src[0]));
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_sum);
ggml_sycl_op_sum(ctx, dst);
}
static void ggml_sycl_sum_rows(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_ASSERT(ggml_is_contiguous(dst->src[0]));
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_sum_rows);
ggml_sycl_op_sum_rows(ctx, dst);
}
static void ggml_sycl_argsort(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_ASSERT(ggml_is_contiguous(dst->src[0]));
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_argsort);
ggml_sycl_op_argsort(ctx, dst);
}
static void ggml_sycl_argmax(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_ASSERT(ggml_is_contiguous(dst->src[0]));
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_argmax);
ggml_sycl_op_argmax(ctx, dst);
}
@@ -3335,7 +3279,7 @@ catch (sycl::exception const &exc) {
std::exit(1);
}
static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct ggml_tensor * dst) {
static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct ggml_tensor * dst) try {
if (!g_sycl_loaded) return false;
if (dst->src[0] != nullptr && ggml_backend_buffer_is_sycl_split(dst->src[0]->buffer)) {
@@ -3528,6 +3472,9 @@ static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct gg
}
return true;
} catch (sycl::exception & e) {
std::cerr << e.what() << "Exception caught at file:" << __FILE__ << ", line:" << __LINE__ << std::endl;
std::exit(1);
}
GGML_API void ggml_backend_sycl_get_device_description(int device, char *description,
+5 -10
View File
@@ -82,10 +82,9 @@ static void im2col_sycl(
}
}
void ggml_sycl_op_im2col(
ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
ggml_tensor *dst, const float *src0_dd, const float *src1_dd, float *dst_dd,
const queue_ptr &main_stream) {
void ggml_sycl_op_im2col(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
GGML_ASSERT(src0->type == GGML_TYPE_F16);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
@@ -115,12 +114,8 @@ void ggml_sycl_op_im2col(
const size_t batch_offset = src1->nb[3] / 4; // nb is byte offset, src is type float32
if (dst->type == GGML_TYPE_F16) {
im2col_sycl(src1_dd, (sycl::half *)dst_dd, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, main_stream);
im2col_sycl((const float *) src1->data, (sycl::half *)dst->data, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, ctx.stream());
} else {
im2col_sycl(src1_dd, (float *)dst_dd, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, main_stream);
im2col_sycl((const float *) src1->data, (float *)dst->data, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, ctx.stream());
}
GGML_UNUSED(src0);
GGML_UNUSED(src0_dd);
GGML_UNUSED(ctx);
}
+1 -3
View File
@@ -16,8 +16,6 @@
#include "common.hpp"
void ggml_sycl_op_im2col(
ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
ggml_tensor *dst, const float *src0_dd, const float *src1_dd, float *dst_dd,
const queue_ptr &main_stream);
ggml_backend_sycl_context & ctx, ggml_tensor *dst);
#endif // GGML_SYCL_IM2COL_HPP
+37 -49
View File
@@ -367,7 +367,7 @@ static void l2_norm_f32_sycl(const float* x, float* dst, const int ncols,
sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims,
block_dims),
[=](sycl::nd_item<3> item_ct1)
[[intel::reqd_sub_group_size(WARP_SIZE)]] {
[[sycl::reqd_sub_group_size(WARP_SIZE)]] {
l2_norm_f32(x, dst, ncols, eps, item_ct1,
nullptr, WARP_SIZE);
});
@@ -389,7 +389,7 @@ static void l2_norm_f32_sycl(const float* x, float* dst, const int ncols,
sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims,
block_dims),
[=](sycl::nd_item<3> item_ct1)
[[intel::reqd_sub_group_size(WARP_SIZE)]] {
[[sycl::reqd_sub_group_size(WARP_SIZE)]] {
l2_norm_f32(x, dst, ncols, eps, item_ct1,
get_pointer(s_sum_acc_ct1), work_group_size);
});
@@ -397,90 +397,78 @@ static void l2_norm_f32_sycl(const float* x, float* dst, const int ncols,
}
}
void ggml_sycl_op_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0, const ggml_tensor* src1,
ggml_tensor* dst, const float* src0_dd,
const float* src1_dd, float* dst_dd,
const queue_ptr& main_stream) {
void ggml_sycl_op_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
const int64_t ne00 = src0->ne[0];
const int64_t nrows = ggml_nrows(src0);
const int64_t ne00 = dst->src[0]->ne[0];
const int64_t nrows = ggml_nrows(dst->src[0]);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
float eps;
memcpy(&eps, dst->op_params, sizeof(float));
norm_f32_sycl(src0_dd, dst_dd, ne00, nrows, eps, main_stream, ctx.device);
(void)src1;
(void)dst;
(void)src1_dd;
}
void ggml_sycl_op_group_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0,
const ggml_tensor* src1, ggml_tensor* dst,
const float* src0_dd, const float* src1_dd,
float* dst_dd,
const queue_ptr& main_stream) {
void ggml_sycl_op_group_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
int num_groups = dst->op_params[0];
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
float eps;
memcpy(&eps, dst->op_params + 1, sizeof(float));
int group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups);
group_norm_f32_sycl(src0_dd, dst_dd, num_groups, eps, group_size, src0->ne[0] * src0->ne[1] * src0->ne[2], main_stream, ctx.device);
(void)src1;
(void)dst;
(void)src1_dd;
GGML_UNUSED(ctx);
int group_size = dst->src[0]->ne[0] * dst->src[0]->ne[1] * ((dst->src[0]->ne[2] + num_groups - 1) / num_groups);
group_norm_f32_sycl(src0_dd, dst_dd, num_groups, eps, group_size, dst->src[0]->ne[0] * dst->src[0]->ne[1] * dst->src[0]->ne[2], main_stream, ctx.device);
}
void ggml_sycl_op_rms_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0,
const ggml_tensor* src1, ggml_tensor* dst,
const float* src0_dd, const float* src1_dd,
float* dst_dd,
const queue_ptr& main_stream) {
void ggml_sycl_op_rms_norm(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
const int64_t ne00 = src0->ne[0];
const int64_t nrows = ggml_nrows(src0);
const int64_t ne00 = dst->src[0]->ne[0];
const int64_t nrows = ggml_nrows(dst->src[0]);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
float eps;
memcpy(&eps, dst->op_params, sizeof(float));
rms_norm_f32_sycl(src0_dd, dst_dd, ne00, nrows, eps, main_stream, ctx.device);
(void)src1;
(void)dst;
(void)src1_dd;
}
void ggml_sycl_op_l2_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0,
const ggml_tensor* src1, ggml_tensor* dst,
const float* src0_dd, const float* src1_dd,
float* dst_dd,
const queue_ptr& main_stream) {
void ggml_sycl_op_l2_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
const int64_t ne00 = src0->ne[0];
const int64_t nrows = ggml_nrows(src0);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const int64_t ne00 = dst->src[0]->ne[0];
const int64_t nrows = ggml_nrows(dst->src[0]);
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
float eps;
memcpy(&eps, dst->op_params, sizeof(float));
l2_norm_f32_sycl(src0_dd, dst_dd, ne00, nrows, eps, main_stream, ctx.device);
(void)src1;
(void)dst;
(void)src1_dd;
}
+4 -19
View File
@@ -15,27 +15,12 @@
#include "common.hpp"
void ggml_sycl_op_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0, const ggml_tensor* src1,
ggml_tensor* dst, const float* src0_dd,
const float* src1_dd, float* dst_dd,
const queue_ptr& main_stream);
void ggml_sycl_op_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst);
void ggml_sycl_op_rms_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0,
const ggml_tensor* src1, ggml_tensor* dst,
const float* src0_dd, const float* src1_dd,
float* dst_dd,
const queue_ptr& main_stream);
void ggml_sycl_op_rms_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst);
void ggml_sycl_op_group_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0,
const ggml_tensor* src1, ggml_tensor* dst,
const float* src0_dd, const float* src1_dd,
float* dst_dd,
const queue_ptr& main_stream);
void ggml_sycl_op_group_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst);
void ggml_sycl_op_l2_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0,
const ggml_tensor* src1, ggml_tensor* dst,
const float* src0_dd, const float* src1_dd,
float* dst_dd,
const queue_ptr& main_stream);
void ggml_sycl_op_l2_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst);
#endif // GGML_SYCL_NORM_HPP
+4 -14
View File
@@ -1,8 +1,5 @@
#include <sycl/sycl.hpp>
#include <oneapi/mkl.hpp>
#include "outprod.hpp"
void ggml_sycl_op_out_prod(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
const ggml_tensor *src0 = dst->src[0];
const ggml_tensor *src1 = dst->src[1];
@@ -34,20 +31,13 @@ void ggml_sycl_op_out_prod(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
// Handle transposition of src1
const bool src1_T = ggml_is_transposed(src1);
const oneapi::mkl::transpose src1_op =
src1_T ? oneapi::mkl::transpose::nontrans : oneapi::mkl::transpose::trans;
const oneapi::math::transpose src1_op = src1_T ? oneapi::math::transpose::nontrans : oneapi::math::transpose::trans;
const int64_t ldb = (src1_T ? nb10 : nb11) / sizeof(float);
try {
// Perform matrix multiplication using oneMKL GEMM
#ifdef GGML_SYCL_NVIDIA
oneapi::mkl::blas::column_major::gemm(oneapi::mkl::backend_selector<oneapi::mkl::backend::cublas>{ *stream },
oneapi::mkl::transpose::nontrans, src1_op, ne0, ne1, ne01, alpha, src0_d,
ne00, src1_d, ldb, beta, dst_d, ne0);
#else
oneapi::mkl::blas::column_major::gemm(*stream, oneapi::mkl::transpose::nontrans, src1_op, ne0, ne1, ne01, alpha,
src0_d, ne00, src1_d, ldb, beta, dst_d, ne0);
#endif
// Perform matrix multiplication using oneMath GEMM
oneapi::math::blas::column_major::gemm(get_onemath_backend(*stream), oneapi::math::transpose::nontrans, src1_op,
ne0, ne1, ne01, alpha, src0_d, ne00, src1_d, ldb, beta, dst_d, ne0);
}
catch (sycl::exception const& exc) {
std::cerr << exc.what() << std::endl;
+20 -25
View File
@@ -192,18 +192,15 @@ static void rope_neox_sycl(
}
}
void ggml_sycl_op_rope(
ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst,
const float *src0_dd, const float *src1_dd, float *dst_dd, const queue_ptr &main_stream) {
const ggml_tensor * src2 = dst->src[2];
void ggml_sycl_op_rope(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
GGML_ASSERT( dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
GGML_ASSERT(src0->type == dst->type);
GGML_ASSERT(dst->src[0]->type == dst->type);
const int64_t ne00 = src0->ne[0];
const int64_t ne01 = src0->ne[1];
const int64_t nr = ggml_nrows(src0);
const int64_t ne00 = dst->src[0]->ne[0];
const int64_t ne01 = dst->src[0]->ne[1];
const int64_t nr = ggml_nrows(dst->src[0]);
//const int n_past = ((int32_t *) dst->op_params)[0];
const int n_dims = ((int32_t *) dst->op_params)[1];
@@ -228,49 +225,47 @@ void ggml_sycl_op_rope(
const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
const int32_t * pos = (const int32_t *) src1_dd;
const int32_t * pos = (const int32_t *) dst->src[1]->data;
const float * freq_factors = nullptr;
if (src2 != nullptr) {
freq_factors = (const float *) src2->data;
if (dst->src[2] != nullptr) {
freq_factors = (const float *) dst->src[2]->data;
}
rope_corr_dims corr_dims;
ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims.v);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
// compute
if (is_neox) {
if (src0->type == GGML_TYPE_F32) {
if (dst->src[0]->type == GGML_TYPE_F32) {
rope_neox_sycl(
(const float *)src0_dd, (float *)dst_dd, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor,
(const float *)dst->src[0]->data, (float *)dst->data, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor,
attn_factor, corr_dims, freq_factors, main_stream
);
} else if (src0->type == GGML_TYPE_F16) {
} else if (dst->src[0]->type == GGML_TYPE_F16) {
rope_neox_sycl(
(const sycl::half *)src0_dd, (sycl::half *)dst_dd, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor,
(const sycl::half *)dst->src[0]->data, (sycl::half *)dst->data, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor,
attn_factor, corr_dims, freq_factors, main_stream
);
} else {
GGML_ABORT("fatal error");
}
} else {
if (src0->type == GGML_TYPE_F32) {
if (dst->src[0]->type == GGML_TYPE_F32) {
rope_norm_sycl(
(const float *)src0_dd, (float *)dst_dd, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor,
(const float *)dst->src[0]->data, (float *)dst->data, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor,
attn_factor, corr_dims, freq_factors, main_stream
);
} else if (src0->type == GGML_TYPE_F16) {
} else if (dst->src[0]->type == GGML_TYPE_F16) {
rope_norm_sycl(
(const sycl::half *)src0_dd, (sycl::half *)dst_dd, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor,
(const sycl::half *)dst->src[0]->data, (sycl::half *)dst->data, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor,
attn_factor, corr_dims, freq_factors, main_stream
);
} else {
GGML_ABORT("fatal error");
}
}
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
}
+1 -3
View File
@@ -15,8 +15,6 @@
#include "common.hpp"
void ggml_sycl_op_rope(
ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst,
const float *src0_dd, const float *src1_dd, float *dst_dd, const queue_ptr &main_stream);
void ggml_sycl_op_rope(ggml_backend_sycl_context & ctx, ggml_tensor *dst);
#endif // GGML_SYCL_ROPE_HPP
+40 -30
View File
@@ -23,40 +23,50 @@ if (Vulkan_FOUND)
../../include/ggml-vulkan.h
)
if(NOT DEFINED GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
# Compile a test shader to determine whether GL_KHR_cooperative_matrix is supported.
# If it's not, there will be an error to stderr.
# If it's supported, set a define to indicate that we should compile those shaders
execute_process(COMMAND ${Vulkan_GLSLC_EXECUTABLE} -o - -fshader-stage=compute --target-env=vulkan1.3 "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_coopmat_support.comp"
OUTPUT_VARIABLE glslc_output
ERROR_VARIABLE glslc_error)
# Compile a test shader to determine whether GL_KHR_cooperative_matrix is supported.
# If it's not, there will be an error to stderr.
# If it's supported, set a define to indicate that we should compile those shaders
execute_process(COMMAND ${Vulkan_GLSLC_EXECUTABLE} -o - -fshader-stage=compute --target-env=vulkan1.3 "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_coopmat_support.comp"
OUTPUT_VARIABLE glslc_output
ERROR_VARIABLE glslc_error)
if (${glslc_error} MATCHES ".*extension not supported: GL_KHR_cooperative_matrix.*")
message(STATUS "GL_KHR_cooperative_matrix not supported by glslc")
set(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT OFF CACHE INTERNAL "Whether coopmat is supported by glslc")
else()
message(STATUS "GL_KHR_cooperative_matrix supported by glslc")
add_compile_definitions(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
set(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT ON CACHE INTERNAL "Whether coopmat is supported by glslc")
endif()
if (${glslc_error} MATCHES ".*extension not supported: GL_KHR_cooperative_matrix.*")
message(STATUS "GL_KHR_cooperative_matrix not supported by glslc")
set(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT OFF)
else()
message(STATUS "GL_KHR_cooperative_matrix supported by glslc")
set(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT ON)
add_compile_definitions(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
endif()
if(NOT DEFINED GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
# Compile a test shader to determine whether GL_NV_cooperative_matrix2 is supported.
# If it's not, there will be an error to stderr.
# If it's supported, set a define to indicate that we should compile those shaders
execute_process(COMMAND ${Vulkan_GLSLC_EXECUTABLE} -o - -fshader-stage=compute --target-env=vulkan1.3 "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_coopmat2_support.comp"
OUTPUT_VARIABLE glslc_output
ERROR_VARIABLE glslc_error)
# Compile a test shader to determine whether GL_NV_cooperative_matrix2 is supported.
# If it's not, there will be an error to stderr.
# If it's supported, set a define to indicate that we should compile those shaders
execute_process(COMMAND ${Vulkan_GLSLC_EXECUTABLE} -o - -fshader-stage=compute --target-env=vulkan1.3 "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_coopmat2_support.comp"
OUTPUT_VARIABLE glslc_output
ERROR_VARIABLE glslc_error)
if (${glslc_error} MATCHES ".*extension not supported: GL_NV_cooperative_matrix2.*")
message(STATUS "GL_NV_cooperative_matrix2 not supported by glslc")
set(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT OFF CACHE INTERNAL "Whether coopmat2 is supported by glslc")
else()
message(STATUS "GL_NV_cooperative_matrix2 supported by glslc")
add_compile_definitions(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
set(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT ON CACHE INTERNAL "Whether coopmat2 is supported by glslc")
endif()
if (${glslc_error} MATCHES ".*extension not supported: GL_NV_cooperative_matrix2.*")
message(STATUS "GL_NV_cooperative_matrix2 not supported by glslc")
set(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT OFF)
else()
message(STATUS "GL_NV_cooperative_matrix2 supported by glslc")
set(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT ON)
add_compile_definitions(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
endif()
# Compile a test shader to determine whether GL_EXT_integer_dot_product is supported.
# If it's not, there will be an error to stderr.
# If it's supported, set a define to indicate that we should compile those shaders
execute_process(COMMAND ${Vulkan_GLSLC_EXECUTABLE} -o - -fshader-stage=compute --target-env=vulkan1.3 "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_integer_dot_support.comp"
OUTPUT_VARIABLE glslc_output
ERROR_VARIABLE glslc_error)
if (${glslc_error} MATCHES ".*extension not supported: GL_EXT_integer_dot_product.*")
message(STATUS "GL_EXT_integer_dot_product not supported by glslc")
else()
message(STATUS "GL_EXT_integer_dot_product supported by glslc")
add_compile_definitions(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
endif()
target_link_libraries(ggml-vulkan PRIVATE Vulkan::Vulkan)
File diff suppressed because it is too large Load Diff
@@ -61,6 +61,10 @@ layout (push_constant) uniform parameter {
uint32_t n_head_log2;
float m0;
float m1;
uint32_t gqa_ratio;
uint32_t split_kv;
uint32_t k_num;
} p;
layout (binding = 0) readonly buffer Q {uint8_t data_q[];};
@@ -103,6 +107,38 @@ ACC_TYPE Max(const in uint32_t row, const in uint32_t col, const in ACC_TYPE ele
#define DECODEFUNC
#endif
// Store the output when doing grouped query attention.
// Rows index by Q's dimension 2, and the first N rows are valid.
D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N)
{
if (r < N && c < D) {
uint32_t offset = (iq2 + r) * D + c;
data_o[o_offset + offset] = D_TYPE(elem);
}
return elem;
}
// Store column zero. This is used to save per-row m and L values for split_k.
ACC_TYPE perElemOpStoreCol0(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N)
{
if (r < N && c == 0) {
uint32_t offset = iq2 + r;
data_o[o_offset + offset] = D_TYPE(elem);
}
return elem;
}
// Load the slope matrix, indexed by Q's dimension 2.
ACC_TYPE perElemOpComputeSlope(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t iq2)
{
const uint32_t h = iq2 + (r & (p.gqa_ratio - 1));
const ACC_TYPE base = ACC_TYPE(h < p.n_head_log2 ? p.m0 : p.m1);
const int exph = int(h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1);
return ACC_TYPE(pow(base, ACC_TYPE(exph)));
}
void main() {
#ifdef NEEDS_INIT_IQ_SHMEM
init_iq_shmem(gl_WorkGroupSize);
@@ -111,12 +147,22 @@ void main() {
const uint32_t N = p.N;
const uint32_t KV = p.KV;
uint32_t i = gl_WorkGroupID.x;
uint32_t split_k_index = 0;
if (p.k_num > 1) {
i = 0;
split_k_index = gl_WorkGroupID.x;
}
const uint32_t Tr = CEIL_DIV(N, Br);
const uint32_t Tc = CEIL_DIV(KV, Bc);
const uint32_t i = gl_WorkGroupID.x;
const uint32_t start_j = split_k_index * p.split_kv / Bc;
const uint32_t end_j = CEIL_DIV(min(KV, (split_k_index + 1) * p.split_kv), Bc);
const uint32_t iq2 = gl_WorkGroupID.y;
// When not using grouped query attention, all rows share the same iq2, equal to gl_WorkGroupID.y.
