AI/diffusers
1
0
Fork 0
mirror of https://github.com/huggingface/diffusers.git synced 2026-10-23 09:09:01 +03:00
Code Activity

[core] deprecate torch_dtype and prefer dtype following transformers. (#14205)

Browse Source 
* deprecate torch_dtype and prefer dtype following transformers.

* address dhruv's comment on introducing a helper.
This commit is contained in:
Sayak Paul
2026-07-21 07:32:08 +05:30
committed by GitHub
parent 86e6dac536
commit 13a7bee487
25 changed files with 255 additions and 151 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
src/diffusers/loaders/ip_adapter.py
View File

@@ -29,6 +29,7 @@ from ..utils import (
is_accelerate_available,
is_torch_version,
is_transformers_available,
is_transformers_version,
logging,
)
from .unet_loader_utils import _maybe_expand_lora_scales
@@ -204,13 +205,19 @@ class IPAdapterMixin:
else:
image_encoder_subfolder = Path(image_encoder_folder).as_posix()
# transformers renamed `torch_dtype` to `dtype` in 4.56.0.
dtype_kwarg = (
{"dtype": self.dtype}
if is_transformers_version(">=", "4.56.0")
else {"torch_dtype": self.dtype}
)
image_encoder = CLIPVisionModelWithProjection.from_pretrained(
pretrained_model_name_or_path_or_dict,
subfolder=image_encoder_subfolder,
low_cpu_mem_usage=low_cpu_mem_usage,
cache_dir=cache_dir,
local_files_only=local_files_only,
torch_dtype=self.dtype,
**dtype_kwarg,
).to(self.device)
self.register_modules(image_encoder=image_encoder)
else:
@@ -1043,11 +1050,17 @@ class SD3IPAdapterMixin:
"cache_dir": cache_dir,
"local_files_only": local_files_only,
}
# transformers renamed `torch_dtype` to `dtype` in 4.56.0.
dtype_kwarg = (
{"dtype": self.dtype}
if is_transformers_version(">=", "4.56.0")
else {"torch_dtype": self.dtype}
)
self.register_modules(
feature_extractor=SiglipImageProcessor.from_pretrained(image_encoder_subfolder, **kwargs),
image_encoder=SiglipVisionModel.from_pretrained(
image_encoder_subfolder, torch_dtype=self.dtype, **kwargs
image_encoder_subfolder, **dtype_kwarg, **kwargs
).to(self.device),
)
else:

6
src/diffusers/modular_pipelines/modular_pipeline.py
View File

@@ -2347,8 +2347,8 @@ class ModularPipeline(ConfigMixin, PushToHubMixin):
default_creation_method == "from_pretrained". If provided as a list or string, will load only the
specified components.
**kwargs: additional kwargs to be passed to `from_pretrained()`.Can be:
- a single value to be applied to all components to be loaded, e.g. torch_dtype=torch.bfloat16
- a dict, e.g. torch_dtype={"unet": torch.bfloat16, "default": torch.float32}
- a single value to be applied to all components to be loaded, e.g. dtype=torch.bfloat16
- a dict, e.g. dtype={"unet": torch.bfloat16, "default": torch.float32}
- if potentially override ComponentSpec if passed a different loading field in kwargs, e.g.
`pretrained_model_name_or_path`, `variant`, `revision`, etc.
- if potentially override ComponentSpec if passed a different loading field in kwargs, e.g.
@@ -2638,7 +2638,7 @@ class ModularPipeline(ConfigMixin, PushToHubMixin):
" is not recommended to move them to `cpu` as running them will fail. Please make"
" sure to use an accelerator to run the pipeline in inference, due to the lack of"
" support for`float16` operations on this device in PyTorch. Please, remove the"
" `torch_dtype=torch.float16` argument, or use another device for inference."
" `dtype=torch.float16` argument, or use another device for inference."
)
return self

9
src/diffusers/modular_pipelines/modular_pipeline_utils.py
View File

@@ -26,7 +26,7 @@ from packaging.specifiers import InvalidSpecifier, SpecifierSet
from ..configuration_utils import ConfigMixin, FrozenDict
from ..loaders.single_file_utils import _is_single_file_path_or_url
from ..utils import DIFFUSERS_LOAD_ID_FIELDS, is_torch_available, logging
from ..utils import DIFFUSERS_LOAD_ID_FIELDS, _resolve_dtype, is_torch_available, logging
from ..utils.import_utils import _is_package_available
@@ -293,6 +293,10 @@ class ComponentSpec:
# YiYi TODO: add guard for type of model, if it is supported by from_pretrained
def load(self, **kwargs) -> Any:
"""Load component using from_pretrained."""
torch_dtype = kwargs.pop("torch_dtype", None)
if torch_dtype is not None:
kwargs["dtype"] = _resolve_dtype(kwargs.get("dtype"), torch_dtype)
# select loading fields from kwargs passed from user: e.g. pretrained_model_name_or_path, subfolder, variant, revision, note the list could change
passed_loading_kwargs = {key: kwargs.pop(key) for key in self.loading_fields() if key in kwargs}
# merge loading field value in the spec with user passed values to create load_kwargs
@@ -311,12 +315,11 @@ class ComponentSpec:
from diffusers import AutoModel
# `dtype`/`torch_dtype` is not an accepted parameter for tokenizers and processors.
# `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:
try:

22
src/diffusers/pipelines/pipeline_loading_utils.py
View File

@@ -563,7 +563,7 @@ def _load_empty_model(
pipelines: Any,
is_pipeline_module: bool,
name: str,
torch_dtype: str | torch.dtype,
dtype: str | torch.dtype,
cached_folder: str | os.PathLike,
**kwargs,
):
@@ -636,7 +636,7 @@ def _load_empty_model(
model = class_obj(config)
if model is not None:
model = model.to(dtype=torch_dtype)
model = model.to(dtype=dtype)
return model
@@ -677,7 +677,7 @@ def _get_final_device_map(device_map, pipeline_class, passed_class_obj, init_dic
# To avoid circular import problem.
from diffusers import pipelines
torch_dtype = kwargs.get("torch_dtype", torch.float32)
dtype = kwargs.get("dtype", torch.float32)
# Load each module in the pipeline on a meta device so that we can derive the device map.
init_empty_modules = {}
@@ -708,9 +708,7 @@ def _get_final_device_map(device_map, pipeline_class, passed_class_obj, init_dic
else:
sub_model_dtype = (
torch_dtype.get(name, torch_dtype.get("default", torch.float32))
if isinstance(torch_dtype, dict)
else torch_dtype
dtype.get(name, dtype.get("default", torch.float32)) if isinstance(dtype, dict) else dtype
)
loaded_sub_model = _load_empty_model(
library_name=library_name,
@@ -720,7 +718,7 @@ def _get_final_device_map(device_map, pipeline_class, passed_class_obj, init_dic
is_pipeline_module=is_pipeline_module,
pipeline_class=pipeline_class,
name=name,
torch_dtype=sub_model_dtype,
dtype=sub_model_dtype,
cached_folder=kwargs.get("cached_folder", None),
force_download=kwargs.get("force_download", None),
proxies=kwargs.get("proxies", None),
@@ -738,9 +736,7 @@ def _get_final_device_map(device_map, pipeline_class, passed_class_obj, init_dic
module_sizes = {
module_name: compute_module_sizes(
module,
dtype=torch_dtype.get(module_name, torch_dtype.get("default", torch.float32))
if isinstance(torch_dtype, dict)
else torch_dtype,
dtype=dtype.get(module_name, dtype.get("default", torch.float32)) if isinstance(dtype, dict) else dtype,
)[""]
for module_name, module in init_empty_modules.items()
if isinstance(module, torch.nn.Module)
@@ -776,7 +772,7 @@ def load_sub_model(
pipelines: Any,
is_pipeline_module: bool,
pipeline_class: Any,
torch_dtype: torch.dtype,
dtype: torch.dtype,
provider: Any,
sess_options: Any,
device_map: dict[str, torch.device] | str | None,
@@ -856,9 +852,9 @@ def load_sub_model(
# For transformers models >= 4.56.0, use 'dtype' instead of 'torch_dtype' to avoid deprecation warnings
if issubclass(class_obj, torch.nn.Module):
if is_transformers_model and transformers_version >= version.parse("4.56.0"):
loading_kwargs["dtype"] = torch_dtype
loading_kwargs["dtype"] = dtype
else:
loading_kwargs["torch_dtype"] = torch_dtype
loading_kwargs["torch_dtype"] = dtype
if issubclass(class_obj, diffusers_module.OnnxRuntimeModel):
loading_kwargs["provider"] = provider
loading_kwargs["sess_options"] = sess_options

34
src/diffusers/pipelines/pipeline_utils.py
View File

@@ -65,6 +65,7 @@ from ..utils import (
PushToHubMixin,
_get_detailed_type,
_is_valid_type,
_resolve_dtype,
deprecate,
is_accelerate_available,
is_accelerate_version,
@@ -592,7 +593,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
" is not recommended to move them to `cpu` as running them will fail. Please make"
" sure to use an accelerator to run the pipeline in inference, due to the lack of"
" support for`float16` operations on this device in PyTorch. Please, remove the"
" `torch_dtype=torch.float16` argument, or use another device for inference."
" `dtype=torch.float16` argument, or use another device for inference."
)
return self
@@ -783,8 +784,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
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
dtype = _resolve_dtype(kwargs.pop("dtype", None), torch_dtype)
custom_pipeline = kwargs.pop("custom_pipeline", None)
custom_revision = kwargs.pop("custom_revision", None)
provider = kwargs.pop("provider", None)
@@ -805,11 +805,9 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
disable_mmap = kwargs.pop("disable_mmap", False)
trust_remote_code = kwargs.pop("trust_remote_code", False)
if torch_dtype is not None and not isinstance(torch_dtype, dict) and not isinstance(torch_dtype, torch.dtype):
torch_dtype = torch.float32
logger.warning(
f"Passed `torch_dtype` {torch_dtype} is not a `torch.dtype`. Defaulting to `torch.float32`."
)
if dtype is not None and not isinstance(dtype, dict) and not isinstance(dtype, torch.dtype):
logger.warning(f"Passed `dtype` {dtype} is not a `torch.dtype`. Defaulting to `torch.float32`.")
dtype = torch.float32
if low_cpu_mem_usage and not is_accelerate_available():
low_cpu_mem_usage = False
@@ -1019,7 +1017,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
init_dict=init_dict,
library=library,
max_memory=max_memory,
torch_dtype=torch_dtype,
dtype=dtype,
cached_folder=cached_folder,
force_download=force_download,
proxies=proxies,
@@ -1067,9 +1065,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
else:
# load sub model
sub_model_dtype = (
torch_dtype.get(name, torch_dtype.get("default", torch.float32))
if isinstance(torch_dtype, dict)
else torch_dtype
dtype.get(name, dtype.get("default", torch.float32)) if isinstance(dtype, dict) else dtype
)
loaded_sub_model = load_sub_model(
library_name=library_name,
@@ -1078,7 +1074,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
pipelines=pipelines,
is_pipeline_module=is_pipeline_module,
pipeline_class=pipeline_class,
torch_dtype=sub_model_dtype,
dtype=sub_model_dtype,
provider=provider,
sess_options=sess_options,
device_map=current_device_map,
@@ -1466,7 +1462,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
>>> from diffusers import DiffusionPipeline
>>> import torch
>>> pipe = DiffusionPipeline.from_pretrained("Qwen/Qwen-Image", torch_dtype=torch.bfloat16)
>>> pipe = DiffusionPipeline.from_pretrained("Qwen/Qwen-Image", dtype=torch.bfloat16)
>>> pipe.enable_group_offload(
... onload_device=torch.device("cuda"),
@@ -2055,7 +2051,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
>>> from diffusers import DiffusionPipeline
>>> from xformers.ops import MemoryEfficientAttentionFlashAttentionOp
>>> pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1", torch_dtype=torch.float16)
>>> pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1", dtype=torch.float16)
>>> pipe = pipe.to("cuda")
>>> pipe.enable_xformers_memory_efficient_attention(attention_op=MemoryEfficientAttentionFlashAttentionOp)
>>> # Workaround for not accepting attention shape using VAE for Flash Attention
@@ -2114,7 +2110,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
>>> pipe = StableDiffusionPipeline.from_pretrained(
... "stable-diffusion-v1-5/stable-diffusion-v1-5",
... torch_dtype=torch.float16,
... dtype=torch.float16,
... use_safetensors=True,
... )
@@ -2167,8 +2163,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
original_config = dict(pipeline.config)
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)
dtype = _resolve_dtype(kwargs.pop("dtype", None), torch_dtype) or torch.float32
trust_remote_code = kwargs.pop("trust_remote_code", False)
# derive the pipeline class to instantiate
@@ -2266,8 +2261,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
new_pipeline.register_to_config(_name_or_path=pretrained_model_name_or_path)
new_pipeline.register_to_config(**unused_original_config)
if torch_dtype is not None:
new_pipeline.to(dtype=torch_dtype)
new_pipeline.to(dtype=dtype)
return new_pipeline

