onnxruntime/.github/workflows/lint.yml
aciddelgado ebd0368bb0
Make Flash Attention work on Windows (#21015)
### 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>
2024-06-24 09:43:49 -07:00

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/"