[Model] Pass param prefix to LLMHead (#24862)

whx-sjtu · web-flow · commit 4a9375fe9dbd · 2025-09-17T16:01:27.000+08:00
Signed-off-by: whx-sjtu &lt;2952154980@qq.com&gt;
diff --git a/vllm/model_executor/models/arctic.py b/vllm/model_executor/models/arctic.py
@@ -427,6 +427,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
             self.vocab_size,
             config.hidden_size,
             quant_config=quant_config,
+            prefix=maybe_prefix(prefix, "lm_head"),
         )
         if self.config.tie_word_embeddings:
             self.lm_head.weight = self.model.embed_tokens.weight
diff --git a/vllm/model_executor/models/aria.py b/vllm/model_executor/models/aria.py
@@ -539,6 +539,7 @@ def __init__(
             config.text_config.hidden_size,
             org_num_embeddings=self.language_model.org_vocab_size,
             quant_config=quant_config,
+            prefix=maybe_prefix(prefix, "lm_head"),
         )
         logit_scale = getattr(config, "logit_scale", 1.0)
         self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
diff --git a/vllm/model_executor/models/baichuan.py b/vllm/model_executor/models/baichuan.py
@@ -51,7 +51,8 @@
 
 from .interfaces import SupportsLoRA, SupportsPP, SupportsQuant
 from .utils import (AutoWeightsLoader, is_pp_missing_parameter,
-                    make_empty_intermediate_tensors_factory, make_layers)
+                    make_empty_intermediate_tensors_factory, make_layers,
+                    maybe_prefix)
 
 
 def _get_alibi_slopes(total_num_heads: int) -> torch.Tensor:
@@ -394,7 +395,8 @@ def __init__(
                                    position_embedding=position_embedding)
         self.lm_head = ParallelLMHead(config.vocab_size,
                                       config.hidden_size,
-                                      quant_config=quant_config)
+                                      quant_config=quant_config,
+                                      prefix=maybe_prefix(prefix, "lm_head"))
         self.lm_head.weight.weight_loader = self.lm_head_weight_loader
         if self.config.tie_word_embeddings:
             self.lm_head.weight = self.model.embed_tokens.weight
diff --git a/vllm/model_executor/models/bamba.py b/vllm/model_executor/models/bamba.py
@@ -514,6 +514,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
             # We need bigger padding if using lora for kernel
             # compatibility
             if not lora_config else lora_config.lora_vocab_padding_size,
+            prefix=maybe_prefix(prefix, "lm_head"),
         )
         # Used to track and store by the Mamba cache between steps.
         self.mamba_cache: Optional[MambaCacheManager] = None
diff --git a/vllm/model_executor/models/bloom.py b/vllm/model_executor/models/bloom.py
@@ -330,7 +330,9 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
             self.lm_head = self.transformer.word_embeddings
         else:
             self.lm_head = ParallelLMHead(self.config.vocab_size,
-                                          self.config.hidden_size)
+                                          self.config.hidden_size,
+                                          prefix=maybe_prefix(
+                                              prefix, "lm_head"))
 
         self.logits_processor = LogitsProcessor(config.vocab_size)
         self.make_empty_intermediate_tensors = (
diff --git a/vllm/model_executor/models/chameleon.py b/vllm/model_executor/models/chameleon.py
@@ -960,6 +960,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         self.lm_head = ParallelLMHead(
             self.unpadded_vocab_size,
             config.hidden_size,
+            prefix=maybe_prefix(prefix, "lm_head"),
         )
         if config.tie_word_embeddings:
             self.lm_head.weight = self.model.embed_tokens.weight
diff --git a/vllm/model_executor/models/dbrx.py b/vllm/model_executor/models/dbrx.py
@@ -438,6 +438,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
             org_num_embeddings=config.vocab_size,
             padding_size=DEFAULT_VOCAB_PADDING_SIZE,
             quant_config=quant_config,
+            prefix=maybe_prefix(prefix, "lm_head"),
         )
         self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
                                                 config.vocab_size)
diff --git a/vllm/model_executor/models/deepseek.py b/vllm/model_executor/models/deepseek.py
@@ -453,9 +453,12 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         self.quant_config = quant_config
         self.model = DeepseekModel(vllm_config=vllm_config,
                                    prefix=maybe_prefix(prefix, "model"))
-        self.lm_head = ParallelLMHead(config.vocab_size,
-                                      config.hidden_size,
-                                      quant_config=quant_config)
+        self.lm_head = ParallelLMHead(
+            config.vocab_size,
+            config.hidden_size,
+            quant_config=quant_config,
+            prefix=maybe_prefix(prefix, "lm_head"),
+        )
         if self.config.tie_word_embeddings:
             self.lm_head.weight = self.model.embed_tokens.weight
         self.logits_processor = LogitsProcessor(config.vocab_size)
diff --git a/vllm/model_executor/models/deepseek_eagle.py b/vllm/model_executor/models/deepseek_eagle.py
@@ -199,7 +199,8 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
 
         self.lm_head = ParallelLMHead(self.config.vocab_size,
                                       self.config.hidden_size,
-                                      quant_config=quant_config)
+                                      quant_config=quant_config,
+                                      prefix=maybe_prefix(prefix, "lm_head"))
 
         logit_scale = getattr(self.config, "logit_scale", 1.0)
         self.logits_processor = LogitsProcessor(self.config.vocab_size,
diff --git a/vllm/model_executor/models/deepseek_v2.py b/vllm/model_executor/models/deepseek_v2.py
@@ -823,9 +823,12 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         self.model = DeepseekV2Model(vllm_config=vllm_config,
                                      prefix=maybe_prefix(prefix, "model"))
         if get_pp_group().is_last_rank:
-            self.lm_head = ParallelLMHead(config.vocab_size,
-                                          config.hidden_size,
-                                          quant_config=quant_config)
+            self.lm_head = ParallelLMHead(
+                config.vocab_size,
+                config.hidden_size,
+                quant_config=quant_config,
+                prefix=maybe_prefix(prefix, "lm_head"),
+            )
         else:
             self.lm_head = PPMissingLayer()
         self.logits_processor = LogitsProcessor(config.vocab_size)
diff --git a/vllm/model_executor/models/dots1.py b/vllm/model_executor/models/dots1.py
@@ -504,7 +504,9 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         if get_pp_group().is_last_rank:
             self.lm_head = ParallelLMHead(config.vocab_size,
                                           config.hidden_size,
-                                          quant_config=quant_config)
+                                          quant_config=quant_config,
+                                          prefix=maybe_prefix(
+                                              prefix, "lm_head"))
         else:
             self.lm_head = PPMissingLayer()
         self.logits_processor = LogitsProcessor(config.vocab_size)
diff --git a/vllm/model_executor/models/ernie45_moe.py b/vllm/model_executor/models/ernie45_moe.py
@@ -562,7 +562,9 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         if get_pp_group().is_last_rank:
             self.lm_head = ParallelLMHead(config.vocab_size,
                                           config.hidden_size,
-                                          quant_config=quant_config)
+                                          quant_config=quant_config,
+                                          prefix=maybe_prefix(
+                                              prefix, "lm_head"))
         else:
             self.lm_head = PPMissingLayer()
 
diff --git a/vllm/model_executor/models/ernie45_vl_moe.py b/vllm/model_executor/models/ernie45_vl_moe.py
@@ -557,7 +557,9 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         if get_pp_group().is_last_rank:
             self.lm_head = ParallelLMHead(config.vocab_size,
                                           config.hidden_size,
-                                          quant_config=quant_config)
+                                          quant_config=quant_config,
+                                          prefix=maybe_prefix(
+                                              prefix, "lm_head"))
         else:
             self.lm_head = PPMissingLayer()
 
diff --git a/vllm/model_executor/models/ernie_mtp.py b/vllm/model_executor/models/ernie_mtp.py
@@ -158,7 +158,8 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
                                               prefix=maybe_prefix(
                                                   prefix, "model"))
         self.lm_head = ParallelLMHead(self.config.vocab_size,
-                                      self.config.hidden_size)
+                                      self.config.hidden_size,
+                                      prefix=maybe_prefix(prefix, "lm_head"))
         self.sampler = get_sampler()
 
         if self.config.tie_word_embeddings:
diff --git a/vllm/model_executor/models/exaone.py b/vllm/model_executor/models/exaone.py
@@ -502,6 +502,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
                 # compatibility
                 if not lora_config else lora_config.lora_vocab_padding_size,
                 quant_config=quant_config,
+                prefix=maybe_prefix(prefix, "lm_head"),
             )
             if config.tie_word_embeddings:
                 self.lm_head.weight = self.transformer.wte.weight
diff --git a/vllm/model_executor/models/exaone4.py b/vllm/model_executor/models/exaone4.py
@@ -485,6 +485,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
                 # compatibility
                 if not lora_config else lora_config.lora_vocab_padding_size,
                 quant_config=quant_config,
+                prefix=maybe_prefix(prefix, "lm_head"),
             )
             if config.tie_word_embeddings:
                 self.lm_head.weight = self.model.embed_tokens.weight
diff --git a/vllm/model_executor/models/falcon.py b/vllm/model_executor/models/falcon.py
@@ -473,6 +473,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
                 config.vocab_size,
                 config.hidden_size,
                 quant_config=quant_config,
+                prefix=maybe_prefix(prefix, "lm_head"),
             )
         self.logits_processor = LogitsProcessor(config.vocab_size)
         self.make_empty_intermediate_tensors = (
diff --git a/vllm/model_executor/models/falcon_h1.py b/vllm/model_executor/models/falcon_h1.py
@@ -607,6 +607,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
                     # compatibility
                     if not lora_config else
                     lora_config.lora_vocab_padding_size),
+                prefix=maybe_prefix(prefix, "lm_head"),
             )
             self.lm_head_multiplier = config.lm_head_multiplier
             if self.tie_word_embeddings:
diff --git a/vllm/model_executor/models/glm4_moe.py b/vllm/model_executor/models/glm4_moe.py
@@ -608,7 +608,9 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         if get_pp_group().is_last_rank:
             self.lm_head = ParallelLMHead(config.vocab_size,
                                           config.hidden_size,
-                                          quant_config=quant_config)
+                                          quant_config=quant_config,
+                                          prefix=maybe_prefix(
+                                              prefix, "lm_head"))
         else:
             self.lm_head = PPMissingLayer()
         self.logits_processor = LogitsProcessor(config.vocab_size)
diff --git a/vllm/model_executor/models/gpt_bigcode.py b/vllm/model_executor/models/gpt_bigcode.py
@@ -302,7 +302,8 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
             self.lm_head = ParallelLMHead(
                 self.transformer.vocab_size,
                 self.transformer.embed_dim,
-                org_num_embeddings=self.config.vocab_size)
+                org_num_embeddings=self.config.vocab_size,
+                prefix=maybe_prefix(prefix, "lm_head"))
         self.unpadded_vocab_size = config.vocab_size
         if lora_config:
             self.unpadded_vocab_size += lora_config.lora_extra_vocab_size
diff --git a/vllm/model_executor/models/gpt_j.py b/vllm/model_executor/models/gpt_j.py
@@ -306,6 +306,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
             config.n_embd,
             bias=True,
             quant_config=quant_config,
+            prefix=maybe_prefix(prefix, "lm_head"),
         )
         self.logits_processor = LogitsProcessor(config.vocab_size)
         self.make_empty_intermediate_tensors = (
diff --git a/vllm/model_executor/models/gpt_oss.py b/vllm/model_executor/models/gpt_oss.py
@@ -655,6 +655,7 @@ def __init__(
         self.lm_head = ParallelLMHead(
             self.config.vocab_size,
             self.config.hidden_size,
+            prefix=maybe_prefix(prefix, "lm_head"),
         )
         self.logits_processor = LogitsProcessor(self.config.vocab_size)
         self.make_empty_intermediate_tensors = (
diff --git a/vllm/model_executor/models/granite.py b/vllm/model_executor/models/granite.py
@@ -434,6 +434,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
                 # compatibility
                 if not lora_config else lora_config.lora_vocab_padding_size,
                 quant_config=quant_config,
+                prefix=maybe_prefix(prefix, "lm_head"),
             )
             if config.tie_word_embeddings:
                 self.lm_head.weight = self.model.embed_tokens.weight
diff --git a/vllm/model_executor/models/granitemoe.py b/vllm/model_executor/models/granitemoe.py
@@ -487,6 +487,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
             # compatibility
             if not lora_config else lora_config.lora_vocab_padding_size,
             quant_config=quant_config,
+            prefix=maybe_prefix(prefix, "lm_head"),
         )
         if config.tie_word_embeddings:
             self.lm_head.weight = self.model.embed_tokens.weight
diff --git a/vllm/model_executor/models/hunyuan_v1.py b/vllm/model_executor/models/hunyuan_v1.py
@@ -58,7 +58,7 @@
 
 from .interfaces import SupportsLoRA, SupportsPP
 from .utils import (AutoWeightsLoader, PPMissingLayer, is_pp_missing_parameter,
-                    make_layers)
+                    make_layers, maybe_prefix)
 
 
 def _is_moe(config: PretrainedConfig) -> bool:
@@ -871,6 +871,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
                 org_num_embeddings=config.vocab_size,
                 padding_size=DEFAULT_VOCAB_PADDING_SIZE,
                 quant_config=quant_config,
+                prefix=maybe_prefix(prefix, "lm_head"),
             )
             if config.tie_word_embeddings:
                 self.lm_head.weight = self.model.embed_tokens.weight
diff --git a/vllm/model_executor/models/idefics3.py b/vllm/model_executor/models/idefics3.py
@@ -606,6 +606,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
             config.text_config.vocab_size,
             config.text_config.hidden_size,
             quant_config=quant_config,
+            prefix=maybe_prefix(prefix, "lm_head"),
         )
         if self.config.text_config.tie_word_embeddings:
             self.lm_head.weight = self.model.text_model.wte.weight
diff --git a/vllm/model_executor/models/jais.py b/vllm/model_executor/models/jais.py
@@ -302,7 +302,9 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
             self.lm_head = self.transformer.wte
         else:
             self.lm_head = ParallelLMHead(self.config.vocab_size,
-                                          self.config.hidden_size)
+                                          self.config.hidden_size,
+                                          prefix=maybe_prefix(
+                                              prefix, "lm_head"))
         if hasattr(config, "width_scale"):
             self.output_logits_scale = config.width_scale
         else:
diff --git a/vllm/model_executor/models/jamba.py b/vllm/model_executor/models/jamba.py
@@ -502,6 +502,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
             # We need bigger padding if using lora for kernel
             # compatibility
             if not lora_config else lora_config.lora_vocab_padding_size,
+            prefix=maybe_prefix(prefix, "lm_head"),
         )
         # Used to track and store by the Mamba cache between steps.
         self.mamba_cache: Optional[MambaCacheManager] = None
diff --git a/vllm/model_executor/models/kimi_vl.py b/vllm/model_executor/models/kimi_vl.py
@@ -328,6 +328,7 @@ def __init__(
                 config.text_config.hidden_size,
                 org_num_embeddings=self.config.text_config.vocab_size,
                 padding_size=DEFAULT_VOCAB_PADDING_SIZE,
+                prefix=maybe_prefix(prefix, "lm_head"),
             )
         else:
             self.lm_head = PPMissingLayer()
diff --git a/vllm/model_executor/models/llama_eagle3.py b/vllm/model_executor/models/llama_eagle3.py
@@ -220,7 +220,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
             self.config.hidden_size,
             org_num_embeddings=self.config.draft_vocab_size,
             padding_size=(DEFAULT_VOCAB_PADDING_SIZE),
-            prefix="")
+            prefix=maybe_prefix(prefix, "lm_head"))
         self.logits_processor = LogitsProcessor(self.config.draft_vocab_size,
                                                 scale=logit_scale)
         self.draft_id_to_target_id = nn.Parameter(
diff --git a/vllm/model_executor/models/mamba.py b/vllm/model_executor/models/mamba.py
@@ -223,6 +223,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
                 # We need bigger padding if using lora for kernel
                 # compatibility
                 if not lora_config else lora_config.lora_vocab_padding_size,
+                prefix=maybe_prefix(prefix, "lm_head"),
             )
 
         # Used to track and store by the Mamba cache between steps.
diff --git a/vllm/model_executor/models/mamba2.py b/vllm/model_executor/models/mamba2.py
@@ -278,6 +278,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
             # We need bigger padding if using lora for kernel
             # compatibility
             if not lora_config else lora_config.lora_vocab_padding_size,
+            prefix=maybe_prefix(prefix, "lm_head"),
         )
         if config.tie_word_embeddings:
             self.lm_head = self.lm_head.tie_weights(self.backbone.embeddings)
diff --git a/vllm/model_executor/models/medusa.py b/vllm/model_executor/models/medusa.py
@@ -15,6 +15,8 @@
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.model_executor.sampling_metadata import SamplingMetadata
 
+from .utils import maybe_prefix
+
 
 class ResidualBlock(nn.Module):
 
@@ -71,6 +73,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None:
                 config.hidden_size,
                 org_num_embeddings=self.truncated_vocab_size,
                 padding_size=DEFAULT_VOCAB_PADDING_SIZE,
+                prefix=maybe_prefix(prefix, "lm_head"),
             )
             self.lm_heads = [
                 self.lm_head for _ in range(self.config.num_heads)
diff --git a/vllm/model_executor/models/mimo_mtp.py b/vllm/model_executor/models/mimo_mtp.py
@@ -158,7 +158,8 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
                                              prefix=maybe_prefix(
                                                  prefix, "model"))
         self.lm_head = ParallelLMHead(self.config.vocab_size,
-                                      self.config.hidden_size)
+                                      self.config.hidden_size,
+                                      prefix=maybe_prefix(prefix, "lm_head"))
 
     def forward(
         self,
diff --git a/vllm/model_executor/models/minicpm.py b/vllm/model_executor/models/minicpm.py
@@ -547,6 +547,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
             # compatibility
             if not lora_config else lora_config.lora_vocab_padding_size,
             quant_config=quant_config,
+            prefix=maybe_prefix(prefix, "lm_head"),
         )
         if config.tie_word_embeddings:
             self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens)
diff --git a/vllm/model_executor/models/minicpm_eagle.py b/vllm/model_executor/models/minicpm_eagle.py
@@ -338,6 +338,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
             # compatibility
             if not lora_config else lora_config.lora_vocab_padding_size,
             quant_config=quant_config,
+            prefix=maybe_prefix(prefix, "lm_head"),
         )
         if config.tie_word_embeddings:
             self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens)
diff --git a/vllm/model_executor/models/minimax_text_01.py b/vllm/model_executor/models/minimax_text_01.py
@@ -702,6 +702,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None:
                 self.config.hidden_size,
                 org_num_embeddings=self.config.vocab_size,
                 padding_size=DEFAULT_VOCAB_PADDING_SIZE,
+                prefix=maybe_prefix(prefix, "lm_head"),
             )
 
             self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
diff --git a/vllm/model_executor/models/mixtral.py b/vllm/model_executor/models/mixtral.py
@@ -507,6 +507,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
             # compatibility
             if not lora_config else lora_config.lora_vocab_padding_size,
             quant_config=quant_config,
+            prefix=maybe_prefix(prefix, "lm_head"),
         )
         if self.config.tie_word_embeddings:
             self.lm_head.weight = self.model.embed_tokens.weight
diff --git a/vllm/model_executor/models/molmo.py b/vllm/model_executor/models/molmo.py
@@ -1403,6 +1403,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
                 config.embedding_size or config.vocab_size,
                 config.hidden_size,
                 quant_config=quant_config,
+                prefix=maybe_prefix(prefix, "lm_head"),
             )
 
         self.logits_processor = LogitsProcessor(config.embedding_size
diff --git a/vllm/model_executor/models/nemotron.py b/vllm/model_executor/models/nemotron.py
@@ -466,6 +466,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
                 # compatibility
                 if not lora_config else lora_config.lora_vocab_padding_size,
                 quant_config=quant_config,
+                prefix=maybe_prefix(prefix, "lm_head"),
             )
             if config.tie_word_embeddings:
                 self.lm_head.weight = self.model.embed_tokens.weight
diff --git a/vllm/model_executor/models/nemotron_h.py b/vllm/model_executor/models/nemotron_h.py
diff --git a/vllm/model_executor/models/olmo.py b/vllm/model_executor/models/olmo.py
diff --git a/vllm/model_executor/models/olmoe.py b/vllm/model_executor/models/olmoe.py
diff --git a/vllm/model_executor/models/opt.py b/vllm/model_executor/models/opt.py
diff --git a/vllm/model_executor/models/orion.py b/vllm/model_executor/models/orion.py
diff --git a/vllm/model_executor/models/persimmon.py b/vllm/model_executor/models/persimmon.py
diff --git a/vllm/model_executor/models/phi.py b/vllm/model_executor/models/phi.py
diff --git a/vllm/model_executor/models/phi4flash.py b/vllm/model_executor/models/phi4flash.py
diff --git a/vllm/model_executor/models/phi4mm.py b/vllm/model_executor/models/phi4mm.py
diff --git a/vllm/model_executor/models/phimoe.py b/vllm/model_executor/models/phimoe.py
diff --git a/vllm/model_executor/models/qwen.py b/vllm/model_executor/models/qwen.py
diff --git a/vllm/model_executor/models/qwen2_moe.py b/vllm/model_executor/models/qwen2_moe.py
diff --git a/vllm/model_executor/models/qwen3_moe.py b/vllm/model_executor/models/qwen3_moe.py
diff --git a/vllm/model_executor/models/qwen3_next.py b/vllm/model_executor/models/qwen3_next.py
diff --git a/vllm/model_executor/models/qwen3_next_mtp.py b/vllm/model_executor/models/qwen3_next_mtp.py
diff --git a/vllm/model_executor/models/solar.py b/vllm/model_executor/models/solar.py
diff --git a/vllm/model_executor/models/step3_text.py b/vllm/model_executor/models/step3_text.py
diff --git a/vllm/model_executor/models/zamba2.py b/vllm/model_executor/models/zamba2.py

Original file line number	Diff line number	Diff line change
`@@ -427,6 +427,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):`
`427`	`427`	`self.vocab_size,`
`428`	`428`	`config.hidden_size,`
`429`	`429`	`quant_config=quant_config,`
	`430`	`+ prefix=maybe_prefix(prefix, "lm_head"),`
`430`	`431`	`)`
`431`	`432`	`if self.config.tie_word_embeddings:`
`432`	`433`	`self.lm_head.weight = self.model.embed_tokens.weight`
Original file line number	Diff line number	Diff line change
`@@ -539,6 +539,7 @@ def __init__(`
`539`	`539`	`config.text_config.hidden_size,`
`540`	`540`	`org_num_embeddings=self.language_model.org_vocab_size,`
`541`	`541`	`quant_config=quant_config,`
	`542`	`+ prefix=maybe_prefix(prefix, "lm_head"),`
`542`	`543`	`)`
`543`	`544`	`logit_scale = getattr(config, "logit_scale", 1.0)`
`544`	`545`	`self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,`
Original file line number	Diff line number	Diff line change
`@@ -514,6 +514,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):`
`514`	`514`	`# We need bigger padding if using lora for kernel`
`515`	`515`	`# compatibility`
`516`	`516`	`if not lora_config else lora_config.lora_vocab_padding_size,`
	`517`	`+ prefix=maybe_prefix(prefix, "lm_head"),`
`517`	`518`	`)`
`518`	`519`	`# Used to track and store by the Mamba cache between steps.`
`519`	`520`	`self.mamba_cache: Optional[MambaCacheManager] = None`
Original file line number	Diff line number	Diff line change
`@@ -960,6 +960,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):`
`960`	`960`	`self.lm_head = ParallelLMHead(`
`961`	`961`	`self.unpadded_vocab_size,`
`962`	`962`	`config.hidden_size,`
	`963`	`+ prefix=maybe_prefix(prefix, "lm_head"),`
`963`	`964`	`)`
`964`	`965`	`if config.tie_word_embeddings:`
`965`	`966`	`self.lm_head.weight = self.model.embed_tokens.weight`
Original file line number	Diff line number	Diff line change
`@@ -438,6 +438,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):`
`438`	`438`	`org_num_embeddings=config.vocab_size,`
`439`	`439`	`padding_size=DEFAULT_VOCAB_PADDING_SIZE,`
`440`	`440`	`quant_config=quant_config,`
	`441`	`+ prefix=maybe_prefix(prefix, "lm_head"),`
`441`	`442`	`)`
`442`	`443`	`self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,`
`443`	`444`	`config.vocab_size)`