happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'master' into compilade/convert-prequant

This commit is contained in:
Francis Couture-Harpin 2025-08-14 17:05:21 -04:00
parent de12f8ac50 df36bce667
commit 1ae6ab7601
292 changed files with 88358 additions and 36487 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

130
.devops/cann.Dockerfile Normal file
View File

@ -0,0 +1,130 @@
# ==============================================================================
# ARGUMENTS
# ==============================================================================
# Define the CANN base image for easier version updates later
ARG CANN_BASE_IMAGE=quay.io/ascend/cann:8.1.rc1-910b-openeuler22.03-py3.10
# ==============================================================================
# BUILD STAGE
# Compile all binary files and libraries
# ==============================================================================
FROM ${CANN_BASE_IMAGE} AS build
# Define the Ascend chip model for compilation. Default is Ascend910B3
ARG ASCEND_SOC_TYPE=Ascend910B3
# -- Install build dependencies --
RUN yum install -y gcc g++ cmake make git libcurl-devel python3 python3-pip && \
yum clean all && \
rm -rf /var/cache/yum
# -- Set the working directory --
WORKDIR /app
# -- Copy project files --
COPY . .
# -- Set CANN environment variables (required for compilation) --
# Using ENV instead of `source` allows environment variables to persist across the entire image layer
ENV ASCEND_TOOLKIT_HOME=/usr/local/Ascend/ascend-toolkit/latest
ENV LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:${LD_LIBRARY_PATH}
ENV PATH=${ASCEND_TOOLKIT_HOME}/bin:${PATH}
ENV ASCEND_OPP_PATH=${ASCEND_TOOLKIT_HOME}/opp
ENV LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/runtime/lib64/stub:$LD_LIBRARY_PATH
# ... You can add other environment variables from the original file as needed ...
# For brevity, only core variables are listed here. You can paste the original ENV list here.
# -- Build llama.cpp --
# Use the passed ASCEND_SOC_TYPE argument and add general build options
RUN source /usr/local/Ascend/ascend-toolkit/set_env.sh --force \
&& \
cmake -B build \
-DGGML_CANN=ON \
-DCMAKE_BUILD_TYPE=Release \
-DSOC_TYPE=${ASCEND_SOC_TYPE} \
. && \
cmake --build build --config Release -j$(nproc)
# -- Organize build artifacts for copying in later stages --
# Create a lib directory to store all .so files
RUN mkdir -p /app/lib && \
find build -name "*.so" -exec cp {} /app/lib \;
# Create a full directory to store all executables and Python scripts
RUN mkdir -p /app/full && \
cp build/bin/* /app/full/ && \
cp *.py /app/full/ && \
cp -r gguf-py /app/full/ && \
cp -r requirements /app/full/ && \
cp requirements.txt /app/full/
# If you have a tools.sh script, make sure it is copied here
# cp .devops/tools.sh /app/full/tools.sh
# ==============================================================================
# BASE STAGE
# Create a minimal base image with CANN runtime and common libraries
# ==============================================================================
FROM ${CANN_BASE_IMAGE} AS base
# -- Install runtime dependencies --
RUN yum install -y libgomp curl && \
yum clean all && \
rm -rf /var/cache/yum
# -- Set CANN environment variables (required for runtime) --
ENV ASCEND_TOOLKIT_HOME=/usr/local/Ascend/ascend-toolkit/latest
ENV LD_LIBRARY_PATH=/app:${ASCEND_TOOLKIT_HOME}/lib64:${LD_LIBRARY_PATH}
ENV PATH=${ASCEND_TOOLKIT_HOME}/bin:${PATH}
ENV ASCEND_OPP_PATH=${ASCEND_TOOLKIT_HOME}/opp
# ... You can add other environment variables from the original file as needed ...
WORKDIR /app
# Copy compiled .so files from the build stage
COPY --from=build /app/lib/ /app
# ==============================================================================
# FINAL STAGES (TARGETS)
# ==============================================================================
### Target: full
# Complete image with all tools, Python bindings, and dependencies
# ==============================================================================
FROM base AS full
COPY --from=build /app/full /app
# Install Python dependencies
RUN yum install -y git python3 python3-pip && \
pip3 install --no-cache-dir --upgrade pip setuptools wheel && \
pip3 install --no-cache-dir -r requirements.txt && \
yum clean all && \
rm -rf /var/cache/yum
# You need to provide a tools.sh script as the entrypoint
ENTRYPOINT ["/app/tools.sh"]
# If there is no tools.sh, you can set the default to start the server
# ENTRYPOINT ["/app/llama-server"]
### Target: light
# Lightweight image containing only llama-cli
# ==============================================================================
FROM base AS light
COPY --from=build /app/full/llama-cli /app
ENTRYPOINT [ "/app/llama-cli" ]
### Target: server
# Dedicated server image containing only llama-server
# ==============================================================================
FROM base AS server
ENV LLAMA_ARG_HOST=0.0.0.0
COPY --from=build /app/full/llama-server /app
HEALTHCHECK --interval=5m CMD [ "curl", "-f", "http://localhost:8080/health" ]
ENTRYPOINT [ "/app/llama-server" ]

22
.devops/cloud-v-pipeline
View File

@ -1,22 +0,0 @@
node('x86_runner1'){ // Running on x86 runner containing latest vector qemu, latest vector gcc and all the necessary libraries
stage('Cleanup'){
cleanWs() // Cleaning previous CI build in workspace
}
stage('checkout repo'){
retry(5){ // Retry if the cloning fails due to some reason
checkout scm // Clone the repo on Runner
}
}
stage('Compiling llama.cpp'){
sh'''#!/bin/bash
make RISCV=1 RISCV_CROSS_COMPILE=1 # Compiling llama for RISC-V
'''
}
stage('Running llama.cpp'){
sh'''#!/bin/bash
module load gnu-bin2/0.1 # loading latest versions of vector qemu and vector gcc
qemu-riscv64 -L /softwares/gnu-bin2/sysroot -cpu rv64,v=true,vlen=256,elen=64,vext_spec=v1.0 ./llama-cli -m /home/alitariq/codellama-7b.Q4_K_M.gguf -p "Anything" -n 9 > llama_log.txt # Running llama.cpp on vector qemu-riscv64
cat llama_log.txt # Printing results
'''
}
}

6
.devops/cpu.Dockerfile
View File

@ -4,8 +4,6 @@ FROM ubuntu:$UBUNTU_VERSION AS build
ARG TARGETARCH
ARG GGML_CPU_ARM_ARCH=armv8-a
RUN apt-get update && \
apt-get install -y build-essential git cmake libcurl4-openssl-dev
@ -13,10 +11,8 @@ WORKDIR /app
COPY . .
RUN if [ "$TARGETARCH" = "amd64" ]; then \
RUN if [ "$TARGETARCH" = "amd64" ] || [ "$TARGETARCH" = "arm64" ]; then \
cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_TESTS=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON; \
elif [ "$TARGETARCH" = "arm64" ]; then \
cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_TESTS=OFF -DGGML_CPU_ARM_ARCH=${GGML_CPU_ARM_ARCH}; \
else \
echo "Unsupported architecture"; \
exit 1; \

3
.devops/cuda.Dockerfile
View File

@ -60,8 +60,7 @@ RUN apt-get update \
git \
python3 \
python3-pip \
&& pip install --upgrade pip setuptools wheel \
&& pip install -r requirements.txt \
&& pip install --break-system-packages -r requirements.txt \
&& apt autoremove -y \
&& apt clean -y \
&& rm -rf /tmp/* /var/tmp/* \

6
.devops/musa.Dockerfile
View File

@ -1,10 +1,10 @@
ARG UBUNTU_VERSION=22.04
# This needs to generally match the container host's environment.
ARG MUSA_VERSION=rc4.0.1
ARG MUSA_VERSION=rc4.2.0
# Target the MUSA build image
ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-mudnn-devel-ubuntu${UBUNTU_VERSION}
ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}-amd64
ARG BASE_MUSA_RUN_CONTAINER=mthreads/musa:${MUSA_VERSION}-mudnn-runtime-ubuntu${UBUNTU_VERSION}
ARG BASE_MUSA_RUN_CONTAINER=mthreads/musa:${MUSA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}-amd64
FROM ${BASE_MUSA_DEV_CONTAINER} AS build

4
.devops/rocm.Dockerfile
View File

@ -1,8 +1,8 @@
ARG UBUNTU_VERSION=24.04
# This needs to generally match the container host's environment.
ARG ROCM_VERSION=6.3
ARG AMDGPU_VERSION=6.3
ARG ROCM_VERSION=6.4
ARG AMDGPU_VERSION=6.4
# Target the CUDA build image
ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete

43
.github/workflows/build-riscv-native.yml vendored Normal file
View File

@ -0,0 +1,43 @@
name: Build on RISCV Linux Machine by Cloud-V
on:
workflow_dispatch:
workflow_call:
jobs:
bianbu-riscv64-native: # Bianbu 2.2
runs-on: self-hosted
steps:
- name: Install prerequisites
run: |
sudo apt-get update || true
sudo apt-get install -y libatomic1
- uses: actions/checkout@v4
- name: Setup Riscv
run: |
sudo apt-get update || true
sudo apt-get install -y --no-install-recommends \
build-essential \
gcc-14-riscv64-linux-gnu \
g++-14-riscv64-linux-gnu \
cmake
- name: Build
run: |
cmake -B build -DLLAMA_CURL=OFF \
-DCMAKE_BUILD_TYPE=Release \
-DGGML_OPENMP=OFF \
-DLLAMA_BUILD_EXAMPLES=ON \
-DLLAMA_BUILD_TOOLS=ON \
-DLLAMA_BUILD_TESTS=OFF \
-DCMAKE_SYSTEM_NAME=Linux \
-DCMAKE_SYSTEM_PROCESSOR=riscv64 \
-DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
-DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14 \
-DCMAKE_POSITION_INDEPENDENT_CODE=ON \
-DCMAKE_FIND_ROOT_PATH=/usr/lib/riscv64-linux-gnu \
-DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
-DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
-DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
cmake --build build --config Release -j $(nproc)

78
.github/workflows/build.yml vendored
View File

@ -159,31 +159,15 @@ jobs:
- name: Dawn Dependency
id: dawn-depends
run: |
ARTIFACTS_JSON=$(curl -s -L \
-H "Accept: application/vnd.github+json" \
-H "Authorization: Bearer ${{ secrets.GITHUB_TOKEN }}" \
-H "X-GitHub-Api-Version: 2022-11-28" \
"https://api.github.com/repos/google/dawn/actions/artifacts")
echo "Finding latest macos-latest-Release artifact..."
DOWNLOAD_URL=$(echo "$ARTIFACTS_JSON" | jq -r '.artifacts
| sort_by(.created_at)
| reverse
| map(select(.name | test("macos-latest-Release$")))
| .[0].archive_download_url')
if [ "$DOWNLOAD_URL" = "null" ] || [ -z "$DOWNLOAD_URL" ]; then
echo "No suitable Dawn artifact found!"
exit 1
fi
echo "Downloading from: $DOWNLOAD_URL"
curl -L \
-H "Accept: application/vnd.github+json" \
-H "Authorization: Bearer ${{ secrets.GITHUB_TOKEN }}" \
-o artifact.zip "$DOWNLOAD_URL"
unzip artifact.zip
DAWN_VERSION="v1.0.0"
DAWN_OWNER="reeselevine"
DAWN_REPO="dawn"
DAWN_ASSET_NAME="Dawn-a1a6b45cced25a3b7f4fb491e0ae70796cc7f22b-macos-latest-Release.tar.gz"
echo "Fetching release asset from https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}"
curl -L -o artifact.tar.gz \
"https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}"
mkdir dawn
tar_file=$(find . -name '*.tar.gz' | head -n 1)
echo "Extracting: $tar_file"
tar -xvf "$tar_file" -C dawn --strip-components=1
tar -xvf artifact.tar.gz -C dawn --strip-components=1
- name: Build
id: cmake_build
@ -433,31 +417,15 @@ jobs:
id: dawn-depends
run: |
sudo apt-get install -y libxrandr-dev libxinerama-dev libxcursor-dev mesa-common-dev libx11-xcb-dev libxi-dev
ARTIFACTS_JSON=$(curl -s -L \
-H "Accept: application/vnd.github+json" \
-H "Authorization: Bearer ${{ secrets.GITHUB_TOKEN }}" \
-H "X-GitHub-Api-Version: 2022-11-28" \
"https://api.github.com/repos/google/dawn/actions/artifacts")
echo "Finding latest ubuntu-latest-Release artifact..."
DOWNLOAD_URL=$(echo "$ARTIFACTS_JSON" | jq -r '.artifacts
| sort_by(.created_at)
| reverse
| map(select(.name | test("ubuntu-latest-Release$")))
| .[0].archive_download_url')
if [ "$DOWNLOAD_URL" = "null" ] || [ -z "$DOWNLOAD_URL" ]; then
echo "No suitable Dawn artifact found!"
exit 1
fi
echo "Downloading from: $DOWNLOAD_URL"
curl -L \
-H "Accept: application/vnd.github+json" \
-H "Authorization: Bearer ${{ secrets.GITHUB_TOKEN }}" \
-o artifact.zip "$DOWNLOAD_URL"
unzip artifact.zip
DAWN_VERSION="v1.0.0"
DAWN_OWNER="reeselevine"
DAWN_REPO="dawn"
DAWN_ASSET_NAME="Dawn-a1a6b45cced25a3b7f4fb491e0ae70796cc7f22b-ubuntu-latest-Release.tar.gz"
echo "Fetching release asset from https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}"
curl -L -o artifact.tar.gz \
"https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}"
mkdir dawn
tar_file=$(find . -name '*.tar.gz' | head -n 1)
echo "Extracting: $tar_file"
tar -xvf "$tar_file" -C dawn --strip-components=1
tar -xvf artifact.tar.gz -C dawn --strip-components=1
- name: Build
id: cmake_build
@ -475,7 +443,7 @@ jobs:
ubuntu-22-cmake-hip:
runs-on: ubuntu-22.04
container: rocm/dev-ubuntu-22.04:6.0.2
container: rocm/dev-ubuntu-22.04:6.1.2
steps:
- name: Clone
@ -503,19 +471,9 @@ jobs:
-DGGML_HIP=ON
cmake --build build --config Release -j $(nproc)
- name: Build with legacy HIP support
id: cmake_build_legacy_hip
run: |
cmake -B build2 -S . \
-DCMAKE_C_COMPILER=hipcc \
-DCMAKE_CXX_COMPILER=hipcc \
-DGGML_HIP_ROCWMMA_FATTN=ON \
-DGGML_HIP=ON
cmake --build build2 --config Release -j $(nproc)
ubuntu-22-cmake-musa:
runs-on: ubuntu-22.04
container: mthreads/musa:rc4.0.1-mudnn-devel-ubuntu22.04
container: mthreads/musa:rc4.2.0-devel-ubuntu22.04-amd64
steps:
- name: Clone

2
.github/workflows/close-issue.yml vendored
View File

@ -17,7 +17,7 @@ jobs:
steps:
- uses: actions/stale@v5
with:
exempt-issue-labels: "refactor,help wanted,good first issue,research,bug,roadmap"
exempt-issue-labels: "refactoring,help wanted,good first issue,research,bug,roadmap"
days-before-issue-stale: 30
days-before-issue-close: 14
stale-issue-label: "stale"

53
.github/workflows/copilot-setup-steps.yml vendored Normal file
View File

@ -0,0 +1,53 @@
name: "Copilot Setup Steps"
# Automatically run the setup steps when they are changed to allow for easy validation, and
# allow manual testing through the repository's "Actions" tab
on:
workflow_dispatch:
push:
paths:
- .github/workflows/copilot-setup-steps.yml
pull_request:
paths:
- .github/workflows/copilot-setup-steps.yml
jobs:
# The job MUST be called `copilot-setup-steps` or it will not be picked up by Copilot.
copilot-setup-steps:
runs-on: ubuntu-latest
# Set the permissions to the lowest permissions possible needed for your steps.
# Copilot will be given its own token for its operations.
permissions:
# If you want to clone the repository as part of your setup steps, for example to install dependencies, you'll need the `contents: read` permission. If you don't clone the repository in your setup steps, Copilot will do this for you automatically after the steps complete.
contents: read
# You can define any steps you want, and they will run before the agent starts.
# If you do not check out your code, Copilot will do this for you.
steps:
- name: Checkout code
uses: actions/checkout@v4
- name: ccache
uses: hendrikmuhs/ccache-action@v1.2.16
with:
key: copilot-setup-steps
evict-old-files: 1d
- name: Dependencies
id: depends
run: |
sudo apt-get update
sudo apt-get install build-essential libcurl4-openssl-dev
- name: Set up Python
uses: actions/setup-python@v5
with:
python-version: '3.11'
- name: Install Python dependencies
run: |
python3 -m venv .venv
.venv/bin/activate
pip install -r requirements/requirements-all.txt -r tools/server/tests/requirements.txt
pip install flake8 pyright

45
.github/workflows/pre-tokenizer-hashes.yml vendored Normal file
View File

@ -0,0 +1,45 @@
name: Check Pre-Tokenizer Hashes
on:
push:
paths:
- 'convert_hf_to_gguf.py'
- 'convert_hf_to_gguf_update.py'
pull_request:
paths:
- 'convert_hf_to_gguf.py'
- 'convert_hf_to_gguf_update.py'
jobs:
pre-tokenizer-hashes:
runs-on: ubuntu-latest
steps:
- name: Checkout repository
uses: actions/checkout@v4
- name: Set up Python
uses: actions/setup-python@v5
with:
python-version: '3.11'
- name: Install Python dependencies
run: |
python3 -m venv .venv
.venv/bin/pip install -r requirements/requirements-convert_hf_to_gguf_update.txt
- name: Update pre-tokenizer hashes
run: |
cp convert_hf_to_gguf.py /tmp
.venv/bin/python convert_hf_to_gguf_update.py --check-missing
- name: Check if committed pre-tokenizer hashes matches generated version
run: |
if ! diff -q convert_hf_to_gguf.py /tmp/convert_hf_to_gguf.py; then
echo "Model pre-tokenizer hashes (in convert_hf_to_gguf.py) do not match generated hashes (from convert_hf_to_gguf_update.py)."
echo "To fix: run ./convert_hf_to_gguf_update.py and commit the updated convert_hf_to_gguf.py along with your changes"
echo "Differences found:"
diff convert_hf_to_gguf.py /tmp/convert_hf_to_gguf.py || true
exit 1
fi
echo "Model pre-tokenizer hashes are up to date."

1
.gitignore vendored
View File

@ -82,6 +82,7 @@ models/*
models-mnt
!models/.editorconfig
!models/ggml-vocab-*.gguf*
!models/templates
# Zig
zig-out/

2
CMakeLists.txt
View File

@ -12,6 +12,8 @@ if (NOT XCODE AND NOT MSVC AND NOT CMAKE_BUILD_TYPE)
set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "MinSizeRel" "RelWithDebInfo")
endif()
message("CMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}")
# Add path to modules
list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")

5
README.md
View File

@ -17,6 +17,8 @@ LLM inference in C/C++
## Hot topics
- **[[FEEDBACK] Better packaging for llama.cpp to support downstream consumers 🤗](https://github.com/ggml-org/llama.cpp/discussions/15313)**
- Support for the `gpt-oss` model with native MXFP4 format has been added | [PR](https://github.com/ggml-org/llama.cpp/pull/15091) | [Collaboration with NVIDIA](https://blogs.nvidia.com/blog/rtx-ai-garage-openai-oss) | [Comment](https://github.com/ggml-org/llama.cpp/discussions/15095)
- Hot PRs: [All](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Apr+label%3Ahot+) | [Open](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Apr+label%3Ahot+is%3Aopen)
- Multimodal support arrived in `llama-server`: [#12898](https://github.com/ggml-org/llama.cpp/pull/12898) | [documentation](./docs/multimodal.md)
- VS Code extension for FIM completions: https://github.com/ggml-org/llama.vscode
@ -239,7 +241,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
<details>
<summary>Infrastructure</summary>
- [Paddler](https://github.com/distantmagic/paddler) - Stateful load balancer custom-tailored for llama.cpp
- [Paddler](https://github.com/intentee/paddler) - Open-source LLMOps platform for hosting and scaling AI in your own infrastructure
- [GPUStack](https://github.com/gpustack/gpustack) - Manage GPU clusters for running LLMs
- [llama_cpp_canister](https://github.com/onicai/llama_cpp_canister) - llama.cpp as a smart contract on the Internet Computer, using WebAssembly
- [llama-swap](https://github.com/mostlygeek/llama-swap) - transparent proxy that adds automatic model switching with llama-server
@ -270,7 +272,6 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
| [CANN](docs/build.md#cann) | Ascend NPU |
| [OpenCL](docs/backend/OPENCL.md) | Adreno GPU |
| [WebGPU [In Progress]](docs/build.md#webgpu) | All |
| [RPC](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) | All |
## Obtaining and quantizing models

2
ci/README.md
View File

@ -54,7 +54,7 @@ docker run --privileged -it \
-v $HOME/llama.cpp/ci-cache:/ci-cache \
-v $HOME/llama.cpp/ci-results:/ci-results \
-v $PWD:/ws -w /ws \
mthreads/musa:rc4.0.1-mudnn-devel-ubuntu22.04
mthreads/musa:rc4.2.0-devel-ubuntu22.04-amd64
```
Inside the container, execute the following commands:

273
common/arg.cpp
View File

@ -24,6 +24,7 @@
#include <cstdarg>
#include <filesystem>
#include <fstream>
#include <list>
#include <regex>
#include <set>
#include <string>
@ -748,6 +749,39 @@ std::pair<long, std::vector<char>> common_remote_get_content(const std::string &
// utils
//
// Helper function to parse tensor buffer override strings
static void parse_tensor_buffer_overrides(const std::string & value, std::vector<llama_model_tensor_buft_override> & overrides) {
std::map<std::string, ggml_backend_buffer_type_t> buft_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");
}
// keep strings alive and avoid leaking memory by storing them in a static vector
static std::list<std::string> buft_overrides;
buft_overrides.push_back(tensor_name);
overrides.push_back({buft_overrides.back().c_str(), buft_list.at(buffer_type)});
}
}
struct handle_model_result {
bool found_mmproj = false;
common_params_model mmproj;
@ -977,6 +1011,10 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
for (auto & seq_breaker : params.sampling.dry_sequence_breakers) {
string_process_escapes(seq_breaker);
}
for (auto & pair : params.speculative.replacements) {
string_process_escapes(pair.first);
string_process_escapes(pair.second);
}
}
if (!params.kv_overrides.empty()) {
@ -988,6 +1026,10 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
params.tensor_buft_overrides.push_back({nullptr, nullptr});
}
if (!params.speculative.tensor_buft_overrides.empty()) {
params.speculative.tensor_buft_overrides.push_back({nullptr, nullptr});
}
if (!params.chat_template.empty() && !common_chat_verify_template(params.chat_template, params.use_jinja)) {
throw std::runtime_error(string_format(
"error: the supplied chat template is not supported: %s%s\n",
@ -1196,6 +1238,7 @@ bool common_params_parse(int argc, char ** argv, common_params & params, llama_e
common_params_print_completion(ctx_arg);
exit(0);
}
params.lr.init();
} catch (const std::invalid_argument & ex) {
fprintf(stderr, "%s\n", ex.what());
ctx_arg.params = params_org;
@ -1464,6 +1507,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.swa_full = true;
}
).set_env("LLAMA_ARG_SWA_FULL"));
add_opt(common_arg(
{"--swa-checkpoints"}, "N",
string_format("max number of SWA checkpoints per slot to create (default: %d)\n"
"[(more info)](https://github.com/ggml-org/llama.cpp/pull/15293)", params.n_swa_checkpoints),
[](common_params & params, int value) {
params.n_swa_checkpoints = value;
}
).set_env("LLAMA_ARG_SWA_CHECKPOINTS").set_examples({LLAMA_EXAMPLE_SERVER}));
add_opt(common_arg(
{"--kv-unified", "-kvu"},
string_format("use single unified KV buffer for the KV cache of all sequences (default: %s)\n"
@ -1612,7 +1663,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
[](common_params & params, const std::string & value) {
params.antiprompt.emplace_back(value);
}
).set_examples({LLAMA_EXAMPLE_MAIN}));
).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}));
add_opt(common_arg(
{"-sp", "--special"},
string_format("special tokens output enabled (default: %s)", params.special ? "true" : "false"),
@ -2091,6 +2142,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.no_kv_offload = true;
}
).set_env("LLAMA_ARG_NO_KV_OFFLOAD"));
add_opt(common_arg(
{"-nr", "--no-repack"},
"disable weight repacking",
[](common_params & params) {
params.no_extra_bufts = true;
}
).set_env("LLAMA_ARG_NO_REPACK"));
add_opt(common_arg(
{"-ctk", "--cache-type-k"}, "TYPE",
string_format(
@ -2337,38 +2395,58 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
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)});
}
parse_tensor_buffer_overrides(value, params.tensor_buft_overrides);
}
));
add_opt(common_arg(
{"--override-tensor-draft", "-otd"}, "<tensor name pattern>=<buffer type>,...",
"override tensor buffer type for draft model", [](common_params & params, const std::string & value) {
parse_tensor_buffer_overrides(value, params.speculative.tensor_buft_overrides);
}
).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
add_opt(common_arg(
{"--cpu-moe", "-cmoe"},
"keep all Mixture of Experts (MoE) weights in the CPU",
[](common_params & params) {
params.tensor_buft_overrides.push_back({"\\.ffn_(up|down|gate)_exps", ggml_backend_cpu_buffer_type()});
}
).set_env("LLAMA_ARG_CPU_MOE"));
add_opt(common_arg(
{"--n-cpu-moe", "-ncmoe"}, "N",
"keep the Mixture of Experts (MoE) weights of the first N layers in the CPU",
[](common_params & params, int value) {
if (value < 0) {
throw std::invalid_argument("invalid value");
}
for (int i = 0; i < value; ++i) {
// keep strings alive and avoid leaking memory by storing them in a static vector
static std::list<std::string> buft_overrides;
buft_overrides.push_back(string_format("blk\\.%d\\.ffn_(up|down|gate)_exps", i));
params.tensor_buft_overrides.push_back({buft_overrides.back().c_str(), ggml_backend_cpu_buffer_type()});
}
}
).set_env("LLAMA_ARG_N_CPU_MOE"));
add_opt(common_arg(
{"--cpu-moe-draft", "-cmoed"},
"keep all Mixture of Experts (MoE) weights in the CPU for the draft model",
[](common_params & params) {
params.speculative.tensor_buft_overrides.push_back({"\\.ffn_(up|down|gate)_exps", ggml_backend_cpu_buffer_type()});
}
).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CPU_MOE_DRAFT"));
add_opt(common_arg(
{"--n-cpu-moe-draft", "-ncmoed"}, "N",
"keep the Mixture of Experts (MoE) weights of the first N layers in the CPU for the draft model",
[](common_params & params, int value) {
if (value < 0) {
throw std::invalid_argument("invalid value");
}
for (int i = 0; i < value; ++i) {
static std::list<std::string> buft_overrides_draft;
buft_overrides_draft.push_back(string_format("blk\\.%d\\.ffn_(up|down|gate)_exps", i));
params.speculative.tensor_buft_overrides.push_back({buft_overrides_draft.back().c_str(), ggml_backend_cpu_buffer_type()});
}
}
).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_N_CPU_MOE_DRAFT"));
add_opt(common_arg(
{"-ngl", "--gpu-layers", "--n-gpu-layers"}, "N",
"number of layers to store in VRAM",
@ -2619,7 +2697,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
[](common_params & params, const std::string & value) {
params.out_file = value;
}
).set_examples({LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_CVECTOR_GENERATOR, LLAMA_EXAMPLE_EXPORT_LORA, LLAMA_EXAMPLE_TTS}));
).set_examples({LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_CVECTOR_GENERATOR, LLAMA_EXAMPLE_EXPORT_LORA, LLAMA_EXAMPLE_TTS, LLAMA_EXAMPLE_FINETUNE}));
add_opt(common_arg(
{"-ofreq", "--output-frequency"}, "N",
string_format("output the imatrix every N iterations (default: %d)", params.n_out_freq),
@ -2627,6 +2705,15 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.n_out_freq = value;
}
).set_examples({LLAMA_EXAMPLE_IMATRIX}));
add_opt(common_arg(
{"--output-format"}, "{gguf,dat}",
string_format("output format for imatrix file (default: %s)", params.imat_dat > 0 ? "dat" : "gguf"),
[](common_params & params, const std::string & value) {
/**/ if (value == "gguf") { params.imat_dat = -1; }
else if (value == "dat") { params.imat_dat = 1; }
else { throw std::invalid_argument("invalid output format"); }
}
).set_examples({LLAMA_EXAMPLE_IMATRIX}));
add_opt(common_arg(
{"--save-frequency"}, "N",
string_format("save an imatrix copy every N iterations (default: %d)", params.n_save_freq),
@ -2655,6 +2742,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.i_chunk = value;
}
).set_examples({LLAMA_EXAMPLE_IMATRIX}));
add_opt(common_arg(
{"--show-statistics"},
string_format("show imatrix statistics and then exit (default: %s)", params.show_statistics ? "true" : "false"),
[](common_params & params) {
params.show_statistics = true;
}
).set_examples({LLAMA_EXAMPLE_IMATRIX}));
add_opt(common_arg(
{"--parse-special"},
string_format("prase special tokens (chat, tool, etc) (default: %s)", params.parse_special ? "true" : "false"),
@ -2895,12 +2989,9 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
"controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned; one of:\n"
"- none: leaves thoughts unparsed in `message.content`\n"
"- deepseek: puts thoughts in `message.reasoning_content` (except in streaming mode, which behaves as `none`)\n"
"(default: deepseek)",
"(default: auto)",
[](common_params & params, const std::string & value) {
/**/ if (value == "deepseek") { params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK; }
else if (value == "deepseek-legacy") { params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY; }
else if (value == "none") { params.reasoning_format = COMMON_REASONING_FORMAT_NONE; }
else { throw std::invalid_argument("invalid value"); }
params.reasoning_format = common_reasoning_format_from_name(value);
}
).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MAIN}).set_env("LLAMA_ARG_THINK"));
add_opt(common_arg(
@ -3081,7 +3172,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.speculative.cpuparams.n_threads = std::thread::hardware_concurrency();
}
}
).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
add_opt(common_arg(
{"-tbd", "--threads-batch-draft"}, "N",
"number of threads to use during batch and prompt processing (default: same as --threads-draft)",
@ -3091,7 +3182,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.speculative.cpuparams_batch.n_threads = std::thread::hardware_concurrency();
}
}
).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
add_opt(common_arg(
{"-Cd", "--cpu-mask-draft"}, "M",
"Draft model CPU affinity mask. Complements cpu-range-draft (default: same as --cpu-mask)",
@ -3242,6 +3333,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.speculative.model.path = value;
}
).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MODEL_DRAFT"));
add_opt(common_arg(
{"--spec-replace"}, "TARGET", "DRAFT",
"translate the string in TARGET into DRAFT if the draft model and main model are not compatible",
[](common_params & params, const std::string & tgt, const std::string & dft) {
params.speculative.replacements.push_back({ tgt, dft });
}
).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
add_opt(common_arg(
{"-ctkd", "--cache-type-k-draft"}, "TYPE",
string_format(
@ -3431,28 +3529,11 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
}
).set_examples({LLAMA_EXAMPLE_SERVER}));
// diffusion parameters
add_opt(common_arg(
{ "--diffusion-steps" }, "N",
string_format("number of diffusion steps (default: %d)", params.diffusion.steps),
[](common_params & params, int value) { params.diffusion.steps = value; }
).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
add_opt(common_arg(
{ "--diffusion-eps" }, "F",
string_format("epsilon for timesteps (default: %.6f)", (double) params.diffusion.eps),
[](common_params & params, const std::string & value) { params.diffusion.eps = std::stof(value); }
).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
add_opt(common_arg(
{ "--diffusion-algorithm" }, "N",
string_format("diffusion algorithm: 0=ORIGIN, 1=MASKGIT_PLUS, 2=TOPK_MARGIN, 3=ENTROPY (default: %d)",
params.diffusion.algorithm),
[](common_params & params, int value) { params.diffusion.algorithm = value; }
).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
add_opt(common_arg(
{ "--diffusion-alg-temp" }, "F",
string_format("algorithm temperature (default: %.3f)", (double) params.diffusion.alg_temp),
[](common_params & params, const std::string & value) { params.diffusion.alg_temp = std::stof(value); }
).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
add_opt(common_arg(
{ "--diffusion-visual" },
string_format("enable visual diffusion mode (show progressive generation) (default: %s)",
@ -3460,5 +3541,85 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
[](common_params & params) { params.diffusion.visual_mode = true; }
).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
add_opt(common_arg(
{ "--diffusion-eps" }, "F",
string_format("epsilon for timesteps (default: %.6f)", (double) params.diffusion.eps),
[](common_params & params, const std::string & value) { params.diffusion.eps = std::stof(value); }
).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
add_opt(common_arg(
{ "--diffusion-algorithm" }, "N",
string_format("diffusion algorithm: 0=ORIGIN, 1=ENTROPY_BASED, 2=MARGIN_BASED, 3=RANDOM, 4=LOW_CONFIDENCE (default: %d)",
params.diffusion.algorithm),
[](common_params & params, int value) { params.diffusion.algorithm = value; }
).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
add_opt(common_arg(
{ "--diffusion-alg-temp" }, "F",
string_format("dream algorithm temperature (default: %.3f)", (double) params.diffusion.alg_temp),
[](common_params & params, const std::string & value) { params.diffusion.alg_temp = std::stof(value); }
).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
add_opt(common_arg(
{ "--diffusion-block-length" }, "N",
string_format("llada block length for generation (default: %d)", params.diffusion.block_length),
[](common_params & params, int value) { params.diffusion.block_length = value; }
).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
add_opt(common_arg(
{ "--diffusion-cfg-scale" }, "F",
string_format("llada classifier-free guidance scale (default: %.3f)", (double) params.diffusion.cfg_scale),
[](common_params & params, const std::string & value) { params.diffusion.cfg_scale = std::stof(value); }
).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
add_opt(common_arg(
{ "--diffusion-add-gumbel-noise" }, "F",
string_format("add gumbel noise to the logits if temp > 0.0 (default: %s)", params.diffusion.add_gumbel_noise ? "true" : "false"),
[](common_params & params, const std::string & value) { params.diffusion.add_gumbel_noise = std::stof(value); }
).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
add_opt(
common_arg({ "-lr", "--learning-rate" }, "ALPHA",
string_format(
"adamw or sgd optimizer alpha (default: %.2g); note: sgd alpha recommended ~10x (no momentum)",
(double) params.lr.lr0),
[](common_params & params, const std::string & value) { params.lr.lr0 = std::stof(value); })
.set_examples({ LLAMA_EXAMPLE_FINETUNE }));
add_opt(
common_arg({ "-lr-min", "--learning-rate-min" }, "ALPHA",
string_format(
"(if >0) final learning rate after decay (if -decay-epochs is set, default=%.2g)",
(double) params.lr.lr_min),
[](common_params & params, const std::string & value) { params.lr.lr_min = std::stof(value); })
.set_examples({ LLAMA_EXAMPLE_FINETUNE }));
add_opt(
common_arg({ "-decay-epochs", "--learning-rate-decay-epochs" }, "ALPHA",
string_format(
"(if >0) decay learning rate to -lr-min after this many epochs (exponential decay, default=%.2g)",
(double) params.lr.decay_epochs),
[](common_params & params, const std::string & value) { params.lr.decay_epochs = std::stof(value); })
.set_examples({ LLAMA_EXAMPLE_FINETUNE }));
add_opt(common_arg(
{ "-wd", "--weight-decay" }, "WD",
string_format(
"adamw or sgd optimizer weight decay (0 is off; recommend very small e.g. 1e-9) (default: %.2g).",
(double) params.lr.wd),
[](common_params & params, const std::string & value) { params.lr.wd = std::stof(value); })
.set_examples({ LLAMA_EXAMPLE_FINETUNE }));
add_opt(common_arg({ "-val-split", "--val-split" }, "FRACTION",
string_format("fraction of data to use as validation set for training (default: %.2g).",
(double) params.val_split),
[](common_params & params, const std::string & value) { params.val_split = std::stof(value); })
.set_examples({ LLAMA_EXAMPLE_FINETUNE }));
add_opt(common_arg({ "-epochs", "--epochs" }, "N",
string_format("optimizer max # of epochs (default: %d)", params.lr.epochs),
[](common_params & params, int epochs) { params.lr.epochs = epochs; })
.set_examples({ LLAMA_EXAMPLE_FINETUNE }));
add_opt(common_arg({ "-opt", "--optimizer" }, "sgd|adamw", "adamw or sgd",
[](common_params & params, const std::string & name) {
params.optimizer = common_opt_get_optimizer(name.c_str());
if (params.optimizer == GGML_OPT_OPTIMIZER_TYPE_COUNT) {
throw std::invalid_argument("invalid --optimizer, valid options: adamw, sgd");
}
})
.set_examples({ LLAMA_EXAMPLE_FINETUNE }));
return ctx_arg;
}

10
common/chat-parser.cpp
View File

@ -55,7 +55,15 @@ bool common_chat_msg_parser::add_tool_call(const std::string & name, const std::
bool common_chat_msg_parser::add_tool_call(const json & tool_call) {
std::string name = tool_call.contains("name") ? tool_call.at("name") : "";
std::string id = tool_call.contains("id") ? tool_call.at("id") : "";
std::string arguments = tool_call.contains("arguments") ? tool_call.at("arguments") : "";
std::string arguments = "";
if (tool_call.contains("arguments")) {
if (tool_call.at("arguments").is_object()) {
arguments = tool_call.at("arguments").dump();
} else {
arguments = tool_call.at("arguments");
}
}
return add_tool_call(name, id, arguments);
}

359
common/chat.cpp
View File

@ -126,6 +126,8 @@ std::vector<common_chat_msg_diff> common_chat_msg_diff::compute_diffs(const comm
typedef minja::chat_template common_chat_template;
struct common_chat_templates {
bool add_bos;
bool add_eos;
bool has_explicit_template; // Model had builtin template or template overridde was specified.
std::unique_ptr<common_chat_template> template_default; // always set (defaults to chatml)
std::unique_ptr<common_chat_template> template_tool_use;
@ -143,6 +145,8 @@ struct templates_params {
bool enable_thinking = true;
std::chrono::system_clock::time_point now = std::chrono::system_clock::now();
json extra_context;
bool add_bos;
bool add_eos;
};
common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::string & tool_choice) {
@ -292,6 +296,7 @@ json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msg
}
if (!msg.reasoning_content.empty()) {
jmsg["reasoning_content"] = msg.reasoning_content;
jmsg["thinking"] = msg.reasoning_content; // gpt-oss
}
if (!msg.tool_name.empty()) {
jmsg["name"] = msg.tool_name;
@ -445,6 +450,8 @@ std::string common_chat_format_single(
common_chat_templates_inputs inputs;
inputs.use_jinja = use_jinja;
inputs.add_bos = tmpls->add_bos;
inputs.add_eos = tmpls->add_eos;
std::string fmt_past_msg;
if (!past_msg.empty()) {
@ -466,9 +473,12 @@ std::string common_chat_format_single(
return ss.str();
}
std::string common_chat_format_example(const struct common_chat_templates * tmpls, bool use_jinja) {
std::string common_chat_format_example(const struct common_chat_templates * tmpls, bool use_jinja, const std::map<std::string, std::string> & chat_template_kwargs) {
common_chat_templates_inputs inputs;
inputs.use_jinja = use_jinja;
inputs.add_bos = tmpls->add_bos;
inputs.add_eos = tmpls->add_eos;
inputs.chat_template_kwargs = chat_template_kwargs;
auto add_simple_msg = [&](auto role, auto content) {
common_chat_msg msg;
msg.role = role;
@ -544,8 +554,21 @@ common_chat_templates_ptr common_chat_templates_init(
default_template_src = CHATML_TEMPLATE_SRC;
}
}
// TODO @ngxson : this is a temporary hack to prevent chat template from throwing an error
// Ref: https://github.com/ggml-org/llama.cpp/pull/15230#issuecomment-3173959633
if (default_template_src.find("<|channel|>") != std::string::npos
// search for the error message and patch it
&& default_template_src.find("in message.content or") != std::string::npos) {
string_replace_all(default_template_src,
"{%- if \"<|channel|>analysis<|message|>\" in message.content or \"<|channel|>final<|message|>\" in message.content %}",
"{%- if false %}");
}
std::string token_bos = bos_token_override;
std::string token_eos = eos_token_override;
bool add_bos = false;
bool add_eos = false;
if (model) {
const auto * vocab = llama_model_get_vocab(model);
const auto get_token = [&](llama_token token, const char * name, const char * jinja_variable_name) {
@ -560,9 +583,13 @@ common_chat_templates_ptr common_chat_templates_init(
};
token_bos = get_token(llama_vocab_bos(vocab), "BOS", "bos_token");
token_eos = get_token(llama_vocab_eos(vocab), "EOS", "eos_token");
add_bos = llama_vocab_get_add_bos(vocab);
add_eos = llama_vocab_get_add_eos(vocab);
}
common_chat_templates_ptr tmpls(new common_chat_templates());
tmpls->has_explicit_template = has_explicit_template;
tmpls->add_bos = add_bos;
tmpls->add_eos = add_eos;
try {
tmpls->template_default = std::make_unique<minja::chat_template>(default_template_src, token_bos, token_eos);
} catch (const std::exception & e) {
@ -592,6 +619,8 @@ const char * common_chat_format_name(common_chat_format format) {
case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1: return "Functionary v3.1 Llama 3.1";
case COMMON_CHAT_FORMAT_HERMES_2_PRO: return "Hermes 2 Pro";
case COMMON_CHAT_FORMAT_COMMAND_R7B: return "Command R7B";
case COMMON_CHAT_FORMAT_GRANITE: return "Granite";
case COMMON_CHAT_FORMAT_GPT_OSS: return "GPT-OSS";
default:
throw std::runtime_error("Unknown chat format");
}
@ -600,13 +629,28 @@ const char * common_chat_format_name(common_chat_format format) {
const char * common_reasoning_format_name(common_reasoning_format format) {
switch (format) {
case COMMON_REASONING_FORMAT_NONE: return "none";
case COMMON_REASONING_FORMAT_AUTO: return "auto";
case COMMON_REASONING_FORMAT_DEEPSEEK: return "deepseek";
case COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY: return "deepseek-legacy";
case COMMON_REASONING_FORMAT_GRANITE: return "granite";
default:
throw std::runtime_error("Unknown reasoning format");
}
}
common_reasoning_format common_reasoning_format_from_name(const std::string & format) {
if (format == "none") {
return COMMON_REASONING_FORMAT_NONE;
} else if (format == "auto") {
return COMMON_REASONING_FORMAT_AUTO;
} else if (format == "deepseek") {
return COMMON_REASONING_FORMAT_DEEPSEEK;
} else if (format == "deepseek-legacy") {
return COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY;
}
throw std::runtime_error("Unknown reasoning format: " + format);
}
static std::string wrap_code_as_arguments(common_chat_msg_parser & builder, const std::string & code) {
std::string arguments;
if (builder.is_partial()) {
@ -748,10 +792,10 @@ static std::string apply(
// instead of using `chat_template_options.use_bos_token = false`, since these tokens
// may be needed inside the template / between messages too.
auto result = tmpl.apply(tmpl_inputs, tmpl_opts);
if (string_starts_with(result, tmpl.bos_token())) {
if (inputs.add_bos && string_starts_with(result, tmpl.bos_token())) {
result = result.substr(tmpl.bos_token().size());
}
if (string_ends_with(result, tmpl.eos_token())) {
if (inputs.add_eos && string_ends_with(result, tmpl.eos_token())) {
result = result.substr(0, result.size() - tmpl.eos_token().size());
}
return result;
@ -1289,6 +1333,174 @@ static void common_chat_parse_deepseek_r1(common_chat_msg_parser & builder) {
tool_calls_end);
}
static common_chat_params common_chat_params_init_gpt_oss(const common_chat_template & tmpl, const struct templates_params & inputs) {
common_chat_params data;
auto prompt = apply(tmpl, inputs);
data.prompt = prompt;
data.format = COMMON_CHAT_FORMAT_GPT_OSS;
// These special tokens are required to parse properly, so we include them
// even if parse_tool_calls is false.
data.preserved_tokens = {
"<|channel|>",
"<|constrain|>",
"<|message|>",
"<|start|>",
"<|end|>",
};
if (inputs.tools.is_array() && !inputs.tools.empty()) {
data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
data.grammar = build_grammar([&](const common_grammar_builder & builder) {
// tool calls can appear in commentary or analysis channels
auto channel = builder.add_rule("channel", "\"<|channel|>\" ( \"commentary\" | \"analysis\" )");
std::vector<std::string> tool_rules_recipient_in_role;
std::vector<std::string> tool_rules_recipient_in_channel;
foreach_function(inputs.tools, [&](const json & tool) {
const auto & function = tool.at("function");
std::string name = function.at("name");
auto parameters = function.at("parameters");
builder.resolve_refs(parameters);
tool_rules_recipient_in_role.push_back(
builder.add_rule(name + "-call",
"\"" + name + "\"" + channel + " \" <|constrain|>json\"? \"<|message|>\" " +
builder.add_schema(name + "-args", parameters)
)
);
tool_rules_recipient_in_channel.push_back(
builder.add_rule(name + "-call",
"\"" + name + "\"" + " \" <|constrain|>json\"? \"<|message|>\" " +
builder.add_schema(name + "-args", parameters)
)
);
});
auto recipient_in_role = builder.add_rule("recipient_in_role",
"\"<|start|>assistant\"? \" to=functions.\" ( " +
string_join(tool_rules_recipient_in_role, " | ") + " )"
);
auto recipient_in_channel = builder.add_rule("recipient_in_channel",
channel + " \" to=functions.\" ( " +
string_join(tool_rules_recipient_in_channel, " | ") + " )"
);
builder.add_rule("root", recipient_in_role + " | " + recipient_in_channel);
// Trigger on tool calls that appear in the commentary channel
data.grammar_triggers.push_back({
COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN,
"<\\|channel\\|>(commentary|analysis) to"
});
// Trigger tool calls that appear in the role section, either at the
// start or in the middle.
data.grammar_triggers.push_back({
COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_FULL,
"^ to"
});
data.grammar_triggers.push_back({
COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN,
"<\\|start\\|>assistant to"
});
});
}
return data;
}
static void common_chat_parse_gpt_oss(common_chat_msg_parser & builder) {
static const std::string constraint = "(?: (<\\|constrain\\|>)?([a-zA-Z0-9_-]+))";
static const std::string recipient("(?: to=functions\\.([^<\\s]+))");
static const common_regex start_regex("<\\|start\\|>assistant");
static const common_regex analysis_regex("<\\|channel\\|>analysis");
static const common_regex final_regex("<\\|channel\\|>final" + constraint + "?");
static const common_regex preamble_regex("<\\|channel\\|>commentary");
static const common_regex tool_call1_regex(recipient + "<\\|channel\\|>(analysis|commentary)" + constraint + "?");
static const common_regex tool_call2_regex("<\\|channel\\|>(analysis|commentary)" + recipient + constraint + "?");
auto consume_end = [&](bool include_end = false) {
if (auto res = builder.try_find_literal("<|end|>")) {
return res->prelude + (include_end ? builder.str(res->groups[0]) : "");
}
return builder.consume_rest();
};
auto handle_tool_call = [&](const std::string & name) {
if (auto args = builder.try_consume_json_with_dumped_args({{}})) {
if (builder.syntax().parse_tool_calls) {
if (!builder.add_tool_call(name, "", args->value) || args->is_partial) {
throw common_chat_msg_partial_exception("incomplete tool call");
}
} else if (args->is_partial) {
throw common_chat_msg_partial_exception("incomplete tool call");
}
}
};
auto regex_match = [](const common_regex & regex, const std::string & input) -> std::optional<common_regex_match> {
auto match = regex.search(input, 0, true);
if (match.type == COMMON_REGEX_MATCH_TYPE_FULL) {
return match;
}
return std::nullopt;
};
do {
auto header_start_pos = builder.pos();
auto content_start = builder.try_find_literal("<|message|>");
if (!content_start) {
throw common_chat_msg_partial_exception("incomplete header");
}
auto header = content_start->prelude;
if (auto match = regex_match(tool_call1_regex, header)) {
auto group = match->groups[1];
auto name = header.substr(group.begin, group.end - group.begin);
handle_tool_call(name);
continue;
}
if (auto match = regex_match(tool_call2_regex, header)) {
auto group = match->groups[2];
auto name = header.substr(group.begin, group.end - group.begin);
handle_tool_call(name);
continue;
}
if (regex_match(analysis_regex, header)) {
builder.move_to(header_start_pos);
if (builder.syntax().reasoning_format == COMMON_REASONING_FORMAT_NONE || builder.syntax().reasoning_in_content) {
builder.add_content(consume_end(true));
} else {
builder.try_parse_reasoning("<|channel|>analysis<|message|>", "<|end|>");
}
continue;
}
if(regex_match(final_regex, header) || regex_match(preamble_regex, header)) {
builder.add_content(consume_end());
continue;
}
// Possibly a malformed message, attempt to recover by rolling
// back to pick up the next <|start|>
LOG_DBG("%s: unknown header from message: %s\n", __func__, header.c_str());
builder.move_to(header_start_pos);
} while (builder.try_find_regex(start_regex, std::string::npos, false));
auto remaining = builder.consume_rest();
if (!remaining.empty()) {
LOG_DBG("%s: content after last message: %s\n", __func__, remaining.c_str());
}
}
static common_chat_params common_chat_params_init_firefunction_v2(const common_chat_template & tmpl, const struct templates_params & inputs) {
LOG_DBG("%s\n", __func__);
common_chat_params data;
@ -1646,7 +1858,7 @@ static void common_chat_parse_hermes_2_pro(common_chat_msg_parser & builder) {
"|<function name=\"([^\"]+)\">" // match 5 (function name again)
);
if (auto res = builder.try_find_regex(open_regex)) {
while (auto res = builder.try_find_regex(open_regex)) {
const auto & block_start = res->groups[1];
std::string block_end = block_start.empty() ? "" : "```";
@ -1668,7 +1880,6 @@ static void common_chat_parse_hermes_2_pro(common_chat_msg_parser & builder) {
builder.consume_literal(block_end);
builder.consume_spaces();
}
builder.add_content(builder.consume_rest());
} else {
throw common_chat_msg_partial_exception("failed to parse tool call");
}
@ -1693,8 +1904,125 @@ static void common_chat_parse_hermes_2_pro(common_chat_msg_parser & builder) {
builder.consume_spaces();
}
}
}
}
builder.add_content(builder.consume_rest());
}
static common_chat_params common_chat_params_init_granite(const common_chat_template & tmpl, const struct templates_params & inputs) {
common_chat_params data;
// Pass thinking context for Granite template
json additional_context = {
{"thinking", inputs.enable_thinking},
};
data.prompt = apply(tmpl, inputs, /* messages_override= */ std::nullopt, /* tools_override= */ std::nullopt, additional_context);
data.format = COMMON_CHAT_FORMAT_GRANITE;
if (string_ends_with(data.prompt, "<think>\n") || string_ends_with(data.prompt, "<think>")) {
if (!inputs.enable_thinking) {
data.prompt += "</think>";
} else {
data.thinking_forced_open = true;
}
}
if (!inputs.tools.is_null()) {
// Granite uses <|tool_call|> followed by JSON list
data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
data.grammar = build_grammar([&](const common_grammar_builder & builder) {
std::vector<std::string> tool_rules;
foreach_function(inputs.tools, [&](const json & tool) {
const auto & function = tool.at("function");
std::string name = function.at("name");
auto parameters = function.at("parameters");
builder.resolve_refs(parameters);
tool_rules.push_back(builder.add_rule(name + "-call", builder.add_schema(name +
"-args", {
{"type", "object"},
{"properties", {
{"name", {{"const", name}}},
{"arguments", parameters},
}},
{"required", json::array({"name", "arguments"})},
})));
});
auto tool_call = builder.add_rule("tool_call", string_join(tool_rules, " | "));
auto tool_list = builder.add_rule("tool_list", "\"[\" space " + tool_call + " (\",\" space " + tool_call + ")* space \"]\"");
if (data.thinking_forced_open) {
builder.add_rule("root", "\"</think>\" space \"<response>\" space [^<]* \"</response>\" space \"<|tool_call|>\" space " + tool_list);
} else {
builder.add_rule("root", "\"<|tool_call|>\" space " + tool_list);
}
data.grammar_triggers.push_back({
COMMON_GRAMMAR_TRIGGER_TYPE_WORD,
"<|tool_call|>"
});
data.preserved_tokens = {
"<think>",
"</think>",
"<response>",
"</response>",
"<|tool_call|>",
};
});
} else {
// Handle thinking tags for non-tool responses
if (data.thinking_forced_open && inputs.enable_thinking) {
data.grammar_lazy = false;
data.grammar = build_grammar([&](const common_grammar_builder & builder) {
builder.add_rule("root", "\"</think>\" space \"<response>\" space .* \"</response>\" space");
});
data.preserved_tokens = {
"<think>",
"</think>",
"<response>",
"</response>",
};
}
}
return data;
}
static void common_chat_parse_granite(common_chat_msg_parser & builder) {
// Parse thinking tags
builder.try_parse_reasoning("<think>", "</think>");
// Parse response tags using regex
static const common_regex response_regex("<response>([\\s\\S]*?)</response>");
if (auto res = builder.try_find_regex(response_regex)) {
// Extract the content between the tags (capture group 1)
auto content = builder.str(res->groups[1]);
builder.add_content(content);
builder.move_to(res->groups[0].end);
}
if (!builder.syntax().parse_tool_calls) {
builder.add_content(builder.consume_rest());
return;
}
// Look for tool calls
static const common_regex tool_call_regex(regex_escape("<|tool_call|>"));
if (auto res = builder.try_find_regex(tool_call_regex)) {
builder.move_to(res->groups[0].end);
// Expect JSON array of tool calls
auto tool_calls_data = builder.consume_json();
if (tool_calls_data.json.is_array()) {
if (!builder.add_tool_calls(tool_calls_data.json)) {
builder.add_content("<|tool_call|>" + tool_calls_data.json.dump());
}
} else {
builder.add_content("<|tool_call|>" + tool_calls_data.json.dump());
}
} else {
builder.add_content(builder.consume_rest());
}
@ -1733,6 +2061,8 @@ static common_chat_params common_chat_templates_apply_jinja(
params.enable_thinking = inputs.enable_thinking;
params.grammar = inputs.grammar;
params.now = inputs.now;
params.add_bos = inputs.add_bos;
params.add_eos = inputs.add_eos;
params.extra_context = json::object();
for (auto el : inputs.chat_template_kwargs) {
@ -1769,11 +2099,21 @@ static common_chat_params common_chat_templates_apply_jinja(
return common_chat_params_init_command_r7b(tmpl, params);
}
// Granite (IBM) - detects thinking / tools support
if (src.find("elif thinking") != std::string::npos && src.find("<|tool_call|>") != std::string::npos) {
return common_chat_params_init_granite(tmpl, params);
}
// Hermes 2/3 Pro, Qwen 2.5 Instruct (w/ tools)
if (src.find("<tool_call>") != std::string::npos && params.json_schema.is_null()) {
return common_chat_params_init_hermes_2_pro(tmpl, params);
}
// GPT-OSS
if (src.find("<|channel|>") != std::string::npos && params.json_schema.is_null()) {
return common_chat_params_init_gpt_oss(tmpl, params);
}
// Use generic handler when mixing tools + JSON schema.
// TODO: support that mix in handlers below.
if ((params.tools.is_array() && params.json_schema.is_object())) {
@ -1824,6 +2164,7 @@ static common_chat_params common_chat_templates_apply_legacy(
int alloc_size = 0;
std::vector<llama_chat_message> chat;
std::vector<std::string> contents;
for (const auto & msg : inputs.messages) {
auto content = msg.content;
for (const auto & part : msg.content_parts) {
@ -1925,6 +2266,12 @@ static void common_chat_parse(common_chat_msg_parser & builder) {
case COMMON_CHAT_FORMAT_COMMAND_R7B:
common_chat_parse_command_r7b(builder);
break;
case COMMON_CHAT_FORMAT_GRANITE:
common_chat_parse_granite(builder);
break;
case COMMON_CHAT_FORMAT_GPT_OSS:
common_chat_parse_gpt_oss(builder);
break;
default:
throw std::runtime_error(std::string("Unsupported format: ") + common_chat_format_name(builder.syntax().format));
}
@ -1944,6 +2291,8 @@ common_chat_msg common_chat_parse(const std::string & input, bool is_partial, co
}
}
auto msg = builder.result();
if (!is_partial) {
LOG_DBG("Parsed message: %s\n", common_chat_msgs_to_json_oaicompat<json>({msg}).at(0).dump().c_str());
}
return msg;
}

8
common/chat.h
View File

@ -109,6 +109,8 @@ enum common_chat_format {
COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1,
COMMON_CHAT_FORMAT_HERMES_2_PRO,
COMMON_CHAT_FORMAT_COMMAND_R7B,
COMMON_CHAT_FORMAT_GRANITE,
COMMON_CHAT_FORMAT_GPT_OSS,
COMMON_CHAT_FORMAT_COUNT, // Not a format, just the # formats
};
@ -127,6 +129,8 @@ struct common_chat_templates_inputs {
bool enable_thinking = true;
std::chrono::system_clock::time_point now = std::chrono::system_clock::now();
std::map<std::string, std::string> chat_template_kwargs;
bool add_bos = false;
bool add_eos = false;
};
struct common_chat_params {
@ -183,10 +187,12 @@ std::string common_chat_format_single(
// Returns an example of formatted chat
std::string common_chat_format_example(
const struct common_chat_templates * tmpls,
bool use_jinja);
bool use_jinja,
const std::map<std::string, std::string> & chat_template_kwargs);
const char* common_chat_format_name(common_chat_format format);
const char* common_reasoning_format_name(common_reasoning_format format);
common_reasoning_format common_reasoning_format_from_name(const std::string & format);
common_chat_msg common_chat_parse(const std::string & input, bool is_partial, const common_chat_syntax & syntax);
common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::string & tool_choice);

55
common/common.cpp
View File

@ -41,6 +41,7 @@
#endif
#include <locale>
#include <windows.h>
#include <string.h>
#include <fcntl.h>
#include <io.h>
#else
@ -1122,6 +1123,7 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
mparams.use_mmap = params.use_mmap;
mparams.use_mlock = params.use_mlock;
mparams.check_tensors = params.check_tensors;
mparams.use_extra_bufts = !params.no_extra_bufts;
if (params.kv_overrides.empty()) {
mparams.kv_overrides = NULL;
@ -1564,3 +1566,56 @@ ggml_opt_dataset_t common_opt_dataset_init(struct llama_context * ctx, const std
return result;
}
ggml_opt_optimizer_params common_opt_lr_pars(void * userdata) {
ggml_opt_optimizer_params result = ggml_opt_get_default_optimizer_params(nullptr);
const lr_opt & d = *(lr_opt *) userdata;
result.adamw.alpha = result.sgd.alpha = d.get_lr(d.epoch);
result.sgd.wd = result.adamw.wd = d.wd;
return result;
}
// TODO make all command line args case-insensitive
static inline bool eq_case_insensitive(char const* a, char const* b) {
return !
#if defined(_MSC_VER)
_stricmp
#else
strcasecmp
#endif // defined(_MSC_VER)
(a, b);
}
enum ggml_opt_optimizer_type common_opt_get_optimizer(const char * n) {
if (eq_case_insensitive("adamw", n)) {
return GGML_OPT_OPTIMIZER_TYPE_ADAMW;
}
if (eq_case_insensitive("sgd", n)) {
return GGML_OPT_OPTIMIZER_TYPE_SGD;
}
return GGML_OPT_OPTIMIZER_TYPE_COUNT;
}
// TODO simplify to use just log and exp
static float const k_log_2 = std::log(2.f);
void lr_opt::init() {
if (lr_min > 0 && lr_min < lr0) {
float nhalf = std::log(lr0 / lr_min) / k_log_2;
float e = epochs;
if (decay_epochs > 0 && decay_epochs < e) {
e = decay_epochs;
} else {
decay_epochs = e;
}
scale_epoch = nhalf / e;
}
}
float lr_opt::get_lr(float epoch) const {
float r = lr_min <= 0 ? lr0 :
epoch >= decay_epochs ? lr_min :
lr0 * std::pow(0.5f, epoch * scale_epoch);
LOG_INF("epoch %.2g lr=%.2g\n", epoch, r);
return r;
}

59
common/common.h
View File

@ -2,14 +2,17 @@
#pragma once
#include "llama-cpp.h"
#include <set>
#include <sstream>
#include <string>
#include <string_view>
#include <vector>
#include <map>
#include <sstream>
#include <cmath>
#include "ggml-opt.h"
#include "llama-cpp.h"
#ifdef _WIN32
#define DIRECTORY_SEPARATOR '\\'
@ -82,6 +85,7 @@ enum llama_example {
LLAMA_EXAMPLE_PARALLEL,
LLAMA_EXAMPLE_TTS,
LLAMA_EXAMPLE_DIFFUSION,
LLAMA_EXAMPLE_FINETUNE,
LLAMA_EXAMPLE_COUNT,
};
@ -201,6 +205,8 @@ struct common_params_speculative {
int32_t n_gpu_layers = -1; // number of layers to store in VRAM for the draft model (-1 - use default)
float p_split = 0.1f; // speculative decoding split probability
float p_min = 0.75f; // minimum speculative decoding probability (greedy)
std::vector<std::pair<std::string, std::string>> replacements; // main to speculative model replacements
std::vector<llama_model_tensor_buft_override> tensor_buft_overrides;
ggml_type cache_type_k = GGML_TYPE_F16; // KV cache data type for the K
ggml_type cache_type_v = GGML_TYPE_F16; // KV cache data type for the V
@ -220,19 +226,46 @@ struct common_params_vocoder {
};
struct common_params_diffusion {
int32_t steps = 64; // number of diffusion steps
float eps = 1e-3f; // epsilon for timesteps
int32_t algorithm = 0; // diffusion algorithm (0=ORIGIN, 1=MASKGIT_PLUS, 2=TOPK_MARGIN, 3=ENTROPY)
int32_t steps = 128;
bool visual_mode = false;
float eps = 0; // epsilon for timesteps
int32_t block_length = 0; // block length for generation
int32_t algorithm = 4; // default algorithm: low-confidence
float alg_temp = 0.0f; // algorithm temperature
bool visual_mode = false; // show progressive diffusion on screen
float cfg_scale = 0; // classifier-free guidance scale
bool add_gumbel_noise = false; // add gumbel noise to the logits if temp > 0.0
};
enum common_reasoning_format {
COMMON_REASONING_FORMAT_NONE,
COMMON_REASONING_FORMAT_AUTO,
COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY, // Extract thinking tag contents and return as `message.reasoning_content`, or leave inline in <think> tags in stream mode
COMMON_REASONING_FORMAT_DEEPSEEK, // Extract thinking tag contents and return as `message.reasoning_content`, including in streaming deltas.
COMMON_REASONING_FORMAT_GRANITE, // Extract thinking tag contents and return as `message.reasoning_content`, including in streaming deltas.
};
struct lr_opt {
float lr0 = 1e-5; // learning rate at first epoch
float lr_min = -1;
float decay_epochs = -1; // if >0, the learning rate starts at lr0 and decays to lr_min after this many epochs
float scale_epoch = 0;
float wd = 0;
unsigned epochs = 2;
unsigned epoch; // set by optimizer outer (epochs) loop
// learning rate decay - constant LR per epoch only for now
float get_lr(float e) const;
float get_lr() const { return get_lr(epoch); }
// must call after arg parse, before get_lr
void init();
};
struct ggml_opt_optimizer_params common_opt_lr_pars(void * userdata);
struct common_params {
int32_t n_predict = -1; // new tokens to predict
int32_t n_ctx = 4096; // context size
@ -352,6 +385,7 @@ struct common_params {
bool warmup = true; // warmup run
bool check_tensors = false; // validate tensor data
bool no_op_offload = false; // globally disable offload host tensor operations to device
bool no_extra_bufts = false; // disable extra buffer types (used for weight repacking)
bool single_turn = false; // single turn chat conversation
@ -366,6 +400,11 @@ struct common_params {
bool no_mmproj = false; // explicitly disable multimodal model
std::vector<std::string> image; // path to image file(s)
// finetune
struct lr_opt lr;
enum ggml_opt_optimizer_type optimizer = GGML_OPT_OPTIMIZER_TYPE_ADAMW;
float val_split = 0.05f; // fraction of the data used for the validation set
// embedding
bool embedding = false; // get only sentence embedding
int32_t embd_normalize = 2; // normalisation for embeddings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
@ -379,6 +418,7 @@ struct common_params {
int32_t timeout_write = timeout_read; // http write timeout in seconds
int32_t n_threads_http = -1; // number of threads to process HTTP requests (TODO: support threadpool)
int32_t n_cache_reuse = 0; // min chunk size to reuse from the cache via KV shifting
int32_t n_swa_checkpoints = 3; // max number of SWA checkpoints per slot
std::string hostname = "127.0.0.1";
std::string public_path = ""; // NOLINT
@ -386,7 +426,7 @@ struct common_params {
std::string chat_template = ""; // NOLINT
bool use_jinja = false; // NOLINT
bool enable_chat_template = true;
common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_AUTO;
int reasoning_budget = -1;
bool prefill_assistant = true; // if true, any trailing assistant message will be prefilled into the response
@ -431,9 +471,11 @@ struct common_params {
int32_t n_out_freq = 10; // output the imatrix every n_out_freq iterations
int32_t n_save_freq = 0; // save the imatrix every n_save_freq iterations
int32_t i_chunk = 0; // start processing from this chunk
int8_t imat_dat = 0; // whether the legacy imatrix.dat format should be output (gguf <= 0 < dat)
bool process_output = false; // collect data for the output tensor
bool compute_ppl = true; // whether to compute perplexity
bool show_statistics = false; // show imatrix statistics per tensor
bool parse_special = false; // whether to parse special tokens during imatrix tokenization
// cvector-generator params
@ -691,3 +733,6 @@ const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
//
ggml_opt_dataset_t common_opt_dataset_init(struct llama_context * ctx, const std::vector<llama_token> & tokens, int64_t stride);
// "adamw" or "sgd" (case insensitive)
enum ggml_opt_optimizer_type common_opt_get_optimizer(const char *);

181
common/speculative.cpp
View File

@ -1,30 +1,39 @@
#include "speculative.h"
#include "ggml.h"
#include "llama.h"
#include "log.h"
#include "common.h"
#include "sampling.h"
#include <cstring>
#include <algorithm>
#include <map>
#define SPEC_VOCAB_MAX_SIZE_DIFFERENCE 128
#define SPEC_VOCAB_CHECK_START_TOKEN_ID 5
struct common_speculative {
struct llama_context * ctx;
struct llama_context * ctx_tgt; // only used for retokenizing from ctx_dft
struct llama_context * ctx_dft;
struct common_sampler * smpl;
llama_batch batch;
llama_tokens prompt;
llama_tokens prompt_dft;
bool vocab_dft_compatible = true; // whether retokenization is needed
std::map<std::string, std::string> tgt_dft_replacements = {};
};
struct common_speculative * common_speculative_init(
struct llama_context * ctx_tgt,
struct llama_context * ctx_dft) {
auto * result = new common_speculative {
/* .ctx = */ ctx_dft,
/* .ctx_tgt = */ ctx_tgt,
/* .ctx_dft = */ ctx_dft,
/* .smpl = */ nullptr,
/* .batch = */ llama_batch_init(llama_n_batch(ctx_dft), 0, 1),
/* .prompt = */ {},
/* .prompt_dft = */ {},
/* .vocab_dft_compatible = */ false,
};
// TODO: optimize or pass from outside?
@ -59,6 +68,9 @@ struct common_speculative * common_speculative_init(
}
#endif
result->vocab_dft_compatible = common_speculative_are_compatible(ctx_tgt, ctx_dft);
LOG_DBG("vocab_dft_compatible = %d\n", result->vocab_dft_compatible);
return result;
}
@ -90,31 +102,32 @@ bool common_speculative_are_compatible(
LOG_DBG("%s: vocab_type dft: %d\n", __func__, vocab_type_dft);
if (vocab_type_tgt != vocab_type_dft) {
LOG_ERR("%s: draft model vocab type must match target model to use speculation but "
"vocab_type_dft = %d while vocab_type_tgt = %d\n", __func__, vocab_type_dft, vocab_type_tgt);
LOG_DBG("%s: draft model vocab type must match target model to use speculation but ", __func__);
LOG_DBG("vocab_type_dft = %d while vocab_type_tgt = %d\n", vocab_type_dft, vocab_type_tgt);
return false;
}
if (llama_vocab_get_add_bos(vocab_tgt) != llama_vocab_get_add_bos(vocab_dft) ||
if (
llama_vocab_get_add_bos(vocab_tgt) != llama_vocab_get_add_bos(vocab_dft) ||
llama_vocab_get_add_eos(vocab_tgt) != llama_vocab_get_add_eos(vocab_dft) ||
llama_vocab_bos(vocab_tgt) != llama_vocab_bos(vocab_dft) ||
llama_vocab_eos(vocab_tgt) != llama_vocab_eos(vocab_dft)) {
LOG_ERR("%s: draft vocab special tokens must match target vocab to use speculation\n", __func__);
LOG_ERR("%s: tgt: bos = %d (%d), eos = %d (%d)\n", __func__, llama_vocab_bos(vocab_tgt), llama_vocab_get_add_bos(vocab_tgt), llama_vocab_eos(vocab_tgt), llama_vocab_get_add_eos(vocab_tgt));
LOG_ERR("%s: dft: bos = %d (%d), eos = %d (%d)\n", __func__, llama_vocab_bos(vocab_dft), llama_vocab_get_add_bos(vocab_dft), llama_vocab_eos(vocab_dft), llama_vocab_get_add_eos(vocab_dft));
llama_vocab_eos(vocab_tgt) != llama_vocab_eos(vocab_dft)
) {
LOG_DBG("%s: draft model special tokens must match target model to use speculation\n", __func__);
return false;
}
{
const int n_vocab_tgt = llama_vocab_n_tokens(vocab_tgt);
const int n_vocab_dft = llama_vocab_n_tokens(vocab_dft);
const int vocab_diff = std::abs(n_vocab_tgt - n_vocab_dft);
const int vocab_diff = n_vocab_tgt > n_vocab_dft
? n_vocab_tgt - n_vocab_dft
: n_vocab_dft - n_vocab_tgt;
if (vocab_diff > SPEC_VOCAB_MAX_SIZE_DIFFERENCE) {
LOG_ERR("%s: draft model vocab must closely match target model to use speculation but "
"target vocab size %d does not match draft vocab size %d - difference %d, max allowed %d\n",
__func__, n_vocab_tgt, llama_vocab_n_tokens(vocab_dft), vocab_diff, SPEC_VOCAB_MAX_SIZE_DIFFERENCE);
LOG_DBG("%s: draft model vocab must closely match target model to use speculation but ", __func__);
LOG_DBG("target vocab size %d does not match draft vocab size %d - difference %d, max allowed %d\n",
n_vocab_tgt, llama_vocab_n_tokens(vocab_dft), vocab_diff, SPEC_VOCAB_MAX_SIZE_DIFFERENCE);
return false;
}
@ -122,8 +135,8 @@ bool common_speculative_are_compatible(
const char * token_text_tgt = llama_vocab_get_text(vocab_tgt, i);
const char * token_text_dft = llama_vocab_get_text(vocab_dft, i);
if (std::strcmp(token_text_tgt, token_text_dft) != 0) {
LOG_ERR("%s: draft vocab vocab must match target vocab to use speculation but "
"token %d content differs - target '%s', draft '%s'\n", __func__, i,
LOG_DBG("%s: draft model vocab must match target model to use speculation but ", __func__);
LOG_DBG("token %d content differs - target '%s', draft '%s'\n", i,
common_token_to_piece(ctx_tgt, i).c_str(),
common_token_to_piece(ctx_dft, i).c_str());
return false;
@ -134,32 +147,93 @@ bool common_speculative_are_compatible(
return true;
}
void common_speculative_add_replacement_tgt_dft(
struct common_speculative * spec,
const char *source, const char *dest) {
spec->tgt_dft_replacements[source] = dest;
}
static std::string replace_to_dft(
struct common_speculative * spec,
const std::string& input) {
std::string result = input;
for (const auto & pair : spec->tgt_dft_replacements) {
size_t pos = result.find(pair.first);
while (pos != std::string::npos) {
result.replace(pos, pair.first.length(), pair.second);
pos = result.find(pair.first, pos + pair.second.length());
}
}
return result;
}
static std::string replace_to_tgt(
struct common_speculative * spec,
const std::string& input) {
std::string result = input;
for (const auto& pair : spec->tgt_dft_replacements) {
size_t pos = result.find(pair.second);
while (pos != std::string::npos) {
result.replace(pos, pair.second.length(), pair.first);
pos = result.find(pair.second, pos + pair.first.length());
}
}
return result;
}
llama_tokens common_speculative_gen_draft(
struct common_speculative * spec,
struct common_speculative_params params,
const llama_tokens & prompt_tgt,
const llama_tokens & prompt_tgt_main_model, // specified in target model vocab
llama_token id_last) {
auto & batch = spec->batch;
auto & ctx = spec->ctx;
auto & ctx_tgt = spec->ctx_tgt;
auto & ctx_dft = spec->ctx_dft;
auto & smpl = spec->smpl;
auto & prompt = spec->prompt;
auto & prompt_dft = spec->prompt_dft;
auto * mem = llama_get_memory(ctx);
auto * mem_dft = llama_get_memory(ctx_dft);
int reuse_i = 0;
int reuse_n = 0;
const int n_ctx = llama_n_ctx(ctx) - params.n_draft;
const int n_ctx = llama_n_ctx(ctx_dft) - params.n_draft;
llama_tokens prompt_tgt_draft_model;
if (!spec->vocab_dft_compatible) {
std::string text;
text = common_detokenize(ctx_tgt, prompt_tgt_main_model, true);
text = replace_to_dft(spec, text);
LOG_DBG("%s: main->draft detokenized string: '%s'\n", __func__, text.c_str());
prompt_tgt_draft_model = common_tokenize(ctx_dft, text, false, true);
// convert id_last to draft vocab. llama_detokenize is called directly to avoid an allocation
const auto * model_tgt = llama_get_model(ctx_tgt);
const auto * vocab_tgt = llama_model_get_vocab(model_tgt);
int32_t n_chars = llama_detokenize(vocab_tgt, &id_last, 1, nullptr, 0, false, false);
GGML_ASSERT(n_chars < 0 && "failed to detokenize id_last");
text.resize(-n_chars);
llama_detokenize(vocab_tgt, &id_last, 1, text.data(), text.size(), false, false);
text = replace_to_dft(spec, text);
LOG_DBG("main->draft detokenized id_last(%d): '%s'\n", id_last, text.c_str());
id_last = common_tokenize(ctx_dft, text, false, true)[0];
}
// prompt_tgt's tokens will always be compatible with ctx_dft
const llama_tokens &prompt_tgt =
spec->vocab_dft_compatible ? prompt_tgt_main_model : prompt_tgt_draft_model;
const int i_start = std::max<int>(0, (int) prompt_tgt.size() - n_ctx);
// reuse as much as possible from the old draft context
// ideally, the draft context should be as big as the target context and we will always reuse the entire prompt
for (int i = 0; i < (int) prompt.size(); ++i) {
for (int i = 0; i < (int) prompt_dft.size(); ++i) {
int cur = 0;
while (i_start + cur < (int) prompt_tgt.size() &&
i + cur < (int) prompt.size() &&
prompt_tgt[i_start + cur] == prompt[i + cur]) {
i + cur < (int) prompt_dft.size() &&
prompt_tgt[i_start + cur] == prompt_dft[i + cur]) {
cur++;
}
@ -169,21 +243,20 @@ llama_tokens common_speculative_gen_draft(
}
}
LOG_DBG("%s: reuse_i = %d, reuse_n = %d, prompt = %d\n", __func__, reuse_i, reuse_n, (int) prompt.size());
LOG_DBG("%s: reuse_i = %d, reuse_n = %d, prompt = %d\n", __func__, reuse_i, reuse_n, (int) prompt_dft.size());
llama_tokens result;
result.reserve(params.n_draft);
if (reuse_n == 0) {
llama_memory_clear(mem, false);
prompt.clear();
llama_memory_clear(mem_dft, false);
prompt_dft.clear();
} else {
// this happens when a previous draft has been discarded (for example, due to being too small), but the
// target model agreed with it. in this case, we simply pass back the previous results to save compute
if (reuse_i + reuse_n < (int) prompt.size() && prompt[reuse_i + reuse_n] == id_last) {
for (int i = reuse_i + reuse_n + 1; i < (int) prompt.size(); ++i) {
result.push_back(prompt[i]);
if (reuse_i + reuse_n < (int) prompt_dft.size() && prompt_dft[reuse_i + reuse_n] == id_last) {
for (int i = reuse_i + reuse_n + 1; i < (int) prompt_dft.size(); ++i) {
result.push_back(prompt_dft[i]);
if (params.n_draft <= (int) result.size()) {
break;
@ -194,16 +267,15 @@ llama_tokens common_speculative_gen_draft(
}
if (reuse_i > 0) {
llama_memory_seq_rm (mem, 0, 0, reuse_i);
llama_memory_seq_add(mem, 0, reuse_i, -1, -reuse_i);
llama_memory_seq_rm (mem_dft, 0, 0, reuse_i);
llama_memory_seq_add(mem_dft, 0, reuse_i, -1, -reuse_i);
prompt.erase(prompt.begin(), prompt.begin() + reuse_i);
prompt_dft.erase(prompt_dft.begin(), prompt_dft.begin() + reuse_i);
}
if (reuse_n < (int) prompt.size()) {
llama_memory_seq_rm (mem, 0, reuse_n, -1);
prompt.erase(prompt.begin() + reuse_n, prompt.end());
if (reuse_n < (int) prompt_dft.size()) {
llama_memory_seq_rm (mem_dft, 0, reuse_n, -1);
prompt_dft.erase(prompt_dft.begin() + reuse_n, prompt_dft.end());
}
}
@ -214,28 +286,28 @@ llama_tokens common_speculative_gen_draft(
//LOG_DBG("i = %d, i_start = %d, reuse_n = %d, i - i_start = %d, id = %6d\n", i, i_start, reuse_n, i - i_start, prompt_tgt[i]);
common_batch_add(batch, prompt_tgt[i], i - i_start, { 0 }, false);
prompt.push_back(prompt_tgt[i]);
prompt_dft.push_back(prompt_tgt[i]);
}
// we should rarely end-up here during normal decoding
if (batch.n_tokens > 0) {
//LOG_DBG("%s: draft prompt batch: %s\n", __func__, string_from(ctx, batch).c_str());
llama_decode(ctx, batch);
llama_decode(ctx_dft, batch);
}
const llama_pos n_past = prompt.size();
const llama_pos n_past = prompt_dft.size();
LOG_DBG("%s: n_past = %d\n", __func__, n_past);
common_batch_clear(batch);
common_batch_add (batch, id_last, n_past, { 0 }, true);
prompt.push_back(id_last);
prompt_dft.push_back(id_last);
//LOG_DBG("%s: draft prompt: %s\n", __func__, string_from(ctx, prompt).c_str());
LOG_DBG("%s: draft prompt: %s\n", __func__, string_from(ctx_dft, prompt_dft).c_str());
llama_decode(ctx, batch);
llama_decode(ctx_dft, batch);
common_sampler_reset(smpl);
@ -243,13 +315,13 @@ llama_tokens common_speculative_gen_draft(
for (int i = 0; i < params.n_draft; ++i) {
common_batch_clear(batch);
common_sampler_sample(smpl, ctx, 0, true);
common_sampler_sample(smpl, ctx_dft, 0, true);
const auto * cur_p = common_sampler_get_candidates(smpl);
for (int k = 0; k < std::min(3, (int) cur_p->size); ++k) {
LOG_DBG(" - draft candidate %3d, pos %3d: %6d (%8.3f) '%s'\n",
k, i, cur_p->data[k].id, cur_p->data[k].p, common_token_to_piece(ctx, cur_p->data[k].id).c_str());
k, i, cur_p->data[k].id, cur_p->data[k].p, common_token_to_piece(ctx_dft, cur_p->data[k].id).c_str());
}
// add drafted token for each sequence
@ -271,10 +343,19 @@ llama_tokens common_speculative_gen_draft(
common_batch_add(batch, id, n_past + i + 1, { 0 }, true);
// evaluate the drafted tokens on the draft model
llama_decode(ctx, batch);
llama_decode(ctx_dft, batch);
prompt.push_back(id);
prompt_dft.push_back(id);
}
if (!spec->vocab_dft_compatible) {
std::string detokenized = common_detokenize(ctx_dft, result, true);
detokenized = replace_to_tgt(spec, detokenized);
LOG_DBG("draft->main detokenized string: '%s'\n", detokenized.c_str());
result = common_tokenize(ctx_tgt, detokenized, false, true);
if (result.size() > (size_t)params.n_draft) {
result.resize(params.n_draft);
}
}
return result;
}

9
common/speculative.h
View File

@ -12,7 +12,10 @@ struct common_speculative_params {
float p_min = 0.75f; // min probability required to accept a token in the draft
};
struct common_speculative * common_speculative_init(struct llama_context * ctx_dft);
struct common_speculative * common_speculative_init(
struct llama_context * ctx_tgt,
struct llama_context * ctx_dft
);
void common_speculative_free(struct common_speculative * spec);
@ -20,6 +23,10 @@ bool common_speculative_are_compatible(
const struct llama_context * ctx_tgt,
const struct llama_context * ctx_dft);
void common_speculative_add_replacement_tgt_dft(
struct common_speculative * spec,
const char *source, const char *dest);
// sample up to n_draft tokens and add them to the batch using the draft model
llama_tokens common_speculative_gen_draft(
struct common_speculative * spec,

959
convert_hf_to_gguf.py
View File

File diff suppressed because it is too large Load Diff

13
convert_hf_to_gguf_update.py
View File

@ -59,6 +59,10 @@ parser.add_argument(
"--full", action="store_true",
help="download full list of models - make sure you have access to all of them",
)
parser.add_argument(
"--check-missing", action="store_true",
help="only check for missing pre-tokenizer hashes",
)
parser.add_argument(
"hf_token",
help="optional HF token",
@ -70,6 +74,10 @@ hf_token = args.hf_token if args.hf_token is not None else hf_token
if hf_token is None:
logger.warning("HF token not found. You can provide it as an argument or set it in ~/.cache/huggingface/token")
if args.check_missing and args.full:
logger.warning("Downloading full list of models requested, ignoring --check-missing!")
args.check_missing = False
# TODO: this string has to exercise as much pre-tokenizer functionality as possible
# will be updated with time - contributions welcome
CHK_TXT = '\n \n\n \n\n\n \t \t\t \t\n \n \n \n \n🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天～ ------======= нещо на Български \'\'\'\'\'\'```````\"\"\"\"......!!!!!!?????? I\'ve been \'told he\'s there, \'RE you sure? \'M not sure I\'ll make it, \'D you like some tea? We\'Ve a\'lL'
@ -130,6 +138,7 @@ models = [
{"name": "midm-2.0", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/K-intelligence/Midm-2.0-Base-Instruct", },
{"name": "lfm2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LiquidAI/LFM2-Tokenizer"},
{"name": "exaone4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/EXAONE-4.0-32B", },
{"name": "mellum", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/JetBrains/Mellum-4b-base", },
]
# some models are known to be broken upstream, so we will skip them as exceptions
@ -138,14 +147,17 @@ pre_computed_hashes = [
{"name": "chatglm-bpe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-chat", "chkhsh": "b6e8e1518dc4305be2fe39c313ed643381c4da5db34a98f6a04c093f8afbe99b"},
{"name": "chatglm-bpe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-chat", "chkhsh": "81d72c7348a9f0ebe86f23298d37debe0a5e71149e29bd283904c02262b27516"},
{"name": "glm4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-hf", "chkhsh": "a1336059768a55c99a734006ffb02203cd450fed003e9a71886c88acf24fdbc2"},
{"name": "glm4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/zai-org/GLM-4.5-Air", "chkhsh": "9ca2dd618e8afaf09731a7cf6e2105b373ba6a1821559f258b272fe83e6eb902"},
{"name": "minerva-7b", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0", "chkhsh": "1431a23e583c97432bc230bff598d103ddb5a1f89960c8f1d1051aaa944d0b35"},
{"name": "hunyuan", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tencent/Hunyuan-A13B-Instruct", "chkhsh": "7e57df22b1fe23a7b1e1c7f3dc4e3f96d43a4eb0836d0c6bdc3436d7b2f1c664"},
{"name": "hunyuan-dense", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tencent/Hunyuan-4B-Instruct", "chkhsh": "bba3b3366b646dbdded5dbc42d59598b849371afc42f7beafa914afaa5b70aa6"},
# falcon-h1 series uses 4 different tokenizers across model sizes (0.5b - 34b), hence we need to define 4 different hashes
{"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-0.5B-Base", "chkhsh": "a6b57017d60e6edb4d88ecc2845188e0eb333a70357e45dcc9b53964a73bbae6"},
{"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-1B-Base", "chkhsh": "60476e1243776c4fb1b993dbd7a5f15ac22f83c80afdf425fa5ae01c8d44ef86"},
{"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-7B-Base", "chkhsh": "3eda48b4c4dc7de733d1a8b3e3b4a85243dbbf704da2ee9d42c6beced8897896"},
{"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-34B-Base", "chkhsh": "48f8e02c0359c0bbdd82f26909171fac1c18a457bb47573ed1fe3bbb2c1cfd4b"},
{"name": "kimi-k2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/moonshotai/Kimi-K2-Base", "chkhsh": "81212dc7cdb7e0c1074ca62c5aeab0d43c9f52b8a737be7b12a777c953027890"},
{"name": "qwen2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen3-Embedding-0.6B", "chkhsh": "d4540891389ea895b53b399da6ac824becc30f2fba0e9ddbb98f92e55ca0e97c"},
]
@ -220,6 +232,7 @@ if not args.full:
all_models = models.copy()
models = [model for model in all_models if model["name"] not in existing_models]
if not args.check_missing:
logging.info(f"Downloading {len(models)} models...")
for model in models:
try:

2
convert_lora_to_gguf.py
View File

@ -340,7 +340,7 @@ if __name__ == '__main__':
sys.exit(1)
else:
logger.info(f"Loading base model: {dir_base_model.name}")
hparams = ModelBase.load_hparams(dir_base_model)
hparams = ModelBase.load_hparams(dir_base_model, False)
with torch.inference_mode():
try:

6
docs/backend/CANN.md
View File

@ -310,5 +310,7 @@ Specifies the memory pool management strategy:
Controls automatic cleanup of the memory pool. This option is only effective when using the prio or leg memory pool strategies.
## TODO
- Support more models and data types.
### GGML_CANN_WEIGHT_NZ
Converting the matmul weight format from ND to NZ can significantly improve performance on the 310I DUO NPU.

46
docs/build-s390x.md
View File

@ -42,14 +42,14 @@ cmake --build build --config Release -j $(nproc)
cmake --build build --config Release -j $(nproc)
```
- By default, NNPA is enabled when available. To disable it (not recommended):
- By default, NNPA is disabled by default. To enable it:
```bash
cmake -S . -B build \
-DCMAKE_BUILD_TYPE=Release \
-DGGML_BLAS=ON \
-DGGML_BLAS_VENDOR=OpenBLAS \
-DGGML_NNPA=OFF
-DGGML_NNPA=ON
cmake --build build --config Release -j $(nproc)
```
@ -84,9 +84,9 @@ All models need to be converted to Big-Endian. You can achieve this in three cas
![File Type - gguf](https://img.shields.io/badge/File_Type-gguf-fff)
You can find popular models pre-converted and verified at [s390x Ready Models](https://huggingface.co/collections/taronaeo/s390x-ready-models-672765393af438d0ccb72a08).
You can find popular models pre-converted and verified at [s390x Verified Models](https://huggingface.co/collections/taronaeo/s390x-verified-models-672765393af438d0ccb72a08) or [s390x Runnable Models](https://huggingface.co/collections/taronaeo/s390x-runnable-models-686e951824198df12416017e).
These models have already been converted from `safetensors` to `GGUF Big-Endian` and their respective tokenizers verified to run correctly on IBM z15 and later system.
These models have already been converted from `safetensors` to `GGUF` Big-Endian and their respective tokenizers verified to run correctly on IBM z15 and later system.
2. **Convert safetensors model to GGUF Big-Endian directly (recommended)**
@ -94,6 +94,14 @@ All models need to be converted to Big-Endian. You can achieve this in three cas
The model you are trying to convert must be in `safetensors` file format (for example [IBM Granite 3.3 2B](https://huggingface.co/ibm-granite/granite-3.3-2b-instruct)). Make sure you have downloaded the model repository for this case.
Ensure that you have installed the required packages in advance
```bash
pip3 install -r requirements.txt
```
Convert the `safetensors` model to `GGUF`
```bash
python3 convert_hf_to_gguf.py \
--outfile model-name-be.f16.gguf \
@ -116,7 +124,7 @@ All models need to be converted to Big-Endian. You can achieve this in three cas
![File Type - gguf](https://img.shields.io/badge/File_Type-gguf-fff)
The model you are trying to convert must be in `gguf` file format (for example [IBM Granite 3.3 2B](https://huggingface.co/ibm-granite/granite-3.3-2b-instruct-GGUF)). Make sure you have downloaded the model file for this case.
The model you are trying to convert must be in `gguf` file format (for example [IBM Granite 3.3 2B GGUF](https://huggingface.co/ibm-granite/granite-3.3-2b-instruct-GGUF)). Make sure you have downloaded the model file for this case.
```bash
python3 gguf-py/gguf/scripts/gguf_convert_endian.py model-name.f16.gguf BIG
@ -141,15 +149,15 @@ Only available in IBM z15 or later system with the `-DGGML_VXE=ON` (turned on by
### 2. NNPA Vector Intrinsics Acceleration
Only available in IBM z16 or later system with the `-DGGML_NNPA=ON` (turned on when available) compile flag. No hardware acceleration is possible with llama.cpp with older systems, such as IBM z15/arch13. In such systems, the APIs can still run but will use a scalar implementation.
Only available in IBM z16 or later system with the `-DGGML_NNPA=ON` (turned off by default) compile flag. No hardware acceleration is possible with llama.cpp with older systems, such as IBM z15/arch13. In such systems, the APIs can still run but will use a scalar implementation.
### 3. zDNN Accelerator
_Only available in IBM z16 or later system. No direction at the moment._
_Only available in IBM z16 / LinuxONE 4 or later system. No support currently available._
### 4. Spyre Accelerator
_No direction at the moment._
_Only available with IBM z17 / LinuxONE 5 or later system. No support currently available._
## Performance Tuning
@ -189,6 +197,26 @@ IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongl
Answer: Please ensure that your GCC compiler is of minimum GCC 15.1.0 version, and have `binutils` updated to the latest version. If this does not fix the problem, kindly open an issue.
4. Failing to install the `sentencepiece` package using GCC 15+
Answer: The `sentencepiece` team are aware of this as seen in [this issue](https://github.com/google/sentencepiece/issues/1108).
As a temporary workaround, please run the installation command with the following environment variables.
```bash
export CXXFLAGS="-include cstdint"
```
For example,
```bash
CXXFLAGS="-include cstdint" pip3 install -r requirements.txt
```
5. `-DGGML_NNPA=ON` generates gibberish output
Answer: We are aware of this as detailed in [this issue](https://github.com/ggml-org/llama.cpp/issues/14877). Please either try reducing the number of threads, or disable the compile option using `-DGGML_NNPA=OFF`.
## Getting Help on IBM Z & LinuxONE
1. **Bugs, Feature Requests**
@ -244,3 +272,5 @@ IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongl
- ✅ - acceleration available
- 🚫 - acceleration unavailable, will still run using scalar implementation
- ❓ - acceleration unknown, please contribute if you can test it yourself
Last Updated by **Aaron Teo (aaron.teo1@ibm.com)** on July 25, 2025.

3
docs/build.md
View File

@ -68,6 +68,9 @@ cmake --build build --config Release
cmake --build build-x64-windows-llvm-release
```
- Curl usage is enabled by default and can be turned off with `-DLLAMA_CURL=OFF`. Otherwise you need to install development libraries for libcurl.
- **Debian / Ubuntu:** `sudo apt-get install libcurl4-openssl-dev` # (or `libcurl4-gnutls-dev` if you prefer GnuTLS)
- **Fedora / RHEL / Rocky / Alma:** `sudo dnf install libcurl-devel`
- **Arch / Manjaro:** `sudo pacman -S curl` # includes libcurl headers
## BLAS Build

21
docs/development/HOWTO-add-model.md
View File

@ -23,11 +23,19 @@ The convert script reads the model configuration, tokenizer, tensor names+data a
The required steps to implement for an HF model are:
1. Define the model `Model.register` annotation in a new `Model` subclass, example:
1. Define the model `ModelBase.register` annotation in a new `TextModel` or `MmprojModel` subclass, example:
```python
@Model.register("MyModelForCausalLM")
class MyModel(Model):
@ModelBase.register("MyModelForCausalLM")
class MyModel(TextModel):
model_arch = gguf.MODEL_ARCH.MYMODEL
```
or
```python
@ModelBase.register("MyModelForConditionalGeneration")
class MyModel(MmprojModel):
model_arch = gguf.MODEL_ARCH.MYMODEL
```
@ -75,9 +83,10 @@ block_mappings_cfg: dict[MODEL_TENSOR, tuple[str, ...]] = {
`transformer.blocks.{bid}.norm_1` will be mapped to `blk.{bid}.attn_norm` in GGUF.
Depending on the model configuration, tokenizer, code and tensors layout, you will have to override:
- `Model#set_gguf_parameters`
- `Model#set_vocab`
- `Model#write_tensors`
- `TextModel#set_gguf_parameters`
- `MmprojModel#set_gguf_parameters`
- `ModelBase#set_vocab`
- `ModelBase#modify_tensors`
NOTE: Tensor names must end with `.weight` or `.bias` suffixes, that is the convention and several tools like `quantize` expect this to proceed the weights.

2
docs/docker.md
View File

@ -110,7 +110,7 @@ You may want to pass in some different `ARGS`, depending on the MUSA environment
The defaults are:
- `MUSA_VERSION` set to `rc4.0.1`
- `MUSA_VERSION` set to `rc4.2.0`
The resulting images, are essentially the same as the non-MUSA images:

3
docs/multimodal.md
View File

@ -97,6 +97,9 @@ NOTE: some models may require large context window, for example: `-c 8192`
# Qwen2-Audio and SeaLLM-Audio
# note: no pre-quantized GGUF this model, as they have very poor result
# ref: https://github.com/ggml-org/llama.cpp/pull/13760
# Mistral's Voxtral
(tool_name) -hf ggml-org/Voxtral-Mini-3B-2507-GGUF
```
**Mixed modalities**:

6
docs/multimodal/minicpmo2.6.md
View File

@ -13,7 +13,7 @@ If there are differences in usage, please refer to the official build [documenta
Clone llama.cpp:
```bash
git clone https://github.com/ggerganov/llama.cpp
git clone https://github.com/ggml-org/llama.cpp
cd llama.cpp
```
@ -29,8 +29,8 @@ cmake --build build --config Release
Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-o-2_6-gguf) by us)
```bash
python ./tools/mtmd/minicpmv-surgery.py -m ../MiniCPM-o-2_6
python ./tools/mtmd/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-o-2_6 --minicpmv-projector ../MiniCPM-o-2_6/minicpmv.projector --output-dir ../MiniCPM-o-2_6/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 4
python ./tools/mtmd/legacy-models/minicpmv-surgery.py -m ../MiniCPM-o-2_6
python ./tools/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-o-2_6 --minicpmv-projector ../MiniCPM-o-2_6/minicpmv.projector --output-dir ../MiniCPM-o-2_6/ --minicpmv_version 4
python ./convert_hf_to_gguf.py ../MiniCPM-o-2_6/model
# quantize int4 version

47
docs/multimodal/minicpmo4.0.md Normal file
View File

@ -0,0 +1,47 @@
## MiniCPM-o 4
### Prepare models and code
Download [MiniCPM-o-4](https://huggingface.co/openbmb/MiniCPM-o-4) PyTorch model from huggingface to "MiniCPM-o-4" folder.
### Build llama.cpp
Readme modification time: 20250206
If there are differences in usage, please refer to the official build [documentation](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md)
Clone llama.cpp:
```bash
git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
```
Build llama.cpp using `CMake`:
```bash
cmake -B build
cmake --build build --config Release
```
### Usage of MiniCPM-o 4
Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-o-4-gguf) by us)
```bash
python ./tools/mtmd/legacy-models/minicpmv-surgery.py -m ../MiniCPM-o-4
python ./tools/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-o-4 --minicpmv-projector ../MiniCPM-o-4/minicpmv.projector --output-dir ../MiniCPM-o-4/ --minicpmv_version 6
python ./convert_hf_to_gguf.py ../MiniCPM-o-4/model
# quantize int4 version
./build/bin/llama-quantize ../MiniCPM-o-4/model/ggml-model-f16.gguf ../MiniCPM-o-4/model/ggml-model-Q4_K_M.gguf Q4_K_M
```
Inference on Linux or Mac
```bash
# run in single-turn mode
./build/bin/llama-mtmd-cli -m ../MiniCPM-o-4/model/ggml-model-f16.gguf --mmproj ../MiniCPM-o-4/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
# run in conversation mode
./build/bin/llama-mtmd-cli -m ../MiniCPM-o-4/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-o-4/mmproj-model-f16.gguf
```

4
docs/multimodal/minicpmv2.5.md
View File

@ -28,8 +28,8 @@ cmake --build build --config Release
Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-Llama3-V-2_5-gguf) by us)
```bash
python ./tools/mtmd/minicpmv-surgery.py -m ../MiniCPM-Llama3-V-2_5
python ./tools/mtmd/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-Llama3-V-2_5 --minicpmv-projector ../MiniCPM-Llama3-V-2_5/minicpmv.projector --output-dir ../MiniCPM-Llama3-V-2_5/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 2
python ./tools/mtmd/legacy-models/minicpmv-surgery.py -m ../MiniCPM-Llama3-V-2_5
python ./tools/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-Llama3-V-2_5 --minicpmv-projector ../MiniCPM-Llama3-V-2_5/minicpmv.projector --output-dir ../MiniCPM-Llama3-V-2_5/ --minicpmv_version 2
python ./convert_hf_to_gguf.py ../MiniCPM-Llama3-V-2_5/model
# quantize int4 version

6
docs/multimodal/minicpmv2.6.md
View File

@ -12,7 +12,7 @@ If there are differences in usage, please refer to the official build [documenta
Clone llama.cpp:
```bash
git clone https://github.com/ggerganov/llama.cpp
git clone https://github.com/ggml-org/llama.cpp
cd llama.cpp
```
@ -28,8 +28,8 @@ cmake --build build --config Release
Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-V-2_6-gguf) by us)
```bash
python ./tools/mtmd/minicpmv-surgery.py -m ../MiniCPM-V-2_6
python ./tools/mtmd/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-V-2_6 --minicpmv-projector ../MiniCPM-V-2_6/minicpmv.projector --output-dir ../MiniCPM-V-2_6/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 3
python ./tools/mtmd/legacy-models/minicpmv-surgery.py -m ../MiniCPM-V-2_6
python ./tools/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-V-2_6 --minicpmv-projector ../MiniCPM-V-2_6/minicpmv.projector --output-dir ../MiniCPM-V-2_6/ --minicpmv_version 3
python ./convert_hf_to_gguf.py ../MiniCPM-V-2_6/model
# quantize int4 version

47
docs/multimodal/minicpmv4.0.md Normal file
View File

@ -0,0 +1,47 @@
## MiniCPM-V 4
### Prepare models and code
Download [MiniCPM-V-4](https://huggingface.co/openbmb/MiniCPM-V-4) PyTorch model from huggingface to "MiniCPM-V-4" folder.
### Build llama.cpp
Readme modification time: 20250206
If there are differences in usage, please refer to the official build [documentation](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md)
Clone llama.cpp:
```bash
git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
```
Build llama.cpp using `CMake`:
```bash
cmake -B build
cmake --build build --config Release
```
### Usage of MiniCPM-V 4
Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-V-4-gguf) by us)
```bash
python ./tools/mtmd/legacy-models/minicpmv-surgery.py -m ../MiniCPM-V-4
python ./tools/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-V-4 --minicpmv-projector ../MiniCPM-V-4/minicpmv.projector --output-dir ../MiniCPM-V-4/ --minicpmv_version 5
python ./convert_hf_to_gguf.py ../MiniCPM-V-4/model
# quantize int4 version
./build/bin/llama-quantize ../MiniCPM-V-4/model/ggml-model-f16.gguf ../MiniCPM-V-4/model/ggml-model-Q4_K_M.gguf Q4_K_M
```
Inference on Linux or Mac
```bash
# run in single-turn mode
./build/bin/llama-mtmd-cli -m ../MiniCPM-V-4/model/ggml-model-f16.gguf --mmproj ../MiniCPM-V-4/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
# run in conversation mode
./build/bin/llama-mtmd-cli -m ../MiniCPM-V-4/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-V-4/mmproj-model-f16.gguf
```

179
docs/ops.md
View File

@ -2,94 +2,101 @@
List of GGML operations and backend support status.
## How to add a backend to this table:
1. Run `test-backend-ops support --output csv` with your backend name and redirect output to a csv file in `docs/ops/` (e.g., `docs/ops/CUDA.csv`)
2. Regenerate `/docs/ops.md` via `./scripts/create_ops_docs.py`
Legend:
- ✅ Fully supported by this backend
- 🟡 Partially supported by this backend
- ❌ Not supported by this backend
| Operation | BLAS | CPU | CUDA | Metal |
|-----------|------|------|------|------|
| ABS | ❌ | ✅ | 🟡 | ❌ |
| ACC | ❌ | ✅ | ✅ | ✅ |
| ADD | ❌ | ✅ | ✅ | 🟡 |
| ADD1 | ❌ | ✅ | ✅ | ❌ |
| ARANGE | ❌ | ✅ | ✅ | ✅ |
| ARGMAX | ❌ | ✅ | ✅ | ✅ |
| ARGSORT | ❌ | ✅ | ✅ | ✅ |
| CLAMP | ❌ | ✅ | ✅ | 🟡 |
| CONCAT | ❌ | ✅ | 🟡 | ✅ |
| CONT | ❌ | ✅ | 🟡 | ✅ |
| CONV_2D_DW | ❌ | ✅ | ✅ | ❌ |
| CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ |
| CONV_TRANSPOSE_2D | ❌ | ✅ | ✅ | ❌ |
| COS | ❌ | ✅ | ✅ | 🟡 |
| COUNT_EQUAL | ❌ | ✅ | ✅ | ❌ |
| CPY | ❌ | 🟡 | 🟡 | 🟡 |
| CROSS_ENTROPY_LOSS | ❌ | ✅ | ✅ | ❌ |
| CROSS_ENTROPY_LOSS_BACK | ❌ | ✅ | ✅ | ❌ |
| DIAG_MASK_INF | ❌ | ✅ | ✅ | 🟡 |
| DIV | ❌ | ✅ | ✅ | 🟡 |
| DUP | ❌ | ✅ | 🟡 | 🟡 |
| ELU | ❌ | ✅ | ❌ | 🟡 |
| EXP | ❌ | ✅ | 🟡 | ❌ |
| FLASH_ATTN_EXT | ❌ | ✅ | 🟡 | 🟡 |
| GATED_LINEAR_ATTN | ❌ | ✅ | ✅ | ❌ |
| GEGLU | ❌ | ✅ | ✅ | 🟡 |
| GEGLU_ERF | ❌ | ✅ | ✅ | 🟡 |
| GEGLU_QUICK | ❌ | ✅ | ✅ | 🟡 |
| GELU | ❌ | ✅ | 🟡 | 🟡 |
| GELU_ERF | ❌ | ✅ | 🟡 | 🟡 |
| GELU_QUICK | ❌ | ✅ | 🟡 | 🟡 |
| GET_ROWS | ❌ | ✅ | 🟡 | ✅ |
| GET_ROWS_BACK | ❌ | 🟡 | 🟡 | ❌ |
| GROUP_NORM | ❌ | ✅ | ✅ | ✅ |
| HARDSIGMOID | ❌ | ✅ | 🟡 | ❌ |
| HARDSWISH | ❌ | ✅ | 🟡 | ❌ |
| IM2COL | ❌ | ✅ | ✅ | 🟡 |
| L2_NORM | ❌ | ✅ | ✅ | ✅ |
| LEAKY_RELU | ❌ | ✅ | ✅ | ✅ |
| LOG | ❌ | ✅ | ✅ | ❌ |
| MEAN | ❌ | ✅ | ✅ | ✅ |
| MUL | ❌ | ✅ | ✅ | 🟡 |
| MUL_MAT | 🟡 | 🟡 | 🟡 | 🟡 |
| MUL_MAT_ID | ❌ | ✅ | ✅ | ✅ |
| NEG | ❌ | ✅ | 🟡 | 🟡 |
| NORM | ❌ | ✅ | ✅ | 🟡 |
| OPT_STEP_ADAMW | ❌ | ✅ | ✅ | ❌ |
| OUT_PROD | 🟡 | 🟡 | 🟡 | ❌ |
| PAD | ❌ | ✅ | ✅ | ✅ |
| PAD_REFLECT_1D | ❌ | ✅ | ❌ | ✅ |
| POOL_2D | ❌ | ✅ | ✅ | ✅ |
| REGLU | ❌ | ✅ | ✅ | 🟡 |
| RELU | ❌ | ✅ | 🟡 | 🟡 |
| REPEAT | ❌ | ✅ | 🟡 | ✅ |
| REPEAT_BACK | ❌ | ✅ | ✅ | ❌ |
| RMS_NORM | ❌ | ✅ | ✅ | 🟡 |
| RMS_NORM_BACK | ❌ | ✅ | ✅ | ❌ |
| RMS_NORM_MUL | ❌ | ✅ | ✅ | ✅ |
| ROPE | ❌ | ✅ | ✅ | ✅ |
| ROPE_BACK | ❌ | ✅ | ✅ | ❌ |
| RWKV_WKV6 | ❌ | ✅ | ✅ | ✅ |
| RWKV_WKV7 | ❌ | ✅ | ✅ | ✅ |
| SCALE | ❌ | ✅ | ✅ | ✅ |
| SET | ❌ | ✅ | ❌ | ✅ |
| SET_ROWS | ❌ | 🟡 | ❌ | 🟡 |
| SGN | ❌ | ✅ | 🟡 | ❌ |
| SIGMOID | ❌ | ✅ | 🟡 | 🟡 |
| SILU | ❌ | ✅ | 🟡 | 🟡 |
| SILU_BACK | ❌ | ✅ | ✅ | ❌ |
| SIN | ❌ | ✅ | ✅ | 🟡 |
| SOFT_MAX | ❌ | ✅ | ✅ | ✅ |
| SOFT_MAX_BACK | ❌ | 🟡 | 🟡 | ❌ |
| SQR | ❌ | ✅ | ✅ | 🟡 |
| SQRT | ❌ | ✅ | ✅ | 🟡 |
| SSM_CONV | ❌ | ✅ | ✅ | ✅ |
| SSM_SCAN | ❌ | ✅ | ✅ | ✅ |
| STEP | ❌ | ✅ | 🟡 | ❌ |
| SUB | ❌ | ✅ | ✅ | 🟡 |
| SUM | ❌ | ✅ | ✅ | ❌ |
| SUM_ROWS | ❌ | ✅ | ✅ | ✅ |
| SWIGLU | ❌ | ✅ | ✅ | 🟡 |
| TANH | ❌ | ✅ | 🟡 | 🟡 |
| TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ |
| UPSCALE | ❌ | ✅ | ✅ | 🟡 |
| Operation | BLAS | CANN | CPU | CUDA | Metal | OpenCL | SYCL | Vulkan |
|-----------|------|------|------|------|------|------|------|------|
| ABS | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ |
| ACC | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ |
| ADD | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ |
| ADD1 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ |
| ARANGE | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
| ARGMAX | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ |
| ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
| CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 |
| CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | 🟡 | ✅ |
| CONT | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 |
| CONV_2D | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ |
| CONV_2D_DW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ |
| CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ |
| CONV_TRANSPOSE_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
| COS | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 |
| COUNT_EQUAL | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ |
| CPY | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
| CROSS_ENTROPY_LOSS | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
| CROSS_ENTROPY_LOSS_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
| DIAG_MASK_INF | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ |
| DIV | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ |
| DUP | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 |
| ELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ |
| EXP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ |
| FLASH_ATTN_EXT | ❌ | 🟡 | ✅ | 🟡 | 🟡 | ❌ | ❌ | 🟡 |
| GATED_LINEAR_ATTN | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ |
| GEGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 |
| GEGLU_ERF | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 |
| GEGLU_QUICK | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 |
| GELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
| GELU_ERF | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
| GELU_QUICK | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
| GET_ROWS | ❌ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 |
| GET_ROWS_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ |
| GROUP_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
| HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ |
| HARDSWISH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ |
| IM2COL | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ |
| L2_NORM | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ |
| LEAKY_RELU | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ |
| LOG | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ |
| MEAN | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
| MUL | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ |
| MUL_MAT | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
| MUL_MAT_ID | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ |
| NEG | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ |
| NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 |
| OPT_STEP_ADAMW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ |
| OUT_PROD | 🟡 | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ |
| PAD | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
| PAD_REFLECT_1D | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ |
| POOL_2D | ❌ | 🟡 | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ |
| REGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 |
| RELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
| REPEAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | 🟡 |
| REPEAT_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ |
| RMS_NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ |
| RMS_NORM_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ |
| RMS_NORM_MUL_ADD | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
| ROLL | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ |
| ROPE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
| ROPE_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ |
| RWKV_WKV6 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ |
| RWKV_WKV7 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ |
| SCALE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
| SET | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ |
| SET_ROWS | ❌ | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
| SGN | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ |
| SIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
| SILU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
| SILU_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ |
| SIN | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 |
| SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | 🟡 | ✅ |
| SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ✅ |
| SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 |
| SQRT | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | ❌ |
| SSM_CONV | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
| SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
| STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ |
| SUB | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ |
| SUM | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ |
| SUM_ROWS | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
| SWIGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 |
| TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | 🟡 |
| TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
| UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ |

14667
docs/ops/BLAS.csv
View File

File diff suppressed because it is too large Load Diff

8133
docs/ops/CANN.csv Normal file
View File

File diff suppressed because it is too large Load Diff

13883
docs/ops/CPU.csv
View File

File diff suppressed because it is too large Load Diff

13883
docs/ops/CUDA.csv
View File

File diff suppressed because it is too large Load Diff

14667
docs/ops/Metal.csv
View File

File diff suppressed because it is too large Load Diff

8133
docs/ops/OpenCL.csv Normal file
View File

File diff suppressed because it is too large Load Diff

8133
docs/ops/SYCL.csv Normal file
View File

File diff suppressed because it is too large Load Diff

8133
docs/ops/Vulkan.csv Normal file
View File

File diff suppressed because it is too large Load Diff

13
examples/diffusion/README.md Normal file
View File

@ -0,0 +1,13 @@
# Diffusion Text Generation
This directory contains implementations for Diffusion LLMs (DLLMs)
More Info:
- https://github.com/ggml-org/llama.cpp/pull/14644
- https://github.com/ggml-org/llama.cpp/pull/14771
Example of using Dream architechture: `llama-diffusion-cli -m dream7b.gguf -p "write code to train MNIST in pytorch" -ub 512 --diffusion-eps 0.001 --diffusion-algorithm 3 --diffusion-steps 256 --diffusion-visual`
Example of using LLaDA architechture: `llama-diffusion-cli -m llada-8b.gguf -p "write code to train MNIST in pytorch" -ub 512 --diffusion-block-length 32 --diffusion-steps 256 --diffusion-visual`

868
examples/diffusion/diffusion-cli.cpp
View File

@ -5,12 +5,22 @@
#include "log.h"
#include <limits.h>
#include <string>
#include <vector>
#include <algorithm>
#include <cmath>
#include <cstring>
#include <limits>
#include <random>
#include <string>
#include <vector>
enum diffusion_algorithm { ORIGIN = 0, ENTROPY_BASED = 1, MARGIN_BASED = 2, RANDOM = 3, CONFIDENCE_BASED = 4 };
// Unified transfer scheduling methods
enum transfer_schedule {
TIMESTEP_BASED = 0, // Dream-style: (1.0 - s/t) * remaining
BLOCK_BASED = 1, // LLaDA-style: process in blocks with get_num_transfer_tokens
};
typedef bool (*diffusion_step_callback_t)(int32_t step,
int32_t total_steps,
@ -18,325 +28,99 @@ typedef bool (*diffusion_step_callback_t)(int32_t step,
int32_t n_tokens,
void * user_data);
enum diffusion_alg {
DIFFUSION_ALG_ORIGIN = 0,
DIFFUSION_ALG_MASKGIT_PLUS = 1,
DIFFUSION_ALG_TOPK_MARGIN = 2,
DIFFUSION_ALG_ENTROPY = 3,
};
struct diffusion_params {
int32_t steps;
float eps;
float temperature;
float top_p;
int32_t top_k;
llama_token mask_token_id;
enum diffusion_alg algorithm;
float alg_temp;
diffusion_step_callback_t step_callback;
void * step_callback_user_data;
int32_t seed;
int32_t steps = 0;
float temperature = 0;
llama_token mask_token_id = LLAMA_TOKEN_NULL;
diffusion_step_callback_t step_callback = nullptr;
void * step_callback_user_data = nullptr;
int32_t seed = 0;
bool visual_mode = false;
bool shift_logits = false; // Shift logits by -1 after decode
float top_p = 0.;
int32_t top_k = 0.;
diffusion_algorithm algorithm = CONFIDENCE_BASED;
transfer_schedule schedule = TIMESTEP_BASED;
float cfg_scale = 0.; // Config scale for classifier-free guidance
float eps = 0.; // Timestep scheduling
int32_t block_length = 0; // Block size (for block scheduling)
float alg_temp = 0; // algorithm temperature (0.0 = deterministic)
bool add_gumbel_noise = false; // Add gumbel noise to the logits if temp > 0.0
int32_t max_length = 0; // Maximum sequence length
};
static diffusion_params diffusion_default_params() {
diffusion_params params = {};
params.steps = 64;
params.eps = 1e-3f;
params.temperature = 0.2f;
params.top_p = 0.95f;
params.top_k = 0;
params.mask_token_id = LLAMA_TOKEN_NULL;
params.algorithm = DIFFUSION_ALG_ORIGIN;
params.alg_temp = 0.0f;
params.step_callback = nullptr;
params.step_callback_user_data = nullptr;
params.seed = 0;
return params;
}
static void diffusion_generate(llama_context * ctx,
const llama_token * input_tokens,
llama_token * output_tokens,
int32_t n_input,
int32_t max_length,
struct diffusion_params params,
int32_t & n_generated) {
n_generated = 0;
if (!ctx || !input_tokens || !output_tokens || n_input <= 0 || max_length <= n_input) {
return;
}
const llama_model * model = llama_get_model(ctx);
// Initialize with input and pad with mask tokens
std::copy(input_tokens, input_tokens + n_input, output_tokens);
std::fill(output_tokens + n_input, output_tokens + max_length, params.mask_token_id);
std::mt19937 rng(params.seed);
std::vector<float> timesteps(params.steps + 1);
for (int32_t i = 0; i <= params.steps; i++) {
timesteps[i] = 1.0f - (float) i / params.steps * (1.0f - params.eps);
}
llama_set_causal_attn(ctx, false);
int32_t n_vocab = llama_vocab_n_tokens(llama_model_get_vocab(model));
std::vector<llama_token_data> candidates(n_vocab);
std::vector<llama_token_data> conf_candidates;
conf_candidates.reserve(max_length);
std::vector<int32_t> mask_positions;
mask_positions.reserve(max_length);
struct llama_sampler * sampler = llama_sampler_chain_init(llama_sampler_chain_default_params());
if (params.top_k > 0) {
llama_sampler_chain_add(sampler, llama_sampler_init_top_k(params.top_k));
}
if (params.top_p < 1.0f) {
llama_sampler_chain_add(sampler, llama_sampler_init_top_p(params.top_p, 1));
}
if (params.temperature > 0.0f) {
llama_sampler_chain_add(sampler, llama_sampler_init_temp(params.temperature));
}
llama_sampler_chain_add(sampler, llama_sampler_init_dist(params.seed));
struct llama_sampler * dist_sampler = llama_sampler_init_dist(params.seed);
llama_batch batch = llama_batch_init(max_length, 0, 1);
batch.n_tokens = max_length;
int64_t total_sampling_time = 0;
int64_t total_time = 0;
int64_t time_start = ggml_time_us();
for (int32_t step = 0; step < params.steps; step++) {
if (params.step_callback) {
if (!params.step_callback(step, params.steps, output_tokens, max_length, params.step_callback_user_data)) {
break;
}
}
for (int32_t i = 0; i < max_length; i++) {
batch.token[i] = output_tokens[i];
batch.pos[i] = i;
batch.n_seq_id[i] = 1;
batch.seq_id[i][0] = 0;
batch.logits[i] = 1;
}
int ret = llama_decode(ctx, batch);
if (ret != 0) {
LOG_ERR("%s: failed to decode at step %d, ret = %d\n", __func__, step, ret);
break;
}
float * raw_logits = llama_get_logits(ctx);
if (!raw_logits) {
LOG_ERR("%s: failed to get logits at step %d\n", __func__, step);
break;
}
auto get_logits_for_pos = [&](int32_t pos) -> const float * {
return pos == 0 ? raw_logits : raw_logits + (pos - 1) * n_vocab;
};
int64_t time_start_sampling = ggml_time_us();
mask_positions.clear();
for (int32_t i = 0; i < max_length; i++) {
if (output_tokens[i] == params.mask_token_id) {
mask_positions.push_back(i);
}
}
if (mask_positions.empty()) {
break;
}
float t = timesteps[step];
float s = timesteps[step + 1];
if (params.algorithm == DIFFUSION_ALG_ORIGIN) {
float p_transfer = (step < params.steps - 1) ? (1.0f - s / t) : 1.0f;
for (int32_t pos : mask_positions) {
if (std::uniform_real_distribution<float>(0.0f, 1.0f)(rng) < p_transfer) {
const float * pos_logits = get_logits_for_pos(pos);
for (int32_t token_id = 0; token_id < n_vocab; token_id++) {
candidates[token_id].id = token_id;
candidates[token_id].logit = pos_logits[token_id];
candidates[token_id].p = 0.0f;
}
llama_token_data_array cur_p = {
/* .data = */ candidates.data(),
/* .size = */ (size_t) n_vocab, // Reset size to full vocab
/* .selected = */ -1,
/* .sorted = */ false,
};
llama_sampler_apply(sampler, &cur_p);
output_tokens[pos] = cur_p.data[cur_p.selected].id;
}
}
} else {
std::vector<std::pair<float, int32_t>> confidences;
std::vector<llama_token> sampled_tokens(mask_positions.size());
for (size_t i = 0; i < mask_positions.size(); i++) {
int32_t pos = mask_positions[i];
const float * pos_logits = get_logits_for_pos(pos);
for (int32_t token_id = 0; token_id < n_vocab; token_id++) {
candidates[token_id].logit = pos_logits[token_id];
candidates[token_id].p = 0.0f;
candidates[token_id].id = token_id;
}
llama_token_data_array cur_p = {
/* .data = */ candidates.data(),
/* .size = */ candidates.size(),
/* .selected = */ -1,
/* .sorted = */ false,
};
llama_sampler_apply(sampler, &cur_p);
llama_token sampled_token = cur_p.data[cur_p.selected].id;
float confidence = 0.0f;
if (params.algorithm == DIFFUSION_ALG_ENTROPY) {
const float epsilon = 1e-10f;
for (size_t j = 0; j < cur_p.size; j++) {
float prob = cur_p.data[j].p;
confidence += prob * logf(prob + epsilon);
}
} else if (params.algorithm == DIFFUSION_ALG_TOPK_MARGIN) {
confidence = cur_p.data[0].p - cur_p.data[1].p;
} else {
confidence = cur_p.data[cur_p.selected].p;
}
sampled_tokens[i] = sampled_token;
confidences.emplace_back(confidence, i);
}
int32_t num_transfer =
(step < params.steps - 1) ? (int32_t) (mask_positions.size() * (1.0f - s / t)) : mask_positions.size();
if (num_transfer > 0) {
if (params.alg_temp == 0.0f) {
std::partial_sort(confidences.begin(), confidences.begin() + num_transfer, confidences.end(),
[](const std::pair<float, int32_t> & a, const std::pair<float, int32_t> & b) {
if (a.first != b.first) {
return a.first > b.first;
}
return a.second < b.second;
});
} else {
conf_candidates.clear();
for (int32_t pos = 0; pos < max_length; pos++) {
float conf_logit = -std::numeric_limits<float>::infinity();
auto it = std::find(mask_positions.begin(), mask_positions.end(), pos);
if (it != mask_positions.end()) {
size_t mask_idx = std::distance(mask_positions.begin(), it);
conf_logit = confidences[mask_idx].first / params.alg_temp; // Apply temperature scaling
}
conf_candidates.emplace_back(llama_token_data{ pos, conf_logit, 0.0f });
}
llama_token_data_array conf_array = {
/* .data = */ conf_candidates.data(),
/* .size = */ conf_candidates.size(),
/* .selected = */ -1,
/* .sorted = */ false,
};
for (int32_t i = 0; i < num_transfer; i++) {
// Apply distribution sampler to get selected index
llama_sampler_apply(dist_sampler, &conf_array);
int selected_idx = conf_array.selected;
confidences[i].second = conf_candidates[selected_idx].id;
conf_candidates[selected_idx].p = 0.0f;
conf_array.selected = -1;
}
}
if (params.alg_temp == 0.0f) {
// Deterministic - use confidence order
for (int32_t i = 0; i < num_transfer; i++) {
int32_t mask_idx = confidences[i].second;
int32_t pos = mask_positions[mask_idx];
llama_token token = sampled_tokens[mask_idx];
output_tokens[pos] = token;
}
} else {
for (int32_t i = 0; i < num_transfer; i++) {
int32_t pos = confidences[i].second;
auto it = std::find(mask_positions.begin(), mask_positions.end(), pos);
if (it != mask_positions.end()) {
int32_t mask_idx = std::distance(mask_positions.begin(), it);
output_tokens[pos] = sampled_tokens[mask_idx];
}
}
}
}
}
int64_t time_end_sampling = ggml_time_us();
total_sampling_time += time_end_sampling - time_start_sampling;
}
int64_t time_end = ggml_time_us();
total_time += time_end - time_start;
LOG_INF("\ntotal time: %0.2fms, time per step: %0.2fms, sampling time per step: %0.2fms\n",
total_time / 1000.0, total_time / 1000.0 / params.steps, total_sampling_time / 1000.0 / params.steps);
llama_batch_free(batch);
llama_sampler_free(sampler);
llama_sampler_free(dist_sampler);
n_generated = max_length;
}
static std::string format_input_text(const std::string & prompt, bool use_chat_template, llama_model * model) {
if (!use_chat_template) {
return prompt;
}
auto chat_templates = common_chat_templates_init(model, "");
common_chat_templates_inputs inputs;
common_chat_msg user_msg;
user_msg.role = "user";
user_msg.content = prompt;
inputs.add_generation_prompt = true;
inputs.messages.push_back(user_msg);
auto result = common_chat_templates_apply(chat_templates.get(), inputs);
return result.prompt;
}
struct callback_data {
const common_params_diffusion * diff_params;
diffusion_params * diff_params;
const llama_vocab * vocab;
int32_t n_input;
};
static float calculate_confidence(const llama_token_data_array & cur_p,
diffusion_algorithm algorithm,
std::mt19937 & rng) {
switch (algorithm) {
case CONFIDENCE_BASED:
return cur_p.data[cur_p.selected].p; // Selected token probability
case ENTROPY_BASED:
{
float entropy = 0.0f;
const float epsilon = 1e-10f;
for (size_t i = 0; i < cur_p.size; i++) {
float prob = cur_p.data[i].p;
entropy += prob * logf(prob + epsilon);
}
return -entropy; // Higher entropy = lower confidence
}
case MARGIN_BASED:
return (cur_p.size > 1) ? cur_p.data[0].p - cur_p.data[1].p : cur_p.data[0].p;
case RANDOM:
{
std::uniform_real_distribution<float> uniform(0.0f, 1.0f);
return uniform(rng); // Random confidence
}
case ORIGIN:
return cur_p.data[cur_p.selected].p;
default:
return 0.0f;
}
}
// Unified transfer count calculation function
static int32_t calculate_transfer_count(int32_t step,
int32_t total_steps,
int32_t remaining_masked,
transfer_schedule schedule,
float eps,
const std::vector<int32_t> & num_transfer_tokens = {}) {
switch (schedule) {
case TIMESTEP_BASED:
{
float t = 1.0f - (float) step / total_steps * (1.0f - eps);
float s = 1.0f - (float) (step + 1) / total_steps * (1.0f - eps);
float p_transfer = (step < total_steps - 1) ? (1.0f - s / t) : 1.0f;
return (int32_t) (remaining_masked * p_transfer);
}
case BLOCK_BASED:
if (!num_transfer_tokens.empty() && step < (int32_t) num_transfer_tokens.size()) {
return num_transfer_tokens[step];
}
return remaining_masked / (total_steps - step); // Fallback
default:
return remaining_masked / (total_steps - step);
}
}
static bool diffusion_step_callback(int32_t step,
int32_t total_steps,
const llama_token * tokens,
@ -391,6 +175,360 @@ static bool diffusion_step_callback(int32_t step,
return true;
}
static void add_gumbel_noise(float * logits, int32_t n_vocab, float temperature, std::mt19937 & rng) {
if (temperature == 0.0f) {
return;
}
std::uniform_real_distribution<double> uniform(0.0, 1.0);
for (int32_t i = 0; i < n_vocab; i++) {
double noise = uniform(rng);
// Prevent log(0)
noise = std::max(noise, 1e-20);
double gumbel_noise = std::pow(-std::log(noise), temperature);
logits[i] = std::exp(logits[i]) / gumbel_noise;
}
}
static std::vector<int32_t> get_num_transfer_tokens(int32_t mask_count, int32_t steps) {
std::vector<int32_t> num_transfer_tokens(steps);
int32_t base = mask_count / steps;
int32_t remainder = mask_count % steps;
for (int32_t i = 0; i < steps; i++) {
num_transfer_tokens[i] = base + (i < remainder ? 1 : 0);
}
return num_transfer_tokens;
}
static void diffusion_generate(llama_context * ctx,
const llama_token * input_tokens,
llama_token * output_tokens,
int32_t n_input,
const diffusion_params & params,
int32_t & n_generated) {
n_generated = 0;
if (!ctx || !input_tokens || !output_tokens || n_input <= 0 || params.max_length <= n_input) {
return;
}
const llama_model * model = llama_get_model(ctx);
// Initialize with input and pad with mask tokens
std::copy(input_tokens, input_tokens + n_input, output_tokens);
std::fill(output_tokens + n_input, output_tokens + params.max_length, params.mask_token_id);
std::mt19937 rng(params.seed);
llama_set_causal_attn(ctx, false);
int32_t n_vocab = llama_vocab_n_tokens(llama_model_get_vocab(model));
std::vector<llama_token_data> candidates(n_vocab);
std::vector<llama_token_data> conf_candidates;
conf_candidates.reserve(params.max_length);
std::vector<int32_t> mask_positions;
mask_positions.reserve(params.max_length);
// Setup sampler chain
struct llama_sampler * sampler = llama_sampler_chain_init(llama_sampler_chain_default_params());
if (params.top_k > 0) {
llama_sampler_chain_add(sampler, llama_sampler_init_top_k(params.top_k));
}
if (params.top_p < 1.0f) {
llama_sampler_chain_add(sampler, llama_sampler_init_top_p(params.top_p, 1));
}
if (params.temperature > 0.0f) {
llama_sampler_chain_add(sampler, llama_sampler_init_temp(params.temperature));
}
llama_sampler_chain_add(sampler, llama_sampler_init_dist(params.seed));
struct llama_sampler * dist_sampler = llama_sampler_init_dist(params.seed);
llama_batch batch = llama_batch_init(params.max_length, 0, 1);
batch.n_tokens = params.max_length;
// Pre-allocate buffers for CFG if needed
int32_t logits_size = n_vocab * params.max_length;
std::vector<float> cond_logits_buffer;
std::vector<llama_token> un_x_buffer;
if (params.cfg_scale > 0.0f) {
cond_logits_buffer.resize(logits_size);
un_x_buffer.resize(params.max_length);
}
// For block-based processing
std::vector<int32_t> num_transfer_tokens;
int32_t num_blocks = 1;
int32_t steps_per_block = params.steps;
if (params.schedule == BLOCK_BASED) {
GGML_ASSERT(params.max_length % params.block_length == 0);
num_blocks = params.max_length / params.block_length;
GGML_ASSERT(params.steps % num_blocks == 0);
steps_per_block = params.steps / num_blocks;
}
std::vector<float> confidence(params.max_length);
int64_t total_sampling_time = 0;
int64_t total_time = 0;
int64_t time_start = ggml_time_us();
for (int block_num = 0; block_num < num_blocks; block_num++) {
int32_t block_start = (params.schedule == BLOCK_BASED) ? n_input + block_num * params.block_length : 0;
int32_t block_end = (params.schedule == BLOCK_BASED) ?
std::min(n_input + (block_num + 1) * params.block_length, params.max_length) :
params.max_length;
// Count masked tokens in current block for block-based processing
if (params.schedule == BLOCK_BASED) {
int32_t block_mask_count = 0;
for (int i = block_start; i < block_end; i++) {
if (output_tokens[i] == params.mask_token_id) {
block_mask_count++;
}
}
num_transfer_tokens = get_num_transfer_tokens(block_mask_count, steps_per_block);
}
for (int32_t step = 0; step < steps_per_block; step++) {
int32_t global_step = block_num * steps_per_block + step;
if (params.step_callback) {
if (!params.step_callback(
global_step, params.steps, output_tokens, params.max_length, params.step_callback_user_data)) {
break;
}
}
// Setup batch
for (int32_t i = 0; i < params.max_length; i++) {
batch.token[i] = output_tokens[i];
batch.pos[i] = i;
batch.n_seq_id[i] = 1;
batch.seq_id[i][0] = 0;
batch.logits[i] = 1;
}
float * logits = nullptr;
if (params.cfg_scale > 0.0f) {
int ret = llama_decode(ctx, batch);
if (ret != 0) {
LOG_ERR("Failed to generate conditional");
break;
}
float * cond_logits_ptr = llama_get_logits(ctx);
std::memcpy(cond_logits_buffer.data(), cond_logits_ptr, logits_size * sizeof(float));
// Unconditional generation (mask input)
std::copy(output_tokens, output_tokens + params.max_length, un_x_buffer.begin());
for (int32_t i = 0; i < n_input; i++) {
un_x_buffer[i] = params.mask_token_id;
}
for (int32_t i = 0; i < params.max_length; i++) {
batch.token[i] = un_x_buffer[i];
}
ret = llama_decode(ctx, batch);
if (ret != 0) {
LOG_ERR("Failed to generate unconditional");
break;
}
float * uncond_logits = llama_get_logits(ctx);
// Apply CFG
for (int32_t i = 0; i < logits_size; i++) {
cond_logits_buffer[i] =
uncond_logits[i] + (params.cfg_scale + 1.0f) * (cond_logits_buffer[i] - uncond_logits[i]);
}
logits = cond_logits_buffer.data();
} else {
int ret = llama_decode(ctx, batch);
if (ret != 0) {
LOG_ERR("%s: failed to decode at step %d, ret = %d\n", __func__, global_step, ret);
break;
}
logits = llama_get_logits(ctx);
}
if (!logits) {
LOG_ERR("%s: failed to get logits at step %d\n", __func__, global_step);
break;
}
auto get_logits_for_pos = [&](int32_t pos) -> const float * {
if (params.shift_logits) {
return pos == 0 ? logits : logits + (pos - 1) * n_vocab;
}
return logits + (pos) *n_vocab;
};
int64_t time_start_sampling = ggml_time_us();
mask_positions.clear();
for (int32_t i = 0; i < params.max_length; i++) {
if (output_tokens[i] == params.mask_token_id) {
// For block-based, only consider current block
if (params.schedule != BLOCK_BASED || (i >= block_start && i < block_end)) {
mask_positions.push_back(i);
}
}
}
if (mask_positions.empty()) {
break;
}
if (params.add_gumbel_noise && params.temperature > 0.0f) {
add_gumbel_noise(logits, n_vocab, params.temperature, rng);
}
if (params.algorithm == ORIGIN) {
int32_t transfer_count = calculate_transfer_count(
step, steps_per_block, mask_positions.size(), params.schedule, params.eps, num_transfer_tokens);
float p_transfer = (float) transfer_count / mask_positions.size();
for (int32_t pos : mask_positions) {
if (std::uniform_real_distribution<float>(0.0f, 1.0f)(rng) < p_transfer) {
const float * pos_logits = get_logits_for_pos(pos);
for (int32_t token_id = 0; token_id < n_vocab; token_id++) {
candidates[token_id].id = token_id;
candidates[token_id].logit = pos_logits[token_id];
candidates[token_id].p = 0.0f;
}
llama_token_data_array cur_p = {
candidates.data(),
(size_t) n_vocab,
-1,
false,
};
llama_sampler_apply(sampler, &cur_p);
output_tokens[pos] = cur_p.data[cur_p.selected].id;
}
}
} else {
std::vector<std::pair<float, int32_t>> confidences;
std::vector<llama_token> sampled_tokens(mask_positions.size());
for (size_t i = 0; i < mask_positions.size(); i++) {
int32_t pos = mask_positions[i];
const float * pos_logits = get_logits_for_pos(pos);
for (int32_t token_id = 0; token_id < n_vocab; token_id++) {
candidates[token_id].logit = pos_logits[token_id];
candidates[token_id].p = 0.0f;
candidates[token_id].id = token_id;
}
llama_token_data_array cur_p = {
candidates.data(),
candidates.size(),
-1,
false,
};
llama_sampler_apply(sampler, &cur_p);
llama_token sampled_token = cur_p.data[cur_p.selected].id;
float conf = calculate_confidence(cur_p, params.algorithm, rng);
sampled_tokens[i] = sampled_token;
confidences.emplace_back(conf, i);
}
int32_t transfer_count = calculate_transfer_count(
step, steps_per_block, mask_positions.size(), params.schedule, params.eps, num_transfer_tokens);
if (transfer_count > 0) {
if (params.alg_temp == 0.0f) {
std::partial_sort(confidences.begin(),
confidences.begin() + std::min(transfer_count, (int32_t) confidences.size()),
confidences.end(),
[](const std::pair<float, int32_t> & a, const std::pair<float, int32_t> & b) {
if (a.first != b.first) {
return a.first > b.first;
}
return a.second < b.second;
});
for (int32_t i = 0; i < std::min(transfer_count, (int32_t) confidences.size()); i++) {
int32_t mask_idx = confidences[i].second;
int32_t pos = mask_positions[mask_idx];
output_tokens[pos] = sampled_tokens[mask_idx];
}
} else {
conf_candidates.clear();
for (size_t i = 0; i < confidences.size(); i++) {
float conf_logit = confidences[i].first / params.alg_temp;
conf_candidates.emplace_back(llama_token_data{ (int32_t) i, conf_logit, 0.0f });
}
llama_token_data_array conf_array = {
conf_candidates.data(),
conf_candidates.size(),
-1,
false,
};
for (int32_t i = 0; i < std::min(transfer_count, (int32_t) confidences.size()); i++) {
llama_sampler_apply(dist_sampler, &conf_array);
int32_t selected_idx = conf_array.selected;
int32_t mask_idx = selected_idx;
int32_t pos = mask_positions[mask_idx];
output_tokens[pos] = sampled_tokens[mask_idx];
conf_candidates[selected_idx].p = 0.0f;
conf_array.selected = -1;
}
}
}
}
int64_t time_end_sampling = ggml_time_us();
total_sampling_time += time_end_sampling - time_start_sampling;
}
}
int64_t time_end = ggml_time_us();
total_time += time_end - time_start;
LOG_INF("\ntotal time: %0.2fms, time per step: %0.2fms, sampling time per step: %0.2fms\n",
total_time / 1000.0,
total_time / 1000.0 / params.steps,
total_sampling_time / 1000.0 / params.steps);
llama_batch_free(batch);
llama_sampler_free(sampler);
llama_sampler_free(dist_sampler);
n_generated = params.max_length;
}
static std::string format_input_text(const std::string & prompt, bool use_chat_template, llama_model * model) {
if (!use_chat_template) {
return prompt;
}
auto chat_templates = common_chat_templates_init(model, "");
common_chat_templates_inputs inputs;
common_chat_msg user_msg;
user_msg.role = "user";
user_msg.content = prompt;
inputs.add_generation_prompt = true;
inputs.messages.push_back(user_msg);
auto result = common_chat_templates_apply(chat_templates.get(), inputs);
return result.prompt;
}
int main(int argc, char ** argv) {
ggml_time_init();
@ -400,11 +538,6 @@ int main(int argc, char ** argv) {
return 1;
}
const char * alg_names[] = { "ORIGIN", "MASKGIT_PLUS", "TOPK_MARGIN", "ENTROPY" };
const char * alg_name = (params.diffusion.algorithm >= 0 && params.diffusion.algorithm <= 3) ?
alg_names[params.diffusion.algorithm] :
"UNKNOWN";
common_init();
llama_backend_init();
@ -421,6 +554,12 @@ int main(int argc, char ** argv) {
return 1;
}
if (!llama_model_is_diffusion(model)) {
LOG_ERR("error: unsupported model for diffusion");
llama_model_free(model);
return 1;
}
llama_context_params ctx_params = llama_context_default_params();
ctx_params.n_ctx = params.n_ctx;
ctx_params.n_batch = params.n_batch;
@ -442,9 +581,11 @@ int main(int argc, char ** argv) {
const llama_vocab * vocab = llama_model_get_vocab(model);
std::string formatted_prompt = format_input_text(params.prompt, params.enable_chat_template, model);
std::vector<llama_token> input_tokens = common_tokenize(vocab, formatted_prompt,
std::vector<llama_token> input_tokens = common_tokenize(vocab,
formatted_prompt,
/*add special tokens*/ true,
/*parse special*/ true);
int n_input = input_tokens.size();
if (n_input >= params.n_ctx) {
@ -454,44 +595,79 @@ int main(int argc, char ** argv) {
return 1;
}
struct diffusion_params ldiff_params = diffusion_default_params();
ldiff_params.steps = params.diffusion.steps;
ldiff_params.eps = params.diffusion.eps;
ldiff_params.temperature = params.sampling.temp;
ldiff_params.top_p = params.sampling.top_p;
ldiff_params.top_k = params.sampling.top_k;
ldiff_params.algorithm = static_cast<enum diffusion_alg>(params.diffusion.algorithm);
ldiff_params.alg_temp = params.diffusion.alg_temp;
ldiff_params.seed = params.sampling.seed;
llama_token mask_token_id = llama_vocab_mask(vocab);
GGML_ASSERT(mask_token_id != LLAMA_TOKEN_NULL);
LOG_INF("diffusion_params: - %-25s llama_token = %d\n", "mask_token_id", mask_token_id);
LOG_INF("diffusion_params: - %-25s u32 = %d\n", "steps", params.diffusion.steps);
LOG_INF("diffusion_params: - %-25s f32 = %.6f\n", "eps", params.diffusion.eps);
LOG_INF("diffusion_params: - %-25s u32 = %d (%s)\n", "algorithm", params.diffusion.algorithm,
alg_name);
LOG_INF("diffusion_params: - %-25s f32 = %.3f\n", "alg_temp", params.diffusion.alg_temp);
ldiff_params.mask_token_id = mask_token_id;
callback_data cb_data = { &params.diffusion, vocab, n_input };
ldiff_params.step_callback = diffusion_step_callback;
ldiff_params.step_callback_user_data = &cb_data;
bool visual_mode = params.diffusion.visual_mode;
int32_t n_generated = 0;
std::vector<llama_token> output_tokens(params.n_ubatch);
diffusion_generate(ctx, input_tokens.data(), output_tokens.data(), n_input, params.n_ubatch,
ldiff_params, n_generated);
struct diffusion_params diff_params;
char shift_logits_str[8];
if (llama_model_meta_val_str(model, "diffusion.shift_logits", shift_logits_str, sizeof(shift_logits_str)) >= 0) {
diff_params.shift_logits = (strcmp(shift_logits_str, "true") == 0);
} else {
diff_params.shift_logits = true;
}
//Use either eps or block length, but not both
GGML_ASSERT((params.diffusion.eps == 0) ^ (params.diffusion.block_length == 0));
if (params.diffusion.eps) {
diff_params.schedule = TIMESTEP_BASED;
diff_params.eps = params.diffusion.eps;
} else if (params.diffusion.block_length) {
diff_params.schedule = BLOCK_BASED;
diff_params.block_length = params.diffusion.block_length;
}
diff_params.mask_token_id = mask_token_id;
diff_params.seed = params.sampling.seed;
diff_params.temperature = params.sampling.temp;
diff_params.steps = params.diffusion.steps;
diff_params.algorithm = static_cast<diffusion_algorithm>(params.diffusion.algorithm);
diff_params.max_length = params.n_ubatch;
diff_params.top_p = params.sampling.top_p;
diff_params.top_k = params.sampling.top_k;
diff_params.visual_mode = params.diffusion.visual_mode;
diff_params.add_gumbel_noise = params.diffusion.add_gumbel_noise;
diff_params.step_callback = diffusion_step_callback;
callback_data cb_data = { &diff_params, vocab, n_input };
diff_params.step_callback_user_data = &cb_data;
const char * alg_names[] = { "ORIGIN", "ENTROPY_BASED", "MARGIN_BASED", "RANDOM", "CONFIDENCE_BASED" };
const char * sched_names[] = { "TIMESTEP_BASED", "BLOCK_BASED" };
const char * alg_name =
(diff_params.algorithm >= 0 && diff_params.algorithm <= 4) ? alg_names[diff_params.algorithm] : "UNKNOWN";
const char * sched_name =
(diff_params.schedule >= 0 && diff_params.schedule <= 1) ? sched_names[diff_params.schedule] : "UNKNOWN";
LOG_INF("diffusion_params: - %-25s llama_token = %d\n", "mask_token_id", mask_token_id);
LOG_INF("diffusion_params: - %-25s u32 = %d\n", "steps", diff_params.steps);
LOG_INF("diffusion_params: - %-25s u32 = %d\n", "max_length", diff_params.max_length);
LOG_INF("diffusion_params: - %-25s enum = %d (%s)\n", "algorithm", diff_params.algorithm, alg_name);
LOG_INF("diffusion_params: - %-25s enum = %d (%s)\n", "schedule", diff_params.schedule, sched_name);
LOG_INF("diffusion_params: - %-25s f32 = %.3f\n", "temperature", diff_params.temperature);
if (diff_params.schedule == TIMESTEP_BASED) {
LOG_INF("diffusion_params: - %-25s f32 = %.6f\n", "eps", diff_params.eps);
LOG_INF("diffusion_params: - %-25s f32 = %.3f\n", "alg_temp", diff_params.alg_temp);
}
if (diff_params.schedule == BLOCK_BASED) {
LOG_INF("diffusion_params: - %-25s u32 = %d\n", "block_length", diff_params.block_length);
LOG_INF("diffusion_params: - %-25s f32 = %.3f\n", "cfg_scale", diff_params.cfg_scale);
}
diffusion_generate(ctx, input_tokens.data(), output_tokens.data(), n_input, diff_params, n_generated);
if (n_generated > 0) {
if (params.diffusion.visual_mode) {
if (visual_mode) {
//clear screen and move cursor to top-left
LOG_INF("\033[2J\033[H");
}
output_tokens.erase(output_tokens.begin(), output_tokens.begin() + n_input);
std::string output_data = common_detokenize(vocab, output_tokens, false);
LOG_INF("\n%s\n", output_data.c_str());

8
examples/embedding/embedding.cpp
View File

@ -81,6 +81,14 @@ int main(int argc, char ** argv) {
params.embedding = true;
// if the number of prompts that would be encoded is known in advance, it's more efficient to specify the
// --parallel argument accordingly. for convenience, if not specified, we fallback to unified KV cache
// in order to support any number of prompts
if (params.n_parallel == 1) {
LOG_INF("%s: n_parallel == 1 -> unified KV cache is enabled\n", __func__);
params.kv_unified = true;
}
// utilize the full context
if (params.n_batch < params.n_ctx) {
LOG_WRN("%s: setting batch size to %d\n", __func__, params.n_ctx);

7
examples/eval-callback/eval-callback.cpp
View File

@ -7,6 +7,7 @@
#include <cstdio>
#include <string>
#include <vector>
#include <numeric>
/**
* This the arbitrary data which will be passed to each callback.
@ -77,6 +78,12 @@ static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne
LOG(" ]\n");
LOG(" sum = %f\n", sum);
}
// TODO: make this abort configurable/optional?
if (std::isnan(sum)) {
LOG_ERR("encountered NaN - aborting\n");
exit(0);
}
}
/**

6
examples/save-load-state/save-load-state.cpp
View File

@ -15,6 +15,12 @@ int main(int argc, char ** argv) {
return 1;
}
if (params.n_parallel == 1) {
// the example uses 2 sequences, so when n_parallel == 1, we need to enable unified kv cache
printf("%s: n_parallel == 1, enabling unified kv cache\n", __func__);
params.kv_unified = true;
}
common_init();
if (params.n_predict < 0) {

9
examples/speculative-simple/speculative-simple.cpp
View File

@ -59,13 +59,15 @@ int main(int argc, char ** argv) {
}
params.cpuparams_batch.n_threads = params.speculative.cpuparams_batch.n_threads;
params.tensor_buft_overrides = params.speculative.tensor_buft_overrides;
common_init_result llama_init_dft = common_init_from_params(params);
//model_dft = llama_init_dft.model.get();
ctx_dft = llama_init_dft.context.get();
if (!common_speculative_are_compatible(ctx_tgt, ctx_dft)) {
return 1;
LOG_INF("the draft model '%s' is not compatible with the target model '%s'. tokens will be translated between the draft and target models.\n", params.speculative.model.path.c_str(), params.model.path.c_str());
}
// Tokenize the prompt
@ -130,7 +132,10 @@ int main(int argc, char ** argv) {
params_spec.n_reuse = llama_n_ctx(ctx_dft) - n_draft;
params_spec.p_min = p_min;
struct common_speculative * spec = common_speculative_init(ctx_dft);
struct common_speculative * spec = common_speculative_init(ctx_tgt, ctx_dft);
for (auto &pair : params.speculative.replacements) {
common_speculative_add_replacement_tgt_dft(spec, pair.first.c_str(), pair.second.c_str());
}
llama_batch batch_tgt = llama_batch_init(llama_n_batch(ctx_tgt), 0, 1);

2
examples/speculative/speculative.cpp
View File

@ -85,6 +85,8 @@ int main(int argc, char ** argv) {
}
params.cpuparams_batch.n_threads = params.speculative.cpuparams_batch.n_threads;
params.tensor_buft_overrides = params.speculative.tensor_buft_overrides;
common_init_result llama_init_dft = common_init_from_params(params);
model_dft = llama_init_dft.model.get();

26
examples/training/finetune.cpp
View File

@ -15,15 +15,15 @@
int main(int argc, char ** argv) {
common_params params;
params.escape = false;
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_PERPLEXITY)) {
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_FINETUNE)) {
return 1;
}
if (params.use_mmap) {
LOG_INF("%s: force disabling memory mapping because it would result in-read-only pointers to the weights\n", __func__);
LOG_INF("%s: force disabling memory mapping because it would result in-read-only pointers to the weights\n",
__func__);
params.use_mmap = false;
}
if (params.cache_type_k != GGML_TYPE_F32) {
@ -38,7 +38,6 @@ int main(int argc, char ** argv) {
common_init();
llama_backend_init();
llama_numa_init(params.numa);
// load the model and apply lora adapter, if any
common_init_result llama_init = common_init_from_params(params);
llama_model_ptr & model = llama_init.model;
@ -55,29 +54,30 @@ int main(int argc, char ** argv) {
LOG_INF("%s\n", common_params_get_system_info(params).c_str());
}
constexpr float val_split = 0.05f;
std::vector<llama_token> tokens = common_tokenize(ctx.get(), params.prompt, true);
ggml_opt_dataset_t dataset = common_opt_dataset_init(ctx.get(), tokens, llama_n_ctx(ctx.get()) / 2);
struct ggml_opt_optimizer_params optimizer_params = ggml_opt_get_default_optimizer_params(nullptr);
optimizer_params.adamw.alpha = 1e-7f; // learning rate
struct lr_opt & lr = params.lr;
LOG_INF("-optimizer %s -lr0 %.2g -wd %.2g -lr-min %.2g -min-epochs %.2g -epochs %d -period %.2g -val %.2g\n",
ggml_opt_optimizer_name(params.optimizer), (double) lr.lr0, (double) lr.wd, (double) lr.lr_min, (double) lr.decay_epochs,
(unsigned) lr.epochs, (double) params.n_batch / params.n_ubatch, (double) params.val_split);
struct llama_opt_params lopt_params{
/*n_ctx_train =*/0,
/*param_filter =*/llama_opt_param_filter_all,
/*param_filter_ud =*/nullptr,
/*get_opt_pars =*/ ggml_opt_get_constant_optimizer_params,
/*get_opt_pars_ud =*/ &optimizer_params,
/*get_opt_pars =*/common_opt_lr_pars,
/*get_opt_pars_ud =*/&params.lr,
/*optimizer_type =*/params.optimizer,
};
llama_opt_init(ctx.get(), model.get(), lopt_params);
const int64_t idata_split = ggml_opt_dataset_ndata(dataset) * (1.0f - val_split);
const int64_t idata_split = ggml_opt_dataset_ndata(dataset) * (1.0f - params.val_split);
ggml_opt_result_t result_train = ggml_opt_result_init();
ggml_opt_result_t result_eval = ggml_opt_result_init();
for (int epoch = 0; epoch < 2; ++epoch) {
for (lr.epoch = 0; lr.epoch < lr.epochs; ++lr.epoch) {
llama_opt_epoch(ctx.get(), dataset, result_train, result_eval, idata_split,
ggml_opt_epoch_callback_progress_bar, ggml_opt_epoch_callback_progress_bar);
fprintf(stderr, "\n");
@ -88,7 +88,7 @@ int main(int argc, char ** argv) {
ggml_opt_result_free(result_train);
ggml_opt_result_free(result_eval);
llama_model_save_to_file(model.get(), "finetuned-model.gguf");
llama_model_save_to_file(model.get(), params.out_file.c_str());
llama_backend_free();

5
flake.nix
View File

@ -36,9 +36,6 @@
# ```
# nixConfig = {
# extra-substituters = [
# # Populated by the CI in ggml-org/llama.cpp
# "https://llama-cpp.cachix.org"
#
# # A development cache for nixpkgs imported with `config.cudaSupport = true`.
# # Populated by https://hercules-ci.com/github/SomeoneSerge/nixpkgs-cuda-ci.
# # This lets one skip building e.g. the CUDA-enabled openmpi.
@ -47,10 +44,8 @@
# ];
#
# # Verify these are the same keys as published on
# # - https://app.cachix.org/cache/llama-cpp
# # - https://app.cachix.org/cache/cuda-maintainers
# extra-trusted-public-keys = [
# "llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc="
# "cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E="
# ];
# };

7
ggml/CMakeLists.txt
View File

@ -41,6 +41,7 @@ endif()
option(BUILD_SHARED_LIBS "ggml: build shared libraries" ${BUILD_SHARED_LIBS_DEFAULT})
option(GGML_BACKEND_DL "ggml: build backends as dynamic libraries (requires BUILD_SHARED_LIBS)" OFF)
set(GGML_BACKEND_DIR "" CACHE PATH "ggml: directory to load dynamic backends from (requires GGML_BACKEND_DL")
#
# option list
@ -131,7 +132,7 @@ option(GGML_RVV "ggml: enable rvv" ON)
option(GGML_RV_ZFH "ggml: enable riscv zfh" OFF)
option(GGML_XTHEADVECTOR "ggml: enable xtheadvector" OFF)
option(GGML_VXE "ggml: enable vxe" ON)
option(GGML_NNPA "ggml: enable nnpa" ON)
option(GGML_NNPA "ggml: enable nnpa" OFF) # temp disabled by default, see: https://github.com/ggml-org/llama.cpp/issues/14877
option(GGML_CPU_ALL_VARIANTS "ggml: build all variants of the CPU backend (requires GGML_BACKEND_DL)" OFF)
set(GGML_CPU_ARM_ARCH "" CACHE STRING "ggml: CPU architecture for ARM")
@ -174,6 +175,10 @@ option(GGML_HIP_GRAPHS "ggml: use HIP graph, experimental,
option(GGML_HIP_NO_VMM "ggml: do not try to use HIP VMM" ON)
option(GGML_HIP_ROCWMMA_FATTN "ggml: enable rocWMMA for FlashAttention" OFF)
