[trainer] feat: add gpt-oss sglang-megatron example

[yiakwy-xpu-ml-framework-team]

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.

This is follow up of #4323 to add SGLang running example

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

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.

py312+torch2.8+vllm0.11+sglang0.5.5+te2.9

docker:

Build Instruction:

vllm:

• training curve

• memory footprint

sglang:

• V1
  

  In our first trial, single node version of SGLang has unusual memory footprint (with the same test codes , just change vLLM to SGlang), we are currently working on improving its memory efficiency.

py312+torch2.9.1+sglang0.5.5+te2.9

sglang:

Benchmark

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 introduces support for the sglang-megatron backend by adding a new example script and modifying an import fallback. My review has identified a critical security vulnerability due to a hardcoded API key in the example script, which must be removed and the key revoked. Additionally, a high-severity typo in the Python code could lead to import errors under certain conditions and should be corrected.

[wuxibin89]

If you want to use Megatron-Bridge instead of mbridge, please set

actor_rollout_ref.actor.megatron.vanilla_mbridge=False

[yiakwy-xpu-ml-framework-team]

Yes we already have it :

Currently we test with standard bridge supported by NVIDIA both verl and megatron. It works well.

[wuxibin89]

Why we need this instead of megatron.param_offload=True?

[yiakwy-xpu-ml-framework-team]

hi @wuxibin89 megatron-bridge supports to load modules directly from cpu :

self.bridge.export_hf_weights(self.actor.actor_module, cpu=load_actor_from_cpu, show_progress=False)

If that is true we don't need to reload parameters from cpu to gpu , then offload megatron module to cpu.