Generate: PT Dynamo without graph breaks in the main greedy/sample lo…

…op (#21648)

src/transformers/generation/configuration_utils.py

@@ -298,6 +298,9 @@ def __init__(self, **kwargs):
         self.validate()
 
     def __eq__(self, other):
+        if not isinstance(other, GenerationConfig):
+            return False
+
         self_dict = self.__dict__.copy()
         other_dict = other.__dict__.copy()
         # ignore metadata

src/transformers/modeling_utils.py

@@ -190,14 +190,14 @@ def get_parameter_dtype(parameter: Union[nn.Module, GenerationMixin, "ModuleUtil
             # Adding fix for https://github.com/pytorch/xla/issues/4152
             # Fixes issue where the model code passes a value that is out of range for XLA_USE_BF16=1
             # and XLA_DOWNCAST_BF16=1 so the conversion would cast it to -inf
-            if is_torch_tpu_available():
-                if XLA_USE_BF16 in ENV_VARS_TRUE_VALUES:
+            # NOTE: `is_torch_tpu_available()` is checked last as it induces a graph break in torch dynamo
+            if XLA_USE_BF16 in ENV_VARS_TRUE_VALUES and is_torch_tpu_available():
+                return torch.bfloat16
+            if XLA_DOWNCAST_BF16 in ENV_VARS_TRUE_VALUES and is_torch_tpu_available():
+                if t.dtype == torch.float:
                     return torch.bfloat16
-                if XLA_DOWNCAST_BF16 in ENV_VARS_TRUE_VALUES:
-                    if t.dtype == torch.float:
-                        return torch.bfloat16
-                    if t.dtype == torch.double:
-                        return torch.float32
+                if t.dtype == torch.double:
+                    return torch.float32
             return t.dtype
 
     if last_dtype is not None: