[Model] Fix missing maybe_prefix (#42280)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>

vllm/model_executor/models/arcee.py

@@ -45,6 +45,7 @@ from .utils import (
     is_pp_missing_parameter,
     make_empty_intermediate_tensors_factory,
     make_layers,
+    maybe_prefix,
 )
 
 
@@ -367,7 +368,10 @@ class ArceeForCausalLM(
         self.config = config
 
         # Initialize the inner Transformer model (ArceeModel)
-        self.model = ArceeModel(vllm_config=vllm_config, prefix=f"{prefix}.model")
+        self.model = ArceeModel(
+            vllm_config=vllm_config,
+            prefix=maybe_prefix(prefix, "model"),
+        )
         # On the last pipeline stage, set up the LM head and logits processor
         if get_pp_group().is_last_rank:
             # Determine vocabulary size (including any LoRA extra tokens
@@ -378,7 +382,7 @@ class ArceeForCausalLM(
                 config.hidden_size,
                 quant_config=vllm_config.quant_config,
                 bias=getattr(config, "lm_head_bias", False),
-                prefix=f"{prefix}.lm_head",
+                prefix=maybe_prefix(prefix, "lm_head"),
             )
             if config.tie_word_embeddings:
                 # Tie output weights with input embedding matrix

vllm/model_executor/models/aria.py

@@ -540,7 +540,7 @@ class AriaForConditionalGeneration(nn.Module, SupportsMultiModal):
             self.vision_tower = AriaVisionTransformer(
                 config.vision_config,
                 quant_config=quant_config,
-                prefix=f"{prefix}.vision_tower",
+                prefix=maybe_prefix(prefix, "vision_tower"),
             )
             self.multi_modal_projector = AriaProjector(
                 config, prefix=maybe_prefix(prefix, "multi_modal_projector")

vllm/model_executor/models/blip2.py

@@ -561,7 +561,7 @@ class Blip2ForConditionalGeneration(
                 config.qformer_config,
                 cache_config=cache_config,
                 quant_config=quant_config,
-                prefix=f"{prefix}.qformer",
+                prefix=maybe_prefix(prefix, "qformer"),
             )
             self.language_projection = nn.Linear(
                 config.qformer_config.hidden_size,

vllm/model_executor/models/clip.py

@@ -777,7 +777,7 @@ class CLIPVisionModel(nn.Module):
             quant_config=quant_config,
             num_hidden_layers_override=num_hidden_layers_override,
             require_post_norm=require_post_norm,
-            prefix=f"{prefix}.vision_model",
+            prefix=maybe_prefix(prefix, "vision_model"),
         )
 
     def forward(

vllm/model_executor/models/cohere_asr.py

@@ -64,7 +64,7 @@ from .interfaces import (
     SupportsMultiModal,
     SupportsTranscription,
 )
-from .utils import AutoWeightsLoader, WeightsMapper, make_layers
+from .utils import AutoWeightsLoader, WeightsMapper, make_layers, maybe_prefix
 
 logger = init_logger(__name__)
 
@@ -1717,7 +1717,8 @@ class CohereASRModel(nn.Module):
         self.encoder = ConformerEncoder(vllm_config=vllm_config)
 
         self.decoder = CohereASRDecoder(
-            vllm_config=vllm_config, prefix=f"{prefix}.decoder"
+            vllm_config=vllm_config,
+            prefix=maybe_prefix(prefix, "decoder"),
         )
 
         if self.encoder.d_model != self.decoder.hidden_size:

vllm/model_executor/models/deepseek_eagle.py

@@ -198,7 +198,9 @@ class EagleDeepseekV3ForCausalLM(DeepseekV3ForCausalLM):
             vllm_config.parallel_config
         )
         self.model = DeepseekV2Model(
-            vllm_config=vllm_config, prefix="model", start_layer_id=target_layer_num
+            vllm_config=vllm_config,
+            prefix=maybe_prefix(prefix, "model"),
+            start_layer_id=target_layer_num,
         )
 
         self.lm_head = ParallelLMHead(

vllm/model_executor/models/deepseek_eagle3.py

@@ -318,7 +318,9 @@ class Eagle3DeepseekV2ForCausalLM(DeepseekV2ForCausalLM):
         self.config.target_layer_count = target_layer_num
 
         self.model = DeepseekV2Eagle3Model(
-            vllm_config=vllm_config, prefix="model", start_layer_id=target_layer_num
+            vllm_config=vllm_config,
+            prefix=maybe_prefix(prefix, "model"),
+            start_layer_id=target_layer_num,
         )
 
         logit_scale = getattr(self.config, "logit_scale", 1.0)

vllm/model_executor/models/granite4_vision.py

@@ -149,7 +149,7 @@ class WindowQFormerDownsampler(nn.Module):
             qformer_config,
             quant_config=quant_config,
             cache_config=cache_config,
-            prefix=f"{prefix}.qformer",
+            prefix=maybe_prefix(prefix, "qformer"),
         )
 
         self.image_side = (

vllm/model_executor/models/granite_speech.py

@@ -620,7 +620,7 @@ class GraniteSpeechForConditionalGeneration(
             self.encoder = GraniteSpeechCTCEncoder(
                 config=config.encoder_config,
                 quant_config=quant_config,
-                prefix=f"{prefix}.encoder",
+                prefix=maybe_prefix(prefix, "encoder"),
             )
 
