[ROCm] [AITER] [Bugfix] Patch for AITER commit 648764942e552a8bb5fe16026703716a81f05374

[tjtanaa]

This PR is a bugfix after PR (#18596). ONLY Merge this after PR (#18596).

This PR also include upgrading the AITER commit of the Dockerfile.

PR (#18596) introduced the use of AITER MHA which depends on a new AITER commit 648764942e552a8bb5fe16026703716a81f05374.

AITER commit: ROCm/aiter@a02a93d has introduced a new enum value in a breaking changes manner.

lm_eval after fix

Qwen/Qwen3-235B-A22B-FP8

vllm (pretrained=Qwen/Qwen3-235B-A22B-FP8,tensor_parallel_size=2,max_model_len=10000,trust_remote_code=True), gen_kwargs: (None), limit: None, num_fewshot: 5, batch_size: auto
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.8613|±  |0.0095|
|     |       |strict-match    |     5|exact_match|↑  |0.8408|±  |0.0101|

mistralai/Mixtral-8x7B-Instruct-v0.1

vllm (pretrained=mistralai/Mixtral-8x7B-Instruct-v0.1,tensor_parallel_size=2,max_model_len=10000,trust_remote_code=True), gen_kwargs: (None), limit: None, num_fewshot: 5, batch_size: auto
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.6399|±  |0.0132|
|     |       |strict-match    |     5|exact_match|↑  |0.6361|±  |0.0133|

mistralai/Mixtral-8x7B-Instruct-v0.1 dynamic fp8 quantization

vllm (pretrained=mistralai/Mixtral-8x7B-Instruct-v0.1,tensor_parallel_size=2,quantization=fp8,max_model_len=10000,trust_remote_code=True), gen_kwargs: (None), limit: None, num_fewshot: 5, batch_size: auto
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.5777|±  |0.0136|
|     |       |strict-match    |     5|exact_match|↑  |0.5754|±  |0.0136|

deepseek-ai/DeepSeek-V3

vllm (pretrained=deepseek-ai/DeepSeek-V3,tensor_parallel_size=8,block_size=1,trust_remote_code=True), gen_kwargs: (None), limit: None, num_fewshot: 5, batch_size: auto
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.9454|±  |0.0063|
|     |       |strict-match    |     5|exact_match|↑  |0.9454|±  |0.0063|

Note:
Even if the AITER commit: 648764942e552a8bb5fe16026703716a81f05374 introduced a new input argument min_seqlen_q to flash_attn_varlen_func. It seems that the default value is set to 0 which retains compatibility with how the MHA is used in the ROCm MLA v1 class. Refer to https://github.com/ROCm/aiter/blob/2c1a21adad9c5b5e02619c7dd05d63f9afda3642/aiter/ops/mha.py#L1369

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[houseroad]

cc: @842974287

[tjtanaa]

@houseroad
The PR #18596 has been merged. Could we try to proceed with this PR to make fixes so that AITER fused moe kernel could run with this new AITER commit?