[Bug]: [DeepSeek-V4-Flash][MTP] CUDA invalid argument during profile_run with DP4 + EP + MegaMoE on B_00
![[Bug]: [DeepSeek-V4-Flash][MTP] CUDA invalid argument during profile_run with DP4 + EP + MegaMoE on B_00](https://www.chat-gpts.plus/wp-content/uploads/2026/06/45099-8bfd8616-768x403.jpg)
在 vllm serve 启动 DeepSeek-V4-Flash 模型时,使用 --speculative-config '{"method":"mtp","num_speculative_tokens":1}' 或 num_speculative_tokens=3 ,并配合 --data-par




![[Bug]: Prefix-read and no-prefix-read paths can yield different greedy answers for the same prompt](https://www.chat-gpts.plus/wp-content/uploads/2026/06/42699-ab6cf9ef-768x403.jpg)
![[Bug]: GLM-5(Sparse MLA / DSA 模型)无法在 sm80 GPU(A100/A800)上运行 — DeepGemm 硬依赖无 fallback](https://www.chat-gpts.plus/wp-content/uploads/2026/06/35021-5544c92b-768x403.jpg)


