Qwen3.5 seems have problem with tp>1 in triton_attention

[gaoxt1983]

I'm using Qwen3.5 with 2 L40-48G. the configuration is like:

{
"log_requests": false,
"enable_cache_report": true,
"mem_fraction_static": 0.85,
"max_running_requests": 64,
"max_prefill_tokens": 8192,
"chunked_prefill_size": 4096,
"tp_size": 2,
"show_time_cost": true,
"enable_metrics": true,
"tool_call_parser": "qwen3_coder",
"trust_remote_code": true,
"reasoning_parser": "qwen3",
"nnodes": 1,
"served_model_name": "duokabushu2"
}

But when inference, the generated text is very short and seems unfinished. I try to read the log which sglang generated. found that there is only prefill phrase, no decode phrase. Is there something wrong with my configuration or my image?