yindo/flash-attention-prebuild-wheels-rocm
1
0
Fork 0
mirror of https://github.com/BillyOutlast/flash-attention-prebuild-wheels-rocm.git synced 2026-07-01 01:27:54 -04:00
Code Issues Packages Projects Releases 2 Wiki Activity

update README.md

Browse Source
This commit is contained in:
Junya Morioka
2025-05-18 02:26:18 +09:00
parent 7228789eab
commit d0d4cbbc6b
2 changed files with 79 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files
README.md
+79 -29
View File
@@ -1,14 +1,13 @@
# flash-attention pre-build wheels # flash-attention pre-build wheels
This repository provides wheels for the pre-built [flash-attention](https://github.com/Dao-AILab/flash-attention). This repository provides wheels for the pre-built [flash-attention](https://github.com/Dao-AILab/flash-attention).
Since building flash-attention takes a **very long time** and is resource-intensive, Since building flash-attention takes a **very long time** and is resource-intensive,
I also build and provide combinations of CUDA and PyTorch that are not officially distributed. I also build and provide combinations of CUDA and PyTorch that are not officially distributed.
The building Github Actions Workflow can be found [here](./.github/workflows/build.yml). The building Github Actions Workflow can be found [here](./.github/workflows/build.yml).
The built packages are available on the [release page](https://github.com/mjun0812/flash-attention-prebuild-wheels/releases). The built packages are available on the [release page](https://github.com/mjun0812/flash-attention-prebuild-wheels/releases).
## Install ## Install
1. Select the versions for Python, CUDA, PyTorch, and flash_attn. 1. Select the versions for Python, CUDA, PyTorch, and flash_attn.
@@ -22,7 +21,7 @@ flash_attn-2.6.3+cu124torch2.5-cp312-cp312-linux_x86_64.whl
2. Find the corresponding version of a wheel from the below table and [releases](https://github.com/mjun0812/flash-attention-prebuild-wheels/releases) 2. Find the corresponding version of a wheel from the below table and [releases](https://github.com/mjun0812/flash-attention-prebuild-wheels/releases)
5. Direct Install or Download and Local Install 3. Direct Install or Download and Local Install
```bash ```bash
# Direct Install # Direct Install
@@ -39,16 +38,16 @@ pip install ./flash_attn-2.6.3+cu124torch2.5-cp312-cp312-linux_x86_64.whl
[Release](https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/tag/v0.0.9) [Release](https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/tag/v0.0.9)
| Flash-Attention | Python | PyTorch | CUDA | | Flash-Attention | Python | PyTorch | CUDA |
|-----------------|--------|---------|------| | ------------------- | ---------------- | ------- | ------ |
| 2.4.3, 2.5.9, 2.6.3 | 3.10, 3.11, 3.12 | 2.7.0 | 12.8.1 | | 2.4.3, 2.5.9, 2.6.3 | 3.10, 3.11, 3.12 | 2.7.0 | 12.8.1 |
### v0.0.8 ### v0.0.8
[Release](https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/tag/v0.0.8) [Release](https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/tag/v0.0.8)
| Flash-Attention | Python | PyTorch | CUDA | | Flash-Attention | Python | PyTorch | CUDA |
|-----------------|--------|---------|------| | -------------------------------- | ---------------- | -------------------------- | ---------------------- |
| 2.4.3, 2.5.9, 2.6.3, 2.7.4.post1 | 3.10, 3.11, 3.12 | 2.4.1, 2.5.1, 2.6.0, 2.7.0 | 11.8.0, 12.4.1, 12.6.3 | | 2.4.3, 2.5.9, 2.6.3, 2.7.4.post1 | 3.10, 3.11, 3.12 | 2.4.1, 2.5.1, 2.6.0, 2.7.0 | 11.8.0, 12.4.1, 12.6.3 |
### v0.0.7 ### v0.0.7
@@ -59,61 +58,112 @@ Skip for experimental reasons.
