feat: nemotron 3.5 cfg (#3442)
* feat(recipes): add nemotron MTP PEFT finetune example Adds a LoRA finetuning example for Nemotron hybrid mamba-attention MoE checkpoints with multi-token prediction, using 2 weight-tied MTP iterations of a single physical MTP layer (mirrors Megatron's --mtp-use-repeated-layer). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Signed-off-by: adil-a <adil.asif2000@hotmail.com> * feat(recipes): use TE backends with gmm experts in nemotron MTP recipe Switch the backend block to attn/linear/rms_norm=te with gmm experts and the hybridep dispatcher (matching the ultra recipe), drop the inert torch.compile block, and raise ep_size to 8. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Signed-off-by: adil-a <adil.asif2000@hotmail.com> * feat(recipes): switch nemotron MTP recipe dispatcher to deepep Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Signed-off-by: adil-a <adil.asif2000@hotmail.com> * fix(nemotron_v3): report view-loaded MTP expert keys under mtp. prefix The MTP pass of from_hf records in-place (strided-view) loaded expert keys in mtp.-stripped form because the shared merge loop only sees stripped keys. The checkpoint loader diffs view_loaded_native_keys against the model's real parameter names, so the stripped entries surfaced as a false 'Checkpoint key mismatch' warning (missing mtp.layers.N.mixer.experts.*, unexpected layers.N.mixer.experts.*) on every MTP checkpoint load — and would raise spuriously under safe custom-MoE TP. Re-prefix the entries the MTP merge adds so the key-diff matches them. Weights were always loaded correctly; this fixes only the bookkeeping. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Signed-off-by: adil-a <adil.asif2000@hotmail.com> * feat(recipes): name Nemotron Nano v3.5 Lightning in MTP PEFT recipe Rename the redacted nemotron MTP recipe to nemotron_nano_v3_5_lightning_hellaswag_peft.yaml and point it at the public nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 checkpoint. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Signed-off-by: adil-a <adil.asif2000@hotmail.com> * docs(recipes): drop Megatron flag mention from lightning recipe comment Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Signed-off-by: adil-a <adil.asif2000@hotmail.com> --------- Signed-off-by: adil-a <adil.asif2000@hotmail.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
A
Adil committed
8cb12a35b65eda421d8f1fa485c658eb06d22c76
Parent: 0ae5d44
Committed by GitHub <noreply@github.com>
on 8/8/2026, 7:21:59 AM