MI100 Support

[LoggerHead22]

Hi, the documentation says that this implementation is compatible only with the MI200 and MI300 GPUs. But what about the MI100 gpu?

The code contains such conditions that formally match the MI100 with the gfx908 architecture.

bool is_gfx90x = dprops->major == 9 && dprops->minor == 0;
bool is_gfx94x = dprops->major == 9 && dprops->minor == 4;
TORCH_CHECK(is_gfx90x || is_gfx94x, "FlashAttention only supports AMD MI200 GPUs or newer.");

Will this code be compatible with MI100 in practice? If not, are there any plans to add such support? Or what are the reasons that keep you from adding support for the MI100?

[dejay-vu]

Hi @LoggerHead22, this code appears to be a logic fault, thanks for noting.

We haven't tested the FA on MI100 since we did most of our testing on MI250&MI300 so we are limiting the support archs. I am not sure whether it will work correctly on MI100 but you can try by adding gfx908 to the valid archs. I suppose the building process will be fine.

[LoggerHead22]

Thanks for the clarification @howiejayz . Your advice really helped, the code is compiled for mi100 and runs.

However, I encountered an error during the build, which is caused by the logic of the patch.

hipified_header_filepath = HIPIFY_FINAL_RESULT[header_filepath].hipified_path
      AttributeError: 'dict' object has no attribute 'hipified_path'

This seems logical, because a dict is being created here and then we try to take its hipified_path attribute.

Replacing dict with an object of the HipifyResult class in patch helped me.

[ccbadd]

Has this patch been merged to the main branch or do we need to apply it in order to test?

[ehartford]

I need mi100 support

[ehartford]

Hi @LoggerHead22, this code appears to be a logic fault, thanks for noting.

We haven't tested the FA on MI100 since we did most of our testing on MI250&MI300 so we are limiting the support archs. I am not sure whether it will work correctly on MI100 but you can try by adding gfx908 to the valid archs. I suppose the building process will be fine.

If you need hardware for testing mi100, I volunteer my server for this purpose. I have 8x mi100 with infinity fabric.

ehartford@gmail.com

[ehartford]

Hi @sabreshao @howiejayz can you please give me a path forward?

I have a bunch of mi100s and I would like them to be hot. Without flash attention, I am blocked.

Maybe you could show me where in the code I would add it? give me some advice?

[dejay-vu]

Hi @LoggerHead22, this code appears to be a logic fault, thanks for noting.

We haven't tested the FA on MI100 since we did most of our testing on MI250&MI300 so we are limiting the support archs. I am not sure whether it will work correctly on MI100 but you can try by adding gfx908 to the valid archs. I suppose the building process will be fine.

Hi @ehartford! Currently I have no time to test FA on MI100 but could you try build and run based on this comment?

[TNT3530]

I was able to compile flash attention for the MI100 using the docker image. Simply adding gfx908 to the target arch array (or in my case, removing everything BUT native and gfx908) makes it run fine. (Note: this also applies to the vLLM ROCm docker image, which was my use case)

Attempts to compile outside of docker seem to fail on ROCm 6.0 due to this issue, though I was unable to downgrade back to 5.7 to test on my machine.

[luizanao]

I managed to build MI100 (gfx908) as well but the env var didn't work @TNT3530 . This is because the setup is protected against unknown architectures and gfx908 is not listed. I will open a PR for adding that since gfx908 definitely works.

[luizanao]

Here's my PR, you folks might benefit from it: #38

[ehartford]

How do I install flash attention for mi100? How is the procedure from the README.md different?

[luizanao]

@ehartford passing the card arch to the build should be enough: export GPU_ARCHS="gfx908"

[gittb]

Also curious if support for Mi100 was finalized.

[ehartford]

This is awesome! Can't wait to try it!