Bug: Fix weight loader error when LM head weights are tied (#3766)

python/sglang/srt/models/llama.py

@@ -458,6 +458,8 @@ class LlamaForCausalLM(nn.Module):
                 continue
             if name.startswith("model.vision_tower") and name not in params_dict:
                 continue
+            if self.config.tie_word_embeddings and "lm_head.weight" in name:
+                continue
             # Handle FP8 kv-scale remapping
             if "scale" in name:
                 name = maybe_remap_kv_scale_name(name, params_dict)

python/sglang/srt/models/minicpm.py

@@ -339,6 +339,8 @@ class MiniCPMForCausalLM(nn.Module):
                 # Models trained using ColossalAI may include these tensors in
                 # the checkpoint. Skip them.
                 continue
+            if self.config.tie_word_embeddings and "lm_head.weight" in name:
+                continue
 
             for param_name, weight_name, shard_id in stacked_params_mapping:
                 if weight_name not in name:

python/sglang/srt/models/minicpm3.py

@@ -603,6 +603,8 @@ class MiniCPM3ForCausalLM(nn.Module):
                 # Models trained using ColossalAI may include these tensors in
                 # the checkpoint. Skip them.
                 continue
+            if self.config.tie_word_embeddings and "lm_head.weight" in name:
+                continue
 
             for param_name, weight_name, shard_id in stacked_params_mapping:
                 if weight_name not in name:

python/sglang/srt/models/olmo.py

@@ -325,6 +325,8 @@ class OlmoForCausalLM(nn.Module):
                 # Models trained using ColossalAI may include these tensors in
                 # the checkpoint. Skip them.
                 continue
+            if self.config.tie_word_embeddings and "lm_head.weight" in name:
+                continue
             for param_name, weight_name, shard_id in stacked_params_mapping:
                 if weight_name not in name:
                     continue

python/sglang/srt/models/phi3_small.py

@@ -433,6 +433,8 @@ class Phi3SmallForCausalLM(nn.Module):
                 continue
             if name.endswith(".bias") and name not in params_dict:
                 continue
+            if self.config.tie_word_embeddings and "lm_head.weight" in name:
+                continue
 
             param = params_dict[name]
             weight_loader = getattr(param, "weight_loader", default_weight_loader)

python/sglang/srt/models/qwen2.py

@@ -377,6 +377,8 @@ class Qwen2ForCausalLM(nn.Module):
                 # Models trained using ColossalAI may include these tensors in
                 # the checkpoint. Skip them.
                 continue
+            if self.config.tie_word_embeddings and "lm_head.weight" in name:
+                continue
             if name.startswith("model.vision_tower") and name not in params_dict:
                 continue
 

python/sglang/srt/models/qwen2_vl.py

@@ -586,6 +586,8 @@ class Qwen2VLForConditionalGeneration(nn.Module):
         for name, loaded_weight in weights:
             if "rotary_emb.inv_freq" in name:
                 continue
+            if self.config.tie_word_embeddings and "lm_head.weight" in name:
+                continue
 
             for param_name, weight_name, shard_id in stacked_params_mapping:
                 if weight_name not in name:

python/sglang/srt/models/torch_native_llama.py

@@ -486,6 +486,8 @@ class TorchNativeLlamaForCausalLM(nn.Module):
                 continue
             if name.startswith("model.vision_tower") and name not in params_dict:
                 continue
+            if self.config.tie_word_embeddings and "lm_head.weight" in name:
+                continue
 
             for param_name, weight_name, shard_id in stacked_params_mapping:
                 if weight_name not in name: