fix: ensure no double load of lm head in nemotron mtp (#47440)
Signed-off-by: Shaun Kotek - Nvidia <skotek@nvidia.com>
Shaun Kotek committed
Jul 7, 2026 at 15:01 UTC
7ff656cc8bc7b30c60ef36f5ad97ecf827f1cd6d
1 file changed
+1
-5
vllm/model_executor/models/nemotron_h_mtp.py
+1
-5
@@ -415,11 +415,7 @@ class NemotronHMTP(nn.Module, SupportsPP):
415
416
for name, loaded_weight in weights:
417
# Only process MTP weights - skip all non-MTP weights
418
- if (
419
- not name.startswith("mtp.")
420
- and "embeddings" not in name
421
- and "lm_head" not in name
422
- ):
418
+ if not name.startswith("mtp.") and "embeddings" not in name:
419
continue
420
# Skip rotary embeddings (computed, not loaded)
421
if "rotary_emb.inv_freq" in name: