torch.dtype alias

src/diffusers/loaders/single_file.py

@@ -281,7 +281,7 @@ def from_single_file(cls, pretrained_model_link_or_path, **kwargs) -> Self:
                     - A link to the `.ckpt` file (for example
                       `"https://huggingface.co/<repo_id>/blob/main/<path_to_file>.ckpt"`) on the Hub.
                     - A path to a *file* containing all pipeline weights.
-            torch_dtype (`str` or `torch.dtype`, *optional*):
+            dtype (`str` or `torch.dtype`, *optional*):
                 Override the default `torch.dtype` and load the model with another dtype.
             force_download (`bool`, *optional*, defaults to `False`):
                 Whether or not to force the (re-)download of the model weights and configuration files, overriding the
@@ -361,6 +361,8 @@ def from_single_file(cls, pretrained_model_link_or_path, **kwargs) -> Self:
         local_files_only = kwargs.pop("local_files_only", False)
         revision = kwargs.pop("revision", None)
         torch_dtype = kwargs.pop("torch_dtype", None)
+        dtype = kwargs.pop("dtype", None)
+        torch_dtype = dtype if dtype is not None else torch_dtype
         disable_mmap = kwargs.pop("disable_mmap", False)
 
         is_legacy_loading = False

src/diffusers/loaders/single_file_model.py

@@ -277,7 +277,7 @@ def from_single_file(cls, pretrained_model_link_or_path_or_dict: str | None = No
             original_config (`str`, *optional*):
                 Dict or path to a yaml file containing the configuration for the model in its original format.
                     If a dict is provided, it will be used to initialize the model configuration.
-            torch_dtype (`torch.dtype`, *optional*):
+            dtype (`torch.dtype`, *optional*):
                 Override the default `torch.dtype` and load the model with another dtype.
             force_download (`bool`, *optional*, defaults to `False`):
                 Whether or not to force the (re-)download of the model weights and configuration files, overriding the
@@ -351,6 +351,8 @@ def from_single_file(cls, pretrained_model_link_or_path_or_dict: str | None = No
         revision = kwargs.pop("revision", None)
         config_revision = kwargs.pop("config_revision", None)
         torch_dtype = kwargs.pop("torch_dtype", None)
+        dtype = kwargs.pop("dtype", None)
+        torch_dtype = dtype if dtype is not None else torch_dtype
         quantization_config = kwargs.pop("quantization_config", None)
         low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT)
         device = kwargs.pop("device", None)

src/diffusers/models/adapter.py

@@ -166,7 +166,7 @@ def from_pretrained(cls, pretrained_model_path: str | os.PathLike | None, **kwar
             pretrained_model_path (`os.PathLike`):
                 A path to a *directory* containing model weights saved using
                 [`~diffusers.models.adapter.MultiAdapter.save_pretrained`], e.g., `./my_model_directory/adapter`.
-            torch_dtype (`torch.dtype`, *optional*):
+            dtype (`torch.dtype`, *optional*):
                 Override the default `torch.dtype` and load the model under this dtype.
             output_loading_info(`bool`, *optional*, defaults to `False`):
                 Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.

src/diffusers/models/auto_model.py

@@ -173,7 +173,7 @@ def from_pretrained(cls, pretrained_model_or_path: str | os.PathLike | None = No
             cache_dir (`str | os.PathLike`, *optional*):
                 Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
                 is not used.
-            torch_dtype (`torch.dtype`, *optional*):
+            dtype (`torch.dtype`, *optional*):
                 Override the default `torch.dtype` and load the model with another dtype.
             force_download (`bool`, *optional*, defaults to `False`):
                 Whether or not to force the (re-)download of the model weights and configuration files, overriding the

src/diffusers/models/controlnets/multicontrolnet.py

@@ -163,7 +163,7 @@ def from_pretrained(cls, pretrained_model_path: str | os.PathLike | None, **kwar
                 A path to a *directory* containing model weights saved using
                 [`~models.controlnets.multicontrolnet.MultiControlNetModel.save_pretrained`], e.g.,
                 `./my_model_directory/controlnet`.
-            torch_dtype (`torch.dtype`, *optional*):
+            dtype (`torch.dtype`, *optional*):
                 Override the default `torch.dtype` and load the model under this dtype.
             output_loading_info(`bool`, *optional*, defaults to `False`):
                 Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.

