[Bugfix] Reland #53020: tie lm_head.weight for Nemotron Parse when checkpoint omits it (#60272)

Signed-off-by: Anish Kumar <aniskumar@nvidia.com>
Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
This commit is contained in:
aniskumar-nv
2026-10-08 04:40:55 +00:00
committed by GitHub
co-authored by Cyrus Leung Claude Fable 5.1
parent 4187ade7b8
commit 775792f2d5
2 changed files with 19 additions and 1 deletions
@@ -568,6 +568,15 @@ class NemotronParseForConditionalGeneration(nn.Module, SupportsMultiModal):
self.lm_head = ParallelLMHead(
config.decoder.vocab_size, config.decoder.d_model, quant_config=quant_config
)
# Some checkpoints (e.g. compact exports) tie the output head to the
# decoder's input embeddings instead of materializing a separate
# lm_head.weight tensor.
tie_word_embeddings = bool(
getattr(config, "tie_word_embeddings", False)
or getattr(config.decoder, "tie_word_embeddings", False)
)
if tie_word_embeddings:
self.lm_head = self.lm_head.tie_weights(self.decoder.embed_tokens)
self.logits_processor = LogitsProcessor(
self.vocab_size, config.decoder.vocab_size
)