Spec-decode on pre-Hopper runs without CUDA graphs on FlashInfer (prefill path, no capture) — is a decode-path spec mask feasible for fa2?

该问题发生在使用 FlashInfer 后端与 speculative decoding 组合时,且运行在 Ampere(如 RTX 3090)等 Pre-Hopper 架构之上。优先排查 vLLM 版本是否为已知重复 Issue #49547 的受影响版本,并建议迁移至 FlashAttentio




![[BUG]: AnythingLLM does not provide write access to second directory](https://www.chat-gpts.plus/wp-content/uploads/2026/09/6090-414b2d90-768x403.jpg)
![[BUG]: Agent clarifying questions does not work while suing the desktop assistant.](https://www.chat-gpts.plus/wp-content/uploads/2026/09/6033-99b743cb-768x403.jpg)


