Conversation
|
Hi, I am the author of #2230. The Triton backend works on RDNA. We merged RDNA CI that runs flash attention tests on gfx1100, see ROCm/aiter#2222 and https://github.com/ROCm/aiter/actions/runs/23300077538/job/67758875394. Can you share your setup so I can reproduce the issue? |
|
Regarding windows support, Can you try this pr #2384? |
I can try that pr (or yours with your changes if available) and the corresponding one in aiter, in around an hour. I did not expect such a quick turn around, or @0xDELUXA to do it, as I had made multiple different attempts with aiter myself over the past several weekends. For my setup I have gfx1150 with 128GB ram on Win11, using TheRock nightly builds. |
|
@astrelsky I have discussed with @0xDELUXA and I'm taking ownership of getting this through. I am working on PRs based on @0xDELUXA's work. I'll ping here when it's ready to test. |
|
@astrelsky I have put up a pr here, #2385 to fix the issue. I have also add a windows smoke test to catch issues like this. See https://github.com/ROCm/aiter/actions/runs/23464138688/job/68272544832 . We can continue discussion there. |
Great thank you. I won't be able to test until tonight. I'll close this in the meantime since it seems like a fix will be a lot sooner than expected. |
Migrating to aiter results in loss of support for ALL AMD GPUs except gfx942 and gfx950. ROCm/aiter#1943 (comment)
While their own codebase contradicts their statement, it appears they have absolutely no intent on supporting anything other than gfx942 and gfx950. The use of aiter also completely removes the ability to use flash-attention on Windows, since aiter does not support Windows. ROCm/aiter#1943
Reverts #2230
Of course long term just reverting isn't great, but it's a huge loss of compatibility otherwise. I'm hoping that by reverting this and maybe someone from Dao-AILab raising the issue, maybe some sense will be knocked into the aiter maintainers.