Deterministic prefill crash at a fixed token position on qwen4exp (Qwen3.8-Flash-Next) under multi-GPU layer split — 4 runtimes, FA on/off,

该问题出现在 llama.cpp / LM Studio 的 CUDA runtime 对 qwen4exp(Qwen3.8-Flash-Next)模型做长上下文 prefill 时,multi-GPU layer split( --split-mode layer )下会在固定 token 位置确
![[Bug]: Responses API streaming regenerates output item id/call_id in response.completed (breaks strict clients)](https://www.chat-gpts.plus/wp-content/uploads/2026/10/59834-4a72fa90-768x403.jpg)


![RuntimeError: Worker failed with error 'MLA kv_data_type torch.uint8 is not supported. Supported dtypes: [torch.float16, torch.bfloat16, torch.float8_e4m3fn].', please check the stack trace abov](https://www.chat-gpts.plus/wp-content/uploads/2026/10/57713-741c6632-768x403.jpg)

![[Performance]: Non-spec Qwen3.5 CUDA GDN wrapper fallback regressed H200 throughput (fixed by #59735)](https://www.chat-gpts.plus/wp-content/uploads/2026/10/59520-c6b19167-768x403.jpg)
![[Model Support] Kimi K3 Tracking Issue](https://www.chat-gpts.plus/wp-content/uploads/2026/10/50001-de375f68-768x403.jpg)

