[Bug]: [ROCm] Qwen4Exp (AMD path): PLE n-gram IDs read a stale query_start_loc tail during PIECEWISE CUDA-graph replay
![[Bug]: [ROCm] Qwen4Exp (AMD path): PLE n-gram IDs read a stale query_start_loc tail during PIECEWISE CUDA-graph replay](https://www.chat-gpts.plus/wp-content/uploads/2026/11/61486-25e2f379-768x403.jpg)
该现象通常出现在 AMD ROCm 后端运行 Qwen4Exp(Qwen3.8-Flash-Next)并使用 PIECEWISE CUDA graph 的场景;当短批次重放时,PLE 的 n-gram ID 会读取到 query_start_loc 中上一次留下的旧尾部数据。应优先确认是否为该竞态,
![[Bug]: V1 thinking budget corrupts a multi-token reasoning_end_str under speculative decoding](https://www.chat-gpts.plus/wp-content/uploads/2026/11/58485-85e0f342-768x403.jpg)



![Misc. bug: [SYCL] CPY to q4_1/q5_1/q8_0 doesn't match CPU reference](https://www.chat-gpts.plus/wp-content/uploads/2026/11/31140-0ca3be80-768x403.jpg)