// When using grouped query attention, each workgroup does gqa_ratio consecutive values of iq2.
const uint32_t iq2 = gl_WorkGroupID.y * p.gqa_ratio;
const uint32_t iq3 = gl_WorkGroupID.z;
// broadcast factors
@@ -149,8 +195,10 @@ void main() {
tensorLayoutK = setTensorLayoutDimensionNV(tensorLayoutK, KV, D);
tensorLayoutV = setTensorLayoutDimensionNV(tensorLayoutV, KV, D);
// nb?1 are already divided by the type size and are in units of elements
uint32_t q_stride = p.nb01;
// nb?1 are already divided by the type size and are in units of elements.
// When using grouped query attention, Q is indexed by iq2, so the stride
// should be nb02 (which is in bytes).
uint32_t q_stride = p.gqa_ratio > 1 ? (p.nb02 / 4) : p.nb01;
uint32_t k_stride = p.nb11;
uint32_t v_stride = p.nb21;
// hint to the compiler that strides are aligned for the aligned variant of the shader
@@ -182,20 +230,15 @@ void main() {
L = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(0);
M = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(-1.0/0.0);
ACC_TYPE slope = ACC_TYPE(1.0);
coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> slopeMat = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(1.0);
// ALiBi
if (p.max_bias > 0.0f) {
const uint32_t h = iq2;
const ACC_TYPE base = ACC_TYPE(h < p.n_head_log2 ? p.m0 : p.m1);
const int exph = int(h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1);
slope = pow(base, ACC_TYPE(exph));
coopMatPerElementNV(slopeMat, slopeMat, perElemOpComputeSlope, iq2);
}
[[dont_unroll]]
for (uint32_t j = 0; j < Tc; ++j) {
for (uint32_t j = start_j; j < end_j; ++j) {
coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> S = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(0);
@@ -215,12 +258,16 @@ void main() {
if (p.mask != 0) {
tensorLayoutNV<2, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutM = createTensorLayoutNV(2, gl_CooperativeMatrixClampModeConstantNV);
tensorLayoutM = setTensorLayoutDimensionNV(tensorLayoutM, p.nem1, KV);
// When using grouped query attention, all rows use the same mask.
if (p.gqa_ratio > 1) {
tensorLayoutM = setTensorLayoutStrideNV(tensorLayoutM, 0, 1);
}
coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> mv;
coopMatLoadTensorNV(mv, data_m, 0, sliceTensorLayoutNV(tensorLayoutM, i * Br, Br, j * Bc, Bc));
S += slope*coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(mv);
S += slopeMat*coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(mv);
}
// Clear padding elements to -inf, so they don't contribute to rowmax
@@ -285,6 +332,20 @@ void main() {
O = coopMatMulAdd(P_A, V, O);
}
// If there is split_k, then the split_k resolve shader does the final
// division by L. Store the intermediate O value and per-row m and L values.
if (p.k_num > 1) {
coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(O);
uint32_t o_offset = D * p.ne1 * split_k_index;
coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N);
o_offset = D * p.ne1 * p.k_num + p.ne1 * split_k_index * 2;
coopMatPerElementNV(L, L, perElemOpStoreCol0, o_offset, iq2, N);
coopMatPerElementNV(M, M, perElemOpStoreCol0, o_offset + p.ne1, iq2, N);
return;
}
coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> Ldiag;
// resize L by using smear/reduce
@@ -297,13 +358,18 @@ void main() {
O = Ldiag*O;
tensorLayoutNV<3, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutD = createTensorLayoutNV(3, gl_CooperativeMatrixClampModeConstantNV);
tensorLayoutD = setTensorLayoutDimensionNV(tensorLayoutD, p.ne2, p.ne1, D);
// permute dimensions
tensorViewNV<3, false, 1, 0, 2> tensorViewPermute = createTensorViewNV(3, false, 1, 0, 2);
uint32_t o_offset = iq3*p.ne2*p.ne1;
coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(O);
coopMatStoreTensorNV(O_D, data_o, o_offset, sliceTensorLayoutNV(tensorLayoutD, i * Br, Br, iq2, 1, 0, D), tensorViewPermute);
if (p.gqa_ratio > 1) {
coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N);
} else {
tensorLayoutNV<3, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutD = createTensorLayoutNV(3, gl_CooperativeMatrixClampModeConstantNV);
tensorLayoutD = setTensorLayoutDimensionNV(tensorLayoutD, p.ne2, p.ne1, D);
// permute dimensions
tensorViewNV<3, false, 1, 0, 2> tensorViewPermute = createTensorViewNV(3, false, 1, 0, 2);
coopMatStoreTensorNV(O_D, data_o, o_offset, sliceTensorLayoutNV(tensorLayoutD, i * Br, Br, iq2, N, 0, D), tensorViewPermute);
}
}
@@ -0,0 +1,59 @@
#version 450
#extension GL_EXT_control_flow_attributes : enable
#define BLOCK_SIZE 32
layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {float data_a[];};
layout (binding = 1) writeonly buffer D {float data_d[];};
layout (push_constant) uniform parameter {
uint D;
uint N;
uint k_num;
} p;
void main() {
// Each workgroup handles a row
const uint n = gl_WorkGroupID.x;
const uint tid = gl_LocalInvocationID.x;
uint D = p.D;
uint N = p.N;
uint k_num = p.k_num;
uint l_offset = D * N * k_num + n;
uint m_offset = D * N * k_num + N + n;
uint lm_stride = N * 2;
// Compute the max m value for the row
float m_max = -1.0/0.0;
[[unroll]] for (uint k = 0; k < k_num; ++k) {
float m = data_a[m_offset + k * lm_stride];
m_max = max(m_max, m);
}
// Compute L based on m_max
float L = 0;
[[unroll]] for (uint k = 0; k < k_num; ++k) {
float l = data_a[l_offset + k * lm_stride];
float m = data_a[m_offset + k * lm_stride];
L += exp(m - m_max) * l;
}
L = 1.0 / L;
// Scale and sum the O contributions based on m_max and store the result to memory
for (uint d = tid; d < D; d += BLOCK_SIZE) {
float O = 0.0;
[[unroll]] for (uint k = 0; k < k_num; ++k) {
uint o_offset = D * N * k + D * n + d;
float m = data_a[m_offset + k * lm_stride];
O += exp(m - m_max) * data_a[o_offset];
}
O *= L;
data_d[D * n + d] = O;
}
}
@@ -212,7 +212,7 @@ void main() {
#else
ACC_TYPE sums[WMITER * TM * WNITER * TN];
FLOAT_TYPE cache_a[WMITER * TM];
FLOAT_TYPE cache_b[WNITER * TN];
FLOAT_TYPE cache_b[TN];
[[unroll]] for (uint i = 0; i < WMITER*TM*WNITER*TN; i++) {
sums[i] = ACC_TYPE(0.0f);
@@ -744,16 +744,14 @@ void main() {
}
[[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {
[[unroll]] for (uint j = 0; j < TN; j++) {
cache_b[wsic * TN + j] = buf_b[(warp_c * WN + wsic * WSUBN + tiwc * TN + j) * SHMEM_STRIDE + i];
cache_b[j] = buf_b[(warp_c * WN + wsic * WSUBN + tiwc * TN + j) * SHMEM_STRIDE + i];
}
}
[[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {
[[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
[[unroll]] for (uint cc = 0; cc < TN; cc++) {
[[unroll]] for (uint cr = 0; cr < TM; cr++) {
const uint sums_idx = (wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr;
sums[sums_idx] = fma(ACC_TYPE(cache_a[wsir * TM + cr]), ACC_TYPE(cache_b[wsic * TN + cc]), sums[sums_idx]);
sums[sums_idx] = fma(ACC_TYPE(cache_a[wsir * TM + cr]), ACC_TYPE(cache_b[cc]), sums[sums_idx]);
}
}
}
@@ -0,0 +1,442 @@
#version 450
#extension GL_EXT_control_flow_attributes : enable
#extension GL_EXT_shader_16bit_storage : require
#extension GL_EXT_shader_explicit_arithmetic_types_int8 : require
#extension GL_EXT_integer_dot_product : require
#ifdef FLOAT16
#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
#endif
#ifdef COOPMAT
#extension GL_KHR_cooperative_matrix : enable
#extension GL_KHR_memory_scope_semantics : enable
#extension GL_KHR_shader_subgroup_basic : enable
#endif
#ifdef MUL_MAT_ID
#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
#endif
#include "types.comp"
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {A_TYPE_PACKED16 data_a[];};
#if defined(A_TYPE_PACKED32)
layout (binding = 0) readonly buffer A_PACKED32 {A_TYPE_PACKED32 data_a_packed32[];};
#endif
layout (binding = 1) readonly buffer B {block_q8_1_packed32 data_b[];};
layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
#ifdef MUL_MAT_ID
layout (binding = 3) readonly buffer IDS {int data_ids[];};
#endif
layout (push_constant) uniform parameter
{
uint M;
uint N;
uint K;
uint stride_a;
uint stride_b;
uint stride_d;
uint batch_stride_a;
uint batch_stride_b;
uint batch_stride_d;
#ifdef MUL_MAT_ID
uint nei0;
uint nei1;
uint nbi1;
uint ne11;
#else
uint k_split;
uint ne02;
uint ne12;
uint broadcast2;
uint broadcast3;
#endif
} p;
layout (constant_id = 0) const uint BLOCK_SIZE = 64;
layout (constant_id = 1) const uint BM = 64;
layout (constant_id = 2) const uint BN = 64;
// layout (constant_id = 3) const uint BK = 32;
layout (constant_id = 4) const uint WM = 32;
layout (constant_id = 5) const uint WN = 32;
