add flash-attn v2.8.3

This commit is contained in:
Junya Morioka
2025-08-23 07:00:13 +09:00
parent 89ce18994e
commit 955b7b8cdf
+98 -97
View File
@@ -22,54 +22,54 @@ jobs:
# #########################################################
# Linux
# #########################################################
# build_wheels_linux:
# name: Build Linux
# needs: create_releases
# strategy:
# fail-fast: false
# matrix:
# flash-attn-version: ["2.7.4"]
# python-version: ["3.10", "3.11", "3.12"]
# torch-version: ["2.5.1", "2.6.0", "2.7.1", "2.8.0"]
# # https://developer.nvidia.com/cuda-toolkit-archive
# cuda-version: ["12.4.1", "12.8.1", "12.9.1"]
# exclude:
# # torch < 2.2 does not support Python 3.12
# - python-version: "3.12"
# torch-version: "2.0.1"
# - python-version: "3.12"
# torch-version: "2.1.2"
# # torch 2.0.1 does not support CUDA 12.x
# - torch-version: "2.0.1"
# cuda-version: "12.1.1"
# - torch-version: "2.0.1"
# cuda-version: "12.4.1"
# - torch-version: "2.0.1"
# cuda-version: "12.6.3"
# - torch-version: "2.0.1"
# cuda-version: "12.8.1"
# # torch 2.6.0 does not support CUDA 12.1
# - torch-version: "2.6.0"
# cuda-version: "12.1.1"
# # torch 2.7.0 does not support CUDA 12.4
# - torch-version: "2.7.0"
# cuda-version: "12.4.1"
# # torch < 2.8 does not support CUDA 12.9
# - torch-version: "2.5.1"
# cuda-version: "12.9.1"
# - torch-version: "2.6.3"
# cuda-version: "12.9.1"
# - torch-version: "2.7.1"
# cuda-version: "12.9.1"
# # flash-attn 2.7.4 does not build in GitHub Hosted Runner
# - flash-attn-version: "2.7.4"
# uses: ./.github/workflows/build_linux.yml
# with:
# flash-attn-version: ${{ matrix.flash-attn-version }}
# python-version: ${{ matrix.python-version }}
# torch-version: ${{ matrix.torch-version }}
# cuda-version: ${{ matrix.cuda-version }}
# secrets: inherit
build_wheels_linux:
name: Build Linux
needs: create_releases
strategy:
fail-fast: false
matrix:
flash-attn-version: ["2.8.3"]
python-version: ["3.10", "3.11", "3.12"]
torch-version: ["2.5.1", "2.6.0", "2.7.1", "2.8.0"]
# https://developer.nvidia.com/cuda-toolkit-archive
cuda-version: ["12.4.1", "12.6.3", "12.8.1", "12.9.1"]
exclude:
# torch < 2.2 does not support Python 3.12
- python-version: "3.12"
torch-version: "2.0.1"
- python-version: "3.12"
torch-version: "2.1.2"
# torch 2.0.1 does not support CUDA 12.x
- torch-version: "2.0.1"
cuda-version: "12.1.1"
- torch-version: "2.0.1"
cuda-version: "12.4.1"
- torch-version: "2.0.1"
cuda-version: "12.6.3"
- torch-version: "2.0.1"
cuda-version: "12.8.1"
# torch 2.6.0 does not support CUDA 12.1
- torch-version: "2.6.0"
cuda-version: "12.1.1"
# torch 2.7.0 does not support CUDA 12.4
- torch-version: "2.7.0"
cuda-version: "12.4.1"
# torch < 2.8 does not support CUDA 12.9
- torch-version: "2.5.1"
cuda-version: "12.9.1"
- torch-version: "2.6.3"
cuda-version: "12.9.1"
- torch-version: "2.7.1"
cuda-version: "12.9.1"
# flash-attn 2.7.4 does not build in GitHub Hosted Runner
- flash-attn-version: "2.7.4"
uses: ./.github/workflows/build_linux.yml
with:
flash-attn-version: ${{ matrix.flash-attn-version }}
python-version: ${{ matrix.python-version }}
torch-version: ${{ matrix.torch-version }}
cuda-version: ${{ matrix.cuda-version }}
secrets: inherit
# build_wheels_linux_self_hosted:
# name: Build Linux (self-hosted)
@@ -77,7 +77,7 @@ jobs:
