Do we have flash_attn-2.8.3 wheel with cu12 + torch2.9 for cp311 ?

[wolfvoid]

My project uses vllm-0.14.1, which depends on torch 2.9.1, however in my python 3.11 version i can't find a flash attention wheel for cu12torch2.9cxx11abiTRUE-cp311-cp311-linux_x86_64. You will help me a lot if you can provide one or help me solve the problem using another way.
Thank you very much.

[cyk2018]

maybe build from source code ? follow the readme doc using python setup.py install

[wolfvoid]

Thanks, I try to use "pip install flash-attn --no-build-isolation" and get 2.8.3. It seems that it builds from begins and takes quite a long time. About 1.5 hours in Linux Server with 64 CPUs, but i use MAX_JOBS=4 to avoid getting out of memory. It will be better if we have cu12torch2.9cxx11abiTRUE-cp311-cp311-linux_x86_64.wheel.

[cyk2018]

Thanks, I try to use "pip install flash-attn --no-build-isolation" and get 2.8.3. It seems that it builds from begins and takes quite a long time. About 1.5 hours in Linux Server with 64 CPUs, but i use MAX_JOBS=4 to avoid getting out of memory. It will be better if we have cu12torch2.9cxx11abiTRUE-cp311-cp311-linux_x86_64.wheel.

this will get from internet the 2.8.3 which is build 5 months ago, So if you don't use something like FA4, this is OK. By the way, MAX_JOBS=4 seems slow for many machines, so maybe you can build without this or set this bigger to get faster.

[Oukaishen]

Hi there, we also need this falsh-attention 2.8.3 for cuda12.9 + torch 2.9 + python 3.11

can you also provide the wheel?

[Oukaishen]

@cyk2018 hi there can we make this wheel for release?