PaddlePaddle / Paddle

PArallel Distributed Deep LEarning: Machine Learning Framework from Industrial Practice (『飞桨』核心框架,深度学习&机器学习高性能单机、分布式训练和跨平台部署)
http://www.paddlepaddle.org/
Apache License 2.0
22.07k stars 5.54k forks source link

Request for support for CUDA 12.4 #67571

Open umarbutler opened 3 weeks ago

umarbutler commented 3 weeks ago

需求描述 Feature Description

I would like to request support for CUDA 12.4, at least on Windows.

Currently, it is impossible to have the latest versions of PyTorch, Flash Attention 2 and PaddlePaddle installed at the same time.

The reason is that PyTorch requires CUDA 11.8, 12.1 or 12.4, Flash Attention 2 requires 12.XX and PaddlePaddle requires 11.2, 11.6, 11.7, 11.8, 12.0 or 12.3.

替代实现 Alternatives

No response

ForFishes commented 2 weeks ago

Currently 12.3 is supported. This version can already support Flash Attention 2. You can give this version a try. There is no obvious difference between 12.4 and 12.3.

umarbutler commented 2 weeks ago

Currently 12.3 is supported. This version can already support Flash Attention 2. You can give this version a try. There is no obvious difference between 12.4 and 12.3.

The only problem is that PyTorch does not have wheels for CUDA 12.3...