option(GGML_HIP_FORCE_ROCWMMA_FATTN_GFX12 "ggml: enable rocWMMA FlashAttention on GFX12" OFF)
option(GGML_HIP_MMQ_MFMA "ggml: enable MFMA MMA for CDNA in MMQ" ON)
option(GGML_HIP_EXPORT_METRICS "ggml: enable kernel perf metrics output" OFF)
option(GGML_MUSA_GRAPHS "ggml: use MUSA graph, experimental, unstable" OFF)
option(GGML_MUSA_MUDNN_COPY "ggml: enable muDNN for accelerated copy" OFF)
option(GGML_VULKAN "ggml: use Vulkan" OFF)
option(GGML_VULKAN_CHECK_RESULTS "ggml: run Vulkan op checks" OFF)
option(GGML_VULKAN_DEBUG "ggml: enable Vulkan debug output" OFF)

161
ggml/cmake/ggml-config.cmake.in
View File

@ -1,12 +1,107 @@
@PACKAGE_INIT@
@GGML_VARIABLES_EXPANDED@
@PACKAGE_INIT@
# Find all dependencies before creating any target.
include(CMakeFindDependencyMacro)
find_dependency(Threads)
if (NOT GGML_SHARED_LIB)
set(GGML_CPU_INTERFACE_LINK_LIBRARIES "")
set(GGML_CPU_INTERFACE_LINK_OPTIONS "")
if (APPLE AND GGML_ACCELERATE)
find_library(ACCELERATE_FRAMEWORK Accelerate)
if(NOT ACCELERATE_FRAMEWORK)
set(${CMAKE_FIND_PACKAGE_NAME}_FOUND 0)
return()
endif()
list(APPEND GGML_CPU_INTERFACE_LINK_LIBRARIES ${ACCELERATE_FRAMEWORK})
endif()
if (GGML_OPENMP_ENABLED)
find_dependency(OpenMP)
list(APPEND GGML_CPU_INTERFACE_LINK_LIBRARIES OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
endif()
if (GGML_CPU_HBM)
find_library(memkind memkind)
if(NOT memkind)
set(${CMAKE_FIND_PACKAGE_NAME}_FOUND 0)
return()
endif()
list(APPEND GGML_CPU_INTERFACE_LINK_LIBRARIES memkind)
endif()
if (GGML_BLAS)
find_dependency(BLAS)
list(APPEND GGML_BLAS_INTERFACE_LINK_LIBRARIES ${BLAS_LIBRARIES})
list(APPEND GGML_BLAS_INTERFACE_LINK_OPTIONS ${BLAS_LINKER_FLAGS})
endif()
if (GGML_CUDA)
set(GGML_CUDA_INTERFACE_LINK_LIBRARIES "")
find_dependency(CUDAToolkit)
if (GGML_STATIC)
list(APPEND GGML_CUDA_INTERFACE_LINK_LIBRARIES $<LINK_ONLY:CUDA::cudart_static>)
if (WIN32)
list(APPEND GGML_CUDA_INTERFACE_LINK_LIBRARIES $<LINK_ONLY:CUDA::cublas> $<LINK_ONLY:CUDA::cublasLt>)
else()
list(APPEND GGML_CUDA_INTERFACE_LINK_LIBRARIES $<LINK_ONLY:CUDA::cublas_static> $<LINK_ONLY:CUDA::cublasLt_static>)
endif()
endif()
if (NOT GGML_CUDA_NO_VMM)
list(APPEND GGML_CUDA_INTERFACE_LINK_LIBRARIES $<LINK_ONLY:CUDA::cuda_driver>)
endif()
endif()
if (GGML_METAL)
find_library(FOUNDATION_LIBRARY Foundation)
find_library(METAL_FRAMEWORK Metal)
find_library(METALKIT_FRAMEWORK MetalKit)
if(NOT FOUNDATION_LIBRARY OR NOT METAL_FRAMEWORK OR NOT METALKIT_FRAMEWORK)
set(${CMAKE_FIND_PACKAGE_NAME}_FOUND 0)
return()
endif()
set(GGML_METAL_INTERFACE_LINK_LIBRARIES
${FOUNDATION_LIBRARY} ${METAL_FRAMEWORK} ${METALKIT_FRAMEWORK})
endif()
if (GGML_OPENCL)
find_dependency(OpenCL)
set(GGML_OPENCL_INTERFACE_LINK_LIBRARIES $<LINK_ONLY:OpenCL::OpenCL>)
endif()
if (GGML_VULKAN)
find_dependency(Vulkan)
set(GGML_VULKAN_INTERFACE_LINK_LIBRARIES $<LINK_ONLY:Vulkan::Vulkan>)
endif()
if (GGML_HIP)
find_dependency(hip)
find_dependency(hipblas)
find_dependency(rocblas)
set(GGML_HIP_INTERFACE_LINK_LIBRARIES hip::host roc::rocblas roc::hipblas)
endif()
if (GGML_SYCL)
set(GGML_SYCL_INTERFACE_LINK_LIBRARIES "")
find_package(DNNL)
if (${DNNL_FOUND} AND GGML_SYCL_TARGET STREQUAL "INTEL")
list(APPEND GGML_SYCL_INTERFACE_LINK_LIBRARIES DNNL::dnnl)
endif()
if (WIN32)
find_dependency(IntelSYCL)
find_dependency(MKL)
list(APPEND GGML_SYCL_INTERFACE_LINK_LIBRARIES IntelSYCL::SYCL_CXX MKL::MKL MKL::MKL_SYCL)
endif()
endif()
endif()
set_and_check(GGML_INCLUDE_DIR "@PACKAGE_GGML_INCLUDE_INSTALL_DIR@")
set_and_check(GGML_LIB_DIR "@PACKAGE_GGML_LIB_INSTALL_DIR@")
#set_and_check(GGML_BIN_DIR "@PACKAGE_GGML_BIN_INSTALL_DIR@")
if(NOT TARGET ggml::ggml)
find_package(Threads REQUIRED)
find_library(GGML_LIBRARY ggml
@ -29,67 +124,8 @@ set_target_properties(ggml::ggml-base
PROPERTIES
IMPORTED_LOCATION "${GGML_BASE_LIBRARY}")
if (NOT GGML_SHARED_LIB)
if (APPLE AND GGML_ACCELERATE)
find_library(ACCELERATE_FRAMEWORK Accelerate REQUIRED)
list(APPEND GGML_CPU_INTERFACE_LINK_LIBRARIES ${ACCELERATE_FRAMEWORK})
endif()
if (GGML_OPENMP)
find_package(OpenMP REQUIRED)
list(APPEND GGML_CPU_INTERFACE_LINK_LIBRARIES OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
endif()
if (GGML_CPU_HBM)
find_library(memkind memkind REQUIRED)
list(APPEND GGML_CPU_INTERFACE_LINK_LIBRARIES memkind)
endif()
if (GGML_BLAS)
find_package(BLAS REQUIRED)
list(APPEND GGML_CPU_INTERFACE_LINK_LIBRARIES ${BLAS_LIBRARIES})
list(APPEND GGML_CPU_INTERFACE_LINK_OPTIONS ${BLAS_LINKER_FLAGS})
endif()
if (GGML_CUDA)
find_package(CUDAToolkit REQUIRED)
endif()
if (GGML_METAL)
find_library(FOUNDATION_LIBRARY Foundation REQUIRED)
find_library(METAL_FRAMEWORK Metal REQUIRED)
find_library(METALKIT_FRAMEWORK MetalKit REQUIRED)
list(APPEND GGML_METAL_INTERFACE_LINK_LIBRARIES
${FOUNDATION_LIBRARY} ${METAL_FRAMEWORK} ${METALKIT_FRAMEWORK})
endif()
if (GGML_VULKAN)
find_package(Vulkan REQUIRED)
list(APPEND GGML_VULKAN_INTERFACE_LINK_LIBRARIES Vulkan::Vulkan)
endif()
if (GGML_HIP)
find_package(hip REQUIRED)
find_package(hipblas REQUIRED)
find_package(rocblas REQUIRED)
list(APPEND GGML_HIP_INTERFACE_LINK_LIBRARIES hip::host roc::rocblas roc::hipblas)
endif()
if (GGML_SYCL)
find_package(DNNL)
if (${DNNL_FOUND} AND GGML_SYCL_TARGET STREQUAL "INTEL")
list(APPEND GGML_SYCL_INTERFACE_LINK_LIBRARIES DNNL::dnnl)
endif()
if (WIN32)
find_package(IntelSYCL REQUIRED)
find_package(MKL REQUIRED)
list(APPEND GGML_SYCL_INTERFACE_LINK_LIBRARIES IntelSYCL::SYCL_CXX MKL::MKL MKL::MKL_SYCL)
endif()
endif()
endif()
set(_ggml_all_targets "")
if (NOT GGML_BACKEND_DL)
foreach(_ggml_backend ${GGML_AVAILABLE_BACKENDS})
string(REPLACE "-" "_" _ggml_backend_pfx "${_ggml_backend}")
string(TOUPPER "${_ggml_backend_pfx}" _ggml_backend_pfx)
@ -138,6 +174,7 @@ foreach(_ggml_backend ${GGML_AVAILABLE_BACKENDS})
list(APPEND _ggml_all_targets ggml::${_ggml_backend})
endforeach()
endif()
list(APPEND GGML_INTERFACE_LINK_LIBRARIES ggml::ggml-base "${_ggml_all_targets}")
set_target_properties(ggml::ggml
@ -149,4 +186,6 @@ set_target_properties(ggml::all
PROPERTIES
INTERFACE_LINK_LIBRARIES "${_ggml_all_targets}")
endif()
check_required_components(ggml)

27
ggml/include/ggml-opt.h
View File

@ -74,16 +74,26 @@ extern "C" {
GGML_OPT_BUILD_TYPE_OPT = 30,
};
enum ggml_opt_optimizer_type {
GGML_OPT_OPTIMIZER_TYPE_ADAMW,
GGML_OPT_OPTIMIZER_TYPE_SGD,
GGML_OPT_OPTIMIZER_TYPE_COUNT
};
// parameters that control which optimizer is used and how said optimizer tries to find the minimal loss
struct ggml_opt_optimizer_params {
// AdamW optimizer parameters
struct {
float alpha; // learning rate
float beta1;
float beta2;
float beta1; // first AdamW momentum
float beta2; // second AdamW momentum
float eps; // epsilon for numerical stability
float wd; // weight decay for AdamW, use 0.0f to disable
float wd; // weight decay - 0.0f to disable
} adamw;
struct {
float alpha; // learning rate
float wd; // weight decay
} sgd;
};
// callback to calculate optimizer parameters prior to a backward pass
@ -114,6 +124,9 @@ extern "C" {
ggml_opt_get_optimizer_params get_opt_pars; // callback for calculating optimizer parameters
void * get_opt_pars_ud; // userdata for calculating optimizer parameters
// only GGML_OPT_OPTIMIZER_TYPE_ADAMW needs m, v momenta per parameter tensor
enum ggml_opt_optimizer_type optimizer;
};
// get parameters for an optimization context with defaults set where possible
@ -142,6 +155,10 @@ extern "C" {
// get the gradient accumulator for a node from the forward graph
GGML_API struct ggml_tensor * ggml_opt_grad_acc(ggml_opt_context_t opt_ctx, struct ggml_tensor * node);
GGML_API enum ggml_opt_optimizer_type ggml_opt_context_optimizer_type(ggml_opt_context_t); //TODO consistent naming scheme
GGML_API const char * ggml_opt_optimizer_name(enum ggml_opt_optimizer_type);
// ====== Optimization Result ======
GGML_API ggml_opt_result_t ggml_opt_result_init(void);
@ -226,12 +243,14 @@ extern "C" {
struct ggml_tensor * outputs, // output tensor, must have shape [ne_label, ndata_batch] if labels are used
ggml_opt_dataset_t dataset, // dataset with data and optionally also labels
enum ggml_opt_loss_type loss_type, // loss to minimize
enum ggml_opt_optimizer_type optimizer, // sgd or adamw
ggml_opt_get_optimizer_params get_opt_pars, // callback to get optimizer params, userdata is pointer to epoch (of type int64_t)
int64_t nepoch, // how many times the dataset should be iterated over
int64_t nbatch_logical, // datapoints optimizer step, must be a multiple of ndata_batch in inputs/outputs
float val_split, // fraction of the dataset to use for validation, must be in [0.0f, 1.0f)
bool silent); // whether or not info prints to stderr should be suppressed
#ifdef __cplusplus
}
#endif

68
ggml/include/ggml.h
View File

@ -241,6 +241,8 @@
#define GGML_ROPE_TYPE_MROPE 8
#define GGML_ROPE_TYPE_VISION 24
#define GGML_MROPE_SECTIONS 4
#define GGML_UNUSED(x) (void)(x)
#define GGML_PAD(x, n) (((x) + (n) - 1) & ~((n) - 1))
@ -304,6 +306,16 @@
GGML_TENSOR_LOCALS(int64_t, ne, dst, ne) \
GGML_TENSOR_LOCALS(size_t, nb, dst, nb)
#define GGML_TENSOR_TERNARY_OP_LOCALS \
GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) \
GGML_TENSOR_LOCALS(size_t, nb0, src0, nb) \
GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne) \
GGML_TENSOR_LOCALS(size_t, nb1, src1, nb) \
GGML_TENSOR_LOCALS(int64_t, ne2, src2, ne) \
GGML_TENSOR_LOCALS(size_t, nb2, src2, nb) \
GGML_TENSOR_LOCALS(int64_t, ne, dst, ne) \
GGML_TENSOR_LOCALS(size_t, nb, dst, nb)
#define GGML_TENSOR_BINARY_OP_LOCALS01 \
GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) \
GGML_TENSOR_LOCALS(size_t, nb0, src0, nb) \
@ -395,7 +407,8 @@ extern "C" {
// GGML_TYPE_IQ4_NL_4_4 = 36,
// GGML_TYPE_IQ4_NL_4_8 = 37,
// GGML_TYPE_IQ4_NL_8_8 = 38,
GGML_TYPE_COUNT = 39,
GGML_TYPE_MXFP4 = 39, // MXFP4 (1 block)
GGML_TYPE_COUNT = 40,
};
// precision
@ -430,6 +443,7 @@ extern "C" {
GGML_FTYPE_MOSTLY_IQ4_XS = 22, // except 1d tensors
GGML_FTYPE_MOSTLY_IQ1_M = 23, // except 1d tensors
GGML_FTYPE_MOSTLY_BF16 = 24, // except 1d tensors
GGML_FTYPE_MOSTLY_MXFP4 = 25, // except 1d tensors
};
// available tensor operations:
@ -438,6 +452,7 @@ extern "C" {
GGML_OP_DUP,
GGML_OP_ADD,
GGML_OP_ADD_ID,
GGML_OP_ADD1,
GGML_OP_ACC,
GGML_OP_SUB,
@ -527,6 +542,7 @@ extern "C" {
GGML_OP_CROSS_ENTROPY_LOSS,
GGML_OP_CROSS_ENTROPY_LOSS_BACK,
GGML_OP_OPT_STEP_ADAMW,
GGML_OP_OPT_STEP_SGD,
GGML_OP_GLU,
@ -557,6 +573,7 @@ extern "C" {
GGML_GLU_OP_REGLU,
GGML_GLU_OP_GEGLU,
GGML_GLU_OP_SWIGLU,
GGML_GLU_OP_SWIGLU_OAI,
GGML_GLU_OP_GEGLU_ERF,
GGML_GLU_OP_GEGLU_QUICK,
@ -831,6 +848,13 @@ extern "C" {
struct ggml_tensor * b,
enum ggml_type type);
// dst[i0, i1, i2] = a[i0, i1, i2] + b[i0, ids[i1, i2]]
GGML_API struct ggml_tensor * ggml_add_id(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
struct ggml_tensor * ids);
GGML_API struct ggml_tensor * ggml_add1(
struct ggml_context * ctx,
struct ggml_tensor * a,
@ -1198,6 +1222,13 @@ extern "C" {
struct ggml_tensor * a,
struct ggml_tensor * b);
GGML_API struct ggml_tensor * ggml_swiglu_oai(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
float alpha,
float limit);
// normalize along rows
GGML_API struct ggml_tensor * ggml_norm(
struct ggml_context * ctx,
@ -1570,6 +1601,10 @@ extern "C" {
float scale,
float max_bias);
GGML_API void ggml_soft_max_add_sinks(
struct ggml_tensor * a,
struct ggml_tensor * sinks);
GGML_API struct ggml_tensor * ggml_soft_max_ext_back(
struct ggml_context * ctx,
struct ggml_tensor * a,
@ -1628,7 +1663,7 @@ extern "C" {
struct ggml_tensor * b,
struct ggml_tensor * c,
int n_dims,
int sections[4],
int sections[GGML_MROPE_SECTIONS],
int mode,
int n_ctx_orig,
float freq_base,
@ -1654,6 +1689,22 @@ extern "C" {
float beta_fast,
float beta_slow);
GGML_API struct ggml_tensor * ggml_rope_multi_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
struct ggml_tensor * c,
int n_dims,
int sections[GGML_MROPE_SECTIONS],
int mode,
int n_ctx_orig,
float freq_base,
float freq_scale,
float ext_factor,
float attn_factor,
float beta_fast,
float beta_slow);
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_rope_custom(
struct ggml_context * ctx,
struct ggml_tensor * a,
@ -2052,6 +2103,10 @@ extern "C" {
GGML_API enum ggml_prec ggml_flash_attn_ext_get_prec(
const struct ggml_tensor * a);
GGML_API void ggml_flash_attn_ext_add_sinks(
struct ggml_tensor * a,
struct ggml_tensor * sinks);
// TODO: needs to be adapted to ggml_flash_attn_ext
GGML_API struct ggml_tensor * ggml_flash_attn_back(
struct ggml_context * ctx,
@ -2257,7 +2312,14 @@ extern "C" {
struct ggml_tensor * grad,
struct ggml_tensor * m,
struct ggml_tensor * v,
struct ggml_tensor * adamw_params); // parameters such a the learning rate
struct ggml_tensor * adamw_params); // parameters such as the learning rate
// stochastic gradient descent step (with weight decay)
GGML_API struct ggml_tensor * ggml_opt_step_sgd(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * grad,
struct ggml_tensor * sgd_params); // alpha, weight decay
//
// automatic differentiation

11
ggml/src/CMakeLists.txt
View File

@ -214,6 +214,13 @@ add_library(ggml
ggml-backend-reg.cpp)
add_library(ggml::ggml ALIAS ggml)
if (GGML_BACKEND_DIR)
if (NOT GGML_BACKEND_DL)
message(FATAL_ERROR "GGML_BACKEND_DIR requires GGML_BACKEND_DL")
endif()
target_compile_definitions(ggml PUBLIC GGML_BACKEND_DIR="${GGML_BACKEND_DIR}")
endif()
target_link_libraries(ggml PUBLIC ggml-base)
if (CMAKE_SYSTEM_NAME MATCHES "Linux")
@ -227,7 +234,11 @@ function(ggml_add_backend_library backend)
set_target_properties(${backend} PROPERTIES LIBRARY_OUTPUT_DIRECTORY ${CMAKE_RUNTIME_OUTPUT_DIRECTORY})
target_compile_definitions(${backend} PRIVATE GGML_BACKEND_DL)
add_dependencies(ggml ${backend})
if (GGML_BACKEND_DIR)
install(TARGETS ${backend} LIBRARY DESTINATION ${GGML_BACKEND_DIR})
else()
install(TARGETS ${backend} LIBRARY DESTINATION ${CMAKE_INSTALL_BINDIR})
endif()
else()
add_library(${backend} ${ARGN})
target_link_libraries(ggml PUBLIC ${backend})

1
ggml/src/ggml-alloc.c
View File

@ -29,6 +29,7 @@ static bool ggml_op_can_inplace(enum ggml_op op) {
case GGML_OP_DIAG_MASK_ZERO:
case GGML_OP_DIAG_MASK_INF:
case GGML_OP_ADD:
case GGML_OP_ADD_ID:
case GGML_OP_ADD1:
case GGML_OP_SUB:
case GGML_OP_MUL:

3
ggml/src/ggml-backend-reg.cpp
View File

@ -498,6 +498,9 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
std::vector<fs::path> search_paths;
if (user_search_path == nullptr) {
#ifdef GGML_BACKEND_DIR
search_paths.push_back(fs::u8path(GGML_BACKEND_DIR));
#endif
// default search paths: executable directory, current directory
search_paths.push_back(get_executable_path());
search_paths.push_back(fs::current_path());

22
ggml/src/ggml-backend.cpp
View File

@ -647,6 +647,7 @@ struct ggml_backend_sched {
// pipeline parallelism support
int n_copies;
int cur_copy;
int next_copy;
ggml_backend_event_t events[GGML_SCHED_MAX_BACKENDS][GGML_SCHED_MAX_COPIES];
struct ggml_tensor * graph_inputs[GGML_SCHED_MAX_SPLIT_INPUTS];
int n_graph_inputs;
@ -1070,6 +1071,11 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
}
}
}
// if the node is still unassigned, assign it to the first backend that supports it
for (int b = 0; b < sched->n_backends && *cur_backend_id == -1; b++) {
ggml_backend_sched_set_if_supported(sched, node, b, cur_backend_id);
}
GGML_ASSERT(*cur_backend_id != -1);
}
// pass 5: split graph, find tensors that need to be copied
@ -1097,7 +1103,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
const int node_backend_id = tensor_backend_id(node);
assert(node_backend_id != -1); // all nodes should be assigned by now, this can happen if there is no CPU fallback
GGML_ASSERT(node_backend_id != -1); // all nodes should be assigned by now, this can happen if there is no CPU fallback
// check if we should start a new split based on the sources of the current node
bool need_new_split = false;
@ -1155,7 +1161,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
size_t src_id = hash_id(src);
const int src_backend_id = sched->hv_tensor_backend_ids[src_id];
assert(src_backend_id != -1); // all inputs should be assigned by now
GGML_ASSERT(src_backend_id != -1); // all inputs should be assigned by now
if (src->flags & GGML_TENSOR_FLAG_INPUT && sched->n_copies > 1) {
if (tensor_id_copy(src_id, src_backend_id, 0) == NULL) {
@ -1433,8 +1439,6 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
}
}
sched->cur_copy = (sched->cur_copy + 1) % sched->n_copies;
return GGML_STATUS_SUCCESS;
}
@ -1535,10 +1539,10 @@ void ggml_backend_sched_reset(ggml_backend_sched_t sched) {
bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph) {
GGML_ASSERT((int)sched->hash_set.size >= measure_graph->n_nodes + measure_graph->n_leafs);
ggml_backend_sched_split_graph(sched, measure_graph);
ggml_backend_sched_synchronize(sched);
ggml_backend_sched_split_graph(sched, measure_graph);
if (!ggml_gallocr_reserve_n(sched->galloc, &sched->graph, sched->node_backend_ids, sched->leaf_backend_ids)) {
return false;
}
@ -1550,6 +1554,10 @@ bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph *
bool ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
GGML_ASSERT((int)sched->hash_set.size >= graph->n_nodes + graph->n_leafs);
GGML_ASSERT(!sched->is_alloc);
sched->cur_copy = sched->next_copy;
sched->next_copy = (sched->next_copy + 1) % sched->n_copies;
ggml_backend_sched_split_graph(sched, graph);
@ -1590,7 +1598,7 @@ void ggml_backend_sched_synchronize(ggml_backend_sched_t sched) {
// if the graph is not already allocated, always use copy 0 after a synchronization
// this ensures that during generation the same copy is used every time,
// which avoids changes in the graph that could cause CUDA or other graphs to be disabled
sched->cur_copy = 0;
sched->next_copy = 0;
}
}

2
ggml/src/ggml-blas/ggml-blas.cpp
View File

@ -282,7 +282,7 @@ ggml_backend_t ggml_backend_blas_init(void) {
ggml_backend_t backend = new ggml_backend {
/* .guid = */ ggml_backend_blas_guid(),
/* .interface = */ blas_backend_i,
/* .iface = */ blas_backend_i,
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_blas_reg(), 0),
/* .context = */ ctx,
};

14
ggml/src/ggml-cann/CMakeLists.txt
View File

@ -31,6 +31,13 @@ string(REGEX MATCH "[0-9]+[a-zA-Z]" SOC_TYPE_MAJOR_SN "${SOC_VERSION}")
set(SOC_TYPE_COMPILE_OPTION "ASCEND_${SOC_TYPE_MAJOR_SN}")
string(TOUPPER ${SOC_TYPE_COMPILE_OPTION} SOC_TYPE_COMPILE_OPTION)
message(STATUS "CANN: SOC_VERSION = ${SOC_VERSION}")
option(USE_ACL_GRAPH "Enable CANN graph execution (ACL graph mode)" OFF)
if(USE_ACL_GRAPH AND (SOC_TYPE_MAJOR_SN STREQUAL "310P" OR SOC_TYPE_COMPILE_OPTION STREQUAL "ASCEND_310P"))
message(FATAL_ERROR
"CANN Graph (ACL graph mode) is not supported on 310P devices. "
"Please build with -DUSE_ACL_GRAPH=OFF or use a supported SOC.")
endif()
if (CANN_INSTALL_DIR)
# Only Support Linux.
@ -68,6 +75,13 @@ if (CANN_INSTALL_DIR)
target_compile_definitions(ggml-cann PRIVATE "-D${SOC_TYPE_COMPILE_OPTION}")
if (USE_ACL_GRAPH)
target_compile_definitions(ggml-cann PRIVATE USE_ACL_GRAPH)
message(STATUS "CANN: USE_ACL_GRAPH is enabled.")
else()
message(STATUS "CANN: USE_ACL_GRAPH is disabled.")
endif()
message(STATUS "CANN: CANN_INCLUDE_DIRS = ${CANN_INCLUDE_DIRS}")
message(STATUS "CANN: CANN_LIBRARIES = ${CANN_LIBRARIES}")
else()

4
ggml/src/ggml-cann/acl_tensor.cpp
View File

@ -77,6 +77,8 @@ aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne,
for (int i = 0; i < final_dims; i++) {
acl_storage_len += (acl_ne[i] - 1) * acl_stride[i];
}
size_t elem_offset = offset / ggml_element_size(tensor);
acl_storage_len += elem_offset;
// Reverse ne and stride.
std::reverse(acl_ne, acl_ne + final_dims);
@ -84,7 +86,7 @@ aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne,
aclTensor* acl_tensor = aclCreateTensor(
acl_ne, final_dims, ggml_cann_type_mapping(tensor->type), acl_stride,
offset / ggml_element_size(tensor), format, &acl_storage_len, 1,
elem_offset, format, &acl_storage_len, 1,
tensor->data);
return acl_tensor;

816
ggml/src/ggml-cann/aclnn_ops.cpp
View File

@ -68,6 +68,8 @@
#include <aclnnop/aclnn_grouped_matmul_v3.h>
#include <aclnnop/aclnn_fused_infer_attention_score_v2.h>
#include <aclnnop/aclnn_zero.h>
#include <aclnnop/aclnn_index_copy.h>
#include <aclnnop/aclnn_index_select.h>
#include <float.h>
#include <cmath>
@ -99,7 +101,7 @@ void bcast_shape(ggml_tensor * src0, ggml_tensor * src1, ggml_tensor * dst, aclT
}
}
void ggml_cann_unary_op(
void ggml_cann_op_unary(
std::function<void(ggml_backend_cann_context&, aclTensor*, aclTensor*)> unary_op,
ggml_backend_cann_context& ctx, ggml_tensor* dst) {
ggml_tensor* src = dst->src[0];
@ -111,6 +113,42 @@ void ggml_cann_unary_op(
ggml_cann_release_resources(ctx, acl_src, acl_dst);
}
void ggml_cann_op_unary_gated(
std::function<void(ggml_backend_cann_context&, aclTensor*, aclTensor*)> unary_op,
ggml_backend_cann_context& ctx, ggml_tensor* dst) {
ggml_tensor* src0 = dst->src[0];
ggml_tensor* src1 = dst->src[1];
GGML_ASSERT(ggml_is_contiguous_1(src0));
GGML_ASSERT(ggml_is_contiguous_1(dst));
const int32_t swapped = ggml_get_op_params_i32(dst, 1);
aclTensor* acl_dst = ggml_cann_create_tensor(dst);
aclTensor *acl_src0 = nullptr, *acl_src1 = nullptr;
if(src1) {
GGML_ASSERT(ggml_is_contiguous_1(src1));
GGML_ASSERT(src0->type == src1->type);
acl_src0 = ggml_cann_create_tensor(src0);
acl_src1 = ggml_cann_create_tensor(src1);
} else {
int64_t ne[] = {src0->ne[0] / 2, src0->ne[1], src0->ne[2], src0->ne[3]};
size_t nb[] = {src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3]};
acl_src0 = ggml_cann_create_tensor(src0, ne, nb, GGML_MAX_DIMS, ACL_FORMAT_ND, 0);
acl_src1 = ggml_cann_create_tensor(src0, ne, nb, GGML_MAX_DIMS, ACL_FORMAT_ND, ne[0] * ggml_element_size(src0));
if (swapped) {
std::swap(acl_src0, acl_src1);
}
}
unary_op(ctx, acl_src0, acl_dst);
GGML_CANN_CALL_ACLNN_OP(ctx, InplaceMul, acl_dst, acl_src1);
ggml_cann_release_resources(ctx, acl_src0, acl_dst);
if(src1)
ggml_cann_release_resources(ctx, acl_src1);
}
/**
* @brief Repeats elements of a tensor along each dimension according to the
* specified repeat array.
@ -715,69 +753,55 @@ 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* src0 = dst->src[0];
if (ggml_are_same_shape(src0, dst)) {
aclTensor* acl_src = ggml_cann_create_tensor(src0);
aclTensor* acl_dst = ggml_cann_create_tensor(dst);
if (ggml_are_same_shape(src0, dst)) {
if (dst->type == src0->type) {
cann_copy(ctx, acl_src, acl_dst);
} else {
aclnn_cast(ctx, acl_src, acl_dst, ggml_cann_type_mapping(dst->type));
}
} else {
if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst)) {
if (dst->type == src0->type) {
size_t cpy_size = ggml_nbytes(dst);
ggml_cann_async_memcpy(ctx, dst->data, src0->data, cpy_size,
ACL_MEMCPY_DEVICE_TO_DEVICE);
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, ggml_cann_type_mapping(dst->type));
size_t cpy_size = ggml_nbytes(dst);
ggml_cann_async_memcpy(ctx, dst->data, src_trans_buffer, cpy_size,
ACL_MEMCPY_DEVICE_TO_DEVICE);
ggml_cann_release_resources(ctx, 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, ggml_cann_type_mapping(dst->type));
size_t cpy_size = ggml_nbytes(dst);
ggml_cann_async_memcpy(ctx, dst->data, src_trans_buffer, cpy_size,
ACL_MEMCPY_DEVICE_TO_DEVICE);
ggml_cann_release_resources(ctx, src_trans_tensor);
return;
} else {
GGML_ABORT("Unsupport dst is not tontiguous.");
}
}
ggml_cann_release_resources(ctx, acl_src, acl_dst);
} else {
void* src_trans_buffer = src0->data;
ggml_cann_pool_alloc src_buffer_allocator;
if (!ggml_is_contiguous(src0)) {
aclTensor* acl_src = ggml_cann_create_tensor(src0);
src_buffer_allocator.alloc(ctx.pool(),
ggml_nelements(src0) * ggml_type_size(src0->type));
src_trans_buffer = src_buffer_allocator.get();
size_t src_trans_nb[GGML_MAX_DIMS];
src_trans_nb[0] = ggml_type_size(src0->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(src0->type),
ggml_type_size(src0->type), src0->ne, src_trans_nb,
GGML_MAX_DIMS);
cann_copy(ctx, acl_src, src_trans_tensor);
ggml_cann_release_resources(ctx, acl_src, src_trans_tensor);
}
size_t src_reshape_nb[GGML_MAX_DIMS];
src_reshape_nb[0] = ggml_type_size(src0->type);
for (int i = 1; i < GGML_MAX_DIMS; i++) {
src_reshape_nb[i] = src_reshape_nb[i - 1] * dst->ne[i - 1];
}
aclTensor* trans_acl_src = ggml_cann_create_tensor(src_trans_buffer,
ggml_cann_type_mapping(src0->type),ggml_type_size(src0->type),
dst->ne, src_reshape_nb, GGML_MAX_DIMS, ACL_FORMAT_ND);
aclTensor* acl_dst = ggml_cann_create_tensor(dst);
if (dst->type == src0->type) {
cann_copy(ctx, trans_acl_src, acl_dst);
} else {
aclnn_cast(ctx, trans_acl_src, acl_dst, ggml_cann_type_mapping(dst->type));
}
ggml_cann_release_resources(ctx, trans_acl_src, acl_dst);
}
return;
}
/**
@ -1292,157 +1316,193 @@ static void aclnn_pow_tensor_tensor(ggml_backend_cann_context& ctx,
}
/**
* @brief Applies the Alibi (Attention with Linear Biases) mechanism to the
* @details This function implements the Alibi mechanism, which introduces
* learnable biases into the attention scores to simulate relative
* position encoding without the need for explicit positional
* embeddings.
* @brief Generate a range of values and apply a scalar base exponentiation.
*
* @param ctx The backend CANN context for executing operations.
* @param acl_src The source tensor representing the query or key.
* @param acl_position The position tensor containing relative positions.
* @param acl_dst The destination tensor where the result will be stored.
* @param n_head The number of attention heads.
* @param src_ne The dimensions of the source tensor.
* @param src_nb0 The byte size of the first dimension of the source
tensor.
* @param max_bias The maximum bias value used in the Alibi mechanism.
* @param dst The destination tensor object for additional metadata.
* This function creates an evenly spaced sequence from `start` to `stop` (exclusive),
* with step size `step`, stores it in a temporary buffer, and then computes:
*
* The function performs the following steps:
* 1. Calculates the logarithm floor of the number of heads to determine the
base for bias calculation.
* 2. Initializes arrays with arithmetic sequences and fills them with bias
values.
* 3. Computes the bias tensor based on the calculated biases and arithmetic
sequences.
* 4. Reshapes the bias tensor to match the dimensions of the input tensors.
* 5. Multiplies the position tensor by the bias tensor.
* 6. Adds the result of the multiplication to the source tensor to produce the
final output.
* @f[
* slope[i] = m^{\left( start + i \cdot step \right)}, \quad 0 \le i < size
* @f]
*
* The results are written to the provided @p slope_buffer.
*
* @param ctx CANN backend context for memory allocation and operator execution.
* @param slope_buffer Pointer to the output buffer (float array) for the computed slope values.
* @param m Scalar base for the exponentiation.
* @param size Number of elements in the generated sequence.
* @param start Starting exponent offset.
* @param stop Stopping exponent offset (exclusive).
* @param step Step size for the exponent increment.
*/
static void aclnn_alibi(ggml_backend_cann_context& ctx, aclTensor* acl_src,
aclTensor* acl_position, aclTensor* acl_dst,
const int n_head, int64_t* src_ne, const size_t src_nb0,
float max_bias, ggml_tensor* dst) {
const int64_t ne2_ne3 = src_ne[2] * src_ne[3];
GGML_ASSERT(src_nb0 == sizeof(float));
GGML_ASSERT(n_head == src_ne[2]);
static void aclnn_get_slope_inner(ggml_backend_cann_context& ctx, void* slope_buffer,
float m, int64_t size, float start, float stop, float step){
int64_t ne[] = {size};
size_t nb[] = {sizeof(float)};
const int n_heads_log2_floor = 1u << (uint32_t)floor(log2(n_head));
ggml_cann_pool_alloc arange_allocator(ctx.pool(), size * sizeof(float));
void* arange_buffer = arange_allocator.get();
float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor);
float m1 = powf(2.0f, -(max_bias / 2.0f) / n_heads_log2_floor);
aclTensor* arange_tensor = ggml_cann_create_tensor(
arange_buffer, ACL_FLOAT, sizeof(float), ne, nb, 1);
aclnn_arange(ctx, arange_tensor, start, stop, step, size);
// init arange
ggml_cann_pool_alloc arange_allocator(ctx.pool(),
ne2_ne3 * ggml_type_size(dst->type));
void* tmp_arange_buffer = arange_allocator.get();
aclTensor* slope_tensor = ggml_cann_create_tensor(
slope_buffer, ACL_FLOAT, sizeof(float), ne, nb, 1);
// arange1: [1, ..., n_heads_log2_floor+1)
float start = 1;
float stop = n_heads_log2_floor + 1;
aclScalar* sc = aclCreateScalar(&m, aclDataType::ACL_FLOAT);
GGML_CANN_CALL_ACLNN_OP(ctx, PowScalarTensor, sc, arange_tensor, slope_tensor);
ggml_cann_release_resources(ctx, sc, arange_tensor, slope_tensor);
}
/**
* @brief Compute slope values for multiple attention heads based on ALiBi bias parameters.
*
* This function generates slope values for each attention head according to the ALiBi
* (Attention with Linear Biases) method. It splits the computation into two ranges depending
* on whether the head index is less than @p n_head_log2 or not, and uses different base values
* (`m0` and `m1`) for the exponentiation.
*
* @f[
* slope[h] =
* \begin{cases}
* m_0^{(h + 1)}, & h < n\_head\_log2 \\
* m_1^{\left( 2 \cdot (h - n\_head\_log2) + 1 \right)}, & h \geq n\_head\_log2
* \end{cases}
* \quad , \quad \text{if } max\_bias > 0
* @f]
*
* If @p max_bias <= 0, all slope values are set to 1.0.
*
* @param ctx CANN backend context for memory allocation and operator execution.
* @param n_head Total number of attention heads.
* @param slope_buffer Pointer to the output buffer (float array) for storing slopes.
* @param max_bias Maximum bias value for slope computation.
*
*/
static void aclnn_get_slope(ggml_backend_cann_context & ctx, int64_t n_head,
void* slope_buffer, float max_bias) {
const int n_head_log2 = 1u << (uint32_t) floor(log2(n_head));
float m0 = powf(2.0f, -(max_bias) / n_head_log2);
float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
// const float slope = (max_bias > 0.0f) ?
// h < n_head_log2 ?
// powf(m0, h + 1) :
// powf(m1, 2*(h - n_head_log2) + 1) :
// 1.0f;
// arange1
float start = 0 + 1;
float end = (n_head_log2 - 1) + 1;
float step = 1;
int64_t n_elements_arange = n_heads_log2_floor;
int64_t tmp_arange1_ne[] = {n_heads_log2_floor};
size_t tmp_arange1_nb[] = {sizeof(dst->type)};
aclTensor* tmp_arange1_tensor = ggml_cann_create_tensor(
tmp_arange_buffer, ggml_cann_type_mapping(dst->type),
ggml_type_size(dst->type), tmp_arange1_ne, tmp_arange1_nb,
GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
aclnn_arange(ctx, tmp_arange1_tensor, start, stop, step, n_elements_arange);
aclTensor* tmp_arange2_tensor = nullptr;
if (n_heads_log2_floor < ne2_ne3) {
// arange2: [1, ..., 2 * (k - n_heads_log2_floor) + 1)
start = 1;
stop = 2 * (ne2_ne3 - n_heads_log2_floor) + 1;
float count = n_head_log2;
// end needs to be +1 because aclnn uses a left-closed, right-open interval.
aclnn_get_slope_inner(ctx, slope_buffer, m0, count, start, end + 1, step);
if (n_head_log2 < n_head) {
// arange2
start = 2 * (n_head_log2 - n_head_log2) + 1;
end = 2 * ((n_head - 1) - n_head_log2) + 1;
step = 2;
n_elements_arange = ne2_ne3 - n_heads_log2_floor;
int64_t tmp_arange2_ne[] = {ne2_ne3 - n_heads_log2_floor};
size_t tmp_arange2_nb[] = {sizeof(dst->type)};
aclTensor* tmp_arange2_tensor = ggml_cann_create_tensor(
(char*)tmp_arange_buffer +
n_heads_log2_floor * ggml_type_size(dst->type),
ggml_cann_type_mapping(dst->type), ggml_type_size(dst->type),
tmp_arange2_ne, tmp_arange2_nb, GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
aclnn_arange(ctx, tmp_arange2_tensor, start, stop, step,
n_elements_arange);
count = n_head - n_head_log2;
aclnn_get_slope_inner(
ctx, (char *) slope_buffer + n_head_log2 * sizeof(float),
m1, count, start, end + 1, step);
}
}
// init mk_base
ggml_cann_pool_alloc mk_base_allocator(ctx.pool(),
ne2_ne3 * ggml_type_size(dst->type));
void* tmp_mk_base_buffer = mk_base_allocator.get();
int64_t tmp_mk_base1_ne[] = {n_heads_log2_floor};
size_t tmp_mk_base1_nb[] = {sizeof(dst->type)};
aclTensor* tmp_mk_base1_tensor = ggml_cann_create_tensor(
tmp_mk_base_buffer, ggml_cann_type_mapping(dst->type),
ggml_type_size(dst->type), tmp_mk_base1_ne, tmp_mk_base1_nb,
GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
/**
* @brief Add ALiBi (Attention with Linear Biases) positional biases to the attention mask.
*
* This function computes the ALiBi slopes for each attention head (if max_bias > 0),
* multiplies them with the attention mask to produce bias tensors, and adds these biases
* to the destination tensor (@p dst).
*
* The function performs necessary broadcasting of the mask and slope tensors to match
* the shape of the destination tensor, then applies element-wise multiplication and addition
* using CANN operators.
*
* @param ctx CANN backend context for memory management and operator execution.
* @param mask Input attention mask tensor, assumed to be contiguous.
* @param dst Destination tensor to which ALiBi biases will be added.
* @param dst_ptr Pointer to the memory of the destination tensor.
* @param max_bias Maximum bias value controlling the slope scaling.
*
* @note
* - Write data into dst_ptr using only the shape information of the dst tensor.
* - `GGML_MAX_DIMS + 2` is used to extend tensor dimensions for broadcasting.
*/
static void aclnn_add_alibi(ggml_backend_cann_context& ctx, ggml_tensor* mask,
ggml_tensor* dst, void* dst_ptr, float max_bias) {
void* slope_buffer = nullptr;
void* bias_buffer = nullptr;
aclnn_fill_scalar(ctx, m0, tmp_mk_base1_tensor);
aclTensor* tmp_mk_base2_tensor = nullptr;
if (n_heads_log2_floor < ne2_ne3) {
int64_t tmp_mk_base2_ne[] = {ne2_ne3 - n_heads_log2_floor};
size_t tmp_mk_base2_nb[] = {sizeof(dst->type)};
aclTensor* tmp_mk_base2_tensor = ggml_cann_create_tensor(
(char*)tmp_mk_base_buffer +
n_heads_log2_floor * ggml_type_size(dst->type),
ggml_cann_type_mapping(dst->type), ggml_type_size(dst->type),
tmp_mk_base2_ne, tmp_mk_base2_nb, GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
aclnn_fill_scalar(ctx, m1, tmp_mk_base2_tensor);
if (max_bias > 0.0f) {
int64_t n_heads = dst->ne[2];
ggml_cann_pool_alloc slope_allocator(ctx.pool(), n_heads * sizeof(float));
slope_buffer = slope_allocator.get();
ggml_cann_pool_alloc bias_allocator(
ctx.pool(), ggml_nelements(dst) * ggml_element_size(dst));
bias_buffer = bias_allocator.get();
aclnn_get_slope(ctx, n_heads, slope_buffer, max_bias);
}
// init mk
int64_t tmp_mk_base_ne[] = {ne2_ne3};
size_t tmp_mk_base_nb[] = {sizeof(dst->type)};
aclTensor* tmp_mk_base_tensor = ggml_cann_create_tensor(
tmp_mk_base_buffer, ggml_cann_type_mapping(dst->type),
ggml_type_size(dst->type), tmp_mk_base_ne, tmp_mk_base_nb,
GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
aclTensor* tmp_arange_tensor = ggml_cann_create_tensor(
tmp_arange_buffer, ggml_cann_type_mapping(dst->type),
ggml_type_size(dst->type), tmp_mk_base_ne, tmp_mk_base_nb,
GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
aclnn_pow_tensor_tensor(ctx, tmp_mk_base_tensor, tmp_arange_tensor);
// broadcast for mask, slop and dst;
int64_t nr2 = dst->ne[2] / mask->ne[2];
int64_t nr3 = dst->ne[3] / mask->ne[3];
// reshape mk
int64_t tmp_mk_ne[] = {1, 1, src_ne[2], src_ne[3]};
size_t tmp_mk_nb[GGML_MAX_DIMS];
tmp_mk_nb[0] = ggml_type_size(dst->type);
for (int i = 1; i < GGML_MAX_DIMS; i++) {
tmp_mk_nb[i] = tmp_mk_nb[i - 1] * tmp_mk_ne[i - 1];
// broadcast the mask across rows
int64_t mask_ne[] = { mask->ne[0], dst->ne[1], mask->ne[2], 1, mask->ne[3], 1 };
size_t mask_nb[] = {
mask_nb[0] = mask->nb[0], mask_nb[1] = mask->nb[1], mask_nb[2] = mask->nb[2],
mask_nb[3] = mask->nb[2], mask_nb[4] = mask->nb[3], mask_nb[5] = mask->nb[3]
};
int64_t dst_ne[] = { dst->ne[0], dst->ne[1], mask->ne[2], nr2, mask->ne[3], nr3 };
size_t dst_nb[] = {
dst_nb[0] = dst->nb[0], dst_nb[1] = dst->nb[1], dst_nb[2] = dst->nb[2],
dst_nb[3] = dst->nb[2], dst_nb[4] = dst->nb[3], dst_nb[5] = dst->nb[3]
};
// slope is a 1 dim tensor, slope.ne2 == dst.ne2
int64_t slope_ne[] = { 1, 1, mask->ne[2], nr2, 1, 1 };
size_t slope_nb[GGML_MAX_DIMS + 2];
slope_nb[0] = sizeof(float);
for (int i = 1; i < GGML_MAX_DIMS + 2; i++) {
slope_nb[i] = slope_nb[i - 1] * slope_ne[i - 1];
}
aclTensor* tmp_mk_tensor = ggml_cann_create_tensor(
tmp_mk_base_buffer, ggml_cann_type_mapping(dst->type),
ggml_type_size(dst->type), tmp_mk_ne, tmp_mk_nb, GGML_MAX_DIMS,
ACL_FORMAT_ND);
// acl_position * mk
int64_t tmp_output_ne[] = {src_ne[0], src_ne[1], src_ne[2], src_ne[3]};
size_t tmp_output_nb[GGML_MAX_DIMS];
tmp_output_nb[0] = ggml_type_size(dst->type);
for (int i = 1; i < GGML_MAX_DIMS; i++) {
tmp_output_nb[i] = tmp_output_nb[i - 1] * tmp_output_ne[i - 1];
aclTensor* acl_slope = ggml_cann_create_tensor(
slope_buffer, ACL_FLOAT, sizeof(float),
slope_ne, slope_nb, GGML_MAX_DIMS + 2);
aclTensor* acl_mask = ggml_cann_create_tensor(
mask, mask_ne, mask_nb, GGML_MAX_DIMS + 2);
// write data into dst_ptr using only the shape information of the dst tensor.
aclTensor* acl_dst = ggml_cann_create_tensor(
dst_ptr, ggml_cann_type_mapping(dst->type),
ggml_type_size(dst->type), dst_ne, dst_nb,
GGML_MAX_DIMS + 2);
if (max_bias > 0.0f) {
int64_t bias_ne[] = { mask->ne[0], dst->ne[1], mask->ne[2], nr2, mask->ne[3], 1 };
size_t bias_nb[GGML_MAX_DIMS + 2];
bias_nb[0] = sizeof(float);
for (int i = 1; i < GGML_MAX_DIMS + 2; i++) {
bias_nb[i] = bias_nb[i - 1] * bias_ne[i - 1];
}
ggml_cann_pool_alloc output_allocator(ctx.pool(), ggml_nbytes(dst));
void* tmp_output_buffer = output_allocator.get();
aclTensor* tmp_output_tensor = ggml_cann_create_tensor(
tmp_output_buffer, ggml_cann_type_mapping(dst->type),
ggml_type_size(dst->type), tmp_output_ne, tmp_output_nb, GGML_MAX_DIMS,
ACL_FORMAT_ND);
aclnn_mul(ctx, acl_position, tmp_mk_tensor, tmp_output_tensor);
aclTensor* bias_tensor = ggml_cann_create_tensor(
bias_buffer, ACL_FLOAT, sizeof(float),
bias_ne, bias_nb, GGML_MAX_DIMS + 2);
// add
aclnn_add(ctx, tmp_output_tensor, acl_src, acl_dst);
ggml_cann_release_resources(ctx, tmp_arange1_tensor, tmp_arange2_tensor,
tmp_mk_base1_tensor, tmp_mk_base2_tensor, tmp_mk_base_tensor,
tmp_arange_tensor, tmp_mk_tensor, tmp_output_tensor);
aclnn_mul(ctx, acl_slope, acl_mask, bias_tensor);
aclnn_add(ctx, acl_dst, bias_tensor);
ggml_cann_release_resources(ctx, bias_tensor);
} else {
aclnn_add(ctx, acl_dst, acl_mask);
}
ggml_cann_release_resources(ctx, acl_slope, acl_mask, acl_dst);
}
void ggml_cann_cpy(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
@ -1463,8 +1523,8 @@ void ggml_cann_cpy(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
* @param acl_dst The destination tensor where the softmax results will be
* stored.
*/
static void aclnn_softmax(ggml_backend_cann_context& ctx, aclTensor* acl_src,
int64_t dim, aclTensor* acl_dst) {
static void aclnn_softmax(ggml_backend_cann_context & ctx,
aclTensor* acl_src, int64_t dim, aclTensor * acl_dst) {
GGML_CANN_CALL_ACLNN_OP(ctx, Softmax, acl_src, dim, acl_dst);
}
@ -1483,145 +1543,115 @@ void ggml_cann_softmax(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
// input mul scale
aclScalar* acl_scale = aclCreateScalar(&scale, aclDataType::ACL_FLOAT);
ggml_cann_pool_alloc src_tensor_allocator(ctx.pool(), ggml_nbytes(src0));
void* src_tensor_buffer = src_tensor_allocator.get();
aclTensor* softmax_tensor = ggml_cann_create_tensor(
src_tensor_buffer, ggml_cann_type_mapping(src0->type),
ggml_element_size(src0), src0->ne, src0->nb,GGML_MAX_DIMS);
size_t n_bytes = ggml_nbytes(src0);
ggml_cann_pool_alloc mul_scale_allocator(ctx.pool(), n_bytes);
void* input_mul_scale_buffer = mul_scale_allocator.get();
aclTensor* acl_input_mul_scale_tensor = ggml_cann_create_tensor(
input_mul_scale_buffer, ACL_FLOAT, ggml_type_size(src0->type), src0->ne,
src0->nb, GGML_MAX_DIMS);
bool inplace = false;
aclnn_muls(ctx, acl_src0, scale, acl_input_mul_scale_tensor, inplace);
aclnn_muls(ctx, acl_src0, scale, softmax_tensor, false);
// mask
aclTensor* acl_src1_fp32_tensor = nullptr;
aclTensor* tmp_mask_tensor = nullptr;
ggml_cann_pool_alloc src1_fp32_allocator(ctx.pool());
if (src1) {
const bool use_f16 = src1->type == GGML_TYPE_F16;
if (use_f16) {
// cast to fp32
size_t n_bytes = ggml_nelements(src1) * sizeof(float_t);
size_t src1_fp32_nb[GGML_MAX_DIMS];
src1_fp32_nb[0] = sizeof(float_t);
for (int i = 1; i < GGML_MAX_DIMS; i++) {
src1_fp32_nb[i] = src1_fp32_nb[i - 1] * src1->ne[i - 1];
aclnn_add_alibi(ctx, src1, src0, src_tensor_buffer, max_bias);
}
src1_fp32_allocator.alloc(n_bytes);
void* src1_fp32_buffer = src1_fp32_allocator.get();
acl_src1_fp32_tensor = ggml_cann_create_tensor(
src1_fp32_buffer, ACL_FLOAT, sizeof(float), src1->ne,
src1_fp32_nb, GGML_MAX_DIMS);
aclTensor* acl_src1 = ggml_cann_create_tensor(src1);
aclnn_cast(ctx, acl_src1, acl_src1_fp32_tensor, ACL_FLOAT);
ggml_cann_release_resources(ctx, acl_src1);
} else {
acl_src1_fp32_tensor = ggml_cann_create_tensor(src1);
}
// broadcast the mask across rows, only use ne11 of ne01 in mask
if (src1->ne[1] != src0->ne[1]) {
// mask shape: [1,1,ne11,ne10]
int64_t tmp_mask_ne[] = {src0->ne[0], src0->ne[1], 1, 1};
size_t tmp_mask_nb[GGML_MAX_DIMS];
tmp_mask_nb[0] = sizeof(float_t);
for (int i = 1; i < GGML_MAX_DIMS; i++) {
tmp_mask_nb[i] = tmp_mask_nb[i - 1] * tmp_mask_ne[i - 1];
}
tmp_mask_tensor = ggml_cann_create_tensor(
src1->data, ACL_FLOAT, sizeof(float), tmp_mask_ne, tmp_mask_nb,
GGML_MAX_DIMS, ACL_FORMAT_ND);
}
// alibi
const int n_head = src0->ne[2];
const size_t src_nb0 = src0->nb[0];
n_bytes = ggml_nbytes(dst);
ggml_cann_pool_alloc output_allocator(ctx.pool(), n_bytes);
void* output_buffer = output_allocator.get();
aclTensor* alibi_output_tensor = ggml_cann_create_tensor(
output_buffer, ACL_FLOAT, ggml_type_size(dst->type), dst->ne,
dst->nb, GGML_MAX_DIMS);
if (max_bias <= 0.0f) {
// slope = 1.0
if (tmp_mask_tensor) {
aclnn_add(ctx, tmp_mask_tensor, acl_input_mul_scale_tensor,
alibi_output_tensor);
} else {
aclnn_add(ctx, acl_src1_fp32_tensor, acl_input_mul_scale_tensor,
alibi_output_tensor);
}
} else {
// slope != 1.0
if (tmp_mask_tensor) {
aclnn_alibi(ctx, acl_input_mul_scale_tensor, tmp_mask_tensor,
alibi_output_tensor, n_head, src0->ne, src_nb0,
max_bias, dst);
} else {
aclnn_alibi(ctx, acl_input_mul_scale_tensor,
acl_src1_fp32_tensor, alibi_output_tensor, n_head,
src0->ne, src_nb0, max_bias, dst);
}
}
// softmax
aclnn_softmax(ctx, alibi_output_tensor, 3, acl_dst);
ggml_cann_release_resources(ctx, alibi_output_tensor);
} else {
aclnn_softmax(ctx, acl_input_mul_scale_tensor, 3, acl_dst);
}
ggml_cann_release_resources(ctx, acl_src0, acl_src1_fp32_tensor, acl_dst,
acl_scale, acl_input_mul_scale_tensor, tmp_mask_tensor);
aclnn_softmax(ctx, softmax_tensor, 3, acl_dst);
ggml_cann_release_resources(ctx, acl_src0, acl_dst, acl_scale, softmax_tensor);
}
/**
* @brief Performs embedding operation on a 4D tensor using the CANN backend.
* @brief Performs index select 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.
* This function applies the `IndexSelect` operation along a specific dimension
* of the source tensor (`src_buffer`) using the indices from the index tensor (`index`).
* It iterates over the last two dimensions of the source tensor, creates the corresponding
* CANN tensors for the source, index, and output slices, and executes the `IndexSelect`
* operation for each slice.
*
* @param ctx The context for CANN backend operations.
* @param src_buffer The source buffer holding the data for the source tensor.
* @param src_buffer The source buffer containing the 4D input tensor data.
* @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.
* @param dst_buffer The destination buffer where the output tensor data will be written.
* @param dst_ne The dimensions of the destination tensor.
* @param dst_nb The strides (byte offsets) of the destination tensor.
* @param index The index tensor specifying the indices to select from the source tensor.
* @param type The data type of the source and destination tensors.
*/
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) {
static void aclnn_index_select_4d(ggml_backend_cann_context& ctx,
void* src_buffer,int64_t* src_ne, size_t* src_nb,
void* dst_buffer, int64_t* dst_ne, size_t* dst_nb,
ggml_tensor* index, ggml_type type) {
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_type_mapping(type), ggml_type_size(type),
src_ne, src_nb, 2);
// 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],
(char*)index->data + (i % index->ne[2]) * index->nb[2] + (j % index->ne[1]) * index->nb[1],
ggml_cann_type_mapping(index->type), ggml_element_size(index),
acl_index_ne, acl_index_nb, 1);
index->ne, 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);
GGML_CANN_CALL_ACLNN_OP(ctx, Embedding, acl_src_tensor, acl_index, acl_out);
(char*)dst_buffer + i * dst_nb[3] + j * dst_nb[2],
ggml_cann_type_mapping(type), ggml_type_size(type),
dst_ne, dst_nb, 2);
GGML_CANN_CALL_ACLNN_OP(ctx, IndexSelect, acl_src_tensor, 0, acl_index, acl_out);
ggml_cann_release_resources(ctx, acl_src_tensor, acl_index, acl_out);
}
}
}
/**
* @brief Performs inplace index copy operation on a 4D tensor using the CANN backend.
*
* This function applies the `IndexCopy` operation along a specific dimension of the
* destination tensor (`dst_buffer`) by copying elements from the source tensor (`src_buffer`)
* to positions specified by the index tensor (`index`).
* It iterates over the last two dimensions of the tensors, creates the corresponding
* CANN tensors for source, index, and destination slices, and performs the index copy
* operation for each slice.
*
* @param ctx The context for CANN backend operations.
* @param src_buffer The source buffer containing the 4D input tensor data to be copied.
* @param src_ne The dimensions of the source tensor.
* @param src_nb The strides (byte offsets) of the source tensor.
* @param dst_buffer The destination buffer where values will be copied to.
* @param dst_ne The dimensions of the destination tensor.
* @param dst_nb The strides (byte offsets) of the destination tensor.
* @param index The index tensor specifying target positions in the destination tensor.
* @param type The data type of the source and destination tensors.
*/
static void aclnn_index_copy_4d(ggml_backend_cann_context& ctx,
void* src_buffer,int64_t* src_ne, size_t* src_nb,
void* dst_buffer, int64_t* dst_ne, size_t* dst_nb,
ggml_tensor* index, ggml_type type) {
for (int64_t i = 0; i < src_ne[3]; i++) {
for (int64_t j = 0; j < src_ne[2]; j++) {
// src
aclTensor* acl_src_tensor = ggml_cann_create_tensor(
(char*)src_buffer + i * src_nb[3] + j * src_nb[2],
ggml_cann_type_mapping(type), ggml_type_size(type),
src_ne, src_nb, 2);
// index
aclTensor* acl_index = ggml_cann_create_tensor(
(char*)index->data + (i % index->ne[2]) * index->nb[2] + (j % index->ne[1]) * index->nb[1],
ggml_cann_type_mapping(index->type), ggml_element_size(index),
index->ne, index->nb, 1);
// out
aclTensor* acl_out = ggml_cann_create_tensor(
(char*)dst_buffer + i * dst_nb[3] + j * dst_nb[2],
ggml_cann_type_mapping(type), ggml_type_size(type),
dst_ne, dst_nb, 2);
GGML_CANN_CALL_ACLNN_OP(ctx, InplaceIndexCopy, acl_out, 0, acl_index, acl_src_tensor);
ggml_cann_release_resources(ctx, acl_src_tensor, acl_index, acl_out);
}
}
@ -1633,8 +1663,9 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
switch (src0->type) {
case GGML_TYPE_F32: {
aclnn_embedding_4d(ctx, src0->data, src0->ne, src0->nb, src1,
dst);
aclnn_index_select_4d(ctx, src0->data, src0->ne, src0->nb,
dst->data, dst->ne, dst->nb,
src1, dst->type);
break;
}
case GGML_TYPE_F16: {
@ -1651,8 +1682,9 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
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, ggml_cann_type_mapping(dst->type));
aclnn_embedding_4d(ctx, src_trans_buffer, src0->ne,
src_trans_nb, src1, dst);
aclnn_index_select_4d(ctx, src_trans_buffer, src0->ne, src_trans_nb,
dst->data, dst->ne, dst->nb,
src1, dst->type);
ggml_cann_release_resources(ctx, acl_src0, src_trans_tensor);
break;
}
@ -1712,8 +1744,10 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
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);
aclnn_index_select_4d(ctx, dequant_buffer_allocator.get(),
dequant_ne, dequant_nb,
dst->data, dst->ne, dst->nb,
src1, dst->type);
ggml_cann_release_resources(ctx, dequant_tensor);
break;
@ -1724,6 +1758,43 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
}
}
void ggml_cann_set_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
ggml_tensor* src0 = dst->src[0]; // src
ggml_tensor* src1 = dst->src[1]; // index
switch (dst->type) {
case GGML_TYPE_F32: {
aclnn_index_copy_4d(ctx, src0->data, src0->ne, src0->nb,
dst->data, dst->ne, dst->nb,
src1, dst->type);
break;
}
case GGML_TYPE_F16: {
aclTensor* acl_src0 = ggml_cann_create_tensor(src0);
ggml_cann_pool_alloc src_buffer_allocator(
ctx.pool(), ggml_nelements(src0) * sizeof(uint16_t));
void* src_trans_buffer = src_buffer_allocator.get();
size_t src_trans_nb[GGML_MAX_DIMS];
src_trans_nb[0] = sizeof(uint16_t);
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, ACL_FLOAT16, ggml_type_size(dst->type),
src0->ne, src_trans_nb, GGML_MAX_DIMS);
aclnn_cast(ctx, acl_src0, src_trans_tensor, ggml_cann_type_mapping(dst->type));
