[Bug]: OffloadingConnector stores but never serves when MTP/EAGLE speculative decoding is enabled (hybrid GDN model, XPU)
![[Bug]: OffloadingConnector stores but never serves when MTP/EAGLE speculative decoding is enabled (hybrid GDN model, XPU)](https://www.chat-gpts.plus/wp-content/uploads/2026/09/52735-70676ec7-768x403.jpg)
该报错发生在 vLLM XPU 环境启用 EAGLE/MTP 推测解码并配合混合 GDN(门控延迟网络)模型使用 OffloadingConnector 时,表现为 KV 缓存只存不取、检索命中率为 0。优先排查 `offloading/scheduler.py` 中 `is_eagle_group



![[Bug] Trellis2UpsampleStage hardcodes lr_resolution = 512; breaks when VaeDecodeStructureTrellis2 resolution is 64 (coordinates overflow by](https://www.chat-gpts.plus/wp-content/uploads/2026/09/16100-35ccd589-768x403.jpg)


![[Go] Ollama: multimodal/vision requests drop images (no vision support)](https://www.chat-gpts.plus/wp-content/uploads/2026/09/17332-b01d4185-768x403.jpg)

