[Bug]: Qwen3.5-4B on NPU returns "!" as the first generated token on xLLM 0.9.0

[xpluspro]

Your environment

Environment:

• Hardware: 910B3
• CANN: 8.5
• Docker: official xLLM docker environment, consistent with the provided setup
• xLLM version: 0.9.0

🐛 Describe the bug

According to RELEASE.md, xLLM 0.9.0 supports Qwen3.5/Qwen3.5-MoE on NPU.

However, in my environment, Qwen3.5-4B consistently generates "!" as the first output token, while Qwen3 works correctly with the same serving setup and request pattern.

Model config highlights:

• architectures: Qwen3_5ForConditionalGeneration
• model_type: qwen3_5
• text_config.model_type: qwen3_5_text
• full_attention_interval: 4
• layer_types present (hybrid linear/full attention)

Minimal reproduction:

curl http://localhost:9977/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Qwen3.5-4B",
    "messages": [{"role":"user","content":"你好"}],
    "temperature": 0,
    "max_tokens": 1
  }'

Observed result:

{"id":"chatcmpl-10922618569676071768-AbnVBNKbWvYDFz7FCZ89Df","object":"chat.completion","created":1775279394,"model":"Qwen3.5-4B","choices":[{"index":0,"message":{"role":"assistant","content":"!"},"tool_calls":[],"finish_reason":"length"}],....

Additional notes:

• Qwen3 works correctly under the same command pattern.
• I also tested code including the Qwen3.5 support chain up to #1160, but the issue remained.
• I checked #1171, but it is still draft and not yet rebased after #1160.

Questions:

1. Is the public checkpoint Qwen/Qwen3.5-4B expected to work on NPU with xLLM 0.9.0?
2. Is there a rebased patch after #1160 / #1171 for this issue?
3. Is there any required model conversion or special launch flag for Qwen3.5 text-only inference on xLLM?

Thanks.