[Release](https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/tag/v0.0.6) [Release](https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/tag/v0.0.6)
| Flash-Attention | Python | PyTorch | CUDA | | Flash-Attention | Python | PyTorch | CUDA |
|-----------------|--------|---------|------| | -------------------------------- | ---------------- | --------------------------------- | -------------- |
| 2.4.3, 2.5.9, 2.6.3, 2.7.4.post1 | 3.10, 3.11, 3.12 | 2.2.2, 2.3.1, 2.4.1, 2.5.1, 2.6.0 | 12.4.1, 12.6.3 | | 2.4.3, 2.5.9, 2.6.3, 2.7.4.post1 | 3.10, 3.11, 3.12 | 2.2.2, 2.3.1, 2.4.1, 2.5.1, 2.6.0 | 12.4.1, 12.6.3 |
### v0.0.5 ### v0.0.5
[Release](https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/tag/v0.0.5) [Release](https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/tag/v0.0.5)
| Flash-Attention | Python | PyTorch | CUDA | | Flash-Attention | Python | PyTorch | CUDA |
|-----------------|--------|---------|------| | ------------------ | ---------------- | ----------------------------------------------- | -------------- |
| 2.6.3, 2.7.4.post1 | 3.10, 3.11, 3.12 | 2.0.1, 2.1.2, 2.2.2, 2.3.1, 2.4.1, 2.5.1, 2.6.0 | 12.4.1, 12.6.3 | | 2.6.3, 2.7.4.post1 | 3.10, 3.11, 3.12 | 2.0.1, 2.1.2, 2.2.2, 2.3.1, 2.4.1, 2.5.1, 2.6.0 | 12.4.1, 12.6.3 |
### v0.0.4 ### v0.0.4
[Release](https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/tag/v0.0.4) [Release](https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/tag/v0.0.4)
| Flash-Attention | Python | PyTorch | CUDA | | Flash-Attention | Python | PyTorch | CUDA |
|-----------------|--------|---------|------| | --------------- | ---------------- | ---------------------------------------- | ---------------------- |
| 2.7.3 | 3.10, 3.11, 3.12 | 2.0.1, 2.1.2, 2.2.2, 2.3.1, 2.4.1, 2.5.1 | 11.8.0, 12.1.1, 12.4.1 | | 2.7.3 | 3.10, 3.11, 3.12 | 2.0.1, 2.1.2, 2.2.2, 2.3.1, 2.4.1, 2.5.1 | 11.8.0, 12.1.1, 12.4.1 |
### v0.0.3 ### v0.0.3
[Release](https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/tag/v0.0.3) [Release](https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/tag/v0.0.3)
| Flash-Attention | Python | PyTorch | CUDA | | Flash-Attention | Python | PyTorch | CUDA |
|-----------------|--------|---------|------| | --------------- | ---------------- | ---------------------------------------- | ---------------------- |
| 2.7.2.post1 | 3.10, 3.11, 3.12 | 2.0.1, 2.1.2, 2.2.2, 2.3.1, 2.4.1, 2.5.1 | 11.8.0, 12.1.1, 12.4.1 | | 2.7.2.post1 | 3.10, 3.11, 3.12 | 2.0.1, 2.1.2, 2.2.2, 2.3.1, 2.4.1, 2.5.1 | 11.8.0, 12.1.1, 12.4.1 |
### v0.0.2 ### v0.0.2
[Release](https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/tag/v0.0.2) [Release](https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/tag/v0.0.2)
| Flash-Attention | Python | PyTorch | CUDA | | Flash-Attention | Python | PyTorch | CUDA |
|-----------------|--------|---------|------| | -------------------------------- | ---------------- | ---------------------------------------- | ---------------------- |
