[Bugfix] Tie lm_head.weight for Nemotron Parse when checkpoint omits it (#53020)

Signed-off-by: Anish Kumar <aniskumar@nvidia.com>
Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>
This commit is contained in:
aniskumar-nv
2026-10-01 09:26:33 -07:00
committed by GitHub
co-authored by Cyrus Leung
parent cfd54ca881
commit 83cadd65d9
2 changed files with 19 additions and 1 deletions
@@ -104,7 +104,16 @@ def run_test(
)
@pytest.mark.parametrize("model", ["nvidia/NVIDIA-Nemotron-Parse-v1.2"])
@pytest.mark.parametrize(
"model",
[
"nvidia/NVIDIA-Nemotron-Parse-v1.2",
# v2.0's checkpoint ties lm_head to the decoder's input embeddings
# instead of shipping a separate lm_head.weight tensor; this
# exercises that path (see tie_weights() in nemotron_parse.py).
"nvidia/NVIDIA-Nemotron-Parse-2.0",
],
)
@pytest.mark.parametrize("dtype", ["bfloat16"])
@pytest.mark.parametrize("num_logprobs", [5])
def test_models(
@@ -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
)