vLLM pads DeepSeek-V4.1-Flash's 2304-wide shared expert to 2560 so it fuses with MegaMoE in all 40 layers
GitHub·medium signal
PR #56568 resolves issue #56217: DeepSeek-V4.1-Flash's shared expert is 2304 wide while MegaMoE's routed experts are padded to 2560, which blocked fusion. The fix pads the shared weights with zeros and unit scales while preserving checkpoint-shaped parameters and the unsupported-layout fallback, and it ships behind VLLM_DISABLE_DSV4_MEGAMOE_SHARED_EXPERT_FUSION so operators can A/B the two paths with a fresh server per arm. The PR discloses AI assistance for implementation, testing and documentation.