mirror of
https://github.com/saymrwulf/onnxruntime.git
synced 2026-05-23 22:13:38 +00:00
### Description Previously, Flash Attention only worked on Linux systems. This PR will make it work and enable it to be built and run on Windows. Limitations of Flash Attention in Windows: Requires CUDA 12. ### Motivation and Context This will significantly increase the performance of Windows-based LLM's with hardware sm>=80. To illustrate the improvement of Flash Attention over Memory Efficient Attention, here are some average benchmark numbers for the GQA operator, run with configurations based on several recent models (Llama, Mixtral, Phi-3). The benchmarks were obtained on RTX4090 GPU using the test script located at (onnxruntime/test/python/transformers/benchmark_gqa_windows.py). * Clarifying Note: These benchmarks are just for the GQA operator, not the entire model. ### Memory Efficient Attention Kernel Benchmarks: | Model Name | Max Sequence Length | Inference Interval (ms) | Throughput (samples/second) | |----------------------------------------|---------------------|-------------------------|-----------------------------| | Llama3-8B (Average Prompt) | 8192 | 0.19790525 | 13105.63425 | | Llama3-8B (Average Token) | 8192 | 0.207775538 | 12025.10172 | | Llama3-70B (Average Prompt) | 8192 | 0.216049167 | 11563.31185 | | Llama3-70B (Average Token) | 8192 | 0.209730731 | 12284.38149 | | Mixtral-8x22B-v0.1 (Average Prompt) | 32768 | 0.371928785 | 7031.440056 | | Mixtral-8x22B-v0.1 (Average Token) | 32768 | 0.2996659 | 7607.947159 | | Phi-3-mini-128k (Average Prompt) | 131072 | 0.183195867 | 15542.0852 | | Phi-3-mini-128k (Average Token) | 131072 | 0.198215688 | 12874.53494 | | Phi-3-small-128k (Average Prompt) | 65536 | 2.9884929 | 2332.584142 | | Phi-3-small-128k (Average Token) | 65536 | 0.845072406 | 2877.85822 | | Phi-3-medium-128K (Average Prompt) | 32768 | 0.324974429 | 8094.909517 | | Phi-3-medium-128K (Average Token) | 32768 | 0.263662567 | 8978.463687 | ### Flash Attention Kernel Benchmarks: | Model Name | Max Sequence Length | Inference Interval (ms) | Throughput (samples/second) | |--------------------------------------|---------------------|-------------------------|-----------------------------| | Llama3-8B (Average Prompt) | 8192 | 0.163566292 | 16213.69057 | | Llama3-8B (Average Token) | 8192 | 0.161643692 | 16196.14715 | | Llama3-70B (Average Prompt) | 8192 | 0.160510375 | 17448.67753 | | Llama3-70B (Average Token) | 8192 | 0.169427308 | 14702.62043 | | Mixtral-8x22B-v0.1 (Average Prompt) | 32768 | 0.164121964 | 15618.51301 | | Mixtral-8x22B-v0.1 (Average Token) | 32768 | 0.1715865 | 14524.32273 | | Phi-3-mini-128k (Average Prompt) | 131072 | 0.167527167 | 14576.725 | | Phi-3-mini-128k (Average Token) | 131072 | 0.175940594 | 15762.051 | | Phi-3-small-128k (Average Prompt) | 65536 | 0.162719733 | 17824.494 | | Phi-3-small-128k (Average Token) | 65536 | 0.14977525 | 16749.19858 | | Phi-3-medium-128K (Average Prompt) | 32768 | 0.156490786 | 17679.2513 | | Phi-3-medium-128K (Average Token) | 32768 | 0.165333833 | 14932.26079 | Flash Attention is consistently faster for every configuration we benchmarked, with improvements in our trials ranging from ~20% to ~650%. In addition to these improvements in performance, Flash Attention has better memory usage. For example, Memory Efficient Attention cannot handle a max sequence length higher than 32,768, but Flash Attention can handle max sequence lengths at least as high as 131,072. --------- Co-authored-by: Tianlei Wu <tlwu@microsoft.com>
114 lines
3.7 KiB
YAML
114 lines
3.7 KiB
YAML
name: Lint
|
|
|
|
on:
|
|
push:
|
|
branches:
|
|
- main
|
|
- rel-*
|
|
pull_request:
|
|
|
|
jobs:
|
|
optional-lint:
|
|
name: Optional Lint
|
|
runs-on: ubuntu-latest
|
|
steps:
|
|
- uses: actions/checkout@v4
|
|
- name: misspell # Check spellings as well
|
|
uses: reviewdog/action-misspell@v1
|
|
with:
|
|
github_token: ${{ secrets.github_token }}
|
|
locale: "US"
|
|
reporter: github-pr-check
|
|
level: info
|
|
filter_mode: diff_context
|
|
- name: shellcheck # Static check shell scripts
|
|
uses: reviewdog/action-shellcheck@v1
|
|
with:
|
|
github_token: ${{ secrets.github_token }}
|
|
reporter: github-pr-check
|
|
level: info
|
|
filter_mode: file
|
|
|
|
lint-python-format:
|
|
# Required workflow
|
|
name: Python format
|
|
runs-on: ubuntu-latest
|
|
steps:
|
|
- uses: actions/checkout@v4
|
|
- name: Setup Python
|
|
uses: actions/setup-python@v5
|
|
with:
|
|
# Version range or exact version of Python to use, using SemVer's version range syntax. Reads from .python-version if unset.