aclnn_index_copy_4d(ctx, src_trans_buffer, src0->ne, src_trans_nb,
dst->data, dst->ne, dst->nb,
src1, dst->type);
ggml_cann_release_resources(ctx, acl_src0, src_trans_tensor);
break;
}
default:
GGML_ABORT("Unsupported tensor type for GGML_OP_SET_ROWS");
break;
}
}
/**
* @brief Repeats elements of a tensor along a specified dimension.
*
@ -1785,8 +1856,25 @@ static void ggml_cann_mat_mul_fp(ggml_backend_cann_context& ctx,
size_t transpose_nb[] = {bcast_weight_nb[1], bcast_weight_nb[0],
bcast_weight_nb[2], bcast_weight_nb[3],
bcast_weight_nb[4], bcast_weight_nb[5]};
aclTensor* acl_weight_tensor =
ggml_cann_create_tensor(weight, transpose_ne, transpose_nb, n_dims);
aclTensor* acl_weight_tensor;
// Only check env once.
static bool weight_to_nz = parse_bool(get_env("GGML_CANN_WEIGHT_NZ").value_or(""));
if (weight_to_nz && is_matmul_weight(weight)) {
int64_t acl_stride[2] = {1, transpose_ne[1]};
// Reverse ne.
std::reverse(transpose_ne, transpose_ne + n_dims);
std::vector<int64_t> storageDims = {transpose_ne[0], transpose_ne[1]};
acl_weight_tensor = aclCreateTensor(
transpose_ne, n_dims, ggml_cann_type_mapping(weight->type), acl_stride,
0, ACL_FORMAT_FRACTAL_NZ, storageDims.data(), 2, weight->data);
} else {
acl_weight_tensor =
ggml_cann_create_tensor(weight, transpose_ne, transpose_nb, n_dims, ACL_FORMAT_ND);
}
aclTensor* acl_dst =
ggml_cann_create_tensor(dst, bcast_dst_ne, bcast_dst_nb, n_dims);
@ -3065,104 +3153,24 @@ void ggml_cann_flash_attn_ext(ggml_backend_cann_context& ctx, ggml_tensor* dst){
// Compute the slope if needed. Derived from ggml_cann_softmax().
if(maxBias != 0.0f){
// alibi
const int64_t ne2_ne3 = src0->ne[2] * src0->ne[3];
const int64_t n_head = src0->ne[2];
const int n_heads_log2_floor = 1u << (uint32_t)floor(log2(n_head));
float m0 = powf(2.0f, -(maxBias) / n_heads_log2_floor);
float m1 = powf(2.0f, -(maxBias / 2.0f) / n_heads_log2_floor);
// init arange
ggml_cann_pool_alloc arange_allocator(ctx.pool(),
ne2_ne3 * faElemSize);
void* tmp_arange_buffer = arange_allocator.get();
const int64_t n_heads = src0->ne[2];
ggml_cann_pool_alloc slope_allocator(ctx.pool(), n_heads * sizeof(float));
void* slope_buffer = slope_allocator.get();
aclnn_get_slope(ctx, n_heads, slope_buffer, maxBias);
// arange1: [1, ..., n_heads_log2_floor+1)
float start = 1;
float stop = n_heads_log2_floor + 1;
float step = 1;
int64_t n_elements_arange = n_heads_log2_floor;
int64_t tmp_arange1_ne[] = {n_heads_log2_floor};
size_t tmp_arange1_nb[] = {faElemSize};
aclTensor* tmp_arange1_tensor = ggml_cann_create_tensor(
tmp_arange_buffer, faDataType, faElemSize,
tmp_arange1_ne, tmp_arange1_nb,
GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
aclnn_arange(ctx, tmp_arange1_tensor, start, stop, step, n_elements_arange);
aclTensor* tmp_arange2_tensor = nullptr;
if (n_heads_log2_floor < ne2_ne3) {
// arange2: [1, ..., 2 * (k - n_heads_log2_floor) + 1)
start = 1;
stop = 2 * (ne2_ne3 - n_heads_log2_floor) + 1;
step = 2;
n_elements_arange = ne2_ne3 - n_heads_log2_floor;
int64_t tmp_arange2_ne[] = {ne2_ne3 - n_heads_log2_floor};
size_t tmp_arange2_nb[] = {faElemSize};
aclTensor* tmp_arange2_tensor = ggml_cann_create_tensor(
(char*)tmp_arange_buffer +
n_heads_log2_floor * faElemSize,
faDataType, faElemSize,
tmp_arange2_ne, tmp_arange2_nb, GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
aclnn_arange(ctx, tmp_arange2_tensor, start, stop, step,
n_elements_arange);
}
// init mk_base
ggml_cann_pool_alloc mk_base_allocator(ctx.pool(),
ne2_ne3 * faElemSize);
void* tmp_mk_base_buffer = mk_base_allocator.get();
int64_t tmp_mk_base1_ne[] = {n_heads_log2_floor};
size_t tmp_mk_base1_nb[] = {faElemSize};
aclTensor* tmp_mk_base1_tensor = ggml_cann_create_tensor(
tmp_mk_base_buffer, faDataType, faElemSize,
tmp_mk_base1_ne, tmp_mk_base1_nb,
GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
aclnn_fill_scalar(ctx, m0, tmp_mk_base1_tensor);
aclTensor* tmp_mk_base2_tensor = nullptr;
if (n_heads_log2_floor < ne2_ne3) {
int64_t tmp_mk_base2_ne[] = {ne2_ne3 - n_heads_log2_floor};
size_t tmp_mk_base2_nb[] = {faElemSize};
aclTensor* tmp_mk_base2_tensor = ggml_cann_create_tensor(
(char*)tmp_mk_base_buffer +
n_heads_log2_floor * faElemSize,
faDataType, faElemSize,
tmp_mk_base2_ne, tmp_mk_base2_nb, GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
aclnn_fill_scalar(ctx, m1, tmp_mk_base2_tensor);
}
// init mk
int64_t tmp_mk_base_ne[] = {ne2_ne3};
size_t tmp_mk_base_nb[] = {faElemSize};
aclTensor* tmp_mk_base_tensor = ggml_cann_create_tensor(
tmp_mk_base_buffer, faDataType, faElemSize,
tmp_mk_base_ne, tmp_mk_base_nb,
GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
aclTensor* tmp_arange_tensor = ggml_cann_create_tensor(
tmp_arange_buffer, faDataType, faElemSize,
tmp_mk_base_ne, tmp_mk_base_nb,
GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
aclnn_pow_tensor_tensor(ctx, tmp_mk_base_tensor, tmp_arange_tensor);
// reshape mk
int64_t tmp_mk_ne[] = {1, 1, src0->ne[2], src0->ne[3]};
size_t tmp_mk_nb[GGML_MAX_DIMS];
tmp_mk_nb[0] = faElemSize;
int64_t slope_ne[] = {1, 1, n_heads, 1};
size_t slope_nb[GGML_MAX_DIMS];
slope_nb[0] = sizeof(float);
for(int i = 1;i<GGML_MAX_DIMS;i++) {
tmp_mk_nb[i] = tmp_mk_nb[i - 1] * tmp_mk_ne[i - 1];
slope_nb[i] = slope_nb[i-1] * slope_ne[0];
}
aclTensor* tmp_mk_tensor = ggml_cann_create_tensor(
tmp_mk_base_buffer, faDataType, faElemSize,
tmp_mk_ne, tmp_mk_nb, GGML_MAX_DIMS,
ACL_FORMAT_ND);
GGML_CANN_CALL_ACLNN_OP(ctx, InplaceMul, bcast_pse_tensor, tmp_mk_tensor);
ggml_cann_release_resources(ctx, tmp_arange1_tensor, tmp_arange2_tensor,
tmp_mk_base1_tensor, tmp_mk_base2_tensor, tmp_mk_base_tensor,
tmp_arange_tensor, tmp_mk_tensor);
aclTensor* slope_tensor = ggml_cann_create_tensor(
slope_buffer, ACL_FLOAT, sizeof(float),
slope_ne, slope_nb, GGML_MAX_DIMS);
GGML_CANN_CALL_ACLNN_OP(ctx, InplaceMul, bcast_pse_tensor, slope_tensor);
ggml_cann_release_resources(ctx, slope_tensor);
}
}

166
ggml/src/ggml-cann/aclnn_ops.h
View File

@ -23,6 +23,7 @@
#ifndef CANN_ACLNN_OPS
#define CANN_ACLNN_OPS
#include <unordered_set>
#include <functional>
#include <aclnnop/aclnn_abs.h>
#include <aclnnop/aclnn_neg.h>
@ -423,15 +424,25 @@ void ggml_cann_softmax(ggml_backend_cann_context& ctx, ggml_tensor* dst);
*
* @details This function retrieves rows from a source tensor src0 according to
* the indices provided in another tensor src1 and stores the result in
* a destination tensor (\p dst). It supports different data types
* including F32, F16, Q4_0, and Q8_0.
* a destination tensor (\p dst).
*
* @param ctx The backend CANN context for executing operations.
* @param dst The destination tensor where the extracted rows will be stored.
* dst->op is `GGML_OP_GET_ROWS`.
*/
void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst);
/**
* @brief Writes specific rows into a tensor at positions specified by indices.
*
* @details This function copies rows from a source tensor into a destination
* tensor (\p dst) at the positions indicated by the indices in another
* tensor.
*
* @param ctx The backend CANN context for executing operations.
* @param dst The destination tensor where the specified rows will be updated.
*/
void ggml_cann_set_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst);
/**
* @brief Executes matrix multiplication for the given tensor.
*
@ -1020,6 +1031,37 @@ inline void ggml_cann_async_memset(ggml_backend_cann_context & ctx, void * buffe
*/
void ggml_cann_mul_mat_id(ggml_backend_cann_context& ctx, ggml_tensor* dst);
/**
* @brief Check whether a tensor is a weight tensor for matrix multiplication.
*
* @details Checks whether the given tensor serves as weight parameters in matrix multiplication operations,
* typically within neural network layers. The function maintains a static set of canonical weight
* naming suffixes from Transformer-based architectures. Uses substring matching to identify weight
* tensors even with hierarchical naming patterns.
*
* @param tensor Pointer to the target ggml_tensor object (const-qualified).
*/
static bool is_matmul_weight(const ggml_tensor* tensor) {
std::string name = ggml_get_name(tensor);
static const std::unordered_set<std::string> weight_suffixes{
"output.weight",
"attn_q.weight",
"attn_k.weight",
"attn_v.weight",
"attn_output.weight",
"ffn_gate.weight",
"ffn_up.weight",
"ffn_down.weight"
};
for (const auto& suffix : weight_suffixes) {
if (name.find(suffix) != std::string::npos) {
return true;
}
}
return false;
}
/**
* @brief Applies a element-wise operation to two input tensors using the CANN
* backend.
@ -1066,7 +1108,7 @@ void ggml_cann_binary_op(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
* @param dst The destination tensor. Its src[0] is treated as the input tensor.
*/
template <void unary_op(ggml_backend_cann_context&, aclTensor*, aclTensor*)>
void ggml_cann_unary_op(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
void ggml_cann_op_unary(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
ggml_tensor* src = dst->src[0];
aclTensor* acl_src = ggml_cann_create_tensor(src);
@ -1079,47 +1121,123 @@ template <void unary_op(ggml_backend_cann_context&, aclTensor*, aclTensor*)>
/**
* @brief Applies a unary operation to a ggml tensor using the CANN backend.
*
* @details This function performs a unary operation on the input tensor using
* a user-provided lambda or callable object `unary_op`, which accepts the CANN
* context and two ACL tensors (source and destination). Internally, this function
* creates ACL representations of the ggml tensors and invokes the unary operation.
* The result is stored in the destination tensor `dst`. This utility abstracts the
* common boilerplate of tensor conversion and cleanup when implementing unary ops.
* @details This function applies a unary operation to the input tensor using
* a user-provided lambda or callable `unary_op`. The lambda receives the
* CANN backend context and two ACL tensors: the source and the destination.
*
* @param unary_op A callable that performs the unary operation using CANN APIs.
* @param ctx The CANN context used for operations.
* @param dst The destination tensor where the result will be stored.
* The source tensor is retrieved from `dst->src[0]`.
* Internally, this function handles the conversion from GGML tensors to ACL tensors,
* calls the provided unary op, and manages resource cleanup. The input is assumed
* to be `dst->src[0]`, and the result is written to `dst`.
*
* This utility simplifies writing unary op wrappers by abstracting tensor preparation.
*
* @param unary_op A callable that performs the unary operation using CANN ACL APIs.
* @param ctx The CANN context for operation execution.
* @param dst The destination ggml_tensor where the result will be stored.
* The input tensor is assumed to be `dst->src[0]`.
*
* @see GGML_CANN_CALL_OP_UNARY
*/
void ggml_cann_unary_op(
void ggml_cann_op_unary(
std::function<void(ggml_backend_cann_context&, aclTensor*, aclTensor*)> unary_op,
ggml_backend_cann_context& ctx, ggml_tensor* dst);
/**
* @brief Helper macro to invoke a unary ACL operation using ggml_cann_unary_op.
* @brief Applies a gated (GLU-style) unary operation using the CANN backend.
*
* This macro defines an inline lambda wrapping a specific ACL operation name,
* and passes it to the templated ggml_cann_unary_op function. It simplifies
* calling unary ops by hiding the lambda boilerplate.
* @details This function performs a gated activation such as GEGLU or ReGLU.
* It supports two input modes:
*
* 1. **Dual input mode**: `dst->src[0]` and `dst->src[1]` are both valid tensors.
* These are used directly as the value and gate tensors.
*
* 2. **Packed input mode**: Only `dst->src[0]` is valid, and it is assumed to
* contain a concatenation of value and gate along the first dimension. This tensor
* will be split into two equal halves to form the value and gate inputs.
*
* The function applies a user-provided unary operation (e.g., GELU) to the value tensor,
* then multiplies the result in-place with the gate tensor:
*
* Internally, the lambda will call:
* @code
* GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);
* dst = unary_op(value) * gate;
* @endcode
*
* The `swapped` parameter (from `dst->op_params[1]`) allows flipping the
* order of value/gate in the packed input case.
*
* @param unary_op A callable that performs the unary operation using CANN ACL APIs.
* It receives (ctx, acl_value_tensor, acl_output_tensor).
* @param ctx The CANN context used for execution.
* @param dst The destination ggml_tensor. Source tensors are in `dst->src[0]` and optionally `src[1]`.
*
* @see GGML_CANN_CALL_OP_UNARY_GATED
*/
void ggml_cann_op_unary_gated(
std::function<void(ggml_backend_cann_context&, aclTensor*, aclTensor*)> unary_op,
ggml_backend_cann_context& ctx, ggml_tensor* dst);
/**
* @brief Helper macro to call a unary ACL operator via ggml_cann_op_unary.
*
* This macro wraps the specified ACLNN unary operator name into a lambda expression,
* and passes it to `ggml_cann_op_unary`, which handles the common logic for executing
* unary ops in the CANN backend.
*
* Internally, this macro expands to a lambda like:
* @code
* [](ggml_backend_cann_context& ctx, aclTensor* acl_src, aclTensor* acl_dst) {
* GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);
* };
* @endcode
*
* This lambda is then passed to `ggml_cann_op_unary`, which applies the operation.
*
* @param OP_NAME The name of the ACL unary operator to invoke via GGML_CANN_CALL_ACLNN_OP.
*
* @see ggml_cann_unary_op
* @see ggml_cann_op_unary
* @see GGML_CANN_CALL_ACLNN_OP
*/
#define GGML_CANN_CALL_UNARY_OP(OP_NAME) \
#define GGML_CANN_CALL_OP_UNARY(OP_NAME) \
do { \
auto lambda = [](ggml_backend_cann_context& ctx, \
aclTensor* acl_src, \
aclTensor* acl_dst) { \
GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst); \
}; \
ggml_cann_unary_op(lambda, ctx, dst); \
ggml_cann_op_unary(lambda, ctx, dst); \
} \
while (0)
/**
* @brief Helper macro to call a gated unary ACL operator via ggml_cann_op_unary_gated.
*
* This macro wraps the specified ACLNN unary operator name into a lambda expression,
* and passes it to `ggml_cann_op_unary_gated`, which handles the common logic for
* executing gated unary ops in the CANN backend.
*
* Internally, this macro expands to a lambda like:
* @code
* [](ggml_backend_cann_context& ctx, aclTensor* acl_src, aclTensor* acl_dst) {
* GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);
* };
* @endcode
*
* This lambda is then passed to `ggml_cann_op_unary_gated`, which applies the operation.
*
* @param OP_NAME The name of the ACL unary operator to invoke via GGML_CANN_CALL_ACLNN_OP.
*
* @see ggml_cann_op_unary_gated
* @see GGML_CANN_CALL_ACLNN_OP
*/
#define GGML_CANN_CALL_OP_UNARY_GATED(OP_NAME) \
do { \
auto lambda = [](ggml_backend_cann_context& ctx, \
aclTensor* acl_src, \
aclTensor* acl_dst) { \
GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst); \
}; \
ggml_cann_op_unary_gated(lambda, ctx, dst); \
} \
while (0)
#endif // CANN_ACLNN_OPS

36
ggml/src/ggml-cann/common.h
View File

@ -337,6 +337,29 @@ private:
int32_t device_;
};
#ifdef USE_ACL_GRAPH
struct ggml_graph_node_properties {
void * node_address;
ggml_op node_op;
int64_t ne[GGML_MAX_DIMS];
size_t nb[GGML_MAX_DIMS];
void * src_address[GGML_MAX_SRC];
int32_t op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)];
};
struct ggml_cann_graph {
~ggml_cann_graph() {
if (graph != nullptr) {
aclmdlRIDestroy(graph);
}
}
aclmdlRI graph = nullptr;
std::vector<ggml_graph_node_properties> ggml_graph_properties;
};
#endif // USE_ACL_GRAPH
/**
* @brief Context for managing CANN backend operations.
*/
@ -345,8 +368,13 @@ struct ggml_backend_cann_context {
std::string name; /**< Name of the device. */
std::string description; /**< Description of the device. */
aclrtEvent copy_event = nullptr; /**< Event for managing copy operations. */
#ifdef USE_ACL_GRAPH
/// Cached CANN ACL graph used for executing the current ggml computation graph.
std::unique_ptr<ggml_cann_graph> cann_graph;
#endif
cann_task_queue task_queue;
bool async_mode;
bool support_set_rows;
aclrtStream streams[GGML_CANN_MAX_STREAMS] = {nullptr}; /**< Array of streams for the device. */
@ -362,6 +390,14 @@ struct ggml_backend_cann_context {
async_mode = parse_bool(get_env("GGML_CANN_ASYNC_MODE").value_or(""));
GGML_LOG_INFO("%s: device %d async operator submission is %s\n", __func__,
device, async_mode ? "ON" : "OFF");
support_set_rows = parse_bool(get_env("LLAMA_SET_ROWS").value_or(""));
GGML_LOG_INFO("%s: LLAMA_SET_ROWS is %s\n", __func__, support_set_rows ? "ON" : "OFF");
if (!support_set_rows) {
GGML_LOG_INFO("%s: CANN Graph currently only supports execution when LLAMA_SET_ROWS is ON. "
"Falling back to eager mode.\n", __func__);
}
}
/**

387
ggml/src/ggml-cann/ggml-cann.cpp
View File

@ -24,6 +24,7 @@
#include <acl/acl.h>
#include <stdarg.h>
#include <aclnnop/aclnn_trans_matmul_weight.h>
#include <cmath>
#include <cstdio>
@ -1115,6 +1116,61 @@ static enum ggml_status ggml_backend_cann_buffer_init_tensor(
return GGML_STATUS_SUCCESS;
}
// ND to NZ Workspace Cache Management. Thread-safety: Not guaranteed
namespace {
void* g_nz_workspace = nullptr;
size_t g_nz_workspace_allocated = 0;
void release_nz_workspace() {
if (g_nz_workspace) {
aclrtFree(g_nz_workspace);
g_nz_workspace = nullptr;
g_nz_workspace_allocated = 0;
}
}
void relloc_nz_workspace(size_t new_size) {
if (new_size > g_nz_workspace_allocated) {
if (g_nz_workspace) {
aclrtFree(g_nz_workspace);
g_nz_workspace = nullptr;
}
ACL_CHECK(aclrtMalloc(&g_nz_workspace, new_size, ACL_MEM_MALLOC_HUGE_FIRST));
g_nz_workspace_allocated = new_size;
}
}
}
/**
* @brief Convert tensor weights to NZ format using Ascend CANN API.
*
* This function creates a transposed tensor descriptor and performs the
* TransMatmulWeight operation. Converting tensor formats can significantly
* improve performance on certain hardware.
*
* @param tensor Pointer to the input ggml_tensor containing the weights.
* @param data Pointer to the raw data buffer for the tensor weights.
* @param offset Byte offset within the tensor data buffer where weights start.
*
* @note The workspace buffer used in this function is managed globally and reused
* across calls. This reduces overhead from repeated memory allocation and deallocation.
*/
static void weight_format_to_nz(ggml_tensor *tensor, const void *data, size_t offset) {
aclTensor* weightTransposed = ggml_cann_create_tensor(tensor, tensor->ne,
tensor->nb, 2, ACL_FORMAT_ND, offset);
uint64_t workspaceSize = 0;
aclOpExecutor *executor;
// TransMatmulWeight
ACL_CHECK(aclnnTransMatmulWeightGetWorkspaceSize(weightTransposed,
&workspaceSize, &executor));
// Avoid frequent malloc/free of the workspace.
relloc_nz_workspace(workspaceSize);
ACL_CHECK(aclnnTransMatmulWeight(g_nz_workspace, workspaceSize, executor, nullptr));
ACL_CHECK(aclDestroyTensor(weightTransposed));
}
// TODO: need handle tensor which has paddings.
/**
* @brief Set tensor data in a CANN buffer.
@ -1139,9 +1195,16 @@ static void ggml_backend_cann_buffer_set_tensor(
// For acl, synchronous functions use this default stream.
// Why aclrtSynchronizeDevice?
// Only check env once.
static bool weight_to_nz = parse_bool(get_env("GGML_CANN_WEIGHT_NZ").value_or(""));
if (!need_transform(tensor->type)) {
ACL_CHECK(aclrtMemcpy((char *)tensor->data + offset, size, data, size,
ACL_MEMCPY_HOST_TO_DEVICE));
if (weight_to_nz && is_matmul_weight((const ggml_tensor*)tensor)) {
GGML_ASSERT(tensor->ne[2] == 1);
GGML_ASSERT(tensor->ne[3] == 1);
weight_format_to_nz(tensor, data, offset);
}
} else {
void *transform_buffer = malloc(size);
ggml_backend_cann_transform(tensor, data, transform_buffer);
@ -1375,20 +1438,32 @@ static size_t ggml_backend_cann_buffer_type_get_alloc_size(
size_t size = ggml_nbytes(tensor);
int64_t ne0 = tensor->ne[0];
// Only check env once.
static bool weight_to_nz = parse_bool(get_env("GGML_CANN_WEIGHT_NZ").value_or(""));
// last line must bigger than 32, because every single op deal at
// least 32 bytes.
// TODO: quantized type?
// int64_t line_size = ne0 * ggml_element_size(tensor);
// int64_t line_size_align_32 = (line_size + 31) & ~31;
// size += (line_size_align_32 - line_size);
// TODO: not support quantized yet.
// TODO: consider un-continue tensor.
if (ggml_is_quantized(tensor->type)) {
if (ne0 % MATRIX_ROW_PADDING != 0) {
size += ggml_row_size(
tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
}
} else if (weight_to_nz && is_matmul_weight((const ggml_tensor*)tensor)) {
// NZ format weight are not support quantized yet.
// If ND tensor transform to NZ, size may changed.
int64_t shape[] = {tensor->ne[1], tensor->ne[0]};
GGML_ASSERT(tensor->ne[2] == 1);
GGML_ASSERT(tensor->ne[3] == 1);
const aclIntArray *acl_shape = aclCreateIntArray(shape, 2);
size_t new_size;
ACL_CHECK(aclnnCalculateMatmulWeightSizeV2(acl_shape,
ggml_cann_type_mapping(tensor->type), &new_size));
ACL_CHECK(aclDestroyIntArray(acl_shape));
size = std::max(size, new_size);
}
return size;
@ -1594,6 +1669,9 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
case GGML_OP_GET_ROWS:
ggml_cann_get_rows(ctx, dst);
break;
case GGML_OP_SET_ROWS:
ggml_cann_set_rows(ctx, dst);
break;
case GGML_OP_DUP:
ggml_cann_dup(ctx, dst);
break;
@ -1616,16 +1694,18 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
case GGML_OP_UNARY:
switch (ggml_get_unary_op(dst)) {
case GGML_UNARY_OP_ABS:
GGML_CANN_CALL_UNARY_OP(Abs);
GGML_CANN_CALL_OP_UNARY(Abs);
break;
case GGML_UNARY_OP_NEG:
GGML_CANN_CALL_UNARY_OP(Neg);
GGML_CANN_CALL_OP_UNARY(Neg);
break;
case GGML_UNARY_OP_GELU:
GGML_CANN_CALL_UNARY_OP(Gelu);
case GGML_UNARY_OP_GELU_ERF:
// aclnnGelu internally uses the erf-based approximation.
GGML_CANN_CALL_OP_UNARY(Gelu);
break;
case GGML_UNARY_OP_SILU:
GGML_CANN_CALL_UNARY_OP(Silu);
GGML_CANN_CALL_OP_UNARY(Silu);
break;
case GGML_UNARY_OP_GELU_QUICK: {
auto lambda = [](ggml_backend_cann_context& ctx,
@ -1633,31 +1713,31 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
aclTensor* acl_dst) {
GGML_CANN_CALL_ACLNN_OP(ctx, GeluV2, acl_src, 0, acl_dst);
};
ggml_cann_unary_op(lambda, ctx, dst);
ggml_cann_op_unary(lambda, ctx, dst);
} break;
case GGML_UNARY_OP_TANH:
GGML_CANN_CALL_UNARY_OP(Tanh);
GGML_CANN_CALL_OP_UNARY(Tanh);
break;
case GGML_UNARY_OP_RELU:
GGML_CANN_CALL_UNARY_OP(Relu);
GGML_CANN_CALL_OP_UNARY(Relu);
break;
case GGML_UNARY_OP_SIGMOID:
GGML_CANN_CALL_UNARY_OP(Sigmoid);
GGML_CANN_CALL_OP_UNARY(Sigmoid);
break;
case GGML_UNARY_OP_HARDSIGMOID:
GGML_CANN_CALL_UNARY_OP(Hardsigmoid);
GGML_CANN_CALL_OP_UNARY(Hardsigmoid);
break;
case GGML_UNARY_OP_HARDSWISH:
GGML_CANN_CALL_UNARY_OP(Hardswish);
GGML_CANN_CALL_OP_UNARY(Hardswish);
break;
case GGML_UNARY_OP_EXP:
GGML_CANN_CALL_UNARY_OP(Exp);
GGML_CANN_CALL_OP_UNARY(Exp);
break;
case GGML_UNARY_OP_ELU:
ggml_cann_elu(ctx, dst);
break;
case GGML_UNARY_OP_SGN:
GGML_CANN_CALL_UNARY_OP(Sign);
GGML_CANN_CALL_OP_UNARY(Sign);
break;
case GGML_UNARY_OP_STEP:
ggml_cann_step(ctx, dst);
@ -1666,6 +1746,31 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
return false;
}
break;
case GGML_OP_GLU:
switch (ggml_get_glu_op(dst)) {
case GGML_GLU_OP_REGLU:
GGML_CANN_CALL_OP_UNARY_GATED(Relu);
break;
case GGML_GLU_OP_GEGLU:
case GGML_GLU_OP_GEGLU_ERF:
// aclnnGelu internally uses the erf-based approximation.
GGML_CANN_CALL_OP_UNARY_GATED(Gelu);
break;
case GGML_GLU_OP_SWIGLU:
GGML_CANN_CALL_OP_UNARY_GATED(Silu);
break;
case GGML_GLU_OP_GEGLU_QUICK: {
auto lambda = [](ggml_backend_cann_context& ctx,
aclTensor* acl_src,
aclTensor* acl_dst) {
GGML_CANN_CALL_ACLNN_OP(ctx, GeluV2, acl_src, 0, acl_dst);
};
ggml_cann_op_unary_gated(lambda, ctx, dst);
} break;
default:
return false;
}
break;
case GGML_OP_NORM:
ggml_cann_norm(ctx, dst);
break;
@ -1708,7 +1813,7 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
ggml_cann_binary_op<aclnn_mul>(ctx, dst);
break;
case GGML_OP_SQRT:
GGML_CANN_CALL_UNARY_OP(Sqrt);
GGML_CANN_CALL_OP_UNARY(Sqrt);
break;
case GGML_OP_CLAMP:
ggml_cann_clamp(ctx, dst);
@ -1753,16 +1858,16 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
ggml_cann_argmax(ctx, dst);
break;
case GGML_OP_COS:
ggml_cann_unary_op<aclnn_cos>(ctx, dst);
ggml_cann_op_unary<aclnn_cos>(ctx, dst);
break;
case GGML_OP_SIN:
ggml_cann_unary_op<aclnn_sin>(ctx, dst);
ggml_cann_op_unary<aclnn_sin>(ctx, dst);
break;
case GGML_OP_CONV_TRANSPOSE_1D:
ggml_cann_conv_transpose_1d(ctx, dst);
break;
case GGML_OP_LOG:
GGML_CANN_CALL_UNARY_OP(Log);
GGML_CANN_CALL_OP_UNARY(Log);
break;
case GGML_OP_MEAN:
ggml_cann_mean(ctx, dst);
@ -1911,6 +2016,9 @@ static bool ggml_backend_cann_cpy_tensor_async(
(ggml_backend_cann_context*)backend_dst->context;
size_t copy_size = ggml_nbytes(dst);
if (copy_size == 0) {
return true;
}
if (backend_src != backend_dst) {
ggml_backend_cann_buffer_context* buf_ctx_src =
(ggml_backend_cann_buffer_context*)buf_src->context;
@ -1967,6 +2075,160 @@ static void ggml_backend_cann_synchronize(ggml_backend_t backend) {
ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
}
#ifdef USE_ACL_GRAPH
/**
* @brief Populate the internal CANN graph node properties from the ggml computation graph.
*
* This function copies all node attributes (operation type, dimensions, strides, input sources,
* and operation parameters) into the cached CANN graph structure for later reuse or comparison.
*
* @param cann_ctx The CANN backend context.
* @param cgraph The ggml computational graph.
*/
static void set_ggml_graph_node_properties(ggml_backend_cann_context * cann_ctx, ggml_cgraph * cgraph) {
for (int node_idx = 0; node_idx < cgraph->n_nodes; node_idx++) {
ggml_tensor * node = cgraph->nodes[node_idx];
cann_ctx->cann_graph->ggml_graph_properties[node_idx].node_address = node->data;
cann_ctx->cann_graph->ggml_graph_properties[node_idx].node_op = node->op;
for (int dim = 0; dim < GGML_MAX_DIMS; dim++) {
cann_ctx->cann_graph->ggml_graph_properties[node_idx].ne[dim] = node->ne[dim];
cann_ctx->cann_graph->ggml_graph_properties[node_idx].nb[dim] = node->nb[dim];
}
for (int src = 0; src < GGML_MAX_SRC; src++) {
cann_ctx->cann_graph->ggml_graph_properties[node_idx].src_address[src] =
node->src[src] ? node->src[src]->data : nullptr;
}
memcpy(cann_ctx->cann_graph->ggml_graph_properties[node_idx].op_params, node->op_params, GGML_MAX_OP_PARAMS);
}
}
/**
* @brief Check if a ggml tensor node matches a previously captured CANN graph node.
*
* This function compares all relevant fields (address, op type, shape, source inputs, op params)
* to determine whether the current node matches a previously recorded version.
*
* @param node The current ggml tensor node.
* @param graph_node_properties The stored properties of a CANN graph node.
* @return true if all fields match (excluding GGML_OP_VIEW); false otherwise.
*/
static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_graph_node_properties * graph_node_properties) {
if (node->data != graph_node_properties->node_address &&
node->op != GGML_OP_VIEW) {
return false;
}
if (node->op != graph_node_properties->node_op) {
return false;
}
for (int i = 0; i < GGML_MAX_DIMS; i++) {
if (node->ne[i] != graph_node_properties->ne[i]) {
return false;
}
if (node->nb[i] != graph_node_properties->nb[i]) {
return false;
}
}
for (int i = 0; i < GGML_MAX_SRC; i++) {
if (node->src[i] &&
node->src[i]->data != graph_node_properties->src_address[i] &&
node->op != GGML_OP_VIEW
) {
return false;
}
}
if (node->op == GGML_OP_SCALE &&
memcmp(graph_node_properties->op_params, node->op_params, GGML_MAX_OP_PARAMS) != 0) {
return false;
}
return true;
}
/**
* @brief Determine if the CANN graph needs to be rebuilt due to graph changes.
*
* This checks whether the number or properties of ggml graph nodes have changed
* compared to the last captured CANN graph. If so, the CANN graph must be re-captured.
*
* @param cann_ctx The CANN backend context.
* @param cgraph The current ggml computation graph.
* @return true if an update is required; false otherwise.
*/
static bool is_cann_graph_update_required(ggml_backend_cann_context * cann_ctx, ggml_cgraph * cgraph) {
// The number of nodes is different, so the graph needs to be reconstructed.
if (cann_ctx->cann_graph->ggml_graph_properties.size() != (size_t)cgraph->n_nodes) {
cann_ctx->cann_graph->ggml_graph_properties.resize(cgraph->n_nodes);
return true;
}
// The number of nodes is the same; iterate over each node to check whether they match.
for (int i = 0; i < cgraph->n_nodes; i++) {
bool has_matching_properties = ggml_graph_node_has_matching_properties(
cgraph->nodes[i], &cann_ctx->cann_graph->ggml_graph_properties[i]);
if(!has_matching_properties) {
return true;
}
}
return false;
}
#endif // USE_ACL_GRAPH
/**
* @brief Evaluate the computation graph and optionally capture or execute it using CANN graph API.
*
* If CANN graph execution is enabled and graph capture is required, this function begins
* graph capture, runs the graph, ends capture, and stores the captured graph.
*
* Otherwise, it falls back to op-by-op execution using the CANN compute kernel dispatcher.
*
* @param cann_ctx The CANN backend context.
* @param cgraph The ggml computation graph.
* @param use_cann_graph Whether to use CANN graph execution.
* @param cann_graph_update_required Whether graph capture is needed due to graph changes.
*/
static void evaluate_and_capture_cann_graph(ggml_backend_cann_context * cann_ctx, ggml_cgraph * cgraph,
bool & use_cann_graph, bool & cann_graph_update_required) {
#ifdef USE_ACL_GRAPH
if (use_cann_graph && cann_graph_update_required) {
if (cann_ctx->cann_graph->graph != nullptr) {
ACL_CHECK(aclmdlRIDestroy(cann_ctx->cann_graph->graph));
cann_ctx->cann_graph->graph = nullptr;
}
ACL_CHECK(aclmdlRICaptureBegin(cann_ctx->stream(), ACL_MODEL_RI_CAPTURE_MODE_GLOBAL));
}
#endif // USE_ACL_GRAPH
// Only perform the graph execution if CANN graphs are not enabled, or we are capturing the graph.
// With the use of CANN graphs, the execution will be performed by the graph launch.
if (!use_cann_graph || cann_graph_update_required) {
for (int i = 0; i < cgraph->n_nodes; i++) {
ggml_tensor * node = cgraph->nodes[i];
if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
continue;
}
bool ok = ggml_cann_compute_forward(*cann_ctx, node);
if (!ok) {
GGML_LOG_ERROR("%s: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
}
GGML_ASSERT(ok);
}
}
#ifdef USE_ACL_GRAPH
if (use_cann_graph && cann_graph_update_required) { // End CANN graph capture
ACL_CHECK(aclmdlRICaptureEnd(cann_ctx->stream(), &cann_ctx->cann_graph->graph));
}
if (use_cann_graph) {
// Execute graph
ACL_CHECK(aclmdlRIExecuteAsync(cann_ctx->cann_graph->graph, cann_ctx->stream()));
}
#endif // USE_ACL_GRAPH
}
/**
* @brief Computes a computational graph using a CANN backend.
*
@ -1983,24 +2245,37 @@ static enum ggml_status ggml_backend_cann_graph_compute(
ggml_backend_t backend, ggml_cgraph* cgraph) {
ggml_backend_cann_context* cann_ctx =
(ggml_backend_cann_context*)backend->context;
ggml_cann_set_device(cann_ctx->device);
release_nz_workspace();
#ifdef USE_ACL_GRAPH
bool use_cann_graph = true;
bool cann_graph_update_required = false;
for (int i = 0; i < cgraph->n_nodes; i++) {
ggml_tensor* node = cgraph->nodes[i];
if (ggml_is_empty(node) || node->op == GGML_OP_NONE) {
continue;
// check environment LLAMA_SET_ROWS
if (!cann_ctx->support_set_rows) {
use_cann_graph = false;
}
bool ok = ggml_cann_compute_forward(*cann_ctx, node);
if (use_cann_graph) {
if (cann_ctx->cann_graph == nullptr) {
cann_ctx->cann_graph.reset(new ggml_cann_graph());
cann_graph_update_required = true;
}
if (!ok) {
GGML_LOG_ERROR("%s: error: op not supported %s (%s)\n", __func__,
node->name, ggml_op_name(node->op));
}
GGML_ASSERT(ok);
cann_graph_update_required = is_cann_graph_update_required(cann_ctx, cgraph);
set_ggml_graph_node_properties(cann_ctx, cgraph);
}
#else
bool use_cann_graph = false;
bool cann_graph_update_required = false;
#endif // USE_ACL_GRAPH
evaluate_and_capture_cann_graph(
cann_ctx,
cgraph,
use_cann_graph,
cann_graph_update_required
);
return GGML_STATUS_SUCCESS;
}
@ -2036,10 +2311,23 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
case GGML_UNARY_OP_ELU:
case GGML_UNARY_OP_SGN:
case GGML_UNARY_OP_STEP:
case GGML_UNARY_OP_GELU_ERF:
return true;
default:
return false;
}
case GGML_OP_GLU:
switch (ggml_get_glu_op(op)) {
case GGML_GLU_OP_REGLU:
case GGML_GLU_OP_GEGLU:
case GGML_GLU_OP_SWIGLU:
case GGML_GLU_OP_GEGLU_ERF:
case GGML_GLU_OP_GEGLU_QUICK:
return true;
default:
return false;
}
break;
case GGML_OP_MUL_MAT: {
switch (op->src[0]->type) {
case GGML_TYPE_F16:
@ -2086,12 +2374,14 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
return false;
}
} break;
case GGML_OP_SET_ROWS:
{
// TODO: add support
// ref: https://github.com/ggml-org/llama.cpp/pull/14274
#pragma message("TODO: implement F32, F16, BF16, Q4_0, Q4_1, Q5_0, Q5_1, Q8_0, IQ4_NL support (https://github.com/ggml-org/llama.cpp/pull/14661)")
case GGML_OP_SET_ROWS: {
switch (op->type) {
case GGML_TYPE_F32:
case GGML_TYPE_F16:
return true;
default:
return false;
}
} break;
case GGML_OP_CPY: {
ggml_tensor *src = op->src[0];
@ -2101,12 +2391,6 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
// 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: {
@ -2172,8 +2456,8 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
// value of paddingW should be at most half of kernelW
return (p0 <= (k0 / 2)) && (p1 <= (k1 / 2));
}
case GGML_OP_SUM:
case GGML_OP_DUP:
case GGML_OP_SUM:
case GGML_OP_IM2COL:
case GGML_OP_CONCAT:
case GGML_OP_REPEAT:
@ -2215,9 +2499,11 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
memcpy(&bias, (float*)op->op_params + 1, sizeof(float));
return bias == 0.0f; // TODO: support bias != 0.0f
case GGML_OP_SOFT_MAX:
// TODO: support broadcast
// ref: https://github.com/ggml-org/llama.cpp/pull/14435
return !op->src[1] || (op->src[1]->ne[2] == 1 && op->src[1]->ne[3] == 1);
// TODO: support attention sinks [TAG_ATTN_SINKS]
if (op->src[2]) {
return false;
}
return true;
case GGML_OP_FLASH_ATTN_EXT:{
// derived from [ggml-cuda.cu]
if(op->src[1]->type != GGML_TYPE_F16 || op->src[2]->type != GGML_TYPE_F16){
@ -2229,6 +2515,10 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
if(op->type != GGML_TYPE_F16 && op->type != GGML_TYPE_F32 && op->type != GGML_TYPE_BF16){
return false;
}
// TODO: support attention sinks [TAG_ATTN_SINKS]
if (op->src[4]) {
return false;
}
if (op->src[1]->ne[0] != op->src[2]->ne[0]) {
// different head sizes of K and V are not supported yet
return false;
@ -2240,11 +2530,6 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
// DeepSeek MLA
return false;
}
// TODO: support broadcast
// ref: https://github.com/ggml-org/llama.cpp/pull/14435
if (op->src[0]->ne[3] != 1) {
return false;
}
float logitSoftcap = 0.0f;
memcpy(&logitSoftcap, (float*)op->op_params + 2, sizeof(float));
if(logitSoftcap != 0.0f) {

17
ggml/src/ggml-common.h
View File

@ -99,6 +99,9 @@ typedef sycl::half2 ggml_half2;
#define QI4_1 (QK4_1 / (4 * QR4_1))
#define QR4_1 2
#define QI_MXFP4 (QK_MXFP4 / (4 * QR_MXFP4))
#define QR_MXFP4 2
#define QI5_0 (QK5_0 / (4 * QR5_0))
#define QR5_0 2
@ -184,6 +187,13 @@ typedef struct {
} block_q4_1;
static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_half) + QK4_1 / 2, "wrong q4_1 block size/padding");
#define QK_MXFP4 32
typedef struct {
uint8_t e; // E8M0
uint8_t qs[QK_MXFP4/2];
} block_mxfp4;
static_assert(sizeof(block_mxfp4) == sizeof(uint8_t) + QK_MXFP4/2, "wrong mxfp4 block size/padding");
#define QK5_0 32
typedef struct {
ggml_half d; // delta
@ -1074,10 +1084,17 @@ GGML_TABLE_BEGIN(uint32_t, iq3s_grid, 512)
0x0f090307, 0x0f090501, 0x0f090b01, 0x0f0b0505, 0x0f0b0905, 0x0f0d0105, 0x0f0d0703, 0x0f0f0101,
GGML_TABLE_END()
// TODO: fix name to kvalues_iq4_nl
GGML_TABLE_BEGIN(int8_t, kvalues_iq4nl, 16)
-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113,
GGML_TABLE_END()
// e2m1 values (doubled)
// ref: https://www.opencompute.org/documents/ocp-microscaling-formats-mx-v1-0-spec-final-pdf
GGML_TABLE_BEGIN(int8_t, kvalues_mxfp4, 16)
0, 1, 2, 3, 4, 6, 8, 12, 0, -1, -2, -3, -4, -6, -8, -12,
GGML_TABLE_END()
#define NGRID_IQ1S 2048
#define IQ1S_DELTA 0.125f
#define IQ1M_DELTA 0.125f

7
ggml/src/ggml-cpu/CMakeLists.txt
View File

@ -70,10 +70,12 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
if (GGML_OPENMP)
find_package(OpenMP)
if (OpenMP_FOUND)
set(GGML_OPENMP_ENABLED "ON" CACHE INTERNAL "")
target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_OPENMP)
target_link_libraries(${GGML_CPU_NAME} PRIVATE OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
else()
set(GGML_OPENMP_ENABLED "OFF" CACHE INTERNAL "")
message(WARNING "OpenMP not found")
endif()
endif()
@ -456,6 +458,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
list(APPEND ARCH_FLAGS -march=z16)
elseif (${S390X_M} MATCHES "9175|9176")
# NOTE: Only available from GCC 15.1.0 onwards. Any z17 machine with compile issues must first verify their GCC version.
# binutils must also be updated to the latest for the -march=z17 flag to work. Otherwise, use -march=arch15.
message(STATUS "z17 target")
list(APPEND ARCH_FLAGS -march=z17)
else()
@ -494,9 +497,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
# Fetch KleidiAI sources:
include(FetchContent)
set(KLEIDIAI_COMMIT_TAG "v1.9.0")
set(KLEIDIAI_COMMIT_TAG "v1.11.0")
set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/archive/refs/tags/${KLEIDIAI_COMMIT_TAG}.tar.gz")
set(KLEIDIAI_ARCHIVE_MD5 "2a8e1bb55d201557553545536489a017")
set(KLEIDIAI_ARCHIVE_MD5 "3fe9e5ab964c375c53839296eb71eaa2")
if (POLICY CMP0135)
cmake_policy(SET CMP0135 NEW)

34
ggml/src/ggml-cpu/arch-fallback.h
View File

@ -13,6 +13,7 @@
#define ggml_vec_dot_q5_0_q8_0_generic ggml_vec_dot_q5_0_q8_0
#define ggml_vec_dot_q5_1_q8_1_generic ggml_vec_dot_q5_1_q8_1
#define ggml_vec_dot_q8_0_q8_0_generic ggml_vec_dot_q8_0_q8_0
#define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
#define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
@ -37,17 +38,25 @@
#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
#elif defined(__aarch64__) || defined(__arm__) || defined(_M_ARM) || defined(_M_ARM64)
// repack.cpp
#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
#elif defined(__x86_64__) || defined(__i386__) || defined(_M_IX86) || defined(_M_X64)
// repack.cpp
#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
@ -64,6 +73,7 @@
#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
#define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
// repack.cpp
#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
@ -72,18 +82,23 @@
#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
#elif defined(__loongarch64)
// quants.c
#define quantize_row_q8_K_generic quantize_row_q8_K
#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
#define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
// repack.cpp
#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
@ -92,12 +107,16 @@
#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
#elif defined(__riscv)
// quants.c
#define quantize_row_q8_K_generic quantize_row_q8_K
@ -112,6 +131,7 @@
#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
#define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
#define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
#define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
// repack.cpp
#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
@ -119,11 +139,15 @@
#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
#elif defined(__s390x__)
// quants.c
#define quantize_row_q8_K_generic quantize_row_q8_K
@ -139,6 +163,7 @@
#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
#define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
// repack.cpp
#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
@ -147,12 +172,16 @@
#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
#elif defined(__wasm__)
// quants.c
#define ggml_vec_dot_q4_1_q8_1_generic ggml_vec_dot_q4_1_q8_1
@ -167,6 +196,7 @@
#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
#define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
#define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
#define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
// repack.cpp
#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
@ -175,10 +205,14 @@
#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
#endif

728
ggml/src/ggml-cpu/arch/arm/quants.c
View File

@ -589,6 +589,67 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
*s = sumf;
}
void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
assert(nrc == 1);
UNUSED(nrc);
UNUSED(bx);
UNUSED(by);
UNUSED(bs);
assert(n % QK_MXFP4 == 0);
static_assert(QK_MXFP4 == QK8_0, "QK_MXFP4 and QK8_0 must be the same");
const block_mxfp4 * GGML_RESTRICT x = vx;
const block_q8_0 * GGML_RESTRICT y = vy;
const int nb = n / QK_MXFP4;
int ib = 0;
float sumf = 0;
#if defined __ARM_NEON
const int8x16_t values = vld1q_s8(kvalues_mxfp4);
const uint8x16_t m4b = vdupq_n_u8(0x0f);
uint8x16x2_t q4bits;
int8x16x4_t q4b;
int8x16x4_t q8b;
int32x4_t prod_1;
int32x4_t prod_2;
for (; ib + 1 < nb; ib += 2) {
q4bits.val[0] = vld1q_u8(x[ib + 0].qs);
q4bits.val[1] = vld1q_u8(x[ib + 1].qs);
q8b.val[0] = vld1q_s8(y[ib + 0].qs);
q8b.val[1] = vld1q_s8(y[ib + 0].qs + 16);
q8b.val[2] = vld1q_s8(y[ib + 1].qs);
q8b.val[3] = vld1q_s8(y[ib + 1].qs + 16);
q4b.val[0] = ggml_vqtbl1q_s8(values, vandq_u8 (q4bits.val[0], m4b));
q4b.val[1] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[0], 4));
q4b.val[2] = ggml_vqtbl1q_s8(values, vandq_u8 (q4bits.val[1], m4b));
q4b.val[3] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[1], 4));
prod_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[0], q8b.val[0]), q4b.val[1], q8b.val[1]);
prod_2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[2], q8b.val[2]), q4b.val[3], q8b.val[3]);
sumf +=
GGML_E8M0_TO_FP32_HALF(x[ib + 0].e) * GGML_CPU_FP16_TO_FP32(y[ib + 0].d) * vaddvq_s32(prod_1) +
GGML_E8M0_TO_FP32_HALF(x[ib + 1].e) * GGML_CPU_FP16_TO_FP32(y[ib + 1].d) * vaddvq_s32(prod_2);
}
#endif
for (; ib < nb; ++ib) {
const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_E8M0_TO_FP32_HALF(x[ib].e);
int sumi1 = 0;
int sumi2 = 0;
for (int j = 0; j < QK_MXFP4/2; ++j) {
sumi1 += y[ib].qs[j + 0] * kvalues_mxfp4[x[ib].qs[j] & 0xf];
sumi2 += y[ib].qs[j + QK_MXFP4/2] * kvalues_mxfp4[x[ib].qs[j] >> 4];
}
sumf += d * (sumi1 + sumi2);
}
*s = sumf;
}
void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
const int qk = QK8_0;
const int nb = n / qk;
@ -1236,44 +1297,10 @@ void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
*s = sumf;
#else
const uint8_t pow3[6] = {1, 3, 9, 27, 81, 243};
float sumf = 0.0f;
for (int i = 0; i < nb; ++i) {
int sum = 0;
for (size_t j = 0; j < sizeof(x->qs) - sizeof(x->qs) % 32; j += 32) {
for (size_t l = 0; l < 5; ++l) {
for (size_t m = 0; m < 32; ++m) {
uint8_t q = x[i].qs[j + m] * pow3[l];
uint16_t xi = ((uint16_t) q * 3) >> 8;
sum += (xi - 1) * y[i].qs[j*5 + l*32 + m];
}
}
}
for (size_t j = sizeof(x->qs) - sizeof(x->qs) % 32; j < sizeof(x->qs); j += 16) {
for (size_t l = 0; l < 5; ++l) {
for (size_t m = 0; m < 16; ++m) {
uint8_t q = x[i].qs[j + m] * pow3[l];
uint16_t xi = ((uint16_t) q * 3) >> 8;
sum += (xi - 1) * y[i].qs[j*5 + l*16 + m];
}
}
}
for (size_t l = 0; l < 4; ++l) {
for (size_t j = 0; j < sizeof(x->qh); ++j) {
uint8_t q = x[i].qh[j] * pow3[l];
uint16_t xi = ((uint16_t) q * 3) >> 8;
sum += (xi - 1) * y[i].qs[sizeof(x->qs)*5 + l*sizeof(x->qh) + j];
}
}
sumf += (float) sum * (GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d);
}
*s = sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_tq1_0_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -1381,25 +1408,10 @@ void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
*s = sumf;
#else
float sumf = 0.0f;
for (int i = 0; i < nb; ++i) {
int32_t sumi = 0;
for (size_t j = 0; j < sizeof(x->qs); j += 32) {
for (size_t l = 0; l < 4; ++l) {
for (size_t k = 0; k < 32; ++k) {
sumi += y[i].qs[j*4 + l*32 + k] * (((x[i].qs[j + k] >> (l*2)) & 3) - 1);
}
}
}
const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
sumf += (float) sumi * d;
}
*s = sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_tq2_0_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -1729,45 +1741,10 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
*s = sum;
#else
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * q2 = x[i].qs;
const int8_t * q8 = y[i].qs;
const uint8_t * sc = x[i].scales;
int summs = 0;
for (int j = 0; j < 16; ++j) {
summs += y[i].bsums[j] * (sc[j] >> 4);
}
const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
int isum = 0;
int is = 0;
int d;
for (int k = 0; k < QK_K/128; ++k) {
int shift = 0;
for (int j = 0; j < 4; ++j) {
d = sc[is++] & 0xF;
int isuml = 0;
for (int l = 0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
isum += d * isuml;
d = sc[is++] & 0xF;
isuml = 0;
for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
isum += d * isuml;
shift += 2;
q8 += 32;
}
q2 += 32;
}
sumf += dall * isum - dmin * summs;
}
*s = sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_q2_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -2057,68 +2034,12 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
*s = sum;
#else
// scalar version
// This function is written like this so the compiler can manage to vectorize most of it
// Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
// manually vectorized version above. Every other version I tried would run at least 4 times slower.
// The ideal situation would be if we could just write the code once, and the compiler would
// automatically produce the best possible set of machine instructions, instead of us having to manually
// write vectorized versions for AVX, ARM_NEON, etc.
int8_t aux8[QK_K];
int16_t aux16[8];
float sums [8];
int32_t aux32[8];
memset(sums, 0, 8*sizeof(float));
uint32_t auxs[4];
const int8_t * scales = (const int8_t*)auxs;
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * GGML_RESTRICT q3 = x[i].qs;
const uint8_t * GGML_RESTRICT hm = x[i].hmask;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
memset(aux32, 0, 8*sizeof(int32_t));
int8_t * GGML_RESTRICT a = aux8;
uint8_t m = 1;
for (int j = 0; j < QK_K; j += 128) {
for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3;
for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
a += 32; m <<= 1;
for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
a += 32; m <<= 1;
for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
a += 32; m <<= 1;
for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
a += 32; m <<= 1;
q3 += 32;
}
a = aux8;
memcpy(auxs, x[i].scales, 12);
uint32_t tmp = auxs[2];
auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
for (int j = 0; j < QK_K/16; ++j) {
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
q8 += 8; a += 8;
}
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
}
for (int l = 0; l < 8; ++l) sumf += sums[l];
*s = sumf;
UNUSED(kmask1);
UNUSED(kmask2);
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_q3_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -2431,61 +2352,14 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
*s = sumf;
#else
const uint8_t * scales = (const uint8_t*)&utmp[0];
const uint8_t * mins = (const uint8_t*)&utmp[2];
int8_t aux8[QK_K];
int16_t aux16[8];
float sums [8];
int32_t aux32[8];
memset(sums, 0, 8*sizeof(float));
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * GGML_RESTRICT q4 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
memset(aux32, 0, 8*sizeof(int32_t));
int8_t * GGML_RESTRICT a = aux8;
for (int j = 0; j < QK_K/64; ++j) {
for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
a += 32;
for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] >> 4);
a += 32; q4 += 32;
}
memcpy(utmp, x[i].scales, 12);
utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
const uint32_t uaux = utmp[1] & kmask1;
utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
utmp[2] = uaux;
utmp[0] &= kmask1;
int sumi = 0;
for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
a = aux8;
int is = 0;
for (int j = 0; j < QK_K/32; ++j) {
int32_t scale = scales[is++];
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
}
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
sumf -= dmin * sumi;
}
for (int l = 0; l < 8; ++l) sumf += sums[l];
*s = sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
UNUSED(kmask1);
UNUSED(kmask2);
UNUSED(kmask3);
UNUSED(utmp);
ggml_vec_dot_q4_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -2578,66 +2452,14 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
*s = sumf;
#else
const uint8_t * scales = (const uint8_t*)&utmp[0];
const uint8_t * mins = (const uint8_t*)&utmp[2];
int8_t aux8[QK_K];
int16_t aux16[8];
float sums [8];
int32_t aux32[8];
memset(sums, 0, 8*sizeof(float));
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * GGML_RESTRICT q4 = x[i].qs;
const uint8_t * GGML_RESTRICT hm = x[i].qh;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
memset(aux32, 0, 8*sizeof(int32_t));
int8_t * GGML_RESTRICT a = aux8;
uint8_t m = 1;
for (int j = 0; j < QK_K/64; ++j) {
for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
a += 32; m <<= 1;
for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] >> 4);
for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
a += 32; m <<= 1;
q4 += 32;
}
memcpy(utmp, x[i].scales, 12);
utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
const uint32_t uaux = utmp[1] & kmask1;
utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
utmp[2] = uaux;
utmp[0] &= kmask1;
int sumi = 0;
for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
a = aux8;
int is = 0;
for (int j = 0; j < QK_K/32; ++j) {
int32_t scale = scales[is++];
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
}
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
sumf -= dmin * sumi;
}
for (int l = 0; l < 8; ++l) sumf += sums[l];
*s = sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
UNUSED(kmask1);
UNUSED(kmask2);
UNUSED(kmask3);
UNUSED(utmp);
ggml_vec_dot_q5_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -3093,47 +2915,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
}
*s = sum;
#else
int8_t aux8[QK_K];
int16_t aux16[8];
float sums [8];
int32_t aux32[8];
memset(sums, 0, 8*sizeof(float));
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * GGML_RESTRICT q4 = x[i].ql;
const uint8_t * GGML_RESTRICT qh = x[i].qh;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
memset(aux32, 0, 8*sizeof(int32_t));
int8_t * GGML_RESTRICT a = aux8;
for (int j = 0; j < QK_K; j += 128) {
for (int l = 0; l < 32; ++l) {
a[l + 0] = (int8_t)((q4[l + 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
a[l + 64] = (int8_t)((q4[l + 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32;
a[l + 96] = (int8_t)((q4[l + 32] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32;
}
a += 128;
q4 += 64;
qh += 32;
}
a = aux8;
int is = 0;
for (int j = 0; j < QK_K/16; ++j) {
int scale = x[i].scales[is++];
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
}
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
}
for (int l = 0; l < 8; ++l) sumf += sums[l];
*s = sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_q6_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -3229,34 +3014,10 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
*s = 0.25f * sumf;
#else
uint32_t aux32[2];
const uint8_t * aux8 = (const uint8_t *)aux32;
float sumf = 0.f;
for (int i = 0; i < nb; ++i) {
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint16_t * GGML_RESTRICT q2 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
int32_t bsum = 0;
for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
memcpy(aux32, q2, 2*sizeof(uint32_t));
q2 += 4;
const uint32_t ls = 2*(aux32[1] >> 28) + 1;
int32_t sumi = 0;
for (int l = 0; l < 4; ++l) {
const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]);
const uint8_t signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127];
for (int j = 0; j < 8; ++j) {
sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
}
q8 += 8;
}
bsum += sumi * ls;
}
sumf += d * bsum;
}
*s = 0.125f * sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_iq2_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -3327,42 +3088,10 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
*s = 0.125f * sumf;
#else
float sumf = 0.f;
for (int i = 0; i < nb; ++i) {
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint16_t * GGML_RESTRICT q2 = x[i].qs;
const uint8_t * GGML_RESTRICT sc = x[i].scales;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
int32_t bsum = 0;
for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
const uint16_t ls1 = 2*(sc[ib32] & 0xf) + 1;
const uint16_t ls2 = 2*(sc[ib32] >> 4) + 1;
int32_t sumi = 0;
for (int l = 0; l < 2; ++l) {
const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
const uint8_t signs = ksigns_iq2xs[q2[l] >> 9];
for (int j = 0; j < 8; ++j) {
sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
}
q8 += 8;
}
bsum += sumi * ls1;
sumi = 0;
for (int l = 2; l < 4; ++l) {
const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
const uint8_t signs = ksigns_iq2xs[q2[l] >> 9];
for (int j = 0; j < 8; ++j) {
sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
}
q8 += 8;
}
bsum += sumi * ls2;
q2 += 4;
}
sumf += d * bsum;
}
*s = 0.125f * sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_iq2_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -3455,45 +3184,10 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
*s = 0.125f * sumf;
#else
float sumf = 0;
for (int i = 0; i < nb; i++) {
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const int8_t * q8 = y[i].qs;
const uint8_t * qs = x[i].qs;
const uint8_t * qh = x[i].qh;
const uint8_t * signs = qs + QK_K/8;
int bsum = 0;
for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
int ls1 = 1 + 2*(x[i].scales[ib32] & 0xf);
int ls2 = 1 + 2*(x[i].scales[ib32] >> 4);
int sumi1 = 0, sumi2 = 0;
for (int l = 0; l < 2; ++l) {
const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
for (int j = 0; j < 8; ++j) {
sumi1 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
}
q8 += 8;
}
for (int l = 2; l < 4; ++l) {
const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
for (int j = 0; j < 8; ++j) {
sumi2 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
}
q8 += 8;
}
bsum += ls1 * sumi1 + ls2 * sumi2;
qs += 4;
signs += 4;
}
sumf += d * bsum;
}
*s = 0.125f * sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_iq2_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -3553,36 +3247,10 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
*s = 0.5f * sumf;
#else
uint32_t aux32;
float sumf = 0.f;
for (int i = 0; i < nb; ++i) {
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint8_t * GGML_RESTRICT q3 = x[i].qs;
const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
int32_t bsum = 0;
for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
memcpy(&aux32, gas, sizeof(uint32_t)); gas += sizeof(uint32_t);
const uint32_t ls = 2*(aux32 >> 28) + 1;
int32_t sumi = 0;
for (int l = 0; l < 4; ++l) {
const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*l+0]);
const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*l+1]);
const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*l) & 127];
for (int j = 0; j < 4; ++j) {
sumi += grid1[j] * q8[j+0] * (signs & kmask_iq2xs[j+0] ? -1 : 1);
sumi += grid2[j] * q8[j+4] * (signs & kmask_iq2xs[j+4] ? -1 : 1);
}
q8 += 8;
}
q3 += 8;
bsum += sumi * ls;
}
sumf += d * bsum;
}
*s = 0.25f * sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_iq3_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -3689,48 +3357,10 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
*s = sumf;
#else
float sumf = 0.f;
for (int i = 0; i < nb; ++i) {
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint8_t * GGML_RESTRICT qs = x[i].qs;
const uint8_t * GGML_RESTRICT qh = x[i].qh;
const uint8_t * GGML_RESTRICT signs = x[i].signs;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
int32_t bsum = 0;
for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
const uint32_t ls1 = 2*(x[i].scales[ib32/2] & 0xf) + 1;
const uint32_t ls2 = 2*(x[i].scales[ib32/2] >> 4) + 1;
int32_t sumi = 0;
for (int l = 0; l < 4; ++l) {
const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+0] << (8-2*l)) & 256)));
const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+0] << (7-2*l)) & 256)));
for (int j = 0; j < 4; ++j) {
sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
}
q8 += 8;
}
qs += 8;
signs += 4;
bsum += sumi * ls1;
sumi = 0;
for (int l = 0; l < 4; ++l) {
const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+1] << (8-2*l)) & 256)));
const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+1] << (7-2*l)) & 256)));
for (int j = 0; j < 4; ++j) {
sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
}
q8 += 8;
}
qs += 8;
signs += 4;
bsum += sumi * ls2;
}
sumf += d * bsum;
}
*s = sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_iq3_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -3793,36 +3423,10 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
*s = sumf;
#else
float sumf = 0;
for (int i = 0; i < nb; i++) {
const int8_t * q8 = y[i].qs;
const uint8_t * qs = x[i].qs;
const uint16_t * qh = x[i].qh;
int sumi = 0, sumi1 = 0;
for (int ib = 0; ib < QK_K/32; ++ib) {
const int ls = 2*((qh[ib] >> 12) & 7) + 1;
const int delta = qh[ib] & 0x8000 ? -1 : 1;
int lsum = 0;
for (int l = 0; l < 4; ++l) {
const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((qh[ib] >> 3*l) & 7) << 8)));
for (int j = 0; j < 8; ++j) {
lsum += q8[j] * grid[j];
}
q8 += 8;
}
sumi += ls * lsum;
sumi1 += ls * delta * (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]);
qs += 4;
}
sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
}
*s = sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_iq1_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -3912,52 +3516,11 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
*s = sumf;
#else
int sum1[2], sum2[2], delta[4];
float sumf = 0;
for (int i = 0; i < nb; i++) {
const int8_t * q8 = y[i].qs;
const uint8_t * qs = x[i].qs;
const uint8_t * qh = x[i].qh;
const uint16_t * sc = (const uint16_t *)x[i].scales;
scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
int sumi1 = 0, sumi2 = 0;
for (int ib = 0; ib < QK_K/32; ++ib) {
delta[0] = qh[0] & 0x08 ? -1 : 1;
delta[1] = qh[0] & 0x80 ? -1 : 1;
delta[2] = qh[1] & 0x08 ? -1 : 1;
delta[3] = qh[1] & 0x80 ? -1 : 1;
sum1[0] = sum1[1] = sum2[0] = sum2[1] = 0;
for (int l = 0; l < 4; ++l) {
const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((uint16_t)qh[l/2] << (8 - 4*(l%2))) & 0x700)));
int lsum1 = 0, lsum2 = 0;
for (int j = 0; j < 8; ++j) {
lsum1 += q8[j] * grid[j];
lsum2 += q8[j];
}
q8 += 8;
sum1[l/2] += lsum1;
sum2[l/2] += lsum2*delta[l];
}
const int ls1 = 2*((sc[ib/2] >> (6*(ib%2)+0)) & 0x7) + 1;
const int ls2 = 2*((sc[ib/2] >> (6*(ib%2)+3)) & 0x7) + 1;
sumi1 += sum1[0] * ls1 + sum1[1] * ls2;
sumi2 += sum2[0] * ls1 + sum2[1] * ls2;
qs += 4;
qh += 2;
}
sumf += GGML_CPU_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
}
*s = sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
UNUSED(scale);
ggml_vec_dot_iq1_m_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -4078,37 +3641,10 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
*s = sumf;
#else
float sumf = 0;
for (int ibl = 0; ibl < nb; ++ibl) {
const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
uint16_t h = x[ibl].scales_h;
const uint8_t * qs = x[ibl].qs;
const int8_t * q8 = y[ibl].qs;
for (int ib = 0; ib < QK_K/32; ib += 2) {
const uint8_t ls1 = (x[ibl].scales_l[ib/2] & 0xf) | ((h << 4) & 0x30);
const uint8_t ls2 = (x[ibl].scales_l[ib/2] >> 4) | ((h << 2) & 0x30);
h >>= 4;
const float d1 = d4d8*(ls1 - 32);
const float d2 = d4d8*(ls2 - 32);
int sumi1 = 0, sumi2 = 0;
for (int j = 0; j < 16; ++j) {
sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >> 4];
}
sumf += d1 * (sumi1 + sumi2);
qs += 16;
q8 += 32;
sumi1 = sumi2 = 0;
for (int j = 0; j < 16; ++j) {
sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >> 4];
}
sumf += d2 * (sumi1 + sumi2);
qs += 16;
q8 += 32;
}
}
*s = sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_iq4_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}

300
ggml/src/ggml-cpu/arch/arm/repack.cpp
View File

@ -86,35 +86,9 @@ void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTR
}
}
#else
// scalar
const int blck_size_interleave = 4;
float srcv[4][QK8_0];
float id[4];
for (int i = 0; i < nb; i++) {
for (int row_iter = 0; row_iter < 4; row_iter++) {
float amax = 0.0f; // absolute max
for (int j = 0; j < QK8_0; j++) {
srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j];
amax = MAX(amax, fabsf(srcv[row_iter][j]));
}
const float d = amax / ((1 << 7) - 1);
id[row_iter] = d ? 1.0f / d : 0.0f;
y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
}
for (int j = 0; j < QK8_0 * 4; j++) {
int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave;
int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave;
src_offset += (j % blck_size_interleave);
float x0 = srcv[src_id][src_offset] * id[src_id];
y[i].qs[j] = roundf(x0);
}
}
UNUSED(nb);
UNUSED(y);
ggml_quantize_mat_q8_0_4x4_generic(x, vy, k);
#endif
}
@ -205,35 +179,9 @@ void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR
}
#else
// scalar
const int blck_size_interleave = 8;
float srcv[4][QK8_0];
float id[4];
for (int i = 0; i < nb; i++) {
for (int row_iter = 0; row_iter < 4; row_iter++) {
float amax = 0.0f; // absolute max
for (int j = 0; j < QK8_0; j++) {
srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j];
amax = MAX(amax, fabsf(srcv[row_iter][j]));
}
const float d = amax / ((1 << 7) - 1);
id[row_iter] = d ? 1.0f / d : 0.0f;
y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
}
for (int j = 0; j < QK8_0 * 4; j++) {
int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave;
int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave;
src_offset += (j % blck_size_interleave);
float x0 = srcv[src_id][src_offset] * id[src_id];
y[i].qs[j] = roundf(x0);
}
}
UNUSED(nb);
UNUSED(y);
ggml_quantize_mat_q8_0_4x8_generic(x, vy, k);
#endif
}
@ -295,29 +243,7 @@ void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
}
return;
#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
float sumf[4];
int sumi;
const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
for (int x = 0; x < nc / ncols_interleaved; x++) {
const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
for (int l = 0; l < nb; l++) {
for (int k = 0; k < (qk / (2 * blocklen)); k++) {
for (int j = 0; j < ncols_interleaved; j++) {
sumi = 0;
for (int i = 0; i < blocklen; ++i) {
const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
}
sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
}
}
}
for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
}
ggml_gemv_q4_0_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc);
}
void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@ -383,29 +309,7 @@ void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
}
return;
#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
float sumf[4];
int sumi;
const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
for (int x = 0; x < nc / ncols_interleaved; x++) {
const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
for (int l = 0; l < nb; l++) {
for (int k = 0; k < (qk / (2 * blocklen)); k++) {
for (int j = 0; j < ncols_interleaved; j++) {
sumi = 0;
for (int i = 0; i < blocklen; ++i) {
const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
}
sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
}
}
}
for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
}
ggml_gemv_q4_0_4x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
}
void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@ -497,31 +401,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
#endif // #if defined(__ARM_FEATURE_SVE)
#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
{
float sumf[8];
int sumi;
const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
for (int x = 0; x < nc / ncols_interleaved; x++) {
const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
for (int l = 0; l < nb; l++) {
for (int k = 0; k < (qk / (2 * blocklen)); k++) {
for (int j = 0; j < ncols_interleaved; j++) {
sumi = 0;
for (int i = 0; i < blocklen; ++i) {
const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
}
sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
}
}
}
for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
}
}
ggml_gemv_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
}
void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@ -591,31 +471,7 @@ void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
}
return;
#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
{
float sumf[4];
int sumi;
const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
for (int x = 0; x < nc / ncols_interleaved; x++) {
const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
for (int l = 0; l < nb; l++) {
for (int k = 0; k < (qk / (2 * blocklen)); k++) {
for (int j = 0; j < ncols_interleaved; j++) {
sumi = 0;
for (int i = 0; i < blocklen; ++i) {
const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
}
sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
}
}
}
for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
}
}
ggml_gemv_iq4_nl_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc);
}
void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@ -1096,40 +952,7 @@ void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
);
return;
#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
{
float sumf[4][4];
int sumi;
for (int y = 0; y < nr / 4; y++) {
const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
for (int x = 0; x < nc / ncols_interleaved; x++) {
const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
}
for (int l = 0; l < nb; l++) {
for (int k = 0; k < (qk / (2 * blocklen)); k++) {
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++) {
sumi = 0;
for (int i = 0; i < blocklen; ++i) {
const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
(v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
}
sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
}
}
}
}
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++)
s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
}
}
}
}
ggml_gemm_q4_0_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc);
}
void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@ -1550,38 +1373,7 @@ void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
);
return;
#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
float sumf[4][4];
int sumi;
for (int y = 0; y < nr / 4; y++) {
const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
for (int x = 0; x < nc / ncols_interleaved; x++) {
const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
}
for (int l = 0; l < nb; l++) {
for (int k = 0; k < (qk / (2 * blocklen)); k++) {
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++) {
sumi = 0;
for (int i = 0; i < blocklen; ++i) {
const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
(v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
}
sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
}
}
}
}
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++)
s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
}
}
}
ggml_gemm_q4_0_4x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
}
void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@ -2019,38 +1811,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
#endif // #if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
float sumf[4][8];
int sumi;
for (int y = 0; y < nr / 4; y++) {
const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
for (int x = 0; x < nc / ncols_interleaved; x++) {
const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
}
for (int l = 0; l < nb; l++) {
for (int k = 0; k < (qk / (2 * blocklen)); k++) {
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++) {
sumi = 0;
for (int i = 0; i < blocklen; ++i) {
const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
(v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
}
sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
}
}
}
}
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++)
s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
}
}
}
ggml_gemm_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
}
void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@ -2126,38 +1887,5 @@ void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
}
return;
#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
{
float sumf[4][4];
int sumi;
for (int y = 0; y < nr / 4; y++) {
const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
for (int x = 0; x < nc / ncols_interleaved; x++) {
const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
}
for (int l = 0; l < nb; l++) {
for (int k = 0; k < (qk / (2 * blocklen)); k++) {
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++) {
sumi = 0;
for (int i = 0; i < blocklen; ++i) {
const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
(v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4]));
}
sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
}
}
}
}
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++)
s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
}
}
}
}
ggml_gemm_iq4_nl_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc);
}

661
ggml/src/ggml-cpu/arch/loongarch/quants.c
View File

@ -544,7 +544,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
__m128 max4 = __lsx_vfmax_s( lasx_extractf128( max_abs, 1 ), lasx_extractf128( max_abs, 0) );
max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vpickod_d((__m128i) max4, (__m128i)max4 ) );
__m128 tmp = max4;
max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vextrins_w((__m128i)tmp, (__m128i)max4, 0x10 ));
max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vextrins_w((__m128i)tmp, (__m128i)max4, 0x1 ));
const float max_scalar = ((v4f32)max4)[0];
// Quantize these floats
@ -821,24 +821,15 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
sumf = hsum_float_8(acc) + summs;
#endif
for (; ib < nb; ++ib) {
int sumi0 = 0;
int sumi1 = 0;
for (int j = 0; j < qk/2; ++j) {
const int v0 = (x[ib].qs[j] & 0x0F);
const int v1 = (x[ib].qs[j] >> 4);
sumi0 += (v0 * y[ib].qs[j]);
sumi1 += (v1 * y[ib].qs[j + qk/2]);
}
int sumi = sumi0 + sumi1;
sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
}
*s = sumf;
#else
UNUSED(nb);
UNUSED(x);
UNUSED(y);
UNUSED(ib);
UNUSED(sumf);
ggml_vec_dot_q4_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@ -883,30 +874,15 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
sumf = hsum_float_8(acc);
#endif
for (; ib < nb; ++ib) {
uint32_t qh;
memcpy(&qh, x[ib].qh, sizeof(qh));
int sumi0 = 0;
int sumi1 = 0;
for (int j = 0; j < qk/2; ++j) {
const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16);
const int32_t x1 = (int8_t)(((x[ib].qs[j] >> 4) | xh_1) - 16);
sumi0 += (x0 * y[ib].qs[j]);
sumi1 += (x1 * y[ib].qs[j + qk/2]);
}
int sumi = sumi0 + sumi1;
sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
}
*s = sumf;
#else
UNUSED(nb);
UNUSED(ib);
UNUSED(sumf);
UNUSED(x);
UNUSED(y);
ggml_vec_dot_q5_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@ -954,30 +930,15 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
sumf = hsum_float_8(acc) + summs;
#endif
for (; ib < nb; ++ib) {
uint32_t qh;
memcpy(&qh, x[ib].qh, sizeof(qh));
int sumi0 = 0;
int sumi1 = 0;
for (int j = 0; j < qk/2; ++j) {
const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10;
const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10;
const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0;
const int32_t x1 = (x[ib].qs[j] >> 4) | xh_1;
sumi0 += (x0 * y[ib].qs[j]);
sumi1 += (x1 * y[ib].qs[j + qk/2]);
}
int sumi = sumi0 + sumi1;
sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
}
*s = sumf;
#else
UNUSED(nb);
UNUSED(ib);
UNUSED(sumf);
UNUSED(x);
UNUSED(y);
ggml_vec_dot_q5_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@ -1016,18 +977,15 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
sumf = hsum_float_8(acc);
#endif
for (; ib < nb; ++ib) {
int sumi = 0;
for (int j = 0; j < qk; j++) {
sumi += x[ib].qs[j]*y[ib].qs[j];
}
sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
}
*s = sumf;
#else
UNUSED(nb);
UNUSED(ib);
UNUSED(sumf);
UNUSED(x);
UNUSED(y);
ggml_vec_dot_q8_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@ -1103,45 +1061,10 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
*s = hsum_float_8(acc);
#else
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * q2 = x[i].qs;
const int8_t * q8 = y[i].qs;
const uint8_t * sc = x[i].scales;
int summs = 0;
for (int j = 0; j < 16; ++j) {
summs += y[i].bsums[j] * (sc[j] >> 4);
}
const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
int isum = 0;
int is = 0;
int d;
for (int k = 0; k < QK_K/128; ++k) {
int shift = 0;
for (int j = 0; j < 4; ++j) {
d = sc[is++] & 0xF;
int isuml = 0;
for (int l = 0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
isum += d * isuml;
d = sc[is++] & 0xF;
isuml = 0;
for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
isum += d * isuml;
shift += 2;
q8 += 32;
}
q2 += 32;
}
sumf += dall * isum - dmin * summs;
}
*s = sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_q2_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -1239,70 +1162,13 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
*s = hsum_float_8(acc);
#else
// scalar version
// This function is written like this so the compiler can manage to vectorize most of it
// Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
// manually vectorized version above. Every other version I tried would run at least 4 times slower.
// The ideal situation would be if we could just write the code once, and the compiler would
// automatically produce the best possible set of machine instructions, instead of us having to manually
// write vectorized versions for AVX, ARM_NEON, etc.
int8_t aux8[QK_K];
int16_t aux16[8];
float sums [8];
int32_t aux32[8];
memset(sums, 0, 8*sizeof(float));
uint32_t auxs[4];
const int8_t * scales = (const int8_t*)auxs;
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * GGML_RESTRICT q3 = x[i].qs;
const uint8_t * GGML_RESTRICT hm = x[i].hmask;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
memset(aux32, 0, 8*sizeof(int32_t));
int8_t * GGML_RESTRICT a = aux8;
uint8_t m = 1;
for (int j = 0; j < QK_K; j += 128) {
for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3;
for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
a += 32; m <<= 1;
for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
a += 32; m <<= 1;
for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
a += 32; m <<= 1;
for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
a += 32; m <<= 1;
q3 += 32;
}
a = aux8;
memcpy(auxs, x[i].scales, 12);
uint32_t tmp = auxs[2];
auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
for (int j = 0; j < QK_K/16; ++j) {
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
q8 += 8; a += 8;
}
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
}
for (int l = 0; l < 8; ++l) sumf += sums[l];
*s = sumf;
UNUSED(kmask1);
UNUSED(kmask2);
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_q3_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@ -1391,61 +1257,14 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
*s = hsum_float_8(acc) + ((v4f32)acc_m)[0];
#else
const uint8_t * scales = (const uint8_t*)&utmp[0];
const uint8_t * mins = (const uint8_t*)&utmp[2];
int8_t aux8[QK_K];
int16_t aux16[8];
float sums [8];
int32_t aux32[8];
memset(sums, 0, 8*sizeof(float));
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * GGML_RESTRICT q4 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
memset(aux32, 0, 8*sizeof(int32_t));
int8_t * GGML_RESTRICT a = aux8;
for (int j = 0; j < QK_K/64; ++j) {
for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
a += 32;
for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] >> 4);
a += 32; q4 += 32;
}
memcpy(utmp, x[i].scales, 12);
utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
const uint32_t uaux = utmp[1] & kmask1;
utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
utmp[2] = uaux;
utmp[0] &= kmask1;
int sumi = 0;
for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
a = aux8;
int is = 0;
for (int j = 0; j < QK_K/32; ++j) {
int32_t scale = scales[is++];
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
}
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
sumf -= dmin * sumi;
}
for (int l = 0; l < 8; ++l) sumf += sums[l];
*s = sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
UNUSED(kmask1);
UNUSED(kmask2);
UNUSED(kmask3);
UNUSED(utmp);
ggml_vec_dot_q4_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -1541,66 +1360,14 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
*s = hsum_float_8(acc) + ((v4f32)acc_m)[0];
#else
const uint8_t * scales = (const uint8_t*)&utmp[0];
const uint8_t * mins = (const uint8_t*)&utmp[2];
int8_t aux8[QK_K];
int16_t aux16[8];
float sums [8];
int32_t aux32[8];
memset(sums, 0, 8*sizeof(float));
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * GGML_RESTRICT q4 = x[i].qs;
const uint8_t * GGML_RESTRICT hm = x[i].qh;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
memset(aux32, 0, 8*sizeof(int32_t));
int8_t * GGML_RESTRICT a = aux8;
uint8_t m = 1;
for (int j = 0; j < QK_K/64; ++j) {
for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
a += 32; m <<= 1;
for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] >> 4);
for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
a += 32; m <<= 1;
q4 += 32;
}
memcpy(utmp, x[i].scales, 12);
utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
const uint32_t uaux = utmp[1] & kmask1;
utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
utmp[2] = uaux;
utmp[0] &= kmask1;
int sumi = 0;
for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
a = aux8;
int is = 0;
for (int j = 0; j < QK_K/32; ++j) {
int32_t scale = scales[is++];
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
}
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
sumf -= dmin * sumi;
}
for (int l = 0; l < 8; ++l) sumf += sums[l];
*s = sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
UNUSED(kmask1);
UNUSED(kmask2);
UNUSED(kmask3);
UNUSED(utmp);
ggml_vec_dot_q5_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -1678,47 +1445,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
*s = hsum_float_8(acc);
#else
int8_t aux8[QK_K];
int16_t aux16[8];
float sums [8];
int32_t aux32[8];
memset(sums, 0, 8*sizeof(float));
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * GGML_RESTRICT q4 = x[i].ql;
const uint8_t * GGML_RESTRICT qh = x[i].qh;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
memset(aux32, 0, 8*sizeof(int32_t));
int8_t * GGML_RESTRICT a = aux8;
for (int j = 0; j < QK_K; j += 128) {
for (int l = 0; l < 32; ++l) {
a[l + 0] = (int8_t)((q4[l + 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
a[l + 64] = (int8_t)((q4[l + 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32;
a[l + 96] = (int8_t)((q4[l + 32] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32;
}
a += 128;
q4 += 64;
qh += 32;
}
a = aux8;
int is = 0;
for (int j = 0; j < QK_K/16; ++j) {
int scale = x[i].scales[is++];
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
}
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
}
for (int l = 0; l < 8; ++l) sumf += sums[l];
*s = sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_q6_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -1815,34 +1545,10 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
*s = 0.125f * hsum_float_8(accumf);
#else
uint32_t aux32[2];
const uint8_t * aux8 = (const uint8_t *)aux32;
float sumf = 0.f;
for (int i = 0; i < nb; ++i) {
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint16_t * GGML_RESTRICT q2 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
int32_t bsum = 0;
for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
memcpy(aux32, q2, 2*sizeof(uint32_t));
q2 += 4;
const uint32_t ls = 2*(aux32[1] >> 28) + 1;
int32_t sumi = 0;
for (int l = 0; l < 4; ++l) {
const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]);
const uint8_t signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127];
for (int j = 0; j < 8; ++j) {
sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
}
q8 += 8;
}
bsum += sumi * ls;
}
sumf += d * bsum;
}
*s = 0.125f * sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_iq2_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -1978,42 +1684,10 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
*s = 0.125f * hsum_float_8(accumf);
#else
float sumf = 0.f;
for (int i = 0; i < nb; ++i) {
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint16_t * GGML_RESTRICT q2 = x[i].qs;
const uint8_t * GGML_RESTRICT sc = x[i].scales;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
int32_t bsum = 0;
for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
const uint16_t ls1 = 2*(sc[ib32] & 0xf) + 1;
const uint16_t ls2 = 2*(sc[ib32] >> 4) + 1;
int32_t sumi = 0;
for (int l = 0; l < 2; ++l) {
const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
const uint8_t signs = ksigns_iq2xs[q2[l] >> 9];
for (int j = 0; j < 8; ++j) {
sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
}
q8 += 8;
}
bsum += sumi * ls1;
sumi = 0;
for (int l = 2; l < 4; ++l) {
const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
const uint8_t signs = ksigns_iq2xs[q2[l] >> 9];
for (int j = 0; j < 8; ++j) {
sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
}
q8 += 8;
}
bsum += sumi * ls2;
q2 += 4;
}
sumf += d * bsum;
}
*s = 0.125f * sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_iq2_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -2105,47 +1779,11 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
*s = 0.125f * hsum_float_8(accumf);
#else
float sumf = 0;
for (int i = 0; i < nb; i++) {
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const int8_t * q8 = y[i].qs;
const uint8_t * qs = x[i].qs;
const uint8_t * qh = x[i].qh;
const uint8_t * signs = qs + QK_K/8;
int bsum = 0;
for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
int ls1 = 1 + 2*(x[i].scales[ib32] & 0xf);
int ls2 = 1 + 2*(x[i].scales[ib32] >> 4);
int sumi1 = 0, sumi2 = 0;
for (int l = 0; l < 2; ++l) {
const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
for (int j = 0; j < 8; ++j) {
sumi1 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
}
q8 += 8;
}
for (int l = 2; l < 4; ++l) {
const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
for (int j = 0; j < 8; ++j) {
sumi2 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
}
q8 += 8;
}
bsum += ls1 * sumi1 + ls2 * sumi2;
qs += 4;
signs += 4;
}
sumf += d * bsum;
}
*s = 0.125f * sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_iq2_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@ -2209,36 +1847,10 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
*s = 0.25f * hsum_float_8(accumf);
#else
uint32_t aux32;
float sumf = 0.f;
for (int i = 0; i < nb; ++i) {
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint8_t * GGML_RESTRICT q3 = x[i].qs;
const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
int32_t bsum = 0;
for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
memcpy(&aux32, gas, sizeof(uint32_t)); gas += sizeof(uint32_t);
const uint32_t ls = 2*(aux32 >> 28) + 1;
int32_t sumi = 0;
for (int l = 0; l < 4; ++l) {
const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*l+0]);
const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*l+1]);
const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*l) & 127];
for (int j = 0; j < 4; ++j) {
sumi += grid1[j] * q8[j+0] * (signs & kmask_iq2xs[j+0] ? -1 : 1);
sumi += grid2[j] * q8[j+4] * (signs & kmask_iq2xs[j+4] ? -1 : 1);
}
q8 += 8;
}
q3 += 8;
bsum += sumi * ls;
}
sumf += d * bsum;
}
*s = 0.25f * sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_iq3_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -2338,48 +1950,10 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
*s = hsum_float_8(accumf);
#else
float sumf = 0.f;
for (int i = 0; i < nb; ++i) {
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint8_t * GGML_RESTRICT qs = x[i].qs;
const uint8_t * GGML_RESTRICT qh = x[i].qh;
const uint8_t * GGML_RESTRICT signs = x[i].signs;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
int32_t bsum = 0;
for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
const uint32_t ls1 = 2*(x[i].scales[ib32/2] & 0xf) + 1;
const uint32_t ls2 = 2*(x[i].scales[ib32/2] >> 4) + 1;
int32_t sumi = 0;
for (int l = 0; l < 4; ++l) {
const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+0] << (8-2*l)) & 256)));
const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+0] << (7-2*l)) & 256)));
for (int j = 0; j < 4; ++j) {
sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
}
q8 += 8;
}
qs += 8;
signs += 4;
bsum += sumi * ls1;
sumi = 0;
for (int l = 0; l < 4; ++l) {
const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+1] << (8-2*l)) & 256)));
const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+1] << (7-2*l)) & 256)));
for (int j = 0; j < 4; ++j) {
sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
}
q8 += 8;
}
qs += 8;
signs += 4;
bsum += sumi * ls2;
}
sumf += d * bsum;
}
*s = sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_iq3_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -2460,36 +2034,10 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
*s = hsum_float_8(accum) + IQ1S_DELTA * accum1;
#else
float sumf = 0;
for (int i = 0; i < nb; i++) {
const int8_t * q8 = y[i].qs;
const uint8_t * qs = x[i].qs;
const uint16_t * qh = x[i].qh;
int sumi = 0, sumi1 = 0;
for (int ib = 0; ib < QK_K/32; ++ib) {
const int ls = 2*((qh[ib] >> 12) & 7) + 1;
const int delta = qh[ib] & 0x8000 ? -1 : 1;
int lsum = 0;
for (int l = 0; l < 4; ++l) {
const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((qh[ib] >> 3*l) & 7) << 8)));
for (int j = 0; j < 8; ++j) {
lsum += q8[j] * grid[j];
}
q8 += 8;
}
sumi += ls * lsum;
sumi1 += ls * delta * (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]);
qs += 4;
}
sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
}
*s = sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_iq1_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -2603,37 +2151,10 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
*s = hsum_float_8(accum);
#else
float sumf = 0;
for (int ibl = 0; ibl < nb; ++ibl) {
const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
uint16_t h = x[ibl].scales_h;
const uint8_t * qs = x[ibl].qs;
const int8_t * q8 = y[ibl].qs;
for (int ib = 0; ib < QK_K/32; ib += 2) {
const uint8_t ls1 = (x[ibl].scales_l[ib/2] & 0xf) | ((h << 4) & 0x30);
const uint8_t ls2 = (x[ibl].scales_l[ib/2] >> 4) | ((h << 2) & 0x30);
h >>= 4;
const float d1 = d4d8*(ls1 - 32);
const float d2 = d4d8*(ls2 - 32);
int sumi1 = 0, sumi2 = 0;
for (int j = 0; j < 16; ++j) {
sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >> 4];
}
sumf += d1 * (sumi1 + sumi2);
qs += 16;
q8 += 32;
sumi1 = sumi2 = 0;
for (int j = 0; j < 16; ++j) {
sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >> 4];
}
sumf += d2 * (sumi1 + sumi2);
qs += 16;
q8 += 32;
}
}
*s = sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_iq4_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}

699
ggml/src/ggml-cpu/arch/powerpc/quants.c
View File

@ -201,24 +201,14 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
sumf = vec_extract(vsumf0, 0);
#endif
for (; ib < nb; ++ib) {
int sumi0 = 0;
int sumi1 = 0;
for (int j = 0; j < qk/2; ++j) {
const int v0 = (x[ib].qs[j] & 0x0F) - 8;
const int v1 = (x[ib].qs[j] >> 4) - 8;
sumi0 += (v0 * y[ib].qs[j]);
sumi1 += (v1 * y[ib].qs[j + qk/2]);
}
int sumi = sumi0 + sumi1;
sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
}
*s = sumf;
#else
UNUSED(x);
UNUSED(y);
UNUSED(ib);
UNUSED(sumf);
ggml_vec_dot_q4_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@ -278,24 +268,14 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
sumf = vec_extract(vsumf0, 0);
#endif
for (; ib < nb; ++ib) {
int sumi0 = 0;
int sumi1 = 0;
for (int j = 0; j < qk/2; ++j) {
const int v0 = (x[ib].qs[j] & 0x0F);
const int v1 = (x[ib].qs[j] >> 4);
sumi0 += (v0 * y[ib].qs[j]);
sumi1 += (v1 * y[ib].qs[j + qk/2]);
}
int sumi = sumi0 + sumi1;
sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
}
*s = sumf;
#else
UNUSED(x);
UNUSED(y);
UNUSED(ib);
UNUSED(sumf);
ggml_vec_dot_q4_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@ -360,30 +340,14 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
sumf = vec_extract(vsumf0, 0);
#endif
for (; ib < nb; ++ib) {
uint32_t qh;
memcpy(&qh, x[ib].qh, sizeof(qh));
int sumi0 = 0;
int sumi1 = 0;
for (int j = 0; j < qk/2; ++j) {
const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16);
const int32_t x1 = (int8_t)(((x[ib].qs[j] >> 4) | xh_1) - 16);
sumi0 += (x0 * y[ib].qs[j]);
sumi1 += (x1 * y[ib].qs[j + qk/2]);
}
int sumi = sumi0 + sumi1;
sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
}
*s = sumf;
#else
UNUSED(ib);
UNUSED(sumf);
UNUSED(x);
UNUSED(y);
ggml_vec_dot_q5_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@ -451,30 +415,15 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
sumf = vec_extract(vsumf0, 0);
#endif
for (; ib < nb; ++ib) {
uint32_t qh;
memcpy(&qh, x[ib].qh, sizeof(qh));
int sumi0 = 0;
int sumi1 = 0;
for (int j = 0; j < qk/2; ++j) {
const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10;
const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10;
const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0;
const int32_t x1 = (x[ib].qs[j] >> 4) | xh_1;
sumi0 += (x0 * y[ib].qs[j]);
sumi1 += (x1 * y[ib].qs[j + qk/2]);
}
int sumi = sumi0 + sumi1;
sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
}
*s = sumf;
#else
UNUSED(nb);
UNUSED(ib);
UNUSED(sumf);
UNUSED(x);
UNUSED(y);
ggml_vec_dot_q5_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@ -535,18 +484,15 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
sumf = vec_extract(vsumf0, 0);
#endif
for (; ib < nb; ++ib) {
int sumi = 0;
for (int j = 0; j < qk; j++) {
sumi += x[ib].qs[j]*y[ib].qs[j];
}
sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
}
*s = sumf;
#else
UNUSED(nb);
UNUSED(x);
UNUSED(y);
UNUSED(ib);
UNUSED(sumf);
ggml_vec_dot_q8_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@ -695,45 +641,10 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
*s = vec_extract(vsumf0, 0);
#else
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * q2 = x[i].qs;
const int8_t * q8 = y[i].qs;
const uint8_t * sc = x[i].scales;
int summs = 0;
for (int j = 0; j < 16; ++j) {
summs += y[i].bsums[j] * (sc[j] >> 4);
}
const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
int isum = 0;
int is = 0;
int d;
for (int k = 0; k < QK_K/128; ++k) {
int shift = 0;
for (int j = 0; j < 4; ++j) {
d = sc[is++] & 0xF;
int isuml = 0;
for (int l = 0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
isum += d * isuml;
d = sc[is++] & 0xF;
isuml = 0;
for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
isum += d * isuml;
shift += 2;
q8 += 32;
}
q2 += 32;
}
sumf += dall * isum - dmin * summs;
}
*s = sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_q2_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -907,70 +818,13 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
*s = vec_extract(vsumf0, 0);
#else
// scalar version
// This function is written like this so the compiler can manage to vectorize most of it
// Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
// manually vectorized version above. Every other version I tried would run at least 4 times slower.
// The ideal situation would be if we could just write the code once, and the compiler would
// automatically produce the best possible set of machine instructions, instead of us having to manually
// write vectorized versions for AVX, ARM_NEON, etc.
int8_t aux8[QK_K];
int16_t aux16[8];
float sums [8];
int32_t aux32[8];
memset(sums, 0, 8*sizeof(float));
uint32_t auxs[4];
const int8_t * scales = (const int8_t*)auxs;
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * GGML_RESTRICT q3 = x[i].qs;
const uint8_t * GGML_RESTRICT hm = x[i].hmask;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
memset(aux32, 0, 8*sizeof(int32_t));
int8_t * GGML_RESTRICT a = aux8;
uint8_t m = 1;
for (int j = 0; j < QK_K; j += 128) {
for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3;
for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
a += 32; m <<= 1;
for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
a += 32; m <<= 1;
for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
a += 32; m <<= 1;
for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
a += 32; m <<= 1;
q3 += 32;
}
a = aux8;
memcpy(auxs, x[i].scales, 12);
uint32_t tmp = auxs[2];
auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
for (int j = 0; j < QK_K/16; ++j) {
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
q8 += 8; a += 8;
}
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
}
for (int l = 0; l < 8; ++l) sumf += sums[l];
*s = sumf;
UNUSED(kmask1);
UNUSED(kmask2);
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_q3_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@ -1130,61 +984,14 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
*s = vec_extract(vsumf0, 0);
#else
const uint8_t * scales = (const uint8_t*)&utmp[0];
const uint8_t * mins = (const uint8_t*)&utmp[2];
int8_t aux8[QK_K];
int16_t aux16[8];
float sums [8];
int32_t aux32[8];
memset(sums, 0, 8*sizeof(float));
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * GGML_RESTRICT q4 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
memset(aux32, 0, 8*sizeof(int32_t));
int8_t * GGML_RESTRICT a = aux8;
for (int j = 0; j < QK_K/64; ++j) {
for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
a += 32;
for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] >> 4);
a += 32; q4 += 32;
}
memcpy(utmp, x[i].scales, 12);
utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
const uint32_t uaux = utmp[1] & kmask1;
utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
utmp[2] = uaux;
utmp[0] &= kmask1;
int sumi = 0;
for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
a = aux8;
int is = 0;
for (int j = 0; j < QK_K/32; ++j) {
int32_t scale = scales[is++];
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
}
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
sumf -= dmin * sumi;
}
for (int l = 0; l < 8; ++l) sumf += sums[l];
*s = sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
UNUSED(kmask1);
UNUSED(kmask2);
UNUSED(kmask3);
UNUSED(utmp);
ggml_vec_dot_q4_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -1342,66 +1149,14 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
*s = vec_extract(vsumf0, 0);
#else
const uint8_t * scales = (const uint8_t*)&utmp[0];
const uint8_t * mins = (const uint8_t*)&utmp[2];
int8_t aux8[QK_K];
int16_t aux16[8];
float sums [8];
int32_t aux32[8];
memset(sums, 0, 8*sizeof(float));
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * GGML_RESTRICT q4 = x[i].qs;
const uint8_t * GGML_RESTRICT hm = x[i].qh;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
memset(aux32, 0, 8*sizeof(int32_t));
int8_t * GGML_RESTRICT a = aux8;
uint8_t m = 1;
for (int j = 0; j < QK_K/64; ++j) {
for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
a += 32; m <<= 1;
for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] >> 4);
for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
a += 32; m <<= 1;
q4 += 32;
}
memcpy(utmp, x[i].scales, 12);
utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
const uint32_t uaux = utmp[1] & kmask1;
utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
utmp[2] = uaux;
utmp[0] &= kmask1;
int sumi = 0;
for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
a = aux8;
int is = 0;
for (int j = 0; j < QK_K/32; ++j) {
int32_t scale = scales[is++];
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
}
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
sumf -= dmin * sumi;
}
for (int l = 0; l < 8; ++l) sumf += sums[l];
*s = sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
UNUSED(kmask1);
UNUSED(kmask2);
UNUSED(kmask3);
UNUSED(utmp);
ggml_vec_dot_q5_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -1556,47 +1311,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
*s = vec_extract(vsumf0, 0);
#else
int8_t aux8[QK_K];
int16_t aux16[8];
float sums [8];
int32_t aux32[8];
memset(sums, 0, 8*sizeof(float));
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * GGML_RESTRICT q4 = x[i].ql;
const uint8_t * GGML_RESTRICT qh = x[i].qh;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
memset(aux32, 0, 8*sizeof(int32_t));
int8_t * GGML_RESTRICT a = aux8;
for (int j = 0; j < QK_K; j += 128) {
for (int l = 0; l < 32; ++l) {
a[l + 0] = (int8_t)((q4[l + 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
a[l + 64] = (int8_t)((q4[l + 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32;
a[l + 96] = (int8_t)((q4[l + 32] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32;
}
a += 128;
q4 += 64;
qh += 32;
}
a = aux8;
int is = 0;
for (int j = 0; j < QK_K/16; ++j) {
int scale = x[i].scales[is++];
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
}
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
}
for (int l = 0; l < 8; ++l) sumf += sums[l];
*s = sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_q6_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -1737,34 +1455,10 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
*s = 0.125f * vec_extract(vsumf0, 0);
#else
uint32_t aux32[2];
const uint8_t * aux8 = (const uint8_t *)aux32;
float sumf = 0.f;
for (int i = 0; i < nb; ++i) {
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint16_t * GGML_RESTRICT q2 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
int32_t bsum = 0;
for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
memcpy(aux32, q2, 2*sizeof(uint32_t));
q2 += 4;
const uint32_t ls = 2*(aux32[1] >> 28) + 1;
