[Bug] Qwen4Exp QSA indexer: per-chunk logits buffer grows with max_seq_len, caching allocator keeps every size, device OOM/hang on unified-m
![[Bug] Qwen4Exp QSA indexer: per-chunk logits buffer grows with max_seq_len, caching allocator keeps every size, device OOM/hang on unified-m](https://www.chat-gpts.plus/wp-content/uploads/2026/09/56457-d01dde83-768x403.jpg)
该报错通常出现在 GB10(SM121)、unified memory 机器上用 vLLM 跑 Qwen4Exp 架构模型做超长 prefill(chunked prefill)时,随着 chunk 推进,QSA indexer 的 per-chunk logits buffer 尺寸逐块增大,Py
![[Bug]: Batch invariance is broken when sequence parallelism / async TP is enabled (`VLLM_BATCH_INVARIANT=1` + `pass_config.enable_sp`)](https://www.chat-gpts.plus/wp-content/uploads/2026/09/56370-536c0b3f-768x403.jpg)
![[Bug] V1 InputBatch condense can leak stale allowed_token_ids mask to recycled row](https://www.chat-gpts.plus/wp-content/uploads/2026/09/43894-a55fb008-768x403.jpg)





