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

Merge remote-tracking branch 'origin/master' into json-fixes

Browse source
This commit is contained in:
ochafik 2024-03-17 22:53:04 +00:00
commit 20869ede26
21 changed files with 940 additions and 174 deletions
Download patch file Download diff file Expand all files Collapse all files

22
.github/workflows/build.yml vendored
View file

@ -48,6 +48,28 @@ jobs:
CC=gcc-8 make tests -j $(nproc)
make test -j $(nproc)
ubuntu-focal-make-curl:
runs-on: ubuntu-20.04
steps:
- name: Clone
id: checkout
uses: actions/checkout@v3
- name: Dependencies
id: depends
run: |
sudo apt-get update
sudo apt-get install build-essential gcc-8 libcurl4-openssl-dev
- name: Build
id: make_build
env:
LLAMA_FATAL_WARNINGS: 1
LLAMA_CURL: 1
run: |
CC=gcc-8 make -j $(nproc)
ubuntu-latest-cmake:
runs-on: ubuntu-latest

23
.github/workflows/close-issue.yml vendored Normal file
View file

@ -0,0 +1,23 @@
name: Close inactive issues
on:
schedule:
- cron: "42 0 * * *"
jobs:
close-issues:
runs-on: ubuntu-latest
permissions:
issues: write
pull-requests: write
steps:
- uses: actions/stale@v5
with:
days-before-issue-stale: 30
days-before-issue-close: 14
stale-issue-label: "stale"
stale-issue-message: "This issue is stale because it has been open for 30 days with no activity."
close-issue-message: "This issue was closed because it has been inactive for 14 days since being marked as stale."
days-before-pr-stale: -1
days-before-pr-close: -1
operations-per-run: 1000
repo-token: ${{ secrets.GITHUB_TOKEN }}

20
.github/workflows/server.yml vendored
View file

@ -57,7 +57,8 @@ jobs:
cmake \
python3-pip \
wget \
language-pack-en
language-pack-en \
libcurl4-openssl-dev
- name: Build
id: cmake_build
@ -67,6 +68,7 @@ jobs:
cmake .. \
-DLLAMA_NATIVE=OFF \
-DLLAMA_BUILD_SERVER=ON \
-DLLAMA_CURL=ON \
-DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
-DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON ;
cmake --build . --config ${{ matrix.build_type }} -j $(nproc) --target server
@ -101,12 +103,21 @@ jobs:
with:
fetch-depth: 0
- name: libCURL
id: get_libcurl
env:
CURL_VERSION: 8.6.0_6
run: |
curl.exe -o $env:RUNNER_TEMP/curl.zip -L "https://curl.se/windows/dl-${env:CURL_VERSION}/curl-${env:CURL_VERSION}-win64-mingw.zip"
mkdir $env:RUNNER_TEMP/libcurl
tar.exe -xvf $env:RUNNER_TEMP/curl.zip --strip-components=1 -C $env:RUNNER_TEMP/libcurl
- name: Build
id: cmake_build
run: |
mkdir build
cd build
cmake .. -DLLAMA_BUILD_SERVER=ON -DCMAKE_BUILD_TYPE=Release ;
cmake .. -DLLAMA_CURL=ON -DCURL_LIBRARY="$env:RUNNER_TEMP/libcurl/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:RUNNER_TEMP/libcurl/include"
cmake --build . --config Release -j ${env:NUMBER_OF_PROCESSORS} --target server
- name: Python setup
@ -120,6 +131,11 @@ jobs:
run: |
pip install -r examples/server/tests/requirements.txt
- name: Copy Libcurl
id: prepare_libcurl
run: |
cp $env:RUNNER_TEMP/libcurl/bin/libcurl-x64.dll ./build/bin/Release/libcurl-x64.dll
- name: Tests
id: server_integration_tests
if: ${{ !matrix.disabled_on_pr || !github.event.pull_request }}

1
CMakeLists.txt
View file

@ -99,6 +99,7 @@ option(LLAMA_CUDA_F16 "llama: use 16 bit floats for some
set(LLAMA_CUDA_KQUANTS_ITER "2" CACHE STRING "llama: iters./thread per block for Q2_K/Q6_K")
set(LLAMA_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
"llama: max. batch size for using peer access")
option(LLAMA_CURL "llama: use libcurl to download model from an URL" OFF)
option(LLAMA_HIPBLAS "llama: use hipBLAS" OFF)
option(LLAMA_HIP_UMA "llama: use HIP unified memory architecture" OFF)
option(LLAMA_CLBLAST "llama: use CLBlast" OFF)

