SIGN IN SIGN UP

Fix nemotron_h save_pretrained emitting singular backbone.embedding.weight (#48075)

Fix nemotron_h original-format save emitting backbone.embedding.weight

The nemotron_h conversion mapping contained
WeightRenaming("embedding.weight", "embeddings.weight"), assuming original
checkpoints use the singular key. All published NemotronH checkpoints
(Nemotron-H-8B-Base-8K, NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16,
NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16) actually store the plural
`backbone.embeddings.weight`, so this rule never matches on load. On
original-format save, however, its reverse rewrites the key to the singular
`backbone.embedding.weight`, producing checkpoints whose embeddings cannot be
loaded back by external NemotronH consumers (vLLM, NeMo Automodel) or by the
original modeling code. Drop the rule.

Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Y
Yuekai Zhang committed
9916b4671fb9e8c543cc565718312275c6ca07be
Parent: 89a7273
Committed by GitHub <noreply@github.com> on 8/24/2026, 4:36:02 PM