[fix gemma] Set default value for output_attentions parameter in Gemma2 and Gemma… (#37633)

* Set default value for output_attentions parameter in Gemma2 and Gemma3 models

* update

* fix

* fix

---------

Co-authored-by: chenin <wangzhichen@encosmart.com>

src/transformers/models/cohere2/modeling_cohere2.py

@@ -660,7 +660,7 @@ class Cohere2Model(Cohere2PreTrainedModel):
         input_tensor: torch.Tensor,
         cache_position: torch.Tensor,
         past_key_values: HybridCache,
-        output_attentions: bool,
+        output_attentions: bool = False,
     ):
         # Flash Attention currently doesn't support static cache but Cohere2 work only with static cache.
         # So we will pass in attention mask as is in any case, not only when ther's padding. Then we'll use its shape

src/transformers/models/gemma2/modeling_gemma2.py

@@ -673,7 +673,7 @@ class Gemma2Model(Gemma2PreTrainedModel):
         input_tensor: torch.Tensor,
         cache_position: torch.Tensor,
         past_key_values: HybridCache,
-        output_attentions: bool,
+        output_attentions: bool = False,
     ):
         # Flash Attention currently doesn't support static cache but Gemma2 work only with static cache.
         # So we will pass in attention mask as is in any case, not only when ther's padding. Then we'll use its shape

src/transformers/models/gemma2/modular_gemma2.py

@@ -540,7 +540,7 @@ class Gemma2Model(GemmaModel):
         input_tensor: torch.Tensor,
         cache_position: torch.Tensor,
         past_key_values: HybridCache,
-        output_attentions: bool,
+        output_attentions: bool = False,
     ):
         # Flash Attention currently doesn't support static cache but Gemma2 work only with static cache.
         # So we will pass in attention mask as is in any case, not only when ther's padding. Then we'll use its shape

src/transformers/models/gemma3/modeling_gemma3.py

@@ -758,7 +758,7 @@ class Gemma3TextModel(Gemma3PreTrainedModel):
         input_tensor: torch.Tensor,
         cache_position: torch.Tensor,
         past_key_values: HybridCache,
-        output_attentions: bool,
+        output_attentions: bool = False,
     ):
         # Flash Attention currently doesn't support static cache but Gemma3Text work only with static cache.
         # So we will pass in attention mask as is in any case, not only when ther's padding. Then we'll use its shape