5
Makefile
View file

@ -596,6 +596,11 @@ include scripts/get-flags.mk
CUDA_CXXFLAGS := $(BASE_CXXFLAGS) $(GF_CXXFLAGS) -Wno-pedantic
endif
ifdef LLAMA_CURL
override CXXFLAGS := $(CXXFLAGS) -DLLAMA_USE_CURL
override LDFLAGS := $(LDFLAGS) -lcurl
endif
#
# Print build information
#

1
README.md
View file

@ -134,6 +134,7 @@ Typically finetunes of the base models below are supported as well.
- Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp)
- JS/TS (llama.cpp server client): [lgrammel/modelfusion](https://modelfusion.dev/integration/model-provider/llamacpp)
- JavaScript/Wasm (works in browser): [tangledgroup/llama-cpp-wasm](https://github.com/tangledgroup/llama-cpp-wasm)
- Typescript/Wasm (nicer API, available on npm): [ngxson/wllama](https://github.com/ngxson/wllama)
- Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb)
- Rust (nicer API): [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp)
- Rust (more direct bindings): [utilityai/llama-cpp-rs](https://github.com/utilityai/llama-cpp-rs)

13
common/CMakeLists.txt
View file

@ -68,6 +68,17 @@ if (BUILD_SHARED_LIBS)
set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
endif()
set(LLAMA_COMMON_EXTRA_LIBS build_info)
# Use curl to download model url
if (LLAMA_CURL)
find_package(CURL REQUIRED)
add_definitions(-DLLAMA_USE_CURL)
include_directories(${CURL_INCLUDE_DIRS})
find_library(CURL_LIBRARY curl REQUIRED)
set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} ${CURL_LIBRARY})
endif ()
target_include_directories(${TARGET} PUBLIC .)
target_compile_features(${TARGET} PUBLIC cxx_std_11)
target_link_libraries(${TARGET} PRIVATE build_info PUBLIC llama)
target_link_libraries(${TARGET} PRIVATE ${LLAMA_COMMON_EXTRA_LIBS} PUBLIC llama)

713
common/common.cpp
View file

File diff suppressed because it is too large Load diff

4
common/common.h
View file

@ -89,6 +89,7 @@ struct gpt_params {
struct llama_sampling_params sparams;
std::string model = "models/7B/ggml-model-f16.gguf"; // model path
std::string model_url = ""; // model url to download
std::string model_draft = ""; // draft model for speculative decoding
std::string model_alias = "unknown"; // model alias
std::string prompt = "";
@ -191,6 +192,9 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
struct llama_model_params llama_model_params_from_gpt_params (const gpt_params & params);
struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params);
struct llama_model * llama_load_model_from_url(const char * model_url, const char * path_model,
struct llama_model_params params);
// Batch utils
void llama_batch_clear(struct llama_batch & batch);

