[trainer] feat: Add Nemo-Automodel as alternative training engine

[HuiyingLi]

What does this PR do?

Add NeMo-Automodel as a training engine. The SFT trainer is tested with Qwen2.5-0.5B.

• automodel engine matches exactly with FSDP engine for SFT trainer (TP1/TP2/rmpad=True/False)
• use_remove_padding=True matches use_remove_padding=False
• EP support tested with kimi moonlight 16B

Relevant PRs:

• RFC [RFC] support Nemo-AutoModel as an alternative training backend #5245
• Add VeOmni [trainer] feat: Implemented VeomniEngine as a alternative training backend #4072
• Add TorchTitan [trainer] feat: Add Torchtitan as alternative training engine #5051

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, veomni, sglang, vllm, rollout, trainer, ci, training_utils, recipe, hardware, deployment, ray, worker, single_controller, misc, perf, model, algo, env, tool, ckpt, doc, data, cfg, reward, fully_async, one_step_off
  • 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

automodel backend 1gpu & 4gpu tp1/tp2 against FSDP backend 1gpu, rmpad true/false.

automodel backend finetuning moonlight 16B with ep8 8H100

automodel backend finetuning Qwen3 30B with ep8 8H100

automodel backend finetuning qwen2.5-7b with 4H100 fsdp2

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).)
• If your PR is related to the recipe submodule, please also update the reference to the submodule commit via git submodule update --remote or cd recipe && git pull origin main.

[CLAassistant]

All committers have signed the CLA.

[gemini-code-assist]

Code Review

This pull request introduces a new automodel SFT backend, which leverages nemo_automodel for distributed training. The changes include adding the engine implementation, configuration files, and test scripts. I've identified a configuration issue in the test script and a maintainability concern in the engine implementation. Overall, this is a significant feature addition.

[gemini-code-assist]

The model's output is conditionally wrapped in a SimpleNamespace if it's a raw tensor. This suggests an inconsistent return type from self.module, which can make the code harder to maintain and reason about. It would be more robust to enforce a consistent, structured return type (like CausalLMOutput) from the model to avoid such conditional handling.

[ISEEKYAN]

hi @HuiyingLi , thanks to your great contribution.
I found the MFU of automodel is lower than FSDP on 0.5B model and the MFU is less than 1% on 16B MoE model, is this supposed to be right? Could you provide some fair comparison on popular models such as 7B dense or 30B MoE?

[HuiyingLi]

hi @HuiyingLi , thanks to your great contribution. I found the MFU of automodel is lower than FSDP on 0.5B model and the MFU is less than 1% on 16B MoE model, is this supposed to be right? Could you provide some fair comparison on popular models such as 7B dense or 30B MoE?

Hi @ISEEKYAN ,
Thank you!

• for 0.5B model, FSDP engine is single GPU while automodel is 4 GPU. I've updated the chart with automodel single GPU for comparison.
• for the low mfu on 16B model: due to very small seqlen and batch size. I've updated the chart with larger seqlen and gbs, and added charts for Qwen 30B MoE and Qwen 7B dense.

[ISEEKYAN]

great, is your exp on H100? If true, the MFU looks good but it will be better to have a fair comparison with FSDP or Megatron. But this is not a blocker of merging this PR. It is just a good reference for the users to adopt AutoModel, and it will be better to add a doc to show the comparison and an example for the user to easily hand on

[ETOgaosion]

@HuiyingLi Thanks for your great contribution! Could you please sign the CLA~

[ETOgaosion]

Maybe in another PR, we should refactor docs/start/install.rst to support all model engines and rollout engine install methods and use some displays for better choices.

[ETOgaosion]

We should also prepare some CI tests for Nemo-Automodel

[HuiyingLi]

@ETOgaosion Thank you for reviewing! I've signed the CLA. Could you please suggest how to pass the ci for this PR? For the license error do you suggest to add Nvidia to the list? Should we add ci tests in this PR? Thank you!

[ETOgaosion]

For this Nemo-Automodel CI, wondering whether verl's docker image contains needed components of Nemo-AutoModel?

Or if there is no dependency conflicts with megatron, we can directly reuse current megatron docker images to run this backend by adding needed packages.

For Adding CI, we can use another PR to test new images, we can merge this first.

With the image, you can enable test in tests/special_e2e/sft/test_sft_engine_all.sh as you implemented here to test.

[ETOgaosion]

License LGTM

[ETOgaosion]

@HuiyingLi For failed CIs, could you rebase main and retry?