Keep relevant weights in fp32 when `model._keep_in_fp32_modules` is set even...

Keep relevant weights in fp32 when `model._keep_in_fp32_modules` is set even when `accelerate` is not installed (#26225)

* fix bug where weight would not be kept in fp32

* nit

* address review comments

* fix test

src/transformers/modeling_utils.py

@@ -2950,26 +2950,18 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
                 dtype_orig = cls._set_default_torch_dtype(torch_dtype)
 
             # Check if `_keep_in_fp32_modules` is not None
-            use_keep_in_fp32_modules = (
-                (cls._keep_in_fp32_modules is not None)
-                and is_accelerate_available()
-                and (torch_dtype == torch.float16 or load_in_4bit or load_in_8bit)
-            )
-            if (
-                (cls._keep_in_fp32_modules is not None)
-                and not is_accelerate_available()
-                and torch_dtype == torch.float16
-            ):
-                logger.warning(
-                    "For stability purposes, it is recommended to have accelerate installed when using this model in"
-                    " torch.float16, please install it with `pip install accelerate`"
+            use_keep_in_fp32_modules = (cls._keep_in_fp32_modules is not None) and (
+                torch_dtype == torch.float16 or load_in_4bit or load_in_8bit
             )
 
             if is_sharded:
                 loaded_state_dict_keys = sharded_metadata["all_checkpoint_keys"]
             else:
                 loaded_state_dict_keys = list(state_dict.keys())
-            if low_cpu_mem_usage or use_keep_in_fp32_modules:
+            if low_cpu_mem_usage or (use_keep_in_fp32_modules and is_accelerate_available()):
+                # In case some weights need to be kept in float32 and accelerate is not installed,
+                # we later on want to take the path where state_dict is not None, that is the one
+                # that do not require accelerate.
                 state_dict = None
 
         config.name_or_path = pretrained_model_name_or_path
@@ -2990,6 +2982,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
 
         # Check first if we are `from_pt`
         if use_keep_in_fp32_modules:
+            if is_accelerate_available():
                 low_cpu_mem_usage = True
             keep_in_fp32_modules = model._keep_in_fp32_modules
         else:
@@ -3465,7 +3458,8 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
         if keep_in_fp32_modules is not None:
             for name, param in model.named_parameters():
                 if any(module_to_keep_in_fp32 in name for module_to_keep_in_fp32 in keep_in_fp32_modules):
-                    param = param.to(torch.float32)
+                    # param = param.to(torch.float32) does not work here as only in the local scope.
+                    param.data = param.data.to(torch.float32)
 
         # Make sure we are able to load base models as well as derived models (with heads)
         start_prefix = ""
@@ -3592,7 +3586,6 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
                     remove_prefix_from_model,
                     ignore_mismatched_sizes,
                 )
-
                 if low_cpu_mem_usage:
                     if not is_fsdp_enabled() or is_fsdp_enabled_and_dist_rank_0():
                         new_error_msgs, offload_index, state_dict_index = _load_state_dict_into_meta_model(

tests/models/t5/test_modeling_t5.py

@@ -1046,7 +1046,22 @@ class T5ModelFp16Tests(unittest.TestCase):
         r"""
         A test to check whether the argument `keep_in_fp32_modules` correctly does its job
         """
+        orig_import = __import__
+        accelerate_mock = unittest.mock.Mock()
+
+        # mock import of accelerate
+        def import_accelerate_mock(name, *args, **kwargs):
+            if name == "accelerate":
+                if accelerate_available:
+                    return accelerate_mock
+                else:
+                    raise ImportError
+            return orig_import(name, *args, **kwargs)
+
         # Load without using `accelerate`
+        with unittest.mock.patch("builtins.__import__", side_effect=import_accelerate_mock):
+            accelerate_available = False
+
             model = T5ForConditionalGeneration.from_pretrained("t5-small", torch_dtype=torch.float16)
             self.assertTrue(model.decoder.block[0].layer[2].DenseReluDense.wo.weight.dtype == torch.float32)
             self.assertTrue(model.decoder.block[0].layer[2].DenseReluDense.wi.weight.dtype == torch.float16)