From 7ff656cc8bc7b30c60ef36f5ad97ecf827f1cd6d Mon Sep 17 00:00:00 2001 From: Shaun Kotek <93727115+shaunkotek@users.noreply.github.com> Date: Tue, 7 Jul 2026 15:01:45 +0300 Subject: [PATCH] fix: ensure no double load of lm head in nemotron mtp (#47440) Signed-off-by: Shaun Kotek - Nvidia --- vllm/model_executor/models/nemotron_h_mtp.py | 6 +----- 1 file changed, 1 insertion(+), 5 deletions(-) diff --git a/vllm/model_executor/models/nemotron_h_mtp.py b/vllm/model_executor/models/nemotron_h_mtp.py index fe737438c30..bd4b908dc61 100644 --- a/vllm/model_executor/models/nemotron_h_mtp.py +++ b/vllm/model_executor/models/nemotron_h_mtp.py @@ -415,11 +415,7 @@ class NemotronHMTP(nn.Module, SupportsPP): for name, loaded_weight in weights: # Only process MTP weights - skip all non-MTP weights - if ( - not name.startswith("mtp.") - and "embeddings" not in name - and "lm_head" not in name - ): + if not name.startswith("mtp.") and "embeddings" not in name: continue # Skip rotary embeddings (computed, not loaded) if "rotary_emb.inv_freq" in name: