vLLM backend: TypeError: MMEncoderAttention.__init__() got unexpected keyword argument 'multimodal_config'

[nchudleigh]

Bug Description

The vLLM backend fails with a TypeError when initializing MMEncoderAttention because it passes a multimodal_config argument that doesn't exist in vLLM's API.

Error Message

TypeError: MMEncoderAttention.__init__() got an unexpected keyword argument 'multimodal_config'

Stack Trace

File "/usr/local/lib/python3.12/site-packages/qwen_asr/core/vllm_backend/qwen3_asr.py", line 197, in __init__
    self.attn = MMEncoderAttention(
                ^^^^^^^^^^^^^^^^^^^
TypeError: MMEncoderAttention.__init__() got an unexpected keyword argument 'multimodal_config'

Root Cause

In qwen_asr/core/vllm_backend/qwen3_asr.py line 197, the code passes multimodal_config=multimodal_config to MMEncoderAttention:

self.attn = MMEncoderAttention(
    num_heads=self.num_local_heads,
    head_size=self.head_dim,
    scale=self.scaling,
    multimodal_config=multimodal_config,  # <-- This argument doesn't exist
)

However, vLLM's MMEncoderAttention.__init__() only accepts:

• num_heads
• head_size
• scale
• num_kv_heads
• prefix

The multimodal_config parameter was never part of the vLLM API for this class.

Environment

• qwen-asr: 0.0.6
• vllm: 0.14.0 (as specified in qwen-asr's pyproject.toml)
• Python: 3.12
• Platform: Modal (nvidia/cuda:12.1.0-devel-ubuntu22.04)

Suggested Fix

Remove the multimodal_config argument from the MMEncoderAttention call in qwen_asr/core/vllm_backend/qwen3_asr.py.

[eschmidbauer]

This branch may help you