Skip to content

Add Nemotron 3.5 Lightning support and testing - #6736

Merged
qgallouedec merged 2 commits into
mainfrom
add-nemotron-3-5-lightning
Aug 17, 2026
Merged

Add Nemotron 3.5 Lightning support and testing#6736
qgallouedec merged 2 commits into
mainfrom
add-nemotron-3-5-lightning

Conversation

@qgallouedec

@qgallouedec qgallouedec commented Aug 13, 2026

Copy link
Copy Markdown
Member

Released 2026-08-11: nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 — 30B/3B-active hybrid Mamba-2 + MoE, 1M context, open weights published as an SFT/RL/distillation starting point.

  • Tiny model script nemotron_3_5_lightning_for_causal_lm.py (NemotronHForCausalLM), mirroring the Ultra script — Mamba weights are bf16 in the checkpoint, so no fp32 restore.
  • Vendored nemotron_3_5_lightning.jinja + _training.jinja (Ultra's template minus the medium_effort hint) and registered both in chat_template_utils.py (reuses qwen3_5_schema / nemotron_3_template).
  • Test params added next to Nemotron Ultra in test_chat_template_utils.py (4 lists) and test_data_utils.py. Trainer tests keep Nano as the single NemotronH representative.
  • Prerequisite: push trl-internal-testing/tiny-NemotronHForCausalLM-3.5-lightning with transformers==5.3.0 before merging.

Note

Low Risk
Additive model-family support (templates, registration, tests, docs) with no changes to existing Nemotron or training paths; merge depends on publishing the tiny Hub model first.

Overview
Adds Nemotron 3.5 Lightning to TRL’s supported chat-template families so SFT/GRPO can auto-swap training templates and tool/response parsing works like other Nemotron-H models.

Vendored nemotron_3_5_lightning.jinja (Ultra-style template without medium_effort) and nemotron_3_5_lightning_training.jinja (assistant {% generation %} wrappers only). chat_template_utils.py loads both, maps Lightning to qwen3_5_schema / nemotron_3_template, and returns the training template from get_training_chat_template.

A new tiny-model script nemotron_3_5_lightning_for_causal_lm.py builds/pushes trl-internal-testing/tiny-NemotronHForCausalLM-3.5-lightning from the NVIDIA checkpoint (transformers ≥ 5.3.0). Docs and README catalog the family; tests add the tiny fixture to chat-template and data-utils parametrizations alongside existing Nemotron Ultra cases.

Reviewed by Cursor Bugbot for commit 1360c7b. Bugbot is set up for automated code reviews on this repo. Configure here.

@bot-ci-comment

Copy link
Copy Markdown

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

@qgallouedec
qgallouedec merged commit 2f80dd7 into main Aug 17, 2026
13 checks passed
@qgallouedec
qgallouedec deleted the add-nemotron-3-5-lightning branch August 17, 2026 19:47
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants