mirror of
https://github.com/saymrwulf/onnxruntime.git
synced 2026-07-06 04:28:32 +00:00
ebd0368bb0
### 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>
118 lines
3.7 KiB
YAML
118 lines
3.7 KiB
YAML
##### start trigger Don't edit it manually, Please do edit set-trigger-rules.py ####
|
|
trigger:
|
|
branches:
|
|
include:
|
|
- main
|
|
- rel-*
|
|
paths:
|
|
exclude:
|
|
- docs/**
|
|
- README.md
|
|
- CONTRIBUTING.md
|
|
- BUILD.md
|
|
- 'js/web'
|
|
- 'onnxruntime/core/providers/js'
|
|
pr:
|
|
branches:
|
|
include:
|
|
- main
|
|
- rel-*
|
|
paths:
|
|
exclude:
|
|
- docs/**
|
|
- README.md
|
|
- CONTRIBUTING.md
|
|
- BUILD.md
|
|
- 'js/web'
|
|
- 'onnxruntime/core/providers/js'
|
|
#### end trigger ####
|
|
|
|
parameters:
|
|
- name: RunOnnxRuntimeTests
|
|
displayName: Run Tests?
|
|
type: boolean
|
|
default: true
|
|
|
|
stages:
|
|
- stage: cuda
|
|
dependsOn: []
|
|
jobs:
|
|
- template: templates/jobs/win-ci-vs-2022-job.yml
|
|
parameters:
|
|
BuildConfig: 'RelWithDebInfo'
|
|
EnvSetupScript: setup_env_cuda.bat
|
|
buildArch: x64
|
|
additionalBuildFlags: >-
|
|
--enable_pybind --build_java --build_nodejs --use_cuda --cuda_home="$(Agent.TempDirectory)\v11.8"
|
|
--enable_cuda_profiling --enable_transformers_tool_test
|
|
--cmake_extra_defines CMAKE_CUDA_ARCHITECTURES=86
|
|
--cmake_extra_defines onnxruntime_BUILD_UNIT_TESTS=ON
|
|
--cmake_extra_defines onnxruntime_ENABLE_CUDA_EP_INTERNAL_TESTS=ON
|
|
msbuildPlatform: x64
|
|
isX86: false
|
|
job_name_suffix: x64_RelWithDebInfo
|
|
RunOnnxRuntimeTests: ${{ parameters.RunOnnxRuntimeTests }}
|
|
ORT_EP_NAME: CUDA
|
|
WITH_CACHE: true
|
|
MachinePool: onnxruntime-Win2022-GPU-A10
|
|
|
|
- stage: training
|
|
dependsOn: []
|
|
jobs:
|
|
- template: templates/jobs/win-ci-vs-2022-job.yml
|
|
parameters:
|
|
BuildConfig: 'RelWithDebInfo'
|
|
EnvSetupScript: setup_env_cuda.bat
|
|
buildArch: x64
|
|
additionalBuildFlags: >-
|
|
--enable_pybind --enable_training --use_cuda --cuda_home="$(Agent.TempDirectory)\v11.8"
|
|
--skip_onnx_tests
|
|
--cmake_extra_defines CMAKE_CUDA_ARCHITECTURES=86
|
|
msbuildPlatform: x64
|
|
isX86: false
|
|
job_name_suffix: x64_RelWithDebInfo
|
|
RunOnnxRuntimeTests: ${{ parameters.RunOnnxRuntimeTests }}
|
|
ORT_EP_NAME: CUDA
|
|
WITH_CACHE: true
|
|
MachinePool: onnxruntime-Win2022-GPU-A10
|
|
isTraining: true
|
|
|
|
- stage: dml
|
|
dependsOn: []
|
|
jobs:
|
|
- template: templates/jobs/win-ci-vs-2022-job.yml
|
|
parameters:
|
|
BuildConfig: 'RelWithDebInfo'
|
|
EnvSetupScript: setup_env.bat
|
|
buildArch: x64
|
|
additionalBuildFlags: --enable_pybind --use_dml --enable_wcos --use_winml
|
|
msbuildPlatform: x64
|
|
isX86: false
|
|
job_name_suffix: x64_RelWithDebInfo
|
|
RunOnnxRuntimeTests: ${{ parameters.RunOnnxRuntimeTests }}
|
|
ORT_EP_NAME: DML
|
|
WITH_CACHE: false
|
|
MachinePool: onnxruntime-Win2022-GPU-dml-A10
|
|
|
|
- stage: kernelDocumentation
|
|
dependsOn: []
|
|
jobs:
|
|
- template: templates/jobs/win-ci-vs-2022-job.yml
|
|
parameters:
|
|
BuildConfig: 'RelWithDebInfo'
|
|
EnvSetupScript: setup_env_cuda.bat
|
|
buildArch: x64
|
|
# note: need to specify `--gen_doc` when creating the build config so it has to be in additionalBuildFlags
|
|
additionalBuildFlags: >-
|
|
--gen_doc validate --skip_tests --enable_pybind --use_dml --use_cuda
|
|
--cuda_home="$(Agent.TempDirectory)\v11.8"
|
|
--cmake_extra_defines CMAKE_CUDA_ARCHITECTURES=86
|
|
--cmake_extra_defines onnxruntime_BUILD_UNIT_TESTS=OFF
|
|
msbuildPlatform: x64
|
|
isX86: false
|
|
job_name_suffix: x64_RelWithDebInfo
|
|
RunOnnxRuntimeTests: false
|
|
GenerateDocumentation: true
|
|
ORT_EP_NAME: CUDA # It doesn't really matter which EP is selected here since this stage is for documentation.
|
|
WITH_CACHE: true
|
|
MachinePool: onnxruntime-Win2022-GPU-A10
|