Qwen3.5-35B-A3B/Qwen3.6-35B-A3B sym_int4 TP=4: runaway-! degeneration returns on intel/llm-scaler-vllm:0.26.0-b1, on 4x Arc Pro B70

error hit both onQwen3.6-35B-A3B.
image:intel/llm-scaler-vllm:0.26.0-b1
MACHINE: xeon-w.
gpu: 4x B70.
OS: edgeai@edgeaihost15:$ uname -a
Linux edgeaihost15 6.17.0-1009-intel https://github.com/intel/llm-scaler/pull/9-Ubuntu SMP PREEMPT_DYNAMIC Mon Apr 6 17:05:13 UTC 2026 x86_64 x86_64 x86_64 GNU/Linux
edgeai@edgeaihost15:$

**VLLM SERVER CMD:**
root@edgeaihost15:/llm-scaler/vllm# TORCH_LLM_ALLREDUCE=1 VLLM_USE_V1=1 W_LONG_MAX_MODEL_LEN=1 VLLM_WORKER_MULTIPROC_METHOD=spawn python3 -m vllm.entrypoints.openai.api_server --model /llm/models/Qwen3.6-35B-A3B --dtype=float16 --enforce-eager --port 8000 --block-size 64 --gpu-memory-util 0.9 --trust-remote-code --max_num_batched_tokens=4096 --max_model_len 50000 --quantization sym_int4 -tp=4

REQUEST CMD:
root@edgeaihost15:/llm-scaler/vllm# curl [http://127.0.01:8000/v1/completions](http://127.0.0.1:8000/v1/completions) -H "Content-Type: application/json" -d '{ "model": "/llm/models/Qwen3.6-35B-A3B", "prompt": "give the info about shanghai city and deliver one ppt FILE", "max_tokens": 2000, "temperature": 0}'

result:

Start by reproducing the issue with the listed intel/llm-scaler-vllm:0.26.0-b1 image, four Arc Pro B70 GPUs, and the vllm.entrypoints.openai.api_server command, then send the shown curl request. Compare the generated output with expected completion behavior and trace the relevant vLLM inference path. Done means the runaway-! degeneration no longer occurs for the specified Qwen3.5/Qwen3.6 sym_int4 TP=4 setup.

Written by the indexing model from the issue text.