int32_t sumi = 0;
for (int l = 0; l < 4; ++l) {
const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]);
const uint8_t signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127];
for (int j = 0; j < 8; ++j) {
sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
}
q8 += 8;
}
bsum += sumi * ls;
}
sumf += d * bsum;
}
*s = 0.125f * sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_iq2_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -1869,42 +1563,10 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
*s = 0.125f * vec_extract(vsumf0, 0);
#else
float sumf = 0.f;
for (int i = 0; i < nb; ++i) {
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint16_t * GGML_RESTRICT q2 = x[i].qs;
const uint8_t * GGML_RESTRICT sc = x[i].scales;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
int32_t bsum = 0;
for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
const uint16_t ls1 = 2*(sc[ib32] & 0xf) + 1;
const uint16_t ls2 = 2*(sc[ib32] >> 4) + 1;
int32_t sumi = 0;
for (int l = 0; l < 2; ++l) {
const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
const uint8_t signs = ksigns_iq2xs[q2[l] >> 9];
for (int j = 0; j < 8; ++j) {
sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
}
q8 += 8;
}
bsum += sumi * ls1;
sumi = 0;
for (int l = 2; l < 4; ++l) {
const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
const uint8_t signs = ksigns_iq2xs[q2[l] >> 9];
for (int j = 0; j < 8; ++j) {
sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
}
q8 += 8;
}
bsum += sumi * ls2;
q2 += 4;
}
sumf += d * bsum;
}
*s = 0.125f * sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_iq2_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -2030,47 +1692,11 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
*s = 0.125f * vec_extract(vsumf0, 0);
#else
float sumf = 0;
for (int i = 0; i < nb; i++) {
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const int8_t * q8 = y[i].qs;
const uint8_t * qs = x[i].qs;
const uint8_t * qh = x[i].qh;
const uint8_t * signs = qs + QK_K/8;
int bsum = 0;
for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
int ls1 = 1 + 2*(x[i].scales[ib32] & 0xf);
int ls2 = 1 + 2*(x[i].scales[ib32] >> 4);
int sumi1 = 0, sumi2 = 0;
for (int l = 0; l < 2; ++l) {
const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
for (int j = 0; j < 8; ++j) {
sumi1 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
}
q8 += 8;
}
for (int l = 2; l < 4; ++l) {
const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
for (int j = 0; j < 8; ++j) {
sumi2 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
}
q8 += 8;
}
bsum += ls1 * sumi1 + ls2 * sumi2;
qs += 4;
signs += 4;
}
sumf += d * bsum;
}
*s = 0.125f * sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_iq2_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@ -2172,36 +1798,10 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
*s = 0.25f * vec_extract(vsumf0, 0);
#else
uint32_t aux32;
float sumf = 0.f;
for (int i = 0; i < nb; ++i) {
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint8_t * GGML_RESTRICT q3 = x[i].qs;
const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
int32_t bsum = 0;
for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
memcpy(&aux32, gas, sizeof(uint32_t)); gas += sizeof(uint32_t);
const uint32_t ls = 2*(aux32 >> 28) + 1;
int32_t sumi = 0;
for (int l = 0; l < 4; ++l) {
const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*l+0]);
const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*l+1]);
const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*l) & 127];
for (int j = 0; j < 4; ++j) {
sumi += grid1[j] * q8[j+0] * (signs & kmask_iq2xs[j+0] ? -1 : 1);
sumi += grid2[j] * q8[j+4] * (signs & kmask_iq2xs[j+4] ? -1 : 1);
}
q8 += 8;
}
q3 += 8;
bsum += sumi * ls;
}
sumf += d * bsum;
}
*s = 0.25f * sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_iq3_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -2327,48 +1927,10 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
*s = vec_extract(vsumf0, 0);
#else
float sumf = 0.f;
for (int i = 0; i < nb; ++i) {
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint8_t * GGML_RESTRICT qs = x[i].qs;
const uint8_t * GGML_RESTRICT qh = x[i].qh;
const uint8_t * GGML_RESTRICT signs = x[i].signs;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
int32_t bsum = 0;
for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
const uint32_t ls1 = 2*(x[i].scales[ib32/2] & 0xf) + 1;
const uint32_t ls2 = 2*(x[i].scales[ib32/2] >> 4) + 1;
int32_t sumi = 0;
for (int l = 0; l < 4; ++l) {
const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+0] << (8-2*l)) & 256)));
const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+0] << (7-2*l)) & 256)));
for (int j = 0; j < 4; ++j) {
sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
}
q8 += 8;
}
qs += 8;
signs += 4;
bsum += sumi * ls1;
sumi = 0;
for (int l = 0; l < 4; ++l) {
const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+1] << (8-2*l)) & 256)));
const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+1] << (7-2*l)) & 256)));
for (int j = 0; j < 4; ++j) {
sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
}
q8 += 8;
}
qs += 8;
signs += 4;
bsum += sumi * ls2;
}
sumf += d * bsum;
}
*s = sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_iq3_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -2481,36 +2043,10 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
*s = vec_extract(vsumf0, 0);
#else
float sumf = 0;
for (int i = 0; i < nb; i++) {
const int8_t * q8 = y[i].qs;
const uint8_t * qs = x[i].qs;
const uint16_t * qh = x[i].qh;
int sumi = 0, sumi1 = 0;
for (int ib = 0; ib < QK_K/32; ++ib) {
const int ls = 2*((qh[ib] >> 12) & 7) + 1;
const int delta = qh[ib] & 0x8000 ? -1 : 1;
int lsum = 0;
for (int l = 0; l < 4; ++l) {
const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((qh[ib] >> 3*l) & 7) << 8)));
for (int j = 0; j < 8; ++j) {
lsum += q8[j] * grid[j];
}
q8 += 8;
}
sumi += ls * lsum;
sumi1 += ls * delta * (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]);
qs += 4;
}
sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
}
*s = sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_iq1_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -2581,17 +2117,15 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
sumf = vec_extract(vsumf0, 0);
#endif
for (; ib < nb; ++ib) {
const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
int sumi1 = 0, sumi2 = 0;
for (int j = 0; j < QK4_NL/2; ++j) {
sumi1 += y[ib].qs[j+ 0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
sumi2 += y[ib].qs[j+QK4_NL/2] * kvalues_iq4nl[x[ib].qs[j] >> 4];
}
sumf += d * (sumi1 + sumi2);
}
*s = sumf;
#else
UNUSED(x);
UNUSED(y);
UNUSED(nb);
UNUSED(ib);
UNUSED(sumf);
ggml_vec_dot_iq4_nl_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@ -2696,37 +2230,10 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
*s = vec_extract(vsumf0, 0);
#else
float sumf = 0;
for (int ibl = 0; ibl < nb; ++ibl) {
const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
uint16_t h = x[ibl].scales_h;
const uint8_t * qs = x[ibl].qs;
const int8_t * q8 = y[ibl].qs;
for (int ib = 0; ib < QK_K/32; ib += 2) {
const uint8_t ls1 = (x[ibl].scales_l[ib/2] & 0xf) | ((h << 4) & 0x30);
const uint8_t ls2 = (x[ibl].scales_l[ib/2] >> 4) | ((h << 2) & 0x30);
h >>= 4;
const float d1 = d4d8*(ls1 - 32);
const float d2 = d4d8*(ls2 - 32);
int sumi1 = 0, sumi2 = 0;
for (int j = 0; j < 16; ++j) {
sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >> 4];
}
sumf += d1 * (sumi1 + sumi2);
qs += 16;
q8 += 32;
sumi1 = sumi2 = 0;
for (int j = 0; j < 16; ++j) {
sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >> 4];
}
sumf += d2 * (sumi1 + sumi2);
qs += 16;
q8 += 32;
}
}
*s = sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_iq4_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}

398
ggml/src/ggml-cpu/arch/riscv/quants.c
View File

@ -116,6 +116,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
//===================================== Dot products =================================
void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
#if defined(__riscv_v)
const int qk = QK8_0;
const int nb = n / qk;
@ -132,7 +133,6 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
int ib = 0;
float sumf = 0;
#if defined(__riscv_v)
size_t vl = qk / 2;
for (; ib < nb; ++ib) {
@ -164,27 +164,14 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
}
#endif
for (; ib < nb; ++ib) {
int sumi0 = 0;
int sumi1 = 0;
for (int j = 0; j < qk/2; ++j) {
const int v0 = (x[ib].qs[j] & 0x0F) - 8;
const int v1 = (x[ib].qs[j] >> 4) - 8;
sumi0 += (v0 * y[ib].qs[j]);
sumi1 += (v1 * y[ib].qs[j + qk/2]);
}
int sumi = sumi0 + sumi1;
sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
}
*s = sumf;
#else
ggml_vec_dot_q4_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
#if defined(__riscv_v)
const int qk = QK8_1;
const int nb = n / qk;
@ -201,7 +188,6 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
int ib = 0;
float sumf = 0;
#if defined(__riscv_v)
size_t vl = qk / 2;
for (; ib < nb; ++ib) {
@ -229,27 +215,14 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
}
#endif
for (; ib < nb; ++ib) {
int sumi0 = 0;
int sumi1 = 0;
for (int j = 0; j < qk/2; ++j) {
const int v0 = (x[ib].qs[j] & 0x0F);
const int v1 = (x[ib].qs[j] >> 4);
sumi0 += (v0 * y[ib].qs[j]);
sumi1 += (v1 * y[ib].qs[j + qk/2]);
}
int sumi = sumi0 + sumi1;
sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
}
*s = sumf;
#else
ggml_vec_dot_q4_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
#if defined(__riscv_v)
const int qk = QK8_0;
const int nb = n / qk;
@ -267,7 +240,6 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
const block_q5_0 * GGML_RESTRICT x = vx;
const block_q8_0 * GGML_RESTRICT y = vy;
#if defined(__riscv_v)
size_t vl;
size_t vlenb = __riscv_vlenb();
@ -297,33 +269,14 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
}
#endif
for (; ib < nb; ++ib) {
uint32_t qh;
memcpy(&qh, x[ib].qh, sizeof(qh));
int sumi0 = 0;
int sumi1 = 0;
for (int j = 0; j < qk/2; ++j) {
const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16);
const int32_t x1 = (int8_t)(((x[ib].qs[j] >> 4) | xh_1) - 16);
sumi0 += (x0 * y[ib].qs[j]);
sumi1 += (x1 * y[ib].qs[j + qk/2]);
}
int sumi = sumi0 + sumi1;
sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
}
*s = sumf;
#else
ggml_vec_dot_q5_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
#if defined(__riscv_v)
const int qk = QK8_1;
const int nb = n / qk;
@ -341,7 +294,6 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
const block_q5_1 * GGML_RESTRICT x = vx;
const block_q8_1 * GGML_RESTRICT y = vy;
#if defined(__riscv_v)
size_t vl;
size_t vlenb = __riscv_vlenb();
@ -370,30 +322,10 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
}
#endif
for (; ib < nb; ++ib) {
uint32_t qh;
memcpy(&qh, x[ib].qh, sizeof(qh));
int sumi0 = 0;
int sumi1 = 0;
for (int j = 0; j < qk/2; ++j) {
const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10;
const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10;
const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0;
const int32_t x1 = (x[ib].qs[j] >> 4) | xh_1;
sumi0 += (x0 * y[ib].qs[j]);
sumi1 += (x1 * y[ib].qs[j + qk/2]);
}
int sumi = sumi0 + sumi1;
sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
}
*s = sumf;
#else
ggml_vec_dot_q5_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@ -431,18 +363,17 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
}
#endif
for (; ib < nb; ++ib) {
int sumi = 0;
for (int j = 0; j < qk; j++) {
sumi += x[ib].qs[j]*y[ib].qs[j];
}
sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
}
*s = sumf;
#else
UNUSED(nb);
UNUSED(x);
UNUSED(y);
UNUSED(ib);
UNUSED(sumf);
ggml_vec_dot_q8_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@ -738,44 +669,11 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
#else
float sumf = 0;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
for (int i = 0; i < nb; ++i) {
const uint8_t * q2 = x[i].qs;
const int8_t * q8 = y[i].qs;
const uint8_t * sc = x[i].scales;
int summs = 0;
for (int j = 0; j < 16; ++j) {
summs += y[i].bsums[j] * (sc[j] >> 4);
}
const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
int isum = 0;
int is = 0;
int d;
for (int k = 0; k < QK_K/128; ++k) {
int shift = 0;
for (int j = 0; j < 4; ++j) {
d = sc[is++] & 0xF;
int isuml = 0;
for (int l = 0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
isum += d * isuml;
d = sc[is++] & 0xF;
isuml = 0;
for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
isum += d * isuml;
shift += 2;
q8 += 32;
}
q2 += 32;
}
sumf += dall * isum - dmin * summs;
}
*s = sumf;
ggml_vec_dot_q2_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -1147,68 +1045,14 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
*s = sumf;
#else
// scalar version
// This function is written like this so the compiler can manage to vectorize most of it
// Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
// manually vectorized version above. Every other version I tried would run at least 4 times slower.
// The ideal situation would be if we could just write the code once, and the compiler would
// automatically produce the best possible set of machine instructions, instead of us having to manually
// write vectorized versions for AVX, ARM_NEON, etc.
int8_t aux8[QK_K];
int16_t aux16[8];
float sums [8];
int32_t aux32[8];
memset(sums, 0, 8*sizeof(float));
uint32_t auxs[4];
const int8_t * scales = (const int8_t*)auxs;
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * GGML_RESTRICT q3 = x[i].qs;
const uint8_t * GGML_RESTRICT hm = x[i].hmask;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
memset(aux32, 0, 8*sizeof(int32_t));
int8_t * GGML_RESTRICT a = aux8;
uint8_t m = 1;
for (int j = 0; j < QK_K; j += 128) {
for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3;
for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
a += 32; m <<= 1;
for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
a += 32; m <<= 1;
for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
a += 32; m <<= 1;
for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
a += 32; m <<= 1;
q3 += 32;
}
a = aux8;
memcpy(auxs, x[i].scales, 12);
uint32_t tmp = auxs[2];
auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
for (int j = 0; j < QK_K/16; ++j) {
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
q8 += 8; a += 8;
}
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
}
for (int l = 0; l < 8; ++l) sumf += sums[l];
*s = sumf;
UNUSED(kmask1);
UNUSED(kmask2);
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_q3_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -1534,60 +1378,15 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
#else
const uint8_t * scales = (const uint8_t*)&utmp[0];
const uint8_t * mins = (const uint8_t*)&utmp[2];
UNUSED(x);
UNUSED(y);
UNUSED(kmask1);
UNUSED(kmask2);
UNUSED(kmask3);
UNUSED(nb);
UNUSED(utmp);
int8_t aux8[QK_K];
int16_t aux16[8];
float sums [8];
int32_t aux32[8];
memset(sums, 0, 8*sizeof(float));
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * GGML_RESTRICT q4 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
memset(aux32, 0, 8*sizeof(int32_t));
int8_t * GGML_RESTRICT a = aux8;
for (int j = 0; j < QK_K/64; ++j) {
for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
a += 32;
for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] >> 4);
a += 32; q4 += 32;
}
memcpy(utmp, x[i].scales, 12);
utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
const uint32_t uaux = utmp[1] & kmask1;
utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
utmp[2] = uaux;
utmp[0] &= kmask1;
int sumi = 0;
for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
a = aux8;
int is = 0;
for (int j = 0; j < QK_K/32; ++j) {
int32_t scale = scales[is++];
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
}
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
sumf -= dmin * sumi;
}
for (int l = 0; l < 8; ++l) sumf += sums[l];
*s = sumf;
ggml_vec_dot_q4_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -1698,65 +1497,15 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
#else
const uint8_t * scales = (const uint8_t*)&utmp[0];
const uint8_t * mins = (const uint8_t*)&utmp[2];
UNUSED(x);
UNUSED(y);
UNUSED(kmask1);
UNUSED(kmask2);
UNUSED(kmask3);
UNUSED(nb);
UNUSED(utmp);
int8_t aux8[QK_K];
int16_t aux16[8];
float sums [8];
int32_t aux32[8];
memset(sums, 0, 8*sizeof(float));
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * GGML_RESTRICT q4 = x[i].qs;
const uint8_t * GGML_RESTRICT hm = x[i].qh;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
memset(aux32, 0, 8*sizeof(int32_t));
int8_t * GGML_RESTRICT a = aux8;
uint8_t m = 1;
for (int j = 0; j < QK_K/64; ++j) {
for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
a += 32; m <<= 1;
for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] >> 4);
for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
a += 32; m <<= 1;
q4 += 32;
}
memcpy(utmp, x[i].scales, 12);
utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
const uint32_t uaux = utmp[1] & kmask1;
utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
utmp[2] = uaux;
utmp[0] &= kmask1;
int sumi = 0;
for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
a = aux8;
int is = 0;
for (int j = 0; j < QK_K/32; ++j) {
int32_t scale = scales[is++];
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
}
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
sumf -= dmin * sumi;
}
for (int l = 0; l < 8; ++l) sumf += sums[l];
*s = sumf;
ggml_vec_dot_q5_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -2024,46 +1773,11 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
#else
int8_t aux8[QK_K];
int16_t aux16[8];
float sums [8];
int32_t aux32[8];
memset(sums, 0, 8*sizeof(float));
UNUSED(x);
UNUSED(y);
UNUSED(nb);
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * GGML_RESTRICT q4 = x[i].ql;
const uint8_t * GGML_RESTRICT qh = x[i].qh;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
memset(aux32, 0, 8*sizeof(int32_t));
int8_t * GGML_RESTRICT a = aux8;
for (int j = 0; j < QK_K; j += 128) {
for (int l = 0; l < 32; ++l) {
a[l + 0] = (int8_t)((q4[l + 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
a[l + 64] = (int8_t)((q4[l + 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32;
a[l + 96] = (int8_t)((q4[l + 32] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32;
}
a += 128;
q4 += 64;
qh += 32;
}
a = aux8;
int is = 0;
for (int j = 0; j < QK_K/16; ++j) {
int scale = x[i].scales[is++];
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
}
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
}
for (int l = 0; l < 8; ++l) sumf += sums[l];
*s = sumf;
ggml_vec_dot_q6_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}

61
ggml/src/ggml-cpu/arch/riscv/repack.cpp
View File

@ -112,31 +112,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
}
#endif
{
float sumf[8];
int sumi;
const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
for (int x = 0; x < nc / ncols_interleaved; x++) {
const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
for (int l = 0; l < nb; l++) {
for (int k = 0; k < (qk / (2 * blocklen)); k++) {
for (int j = 0; j < ncols_interleaved; j++) {
sumi = 0;
for (int i = 0; i < blocklen; ++i) {
const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
}
sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
}
}
}
for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
}
}
ggml_gemv_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
}
void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@ -361,37 +337,6 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
return;
}
#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
float sumf[4][8];
int sumi;
for (int y = 0; y < nr / 4; y++) {
const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
for (int x = 0; x < nc / ncols_interleaved; x++) {
const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
}
for (int l = 0; l < nb; l++) {
for (int k = 0; k < (qk / (2 * blocklen)); k++) {
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++) {
sumi = 0;
for (int i = 0; i < blocklen; ++i) {
const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
(v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
}
sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
}
}
}
}
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++)
s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
}
}
}
#endif
ggml_gemm_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
}

367
ggml/src/ggml-cpu/arch/s390/quants.c
View File

@ -172,24 +172,15 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
sumf = acc[0] + acc[1] + acc[2] + acc[3];
#endif
for (; ib < nb; ++ib) {
int sumi0 = 0;
int sumi1 = 0;
for (int j = 0; j < qk/2; ++j) {
const int v0 = (x[ib].qs[j] & 0x0F) - 8;
const int v1 = (x[ib].qs[j] >> 4) - 8;
sumi0 += (v0 * y[ib].qs[j]);
sumi1 += (v1 * y[ib].qs[j + qk/2]);
}
int sumi = sumi0 + sumi1;
sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
}
*s = sumf;
#else
UNUSED(nb);
UNUSED(x);
UNUSED(y);
UNUSED(ib);
UNUSED(sumf);
ggml_vec_dot_q4_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@ -239,24 +230,15 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
sumf = acc[0] + acc[1] + acc[2] + acc[3] + summs;
#endif
for (; ib < nb; ++ib) {
int sumi0 = 0;
int sumi1 = 0;
for (int j = 0; j < qk/2; ++j) {
const int v0 = (x[ib].qs[j] & 0x0F);
const int v1 = (x[ib].qs[j] >> 4);
sumi0 += (v0 * y[ib].qs[j]);
sumi1 += (v1 * y[ib].qs[j + qk/2]);
}
int sumi = sumi0 + sumi1;
sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
}
*s = sumf;
#else
UNUSED(nb);
UNUSED(x);
UNUSED(y);
UNUSED(ib);
UNUSED(sumf);
ggml_vec_dot_q4_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@ -298,18 +280,15 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
sumf = acc[0] + acc[1] + acc[2] + acc[3];
#endif
for (; ib < nb; ++ib) {
int sumi = 0;
for (int j = 0; j < qk; j++) {
sumi += x[ib].qs[j]*y[ib].qs[j];
}
sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
}
*s = sumf;
#else
UNUSED(nb);
UNUSED(x);
UNUSED(y);
UNUSED(ib);
UNUSED(sumf);
ggml_vec_dot_q8_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@ -442,70 +421,13 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
*s = sum;
#else
// scalar version
// This function is written like this so the compiler can manage to vectorize most of it
// Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
// manually vectorized version above. Every other version I tried would run at least 4 times slower.
// The ideal situation would be if we could just write the code once, and the compiler would
// automatically produce the best possible set of machine instructions, instead of us having to manually
// write vectorized versions for AVX, ARM_NEON, etc.
int8_t aux8[QK_K];
int16_t aux16[8];
float sums [8];
int32_t aux32[8];
memset(sums, 0, 8*sizeof(float));
uint32_t auxs[4];
const int8_t * scales = (const int8_t*)auxs;
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * GGML_RESTRICT q3 = x[i].qs;
const uint8_t * GGML_RESTRICT hm = x[i].hmask;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
memset(aux32, 0, 8*sizeof(int32_t));
int8_t * GGML_RESTRICT a = aux8;
uint8_t m = 1;
for (int j = 0; j < QK_K; j += 128) {
for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3;
for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
a += 32; m <<= 1;
for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
a += 32; m <<= 1;
for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
a += 32; m <<= 1;
for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
a += 32; m <<= 1;
q3 += 32;
}
a = aux8;
memcpy(auxs, x[i].scales, 12);
uint32_t tmp = auxs[2];
auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
for (int j = 0; j < QK_K/16; ++j) {
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
q8 += 8; a += 8;
}
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
}
for (int l = 0; l < 8; ++l) sumf += sums[l];
*s = sumf;
UNUSED(kmask1);
UNUSED(kmask2);
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_q3_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@ -600,61 +522,14 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
*s = sumf;
#else
const uint8_t * scales = (const uint8_t*)&utmp[0];
const uint8_t * mins = (const uint8_t*)&utmp[2];
int8_t aux8[QK_K];
int16_t aux16[8];
float sums [8];
int32_t aux32[8];
memset(sums, 0, 8*sizeof(float));
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * GGML_RESTRICT q4 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
memset(aux32, 0, 8*sizeof(int32_t));
int8_t * GGML_RESTRICT a = aux8;
for (int j = 0; j < QK_K/64; ++j) {
for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
a += 32;
for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] >> 4);
a += 32; q4 += 32;
}
memcpy(utmp, x[i].scales, 12);
utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
const uint32_t uaux = utmp[1] & kmask1;
utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
utmp[2] = uaux;
utmp[0] &= kmask1;
int sumi = 0;
for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
a = aux8;
int is = 0;
for (int j = 0; j < QK_K/32; ++j) {
int32_t scale = scales[is++];
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
}
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
sumf -= dmin * sumi;
}
for (int l = 0; l < 8; ++l) sumf += sums[l];
*s = sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
UNUSED(kmask1);
UNUSED(kmask2);
UNUSED(kmask3);
UNUSED(utmp);
ggml_vec_dot_q4_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -767,66 +642,14 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
*s = sumf;
#else
const uint8_t * scales = (const uint8_t*)&utmp[0];
const uint8_t * mins = (const uint8_t*)&utmp[2];
int8_t aux8[QK_K];
int16_t aux16[8];
float sums [8];
int32_t aux32[8];
memset(sums, 0, 8*sizeof(float));
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * GGML_RESTRICT q4 = x[i].qs;
const uint8_t * GGML_RESTRICT hm = x[i].qh;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
memset(aux32, 0, 8*sizeof(int32_t));
int8_t * GGML_RESTRICT a = aux8;
uint8_t m = 1;
for (int j = 0; j < QK_K/64; ++j) {
for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
a += 32; m <<= 1;
for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] >> 4);
for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
a += 32; m <<= 1;
q4 += 32;
}
memcpy(utmp, x[i].scales, 12);
utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
const uint32_t uaux = utmp[1] & kmask1;
utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
utmp[2] = uaux;
utmp[0] &= kmask1;
int sumi = 0;
for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
a = aux8;
int is = 0;
for (int j = 0; j < QK_K/32; ++j) {
int32_t scale = scales[is++];
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
}
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
sumf -= dmin * sumi;
}
for (int l = 0; l < 8; ++l) sumf += sums[l];
*s = sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
UNUSED(kmask1);
UNUSED(kmask2);
UNUSED(kmask3);
UNUSED(utmp);
ggml_vec_dot_q5_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -969,47 +792,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
*s = sum;
#else
int8_t aux8[QK_K];
int16_t aux16[8];
float sums [8];
int32_t aux32[8];
memset(sums, 0, 8*sizeof(float));
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * GGML_RESTRICT q4 = x[i].ql;
const uint8_t * GGML_RESTRICT qh = x[i].qh;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
memset(aux32, 0, 8*sizeof(int32_t));
int8_t * GGML_RESTRICT a = aux8;
for (int j = 0; j < QK_K; j += 128) {
for (int l = 0; l < 32; ++l) {
a[l + 0] = (int8_t)((q4[l + 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
a[l + 64] = (int8_t)((q4[l + 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32;
a[l + 96] = (int8_t)((q4[l + 32] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32;
}
a += 128;
q4 += 64;
qh += 32;
}
a = aux8;
int is = 0;
for (int j = 0; j < QK_K/16; ++j) {
int scale = x[i].scales[is++];
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
}
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
}
for (int l = 0; l < 8; ++l) sumf += sums[l];
*s = sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_q6_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -1186,17 +972,15 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
sumf += GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d) * (v_xy[0] + v_xy[1] + v_xy[2] + v_xy[3]);
}
#endif
for (; ib < nb; ++ib) {
const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
int sumi1 = 0, sumi2 = 0;
for (int j = 0; j < QK4_NL/2; ++j) {
sumi1 += y[ib].qs[j+ 0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
sumi2 += y[ib].qs[j+QK4_NL/2] * kvalues_iq4nl[x[ib].qs[j] >> 4];
}
sumf += d * (sumi1 + sumi2);
}
*s = sumf;
#else
UNUSED(x);
UNUSED(y);
UNUSED(nb);
UNUSED(ib);
UNUSED(sumf);
ggml_vec_dot_iq4_nl_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@ -1264,37 +1048,10 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
*s = sumf;
#else
float sumf = 0;
for (int ibl = 0; ibl < nb; ++ibl) {
const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
uint16_t h = x[ibl].scales_h;
const uint8_t * qs = x[ibl].qs;
const int8_t * q8 = y[ibl].qs;
for (int ib = 0; ib < QK_K/32; ib += 2) {
const uint8_t ls1 = (x[ibl].scales_l[ib/2] & 0xf) | ((h << 4) & 0x30);
const uint8_t ls2 = (x[ibl].scales_l[ib/2] >> 4) | ((h << 2) & 0x30);
h >>= 4;
const float d1 = d4d8*(ls1 - 32);
const float d2 = d4d8*(ls2 - 32);
int sumi1 = 0, sumi2 = 0;
for (int j = 0; j < 16; ++j) {
sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >> 4];
}
sumf += d1 * (sumi1 + sumi2);
qs += 16;
q8 += 32;
sumi1 = sumi2 = 0;
for (int j = 0; j < 16; ++j) {
sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >> 4];
}
sumf += d2 * (sumi1 + sumi2);
qs += 16;
q8 += 32;
}
}
*s = sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_iq4_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}

368
ggml/src/ggml-cpu/arch/wasm/quants.c
View File

@ -435,30 +435,15 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3);
#endif
for (; ib < nb; ++ib) {
uint32_t qh;
memcpy(&qh, x[ib].qh, sizeof(qh));
int sumi0 = 0;
int sumi1 = 0;
for (int j = 0; j < qk/2; ++j) {
const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16);
const int32_t x1 = (int8_t)(((x[ib].qs[j] >> 4) | xh_1) - 16);
sumi0 += (x0 * y[ib].qs[j]);
sumi1 += (x1 * y[ib].qs[j + qk/2]);
}
int sumi = sumi0 + sumi1;
sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
}
*s = sumf;
#else
UNUSED(nb);
UNUSED(ib);
UNUSED(sumf);
UNUSED(x);
UNUSED(y);
ggml_vec_dot_q5_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@ -545,30 +530,15 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3) + summs;
#endif
for (; ib < nb; ++ib) {
uint32_t qh;
memcpy(&qh, x[ib].qh, sizeof(qh));
int sumi0 = 0;
int sumi1 = 0;
for (int j = 0; j < qk/2; ++j) {
const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10;
const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10;
const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0;
const int32_t x1 = (x[ib].qs[j] >> 4) | xh_1;
sumi0 += (x0 * y[ib].qs[j]);
sumi1 += (x1 * y[ib].qs[j + qk/2]);
}
int sumi = sumi0 + sumi1;
sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
}
*s = sumf;
#else
UNUSED(nb);
UNUSED(ib);
UNUSED(sumf);
UNUSED(x);
UNUSED(y);
ggml_vec_dot_q5_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@ -628,18 +598,15 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3);
#endif
for (; ib < nb; ++ib) {
int sumi = 0;
for (int j = 0; j < qk; j++) {
sumi += x[ib].qs[j]*y[ib].qs[j];
}
sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
}
*s = sumf;
#else
UNUSED(nb);
UNUSED(x);
UNUSED(y);
UNUSED(ib);
UNUSED(sumf);
ggml_vec_dot_q8_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@ -755,45 +722,10 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
*s = sumf;
#else
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * q2 = x[i].qs;
const int8_t * q8 = y[i].qs;
const uint8_t * sc = x[i].scales;
int summs = 0;
for (int j = 0; j < 16; ++j) {
summs += y[i].bsums[j] * (sc[j] >> 4);
}
const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
int isum = 0;
int is = 0;
int d;
for (int k = 0; k < QK_K/128; ++k) {
int shift = 0;
for (int j = 0; j < 4; ++j) {
d = sc[is++] & 0xF;
int isuml = 0;
for (int l = 0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
isum += d * isuml;
d = sc[is++] & 0xF;
isuml = 0;
for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
isum += d * isuml;
shift += 2;
q8 += 32;
}
q2 += 32;
}
sumf += dall * isum - dmin * summs;
}
*s = sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_q2_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -902,68 +834,12 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
*s = sumf;
#else
// scalar version
// This function is written like this so the compiler can manage to vectorize most of it
// Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
// manually vectorized version above. Every other version I tried would run at least 4 times slower.
// The ideal situation would be if we could just write the code once, and the compiler would
// automatically produce the best possible set of machine instructions, instead of us having to manually
// write vectorized versions for AVX, ARM_NEON, etc.
int8_t aux8[QK_K];
int16_t aux16[8];
float sums [8];
int32_t aux32[8];
memset(sums, 0, 8*sizeof(float));
uint32_t auxs[4];
const int8_t * scales = (const int8_t*)auxs;
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * GGML_RESTRICT q3 = x[i].qs;
const uint8_t * GGML_RESTRICT hm = x[i].hmask;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
memset(aux32, 0, 8*sizeof(int32_t));
int8_t * GGML_RESTRICT a = aux8;
uint8_t m = 1;
for (int j = 0; j < QK_K; j += 128) {
for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3;
for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
a += 32; m <<= 1;
for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
a += 32; m <<= 1;
for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
a += 32; m <<= 1;
for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
a += 32; m <<= 1;
q3 += 32;
}
a = aux8;
memcpy(auxs, x[i].scales, 12);
uint32_t tmp = auxs[2];
auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
for (int j = 0; j < QK_K/16; ++j) {
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
q8 += 8; a += 8;
}
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
}
for (int l = 0; l < 8; ++l) sumf += sums[l];
*s = sumf;
UNUSED(kmask1);
UNUSED(kmask2);
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_q3_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -1089,61 +965,14 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
*s = sumf;
#else
const uint8_t * scales = (const uint8_t*)&utmp[0];
const uint8_t * mins = (const uint8_t*)&utmp[2];
int8_t aux8[QK_K];
int16_t aux16[8];
float sums [8];
int32_t aux32[8];
memset(sums, 0, 8*sizeof(float));
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * GGML_RESTRICT q4 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
memset(aux32, 0, 8*sizeof(int32_t));
int8_t * GGML_RESTRICT a = aux8;
for (int j = 0; j < QK_K/64; ++j) {
for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
a += 32;
for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] >> 4);
a += 32; q4 += 32;
}
memcpy(utmp, x[i].scales, 12);
utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
const uint32_t uaux = utmp[1] & kmask1;
utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
utmp[2] = uaux;
utmp[0] &= kmask1;
int sumi = 0;
for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
a = aux8;
int is = 0;
for (int j = 0; j < QK_K/32; ++j) {
int32_t scale = scales[is++];
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
}
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
sumf -= dmin * sumi;
}
for (int l = 0; l < 8; ++l) sumf += sums[l];
*s = sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
UNUSED(kmask1);
UNUSED(kmask2);
UNUSED(kmask3);
UNUSED(utmp);
ggml_vec_dot_q4_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -1279,66 +1108,14 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
*s = sumf;
#else
const uint8_t * scales = (const uint8_t*)&utmp[0];
const uint8_t * mins = (const uint8_t*)&utmp[2];
int8_t aux8[QK_K];
int16_t aux16[8];
float sums [8];
int32_t aux32[8];
memset(sums, 0, 8*sizeof(float));
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * GGML_RESTRICT q4 = x[i].qs;
const uint8_t * GGML_RESTRICT hm = x[i].qh;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
memset(aux32, 0, 8*sizeof(int32_t));
int8_t * GGML_RESTRICT a = aux8;
uint8_t m = 1;
for (int j = 0; j < QK_K/64; ++j) {
for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
a += 32; m <<= 1;
for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] >> 4);
for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
a += 32; m <<= 1;
q4 += 32;
}
memcpy(utmp, x[i].scales, 12);
utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
const uint32_t uaux = utmp[1] & kmask1;
utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
utmp[2] = uaux;
utmp[0] &= kmask1;
int sumi = 0;
for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
a = aux8;
int is = 0;
for (int j = 0; j < QK_K/32; ++j) {
int32_t scale = scales[is++];
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
}
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
sumf -= dmin * sumi;
}
for (int l = 0; l < 8; ++l) sumf += sums[l];
*s = sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
UNUSED(kmask1);
UNUSED(kmask2);
UNUSED(kmask3);
UNUSED(utmp);
ggml_vec_dot_q5_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
@ -1435,47 +1212,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
*s = sumf;
#else
int8_t aux8[QK_K];
int16_t aux16[8];
float sums [8];
int32_t aux32[8];
memset(sums, 0, 8*sizeof(float));
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * GGML_RESTRICT q4 = x[i].ql;
const uint8_t * GGML_RESTRICT qh = x[i].qh;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
memset(aux32, 0, 8*sizeof(int32_t));
int8_t * GGML_RESTRICT a = aux8;
for (int j = 0; j < QK_K; j += 128) {
for (int l = 0; l < 32; ++l) {
a[l + 0] = (int8_t)((q4[l + 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
a[l + 64] = (int8_t)((q4[l + 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32;
a[l + 96] = (int8_t)((q4[l + 32] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32;
}
a += 128;
q4 += 64;
qh += 32;
}
a = aux8;
int is = 0;
for (int j = 0; j < QK_K/16; ++j) {
int scale = x[i].scales[is++];
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
}
const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
}
for (int l = 0; l < 8; ++l) sumf += sums[l];
*s = sumf;
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_q6_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}

859
ggml/src/ggml-cpu/arch/x86/quants.c
View File

File diff suppressed because it is too large Load Diff

4896
ggml/src/ggml-cpu/arch/x86/repack.cpp
View File

File diff suppressed because it is too large Load Diff

21
ggml/src/ggml-cpu/ggml-cpu.c
View File

@ -253,6 +253,12 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
.vec_dot_type = GGML_TYPE_Q8_1,
.nrows = 1,
},
[GGML_TYPE_MXFP4] = {
.from_float = quantize_row_mxfp4,
.vec_dot = ggml_vec_dot_mxfp4_q8_0,
.vec_dot_type = GGML_TYPE_Q8_0,
.nrows = 1,
},
[GGML_TYPE_Q2_K] = {
.from_float = quantize_row_q2_K,
.vec_dot = ggml_vec_dot_q2_K_q8_K,
@ -1670,6 +1676,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
{
ggml_compute_forward_add(params, tensor);
} break;
case GGML_OP_ADD_ID:
{
ggml_compute_forward_add_id(params, tensor);
} break;
case GGML_OP_ADD1:
{
ggml_compute_forward_add1(params, tensor);
@ -1924,7 +1934,7 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
} break;
case GGML_OP_FLASH_ATTN_EXT:
{
ggml_compute_forward_flash_attn_ext(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor->src[3], tensor);
ggml_compute_forward_flash_attn_ext(params, tensor);
} break;
case GGML_OP_FLASH_ATTN_BACK:
{
@ -2012,6 +2022,11 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
ggml_compute_forward_opt_step_adamw(params, tensor);
}
break;
case GGML_OP_OPT_STEP_SGD:
{
ggml_compute_forward_opt_step_sgd(params, tensor);
}
break;
case GGML_OP_NONE:
{
// nop
@ -2111,6 +2126,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
case GGML_OP_DUP:
case GGML_OP_CONT:
case GGML_OP_ADD:
case GGML_OP_ADD_ID:
case GGML_OP_ADD1:
case GGML_OP_ACC:
{
@ -2172,6 +2188,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
case GGML_GLU_OP_REGLU:
case GGML_GLU_OP_GEGLU:
case GGML_GLU_OP_SWIGLU:
case GGML_GLU_OP_SWIGLU_OAI:
case GGML_GLU_OP_GEGLU_ERF:
case GGML_GLU_OP_GEGLU_QUICK:
{
@ -2313,6 +2330,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
case GGML_OP_CROSS_ENTROPY_LOSS:
case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
case GGML_OP_OPT_STEP_ADAMW:
case GGML_OP_OPT_STEP_SGD:
{
n_tasks = n_threads;
} break;
@ -2673,6 +2691,7 @@ struct ggml_cplan ggml_graph_plan(
}
} break;
case GGML_OP_ADD:
case GGML_OP_ADD_ID:
case GGML_OP_ADD1:
{
if (ggml_is_quantized(node->src[0]->type)) {

39
ggml/src/ggml-cpu/ggml-cpu.cpp
View File

@ -35,7 +35,7 @@
// ggml-backend interface
std::vector<ggml_backend_buffer_type_t>& ggml_backend_cpu_get_extra_buffers_type() {
std::vector<ggml_backend_buffer_type_t> & ggml_backend_cpu_get_extra_buffer_types() {
static std::vector<ggml_backend_buffer_type_t> bufts = []() {
std::vector<ggml_backend_buffer_type_t> bufts;
@ -57,8 +57,6 @@ std::vector<ggml_backend_buffer_type_t>& ggml_backend_cpu_get_extra_buffers_type
}
#endif
bufts.push_back(NULL);
return bufts;
}();
@ -66,14 +64,20 @@ std::vector<ggml_backend_buffer_type_t>& ggml_backend_cpu_get_extra_buffers_type
}
static ggml_backend_buffer_type_t * ggml_backend_cpu_device_get_extra_buffers_type(ggml_backend_dev_t device) {
return ggml_backend_cpu_get_extra_buffers_type().data();
static std::vector<ggml_backend_buffer_type_t> extra_bufts = [] {
std::vector<ggml_backend_buffer_type_t> bufts = ggml_backend_cpu_get_extra_buffer_types();
bufts.push_back(nullptr);
return bufts;
}();
return extra_bufts.data();
GGML_UNUSED(device);
}
static bool ggml_backend_cpu_is_extra_buffer_type(ggml_backend_buffer_type_t buft) {
for (auto * extra : ggml_backend_cpu_get_extra_buffers_type()) {
if (extra && extra == buft) {
for (auto * extra : ggml_backend_cpu_get_extra_buffer_types()) {
if (extra == buft) {
return true;
}
}
@ -211,7 +215,7 @@ ggml_backend_t ggml_backend_cpu_init(void) {
ggml_backend_t cpu_backend = new ggml_backend {
/* .guid = */ ggml_backend_cpu_guid(),
/* .interface = */ ggml_backend_cpu_i,
/* .iface = */ ggml_backend_cpu_i,
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
/* .context = */ ctx,
};
@ -397,20 +401,13 @@ static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const st
return true;
}
// extra_buffer_op?
for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) {
if (extra) {
auto buf_extra = (ggml::cpu::extra_buffer_type*) extra->context;
if (buf_extra && buf_extra->supports_op(dev, op)) {
return true;
}
}
}
// the other case need host buffer.
for (int i = 0; i < GGML_MAX_SRC; i++) {
if (op->src[i] && op->src[i]->buffer && !ggml_backend_buft_is_host(op->src[i]->buffer->buft)) {
return false;
// check extra buffer types
// note: only the first sources are checked for extra buffer types to reduce overhead, increase if necessary
for (int i = 0; i < 4; i++) {
if (op->src[i] && op->src[i]->buffer &&
ggml_backend_cpu_is_extra_buffer_type(op->src[i]->buffer->buft)) {
auto * buf_extra = (ggml::cpu::extra_buffer_type *) op->src[i]->buffer->buft->context;
return buf_extra->supports_op(dev, op);
}
}

97
ggml/src/ggml-cpu/kleidiai/kernels.cpp
View File

@ -22,9 +22,94 @@
#include "kai_common.h"
#include "simd-mappings.h"
#include "kernels.h"
#define NELEMS(x) sizeof(x) / sizeof(*x)
static const size_t INT4_PER_BYTE = 2;
static const size_t INT4_BITS = 4;
static const int Q4_0_ZERO_POINT = 8;
const size_t INT4_PER_UINT16 = 4;
static void dequantize_row_qsi4c32pscalef16(
const void *packed_data,
int32_t row_idx,
int64_t nc,
float *out,
size_t nr_pack,
size_t packed_row_stride,
size_t kr,
size_t bl,
size_t num_bytes_multiplier
) {
size_t group_idx = row_idx / nr_pack;
size_t row_in_group = row_idx % nr_pack;
const uint8_t *packed_group = (const uint8_t *)packed_data + group_idx * packed_row_stride;
size_t num_blocks = nc / bl;
const uint8_t *block_ptr = packed_group;
for (size_t b = 0; b < num_blocks; ++b) {
uint16_t scale_f16 = *((const uint16_t *)(block_ptr + row_in_group * num_bytes_multiplier));
float scale = GGML_CPU_FP16_TO_FP32(scale_f16);
const uint8_t *segment_ptr = block_ptr + nr_pack * num_bytes_multiplier;
size_t num_segments = bl / kr;
size_t num_bytes_per_segment = kr / INT4_PER_BYTE;
for (size_t s = 0; s < num_segments; ++s) {
const uint8_t *seg_base = segment_ptr + s * nr_pack * num_bytes_per_segment;
const uint8_t *qbytes = seg_base + row_in_group * num_bytes_per_segment;
for (size_t k = 0; k < num_bytes_per_segment; ++k) {
uint8_t byte = qbytes[k] ^ 0x88;
int x0 = (byte & 0x0F) - Q4_0_ZERO_POINT;
int x1 = (byte >> INT4_BITS) - Q4_0_ZERO_POINT;
out[b * bl + s * num_bytes_per_segment + k] = x0 * scale;
out[b * bl + s * num_bytes_per_segment + k + bl/2] = x1 * scale;
}
}
block_ptr += nr_pack * num_bytes_multiplier + num_segments * nr_pack * num_bytes_per_segment;
}
}
static void dequantize_row_qsi4c32ps1s0scalef16(
const void *packed_data,
int32_t row_idx,
int64_t k,
float *out,
size_t nr,
size_t packed_row_stride,
size_t kr,
size_t bl,
size_t num_bytes_multiplier
) {
const size_t num_blocks = k / bl;
const size_t bl4 = bl / INT4_PER_UINT16;
size_t group_idx = row_idx / nr;
size_t row_in_group = row_idx % nr;
const uint8_t *packed_group = (const uint8_t *)packed_data + group_idx * packed_row_stride;
const uint16_t *qdata = (const uint16_t *)packed_group;
const uint16_t *scales = (const uint16_t *)(packed_group + packed_row_stride - (nr * num_blocks * num_bytes_multiplier));
for (size_t block_idx = 0; block_idx < num_blocks; ++block_idx) {
uint16_t scale_f16 = scales[row_in_group + block_idx * nr];
float scale = GGML_CPU_FP16_TO_FP32(scale_f16);
for (size_t bl4_idx = 0; bl4_idx < bl4; ++bl4_idx) {
uint16_t q = qdata[(block_idx * bl4 + bl4_idx) * nr + row_in_group];
for (size_t qidx = 0; qidx < INT4_PER_UINT16; ++qidx) {
int v = ((q >> (qidx * 4)) & 0xF) - Q4_0_ZERO_POINT;
out[block_idx * bl + bl4_idx * INT4_BITS + qidx] = v * scale;
}
}
}
GGML_UNUSED(kr);
}
static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
#if defined(__ARM_FEATURE_SME)
{
@ -64,7 +149,9 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
},
/* .rhs_info = */ {
/* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
/* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
/* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
/* .to_float = */ dequantize_row_qsi4c32ps1s0scalef16,
},
/* .required_cpu = */ CPU_FEATURE_SME,
/* .lhs_type = */ GGML_TYPE_F32,
@ -108,7 +195,9 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
},
/* .rhs_info = */ {
/* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme,
/* .packed_stride = */ NULL,
/* .pack_func = */ kai_run_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme,
/* .to_float = */ NULL,
},
/* .required_cpu = */ CPU_FEATURE_SME,
/* .lhs_type = */ GGML_TYPE_F32,
@ -155,7 +244,9 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
},
/* .rhs_info = */ {
/* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
/* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
/* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
/* .to_float = */ dequantize_row_qsi4c32pscalef16,
},
/* .required_cpu = */ CPU_FEATURE_DOTPROD,
/* .lhs_type = */ GGML_TYPE_F32,
@ -201,7 +292,9 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
},
/* .rhs_info = */ {
/* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
/* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
/* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
/* .to_float = */ dequantize_row_qsi4c32pscalef16,
},
/* .required_cpu = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM,
/* .lhs_type = */ GGML_TYPE_F32,
@ -248,7 +341,9 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
},
/* .rhs_info = */ {
/* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
/* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
/* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
/* .to_float = */ dequantize_row_qsi4c32pscalef16,
},
/* .required_cpu = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM,
/* .lhs_type = */ GGML_TYPE_F32,
@ -294,7 +389,9 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
},
/* .rhs_info = */ {
/* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
/* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
/* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
/* .to_float = */ dequantize_row_qsi4c32pscalef16,
},
/* .required_cpu = */ CPU_FEATURE_DOTPROD,
/* .lhs_type = */ GGML_TYPE_F32,

3
ggml/src/ggml-cpu/kleidiai/kernels.h
