[skyrl-train] Add example for 235B LoRA training with Megatron on 4 H100 nodes#1000
[skyrl-train] Add example for 235B LoRA training with Megatron on 4 H100 nodes#1000erictang000 merged 6 commits intoNovaSky-AI:mainfrom
Conversation
![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 a new example script for 235B LoRA training with Megatron on 4 H100 nodes using DAPO. It also updates an existing Megatron training script with a new configuration and adds LoRA parameters. Additionally, it enhances the evaluation process by logging more detailed rollout metrics.
My review focuses on improving the clarity and correctness of the new and modified shell scripts. I've pointed out a redundant variable definition and inconsistencies between comments and code in the configuration parameters. These changes will help improve the maintainability and readability of the scripts.
| # Qwen3-235B-A22B has 94 blocks, so we need to set the last pipeline stage layer to use 4 blocks | ||
| MEGATRON_LAST_PIPELINE_STAGE_LAYER=16 |
There was a problem hiding this comment.
The comment on line 30 states that the last pipeline stage should use 4 blocks, but the value of MEGATRON_LAST_PIPELINE_STAGE_LAYER is set to 16 on line 31. This is inconsistent. Please update either the comment or the value to reflect the correct configuration.
| MEGATRON_ETP=1 | ||
| # Qwen3-235B-A22B has 94 blocks, so we need to set the last pipeline stage layer to use 4 blocks | ||
| MEGATRON_LAST_PIPELINE_STAGE_LAYER=4 | ||
| MEGATRON_LAST_PIPELINE_STAGE_LAYER=10 |
There was a problem hiding this comment.
The comment on line 29 states that the last pipeline stage should use 4 blocks, but the value of MEGATRON_LAST_PIPELINE_STAGE_LAYER is changed to 10. This is inconsistent. Please update either the comment or the value to reflect the correct configuration.
1 participant
Adds example script for Qwen3-235-A3B-Instruct-2507 DAPO training with the Megatron backend + LoRA. Mean @ 32 increases from ~60% -> 67.5%.
Also adds eval generation length metrics to the evaluate function by default - this change affects both the generation and training entry points.
Training Curve
Metrics