# strategy:
# fail-fast: false
# matrix:
# flash-attn-version: ["2.7.4"]
# flash-attn-version: ["2.8.3"]
# python-version: ["3.10", "3.11", "3.12"]
# torch-version: ["2.5.1", "2.6.0", "2.7.1", "2.8.0"]
# # https://developer.nvidia.com/cuda-toolkit-archive
@@ -168,58 +168,59 @@ jobs:
# cuda-version: ${{ matrix.cuda-version }}
# secrets: inherit
build_wheels_windows_code_build:
name: Build Windows (AWS CodeBuild)
needs: create_releases
strategy:
fail-fast: false
matrix:
flash-attn-version: ["2.7.4", "2.8.2"]
python-version: ["3.10", "3.11", "3.12"]
torch-version: ["2.7.1", "2.8.0"]
# https://developer.nvidia.com/cuda-toolkit-archive
cuda-version: ["12.8.1"]
exclude:
# torch < 2.2 does not support Python 3.12
- python-version: "3.12"
torch-version: "2.0.1"
- python-version: "3.12"
torch-version: "2.1.2"
# torch 2.0.1 does not support CUDA 12.x
- torch-version: "2.0.1"
cuda-version: "12.1.1"
- torch-version: "2.0.1"
cuda-version: "12.4.1"
- torch-version: "2.0.1"
cuda-version: "12.6.3"
- torch-version: "2.0.1"
cuda-version: "12.8.1"
# torch 2.6.0 does not support CUDA 12.1
- torch-version: "2.6.0"
cuda-version: "12.1.1"
# torch 2.7.0 does not support CUDA 12.4
- torch-version: "2.7.0"
cuda-version: "12.4.1"
# torch < 2.8 does not support CUDA 12.9
- torch-version: "2.5.1"
cuda-version: "12.9.1"
- torch-version: "2.6.3"
cuda-version: "12.9.1"
- torch-version: "2.7.1"
cuda-version: "12.9.1"
uses: ./.github/workflows/build_windows_code_build.yml
with:
flash-attn-version: ${{ matrix.flash-attn-version }}
python-version: ${{ matrix.python-version }}
torch-version: ${{ matrix.torch-version }}
cuda-version: ${{ matrix.cuda-version }}
secrets: inherit
# build_wheels_windows_code_build:
# name: Build Windows (AWS CodeBuild)
# needs: create_releases
# strategy:
# fail-fast: false
# matrix:
# flash-attn-version: ["2.7.4", "2.8.2"]
# python-version: ["3.10", "3.11", "3.12"]
# torch-version: ["2.7.1", "2.8.0"]
# # https://developer.nvidia.com/cuda-toolkit-archive
# cuda-version: ["12.8.1"]
# exclude:
# # torch < 2.2 does not support Python 3.12
# - python-version: "3.12"
# torch-version: "2.0.1"
# - python-version: "3.12"
# torch-version: "2.1.2"
# # torch 2.0.1 does not support CUDA 12.x
# - torch-version: "2.0.1"
# cuda-version: "12.1.1"
# - torch-version: "2.0.1"
# cuda-version: "12.4.1"
# - torch-version: "2.0.1"
# cuda-version: "12.6.3"
# - torch-version: "2.0.1"
# cuda-version: "12.8.1"
# # torch 2.6.0 does not support CUDA 12.1
# - torch-version: "2.6.0"
# cuda-version: "12.1.1"
# # torch 2.7.0 does not support CUDA 12.4
# - torch-version: "2.7.0"
# cuda-version: "12.4.1"
# # torch < 2.8 does not support CUDA 12.9
# - torch-version: "2.5.1"
# cuda-version: "12.9.1"
# - torch-version: "2.6.3"
# cuda-version: "12.9.1"
# - torch-version: "2.7.1"
# cuda-version: "12.9.1"
# uses: ./.github/workflows/build_windows_code_build.yml
# with:
# flash-attn-version: ${{ matrix.flash-attn-version }}
# python-version: ${{ matrix.python-version }}
# torch-version: ${{ matrix.torch-version }}
# cuda-version: ${{ matrix.cuda-version }}
# secrets: inherit
update_release_notes:
name: Update Release Notes
# needs:
needs:
# [build_wheels_linux, build_wheels_linux_self_hosted, build_wheels_windows]
needs: [build_wheels_windows_code_build]
[build_wheels_linux]
# needs: [build_wheels_windows_code_build]
if: always()
runs-on: ubuntu-latest
steps: