[recipe] fix: Update the grpo training script for gpt-oss models

[HJSang]

What does this PR do?

Add concise overview of what this PR aims to achieve or accomplish. Reference related GitHub issues and PRs that help with the review.

• remove customized python packages since they are already supported
• Add reasoning_effort input
• recommend a setup for batch size to avoid MOE instability.

Checklist Before Starting

• Search for similar PRs. Paste at least one query link here: ...
• Format the PR title as [{modules}] {type}: {description} (This will be checked by the CI)
  • {modules} include fsdp, megatron, sglang, vllm, rollout, trainer, ci, training_utils, recipe, hardware, deployment, ray, worker, single_controller, misc, perf, model, algo, env, tool, ckpt, doc, data
  • If this PR involves multiple modules, separate them with , like [megatron, fsdp, doc]
  • {type} is in feat, fix, refactor, chore, test
  • If this PR breaks any API (CLI arguments, config, function signature, etc.), add [BREAKING] to the beginning of the title.
  • Example: [BREAKING][fsdp, megatron] feat: dynamic batching

Test

Test offline: run a training job

For changes that can not be tested by CI (e.g., algorithm implementation, new model support), validate by experiment(s) and show results like training curve plots, evaluation results, etc.

API and Usage Example

Demonstrate how the API changes if any, and provide usage example(s) if possible.

# Add code snippet or script demonstrating how to use this

Design & Code Changes

Demonstrate the high-level design if this PR is complex, and list the specific changes.

Checklist Before Submitting

Important

Please check all the following items before requesting a review, otherwise the reviewer might deprioritize this PR for review.

• Read the Contribute Guide.
• Apply pre-commit checks: pre-commit install && pre-commit run --all-files --show-diff-on-failure --color=always
• Add / Update the documentation.
• Add unit or end-to-end test(s) to the CI workflow to cover all the code. If not feasible, explain why: ...
• Once your PR is ready for CI, send a message in the ci-request channel in the verl Slack workspace. (If not accessible, please try the Feishu group (飞书群).)

[gemini-code-assist]

Code Review

This pull request updates the GRPO training script for gpt-oss models by removing obsolete package installation steps and tuning several training parameters. The changes include adjusting batch sizes for MoE stability, increasing the maximum response length, and adding a reasoning_effort parameter.

While the changes are generally positive, I have a critical concern regarding performance. The get_model.py script within run_gptoss_20b.sh configures the model to use attn_implementation=\"eager\". This setting will persist in the saved model's config.json and will be used during the main FSDP training loop. For a large model like gpt-oss-20b, using the unoptimized eager attention mechanism will be significantly slower and more memory-intensive than optimized alternatives like Flash Attention, severely impacting training efficiency.

While eager might be required for the Mxfp4Config quantization with device_map=\"auto\", it's detrimental for training. I recommend investigating if a more performant attention implementation like flash_attention_2 can be used. If eager is strictly necessary for the initial model download and quantization, consider patching the config.json file after the get_model.py script runs to switch to a faster implementation for the training stage. For example, you could add the following line to the shell script:

sed -i 's/\"attn_implementation\": \"eager\"/\"attn_implementation\": \"flash_attention_2\"/' \"${HOME}/models/gpt-oss-20b-bf16/config.json\"

This would ensure the training part of the script benefits from optimized attention, while the initial model preparation remains unchanged.