src/diffusers/models/controlnets/multicontrolnet_union.py

@@ -180,7 +180,7 @@ def from_pretrained(cls, pretrained_model_path: str | os.PathLike | None, **kwar
                 A path to a *directory* containing model weights saved using
                 [`~models.controlnets.multicontrolnet.MultiControlNetUnionModel.save_pretrained`], e.g.,
                 `./my_model_directory/controlnet`.
-            torch_dtype (`torch.dtype`, *optional*):
+            dtype (`torch.dtype`, *optional*):
                 Override the default `torch.dtype` and load the model under this dtype.
             output_loading_info(`bool`, *optional*, defaults to `False`):
                 Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.

src/diffusers/models/modeling_utils.py

@@ -892,7 +892,7 @@ def from_pretrained(cls, pretrained_model_name_or_path: str | os.PathLike | None
             cache_dir (`str | os.PathLike`, *optional*):
                 Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
                 is not used.
-            torch_dtype (`torch.dtype`, *optional*):
+            dtype (`torch.dtype`, *optional*):
                 Override the default `torch.dtype` and load the model with another dtype.
             force_download (`bool`, *optional*, defaults to `False`):
                 Whether or not to force the (re-)download of the model weights and configuration files, overriding the
@@ -1018,6 +1018,8 @@ def from_pretrained(cls, pretrained_model_name_or_path: str | os.PathLike | None
         token = kwargs.pop("token", None)
         revision = kwargs.pop("revision", None)
         torch_dtype = kwargs.pop("torch_dtype", None)
+        dtype = kwargs.pop("dtype", None)
+        torch_dtype = dtype if dtype is not None else torch_dtype
         subfolder = kwargs.pop("subfolder", None)
         device_map = kwargs.pop("device_map", None)
         max_memory = kwargs.pop("max_memory", None)

src/diffusers/modular_pipelines/modular_pipeline_utils.py

@@ -311,10 +311,11 @@ def load(self, **kwargs) -> Any:
 
         from diffusers import AutoModel
 
-        # `torch_dtype` is not an accepted parameter for tokenizers and processors.
+        # `dtype`/`torch_dtype` is not an accepted parameter for tokenizers and processors.
         # As a result, it gets stored in `init_kwargs`, which are written to the config
         # during save. This causes JSON serialization to fail when saving the component.
         if self.type_hint is not None and not issubclass(self.type_hint, (torch.nn.Module, AutoModel)):
+            kwargs.pop("dtype", None)
             kwargs.pop("torch_dtype", None)
 
         if self.type_hint is None:

