fix qwen25-vl grad acc (huggingface#40333)

* fix qwen25—vl grad acc

* fix Qwen2_5_VLForConditionalGeneration for accepts_loss_kwargs

* fix ci

* fix ci

* fix typo

* fix CI

Author: iMountTai

src/transformers/models/glm4v/modeling_glm4v.py

@@ -893,6 +893,8 @@ def forward(
 class Glm4vModel(Glm4vPreTrainedModel):
     base_model_prefix = ""
     _checkpoint_conversion_mapping = {}
+    # Reference: fix gemma3 grad acc #37208
+    accepts_loss_kwargs = False
     config: Glm4vConfig
     _no_split_modules = ["Glm4vTextDecoderLayer", "Glm4vVisionBlock"]
 
@@ -1329,6 +1331,8 @@ class Glm4vCausalLMOutputWithPast(ModelOutput):
 class Glm4vForConditionalGeneration(Glm4vPreTrainedModel, GenerationMixin):
     _checkpoint_conversion_mapping = {}
     _tied_weights_keys = ["lm_head.weight"]
+    # Reference: fix gemma3 grad acc #37208
+    accepts_loss_kwargs = False
 
     def __init__(self, config):
         super().__init__(config)

src/transformers/models/glm4v_moe/modeling_glm4v_moe.py

@@ -1009,6 +1009,8 @@ def forward(
 class Glm4vMoeModel(Glm4vMoePreTrainedModel):
     base_model_prefix = ""
     _checkpoint_conversion_mapping = {}
+    # Reference: fix gemma3 grad acc #37208
+    accepts_loss_kwargs = False
     config: Glm4vMoeConfig
     _no_split_modules = ["Glm4vMoeTextDecoderLayer", "Glm4vMoeVisionBlock"]
 
@@ -1445,6 +1447,8 @@ class Glm4vMoeCausalLMOutputWithPast(ModelOutput):
 class Glm4vMoeForConditionalGeneration(Glm4vMoePreTrainedModel, GenerationMixin):
     _checkpoint_conversion_mapping = {}
     _tied_weights_keys = ["lm_head.weight"]
+    # Reference: fix gemma3 grad acc #37208
+    accepts_loss_kwargs = False
 
     def __init__(self, config):
         super().__init__(config)

src/transformers/models/qwen2_5_vl/modeling_qwen2_5_vl.py

@@ -938,6 +938,8 @@ def forward(
 class Qwen2_5_VLModel(Qwen2_5_VLPreTrainedModel):
     base_model_prefix = ""
     _checkpoint_conversion_mapping = {"^model": "language_model"}
+    # Reference: fix gemma3 grad acc #37208
+    accepts_loss_kwargs = False
     config: Qwen2_5_VLConfig
     _no_split_modules = ["Qwen2_5_VLDecoderLayer", "Qwen2_5_VLVisionBlock"]
 
@@ -1368,6 +1370,8 @@ class Qwen2_5_VLForConditionalGeneration(Qwen2_5_VLPreTrainedModel, GenerationMi
         r"^model(?!\.(language_model|visual))": "model.language_model",
     }
     _tied_weights_keys = ["lm_head.weight"]
+    # Reference: fix gemma3 grad acc #37208
+    accepts_loss_kwargs = False
 
     def __init__(self, config):
         super().__init__(config)

src/transformers/models/qwen2_5_vl/modular_qwen2_5_vl.py

@@ -346,6 +346,8 @@ class Qwen2_5_VLModel(Qwen2VLModel):
     config: Qwen2_5_VLConfig
     base_model_prefix = ""
     _no_split_modules = ["Qwen2_5_VLDecoderLayer", "Qwen2_5_VLVisionBlock"]
+    # Reference: fix gemma3 grad acc #37208
+    accepts_loss_kwargs = False
 
     def __init__(self, config):
         super().__init__(config)
@@ -651,6 +653,9 @@ class Qwen2_5_VLCausalLMOutputWithPast(Qwen2VLCausalLMOutputWithPast):
 
 
 class Qwen2_5_VLForConditionalGeneration(Qwen2VLForConditionalGeneration):
+    # Reference: fix gemma3 grad acc #37208
+    accepts_loss_kwargs = False
+
     def forward(
         self,
         input_ids: torch.LongTensor = None,

src/transformers/models/qwen2_vl/modeling_qwen2_vl.py

@@ -911,6 +911,8 @@ def forward(
 class Qwen2VLModel(Qwen2VLPreTrainedModel):
     base_model_prefix = ""
     _checkpoint_conversion_mapping = {"^model": "language_model"}
+    # Reference: fix gemma3 grad acc #37208
+    accepts_loss_kwargs = False
 
     def __init__(self, config: Qwen2VLConfig):
         super().__init__(config)