Fix incorrect comments about atten mask for pytorch backend (#18728)

* fix incorrect comments about atten mask

* typo

* Update for CodeGen
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

src/transformers/modeling_utils.py

@@ -756,7 +756,7 @@ class ModuleUtilsMixin:
 
         # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
         # masked positions, this operation will create a tensor which is 0.0 for
-        # positions we want to attend and -10000.0 for masked positions.
+        # positions we want to attend and the dtype's smallest value for masked positions.
         # Since we are adding it to the raw scores before the softmax, this is
         # effectively the same as removing these entirely.
         extended_attention_mask = extended_attention_mask.to(dtype=dtype)  # fp16 compatibility

src/transformers/models/canine/modeling_canine.py

@@ -466,7 +466,7 @@ class CanineSelfAttention(nn.Module):
                 attention_mask = torch.unsqueeze(attention_mask, dim=1)
                 # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
                 # masked positions, this operation will create a tensor which is 0.0 for
-                # positions we want to attend and -10000.0 for masked positions.
+                # positions we want to attend and the dtype's smallest value for masked positions.
                 attention_mask = (1.0 - attention_mask.float()) * torch.finfo(attention_scores.dtype).min
             # Apply the attention mask (precomputed for all layers in CanineModel forward() function)
             attention_scores = attention_scores + attention_mask

src/transformers/models/codegen/modeling_codegen.py

@@ -518,11 +518,11 @@ class CodeGenModel(CodeGenPreTrainedModel):
 
             # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
             # masked positions, this operation will create a tensor which is 0.0 for
-            # positions we want to attend and -10000.0 for masked positions.
+            # positions we want to attend and the dtype's smallest value for masked positions.
             # Since we are adding it to the raw scores before the softmax, this is
             # effectively the same as removing these entirely.
             attention_mask = attention_mask.to(dtype=self.dtype)  # fp16 compatibility
-            attention_mask = (1.0 - attention_mask) * -10000.0
+            attention_mask = (1.0 - attention_mask) * torch.finfo(self.dtype).min
 
         # Prepare head mask if needed
         # 1.0 in head_mask indicate we keep the head

src/transformers/models/ctrl/modeling_ctrl.py

@@ -431,7 +431,7 @@ class CTRLModel(CTRLPreTrainedModel):
 
             # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
             # masked positions, this operation will create a tensor which is 0.0 for
-            # positions we want to attend and -10000.0 for masked positions.
+            # positions we want to attend and the dtype's smallest value for masked positions.
             # Since we are adding it to the raw scores before the softmax, this is
             # effectively the same as removing these entirely.
             attention_mask = attention_mask.to(dtype=self.dtype)  # fp16 compatibility

src/transformers/models/decision_transformer/modeling_decision_transformer.py

@@ -571,7 +571,7 @@ class DecisionTransformerGPT2Model(DecisionTransformerGPT2PreTrainedModel):
 
             # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
             # masked positions, this operation will create a tensor which is 0.0 for
-            # positions we want to attend and -10000.0 for masked positions.
+            # positions we want to attend and the dtype's smallest value for masked positions.
             # Since we are adding it to the raw scores before the softmax, this is
             # effectively the same as removing these entirely.
             attention_mask = attention_mask.to(dtype=self.dtype)  # fp16 compatibility

src/transformers/models/gpt2/modeling_gpt2.py

@@ -805,7 +805,7 @@ class GPT2Model(GPT2PreTrainedModel):
 
             # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
             # masked positions, this operation will create a tensor which is 0.0 for
-            # positions we want to attend and -10000.0 for masked positions.
+            # positions we want to attend and the dtype's smallest value for masked positions.
             # Since we are adding it to the raw scores before the softmax, this is
             # effectively the same as removing these entirely.
             attention_mask = attention_mask.to(dtype=self.dtype)  # fp16 compatibility

src/transformers/models/gpt_neo/modeling_gpt_neo.py

@@ -565,7 +565,7 @@ class GPTNeoModel(GPTNeoPreTrainedModel):
 
             # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
             # masked positions, this operation will create a tensor which is 0.0 for
-            # positions we want to attend and -10000.0 for masked positions.
+            # positions we want to attend and the dtype's smallest value for masked positions.
             # Since we are adding it to the raw scores before the softmax, this is
             # effectively the same as removing these entirely.
             attention_mask = attention_mask.to(dtype=self.dtype)  # fp16 compatibility

src/transformers/models/gpt_neox/modeling_gpt_neox.py

@@ -484,7 +484,7 @@ class GPTNeoXModel(GPTNeoXPreTrainedModel):
 
             # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
             # masked positions, this operation will create a tensor which is 0.0 for
-            # positions we want to attend and -10000.0 for masked positions.
+            # positions we want to attend and the dtype's smallest value for masked positions.
             # Since we are adding it to the raw scores before the softmax, this is
             # effectively the same as removing these entirely.
             attention_mask = attention_mask.to(dtype=self.dtype)  # fp16 compatibility

src/transformers/models/gptj/modeling_gptj.py

@@ -606,7 +606,7 @@ class GPTJModel(GPTJPreTrainedModel):
 
             # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
             # masked positions, this operation will create a tensor which is 0.0 for
-            # positions we want to attend and -10000.0 for masked positions.
+            # positions we want to attend and the dtype's smallest value for masked positions.
             # Since we are adding it to the raw scores before the softmax, this is
             # effectively the same as removing these entirely.
             attention_mask = attention_mask.to(dtype=self.dtype)  # fp16 compatibility

src/transformers/models/imagegpt/modeling_imagegpt.py

@@ -757,7 +757,7 @@ class ImageGPTModel(ImageGPTPreTrainedModel):
 
             # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
             # masked positions, this operation will create a tensor which is 0.0 for
-            # positions we want to attend and -10000.0 for masked positions.
+            # positions we want to attend and the dtype's smallest value for masked positions.
             # Since we are adding it to the raw scores before the softmax, this is
             # effectively the same as removing these entirely.
             attention_mask = attention_mask.to(dtype=self.dtype)  # fp16 compatibility

src/transformers/models/lxmert/modeling_lxmert.py

@@ -955,7 +955,7 @@ class LxmertModel(LxmertPreTrainedModel):
 
         # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
         # masked positions, this operation will create a tensor which is 0.0 for
-        # positions we want to attend and -10000.0 for masked positions.
+        # positions we want to attend and the dtype's smallest value for masked positions.
         # Since we are adding it to the raw scores before the softmax, this is
         # effectively the same as removing these entirely.
         extended_attention_mask = extended_attention_mask.to(dtype=self.dtype)

src/transformers/models/openai/modeling_openai.py

@@ -475,7 +475,7 @@ class OpenAIGPTModel(OpenAIGPTPreTrainedModel):
 
             # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
             # masked positions, this operation will create a tensor which is 0.0 for
-            # positions we want to attend and -10000.0 for masked positions.
+            # positions we want to attend and the dtype's smallest value for masked positions.
             # Since we are adding it to the raw scores before the softmax, this is
             # effectively the same as removing these entirely.
             attention_mask = attention_mask.to(dtype=next(self.parameters()).dtype)  # fp16 compatibility