src/diffusers/pipelines/auto_pipeline.py

@@ -426,7 +426,7 @@ def from_pretrained(cls, pretrained_model_or_path, **kwargs):
                     - A path to a *directory* (for example `./my_pipeline_directory/`) containing pipeline weights
                       saved using
                     [`~DiffusionPipeline.save_pretrained`].
-            torch_dtype (`torch.dtype`, *optional*):
+            dtype (`torch.dtype`, *optional*):
                 Override the default `torch.dtype` and load the model with another dtype.
             force_download (`bool`, *optional*, defaults to `False`):
                 Whether or not to force the (re-)download of the model weights and configuration files, overriding the
@@ -717,7 +717,7 @@ def from_pretrained(cls, pretrained_model_or_path, **kwargs):
                     - A path to a *directory* (for example `./my_pipeline_directory/`) containing pipeline weights
                       saved using
                     [`~DiffusionPipeline.save_pretrained`].
-            torch_dtype (`str` or `torch.dtype`, *optional*):
+            dtype (`str` or `torch.dtype`, *optional*):
                 Override the default `torch.dtype` and load the model with another dtype.
             force_download (`bool`, *optional*, defaults to `False`):
                 Whether or not to force the (re-)download of the model weights and configuration files, overriding the
@@ -1022,7 +1022,7 @@ def from_pretrained(cls, pretrained_model_or_path, **kwargs):
                     - A path to a *directory* (for example `./my_pipeline_directory/`) containing pipeline weights
                       saved using
                     [`~DiffusionPipeline.save_pretrained`].
-            torch_dtype (`str` or `torch.dtype`, *optional*):
+            dtype (`str` or `torch.dtype`, *optional*):
                 Override the default `torch.dtype` and load the model with another dtype.
             force_download (`bool`, *optional*, defaults to `False`):
                 Whether or not to force the (re-)download of the model weights and configuration files, overriding the
@@ -1312,7 +1312,7 @@ def from_pretrained(cls, pretrained_model_or_path, **kwargs):
                     - A path to a *directory* (for example `./my_pipeline_directory/`) containing pipeline weights
                       saved using
                     [`~DiffusionPipeline.save_pretrained`].
-            torch_dtype (`torch.dtype`, *optional*):
+            dtype (`torch.dtype`, *optional*):
                 Override the default `torch.dtype` and load the model with another dtype.
             force_download (`bool`, *optional*, defaults to `False`):
                 Whether or not to force the (re-)download of the model weights and configuration files, overriding the

src/diffusers/pipelines/pipeline_utils.py

@@ -639,7 +639,7 @@ def from_pretrained(cls, pretrained_model_name_or_path: str | os.PathLike, **kwa
                       saved using
                     [`~DiffusionPipeline.save_pretrained`].
                     - A path to a *directory* (for example `./my_pipeline_directory/`) containing a dduf file
-            torch_dtype (`torch.dtype` or `dict[str, Union[str, torch.dtype]]`, *optional*):
+            dtype (`torch.dtype` or `dict[str, Union[str, torch.dtype]]`, *optional*):
                 Override the default `torch.dtype` and load the model with another dtype. To load submodels with
                 different dtype pass a `dict` (for example `{'transformer': torch.bfloat16, 'vae': torch.float16}`).
                 Set the default dtype for unspecified components with `default` (for example `{'transformer':
@@ -769,6 +769,8 @@ def from_pretrained(cls, pretrained_model_name_or_path: str | os.PathLike, **kwa
         revision = kwargs.pop("revision", None)
         from_flax = kwargs.pop("from_flax", False)
         torch_dtype = kwargs.pop("torch_dtype", None)
+        dtype = kwargs.pop("dtype", None)
+        torch_dtype = dtype if dtype is not None else torch_dtype
         custom_pipeline = kwargs.pop("custom_pipeline", None)
         custom_revision = kwargs.pop("custom_revision", None)
         provider = kwargs.pop("provider", None)
@@ -2121,7 +2123,9 @@ def from_pipe(cls, pipeline, **kwargs):
         """
 
         original_config = dict(pipeline.config)
-        torch_dtype = kwargs.pop("torch_dtype", torch.float32)
+        torch_dtype = kwargs.pop("torch_dtype", None)
+        dtype = kwargs.pop("dtype", None)
+        torch_dtype = dtype if dtype is not None else (torch_dtype if torch_dtype is not None else torch.float32)
         trust_remote_code = kwargs.pop("trust_remote_code", False)
 
         # derive the pipeline class to instantiate

tests/models/testing_utils/common.py

@@ -379,6 +379,21 @@ def test_from_save_pretrained_dtype(self, tmp_path, dtype):
             new_model = self.model_class.from_pretrained(tmp_path, low_cpu_mem_usage=False, torch_dtype=dtype)
             assert new_model.dtype == dtype
 