| 2.4.3, 2.5.6, 2.6.3, 2.7.0.post2 | 3.10, 3.11, 3.12 | 2.0.1, 2.1.2, 2.2.2, 2.3.1, 2.4.1, 2.5.1 | 11.8.0, 12.1.1, 12.4.1 | | 2.4.3, 2.5.6, 2.6.3, 2.7.0.post2 | 3.10, 3.11, 3.12 | 2.0.1, 2.1.2, 2.2.2, 2.3.1, 2.4.1, 2.5.1 | 11.8.0, 12.1.1, 12.4.1 |
### v0.0.1 ### v0.0.1
[Release](https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/tag/v0.0.1) [Release](https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/tag/v0.0.1)
|flash-attention|Python|PyTorch|CUDA| | flash-attention | Python | PyTorch | CUDA |
|-|-|-|-| | --------------------------------- | ---------------- | ---------------------------------------- | ---------------------- |
|1.0.9, 2.4.3, 2.5.6, 2.5.9, 2.6.3|3.10, 3.11, 3.12|2.0.1, 2.1.2, 2.2.2, 2.3.1, 2.4.1, 2.5.0|11.8.0, 12.1.1, 12.4.1| | 1.0.9, 2.4.3, 2.5.6, 2.5.9, 2.6.3 | 3.10, 3.11, 3.12 | 2.0.1, 2.1.2, 2.2.2, 2.3.1, 2.4.1, 2.5.0 | 11.8.0, 12.1.1, 12.4.1 |
### v0.0.0 ### v0.0.0
[Release](https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/tag/v0.0.0) [Release](https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/tag/v0.0.0)
|flash-attention|Python|PyTorch|CUDA| | flash-attention | Python | PyTorch | CUDA |
|-|-|-|-| | -------------------------- | ---------- | ---------------------------------------- | ---------------------- |
|2.4.3, 2.5.6, 2.5.9, 2.6.3|3.11, 3.12|2.0.1, 2.1.2, 2.2.2, 2.3.1, 2.4.1, 2.5.0|11.8.0, 12.1.1, 12.4.1| | 2.4.3, 2.5.6, 2.5.9, 2.6.3 | 3.11, 3.12 | 2.0.1, 2.1.2, 2.2.2, 2.3.1, 2.4.1, 2.5.0 | 11.8.0, 12.1.1, 12.4.1 |
## Self build
## Original If you want to build the wheels yourself, you can folk this repository and run the build workflow.
1. Fork this repository
2. Edit workflow file `.github/workflows/build.yml` to set the version you want to build.
3. Add tag `v*.*.*` to trigger the build workflow.
### Self-Hosted Runner Build
In some version combinations, you cannot build wheels on GitHub-hosted runners due to job time limitations.
To build the wheels for these versions, you can use self-hosted runners.
```bash
git clone https://github.com/mjun0812/flash-attention-prebuild-wheels.git
cd self-hosted-runner
cp env.template env
```
Edit `env` file to set the environment variables.
```bash
# Edit env
PERSONAL_ACCESS_TOKEN=[Github Personal Access Token]
```
Edit compose.yml file if you use repository folked from this repository.
```yaml
services:
runner:
privileged: true
build:
context: .
dockerfile: Dockerfile
args:
REPOSITORY_URL: [Target Repository URL]
PERSONAL_ACCESS_TOKEN: $PERSONAL_ACCESS_TOKEN
GH_RUNNER_VERSION: 2.324.0
RUNNER_NAME: self-hosted-runner
RUNNER_GROUP: default
RUNNER_LABELS: self-hosted
TARGET_ARCH: x64
```
Then, build and run the docker container.
```bash
# Build and run
docker compose build
docker compose up -d
```
## Original Repository
[repo](https://github.com/Dao-AILab/flash-attention) [repo](https://github.com/Dao-AILab/flash-attention)
self-hosted-runner/env.template
-1
View File
@@ -1,2 +1 @@
PERSONAL_ACCESS_TOKEN= PERSONAL_ACCESS_TOKEN=
DOCKER_GID=