layout (constant_id = 6) const uint WMITER = 2;
layout (constant_id = 7) const uint TM = 4;
layout (constant_id = 8) const uint TN = 2;
layout (constant_id = 9) const uint TK = 1; // Only needed for coopmat
layout (constant_id = 10) const uint WARP = 32;
#define BK 32
#ifdef COOPMAT
#define SHMEM_STRIDE (BK / 4 + 4)
#else
#define SHMEM_STRIDE (BK / 4 + 1)
#endif
shared int32_t buf_a_qs[BM * SHMEM_STRIDE];
#ifndef COOPMAT
#if QUANT_AUXF == 1
shared FLOAT_TYPE buf_a_dm[BM];
#else
shared FLOAT_TYPE_VEC2 buf_a_dm[BM];
#endif
#endif
shared int32_t buf_b_qs[BN * SHMEM_STRIDE];
#ifndef COOPMAT
shared FLOAT_TYPE_VEC2 buf_b_ds[BN];
#endif
#define LOAD_VEC_A (4 * QUANT_R)
#define LOAD_VEC_B 4
#ifdef MUL_MAT_ID
shared u16vec2 row_ids[3072];
#endif // MUL_MAT_ID
#define NUM_WARPS (BLOCK_SIZE / WARP)
#ifdef COOPMAT
shared ACC_TYPE coopmat_stage[TM * TN * NUM_WARPS];
#endif
#include "mul_mmq_funcs.comp"
void main() {
#ifdef NEEDS_INIT_IQ_SHMEM
init_iq_shmem(gl_WorkGroupSize);
#endif
#ifdef MUL_MAT_ID
const uint expert_idx = gl_GlobalInvocationID.z;
#else
const uint batch_idx = gl_GlobalInvocationID.z;
const uint i13 = batch_idx / p.ne12;
const uint i12 = batch_idx % p.ne12;
const uint i03 = i13 / p.broadcast3;
const uint i02 = i12 / p.broadcast2;
const uint batch_idx_a = i03 * p.ne02 + i02;
#endif
const uint blocks_m = (p.M + BM - 1) / BM;
const uint ir = gl_WorkGroupID.x % blocks_m;
const uint ik = gl_WorkGroupID.x / blocks_m;
const uint ic = gl_WorkGroupID.y;
const uint WNITER = (WM * WN) / (WARP * TM * TN * WMITER);
const uint WSUBM = WM / WMITER;
const uint WSUBN = WN / WNITER;
#ifdef COOPMAT
const uint warp_i = gl_SubgroupID;
const uint tiw = gl_SubgroupInvocationID;
const uint cms_per_row = WM / TM;
const uint cms_per_col = WN / TN;
const uint storestride = WARP / TM;
const uint store_r = tiw % TM;
const uint store_c = tiw / TM;
#else
const uint warp_i = gl_LocalInvocationID.x / WARP;
const uint tiw = gl_LocalInvocationID.x % WARP;
const uint tiwr = tiw % (WSUBM / TM);
const uint tiwc = tiw / (WSUBM / TM);
#endif
const uint warp_r = warp_i % (BM / WM);
const uint warp_c = warp_i / (BM / WM);
const uint loadr_a = gl_LocalInvocationID.x % (BK / LOAD_VEC_A);
const uint loadc_a = gl_LocalInvocationID.x / (BK / LOAD_VEC_A);
const uint loadr_b = gl_LocalInvocationID.x % (BK / LOAD_VEC_B);
const uint loadc_b = gl_LocalInvocationID.x / (BK / LOAD_VEC_B);
const uint loadstride_a = BLOCK_SIZE * LOAD_VEC_A / BK;
const uint loadstride_b = BLOCK_SIZE * LOAD_VEC_B / BK;
#ifdef MUL_MAT_ID
uint _ne1 = 0;
for (uint ii1 = 0; ii1 < p.nei1; ii1++) {
for (uint ii0 = 0; ii0 < p.nei0; ii0++) {
if (data_ids[ii1*p.nbi1 + ii0] == expert_idx) {
row_ids[_ne1] = u16vec2(ii0, ii1);
_ne1++;
}
}
}
barrier();
// Workgroup has no work
if (ic * BN >= _ne1) return;
#endif
#ifdef MUL_MAT_ID
const uint start_k = 0;
const uint end_k = p.K;
#else
const uint start_k = ik * p.k_split;
const uint end_k = min(p.K, (ik + 1) * p.k_split);
#endif
uint pos_a_ib = (
#ifdef MUL_MAT_ID
expert_idx * p.batch_stride_a +
#else
batch_idx_a * p.batch_stride_a +
#endif
ir * BM * p.stride_a + start_k) / BK;
#ifdef MUL_MAT_ID
uint pos_b_ib = 0;
#else
uint pos_b_ib = (batch_idx * p.batch_stride_b + ic * BN * p.stride_b + start_k) / BK;
#endif
#ifdef COOPMAT
coopmat<int8_t, gl_ScopeSubgroup, TM, TK, gl_MatrixUseA> cache_a;
coopmat<int8_t, gl_ScopeSubgroup, TK, TN, gl_MatrixUseB> cache_b;
coopmat<int32_t, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator> cm_result;
coopmat<ACC_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator> factors[cms_per_row * cms_per_col];
coopmat<ACC_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator> sums[cms_per_row * cms_per_col];
[[unroll]] for (uint i = 0; i < cms_per_row * cms_per_col; i++) {
sums[i] = coopmat<ACC_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator>(0.0f);
}
#else
int32_t cache_a_qs[WMITER * TM * BK / 4];
int32_t cache_b_qs[TN * BK / 4];
ACC_TYPE sums[WMITER * TM * WNITER * TN];
[[unroll]] for (uint i = 0; i < WMITER*TM*WNITER*TN; i++) {
sums[i] = ACC_TYPE(0.0f);
}
#endif
#if QUANT_AUXF == 1
FLOAT_TYPE cache_a_dm[WMITER * TM];
#else
FLOAT_TYPE_VEC2 cache_a_dm[WMITER * TM];
#endif
FLOAT_TYPE_VEC2 cache_b_ds[TN];
for (uint block = start_k; block < end_k; block += BK) {
[[unroll]] for (uint l = 0; loadc_a + l < BM; l += loadstride_a) {
const uint ib = pos_a_ib + (loadc_a + l) * p.stride_a / BK;
const uint iqs = loadr_a;
const uint buf_ib = loadc_a + l;
if (iqs == 0) {
#if QUANT_AUXF == 1
buf_a_dm[buf_ib] = get_d(ib);
#else
buf_a_dm[buf_ib] = get_dm(ib);
#endif
}
#if QUANT_R == 1
buf_a_qs[buf_ib * SHMEM_STRIDE + iqs] = repack(ib, iqs);
#else
const i32vec2 vals = repack(ib, iqs);
buf_a_qs[buf_ib * SHMEM_STRIDE + iqs ] = vals.x;
buf_a_qs[buf_ib * SHMEM_STRIDE + iqs + 4] = vals.y;
#endif
}
[[unroll]] for (uint l = 0; loadc_b + l < BN; l += loadstride_b) {
#ifdef MUL_MAT_ID
const u16vec2 row_idx = row_ids[ic * BN + loadc_b + l];
const uint idx = pos_b_ib + row_idx.y * p.batch_stride_b / LOAD_VEC_B + (row_idx.x % p.ne11) * p.stride_b / LOAD_VEC_B + loadr_b;
const uint ib = idx / 8;
const uint iqs = idx & 0x7;
#else
const uint ib = pos_b_ib + (loadc_b + l) * p.stride_b / BK;
const uint iqs = loadr_b;
#endif
const uint buf_ib = loadc_b + l;
if (iqs == 0) {
buf_b_ds[buf_ib] = FLOAT_TYPE_VEC2(data_b[ib].ds);
}
buf_b_qs[buf_ib * SHMEM_STRIDE + iqs] = data_b[ib].qs[iqs];
}
barrier();
pos_a_ib += 1;
pos_b_ib += 1;
#ifdef COOPMAT
[[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) {
const uint ib_a = warp_r * WM + cm_row * TM;
// Load from shared into cache
coopMatLoad(cache_a, buf_a_qs, ib_a * SHMEM_STRIDE, SHMEM_STRIDE, gl_CooperativeMatrixLayoutRowMajor);
// TODO: only cache values that are actually needed
[[unroll]] for (uint t_idx = 0; t_idx < TM; t_idx++) {
cache_a_dm[t_idx] = buf_a_dm[ib_a + t_idx];
}
[[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) {
const uint ib_b = warp_c * WN + cm_col * TN;
coopMatLoad(cache_b, buf_b_qs, ib_b * SHMEM_STRIDE, SHMEM_STRIDE, gl_CooperativeMatrixLayoutColumnMajor);
// TODO: only cache values that are actually needed
[[unroll]] for (uint t_idx = 0; t_idx < TN; t_idx++) {
cache_b_dm[t_idx] = buf_b_d[ib_b + t_idx];
}
cm_result = coopmat<int32_t, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator>(0);
cm_result = coopMatMulAdd(cache_a, cache_b, cm_result);
[[unroll]] for (uint col = 0; col < TN; col += storestride) {
coopmat_stage[warp_i * TM * TN + (store_c + col) * TM + store_r] = ACC_TYPE(float(cache_a_d[store_r]) * float(cache_b_d[store_c + col]));
}
coopMatLoad(factors, coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor);
sums[cm_col * cms_per_row + cm_row] += factors * coopmat<ACC_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator>(cm_result);
}
}
#else
// Load from shared into cache
[[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
[[unroll]] for (uint cr = 0; cr < TM; cr++) {
const uint ib = warp_r * WM + wsir * WSUBM + tiwr * TM + cr;
cache_a_dm[wsir * TM + cr] = buf_a_dm[ib];
[[unroll]] for (uint idx_k = 0; idx_k < BK / 4; idx_k++) {
cache_a_qs[(wsir * TM + cr) * (BK / 4) + idx_k] = buf_a_qs[ib * SHMEM_STRIDE + idx_k];
}
}
}
[[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {
[[unroll]] for (uint cc = 0; cc < TN; cc++) {
const uint ib = warp_c * WN + wsic * WSUBN + tiwc * TN + cc;
cache_b_ds[cc] = buf_b_ds[ib];
[[unroll]] for (uint idx_k = 0; idx_k < BK / 4; idx_k++) {
cache_b_qs[cc * (BK / 4) + idx_k] = buf_b_qs[ib * SHMEM_STRIDE + idx_k];
}
}
[[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
[[unroll]] for (uint cc = 0; cc < TN; cc++) {
[[unroll]] for (uint cr = 0; cr < TM; cr++) {
const uint cache_a_idx = wsir * TM + cr;
const uint sums_idx = (wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr;
int32_t q_sum = 0;
[[unroll]] for (uint idx_k = 0; idx_k < BK / 4; idx_k++) {
q_sum += dotPacked4x8EXT(cache_a_qs[cache_a_idx * (BK / 4) + idx_k],
cache_b_qs[cc * (BK / 4) + idx_k]);
}
sums[sums_idx] += mul_q8_1(q_sum, cache_a_dm[cache_a_idx], cache_b_ds[cc]);
}
}
}
}
#endif
barrier();
}
const uint dr = ir * BM + warp_r * WM;
const uint dc = ic * BN + warp_c * WN;
#ifndef MUL_MAT_ID
const uint offsets = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * gl_NumWorkGroups.z;
#endif
#ifdef COOPMAT
#ifdef MUL_MAT_ID
[[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) {
[[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) {
coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor);
[[unroll]] for (uint col = 0; col < BN; col += storestride) {
const uint row_i = dc + cm_col * TN + col + store_c;
if (row_i >= _ne1) break;
const u16vec2 row_idx = row_ids[row_i];
data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]);
}
}
}
#else
const bool is_aligned = p.stride_d % 4 == 0; // Assumption: D_TYPE == float
[[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) {
[[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) {
const bool is_in_bounds = dr + (cm_row + 1) * TM <= p.M && dc + (cm_col + 1) * TN <= p.N;
if (is_aligned && is_in_bounds) {
// Full coopMat is within bounds and stride_d is aligned with 16B
coopmat<D_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator> cm_dtype = coopmat<D_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator>(sums[cm_col * cms_per_row + cm_row]);
coopMatStore(cm_dtype, data_d, offsets + (dc + cm_col * TN) * p.stride_d + dr + cm_row * TM, p.stride_d, gl_CooperativeMatrixLayoutColumnMajor);
} else if (is_in_bounds) {
// Full coopMat is within bounds, but stride_d is not aligned
coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor);
[[unroll]] for (uint col = 0; col < TN; col += storestride) {
data_d[offsets + (dc + cm_col * TN + col + store_c) * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]);
}
} else if (dr + cm_row * TM < p.M && dc + cm_col * TN < p.N) {
// Partial coopMat is within bounds
coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor);
[[unroll]] for (uint col = 0; col < TN; col += storestride) {
if (dr + cm_row * TM + store_r < p.M && dc + cm_col * TN + col + store_c < p.N) {
data_d[offsets + (dc + cm_col * TN + col + store_c) * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]);
}
}
}
}
}
#endif // MUL_MAT_ID
#else
[[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {
[[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
const uint dr_warp = dr + wsir * WSUBM + tiwr * TM;
const uint dc_warp = dc + wsic * WSUBN + tiwc * TN;
[[unroll]] for (uint cc = 0; cc < TN; cc++) {
#ifdef MUL_MAT_ID
const uint row_i = dc_warp + cc;
if (row_i >= _ne1) break;
const u16vec2 row_idx = row_ids[row_i];
#endif // MUL_MAT_ID
[[unroll]] for (uint cr = 0; cr < TM; cr++) {
#ifdef MUL_MAT_ID
data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]);
#else
if (dr_warp + cr < p.M && dc_warp + cc < p.N) {
data_d[offsets + (dc_warp + cc) * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]);
}
#endif // MUL_MAT_ID
}
}
}
}
#endif // COOPMAT
}
@@ -0,0 +1,99 @@
#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
#extension GL_EXT_shader_explicit_arithmetic_types_int8 : require
#include "types.comp"
// Each iqs value maps to a 32-bit integer
#if defined(DATA_A_Q4_0)
i32vec2 repack(uint ib, uint iqs) {
// Use 2-byte loads since a q4_0 block (18 bytes) is not divisible by 4
const u16vec2 quants = u16vec2(data_a[ib].qs[iqs * 2 ],
data_a[ib].qs[iqs * 2 + 1]);
const uint32_t vui = pack32(quants);
return i32vec2( vui & 0x0F0F0F0F,
(vui >> 4) & 0x0F0F0F0F);
}
ACC_TYPE mul_q8_1(int32_t q_sum, float da, vec2 dsb) {
return ACC_TYPE(da * (float(q_sum) * dsb.x - 8.0f * dsb.y));
}
#endif
#if defined(DATA_A_Q4_1)
i32vec2 repack(uint ib, uint iqs) {
// Use 4-byte loads since a q4_1 block (20 bytes) is divisible by 4
const uint32_t vui = data_a_packed32[ib].qs[iqs];
return i32vec2( vui & 0x0F0F0F0F,
(vui >> 4) & 0x0F0F0F0F);
}
ACC_TYPE mul_q8_1(int32_t q_sum, vec2 dma, vec2 dsb) {
return ACC_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y);
}
#endif
#if defined(DATA_A_Q5_0)
i32vec2 repack(uint ib, uint iqs) {
// Use 2-byte loads since a q5_0 block (22 bytes) is not divisible by 4
const u16vec2 quants = u16vec2(data_a[ib].qs[iqs * 2 ],
data_a[ib].qs[iqs * 2 + 1]);
const uint32_t vui = pack32(quants);
const int32_t qh = int32_t((uint32_t(data_a[ib].qh[1]) << 16 | data_a[ib].qh[0]) >> (4 * iqs));
const int32_t v0 = int32_t(vui & 0x0F0F0F0F)
| ((qh & 0xF) * 0x02040810) & 0x10101010; // (0,1,2,3) -> (4,12,20,28)
const int32_t v1 = int32_t((vui >> 4) & 0x0F0F0F0F)
| (((qh >> 16) & 0xF) * 0x02040810) & 0x10101010; // (16,17,18,19) -> (4,12,20,28)
return i32vec2(v0, v1);
}
ACC_TYPE mul_q8_1(int32_t q_sum, float da, vec2 dsb) {
return ACC_TYPE(da * (float(q_sum) * dsb.x - 16.0f * dsb.y));
}
#endif
#if defined(DATA_A_Q5_1)
i32vec2 repack(uint ib, uint iqs) {
// Use 4-byte loads since a q5_1 block (24 bytes) is divisible by 4
const uint32_t vui = data_a_packed32[ib].qs[iqs];
const int32_t qh = int32_t(data_a_packed32[ib].qh >> (4 * iqs));
const int32_t v0 = int32_t(vui & 0x0F0F0F0F)
| ((qh & 0xF) * 0x02040810) & 0x10101010; // (0,1,2,3) -> (4,12,20,28)
const int32_t v1 = int32_t((vui >> 4) & 0x0F0F0F0F)
| (((qh >> 16) & 0xF) * 0x02040810) & 0x10101010; // (16,17,18,19) -> (4,12,20,28)
return i32vec2(v0, v1);
}
ACC_TYPE mul_q8_1(int32_t q_sum, vec2 dma, vec2 dsb) {
return ACC_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y);
}
#endif
#if defined(DATA_A_Q8_0)
int32_t repack(uint ib, uint iqs) {
// Use 2-byte loads since a q8_0 block (34 bytes) is not divisible by 4
return pack32(i16vec2(data_a[ib].qs[iqs * 2 ],
data_a[ib].qs[iqs * 2 + 1]));
}
ACC_TYPE mul_q8_1(int32_t q_sum, float da, vec2 dsb) {
return ACC_TYPE(float(q_sum) * da * dsb.x);
}
#endif
#if defined(DATA_A_Q4_0) || defined(DATA_A_Q5_0) || defined(DATA_A_Q8_0) || defined(DATA_A_IQ1_S) || defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL)
FLOAT_TYPE get_d(uint ib) {
return FLOAT_TYPE(data_a[ib].d);
}
#endif
#if defined(DATA_A_Q4_1) || defined(DATA_A_Q5_1)
FLOAT_TYPE_VEC2 get_dm(uint ib) {
return FLOAT_TYPE_VEC2(data_a_packed32[ib].dm);
}
#endif
@@ -0,0 +1,77 @@
#version 450
#extension GL_EXT_control_flow_attributes : require
#extension GL_EXT_shader_16bit_storage : require
layout (push_constant) uniform parameter
{
uint ne;
} p;
#include "types.comp"
layout(constant_id = 0) const uint GROUP_SIZE = 32;
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {vec4 data_a[];};
layout (binding = 1) writeonly buffer D {block_q8_1_packed32 data_b[];};
shared float shmem[GROUP_SIZE];
void quantize() {
const uint wgid = gl_WorkGroupID.x;
const uint tid = gl_LocalInvocationID.x;
// Each thread handles a vec4, so 8 threads handle a block
const uint blocks_per_group = GROUP_SIZE / 8;
const uint block_in_wg = tid / 8;
const uint ib = wgid * blocks_per_group + block_in_wg;
const uint iqs = tid % 8;
if (ib >= gl_NumWorkGroups.x * blocks_per_group) {
return;
}
const uint a_idx = ib * 8 + iqs;
vec4 vals = a_idx < p.ne ? data_a[a_idx] : vec4(0.0f);
const vec4 abs_vals = abs(vals);
// Find absolute max for each block
shmem[tid] = max(max(abs_vals.x, abs_vals.y), max(abs_vals.z, abs_vals.w));
barrier();
[[unroll]] for (uint s = 4; s > 0; s >>= 1) {
if (iqs < s) {
shmem[tid] = max(shmem[tid], shmem[tid + s]);
}
barrier();
}
const float amax = shmem[block_in_wg * 8];
const float d = amax / 127.0;
const float d_inv = d != 0.0 ? 1.0 / d : 0.0;
vals = round(vals * d_inv);
data_b[ib].qs[iqs] = pack32(i8vec4(round(vals)));
barrier();
// Calculate the sum for each block
shmem[tid] = vals.x + vals.y + vals.z + vals.w;
barrier();
[[unroll]] for (uint s = 4; s > 0; s >>= 1) {
if (iqs < s) {
shmem[tid] += shmem[tid + s];
}
barrier();
}
if (iqs == 0) {
const float sum = shmem[tid];
data_b[ib].ds = f16vec2(vec2(d, sum * d));
}
}
void main() {
quantize();
}
@@ -0,0 +1,7 @@
#version 460
#extension GL_EXT_integer_dot_product : require
void main()
{
}
+45 -1
View File
@@ -1,4 +1,3 @@
#if !defined(GGML_TYPES_COMP)
#define GGML_TYPES_COMP
@@ -51,6 +50,7 @@ struct block_q4_0_packed16
#if defined(DATA_A_Q4_0)
#define QUANT_K QUANT_K_Q4_0
#define QUANT_R QUANT_R_Q4_0
#define QUANT_AUXF 1
#define A_TYPE block_q4_0
#define A_TYPE_PACKED16 block_q4_0_packed16
#endif
@@ -72,11 +72,19 @@ struct block_q4_1_packed16
uint16_t qs[16/2];
};
struct block_q4_1_packed32
{
f16vec2 dm;
uint32_t qs[16/4];
};
#if defined(DATA_A_Q4_1)
#define QUANT_K QUANT_K_Q4_1
#define QUANT_R QUANT_R_Q4_1
#define QUANT_AUXF 2
#define A_TYPE block_q4_1
#define A_TYPE_PACKED16 block_q4_1_packed16
#define A_TYPE_PACKED32 block_q4_1_packed32
#endif
#define QUANT_K_Q5_0 32
@@ -99,6 +107,7 @@ struct block_q5_0_packed16
#if defined(DATA_A_Q5_0)
#define QUANT_K QUANT_K_Q5_0
#define QUANT_R QUANT_R_Q5_0
#define QUANT_AUXF 1
#define A_TYPE block_q5_0
#define A_TYPE_PACKED16 block_q5_0_packed16
#endif
@@ -122,11 +131,20 @@ struct block_q5_1_packed16
uint16_t qs[16/2];
};
struct block_q5_1_packed32
{
f16vec2 dm;
uint qh;
uint32_t qs[16/4];
};
#if defined(DATA_A_Q5_1)
#define QUANT_K QUANT_K_Q5_1
#define QUANT_R QUANT_R_Q5_1
#define QUANT_AUXF 2
#define A_TYPE block_q5_1
#define A_TYPE_PACKED16 block_q5_1_packed16
#define A_TYPE_PACKED32 block_q5_1_packed32
#endif
#define QUANT_K_Q8_0 32
@@ -142,14 +160,40 @@ struct block_q8_0_packed16
float16_t d;
int16_t qs[32/2];
};
struct block_q8_0_packed32
{
float16_t d;
int32_t qs[32/4];
};
#if defined(DATA_A_Q8_0)
#define QUANT_K QUANT_K_Q8_0
#define QUANT_R QUANT_R_Q8_0
#define QUANT_AUXF 1
#define A_TYPE block_q8_0
#define A_TYPE_PACKED16 block_q8_0_packed16
#define A_TYPE_PACKED32 block_q8_0_packed32
#endif
#define QUANT_K_Q8_1 32
#define QUANT_R_Q8_1 1
struct block_q8_1
{
f16vec2 ds;
int8_t qs[32];
};
struct block_q8_1_packed16
{
f16vec2 ds;
int16_t qs[16];
};
struct block_q8_1_packed32
{
f16vec2 ds;
int32_t qs[8];
};
// K-quants
#define QUANT_K_Q2_K 256
@@ -295,7 +295,10 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool
std::string aligned_b_type_f32 = coopmat2 ? "float" : fp16 ? "mat2x4" : "vec4";
std::string aligned_b_type_f16 = coopmat2 ? "float16_t" : fp16 ? "f16mat2x4" : "f16vec4";
std::map<std::string, std::string> base_dict = {{"FLOAT_TYPE", (coopmat2 || fp16) ? "float16_t" : "float"}};
std::map<std::string, std::string> base_dict = {
{"FLOAT_TYPE", (coopmat2 || fp16) ? "float16_t" : "float"},
{"FLOAT_TYPE_VEC2", (coopmat2 || fp16) ? "f16vec2" : "vec2"},
};
std::string shader_name = "matmul";
if (matmul_id) {
@@ -313,9 +316,7 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool
base_dict["COOPMAT"] = "1";
}
base_dict["ACC_TYPE"] = f16acc ? "float16_t" : "float";
std::string source_name = coopmat2 ? "mul_mm_cm2.comp" : "mul_mm.comp";
const std::string source_name = coopmat2 ? "mul_mm_cm2.comp" : "mul_mm.comp";
// Shaders with f16 B_TYPE
string_to_spv(shader_name + "_f32_f16", source_name, merge_maps(base_dict, {{"DATA_A_F32", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}, }), fp16, coopmat, coopmat2, f16acc);
@@ -339,14 +340,20 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool
// don't generate f32 variants for coopmat2
if (!coopmat2) {
string_to_spv(shader_name + "_" + tname + "_f32", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}}), fp16, coopmat, coopmat2, f16acc);
string_to_spv(shader_name + "_" + tname + "_f32_aligned", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
string_to_spv(shader_name + "_" + tname + "_f32", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
string_to_spv(shader_name + "_" + tname + "_f32_aligned", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
}
if (tname != "f16" && tname != "f32") {
string_to_spv(shader_name + "_" + tname + "_f16", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}}), fp16, coopmat, coopmat2, f16acc);
string_to_spv(shader_name + "_" + tname + "_f16_aligned", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
string_to_spv(shader_name + "_" + tname + "_f16", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
string_to_spv(shader_name + "_" + tname + "_f16_aligned", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
}
#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
if (!coopmat && !coopmat2 && !matmul_id && (tname == "q4_0" || tname == "q4_1" || tname == "q5_0" || tname == "q5_1" || tname == "q8_0")) {
string_to_spv(shader_name + "_" + tname + "_q8_1", "mul_mmq.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float"},}), fp16, coopmat, coopmat2, f16acc);
}
#endif
}
}
@@ -458,6 +465,8 @@ void process_shaders() {
string_to_spv("acc_f32", "acc.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
string_to_spv("split_k_reduce", "mul_mat_split_k_reduce.comp", {});
string_to_spv("fa_split_k_reduce", "flash_attn_split_k_reduce.comp", {});
string_to_spv("quantize_q8_1", "quantize_q8_1.comp", {});
string_to_spv("mul_f32", "mul.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
+6
View File
@@ -1159,6 +1159,12 @@ int64_t ggml_nrows(const struct ggml_tensor * tensor) {
}
size_t ggml_nbytes(const struct ggml_tensor * tensor) {
for (int i = 0; i < GGML_MAX_DIMS; ++i) {
if (tensor->ne[i] <= 0) {
return 0;
}
}
size_t nbytes;
const size_t blck_size = ggml_blck_size(tensor->type);
if (blck_size == 1) {
+1
View File
@@ -932,6 +932,7 @@ static void gguf_check_reserved_keys(const std::string & key, const T val) {
if constexpr (std::is_same<T, uint32_t>::value) {
GGML_ASSERT(val > 0 && (val & (val - 1)) == 0 && GGUF_KEY_GENERAL_ALIGNMENT " must be power of 2");
} else {
GGML_UNUSED(val);
GGML_ABORT(GGUF_KEY_GENERAL_ALIGNMENT " must be type u32");
}
}
+8
View File
@@ -280,10 +280,18 @@ extern "C" {
};
};
struct llama_model_tensor_buft_override {
const char * pattern;
ggml_backend_buffer_type_t buft;
};
struct llama_model_params {
// NULL-terminated list of devices to use for offloading (if NULL, all available devices are used)
ggml_backend_dev_t * devices;
// NULL-terminated list of buffer types to use for tensors that match a pattern
const struct llama_model_tensor_buft_override * tensor_buft_overrides;
int32_t n_gpu_layers; // number of layers to store in VRAM
enum llama_split_mode split_mode; // how to split the model across multiple GPUs
+1 -1
View File
@@ -1 +1 @@
d53795ee70aa545464569d71caa46f38c05c1982
f06264eda2e2bf6e814db5a32bbf42e0b2b1ed98
+1
View File
@@ -75,6 +75,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_GENERAL_ARCHITECTURE, "general.architecture" },
{ LLM_KV_GENERAL_QUANTIZATION_VERSION, "general.quantization_version" },
{ LLM_KV_GENERAL_ALIGNMENT, "general.alignment" },
{ LLM_KV_GENERAL_FILE_TYPE, "general.file_type" },
{ LLM_KV_GENERAL_NAME, "general.name" },
{ LLM_KV_GENERAL_AUTHOR, "general.author" },
{ LLM_KV_GENERAL_VERSION, "general.version" },
+1
View File
@@ -79,6 +79,7 @@ enum llm_kv {
LLM_KV_GENERAL_ARCHITECTURE,
LLM_KV_GENERAL_QUANTIZATION_VERSION,
LLM_KV_GENERAL_ALIGNMENT,
LLM_KV_GENERAL_FILE_TYPE,
LLM_KV_GENERAL_NAME,
LLM_KV_GENERAL_AUTHOR,
LLM_KV_GENERAL_VERSION,
+10 -52
View File
@@ -255,7 +255,8 @@ llama_context::llama_context(
model.n_devices() > 1 &&
model.params.n_gpu_layers > (int) model.hparams.n_layer &&
model.params.split_mode == LLAMA_SPLIT_MODE_LAYER &&
cparams.offload_kqv;
cparams.offload_kqv &&
!model.has_tensor_overrides();
// pipeline parallelism requires support for async compute and events in all devices
if (pipeline_parallel) {
@@ -1201,33 +1202,7 @@ int llama_context::decode(llama_batch & inp_batch) {
const int64_t n_tokens_all = batch.n_tokens;
const int64_t n_embd = hparams.n_embd;
// TODO: remove this stuff
class batch_guard {
public:
batch_guard(llama_kv_cache_unified & kv_self) : kv_slot_restorer(kv_self) {
}
~batch_guard() {
if (!is_done) {
kv_slot_restorer.restore();
}
}
void done() {
is_done = true;
}
void save(const llama_kv_cache_slot_info & slot_info) {
kv_slot_restorer.save(slot_info);
}
private:
bool is_done = false;
llama_kv_slot_restorer kv_slot_restorer;
};
batch_guard bg(*kv_self);
llama_kv_cache_guard kv_guard(kv_self.get());
GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT
@@ -1280,6 +1255,9 @@ int llama_context::decode(llama_batch & inp_batch) {
return -2;
};
// handle any pending defrags/shifts
kv_self_update();
int64_t n_outputs_prev = 0;
while (sbatch.n_tokens > 0) {
@@ -1319,22 +1297,12 @@ int llama_context::decode(llama_batch & inp_batch) {
// find KV slot
{
kv_self_update();
if (!kv_self->find_slot(ubatch)) {
LLAMA_LOG_WARN("%s: failed to find KV cache slot for ubatch of size %d\n", __func__, ubatch.n_tokens);
// if we have enough unused cells before the current head ->
// better to start searching from the beginning of the cache, hoping to fill it
if (kv_self->head > kv_self->used + 2*ubatch.n_tokens) {
kv_self->head = 0;
return 1;
}
const auto slot_info = kv_self->find_slot(ubatch);
if (!slot_info) {
LLAMA_LOG_ERROR("%s: failed to prepare ubatch\n", __func__);
return -3;
}
bg.save(slot_info);
if (!kv_self->recurrent) {
// a heuristic, to avoid attending the full cache if it is not yet utilized
// after enough generations, the benefit from this heuristic disappears
@@ -1371,16 +1339,6 @@ int llama_context::decode(llama_batch & inp_batch) {
}
}
// update the kv ring buffer
{
kv_self->head += ubatch.n_tokens;
// Ensure kv cache head points to a valid index.
if (kv_self->head >= kv_self->size) {
kv_self->head = 0;
}
}
// plot the computation graph in dot format (for debugging purposes)
//if (n_past%100 == 0) {
// ggml_graph_dump_dot(gf, NULL, "llama.dot");
@@ -1467,7 +1425,7 @@ int llama_context::decode(llama_batch & inp_batch) {
}
// finalize the batch processing
bg.done();
kv_guard.commit();
// set output mappings
{
+61 -8
View File
@@ -11,8 +11,6 @@
#include <map>
#include <stdexcept>
static const llama_kv_cache_slot_info llama_kv_cache_slot_info_failed{false};
llama_kv_cache_unified::llama_kv_cache_unified(const llama_hparams & hparams, callbacks cbs) : hparams(hparams), cbs(std::move(cbs)) {
}
@@ -206,6 +204,8 @@ bool llama_kv_cache_unified::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos
return false;
}
}
return true;
}
for (uint32_t i = 0; i < size; ++i) {
@@ -446,16 +446,66 @@ void llama_kv_cache_unified::defrag() {
}
}
void llama_kv_cache_unified::restore() {
if (pending.ranges.empty()) {
return;
}
// TODO: tmp - move to llama_kv_cache_recurrent
if (recurrent) {
seq_rm(-1, -1, -1);
return;
}
uint32_t new_head = size;
for (auto & range : pending.ranges) {
for (uint32_t i = range.c0; i < range.c1; ++i) {
cells[i].seq_id.clear();
// keep count of the number of used cells
if (cells[i].pos >= 0) {
used--;
}
cells[i].pos = -1;
cells[i].src = -1;
}
new_head = std::min(new_head, range.c0);
}
if (new_head != size && new_head < head) {
head = new_head;
}
}
void llama_kv_cache_unified::commit() {
if (pending.ranges.empty()) {
LLAMA_LOG_WARN("%s: no pending KV cache updates to commit - might indicate a bug (ref: %s)\n",
__func__, "https://github.com/ggml-org/llama.cpp/pull/12695");
return;
}
pending.ranges.clear();
}
bool llama_kv_cache_unified::get_can_shift() const {
return can_shift;
}
llama_kv_cache_slot_info llama_kv_cache_unified::find_slot(
bool llama_kv_cache_unified::find_slot(
const llama_ubatch & ubatch) {
const uint32_t n_tokens = ubatch.n_tokens;
const uint32_t n_seqs = ubatch.n_seqs;
const uint32_t n_seq_tokens = ubatch.n_seq_tokens;
// if we have enough unused cells before the current head ->
// better to start searching from the beginning of the cache, hoping to fill it
if (head > used + 2*ubatch.n_tokens) {
head = 0;
}
if (recurrent) {
// For recurrent state architectures (like Mamba or RWKV),
// each cache cell can store the state for a whole sequence.
@@ -477,7 +527,7 @@ llama_kv_cache_slot_info llama_kv_cache_unified::find_slot(
// too big seq_id
// TODO: would it be possible to resize the cache instead?
LLAMA_LOG_ERROR("%s: seq_id=%d >= n_seq_max=%d Try using a bigger --parallel value\n", __func__, seq_id, size);
return llama_kv_cache_slot_info_failed;
return false;
}
if (j > 0) {
llama_kv_cell & seq = cells[seq_id];
@@ -616,14 +666,14 @@ llama_kv_cache_slot_info llama_kv_cache_unified::find_slot(
[](const llama_kv_cell& cell){ return !cell.is_empty(); });
// sanity check
return llama_kv_cache_slot_info(n >= n_seqs);
return n >= n_seqs;
}
// otherwise, one cell per token.
if (n_tokens > size) {
LLAMA_LOG_ERROR("%s: n_tokens = %d > size = %d\n", __func__, n_tokens, size);
return llama_kv_cache_slot_info_failed;
return false;
}
uint32_t n_tested = 0;
@@ -651,7 +701,7 @@ llama_kv_cache_slot_info llama_kv_cache_unified::find_slot(
if (n_tested >= size) {
//LLAMA_LOG_ERROR("%s: failed to find a slot for %d tokens\n", __func__, n_tokens);
return llama_kv_cache_slot_info_failed;
return false;
}
}
@@ -668,7 +718,9 @@ llama_kv_cache_slot_info llama_kv_cache_unified::find_slot(
used += n_tokens;
return llama_kv_cache_slot_info(head, head + n_tokens);
pending.ranges.push_back({head, head + n_tokens});
return true;
}
uint32_t llama_kv_cache_unified::get_padding(const llama_cparams & cparams) const {
@@ -1033,6 +1085,7 @@ bool llama_kv_cache_unified::state_read_meta(llama_io_read_i & io, uint32_t cell
LLAMA_LOG_ERROR("%s: failed to find available cells in kv cache\n", __func__);
return false;
}
commit();
// DEBUG CHECK: kv.head should be our first cell, kv.head + cell_count - 1 should be our last cell (verify seq_id and pos values)
// Assume that this is one contiguous block of cells
+36 -68
View File
@@ -17,6 +17,9 @@ struct llama_ubatch;
struct llama_kv_cache : public llama_memory_i {
using llama_memory_i::llama_memory_i;
virtual void restore() = 0; // call if batch processing fails - restores the cache state
virtual void commit() = 0; // call after successful batch processing - clears any pending state
virtual int32_t get_n_tokens() const = 0;
virtual uint32_t get_used_cells() const = 0; // TODO: remove, this is too-specific to the unified cache
@@ -25,9 +28,24 @@ struct llama_kv_cache : public llama_memory_i {
bool get_can_edit() const override { return get_can_shift(); }
};
struct llama_kv_cache_guard {
llama_kv_cache_guard(llama_kv_cache * kv) : kv(kv) {}
~llama_kv_cache_guard() {
kv->restore();
}
void commit() {
kv->commit();
}
private:
llama_kv_cache * kv;
};
struct llama_kv_cell {
llama_pos pos = -1;
llama_pos delta = 0;
llama_pos delta = 0;
int32_t src = -1; // used by recurrent state models to copy states
int32_t tail = -1;
@@ -46,17 +64,6 @@ struct llama_kv_cell {
}
};
// a structure holds information about the slot found in llama_kv_cache_find_slot
struct llama_kv_cache_slot_info {
std::pair<uint32_t, uint32_t> boundaries; // slot boundaries [begin, end)
bool found = false; // the slot was found
explicit llama_kv_cache_slot_info(bool found_) : found{found_} {}
llama_kv_cache_slot_info(uint32_t begin, uint32_t end) : boundaries{begin, end}, found{true} {}
operator bool() const { return found; }
};
// ring-buffer of cached KV data
// TODO: pimpl
// TODO: add notion of max sequences
@@ -93,6 +100,9 @@ public:
void clear() override;
void defrag() override;
virtual void restore() override;
virtual void commit() override;
bool seq_rm (llama_seq_id seq_id, llama_pos p0, llama_pos p1) override;
void seq_cp (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) override;
void seq_keep(llama_seq_id seq_id) override;
@@ -105,10 +115,9 @@ public:
// find an empty slot of size "n_tokens" in the cache
// updates the cache head
// returns a structure holding information about the slot found
// Note: On success, it's important that cache.head points
// to the first cell of the slot.
llama_kv_cache_slot_info find_slot(const llama_ubatch & batch);
bool find_slot(const llama_ubatch & batch);
// TODO: maybe not needed
uint32_t get_padding(const llama_cparams & cparams) const;
@@ -128,7 +137,19 @@ public:
// return true if cells have been moved
bool defrag_prepare(int32_t n_max_nodes);
// state save/load
// commit/restore cache
struct slot_range {
uint32_t c0 = 0; // note: these are cell indices, not sequence positions
uint32_t c1 = 0;
};
// pending cell updates that are not yet committed
struct {
std::vector<slot_range> ranges;
} pending;
// state write/load
void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const;
void state_read (llama_io_read_i & io, llama_seq_id seq_id = -1);
@@ -183,59 +204,6 @@ private:
// using llama_kv_cache_unified::llama_kv_cache_unified;
//};
//
// kv cache restore
//
// saves the kv_cache state for future recovery.
// used to rollback llama_kv_cache_find_slot changes.