+    @pytest.mark.parametrize("dtype", [torch.float32, torch.float16, torch.bfloat16], ids=["fp32", "fp16", "bf16"])
+    def test_from_pretrained_dtype_alias(self, tmp_path, dtype):
+        # `dtype` is an alias for `torch_dtype` in `from_pretrained`.
+        if torch_device == "mps" and dtype == torch.bfloat16:
+            pytest.skip(reason=f"{dtype} is not supported on {torch_device}")
+
+        model = self.model_class(**self.get_init_dict())
+        model.to(torch_device)
+        model.eval()
+
+        model.to(dtype)
+        model.save_pretrained(tmp_path)
+        new_model = self.model_class.from_pretrained(tmp_path, low_cpu_mem_usage=True, dtype=dtype)
+        assert new_model.dtype == dtype
+
     @torch.no_grad()
     def test_determinism(self, atol=1e-5, rtol=0):
         model = self.model_class(**self.get_init_dict())

tests/pipelines/cosmos/test_cosmos.py

@@ -349,6 +349,27 @@ def test_torch_dtype_dict(self):
                     f"Component '{name}' has dtype {component.dtype} but expected {expected_dtype}",
                 )
 
+    def test_dtype_alias(self):
+        # `dtype` is an alias for `torch_dtype` in `from_pretrained`.
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+
+        with tempfile.TemporaryDirectory(ignore_cleanup_errors=True) as tmpdirname:
+            pipe.save_pretrained(tmpdirname, safe_serialization=False)
+            loaded_pipe = self.pipeline_class.from_pretrained(
+                tmpdirname, safety_checker=DummyCosmosSafetyChecker(), dtype=torch.float16
+            )
+
+        for name, component in loaded_pipe.components.items():
+            if name == "safety_checker":
+                continue
+            if isinstance(component, torch.nn.Module) and hasattr(component, "dtype"):
+                self.assertEqual(
+                    component.dtype,
+                    torch.float16,
+                    f"Component '{name}' has dtype {component.dtype} but expected {torch.float16}",
+                )
+
     @unittest.skip(
         "The pipeline should not be runnable without a safety checker. The test creates a pipeline without passing in "
         "a safety checker, which makes the pipeline default to the actual Cosmos Guardrail. The Cosmos Guardrail is "

tests/pipelines/cosmos/test_cosmos2_5_predict.py

@@ -328,6 +328,27 @@ def test_torch_dtype_dict(self):
                     f"Component '{name}' has dtype {component.dtype} but expected {expected_dtype}",
                 )
 
+    def test_dtype_alias(self):
+        # `dtype` is an alias for `torch_dtype` in `from_pretrained`.
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+
+        with tempfile.TemporaryDirectory(ignore_cleanup_errors=True) as tmpdirname:
+            pipe.save_pretrained(tmpdirname, safe_serialization=False)
+            loaded_pipe = self.pipeline_class.from_pretrained(
+                tmpdirname, safety_checker=DummyCosmosSafetyChecker(), dtype=torch.float16
+            )
+
+        for name, component in loaded_pipe.components.items():
+            if name == "safety_checker":
+                continue
+            if isinstance(component, torch.nn.Module) and hasattr(component, "dtype"):
+                self.assertEqual(
+                    component.dtype,
+                    torch.float16,
+                    f"Component '{name}' has dtype {component.dtype} but expected {torch.float16}",
+                )
+
     @unittest.skip(
         "The pipeline should not be runnable without a safety checker. The test creates a pipeline without passing in "
         "a safety checker, which makes the pipeline default to the actual Cosmos Guardrail. The Cosmos Guardrail is "

tests/pipelines/cosmos/test_cosmos2_5_transfer.py

@@ -426,6 +426,28 @@ def test_torch_dtype_dict(self):
                     f"Component '{name}' has dtype {component.dtype} but expected {expected_dtype}",
                 )
 
+    def test_dtype_alias(self):
+        # `dtype` is an alias for `torch_dtype` in `from_pretrained`.
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+
+        with tempfile.TemporaryDirectory(ignore_cleanup_errors=True) as tmpdirname:
+            pipe.save_pretrained(tmpdirname, safe_serialization=False)
+            loaded_pipe = self.pipeline_class.from_pretrained(
+                tmpdirname, safety_checker=DummyCosmosSafetyChecker(), dtype=torch.float16
+            )
+
+        for name, component in loaded_pipe.components.items():
+            # Skip components that are not loaded from disk or have special handling
+            if name == "safety_checker":
+                continue
+            if isinstance(component, torch.nn.Module) and hasattr(component, "dtype"):
+                self.assertEqual(
+                    component.dtype,
+                    torch.float16,
+                    f"Component '{name}' has dtype {component.dtype} but expected {torch.float16}",
+                )
+
     def test_save_load_optional_components(self, expected_max_difference=1e-4):
         self.pipeline_class._optional_components.remove("safety_checker")
         super().test_save_load_optional_components(expected_max_difference=expected_max_difference)

tests/pipelines/cosmos/test_cosmos2_text2image.py

@@ -333,6 +333,27 @@ def test_torch_dtype_dict(self):
                     f"Component '{name}' has dtype {component.dtype} but expected {expected_dtype}",
                 )
 
+    def test_dtype_alias(self):
+        # `dtype` is an alias for `torch_dtype` in `from_pretrained`.
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+
+        with tempfile.TemporaryDirectory(ignore_cleanup_errors=True) as tmpdirname:
+            pipe.save_pretrained(tmpdirname, safe_serialization=False)
+            loaded_pipe = self.pipeline_class.from_pretrained(
+                tmpdirname, safety_checker=DummyCosmosSafetyChecker(), dtype=torch.float16
+            )
+
+        for name, component in loaded_pipe.components.items():
+            if name == "safety_checker":
+                continue
+            if isinstance(component, torch.nn.Module) and hasattr(component, "dtype"):
+                self.assertEqual(
+                    component.dtype,
+                    torch.float16,
+                    f"Component '{name}' has dtype {component.dtype} but expected {torch.float16}",
+                )
+
     @unittest.skip(
         "The pipeline should not be runnable without a safety checker. The test creates a pipeline without passing in "
         "a safety checker, which makes the pipeline default to the actual Cosmos Guardrail. The Cosmos Guardrail is "

tests/pipelines/cosmos/test_cosmos2_video2world.py

@@ -347,6 +347,27 @@ def test_torch_dtype_dict(self):
                     f"Component '{name}' has dtype {component.dtype} but expected {expected_dtype}",
                 )
 
+    def test_dtype_alias(self):
+        # `dtype` is an alias for `torch_dtype` in `from_pretrained`.
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+
+        with tempfile.TemporaryDirectory(ignore_cleanup_errors=True) as tmpdirname:
+            pipe.save_pretrained(tmpdirname, safe_serialization=False)
+            loaded_pipe = self.pipeline_class.from_pretrained(
+                tmpdirname, safety_checker=DummyCosmosSafetyChecker(), dtype=torch.float16
+            )
+
+        for name, component in loaded_pipe.components.items():
+            if name == "safety_checker":
+                continue
+            if isinstance(component, torch.nn.Module) and hasattr(component, "dtype"):
+                self.assertEqual(
+                    component.dtype,
+                    torch.float16,
+                    f"Component '{name}' has dtype {component.dtype} but expected {torch.float16}",
+                )
+
     @unittest.skip(
         "The pipeline should not be runnable without a safety checker. The test creates a pipeline without passing in "
         "a safety checker, which makes the pipeline default to the actual Cosmos Guardrail. The Cosmos Guardrail is "