62
examples/gritlm/README.md Normal file
View file

@ -0,0 +1,62 @@
## Generative Representational Instruction Tuning (GRIT) Example
[gritlm] a model which can generate embeddings as well as "normal" text
generation depending on the instructions in the prompt.
* Paper: https://arxiv.org/pdf/2402.09906.pdf
### Retrieval-Augmented Generation (RAG) use case
One use case for `gritlm` is to use it with RAG. If we recall how RAG works is
that we take documents that we want to use as context, to ground the large
language model (LLM), and we create token embeddings for them. We then store
these token embeddings in a vector database.
When we perform a query, prompt the LLM, we will first create token embeddings
for the query and then search the vector database to retrieve the most
similar vectors, and return those documents so they can be passed to the LLM as
context. Then the query and the context will be passed to the LLM which will
have to _again_ create token embeddings for the query. But because gritlm is used
the first query can be cached and the second query tokenization generation does
not have to be performed at all.
### Running the example
Download a Grit model:
```console
$ scripts/hf.sh --repo cohesionet/GritLM-7B_gguf --file gritlm-7b_q4_1.gguf
```
Run the example using the downloaded model:
```console
$ ./gritlm -m gritlm-7b_q4_1.gguf
Cosine similarity between "Bitcoin: A Peer-to-Peer Electronic Cash System" and "A purely peer-to-peer version of electronic cash w" is: 0.605
Cosine similarity between "Bitcoin: A Peer-to-Peer Electronic Cash System" and "All text-based language problems can be reduced to" is: 0.103
Cosine similarity between "Generative Representational Instruction Tuning" and "A purely peer-to-peer version of electronic cash w" is: 0.112
Cosine similarity between "Generative Representational Instruction Tuning" and "All text-based language problems can be reduced to" is: 0.547
Oh, brave adventurer, who dared to climb
The lofty peak of Mt. Fuji in the night,
When shadows lurk and ghosts do roam,
And darkness reigns, a fearsome sight.
Thou didst set out, with heart aglow,
To conquer this mountain, so high,
And reach the summit, where the stars do glow,
And the moon shines bright, up in the sky.
Through the mist and fog, thou didst press on,
With steadfast courage, and a steadfast will,
Through the darkness, thou didst not be gone,
But didst climb on, with a steadfast skill.
At last, thou didst reach the summit's crest,
And gazed upon the world below,
And saw the beauty of the night's best,
And felt the peace, that only nature knows.
Oh, brave adventurer, who dared to climb
The lofty peak of Mt. Fuji in the night,
Thou art a hero, in the eyes of all,
For thou didst conquer this mountain, so bright.
```
[gritlm]: https://github.com/ContextualAI/gritlm

1
examples/main/README.md
View file

@ -67,6 +67,7 @@ main.exe -m models\7B\ggml-model.bin --ignore-eos -n -1 --random-prompt
In this section, we cover the most commonly used options for running the `main` program with the LLaMA models:
- `-m FNAME, --model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.bin`).
- `-mu MODEL_URL --model-url MODEL_URL`: Specify a remote http url to download the file (e.g https://huggingface.co/ggml-org/models/resolve/main/phi-2/ggml-model-q4_0.gguf).
- `-i, --interactive`: Run the program in interactive mode, allowing you to provide input directly and receive real-time responses.
- `-ins, --instruct`: Run the program in instruction mode, which is particularly useful when working with Alpaca models.
- `-n N, --n-predict N`: Set the number of tokens to predict when generating text. Adjusting this value can influence the length of the generated text.

1
examples/server/README.md
View file

@ -20,6 +20,7 @@ The project is under active development, and we are [looking for feedback and co
- `-tb N, --threads-batch N`: Set the number of threads to use during batch and prompt processing. If not specified, the number of threads will be set to the number of threads used for generation.
- `--threads-http N`: number of threads in the http server pool to process requests (default: `max(std::thread::hardware_concurrency() - 1, --parallel N + 2)`)
- `-m FNAME`, `--model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.gguf`).
- `-mu MODEL_URL --model-url MODEL_URL`: Specify a remote http url to download the file (e.g https://huggingface.co/ggml-org/models/resolve/main/phi-2/ggml-model-q4_0.gguf).
- `-a ALIAS`, `--alias ALIAS`: Set an alias for the model. The alias will be returned in API responses.
- `-c N`, `--ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference. The size may differ in other models, for example, baichuan models were build with a context of 4096.
- `-ngl N`, `--n-gpu-layers N`: When compiled with appropriate support (currently CLBlast or cuBLAS), this option allows offloading some layers to the GPU for computation. Generally results in increased performance.

8
examples/server/server.cpp
View file

@ -2202,6 +2202,8 @@ static void server_print_usage(const char * argv0, const gpt_params & params, co
}
printf(" -m FNAME, --model FNAME\n");
printf(" model path (default: %s)\n", params.model.c_str());
printf(" -mu MODEL_URL, --model-url MODEL_URL\n");
printf(" model download url (default: %s)\n", params.model_url.c_str());
printf(" -a ALIAS, --alias ALIAS\n");
printf(" set an alias for the model, will be added as `model` field in completion response\n");
printf(" --lora FNAME apply LoRA adapter (implies --no-mmap)\n");
@ -2324,6 +2326,12 @@ static void server_params_parse(int argc, char ** argv, server_params & sparams,
break;
}
params.model = argv[i];
} else if (arg == "-mu" || arg == "--model-url") {
if (++i >= argc) {
invalid_param = true;
break;
}
params.model_url = argv[i];
} else if (arg == "-a" || arg == "--alias") {
if (++i >= argc) {
invalid_param = true;

2
examples/server/tests/README.md
View file

@ -57,7 +57,7 @@ Feature or Scenario must be annotated with `@llama.cpp` to be included in the de
To run a scenario annotated with `@bug`, start:
```shell
DEBUG=ON ./tests.sh --no-skipped --tags bug
DEBUG=ON ./tests.sh --no-skipped --tags bug --stop
```
After changing logic in `steps.py`, ensure that `@bug` and `@wrong_usage` scenario are updated.