struct llama_kv_slot_restorer {
struct llama_kv_cache_state {
uint32_t head = 0;
uint32_t n = 0;
} old_state;
// for non-recurrent models only
// list of slots to restore
std::vector<std::pair<uint32_t, uint32_t>> slot_boundaries;
bool do_restore = false;
llama_kv_cache_unified & cache;
explicit llama_kv_slot_restorer(llama_kv_cache_unified & cache) : cache(cache) {
old_state.head = cache.head;
old_state.n = cache.n;
}
// saves a slot information for future restoration
void save(const llama_kv_cache_slot_info & slot) {
if (slot) {
do_restore = true;
if (slot.boundaries.first != slot.boundaries.second) {
slot_boundaries.push_back(slot.boundaries);
}
}
}
// must be explicitly called to restore the kv_cache state
// and rollback changes from all llama_kv_cache_find_slot calls
void restore() {
if (do_restore) {
cache.head = old_state.head;
cache.n = old_state.n;
if (cache.recurrent) { // recurrent models like Mamba or RWKV can't have a state partially erased
cache.seq_rm(-1, -1, -1);
} else {
for (auto & slot : slot_boundaries) {
cache.seq_rm(-1, slot.first, slot.second);
}
}
}
}
};
// TODO: maybe become part of the public llama_kv_cache in the future
int32_t llama_kv_cache_n_tokens(const llama_kv_cache * kv);
+10 -5
View File
@@ -445,7 +445,8 @@ llama_model_loader::llama_model_loader(
std::vector<std::string> & splits,
bool use_mmap,
bool check_tensors,
const struct llama_model_kv_override * param_overrides_p) {
const llama_model_kv_override * param_overrides_p,
const llama_model_tensor_buft_override * param_tensor_buft_overrides_p) {
int trace = 0;
if (getenv("LLAMA_TRACE")) {
trace = atoi(getenv("LLAMA_TRACE"));
@@ -457,6 +458,8 @@ llama_model_loader::llama_model_loader(
}
}
tensor_buft_overrides = param_tensor_buft_overrides_p;
// Load the main GGUF
struct ggml_context * ctx = NULL;
struct gguf_init_params params = {
@@ -600,7 +603,9 @@ llama_model_loader::llama_model_loader(
if (trace > 0) {
const uint16_t sid = w.idx;
LLAMA_LOG_INFO("%s: - tensor split %2d: %32s %-8s [ %s ]\n", __func__, sid, ggml_get_name(tensor), ggml_type_name(type), llama_format_tensor_shape(tensor).c_str());
LLAMA_LOG_INFO("%s: - tensor split %2d: %32s %-8s [ %s ] %8.2f MiB\n", __func__,
sid, ggml_get_name(tensor), ggml_type_name(type), llama_format_tensor_shape(tensor).c_str(),
ggml_nbytes(tensor)/1024.0f/1024.0f);
}
}
@@ -640,9 +645,9 @@ llama_model_loader::llama_model_loader(
ftype = (llama_ftype) (ftype | LLAMA_FTYPE_GUESSED);
{
const int kid = gguf_find_key(meta.get(), "general.file_type"); // TODO: use LLM_KV
if (kid >= 0) {
ftype = (llama_ftype) gguf_get_val_u32(meta.get(), kid);
uint32_t ftype_val = 0;
if (get_key(LLM_KV_GENERAL_FILE_TYPE, ftype_val, false)) {
ftype = (llama_ftype) ftype_val;
}
}
+5 -3
View File
@@ -77,8 +77,9 @@ struct llama_model_loader {
llama_mmaps mappings;
std::map<std::string, struct llama_tensor_weight, weight_name_comparer> weights_map;
std::unordered_map<std::string, struct llama_model_kv_override> kv_overrides;
std::map<std::string, llama_tensor_weight, weight_name_comparer> weights_map;
std::unordered_map<std::string, llama_model_kv_override> kv_overrides;
const llama_model_tensor_buft_override * tensor_buft_overrides;
gguf_context_ptr meta;
std::vector<ggml_context_ptr> contexts;
@@ -95,7 +96,8 @@ struct llama_model_loader {
std::vector<std::string> & splits, // optional, only need if the split does not follow naming scheme
bool use_mmap,
bool check_tensors,
const struct llama_model_kv_override * param_overrides_p);
const llama_model_kv_override * param_overrides_p,
const llama_model_tensor_buft_override * param_tensor_buft_overrides_p);
template<typename T>
typename std::enable_if<std::is_integral<T>::value, bool>::type
+28 -2
View File
@@ -17,6 +17,7 @@
#include <cmath>
#include <functional>
#include <map>
#include <regex>
#include <sstream>
#include <stdexcept>
@@ -378,9 +379,12 @@ struct llama_model::impl {
layer_dev dev_input = {};
layer_dev dev_output = {};
std::vector<layer_dev> dev_layer;
bool has_tensor_overrides;
};
llama_model::llama_model(const llama_model_params & params) : params(params), pimpl(std::make_unique<impl>()) {
pimpl->has_tensor_overrides = params.tensor_buft_overrides && params.tensor_buft_overrides[0].pattern;
}
llama_model::~llama_model() {}
@@ -1571,9 +1575,26 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
GGML_ABORT("invalid layer %d for tensor %s", info.layer, tn.str().c_str());
}
ggml_backend_buffer_type_t buft = select_weight_buft(hparams, t_meta, op, *buft_list);
ggml_backend_buffer_type_t buft = nullptr;
// check overrides
if (ml.tensor_buft_overrides) {
std::string tensor_name = tn.str();
for (const auto * overrides = ml.tensor_buft_overrides; overrides->pattern != nullptr; ++overrides) {
std::regex pattern(overrides->pattern);
if (std::regex_search(tensor_name, pattern)) {
LLAMA_LOG_DEBUG("tensor %s buffer type overriden to %s\n", tensor_name.c_str(), ggml_backend_buft_name(overrides->buft));
buft = overrides->buft;
break;
}
}
}
if (!buft) {
throw std::runtime_error(format("failed to find a compatible buffer type for tensor %s", tn.str().c_str()));
buft = select_weight_buft(hparams, t_meta, op, *buft_list);
if (!buft) {
throw std::runtime_error(format("failed to find a compatible buffer type for tensor %s", tn.str().c_str()));
}
}
// avoid using a host buffer when using mmap
@@ -4151,6 +4172,10 @@ ggml_backend_buffer_type_t llama_model::select_buft(int il) const {
});
}
bool llama_model::has_tensor_overrides() const {
return pimpl->has_tensor_overrides;
}
const ggml_tensor * llama_model::get_tensor(const char * name) const {
auto it = std::find_if(tensors_by_name.begin(), tensors_by_name.end(),
[name](const std::pair<std::string, ggml_tensor *> & it) {
@@ -12319,6 +12344,7 @@ llm_graph_result_ptr llama_model::build_graph(
llama_model_params llama_model_default_params() {
llama_model_params result = {
/*.devices =*/ nullptr,
/*.tensor_buft_overrides =*/ nullptr,
/*.n_gpu_layers =*/ 0,
/*.split_mode =*/ LLAMA_SPLIT_MODE_LAYER,
/*.main_gpu =*/ 0,
+2
View File
@@ -382,6 +382,8 @@ struct llama_model {
ggml_backend_buffer_type_t select_buft(int il) const;
bool has_tensor_overrides() const;
const struct ggml_tensor * get_tensor(const char * name) const;
// TODO: move this to new llm_arch_model_i interface
+1 -1
View File
@@ -527,7 +527,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
}
std::vector<std::string> splits = {};
llama_model_loader ml(fname_inp, splits, use_mmap, /*check_tensors*/ true, kv_overrides);
llama_model_loader ml(fname_inp, splits, use_mmap, /*check_tensors*/ true, kv_overrides, nullptr);
ml.init_mappings(false); // no prefetching
llama_model model(llama_model_default_params());
+7 -1
View File
@@ -411,7 +411,8 @@ struct llm_tokenizer_bpe : llm_tokenizer {
regex_exprs = {
// original regex from tokenizer.json
// "'(?i:[sdmt]|ll|ve|re)|[^\\r\\n\\p{L}\\p{N}]?+\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]++[\\r\\n]*|\\s*[\\r\\n]|\\s+(?!\\S)|\\s+"
"'(?:[sSdDmMtT]|[lL][lL]|[vV][eE]|[rR][eE])|[^\\r\\n\\p{L}\\p{N}]?+\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]++[\\r\\n]*|\\s*[\\r\\n]|\\s+(?!\\S)|\\s+",
// FIXME? Changed possessive quantifiers (?+ and ++) to greedy to avoid errors and imatrix hanging (tried atomic grouping but it's not supported?)
"'(?:[sSdDmMtT]|[lL][lL]|[vV][eE]|[rR][eE])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]|\\s+(?!\\S)|\\s+",
};
break;
default:
@@ -1807,6 +1808,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
|| t.first == "<end_of_turn>"
|| t.first == "<|endoftext|>"
|| t.first == "<EOT>"
|| t.first == "_<EOT>"
|| t.first == "<|end▁of▁sentence|>" // DeepSeek
) {
special_eot_id = t.second;
@@ -1839,6 +1841,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
|| t.first == "<fim-prefix>"
|| t.first == "<|fim▁begin|>" // DeepSeek
|| t.first == "<PRE>"
|| t.first == "▁<PRE>" // CodeLlama
) {
special_fim_pre_id = t.second;
if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
@@ -1856,6 +1859,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
|| t.first == "<fim-suffix>"
|| t.first == "<|fim▁hole|>" // DeepSeek
|| t.first == "<SUF>"
|| t.first == "▁<SUF>" // CodeLlama
) {
special_fim_suf_id = t.second;
if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
@@ -1873,6 +1877,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
|| t.first == "<fim-middle>"
|| t.first == "<|fim▁end|>" // DeepSeek
|| t.first == "<MID>"
|| t.first == "▁<MID>" // CodeLlama
) {
special_fim_mid_id = t.second;
if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
@@ -1957,6 +1962,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
|| t.first == "<|endoftext|>"
|| t.first == "<|eom_id|>"
|| t.first == "<EOT>"
|| t.first == "_<EOT>"
) {
special_eog_ids.insert(t.second);
if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
+1 -1
View File
@@ -92,7 +92,7 @@ static int llama_model_load(const std::string & fname, std::vector<std::string>
model.t_start_us = tm.t_start_us;
try {
llama_model_loader ml(fname, splits, params.use_mmap, params.check_tensors, params.kv_overrides);
llama_model_loader ml(fname, splits, params.use_mmap, params.check_tensors, params.kv_overrides, params.tensor_buft_overrides);
ml.print_info();
+4 -4
View File
@@ -77,7 +77,7 @@ int main(void) {
argv = {"binary_name", "-m", "model_file.gguf"};
assert(true == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON));
assert(params.model == "model_file.gguf");
assert(params.model.path == "model_file.gguf");
argv = {"binary_name", "-t", "1234"};
assert(true == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON));
@@ -89,7 +89,7 @@ int main(void) {
argv = {"binary_name", "-m", "abc.gguf", "--predict", "6789", "--batch-size", "9090"};
assert(true == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON));
assert(params.model == "abc.gguf");
assert(params.model.path == "abc.gguf");
assert(params.n_predict == 6789);
assert(params.n_batch == 9090);
@@ -112,7 +112,7 @@ int main(void) {
setenv("LLAMA_ARG_THREADS", "1010", true);
argv = {"binary_name"};
assert(true == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON));
assert(params.model == "blah.gguf");
assert(params.model.path == "blah.gguf");
assert(params.cpuparams.n_threads == 1010);
@@ -122,7 +122,7 @@ int main(void) {
setenv("LLAMA_ARG_THREADS", "1010", true);
argv = {"binary_name", "-m", "overwritten.gguf"};
assert(true == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON));
assert(params.model == "overwritten.gguf");
assert(params.model.path == "overwritten.gguf");
assert(params.cpuparams.n_threads == 1010);
#endif // _WIN32
+6
View File
@@ -4516,6 +4516,12 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
}
}
for (int kv : { 4096, 8192, 16384, }) {
for (int hs : { 64, 128, }) {
test_cases.emplace_back(new test_flash_attn_ext(hs, hs, 8, 4, kv, 1, true, 0, 0, GGML_PREC_F32, GGML_TYPE_F16));
}
}
return test_cases;
}