[ROCm] [Hardware][AMD] Remove xformer patches and ray issue fix

[hongxiayang]

Thanks for previous pull requests (#3005) which can support flash attention without xformers.

This PR is to clean up xformer dependencies from ROCm path:

• Removed xformer patches.
• Ray fix for ROCm. need to pin the version 2.9.3 as the latest version breaks for ROCm (and my patch to vllm worker code to fix this ray issue was too ugly to put in the pull request).

This gives a noticeable throughput bump after running the benchmark script on MI250x.

BEFORE SUBMITTING, PLEASE READ THE CHECKLIST BELOW AND FILL IN THE DESCRIPTION ABOVE

---

PR Checklist (Click to Expand)

Thank you for your contribution to vLLM! Before submitting the pull request, please ensure the PR meets the following criteria. This helps vLLM maintain the code quality and improve the efficiency of the review process.

PR Title and Classification

Only specific types of PRs will be reviewed. The PR title is prefixed appropriately to indicate the type of change. Please use one of the following:

• [Bugfix] for bug fixes.
• [CI/Build] for build or continuous integration improvements.
• [Doc] for documentation fixes and improvements.
• [Model] for adding a new model or improving an existing model. Model name should appear in the title.
• [Frontend] For changes on the vLLM frontend (e.g., OpenAI API server, LLM class, etc.)
• [Kernel] for changes affecting CUDA kernels or other compute kernels.
• [Core] for changes in the core vLLM logic (e.g., LLMEngine, AsyncLLMEngine, Scheduler, etc.)
• [Hardware][Vendor] for hardware-specific changes. Vendor name should appear in the prefix (e.g., [Hardware][AMD]).
• [Misc] for PRs that do not fit the above categories. Please use this sparingly.

Note: If the PR spans more than one category, please include all relevant prefixes.

Code Quality

The PR need to meet the following code quality standards:

• We adhere to Google Python style guide and Google C++ style guide.
• Pass all linter checks. Please use format.sh to format your code.
• The code need to be well-documented to ensure future contributors can easily understand the code.
• Include sufficient tests to ensure the project to stay correct and robust. This includes both unit tests and integration tests.
• Please add documentation to docs/source/ if the PR modifies the user-facing behaviors of vLLM. It helps vLLM user understand and utilize the new features or changes.

Notes for Large Changes

Please keep the changes as concise as possible. For major architectural changes (>500 LOC excluding kernel/data/config/test), we would expect a GitHub issue (RFC) discussing the technical design and justification. Otherwise, we will tag it with rfc-required and might not go through the PR.

What to Expect for the Reviews

The goal of the vLLM team is to be a transparent reviewing machine. We would like to make the review process transparent and efficient and make sure no contributor feel confused or frustrated. However, the vLLM team is small, so we need to prioritize some PRs over others. Here is what you can expect from the review process:

• After the PR is submitted, the PR will be assigned to a reviewer. Every reviewer will pick up the PRs based on their expertise and availability.
• After the PR is assigned, the reviewer will provide status update every 2-3 days. If the PR is not reviewed within 7 days, please feel free to ping the reviewer or the vLLM team.
• After the review, the reviewer will put an action-required label on the PR if there are changes required. The contributor should address the comments and ping the reviewer to re-review the PR.
• Please respond to all comments within a reasonable time frame. If a comment isn't clear or you disagree with a suggestion, feel free to ask for clarification or discuss the suggestion.

Thank You

Finally, thank you for taking the time to read these guidelines and for your interest in contributing to vLLM. Your contributions make vLLM a great tool for everyone!

[WoosukKwon]

@hongxiayang Thanks for the PR! Removing the ROCm patch is really great!

I have two concerns/questions on this PR:

1. I recently heard that xFormers now officially supports some AMD GPUs, and it will support head sizes (e.g., 256) that are not supported by the ROCm flash attn. Is there no performance difference between the ROCm xFormers and flash-attn, can we just use the ROCm xFormers (without the patch)?
2. IIRC, we use naive PyTorch implementation of attention for some old AMD GPUs that are not supported by ROCm flash-attn. You can find this code in

   vllm/vllm/model_executor/layers/attention/backends/xformers.py

   Line 59 in f721096

   self.use_ref_attention = _check_use_ref_attention()

   
   I think this PR will break the support for these old GPUs.

[hongxiayang]

@hongxiayang Thanks for the PR! Removing the ROCm patch is really great!

I have two concerns/questions on this PR:

1. I recently heard that xFormers now officially supports some AMD GPUs, and it will support head sizes (e.g., 256) that are not supported by the ROCm flash attn. Is there no performance difference between the ROCm xFormers and flash-attn, can we just use the ROCm xFormers (without the patch)?

