when will rtx50 series GPU be supported?

[nianqiu-liu]

thanks!

[tridao]

It's already supported

[HDANILO]

Double checking torch functionality:

(base) hdanilo@DragonRollDev:/mnt/c/Users/helto/diffusion-pipe/flash-attention$ python
Python 3.12.9 | packaged by Anaconda, Inc. | (main, Feb  6 2025, 18:56:27) [GCC 11.2.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch
>>> print(torch.cuda.is_available())
True
>>> print(torch.cuda.get_arch_list())
['sm_75', 'sm_80', 'sm_86', 'sm_90', 'sm_100', 'sm_120', 'compute_120']
>>> zeros_tensor_gpu = torch.zeros((50, 50), device='cuda')
>>> zeros_tensor_gpu
tensor([[0., 0., 0.,  ..., 0., 0., 0.],
        [0., 0., 0.,  ..., 0., 0., 0.],
        [0., 0., 0.,  ..., 0., 0., 0.],
        ...,
        [0., 0., 0.,  ..., 0., 0., 0.],
        [0., 0., 0.,  ..., 0., 0., 0.],
        [0., 0., 0.,  ..., 0., 0., 0.]], device='cuda:0')

[nuxxor]

@HDANILO is right it's not working on RTX 5090.

[jkpjkpjkp]

is there updates on this? after following the hopper/ installation guide tests failed with CUDA error (/home/jkp/g/flash-attention/hopper/flash_fwd_launch_template.h:192): no kernel image is available for execution on the device rtx5080

[CristyNel]

I spent 3 hours building FA3, only to realize it doesn't work on the 5090 (sm120).

[alecco]

It's already supported

AFAIK the MLIR compiler (proprietary) doesn't support [some] SM120/SM121 (RTX 50XX/6000, DGX Spark) so CuTeDSL can't target them.

Edit: I was being too aggressive in my approach, perhaps it's doable without anything fancy.

[alecco]

I'm working on native support for SM120 (RTX 50/GeForce) with proper TMA. But it's going to be a huge PR (40 commits and nowhere near done). And that's without anything fancy like FP8.

I'm already concerned for the poor soul who dares taking a shot at reviewing it. (cc @tridao)

[tridao]

there might be an ongoing PR doing this: #2406