3
examples/server/tests/features/embeddings.feature
View file

@ -4,7 +4,8 @@ Feature: llama.cpp server
Background: Server startup
Given a server listening on localhost:8080
And a model file bert-bge-small/ggml-model-f16.gguf from HF repo ggml-org/models
And a model url https://huggingface.co/ggml-org/models/resolve/main/bert-bge-small/ggml-model-f16.gguf
And a model file ggml-model-f16.gguf
And a model alias bert-bge-small
And 42 as server seed
And 2 slots

93
examples/server/tests/features/environment.py
View file

@ -1,10 +1,12 @@
import errno
import os
import socket
import subprocess
import time
from contextlib import closing
import signal
import socket
import sys
import time
import traceback
from contextlib import closing
import psutil
def before_scenario(context, scenario):
@ -20,33 +22,40 @@ def before_scenario(context, scenario):
def after_scenario(context, scenario):
if context.server_process is None:
return
if scenario.status == "failed":
if 'GITHUB_ACTIONS' in os.environ:
print(f"\x1b[33;101mSCENARIO FAILED: {scenario.name} server logs:\x1b[0m\n\n")
if os.path.isfile('llama.log'):
with closing(open('llama.log', 'r')) as f:
for line in f:
print(line)
if not is_server_listening(context.server_fqdn, context.server_port):
print("\x1b[33;101mERROR: Server stopped listening\x1b[0m\n")
try:
if 'server_process' not in context or context.server_process is None:
return
if scenario.status == "failed":
if 'GITHUB_ACTIONS' in os.environ:
print(f"\x1b[33;101mSCENARIO FAILED: {scenario.name} server logs:\x1b[0m\n\n")
if os.path.isfile('llama.log'):
with closing(open('llama.log', 'r')) as f:
for line in f:
print(line)
if not is_server_listening(context.server_fqdn, context.server_port):
print("\x1b[33;101mERROR: Server stopped listening\x1b[0m\n")
if not pid_exists(context.server_process.pid):
assert False, f"Server not running pid={context.server_process.pid} ..."
if not pid_exists(context.server_process.pid):
assert False, f"Server not running pid={context.server_process.pid} ..."
server_graceful_shutdown(context)
server_graceful_shutdown(context)
# Wait few for socket to free up
time.sleep(0.05)
# Wait few for socket to free up
time.sleep(0.05)
attempts = 0
while pid_exists(context.server_process.pid) or is_server_listening(context.server_fqdn, context.server_port):
server_kill(context)
time.sleep(0.1)
attempts += 1
if attempts > 5:
server_kill_hard(context)
attempts = 0
while pid_exists(context.server_process.pid) or is_server_listening(context.server_fqdn, context.server_port):
server_kill(context)
time.sleep(0.1)
attempts += 1
if attempts > 5:
server_kill_hard(context)
except:
exc = sys.exception()
print("error in after scenario: \n")
print(exc)
print("*** print_tb: \n")
traceback.print_tb(exc.__traceback__, file=sys.stdout)
def server_graceful_shutdown(context):
@ -67,11 +76,11 @@ def server_kill_hard(context):
path = context.server_path
print(f"Server dangling exits, hard killing force {pid}={path}...\n")
if os.name == 'nt':
process = subprocess.check_output(['taskkill', '/F', '/pid', str(pid)]).decode()
print(process)
else:
os.kill(-pid, signal.SIGKILL)
try:
psutil.Process(pid).kill()
except psutil.NoSuchProcess:
return False
return True
def is_server_listening(server_fqdn, server_port):
@ -84,17 +93,9 @@ def is_server_listening(server_fqdn, server_port):
def pid_exists(pid):
"""Check whether pid exists in the current process table."""
if pid < 0:
try:
psutil.Process(pid)
except psutil.NoSuchProcess:
return False
if os.name == 'nt':
output = subprocess.check_output(['TASKLIST', '/FI', f'pid eq {pid}']).decode()
print(output)
return "No tasks are running" not in output
else:
try:
os.kill(pid, 0)
except OSError as e:
return e.errno == errno.EPERM
else:
return True
return True

3
examples/server/tests/features/server.feature
View file

@ -4,7 +4,8 @@ Feature: llama.cpp server
Background: Server startup
Given a server listening on localhost:8080
And a model file tinyllamas/stories260K.gguf from HF repo ggml-org/models
And a model url https://huggingface.co/ggml-org/models/resolve/main/tinyllamas/stories260K.gguf
And a model file stories260K.gguf
And a model alias tinyllama-2
And 42 as server seed
# KV Cache corresponds to the total amount of tokens

37
examples/server/tests/features/steps/steps.py
View file

@ -5,6 +5,8 @@ import os
import re
import socket
import subprocess
import sys
import threading
import time
from contextlib import closing
from re import RegexFlag
@ -32,6 +34,8 @@ def step_server_config(context, server_fqdn, server_port):
context.base_url = f'http://{context.server_fqdn}:{context.server_port}'
context.model_alias = None
context.model_file = None
context.model_url = None
context.n_batch = None
context.n_ubatch = None
context.n_ctx = None
@ -66,6 +70,16 @@ def step_download_hf_model(context, hf_file, hf_repo):
print(f"model file: {context.model_file}\n")
@step('a model file {model_file}')
def step_model_file(context, model_file):
context.model_file = model_file
@step('a model url {model_url}')
def step_model_url(context, model_url):
context.model_url = model_url
@step('a model alias {model_alias}')
def step_model_alias(context, model_alias):
context.model_alias = model_alias
@ -142,7 +156,8 @@ def step_start_server(context):
async def step_wait_for_the_server_to_be_started(context, expecting_status):
match expecting_status:
case 'healthy':
await wait_for_health_status(context, context.base_url, 200, 'ok')
await wait_for_health_status(context, context.base_url, 200, 'ok',
timeout=30)
case 'ready' | 'idle':
await wait_for_health_status(context, context.base_url, 200, 'ok',
@ -1055,8 +1070,11 @@ def start_server_background(context):
server_args = [
'--host', server_listen_addr,
'--port', context.server_port,
'--model', context.model_file
]
if context.model_file:
server_args.extend(['--model', context.model_file])
if context.model_url:
server_args.extend(['--model-url', context.model_url])
if context.n_batch:
server_args.extend(['--batch-size', context.n_batch])
if context.n_ubatch:
@ -1096,8 +1114,23 @@ def start_server_background(context):
pkwargs = {
'creationflags': flags,
'stdout': subprocess.PIPE,
'stderr': subprocess.PIPE
}
context.server_process = subprocess.Popen(
[str(arg) for arg in [context.server_path, *server_args]],
**pkwargs)
def log_stdout(process):
for line in iter(process.stdout.readline, b''):
print(line.decode('utf-8'), end='')
thread_stdout = threading.Thread(target=log_stdout, args=(context.server_process,))
thread_stdout.start()
def log_stderr(process):
for line in iter(process.stderr.readline, b''):
print(line.decode('utf-8'), end='', file=sys.stderr)
thread_stderr = threading.Thread(target=log_stderr, args=(context.server_process,))
thread_stderr.start()
print(f"server pid={context.server_process.pid}, behave pid={os.getpid()}")

1
examples/server/tests/requirements.txt
View file

@ -3,4 +3,5 @@ behave~=1.2.6
huggingface_hub~=0.20.3
numpy~=1.24.4
openai~=0.25.0
psutil~=5.9.8
prometheus-client~=0.20.0

6
ggml-vulkan.cpp
View file

@ -710,6 +710,12 @@ static uint32_t ggml_vk_find_queue_family_index(std::vector<vk::QueueFamilyPrope
}
}
// All commands that are allowed on a queue that supports transfer operations are also allowed on a queue that supports either graphics or compute operations.
// Thus, if the capabilities of a queue family include VK_QUEUE_GRAPHICS_BIT or VK_QUEUE_COMPUTE_BIT, then reporting the VK_QUEUE_TRANSFER_BIT capability separately for that queue family is optional.
if (compute_index >= 0) {
return compute_index;
}
std::cerr << "ggml_vulkan: No suitable queue family index found." << std::endl;
for(auto &q_family : queue_family_props) {

