Fused Moe Config Lifecycle
Avoid false missing-config warnings in modular fused-MoE profiling and forward passes.
Profiling a configured DP+EP fused-MoE layer through the normal factory path emits Current vLLM config is not set., and the same warning can appear during forward even though the layer has a usable parallel configuration. Fix construction, profiling, and forward to use the active layer's effective configuration, including explicit parallel-size overrides, when the temporary global context is absent or belongs to another layer. Configured layers should not emit this warning, but accesses with genuinely missing configuration must still warn. Fix the configuration handling without suppressing, downgrading, filtering, or globally monkeypatching the diagnostic.
Preserve the real Triton MoE forward behavior and the workspace requirements of each layer. DP+EP profiling must retain the backend's worst-case reservation of 16,384 routing tokens even for a small local batch; an ordinary layer must retain its ordinary workspace requirements under a conflicting DP+EP context. Check these behaviors through the normal factory, profiling, and forward lifecycle as the context changes. How configuration is passed and stored, the internal workspace layout, and the number of workspace queries are up to you.