[megatron] Convert fused-MoE expert LoRA to vLLM layout on adapter sync#1857
Merged
erictang000 merged 2 commits intoJul 2, 2026
Merged
Conversation
Megatron-Bridge exports fused-MoE expert LoRA as 3D (num_experts-leading) tensors keyed `...mlp.experts.gate_up_proj` / `...mlp.experts.down_proj`, but vLLM's fused-MoE LoRA loader (FusedMoE3DWithLoRA / _stack_moe_lora_weights) expects the flat PEFT layout keyed `...mlp.experts.base_layer` (w13) and `...mlp.experts` (w2). Without the rewrite, merge_lora=False on-policy sync is rejected because load_lora_adapter cannot resolve `experts.down_proj`. Add `_convert_moe_experts_lora_to_vllm` to megatron_utils and apply it in MegatronPolicyWorkerBase._save_lora_adapters_and_sync before inferring target modules (dropping the synthetic `base_layer` module).
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Code Review
This pull request introduces a helper function _convert_moe_experts_lora_to_vllm to convert fused-MoE expert LoRA tensors from Megatron-Bridge's 3D layout to the flat PEFT layout expected by vLLM, ensuring compatibility during on-policy synchronization. Feedback suggests unpacking the tensor shapes explicitly before reshaping or permuting to improve code readability, prevent potential bugs during future refactoring, and avoid relying on tensor.shape after partial reassignments.
Condense the docstrings and inline comments added in this PR while preserving their technical content. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes #1852 @erictang000
Megatron-Bridge exports fused-MoE expert LoRA as 3D (num_experts-leading) tensors keyed
...mlp.experts.gate_up_proj/...mlp.experts.down_proj, but vLLM's fused-MoE LoRA loader (FusedMoE3DWithLoRA / _stack_moe_lora_weights) expects the flat PEFT layout keyed...mlp.experts.base_layer(w13) and...mlp.experts(w2). Without the rewrite, merge_lora=False on-policy sync is rejected because load_lora_adapter cannot resolveexperts.down_proj.Add
_convert_moe_experts_lora_to_vllmto megatron_utils and apply it in MegatronPolicyWorkerBase._save_lora_adapters_and_sync before inferring target modules (dropping the syntheticbase_layermodule).