|
|
python-version: "3.10"
|
|
- name: Setup Rust
|
|
uses: actions-rs/toolchain@v1
|
|
with:
|
|
toolchain: stable
|
|
components: rustfmt
|
|
- name: Install dependencies
|
|
run: |
|
|
python -m pip install -r requirements-dev.txt
|
|
python -m pip install lintrunner lintrunner-adapters
|
|
lintrunner init
|
|
- name: Run lintrunner on all files
|
|
run: |
|
|
set +e
|
|
if ! lintrunner --force-color --all-files --tee-json=lint.json -v; then
|
|
echo ""
|
|
echo -e "\e[1m\e[36mYou can reproduce these results locally by using \`lintrunner\`. To set up lintrunner locally, see https://github.com/microsoft/onnxruntime/blob/main/docs/Coding_Conventions_and_Standards.md#linting .\e[0m"
|
|
exit 1
|
|
fi
|
|
- name: Produce SARIF
|
|
if: always()
|
|
run: |
|
|
python -m lintrunner_adapters to-sarif lint.json lintrunner.sarif
|
|
- name: Upload SARIF file
|
|
if: always()
|
|
continue-on-error: true
|
|
uses: github/codeql-action/upload-sarif@v3
|
|
with:
|
|
# Path to SARIF file relative to the root of the repository
|
|
sarif_file: lintrunner.sarif
|
|
category: lintrunner
|
|
checkout_path: ${{ github.workspace }}
|
|
|
|
lint-cpp:
|
|
name: Lint C++
|
|
runs-on: ubuntu-latest
|
|
steps:
|
|
- uses: actions/checkout@master
|
|
- name: Install ninja
|
|
run: python -m pip install --upgrade ninja
|
|
- name: Generate compile_commands.json
|
|
run: |
|
|
python tools/ci_build/build.py \
|
|
--cmake_generator "Ninja" \
|
|
--build_dir build \
|
|
--update \
|
|
--cmake_extra_defines CMAKE_EXPORT_COMPILE_COMMANDS=ON
|
|
- name: Generate ONNX protobuf files
|
|
run: cmake --build build/Debug --config Debug --target onnx_proto
|
|
- uses: reviewdog/action-cpplint@master
|
|
with:
|
|
github_token: ${{ secrets.github_token }}
|
|
reporter: github-pr-check
|
|
level: warning
|
|
flags: --linelength=120
|
|
--exclude=java/src/main/native/*.c
|
|
--exclude=onnxruntime/core/mlas/inc/*
|
|
--exclude=onnxruntime/core/mlas/lib/*
|
|
--exclude=onnxruntime/contrib_ops/cuda/bert/flash_attention/*
|
|
filter: "-runtime/references"
|
|
|
|
lint-js:
|
|
name: Lint JavaScript
|
|
runs-on: ubuntu-latest
|
|
steps:
|
|
- uses: actions/checkout@v4
|
|
- uses: reviewdog/action-eslint@v1
|
|
with:
|
|
reporter: github-pr-check
|
|
level: error
|
|
filter_mode: file
|
|
eslint_flags: "--ext .ts --ext .tsx"
|
|
workdir: "js/"
|