Fix Nemotron 3 SFT example: add Mamba modules to LoRA targets

[davidnichols-ops]

Summary

• Replaced gate_proj with in_proj and out_proj in the LoRA target modules for the Nemotron 3 SFT example
• Added explanatory comment about NemotronH's hybrid Mamba-Attention architecture

Why

NemotronH is a hybrid Mamba-Attention architecture. The previous target modules (q_proj k_proj v_proj o_proj gate_proj up_proj down_proj) had two problems:

1. gate_proj does not exist in NemotronH. The MLP layers use up_proj/down_proj (standard MLP) or MoE variants, not gate_proj.
2. Mamba layers were missing entirely. NemotronH models have Mamba SSM layers with in_proj/out_proj modules. In Nemotron-Nano-9B-v2, 52 of 56 layers are Mamba. Omitting these from LoRA targets means 93% of the model is frozen during fine-tuning, resulting in zero learning (DPO loss stays at ln(2)).

See huggingface/peft#3554 for the full analysis.

What Changed

• examples/scripts/sft_nemotron_3.py: Changed --lora_target_modules from q_proj k_proj v_proj o_proj gate_proj up_proj down_proj to q_proj k_proj v_proj o_proj in_proj out_proj up_proj down_proj
• Added a comment block explaining the hybrid architecture and why Mamba modules must be included

Test plan

• Example runs without errors on Nemotron-3-Nano-30B-A3B
• LoRA adapter targets both Mamba and Attention layers

---

Note

Low Risk
Documentation and example CLI defaults only; no library or training runtime code changes.

Overview
Updates the Nemotron 3 SFT example so LoRA actually trains the hybrid model instead of mostly freezing it.

The sample accelerate command now uses in_proj and out_proj instead of gate_proj, and keeps attention/MLP targets (q_proj, k_proj, v_proj, o_proj, up_proj, down_proj). gate_proj is not part of NemotronH, and without the Mamba projections most layers stay frozen.

A short module docstring explains why NemotronH needs both Mamba and attention/MLP modules in --lora_target_modules, with a link to the PEFT issue.

^{Reviewed by Cursor Bugbot for commit a3e38df. Bugbot is set up for automated code reviews on this repo. Configure here.}

[github-actions]

Your PR has been closed because it does not follow the required PR template. Please open a new PR following the template. See CONTRIBUTING.md for guidance.

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

Want higher recall? High effort reviews run extra passes and find more bugs. A team admin can switch effort levels in the Cursor dashboard.

^{Reviewed by Cursor Bugbot for commit a3e38df. Configure here.}

[cursor]

Invalid Mamba LoRA out_proj target

High Severity

Adding out_proj to --lora_target_modules breaks NemotronH LoRA. Mamba mixers feed out_proj.weight into fused kernels, so LoRA on that module is skipped or rejected by PEFT, and adapters can fail at merge. Including in_proj is the useful Mamba change here; out_proj is not a valid target for this architecture.

Additional Locations (1)

• examples/scripts/sft_nemotron_3.py#L27-L33

 

^{Reviewed by Cursor Bugbot for commit a3e38df. Configure here.}