             # Blip2 QFormer
@@ -628,7 +628,7 @@ class GraniteSpeechForConditionalGeneration(
                 config=config,
                 quant_config=quant_config,
                 cache_config=cache_config,
-                prefix=f"{prefix}.projector",
+                prefix=maybe_prefix(prefix, "projector"),
             )
 
         self.make_empty_intermediate_tensors = (

vllm/model_executor/models/hunyuan_v1.py

@@ -930,7 +930,10 @@ class HunyuanV1ModelBase(
         self.config = config
         self.quant_config = quant_config
 
-        self.model = HunYuanModel(vllm_config=vllm_config, prefix="model")
+        self.model = HunYuanModel(
+            vllm_config=vllm_config,
+            prefix=maybe_prefix(prefix, "model"),
+        )
         if get_pp_group().is_last_rank:
             self.lm_head = ParallelLMHead(
                 config.vocab_size,

vllm/model_executor/models/keye.py

@@ -729,7 +729,7 @@ class KeyeSiglipVisionModel(nn.Module):
         self.vision_model = KeyeSiglipVisionTransformer(
             config,
             quant_config=quant_config,
-            prefix=f"{prefix}.vision_model",
+            prefix=maybe_prefix(prefix, "vision_model"),
         )
         self.quant_config = quant_config
 

vllm/model_executor/models/lfm2_siglip2.py

@@ -25,6 +25,7 @@ from vllm.model_executor.layers.linear import (
 from vllm.model_executor.layers.quantization import QuantizationConfig
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
+from .utils import maybe_prefix
 from .vision import (
     is_vit_use_data_parallel,
     resolve_visual_encoder_outputs,
@@ -472,7 +473,7 @@ class Siglip2Model(torch.nn.Module):
             quant_config=quant_config,
             num_hidden_layers_override=num_hidden_layers_override,
             require_post_norm=require_post_norm,
-            prefix=f"{prefix}.vision_model",
+            prefix=maybe_prefix(prefix, "vision_model"),
         )
 
     def forward(

vllm/model_executor/models/llama4_eagle.py

@@ -176,7 +176,7 @@ class EagleLlama4ForCausalLM(Llama4ForCausalLM):
         )
         self.model = LlamaModel(
             vllm_config=vllm_config,
-            prefix="model",
+            prefix=maybe_prefix(prefix, "model"),
             start_layer_id=target_layer_num,
             quant_config=quant_config,
         )

vllm/model_executor/models/llama_eagle.py

@@ -174,7 +174,9 @@ class EagleLlamaForCausalLM(LlamaForCausalLM):
             vllm_config.parallel_config
         )
         self.model = LlamaModel(
-            vllm_config=vllm_config, prefix="model", start_layer_id=target_layer_num
+            vllm_config=vllm_config,
+            prefix=maybe_prefix(prefix, "model"),
+            start_layer_id=target_layer_num,
         )
 
         logit_scale = getattr(self.config, "logit_scale", 1.0)

vllm/model_executor/models/llama_eagle3.py

@@ -287,7 +287,9 @@ class Eagle3LlamaForCausalLM(LlamaForCausalLM):
         # proper layer_types indexing in draft models
         self.config.target_layer_count = target_layer_num
         self.model = LlamaModel(
-            vllm_config=vllm_config, prefix="model", start_layer_id=target_layer_num
+            vllm_config=vllm_config,
+            prefix=maybe_prefix(prefix, "model"),
+            start_layer_id=target_layer_num,
         )
 
         logit_scale = getattr(self.config, "logit_scale", 1.0)

vllm/model_executor/models/mistral_eagle.py

@@ -128,7 +128,9 @@ class EagleMistralForCausalLM(MistralForCausalLM):
             vllm_config.parallel_config
         )
         self.model = EagleMistralModel(
-            vllm_config=vllm_config, prefix="model", start_layer_id=target_layer_num
+            vllm_config=vllm_config,
+            prefix=maybe_prefix(prefix, "model"),
+            start_layer_id=target_layer_num,
         )
 
         logit_scale = getattr(self.config, "logit_scale", 1.0)

