Fix base model prefix in VLMs

src/transformers/models/cohere2_vision/modeling_cohere2_vision.py

@@ -129,7 +129,7 @@ class Cohere2VisionCausalLMOutputWithPast(ModelOutput):
 @auto_docstring
 class Cohere2VisionPreTrainedModel(PreTrainedModel):
     config: Cohere2VisionConfig
-    base_model_prefix = ""
+    base_model_prefix = "model"
     input_modalities = ["image", "text"]
     supports_gradient_checkpointing = True
     _skip_keys_device_placement = "past_key_values"

src/transformers/models/cohere2_vision/modular_cohere2_vision.py

@@ -26,6 +26,7 @@
     AyaVisionForConditionalGeneration,
     AyaVisionModel,
     AyaVisionModelOutputWithPast,
+    AyaVisionPreTrainedModel,
 )
 from transformers.models.got_ocr2.image_processing_got_ocr2_fast import GotOcr2ImageProcessorFast
 
@@ -89,6 +90,10 @@ class Cohere2VisionCausalLMOutputWithPast(AyaVisionCausalLMOutputWithPast):
     pass
 
 
+class Cohere2VisionPreTrainedModel(AyaVisionPreTrainedModel):
+    base_model_prefix = "model"
+
+
 class Cohere2VisionModel(AyaVisionModel):
     _checkpoint_conversion_mapping = {}
 

src/transformers/models/florence2/modeling_florence2.py

@@ -615,7 +615,7 @@ class Florence2Seq2SeqLMOutput(Seq2SeqLMOutput):
 @auto_docstring
 class Florence2PreTrainedModel(PreTrainedModel):
     config: Florence2Config
-    base_model_prefix = ""
+    base_model_prefix = "model"
     input_modalities = ["image", "text"]
     supports_gradient_checkpointing = True
     _skip_keys_device_placement = "past_key_values"

src/transformers/models/florence2/modular_florence2.py

@@ -1500,6 +1500,7 @@ class Florence2Seq2SeqLMOutput(Seq2SeqLMOutput):
 @auto_docstring
 class Florence2PreTrainedModel(LlavaPreTrainedModel):
     config_class = Florence2Config
+    base_model_prefix = "model"
 
     _supports_attention_backend = False
 

src/transformers/models/gemma3n/modeling_gemma3n.py

@@ -1584,7 +1584,7 @@ def forward(
 @auto_docstring
 class Gemma3nPreTrainedModel(PreTrainedModel):
     config: Gemma3nConfig
-    base_model_prefix = ""
+    base_model_prefix = "model"
     supports_gradient_checkpointing = True
     _no_split_modules = ["Gemma3nTextDecoderLayer"]
     _skip_keys_device_placement = ["past_key_values"]

src/transformers/models/gemma3n/modular_gemma3n.py

@@ -1872,7 +1872,6 @@ def forward(
 
 class Gemma3nPreTrainedModel(Gemma2PreTrainedModel):
     config: Gemma3nConfig
-    base_model_prefix = ""
     input_modalities = ["image", "text", "audio"]
     _no_split_modules = ["Gemma3nTextDecoderLayer"]
 

src/transformers/models/lfm2_vl/modeling_lfm2_vl.py

@@ -76,7 +76,7 @@ def pixel_unshuffle(self, hidden_states: torch.Tensor):
 @auto_docstring
 class Lfm2VlPreTrainedModel(PreTrainedModel):
     config: Lfm2VlConfig
-    base_model_prefix = ""
+    base_model_prefix = "model"
     input_modalities = ["image", "text"]
     supports_gradient_checkpointing = True
     _skip_keys_device_placement = "past_key_values"

src/transformers/models/lfm2_vl/modular_lfm2_vl.py

@@ -75,6 +75,7 @@ def pixel_unshuffle(self, hidden_states: torch.Tensor):
 
 class Lfm2VlPreTrainedModel(LlavaPreTrainedModel):
     _can_compile_fullgraph = False
+    base_model_prefix = "model"
 
 
 class Lfm2VlCausalLMOutputWithPast(LlavaCausalLMOutputWithPast):

src/transformers/models/qwen3_vl/modeling_qwen3_vl.py

@@ -915,7 +915,7 @@ def _deepstack_process(
 
 @auto_docstring
 class Qwen3VLModel(Qwen3VLPreTrainedModel):
-    base_model_prefix = ""
+    base_model_prefix = "model"
     _checkpoint_conversion_mapping = {}
     # Reference: fix gemma3 grad acc #37208
     accepts_loss_kwargs = False

src/transformers/models/qwen3_vl/modular_qwen3_vl.py

@@ -804,6 +804,7 @@ def forward(
 @auto_docstring
 class Qwen3VLModel(Qwen2_5_VLModel):
     config: Qwen3VLConfig
+    base_model_prefix = "model"
     _checkpoint_conversion_mapping = {}
     _no_split_modules = ["Qwen3VLTextDecoderLayer", "Qwen3VLVisionBlock"]
 

src/transformers/models/qwen3_vl_moe/modeling_qwen3_vl_moe.py

@@ -1049,7 +1049,7 @@ class Qwen3VLMoeModelOutputWithPast(ModelOutput):
 
 @auto_docstring
 class Qwen3VLMoeModel(Qwen3VLMoePreTrainedModel):
-    base_model_prefix = ""
+    base_model_prefix = "model"
     _checkpoint_conversion_mapping = {}
     # Reference: fix gemma3 grad acc #37208
     accepts_loss_kwargs = False