[5/7][multi-lora]: rollout request routing - per sample lora_path/rid, per-adapter rewards, prefill logprob grouping#1746
Conversation
…ml config, shared helpers, CLI flags and validation
…d, control-plane HTTP API, named Ray actor
…schedules, slot lifecycle, batch routing in get_batch
…ert push to engines, reconcile/train/save hooks
… per-adapter rewards, prefill logprob grouping
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request introduces multi-LoRA support to enable concurrent training and serving of multiple LoRA adapters in Megatron-LM, adding a controller, registry, HTTP server, per-slot Adam optimizers, and per-adapter schedulers. The code review highlights several critical improvements: fixing potential distributed deadlocks during weight synchronization by ensuring proper collective broadcast participation across ranks, preventing post-save hook failures in multi-LoRA mode, defensively handling standard optimizer wrappers to avoid AttributeError, optimizing Ray actor calls during adapter cleanup via broadcasting, supporting cloud storage URIs in configuration validation, and guarding against IndexError when adapter_slots is empty.
Important
The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.
yushengsu-thu
left a comment
yushengsu-thu
left a comment
There was a problem hiding this comment.
Reviewed extensively on the umbrella PR #1638 (2xH200 E2E validated); layer verified independently (imports + fast suites green on this branch's own tree).
2 participants
#1638