2. IIRC, we use naive PyTorch implementation of attention for some old AMD GPUs that are not supported by ROCm flash-attn. You can find this code in

   vllm/vllm/model_executor/layers/attention/backends/xformers.py

   Line 59 in f721096

   self.use_ref_attention = _check_use_ref_attention()

   I think this PR will break the support for these old GPUs.

@WoosukKwon Thank you for your comment and review.

For 1: I need to check whether we can integrate with xformers directly without patch, and its performance.

For 2: It should not break. See the note in my previous pull request to enable the naive pytorch attention for those GPUs where flash-attention support is not good: #2768. On those GPUs (not old GPUs, by the way, Radeon™ 7900 series (gfx1100) is a kind of relatively new Consumer AI GPUs ). When building the docker for this arch target, the user is suggested not to build/install flash-attention at all, so the check for self.use_ref_attention will return True.

[hongxiayang]

Put the un-versioned xformer package in Dockerfile.rocm. Regarding the naive pytorch attention, it looks like ref-masked -attention is part of xformer backend now in the code (that is why you mentioned it might break). Since it is a pure pytorch based attention, should it not rely on any of the flash-attn, or xformers backend? Should we refactor the backend code?

[WoosukKwon]

For 1: I need to check whether we can integrate with xformers directly without patch, and its performance.

@hongxiayang Have you got a chance to look into this? I believe the latest version (v0.0.25) of xFormers supports AMD GPUs. If this performs well, I think we should use xFormers instead of the ROCm flash attn, given that xFormers will be the superset of the flash attn (unlike the case for NVIDIA GPUs).

[hongxiayang]

For 1: I need to check whether we can integrate with xformers directly without patch, and its performance.

@hongxiayang Have you got a chance to look into this? I believe the latest version (v0.0.25) of xFormers supports AMD GPUs. If this performs well, I think we should use xFormers instead of the ROCm flash attn, given that xFormers will be the superset of the flash attn (unlike the case for NVIDIA GPUs).

I checked. The xformers v0.0.25 does not show flash-attn available:

/app# python -m xformers.info
WARNING[XFORMERS]: xFormers can't load C++/CUDA extensions. xFormers was built for:
    PyTorch 2.2.1+cu121 with CUDA 1201 (you have 2.1.1+git011de5c)
    Python  3.9.18 (you have 3.9.18)
  Please reinstall xformers (see https://github.com/facebookresearch/xformers#installing-xformers)
  Memory-efficient attention, SwiGLU, sparse and more won't be available.
  Set XFORMERS_MORE_DETAILS=1 for more details
/opt/conda/envs/py_3.9/lib/python3.9/site-packages/torch/cuda/__init__.py:611: UserWarning: Can't initialize NVML
  warnings.warn("Can't initialize NVML")
Unable to find python bindings at /usr/local/dcgm/bindings/python3. No data will be captured.
xFormers 0.0.25
memory_efficient_attention.ckF:                    unavailable
memory_efficient_attention.ckB:                    unavailable
memory_efficient_attention.ck_decoderF:            unavailable
memory_efficient_attention.ck_splitKF:             unavailable
memory_efficient_attention.cutlassF:               unavailable
memory_efficient_attention.cutlassB:               unavailable
memory_efficient_attention.decoderF:               unavailable
memory_efficient_attention.flshattF@2.0.4:         unavailable
memory_efficient_attention.flshattB@2.0.4:         unavailable

As opposed to the previous version before, which shows it as available since that was patched.

# python -m xformers.info
WARNING[XFORMERS]: xFormers can't load C++/CUDA extensions. xFormers was built for:
    PyTorch 2.1.1+cu121 with CUDA 1201 (you have 2.1.1+git011de5c)
    Python  3.9.18 (you have 3.9.18)
  Please reinstall xformers (see https://github.com/facebookresearch/xformers#installing-xformers)
  Memory-efficient attention, SwiGLU, sparse and more won't be available.
  Set XFORMERS_MORE_DETAILS=1 for more details
xFormers 0.0.23
/opt/conda/envs/py_3.9/lib/python3.9/site-packages/torch/cuda/__init__.py:611: UserWarning: Can't initialize NVML
  warnings.warn("Can't initialize NVML")
memory_efficient_attention.cutlassF:               unavailable
memory_efficient_attention.cutlassB:               unavailable
memory_efficient_attention.decoderF:               unavailable
memory_efficient_attention.flshattF@2.0.4:         available

[hongxiayang]

close this pull request as most of changes are available in #3643