Note

Low Risk
Scoped to vLLM LoRA profiling/warmup in inference patches; no changes to training, auth, or persisted data paths.

Overview
Adds a vLLM inference monkey patch so dummy LoRA warmup during profiling no longer crashes on FusedMoEWithLoRA (and related MoE LoRA layers) under expert parallelism.

The new monkey_patch_fused_moe_dummy_lora replaces LoRAModelManager.create_dummy_lora and is registered from transformers_v5_compat alongside the other vLLM patches. For FusedMoEWithLoRA, it limits packed “replacement” slices to the local EP experts and sets n_slices from that narrowed list so dummy weight creation does not index past module.lora_a_stacked (vLLM 0.21.0 bug). It also keeps explicit handling for FusedMoE3DWithLoRA and MoE PackedLoRALayerWeights packing.

^{Reviewed by Cursor Bugbot for commit e6010e3. Bugbot is set up for automated code reviews on this repo. Configure here.}