vllm/model_executor/models/molmo2.py

@@ -489,7 +489,7 @@ class Molmo2VisionTransformer(nn.Module):
         self.transformer = Molmo2VisionBlockCollection(
             config,
             quant_config,
-            prefix=f"{prefix}.transformer",
+            prefix=maybe_prefix(prefix, "transformer"),
         )
 
     def add_pos_emb(self, x: torch.Tensor, patch_num: int) -> torch.Tensor:

vllm/model_executor/models/openpangu_vl.py

@@ -857,7 +857,7 @@ class OpenPanguVLForConditionalGeneration(
         with self._mark_language_model(vllm_config):
             self.language_model = init_vllm_registered_model(
                 vllm_config=vllm_config,
-                prefix=maybe_prefix("openpangu", "language_model"),
+                prefix=maybe_prefix(prefix, "openpangu.language_model"),
                 architectures=["PanguEmbeddedForCausalLM"],
             )
 

vllm/model_executor/models/ovis.py

@@ -96,7 +96,7 @@ class VisualTokenizer(torch.nn.Module):
         self.backbone = self._init_backbone(
             config=config,
             quant_config=quant_config,
-            prefix=f"{prefix}.backbone",
+            prefix=maybe_prefix(prefix, "backbone"),
         )
         # reserved tokens for IMAGE_INDICATORS
         head_dim = config.vocab_size - len(IMAGE_INDICATOR_IDS)
@@ -442,7 +442,7 @@ class Ovis(nn.Module, SupportsMultiModal, SupportsPP):
             self.visual_tokenizer = VisualTokenizer(
                 config=config.visual_tokenizer_config,
                 quant_config=quant_config,
-                prefix=f"{prefix}.visual_tokenizer",
+                prefix=maybe_prefix(prefix, "visual_tokenizer"),
             )
             self.vte = VisualEmbedding(
                 self.config.visual_tokenizer_config.vocab_size, self.config.hidden_size

vllm/model_executor/models/ovis2_5.py

@@ -465,7 +465,7 @@ class Ovis2_5(nn.Module, SupportsMultiModal, SupportsPP):
                 config=config.vit_config,
                 visual_vocab_size=config.visual_vocab_size,
                 quant_config=quant_config,
-                prefix=f"{prefix}.visual_tokenizer",
+                prefix=maybe_prefix(prefix, "visual_tokenizer"),
             )
             self.vte = VisualEmbedding(config.visual_vocab_size, config.hidden_size)
 

vllm/model_executor/models/paddleocr_vl.py

@@ -895,7 +895,7 @@ class SiglipVisionModel(nn.Module):
         self.vision_model = SiglipVisionTransformer(
             config,
             quant_config=quant_config,
-            prefix=f"{prefix}.vision_model",
+            prefix=maybe_prefix(prefix, "vision_model"),
         )
         self.quant_config = quant_config
 

vllm/model_executor/models/phi4mm.py

@@ -1034,7 +1034,7 @@ class Phi4MMForCausalLM(nn.Module, SupportsLoRA, SupportsMultiModal):
             self.vision_encoder = Phi4MMImageEncoder(
                 config,
                 quant_config,
-                prefix="model.vision_embed_tokens",
+                prefix=maybe_prefix(prefix, "model.vision_embed_tokens"),
                 model_dir=config._name_or_path,
             )
 

vllm/model_executor/models/qwen3_dflash.py

@@ -510,7 +510,7 @@ class DFlashQwen3ForCausalLM(Qwen3ForCausalLM):
         self.config.target_layer_count = target_layer_num
         self.model = DFlashQwen3Model(
             vllm_config=vllm_config,
-            prefix="model",
+            prefix=maybe_prefix(prefix, "model"),
             start_layer_id=target_layer_num,
         )
 

vllm/model_executor/models/siglip.py

@@ -867,7 +867,7 @@ class SiglipVisionModel(nn.Module):
             quant_config=quant_config,
             num_hidden_layers_override=num_hidden_layers_override,
             require_post_norm=require_post_norm,
-            prefix=f"{prefix}.vision_model",
+            prefix=maybe_prefix(prefix, "vision_model"),
             use_head=use_head,
         )
 

vllm/model_executor/models/siglip2navit.py

@@ -29,6 +29,7 @@ from vllm.model_executor.layers.rotary_embedding.common import (
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.platforms import current_platform
 
+from .utils import maybe_prefix
 from .vision import is_vit_use_data_parallel
 
 
@@ -598,7 +599,7 @@ class Siglip2NavitModel(torch.nn.Module):
         self.vision_model = Siglip2VisionTransformer(
             config,
             quant_config=quant_config,
-            prefix=f"{prefix}.vision_model",
+            prefix=maybe_prefix(prefix, "vision_model"),
         )
 
     def forward(