[Bug]: AiterFlashAttentionImpl.__init__() got multiple values for argument 'use_irope' for llama4 model

[hongxiayang]

Your current environment

The output of python collect_env.py

Your output of `python collect_env.py` here

🐛 Describe the bug

We hit an exception on running llama4 models with latest code on ROCm V1:

(VllmWorker rank=2 pid=267) ERROR 06-19 01:00:39 [multiproc_executor.py:488] TypeError: AiterFlashAttentionImpl.__init__() got multiple values for argument 'use_irope'

Current work-around:
To turn off AITER_MHA, with VLLM_ROCM_USE_AITER_MHA=0

Proposal:

• Fix the bug (the team is working on it)
• Add a end-to-end test for one of the small llama4 models
• [ ]

The motivation for adding an end to end test for a small version of llama4 models, is that we have seen issues of breaking llama4 models in the past because of lacking such tests.

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[tjtanaa]

This PR #18212 introduced a new argument to AttentionImpl init function (kv_sharing_target_layer_name: Optional[str] = None,)
Bug fix is undergoing.

The MHA backend has to fallback to SDPA on rocm platform

vllm/vllm/attention/layer.py

Line 310 in 01220ce

backend = _Backend.XFORMERS