95
ggml.c
View file

@ -931,6 +931,101 @@ inline static float vaddvq_f32(float32x4_t v) {
#define GGML_F16_VEC_REDUCE GGML_F32Cx4_REDUCE
#endif
#elif defined(__AVX512F__)
#define GGML_SIMD
// F32 AVX512
#define GGML_F32_STEP 64
#define GGML_F32_EPR 16
#define GGML_F32x16 __m512
#define GGML_F32x16_ZERO _mm512_setzero_ps()
#define GGML_F32x16_SET1(x) _mm512_set1_ps(x)
#define GGML_F32x16_LOAD _mm512_loadu_ps
#define GGML_F32x16_STORE _mm512_storeu_ps
// _mm512_fmadd_ps is defined in AVX512F so no guard is required
#define GGML_F32x16_FMA(a, b, c) _mm512_fmadd_ps(b, c, a)
#define GGML_F32x16_ADD _mm512_add_ps
#define GGML_F32x16_MUL _mm512_mul_ps
#define GGML_F32x16_REDUCE(res, x) \
do { \
int offset = GGML_F32_ARR >> 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = _mm512_add_ps(x[i], x[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = _mm512_add_ps(x[i], x[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = _mm512_add_ps(x[i], x[offset+i]); \
} \
res = _mm512_reduce_add_ps(x[0]); \
} while (0)
// TODO: is this optimal ?
#define GGML_F32_VEC GGML_F32x16
#define GGML_F32_VEC_ZERO GGML_F32x16_ZERO
#define GGML_F32_VEC_SET1 GGML_F32x16_SET1
#define GGML_F32_VEC_LOAD GGML_F32x16_LOAD
#define GGML_F32_VEC_STORE GGML_F32x16_STORE
#define GGML_F32_VEC_FMA GGML_F32x16_FMA
#define GGML_F32_VEC_ADD GGML_F32x16_ADD
#define GGML_F32_VEC_MUL GGML_F32x16_MUL
#define GGML_F32_VEC_REDUCE GGML_F32x16_REDUCE
// F16 AVX512
// F16 AVX
#define GGML_F16_STEP 64
#define GGML_F16_EPR 16
// AVX512 has FP16 extension (AVX512_FP16) but I don't have it on my machine so I use FP32 instead
#define GGML_F32Cx16 __m512
#define GGML_F32Cx16_ZERO _mm512_setzero_ps()
#define GGML_F32Cx16_SET1(x) _mm512_set1_ps(x)
// unlike _mm256_cvt intrinsics that require F16C, _mm512_cvt is defined in AVX512F
// so F16C guard isn't required
#define GGML_F32Cx16_LOAD(x) _mm512_cvtph_ps(_mm256_loadu_si256((__m256i *)(x)))
#define GGML_F32Cx16_STORE(x, y) _mm256_storeu_si256((__m256i *)(x), _mm512_cvtps_ph(y, 0))
#define GGML_F32Cx16_FMA(a, b, c) _mm512_fmadd_ps(b, c, a)
#define GGML_F32Cx16_ADD _mm512_add_ps
#define GGML_F32Cx16_MUL _mm512_mul_ps
#define GGML_F32Cx16_REDUCE(res, x) \
do { \
int offset = GGML_F32_ARR >> 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = _mm512_add_ps(x[i], x[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = _mm512_add_ps(x[i], x[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = _mm512_add_ps(x[i], x[offset+i]); \
} \
res = _mm512_reduce_add_ps(x[0]); \
} while (0)
#define GGML_F16_VEC GGML_F32Cx16
#define GGML_F16_VEC_ZERO GGML_F32Cx16_ZERO
#define GGML_F16_VEC_SET1 GGML_F32Cx16_SET1
#define GGML_F16_VEC_LOAD(p, i) GGML_F32Cx16_LOAD(p)
#define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx16_STORE(p, r[i])
#define GGML_F16_VEC_FMA GGML_F32Cx16_FMA
#define GGML_F16_VEC_ADD GGML_F32Cx16_ADD
#define GGML_F16_VEC_MUL GGML_F32Cx16_MUL
#define GGML_F16_VEC_REDUCE GGML_F32Cx16_REDUCE
#elif defined(__AVX__)
#define GGML_SIMD