View File

@ -71,12 +71,15 @@ struct rhs_packing_info {
std::function<size_t(size_t n, size_t k, size_t nr, size_t kr, size_t bl)>,
std::function<size_t(size_t n, size_t k)>
> packed_size;
size_t (*packed_stride)(size_t k, size_t nr, size_t kr, size_t bl);
std::variant<
std::function<void(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t bl, const uint8_t* rhs,
const float* bias, void* rhs_packed, size_t extra_bytes, const struct kai_rhs_pack_qs4cxs1s0_param* params)>,
std::function<void(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t rhs_stride, const void* rhs,
const void* bias, const void* scale, void* rhs_packed, size_t extra_bytes, const void* params)>
> pack_func;
void (*to_float)(const void *packed_data, int32_t row_idx, int64_t nc, float *out, size_t nr_pack, size_t packed_row_stride,
size_t kr, size_t bl, size_t num_bytes_multiplier);
};
struct ggml_kleidiai_kernels {

113
ggml/src/ggml-cpu/kleidiai/kleidiai.cpp
View File

@ -40,6 +40,17 @@ struct ggml_kleidiai_context {
ggml_kleidiai_kernels * kernels;
} static ctx = { CPU_FEATURE_NONE, NULL };
static const char* cpu_feature_to_string(cpu_feature f) {
switch (f) {
case CPU_FEATURE_NONE: return "NONE";
case CPU_FEATURE_DOTPROD: return "DOTPROD";
case CPU_FEATURE_I8MM: return "I8MM";
case CPU_FEATURE_SVE: return "SVE";
case CPU_FEATURE_SME: return "SME";
default: return "UNKNOWN";
}
}
static void init_kleidiai_context(void) {
ggml_critical_section_start();
@ -62,6 +73,11 @@ static void init_kleidiai_context(void) {
ctx.features |= ggml_cpu_has_sme() ? CPU_FEATURE_SME : CPU_FEATURE_NONE;
}
ctx.kernels = ggml_kleidiai_select_kernels_q4_0(ctx.features);
#ifndef NDEBUG
if (ctx.kernels) {
GGML_LOG_DEBUG("kleidiai: using kernel with CPU feature %s\n", cpu_feature_to_string(ctx.kernels->required_cpu));
}
#endif
}
ggml_critical_section_end();
}
@ -102,6 +118,9 @@ static void transpose_f32kxn_f16nxk(size_t n, size_t k, float * dst, const uint1
class tensor_traits : public ggml::cpu::tensor_traits {
bool work_size(int /* n_threads */, const struct ggml_tensor * op, size_t & size) override {
if (op->op != GGML_OP_MUL_MAT) {
return false;
}
ggml_kleidiai_kernels *kernels = ggml_kleidiai_select_kernels(ctx.features, op);
GGML_ASSERT(kernels);
kernel_info * kernel = op->src[1]->ne[1] == 1 ? &kernels->gemv : &kernels->gemm;
@ -135,6 +154,10 @@ class tensor_traits : public ggml::cpu::tensor_traits {
} else if (dst->src[0]->type == GGML_TYPE_F16) {
return compute_forward_kv_cache(params, dst);
}
} else if (dst->op == GGML_OP_GET_ROWS) {
if (dst->src[0]->type == GGML_TYPE_Q4_0) {
return compute_forward_get_rows(params, dst);
}
}
return false;
}
@ -236,7 +259,10 @@ class tensor_traits : public ggml::cpu::tensor_traits {
const int64_t m_start = 0;
const int64_t n_step = static_cast<int64_t>(kernel->get_n_step());
const int64_t num_threads = KAI_MIN(n / n_step, nth);
int64_t num_threads = KAI_MIN(n / n_step, nth);
if (num_threads <= 0) {
num_threads = 1;
}
if (ith < num_threads) {
const int64_t num_n_per_thread0 = round_down(n / num_threads, n_step);
@ -270,6 +296,8 @@ class tensor_traits : public ggml::cpu::tensor_traits {
}
bool compute_forward_q4_0(struct ggml_compute_params * params, struct ggml_tensor * dst) {
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_Q4_0);
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
@ -284,7 +312,8 @@ class tensor_traits : public ggml::cpu::tensor_traits {
GGML_ASSERT(kernel);
const int ith = params->ith;
const int nth = params->nth;
const int nth_raw = params->nth;
const int nth = nth_raw > 0 ? nth_raw : 1;
const size_t k = ne00;
const size_t m = ne11;
@ -302,10 +331,13 @@ class tensor_traits : public ggml::cpu::tensor_traits {
const size_t num_n_per_thread = kai_roundup(kai_roundup(n, nth) / nth, n_step);
const size_t n_start = ith * num_n_per_thread;
size_t n_to_process = num_n_per_thread;
size_t n_to_process = 0;
if (n_start < n) {
n_to_process = num_n_per_thread;
if ((n_start + n_to_process) > n) {
n_to_process = n - n_start;
}
}
// Calculate number of columns to be processed per thread
const size_t num_m_per_thread = kai_roundup(m, mr * nth) / nth;
@ -336,14 +368,57 @@ class tensor_traits : public ggml::cpu::tensor_traits {
const void* lhs_ptr = (const void*)((const char *)lhs_packed + lhs_packed_offset);
float *dst_ptr = reinterpret_cast<float *>(static_cast<uint8_t *>(dst->data) + dst_offset);
if (n_to_process > 0) {
variant_call<void>(kernel->run_kernel, m, n_to_process, k, QK4_0, lhs_ptr, rhs_ptr, dst_ptr, dst_stride,
sizeof(float), -FLT_MAX, FLT_MAX);
}
return true;
}
bool compute_forward_get_rows(struct ggml_compute_params * params, struct ggml_tensor * dst) {
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_Q4_0);
GGML_ASSERT(ctx.kernels);
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
GGML_TENSOR_BINARY_OP_LOCALS
rhs_packing_info * rhs_info = &ctx.kernels->rhs_info;
kernel_info * kernel = &ctx.kernels->gemm;
const int64_t nc = ne00;
const int64_t nr = ggml_nelements(src1);
const size_t block_rows = kernel->get_nr();
const size_t kr = kernel->get_kr();
const size_t num_bytes_multiplier = sizeof(uint16_t);
const size_t packed_stride = rhs_info->packed_stride(nc, block_rows, kr, QK4_0);
const int ith = params->ith;
const int nth = params->nth;
const int dr = (nr + nth - 1) / nth;
const int ir0 = dr * ith;
const int ir1 = MIN(ir0 + dr, nr);
for (int64_t i = ir0; i < ir1; ++i) {
GGML_ASSERT(src1->type == GGML_TYPE_I32);
int64_t row_idx = ((const int32_t *)src1->data)[i];
GGML_ASSERT(row_idx >= 0 && row_idx < src0->ne[1]);
float *out = (float *)((char *)dst->data + i * nb1);
rhs_info->to_float(src0->data, row_idx, nc, out, block_rows, packed_stride, kr, QK4_0, num_bytes_multiplier);
}
return true;
}
public:
int repack(struct ggml_tensor * tensor, const void * data, size_t data_size) {
GGML_ASSERT(tensor->type == GGML_TYPE_Q4_0);
GGML_ASSERT(ctx.kernels);
const size_t n = tensor->ne[1];
const size_t k = tensor->ne[0];
@ -351,17 +426,12 @@ public:
size_t kr = ctx.kernels->gemm.get_kr();
size_t sr = ctx.kernels->gemm.get_sr();
#ifndef NDEBUG
const size_t repacked_size = variant_call<size_t>(ctx.kernels->rhs_info.packed_size, n, k, nr, kr, QK4_0);
GGML_ASSERT(repacked_size <= data_size && "repacked size larger than the packed size!");
#endif
struct kai_rhs_pack_qs4cxs1s0_param params;
params.lhs_zero_point = 1;
params.rhs_zero_point = 8;
variant_call<void>(ctx.kernels->rhs_info.pack_func, 1, n, k, nr, kr, sr, QK4_0, (const uint8_t*)data, nullptr, tensor->data, 0, &params);
return 0;
GGML_UNUSED(data_size);
}
};
@ -375,8 +445,8 @@ static ggml::cpu::tensor_traits * get_tensor_traits(ggml_backend_buffer_t, struc
static enum ggml_status ggml_backend_cpu_kleidiai_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
tensor->extra = (void *) ggml::cpu::kleidiai::get_tensor_traits(buffer, tensor);
GGML_UNUSED(buffer);
return GGML_STATUS_SUCCESS;
GGML_UNUSED(buffer);
}
static void ggml_backend_cpu_kleidiai_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor,
@ -418,18 +488,35 @@ static size_t ggml_backend_cpu_kleidiai_buffer_type_get_alignment(ggml_backend_b
GGML_UNUSED(buft);
}
static size_t ggml_backend_cpu_kleidiai_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor) {
GGML_ASSERT(tensor->type == GGML_TYPE_Q4_0);
GGML_ASSERT(ctx.kernels);
const size_t n = tensor->ne[1];
const size_t k = tensor->ne[0];
const size_t nr = ctx.kernels->gemm.get_nr();
const size_t kr = ctx.kernels->gemm.get_kr();
return variant_call<size_t>(ctx.kernels->rhs_info.packed_size, n, k, nr, kr, QK4_0);
GGML_UNUSED(buft);
}
namespace ggml::cpu::kleidiai {
class extra_buffer_type : ggml::cpu::extra_buffer_type {
bool supports_op(ggml_backend_dev_t, const struct ggml_tensor * op) override {
if (op->op == GGML_OP_MUL_MAT &&
if ((op->op == GGML_OP_MUL_MAT || op->op == GGML_OP_GET_ROWS) &&
op->src[0]->type == GGML_TYPE_Q4_0 &&
op->src[0]->buffer &&
(ggml_n_dims(op->src[0]) == 2) &&
op->src[0]->buffer->buft == ggml_backend_cpu_kleidiai_buffer_type() && ctx.kernels) {
if (op->op == GGML_OP_GET_ROWS && op->src[1]->ne[0] != 8) {
return false;
}
if (op->src[1]->buffer && !ggml_backend_buft_is_host(op->src[1]->buffer->buft)) {
return false;
}
if (op->src[1]->type == GGML_TYPE_F32 &&
if ((op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == GGML_TYPE_I32) &&
ggml_ne(op->src[1], 2) == 1 && ggml_ne(op->src[1], 3) == 1) {
return true;
}
@ -438,7 +525,7 @@ class extra_buffer_type : ggml::cpu::extra_buffer_type {
}
ggml::cpu::tensor_traits * get_tensor_traits(const struct ggml_tensor * op) override {
if (op->op == GGML_OP_MUL_MAT) {
if (op->op == GGML_OP_MUL_MAT || op->op == GGML_OP_GET_ROWS) {
if (op->src[0]->buffer && op->src[0]->buffer->buft == ggml_backend_cpu_kleidiai_buffer_type()) {
return (ggml::cpu::tensor_traits *) op->src[0]->extra;
}
@ -469,7 +556,7 @@ ggml_backend_buffer_type_t ggml_backend_cpu_kleidiai_buffer_type(void) {
/* .alloc_buffer = */ ggml_backend_cpu_kleidiai_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_cpu_kleidiai_buffer_type_get_alignment,
/* .get_max_size = */ nullptr, // defaults to SIZE_MAX
/* .get_alloc_size = */ nullptr, // defaults to ggml_nbytes
/* .get_alloc_size = */ ggml_backend_cpu_kleidiai_buffer_type_get_alloc_size,
/* .is_host = */ nullptr,
},
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),

281
ggml/src/ggml-cpu/ops.cpp
View File

@ -8,6 +8,7 @@
#include "vec.h"
#include <float.h>
#include <algorithm>
// ggml_compute_forward_dup
@ -1283,6 +1284,7 @@ void ggml_compute_forward_add(
case GGML_TYPE_Q5_0:
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
case GGML_TYPE_MXFP4:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q4_K:
@ -1309,6 +1311,77 @@ void ggml_compute_forward_add(
}
}
// ggml_compute_forward_add_id
static void ggml_compute_forward_add_id_f32(
const ggml_compute_params * params,
ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
const ggml_tensor * src2 = dst->src[2];
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT(src2->type == GGML_TYPE_I32);
GGML_ASSERT(src0->nb[0] == sizeof(float));
GGML_ASSERT(src1->nb[0] == sizeof(float));
const int ith = params->ith;
const int nth = params->nth;
const int nr = ggml_nrows(src0);
GGML_TENSOR_TERNARY_OP_LOCALS
GGML_ASSERT( nb0 == sizeof(float));
GGML_ASSERT(nb10 == sizeof(float));
// rows per thread
const int dr = (nr + nth - 1)/nth;
// row range for this thread
const int ir0 = dr*ith;
const int ir1 = MIN(ir0 + dr, nr);
for (int ir = ir0; ir < ir1; ++ir) {
// src0 indices
const int i3 = ir/(ne2*ne1);
const int i2 = (ir - i3*ne2*ne1)/ne1;
const int i1 = (ir - i3*ne2*ne1 - i2*ne1);
// src1 indices
const int i11 = *(int32_t *) ((char *) src2->data + i1*nb20 + i2*nb21);
GGML_ASSERT(i11 >= 0 && i11 < ne11);
ggml_vec_add_f32(ne0,
(float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ),
(float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01),
(float *) ((char *) src1->data + i11*nb11));
}
}
void ggml_compute_forward_add_id(
const ggml_compute_params * params,
ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
switch (src0->type) {
case GGML_TYPE_F32:
{
ggml_compute_forward_add_id_f32(params, dst);
} break;
default:
{
GGML_ABORT("unsupported type for ggml_compute_forward_add_id: %s", ggml_type_name(src0->type));
}
}
}
// ggml_compute_forward_add1
static void ggml_compute_forward_add1_f32(
@ -1660,6 +1733,7 @@ void ggml_compute_forward_add1(
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q8_1:
case GGML_TYPE_MXFP4:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q4_K:
@ -1787,6 +1861,7 @@ void ggml_compute_forward_acc(
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q8_1:
case GGML_TYPE_MXFP4:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q4_K:
@ -3614,6 +3689,93 @@ static void ggml_compute_forward_swiglu(
}
}
// ggml_compute_forward_swiglu_oai
static void ggml_compute_forward_swiglu_oai_f32(
const ggml_compute_params * params,
ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
char * src0_d = (char *) src0->data;
char * src1_d = (char *) (src1 ? src1->data : src0->data);
const size_t src0_o = src0->nb[1];
const size_t src1_o = src1 ? src1->nb[1] : src0->nb[1];
GGML_ASSERT(ggml_is_contiguous_1(src0));
GGML_ASSERT(ggml_is_contiguous_1(dst));
if (src1) {
GGML_ASSERT(ggml_is_contiguous_1(src1));
GGML_ASSERT(src0->type == src1->type);
}
const int ith = params->ith;
const int nth = params->nth;
const int nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
const int nr = ggml_nrows(src0);
GGML_ASSERT(dst->ne[0] == nc);
GGML_ASSERT(ggml_nrows(dst) == nr);
const int32_t swapped = ggml_get_op_params_i32(dst, 1);
const float alpha = ggml_get_op_params_f32(dst, 2);
const float limit = ggml_get_op_params_f32(dst, 3);
// rows per thread
const int dr = (nr + nth - 1)/nth;
// row range for this thread
const int ir0 = dr*ith;
const int ir1 = MIN(ir0 + dr, nr);
for (int i1 = ir0; i1 < ir1; i1++) {
float * src0_p = (float *) (src0_d + i1*src0_o);
float * src1_p = (float *) (src1_d + i1*src1_o);
float * dst_p = (float *) ((char *) dst->data + i1*(dst->nb[1]));
if (!src1) {
src0_p += swapped ? nc : 0;
src1_p += swapped ? 0 : nc;
}
for (int k = 0; k < nc; k++) {
const float x = std::min(src0_p[k], limit);
const float y = std::clamp(src1_p[k], -limit, limit);
const float out_glu = x / (1.f + expf(alpha * (-x)));
dst_p[k] = out_glu * (y + 1.f);
}
#ifndef NDEBUG
for (int k = 0; k < nc; k++) {
const float x = dst_p[k];
GGML_UNUSED(x);
assert(!isnan(x));
assert(!isinf(x));
}
#endif
}
}
static void ggml_compute_forward_swiglu_oai(
const ggml_compute_params * params,
ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
switch (src0->type) {
case GGML_TYPE_F32:
{
ggml_compute_forward_swiglu_oai_f32(params, dst);
} break;
default:
{
GGML_ABORT("fatal error");
}
}
}
// ggml_compute_forward_geglu_erf
static void ggml_compute_forward_geglu_erf_f32(
@ -4599,6 +4761,7 @@ void ggml_compute_forward_out_prod(
case GGML_TYPE_Q5_0:
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
case GGML_TYPE_MXFP4:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q4_K:
@ -4873,6 +5036,7 @@ void ggml_compute_forward_set(
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q8_1:
case GGML_TYPE_MXFP4:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q4_K:
@ -5134,6 +5298,7 @@ void ggml_compute_forward_get_rows(
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q8_1:
case GGML_TYPE_MXFP4:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q4_K:
@ -5523,6 +5688,7 @@ static void ggml_compute_forward_soft_max_f32(
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
const ggml_tensor * src2 = dst->src[2];
assert(ggml_is_contiguous(dst));
assert(ggml_are_same_shape(src0, dst));
@ -5557,6 +5723,9 @@ static void ggml_compute_forward_soft_max_f32(
const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16);
// sinks
const float * sk = src2 ? (float *)((char *) src2->data) : nullptr;
for (int64_t i03 = 0; i03 < ne03; i03++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
for (int64_t i01 = ith; i01 < ne01; i01 += nth) {
@ -5599,9 +5768,18 @@ static void ggml_compute_forward_soft_max_f32(
float max = -INFINITY;
ggml_vec_max_f32(ne00, &max, wp);
// if we have sinks, make a correction as if they were included in the softmax
if (sk) {
max = MAX(max, sk[i02]);
}
ggml_float sum = ggml_vec_soft_max_f32(ne00, dp, wp, max);
assert(sum > 0.0);
if (sk) {
sum += (ggml_float) expf(sk[i02] - max);
}
sum = 1.0/sum;
ggml_vec_scale_f32(ne00, dp, sum);
@ -5836,6 +6014,7 @@ void ggml_compute_forward_clamp(
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q8_1:
case GGML_TYPE_MXFP4:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q4_K:
@ -7989,12 +8168,14 @@ void ggml_compute_forward_argsort(
static void ggml_compute_forward_flash_attn_ext_f16(
const ggml_compute_params * params,
const ggml_tensor * q,
const ggml_tensor * k,
const ggml_tensor * v,
const ggml_tensor * mask,
ggml_tensor * dst) {
const ggml_tensor * q = dst->src[0];
const ggml_tensor * k = dst->src[1];
const ggml_tensor * v = dst->src[2];
const ggml_tensor * mask = dst->src[3];
const ggml_tensor * sinks = dst->src[4];
GGML_TENSOR_LOCALS(int64_t, neq, q, ne)
GGML_TENSOR_LOCALS(size_t, nbq, q, nb)
GGML_TENSOR_LOCALS(int64_t, nek, k, ne)
@ -8189,6 +8370,23 @@ static void ggml_compute_forward_flash_attn_ext_f16(
}
}
// sinks
if (sinks) {
const float s = ((float *)((char *) sinks->data))[h];
float ms = 1.0f;
float vs = 1.0f;
if (s > M) {
ms = expf(M - s);
ggml_vec_scale_f32(DV, VKQ32, ms);
} else {
vs = expf(s - M);
}
S = S*ms + vs;
}
// V /= S
const float S_inv = 1.0f/S;
ggml_vec_scale_f32(DV, VKQ32, S_inv);
@ -8208,17 +8406,13 @@ static void ggml_compute_forward_flash_attn_ext_f16(
void ggml_compute_forward_flash_attn_ext(
const ggml_compute_params * params,
const ggml_tensor * q,
const ggml_tensor * k,
const ggml_tensor * v,
const ggml_tensor * mask,
ggml_tensor * dst) {
switch (dst->op_params[3]) {
case GGML_PREC_DEFAULT:
case GGML_PREC_F32:
{
// uses F32 accumulators
ggml_compute_forward_flash_attn_ext_f16(params, q, k, v, mask, dst);
ggml_compute_forward_flash_attn_ext_f16(params, dst);
} break;
default:
{
@ -9080,6 +9274,10 @@ void ggml_compute_forward_glu(
{
ggml_compute_forward_swiglu(params, dst);
} break;
case GGML_GLU_OP_SWIGLU_OAI:
{
ggml_compute_forward_swiglu_oai(params, dst);
} break;
case GGML_GLU_OP_GEGLU_ERF:
{
ggml_compute_forward_geglu_erf(params, dst);
@ -10132,6 +10330,7 @@ static void ggml_compute_forward_opt_step_adamw_f32(
const int ir1 = MIN(ir0 + dr, nr);
const float * adamw_params_ptr = ggml_get_data_f32(adamw_params);
const float alpha = adamw_params_ptr[0];
const float beta1 = adamw_params_ptr[1];
const float beta2 = adamw_params_ptr[2];
@ -10139,7 +10338,7 @@ static void ggml_compute_forward_opt_step_adamw_f32(
const float wd = adamw_params_ptr[4];
const float beta1h = adamw_params_ptr[5];
const float beta2h = adamw_params_ptr[6];
const float keep = 1.f - alpha * wd;
for (int ir = ir0; ir < ir1; ++ir) {
const int64_t i03 = ir/(ne02*ne01);
const int64_t i02 = (ir - i03*ne02*ne01)/ne01;
@ -10162,7 +10361,7 @@ static void ggml_compute_forward_opt_step_adamw_f32(
// The weight decay is applied independently of the Adam momenta m and v.
// This is NOT equivalent to l2 regularization that adds w[i00]*w[i00] to the loss.
// See: https://arxiv.org/pdf/1711.05101v3.pdf
w[i00] = w[i00]*(1.0f - alpha*wd) - alpha*mh/vh;
w[i00] = w[i00] * keep - alpha * mh / vh;
}
}
}
@ -10184,3 +10383,63 @@ void ggml_compute_forward_opt_step_adamw(
}
}
}
static void ggml_compute_forward_opt_step_sgd_f32(const ggml_compute_params * params, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src0_grad = dst->src[1];
const ggml_tensor * sgd_params = dst->src[2];
GGML_ASSERT(ggml_are_same_shape(src0, src0_grad));
GGML_ASSERT(ggml_nelements(sgd_params) == 2);
const int ith = params->ith;
const int nth = params->nth;
const int nr = ggml_nrows(src0);
GGML_TENSOR_UNARY_OP_LOCALS
GGML_ASSERT(nb00 == sizeof(float));
// rows per thread
const int dr = (nr + nth - 1) / nth;
// row range for this thread
const int ir0 = dr * ith;
const int ir1 = MIN(ir0 + dr, nr);
// using adamw param subset we care about - alpha, wd - could have a separate struct
const float * sgd_params_ptr = ggml_get_data_f32(sgd_params);
const float alpha = sgd_params_ptr[0];
const float keep = 1.f - alpha * sgd_params_ptr[1];
for (int ir = ir0; ir < ir1; ++ir) {
const int64_t i03 = ir / (ne02 * ne01);
const int64_t i02 = (ir - i03 * ne02 * ne01) / ne01;
const int64_t i01 = (ir - i03 * ne02 * ne01 - i02 * ne01);
const size_t offset = i03 * nb03 + i02 * nb02 + i01 * nb01;
float * w = (float *) ((char *) src0->data + offset); // weight
const float * g = (const float *) ((const char *) src0_grad->data + offset); // grad
for (int i00 = 0; i00 < ne00; ++i00) {
w[i00] = w[i00] * keep - alpha * g[i00];
}
}
}
void ggml_compute_forward_opt_step_sgd(const ggml_compute_params * params, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
switch (src0->type) {
case GGML_TYPE_F32:
{
ggml_compute_forward_opt_step_sgd_f32(params, dst);
}
break;
default:
{
GGML_ABORT("fatal error - sgd is F32 only");
}
}
}

11
ggml/src/ggml-cpu/ops.h
View File

@ -29,6 +29,7 @@ extern "C" {
void ggml_compute_forward_dup(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_add(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_add_id(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_add1(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_acc(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_sum(const struct ggml_compute_params * params, struct ggml_tensor * dst);
@ -82,13 +83,7 @@ void ggml_compute_forward_arange(const struct ggml_compute_params * params, stru
void ggml_compute_forward_timestep_embedding(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_argsort(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_leaky_relu(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_flash_attn_ext(
const struct ggml_compute_params * params,
const struct ggml_tensor * q,
const struct ggml_tensor * k,
const struct ggml_tensor * v,
const struct ggml_tensor * mask,
struct ggml_tensor * dst);
void ggml_compute_forward_flash_attn_ext(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_flash_attn_back(
const struct ggml_compute_params * params,
const bool masked,
@ -112,7 +107,7 @@ void ggml_compute_forward_cross_entropy_loss(const struct ggml_compute_params *
void ggml_compute_forward_cross_entropy_loss_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_opt_step_adamw(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_mul_mat(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_opt_step_sgd(const struct ggml_compute_params * params, struct ggml_tensor * dst);
#ifdef __cplusplus
}
#endif

35
ggml/src/ggml-cpu/quants.c
View File

@ -46,6 +46,10 @@ void quantize_row_q8_1_generic(const float * GGML_RESTRICT x, void * GGML_RESTRI
quantize_row_q8_1_ref(x, y, k);
}
void quantize_row_mxfp4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
quantize_row_mxfp4_ref(x, y, k);
}
//
// 2-6 bit quantization in super-blocks
//
@ -181,6 +185,37 @@ void ggml_vec_dot_q4_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, c
*s = sumf;
}
void ggml_vec_dot_mxfp4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
assert(nrc == 1);
UNUSED(nrc);
UNUSED(bx);
UNUSED(by);
UNUSED(bs);
assert(n % QK_MXFP4 == 0);
static_assert(QK_MXFP4 == QK8_0, "QK_MXFP4 and QK8_0 must be the same");
const block_mxfp4 * GGML_RESTRICT x = vx;
const block_q8_0 * GGML_RESTRICT y = vy;
const int nb = n / QK_MXFP4;
int ib = 0;
float sumf = 0;
for (; ib < nb; ++ib) {
const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_E8M0_TO_FP32_HALF(x[ib].e);
int sumi1 = 0;
int sumi2 = 0;
for (int j = 0; j < QK_MXFP4/2; ++j) {
sumi1 += y[ib].qs[j + 0] * kvalues_mxfp4[x[ib].qs[j] & 0xf];
sumi2 += y[ib].qs[j + QK_MXFP4/2] * kvalues_mxfp4[x[ib].qs[j] >> 4];
}
sumf += d * (sumi1 + sumi2);
}
*s = sumf;
}
void ggml_vec_dot_q5_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
const int qk = QK8_0;
const int nb = n / qk;

8
ggml/src/ggml-cpu/quants.h
View File

@ -19,6 +19,8 @@ void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, in
void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_mxfp4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
@ -39,6 +41,8 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
@ -67,8 +71,12 @@ void ggml_vec_dot_q4_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, c
void ggml_vec_dot_q5_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q5_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q8_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_mxfp4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_tq1_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_tq2_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q2_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q3_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q4_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);

433
ggml/src/ggml-cpu/repack.cpp
View File

@ -14,7 +14,6 @@
#include <cmath>
#include <cstring>
#include <cassert>
#include <cstdlib> // for qsort
#include <cstdio> // for GGML_ASSERT
#include "repack.h"
@ -207,6 +206,7 @@ void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
const int ncols_interleaved = 4;
const int blocklen = 4;
assert(nr == 1);
assert(n % qk == 0);
assert(nc % ncols_interleaved == 0);
@ -308,7 +308,6 @@ void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
UNUSED(ncols_interleaved);
UNUSED(blocklen);
{
float sumf[8];
int sumi;
@ -333,7 +332,6 @@ void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
}
}
}
void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
const int qk = QK_K;
@ -413,11 +411,11 @@ void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
}
}
void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
const int qk = QK8_0;
void ggml_gemv_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
const int qk = QK_K;
const int nb = n / qk;
const int ncols_interleaved = 4;
const int blocklen = 4;
const int ncols_interleaved = 8;
const int blocklen = 8;
assert (n % qk == 0);
assert (nc % ncols_interleaved == 0);
@ -432,7 +430,76 @@ void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
UNUSED(ncols_interleaved);
UNUSED(blocklen);
{
float sumf[8];
float sum_minf[8];
int sumi1,sumi2,sumi3,sumi4;
int sumi;
const block_q8_K * a_ptr = (const block_q8_K *)vy;
for(int x = 0; x < nc / ncols_interleaved; x++) {
const block_q2_Kx8 * b_ptr = (const block_q2_Kx8 *) vx + (x * nb);
for (int j = 0; j < ncols_interleaved; j++) {
sumf[j] = 0.0;
sum_minf[j] = 0.0;
}
for (int l = 0; l < nb; l++) {
for (int k = 0; k < (qk / (4 * blocklen)); k++) {
const uint8_t *scales_0 = b_ptr[l].scales + (k / 4) * 64 ;
const uint8_t *scales_1 = b_ptr[l].scales + (k / 4) * 64 + 16;
const uint8_t *scales_2 = b_ptr[l].scales + (k / 4) * 64 + 32;
const uint8_t *scales_3 = b_ptr[l].scales + (k / 4) * 64 + 48;
for (int j = 0; j < ncols_interleaved; j++) {
sumi1 = 0;
sumi2 = 0;
sumi3 = 0;
sumi4 = 0;
sumi = 0;
int offset = ((k / 2) % 2) + j * 2;
for (int i = 0; i < blocklen; ++i){
const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 3);
const int v1 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 2 ) & 3);
const int v2 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4 ) & 3);
const int v3 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 6 ) & 3);
sumi1 = (v0 * a_ptr[l].qs[(k >> 2) * 128 + (k % 4) * blocklen + i]);
sumi2 = (v1 * a_ptr[l].qs[(k >> 2) * 128 + (k % 4) * blocklen + i + 32]);
sumi3 = (v2 * a_ptr[l].qs[(k >> 2) * 128 + (k % 4) * blocklen + i + 64]);
sumi4 = (v3 * a_ptr[l].qs[(k >> 2) * 128 + (k % 4) * blocklen + i + 96]);
sumi1 = sumi1 * (scales_0[offset] & 0xF);
sumi2 = sumi2 * (scales_1[offset] & 0xF);
sumi3 = sumi3 * (scales_2[offset] & 0xF);
sumi4 = sumi4 * (scales_3[offset] & 0xF);
sumi += sumi1 + sumi2 + sumi3 + sumi4;
}
sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
}
}
for(int sb = 0; sb < 8; sb++) {
const uint8_t *mins = b_ptr[l].scales + sb * 16;
for(int j = 0; j < ncols_interleaved; j++){
sum_minf[j] += ((mins[j * 2] >> 4) * a_ptr[l].bsums[sb * 2] + (mins[(j * 2)+ 1] >> 4) * a_ptr[l].bsums[sb * 2 + 1]) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
}
}
}
for (int j = 0; j < ncols_interleaved; j++) {
s[x * ncols_interleaved + j] = sumf[j] - sum_minf[j];
}
}
}
void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
const int qk = QK8_0;
const int nb = n / qk;
const int ncols_interleaved = 4;
const int blocklen = 4;
assert(nr == 1);
assert(n % qk == 0);
assert(nc % ncols_interleaved == 0);
UNUSED(bs);
UNUSED(nr);
float sumf[4];
int sumi;
@ -457,6 +524,43 @@ void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
}
}
void ggml_gemv_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
const int qk = QK8_0;
const int nb = n / qk;
const int ncols_interleaved = 8;
const int blocklen = 8;
assert(nr == 1);
assert(n % qk == 0);
assert(nc % ncols_interleaved == 0);
UNUSED(bs);
UNUSED(nr);
float sumf[8];
int sumi;
const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
for (int x = 0; x < nc / ncols_interleaved; x++) {
const block_iq4_nlx8 * b_ptr = (const block_iq4_nlx8 *) vx + (x * nb);
for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
for (int l = 0; l < nb; l++) {
for (int k = 0; k < (qk / (2 * blocklen)); k++) {
for (int j = 0; j < ncols_interleaved; j++) {
sumi = 0;
for (int i = 0; i < blocklen; ++i) {
const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
}
sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
}
}
}
for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
}
}
void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@ -712,6 +816,97 @@ void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
}
}
void ggml_gemm_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
const int qk = QK_K;
const int nb = n / qk;
const int ncols_interleaved = 8;
const int blocklen = 8;
assert (n % qk == 0);
assert (nr % 4 == 0);
assert (nc % ncols_interleaved == 0);
UNUSED(s);
UNUSED(bs);
UNUSED(vx);
UNUSED(vy);
UNUSED(nr);
UNUSED(nc);
UNUSED(nb);
UNUSED(ncols_interleaved);
UNUSED(blocklen);
float sumf[4][8];
float sum_minf[4][8];
int sumi1, sumi2, sumi3, sumi4;
int sumi;
for (int y = 0; y < nr / 4; y++) {
const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
for (int x = 0; x < nc / ncols_interleaved; x++) {
const block_q2_Kx8 * b_ptr = (const block_q2_Kx8 *) vx + (x * nb);
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++) {
sumf[m][j] = 0.0;
sum_minf[m][j] = 0.0;
}
}
for (int l = 0; l < nb; l++) {
for (int k = 0; k < (qk / (4 * blocklen)); k++) {
const uint8_t *scales_0 = b_ptr[l].scales + (k / 4) * 64 ;
const uint8_t *scales_1 = b_ptr[l].scales + (k / 4) * 64 + 16;
const uint8_t *scales_2 = b_ptr[l].scales + (k / 4) * 64 + 32;
const uint8_t *scales_3 = b_ptr[l].scales + (k / 4) * 64 + 48;
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++) {
sumi1 = 0;
sumi2 = 0;
sumi3 = 0;
sumi4 = 0;
sumi = 0;
int offset = ((k / 2) % 2) + j * 2;
for (int i = 0; i < blocklen; ++i){
const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 3);
const int v1 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 2 ) & 3);
const int v2 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4 ) & 3);
const int v3 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 6 ) & 3);
sumi1 = (v0 * a_ptr[l].qs[(k >> 2) * 512 + (k % 4) * 4 * blocklen + m * blocklen + i]);
sumi2 = (v1 * a_ptr[l].qs[(k >> 2) * 512 + (k % 4) * 4 * blocklen + m * blocklen + i + 128]);
sumi3 = (v2 * a_ptr[l].qs[(k >> 2) * 512 + (k % 4) * 4 * blocklen + m * blocklen + i + 256]);
sumi4 = (v3 * a_ptr[l].qs[(k >> 2) * 512 + (k % 4) * 4 * blocklen + m * blocklen + i + 384]);
sumi1 = sumi1 * (scales_0[offset] & 0xF);
sumi2 = sumi2 * (scales_1[offset] & 0xF);
sumi3 = sumi3 * (scales_2[offset] & 0xF);
sumi4 = sumi4 * (scales_3[offset] & 0xF);
sumi += sumi1 + sumi2 + sumi3 + sumi4;
}
sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
}
}
}
for(int sb = 0; sb < 8; sb++) {
const uint8_t *mins = b_ptr[l].scales + sb * 16;
for(int m = 0; m < 4; m++) {
const int16_t *bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6);
for(int j = 0; j < ncols_interleaved; j++) {
int mins_prod = ((mins[j * 2] >> 4) * bsums[0] + (mins[(j * 2)+ 1] >> 4) * bsums[1]);
sum_minf[m][j] += (mins_prod) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
}
}
}
}
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++) {
s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j] - sum_minf[m][j];
}
}
}
}
}
void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
const int qk = QK8_0;
const int nb = n / qk;
@ -768,6 +963,50 @@ void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
}
}
void ggml_gemm_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
const int qk = QK8_0;
const int nb = n / qk;
const int ncols_interleaved = 8;
const int blocklen = 8;
assert(n % qk == 0);
assert(nr % 4 == 0);
assert(nc % ncols_interleaved == 0);
float sumf[4][8];
int sumi;
for (int y = 0; y < nr / 4; y++) {
const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
for (int x = 0; x < nc / ncols_interleaved; x++) {
const block_iq4_nlx8 * b_ptr = (const block_iq4_nlx8 *) vx + (x * nb);
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
}
for (int l = 0; l < nb; l++) {
for (int k = 0; k < (qk / (2 * blocklen)); k++) {
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++) {
sumi = 0;
for (int i = 0; i < blocklen; ++i) {
const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
(v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4]));
}
sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
}
}
}
}
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++)
s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
}
}
}
}
} // extern "C"
static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave) {
@ -915,6 +1154,50 @@ static block_q4_Kx8 make_block_q4_Kx8(block_q4_K * in, unsigned int blck_size_in
return out;
}
static block_q2_Kx8 make_block_q2_Kx8(block_q2_K * in, unsigned int blck_size_interleave) {
block_q2_Kx8 out;
// Delta(scale) and dmin values of the eight Q2_K structures are copied onto the output interleaved structure
for (int i = 0; i < 8; i++) {
out.d[i] = in[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.d;
}
for (int i = 0; i < 8; i++) {
out.dmin[i] = in[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.dmin;
}
const int end = QK_K * 2 / blck_size_interleave;
// Interleave Q2_K quants by taking 8 bytes at a time
for (int i = 0; i < end; ++i) {
int src_id = i % 8;
int src_offset = (i / 8) * blck_size_interleave;
int dst_offset = i * blck_size_interleave;
uint64_t elems;
memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
}
// The below logic is designed so as to unpack and rearrange scales and mins values in Q2_K
// Currently the Q2_K structure has 16 scales and 16 mins packed in 16 bytes ( 4 bits for each value)
// The output Q2_Kx8 structure has 128 bytes for storing scales and mins
// Every 16 byte is packed such that it contains scales and mins for corresponding sub blocks from Q2_K structure
// For eg - First 16 bytes contains 16 scales and 16 mins - each of first and second sub blocks from different Q2_K structures
for(int i = 0; i < 128; i++){
// Index for selecting which q2k super block
int src1 = (i % 16) / 2;
// Index for selecting scale
int src2 = ((i / 16) * 2) + (i % 2);
out.scales[i] = in[src1].scales[src2];
}
return out;
}
static int repack_q4_0_to_q4_0_4_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
GGML_ASSERT(t->type == GGML_TYPE_Q4_0);
GGML_ASSERT(interleave_block == 4 || interleave_block == 8);
@ -976,6 +1259,37 @@ static int repack_q4_K_to_q4_K_8_bl(struct ggml_tensor * t, int interleave_block
GGML_UNUSED(data_size);
}
static int repack_q2_K_to_q2_K_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
GGML_ASSERT(t->type == GGML_TYPE_Q2_K);
GGML_ASSERT(interleave_block == 8);
constexpr int nrows_interleaved = 8;
block_q2_Kx8 * dst = (block_q2_Kx8*)t->data;
const block_q2_K * src = (const block_q2_K*) data;
block_q2_K dst_tmp[8];
int nrow = ggml_nrows(t);
int nblocks = t->ne[0] / QK_K;
GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q2_K));
if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) {
return -1;
}
for (int b = 0; b < nrow; b += nrows_interleaved) {
for (int64_t x = 0; x < nblocks; x++) {
for (int i = 0; i < nrows_interleaved; i++ ) {
dst_tmp[i] = src[x + i * nblocks];
}
*dst++ = make_block_q2_Kx8(dst_tmp, interleave_block);
}
src += nrows_interleaved * nblocks;
}
return 0;
GGML_UNUSED(data_size);
}
static int repack_q4_0_to_q4_0_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
GGML_ASSERT(t->type == GGML_TYPE_Q4_0);
GGML_ASSERT(interleave_block == 8);
@ -1044,15 +1358,16 @@ static block_iq4_nlx4 make_block_iq4_nlx4(block_iq4_nl * in, unsigned int blck_s
static int repack_iq4_nl_to_iq4_nl_4_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
GGML_ASSERT(t->type == GGML_TYPE_IQ4_NL);
//GGML_ASSERT(interleave_block == 4 || interleave_block == 8);
GGML_ASSERT(interleave_block == 4);
block_iq4_nlx4 * dst = (block_iq4_nlx4 *)t->data;
const block_iq4_nl * src = (const block_iq4_nl *)data;
block_iq4_nlx4 * dst = ( block_iq4_nlx4 *)t->data;
block_iq4_nl dst_tmp[4];
int nrow = ggml_nrows(t);
int nrows_interleaved = 4;
int nblocks = t->ne[0] / QK4_0;
int nblocks = t->ne[0] / QK4_NL;
GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_iq4_nl));
@ -1074,6 +1389,63 @@ static int repack_iq4_nl_to_iq4_nl_4_bl(struct ggml_tensor * t, int interleave_b
GGML_UNUSED(data_size);
}
static block_iq4_nlx8 make_block_iq4_nlx8(block_iq4_nl * in, unsigned int blck_size_interleave) {
block_iq4_nlx8 out;
for (int i = 0; i < 8; i++) {
out.d[i] = in[i].d;
}
const int end = QK4_NL * 4 / blck_size_interleave;
if (blck_size_interleave == 8) {
for (int i = 0; i < end; ++i) {
int src_id = i % 8;
int src_offset = (i / 8) * blck_size_interleave;
int dst_offset = i * blck_size_interleave;
memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], sizeof(uint64_t));
}
} else {
GGML_ASSERT(false);
}
return out;
}
static int repack_iq4_nl_to_iq4_nl_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
GGML_ASSERT(t->type == GGML_TYPE_IQ4_NL);
GGML_ASSERT(interleave_block == 8);
const block_iq4_nl * src = (const block_iq4_nl *)data;
block_iq4_nlx8 * dst = ( block_iq4_nlx8 *)t->data;
block_iq4_nl dst_tmp[8];
int nrow = ggml_nrows(t);
int nrows_interleaved = 8;
int nblocks = t->ne[0] / QK4_NL;
GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_iq4_nl));
if (t->ne[1] % nrows_interleaved != 0) {
return -1;
}
for (int b = 0; b < nrow; b += nrows_interleaved) {
for (int64_t x = 0; x < nblocks; x++) {
for (int i = 0; i < nrows_interleaved; i++) {
dst_tmp[i] = src[x + i * nblocks];
}
*dst++ = make_block_iq4_nlx8(dst_tmp, interleave_block);
}
src += nrows_interleaved * nblocks;
}
return 0;
GGML_UNUSED(data_size);
}
namespace ggml::cpu::repack {
// repack
template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS>
@ -1096,6 +1468,10 @@ template <> int repack<block_q4_K, 8, 8>(struct ggml_tensor * t, const void * da
return repack_q4_K_to_q4_K_8_bl(t, 8, data, data_size);
}
template <> int repack<block_q2_K, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
return repack_q2_K_to_q2_K_8_bl(t, 8, data, data_size);
}
template <> int repack<block_iq4_nl, 4, 4>(struct ggml_tensor * t, const void * data, size_t data_size) {
return repack_iq4_nl_to_iq4_nl_4_bl(t, 4, data, data_size);
}
@ -1105,6 +1481,10 @@ template <> int repack<block_iq4_nl, 4, 4>(struct ggml_tensor * t, const void *
// return repack_iq4_nl_to_iq4_nl_4_bl(t, 8, data, data_size);
//}
template <> int repack<block_iq4_nl, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
return repack_iq4_nl_to_iq4_nl_8_bl(t, 8, data, data_size);
}
// gemv
template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PARAM_TYPE>
void gemv(int, float *, size_t, const void *, const void *, int, int);
@ -1125,10 +1505,18 @@ template <> void gemv<block_q4_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t
ggml_gemv_q4_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
}
template <> void gemv<block_q2_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
ggml_gemv_q2_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
}
template <> void gemv<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
ggml_gemv_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
}
template <> void gemv<block_iq4_nl, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
ggml_gemv_iq4_nl_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
}
// gemm
template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PARAM_TYPE>
void gemm(int, float *, size_t, const void *, const void *, int, int);
@ -1149,10 +1537,18 @@ template <> void gemm<block_q4_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t
ggml_gemm_q4_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
}
template <> void gemm<block_q2_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
ggml_gemm_q2_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
}
template <> void gemm<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
ggml_gemm_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
}
template <> void gemm<block_iq4_nl, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
ggml_gemm_iq4_nl_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
}
class tensor_traits_base : public ggml::cpu::tensor_traits {
public:
virtual int repack(struct ggml_tensor * t, const void * data, size_t data_size) = 0;
@ -1422,8 +1818,12 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
static const ggml::cpu::repack::tensor_traits<block_q4_0, 8, 8, GGML_TYPE_Q8_0> q4_0_8x8_q8_0;
static const ggml::cpu::repack::tensor_traits<block_q4_K, 8, 8, GGML_TYPE_Q8_K> q4_K_8x8_q8_K;
// instance for Q2
static const ggml::cpu::repack::tensor_traits<block_q2_K, 8, 8, GGML_TYPE_Q8_K> q2_K_8x8_q8_K;
// instance for IQ4
static const ggml::cpu::repack::tensor_traits<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0> iq4_nl_4x4_q8_0;
static const ggml::cpu::repack::tensor_traits<block_iq4_nl, 8, 8, GGML_TYPE_Q8_0> iq4_nl_8x8_q8_0;
if (cur->type == GGML_TYPE_Q4_0) {
if (ggml_cpu_has_avx2() || (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_0)) {
@ -1447,7 +1847,18 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
return &q4_K_8x8_q8_K;
}
}
} else if (cur->type == GGML_TYPE_Q2_K) {
if (ggml_cpu_has_avx512()) {
if (cur->ne[1] % 8 == 0) {
return &q2_K_8x8_q8_K;
}
}
} else if (cur->type == GGML_TYPE_IQ4_NL) {
if (ggml_cpu_has_avx2()) {
if (cur->ne[1] % 8 == 0) {
return &iq4_nl_8x8_q8_0;
}
}
if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
if (cur->ne[1] % 4 == 0) {
return &iq4_nl_4x4_q8_0;

22
ggml/src/ggml-cpu/repack.h
View File

@ -44,7 +44,14 @@ struct block_q4_Kx8 {
};
static_assert(sizeof(block_q4_Kx8) == sizeof(ggml_half) * 16 + K_SCALE_SIZE * 8 + QK_K * 4, "wrong q4_K block size/padding");
struct block_q2_Kx8 {
ggml_half d[8]; // super-block scale for quantized scales
ggml_half dmin[8]; // super-block scale for quantized mins
uint8_t scales[128]; // scales and mins, quantized with 4 bits
uint8_t qs[512]; // 2--bit quants
};
static_assert(sizeof(block_q2_Kx8) == sizeof(ggml_half) * 16 + QK_K/2 + QK_K * 2, "wrong q2_K block size/padding");
struct block_q8_Kx4 {
float d[4]; // delta
int8_t qs[QK_K * 4]; // quants
@ -60,6 +67,13 @@ struct block_iq4_nlx4 {
static_assert(sizeof(block_iq4_nlx4) == 4 * sizeof(ggml_half) + QK4_NL * 2, "wrong iq4_nlx4 block size/padding");
struct block_iq4_nlx8 {
ggml_half d[8]; // deltas for 8 iq4_nl blocks
uint8_t qs[QK4_NL * 4]; // nibbles / quants for 8 iq4_nl blocks
};
static_assert(sizeof(block_iq4_nlx8) == 8 * sizeof(ggml_half) + QK4_NL * 4, "wrong iq4_nlx8 block size/padding");
#if defined(__cplusplus)
extern "C" {
#endif
@ -71,12 +85,16 @@ void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
// Native implementations
void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
@ -86,12 +104,16 @@ void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
void ggml_gemv_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
#if defined(__cplusplus)
} // extern "C"

4
ggml/src/ggml-cpu/traits.cpp
View File

@ -10,7 +10,7 @@ extra_buffer_type::~extra_buffer_type() {}
} // namespace ggml::cpu
bool ggml_cpu_extra_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) {
for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) {
for (auto extra : ggml_backend_cpu_get_extra_buffer_types()) {
if (extra && extra->context) {
auto buf_extra = (ggml::cpu::extra_buffer_type *) extra->context;
auto tensor_traits = buf_extra->get_tensor_traits(op);
@ -23,7 +23,7 @@ bool ggml_cpu_extra_compute_forward(struct ggml_compute_params * params, struct
}
bool ggml_cpu_extra_work_size(int n_threads, const struct ggml_tensor * op, size_t * size) {
for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) {
for (auto extra : ggml_backend_cpu_get_extra_buffer_types()) {
if (extra && extra->context) {
auto buf_extra = (ggml::cpu::extra_buffer_type *) extra->context;
auto tensor_traits = buf_extra->get_tensor_traits(op);

2
ggml/src/ggml-cpu/traits.h
View File

@ -33,6 +33,6 @@ class extra_buffer_type {
} // namespace ggml::cpu
// implemented in ggml-cpu.cpp.
std::vector<ggml_backend_buffer_type_t> & ggml_backend_cpu_get_extra_buffers_type();
std::vector<ggml_backend_buffer_type_t> & ggml_backend_cpu_get_extra_buffer_types();
#endif

23
ggml/src/ggml-cpu/vec.h
View File

@ -55,7 +55,22 @@ inline static void ggml_vec_cpy_i32(const int n, int32_t * y, const int32_t * x)
inline static void ggml_vec_set_f16(const int n, ggml_fp16_t * x, const ggml_fp16_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
inline static void ggml_vec_set_bf16(const int n, ggml_bf16_t * x, const ggml_bf16_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
inline static void ggml_vec_add_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i] + y[i]; }
inline static void ggml_vec_add_f32 (const int n, float * z, const float * x, const float * y) {
int i = 0;
#if defined(__AVX2__)
for (; i + 7 < n; i += 8) {
__m256 vx = _mm256_loadu_ps(x + i);
__m256 vy = _mm256_loadu_ps(y + i);
__m256 vz = _mm256_add_ps(vx, vy);
_mm256_storeu_ps(z + i, vz);
}
#endif
for (; i < n; ++i) {
z[i] = x[i] + y[i];
}
}
inline static void ggml_vec_add_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) {
for (int i = 0; i < n; ++i) {
z[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(x[i]) + GGML_CPU_FP16_TO_FP32(y[i]));
@ -992,9 +1007,9 @@ void ggml_vec_swiglu_f32(const int n, float * y, const float * x, const float *
inline static void ggml_vec_swiglu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const ggml_fp16_t * g) {
for (int i = 0; i < n; ++i) {
float v = GGML_CPU_FP16_TO_FP32(x[i]);
float w = GGML_CPU_FP16_TO_FP32(g[i]);
y[i] = GGML_CPU_FP32_TO_FP16((v/(1.0f + expf(-v))) * w);
float xi = GGML_CPU_FP16_TO_FP32(x[i]);
float gi = GGML_CPU_FP16_TO_FP32(g[i]);
y[i] = GGML_CPU_FP32_TO_FP16((xi/(1.0f + expf(-xi))) * gi);
}
}

10
ggml/src/ggml-cuda/CMakeLists.txt
View File

@ -102,12 +102,12 @@ if (CUDAToolkit_FOUND)
if (GGML_STATIC)
if (WIN32)
# As of 12.3.1 CUDA Toolkit for Windows does not offer a static cublas library
target_link_libraries(ggml-cuda PRIVATE CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
target_link_libraries(ggml-cuda PRIVATE CUDA::cudart_static CUDA::cublas)
else ()
target_link_libraries(ggml-cuda PRIVATE CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
target_link_libraries(ggml-cuda PRIVATE CUDA::cudart_static CUDA::cublas_static)
endif()
else()
target_link_libraries(ggml-cuda PRIVATE CUDA::cudart CUDA::cublas CUDA::cublasLt)
target_link_libraries(ggml-cuda PRIVATE CUDA::cudart CUDA::cublas)
endif()
if (GGML_CUDA_NO_VMM)
@ -120,6 +120,10 @@ if (CUDAToolkit_FOUND)
set(CUDA_FLAGS -use_fast_math -extended-lambda)
if (GGML_CUDA_DEBUG)
list(APPEND CUDA_FLAGS -lineinfo)
endif()
if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "12.8")
# Options are:
# - none (not recommended)

58
ggml/src/ggml-cuda/add-id.cu Normal file
View File

@ -0,0 +1,58 @@
#include "add-id.cuh"
static __global__ void add_id_kernel(
const float * src0, const float * src1, const int32_t * src2, float * dst,
int64_t ne0, int64_t ne1,
size_t nb01, size_t nb02,
size_t nb11,
size_t nb21
) {
const int64_t i1 = blockIdx.x;
const int64_t i2 = blockIdx.y;
const int i11 = *(int32_t *) ((char *) src2 + i1*sizeof(int32_t) + i2*nb21);
const size_t nb1 = ne0 * sizeof(float);
const size_t nb2 = ne1 * nb1;
float * dst_row = (float *)((char *)dst + i1*nb1 + i2*nb2);
const float * src0_row = (const float *)((char *)src0 + i1*nb01 + i2*nb02);
const float * src1_row = (const float *)((char *)src1 + i11*nb11);
for (int64_t i0 = threadIdx.x; i0 < ne0; i0 += blockDim.x) {
dst_row[i0] = src0_row[i0] + src1_row[i0];
}
}
void ggml_cuda_op_add_id(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
const ggml_tensor * src2 = dst->src[2];
GGML_TENSOR_TERNARY_OP_LOCALS
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT(src2->type == GGML_TYPE_I32);
GGML_ASSERT(nb00 == sizeof(float));
GGML_ASSERT(nb10 == sizeof(float));
GGML_ASSERT(nb20 == sizeof(int32_t));
const float * src0_d = (const float *)src0->data;
const float * src1_d = (const float *)src1->data;
const int32_t * src2_d = (const int32_t *)src2->data;
float * dst_d = (float *)dst->data;
int threads = std::min((int)ne00, 768); // cols
dim3 blocks(ne01, ne02); // n_experts_used, n_tokens
add_id_kernel<<<blocks, threads, 0, ctx.stream()>>>(
src0_d, src1_d, src2_d, dst_d,
ne0, ne1,
nb01, nb02,
nb11,
nb21
);
}

3
ggml/src/ggml-cuda/add-id.cuh Normal file
View File

@ -0,0 +1,3 @@
#include "common.cuh"
void ggml_cuda_op_add_id(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

Some files were not shown because too many files have changed in this diff Show More