2
src/diffusers/utils/__init__.py
View File

@@ -41,7 +41,7 @@ from .constants import (
WEIGHTS_INDEX_NAME,
WEIGHTS_NAME,
)
from .deprecation_utils import _maybe_remap_transformers_class, deprecate
from .deprecation_utils import _maybe_remap_transformers_class, _resolve_dtype, deprecate
from .doc_utils import replace_example_docstring
from .dynamic_modules_utils import get_class_from_dynamic_module
from .export_utils import encode_video, export_to_gif, export_to_obj, export_to_ply, export_to_video

15
src/diffusers/utils/deprecation_utils.py
View File

@@ -9,6 +9,21 @@ from ..utils import logging
logger = logging.get_logger(__name__)
_TORCH_DTYPE_DEPRECATION_MESSAGE = "Please use `dtype` instead."
def _resolve_dtype(dtype, torch_dtype):
"""Resolve the deprecated `torch_dtype` argument into `dtype`."""
if torch_dtype is None:
return dtype
if dtype is not None:
raise ValueError(
"You have passed both `dtype` and `torch_dtype`. Please only pass `dtype`, `torch_dtype` is deprecated."
)
deprecate("torch_dtype", "1.0.0", _TORCH_DTYPE_DEPRECATION_MESSAGE)
return torch_dtype
# Mapping for deprecated Transformers classes to their replacements
# This is used to handle models that reference deprecated class names in their configs
# Reference: https://github.com/huggingface/transformers/issues/40822

4
tests/modular_pipelines/anima/test_modular_pipeline_anima.py
View File

@@ -177,10 +177,10 @@ class TestAnimaModularPipelineFast(ModularPipelineTesterMixin, ModularGuiderTest
batch_params = frozenset(["prompt", "negative_prompt"])
expected_workflow_blocks = ANIMA_TEXT2IMAGE_WORKFLOWS
def get_pipeline(self, components_manager=None, torch_dtype=torch.float32):
def get_pipeline(self, components_manager=None, dtype=torch.float32):
pipe = self.pipeline_blocks_class().init_pipeline(components_manager=components_manager)
pipe.update_components(**get_dummy_components())
pipe.to(dtype=torch_dtype)
pipe.to(dtype=dtype)
pipe.set_progress_bar_config(disable=None)
return pipe

6
tests/modular_pipelines/cosmos/test_modular_pipeline_cosmos3.py
View File

@@ -125,8 +125,8 @@ class TestCosmos3OmniModularPipelineFast(ModularPipelineTesterMixin):
output_name = "videos"
expected_workflow_blocks = COSMOS3_OMNI_WORKFLOWS
def get_pipeline(self, components_manager=None, torch_dtype=torch.float32):
pipe = super().get_pipeline(components_manager, torch_dtype)
def get_pipeline(self, components_manager=None, dtype=torch.float32):
pipe = super().get_pipeline(components_manager, dtype)
pipe.disable_safety_checker()
return pipe
@@ -164,7 +164,7 @@ class TestCosmos3OmniModularPipelineFast(ModularPipelineTesterMixin):
base_pipe.save_pretrained(str(tmp_path))
loaded_pipe = ModularPipeline.from_pretrained(str(tmp_path))
loaded_pipe.load_components(torch_dtype=torch.float32)
loaded_pipe.load_components(dtype=torch.float32)
loaded_pipe.disable_safety_checker()
loaded_pipe.to(torch_device)

8
tests/modular_pipelines/flux/test_modular_pipeline_flux.py
View File

@@ -98,8 +98,8 @@ class TestFluxImg2ImgModularPipelineFast(ModularPipelineTesterMixin):
batch_params = frozenset(["prompt", "image"])
expected_workflow_blocks = FLUX_IMAGE2IMAGE_WORKFLOWS
def get_pipeline(self, components_manager=None, torch_dtype=torch.float32):
pipeline = super().get_pipeline(components_manager, torch_dtype)
def get_pipeline(self, components_manager=None, dtype=torch.float32):
pipeline = super().get_pipeline(components_manager, dtype)
# Override `vae_scale_factor` here as currently, `image_processor` is initialized with
# fixed constants instead of
@@ -135,7 +135,7 @@ class TestFluxImg2ImgModularPipelineFast(ModularPipelineTesterMixin):
base_pipe.save_pretrained(str(tmp_path))
pipe = ModularPipeline.from_pretrained(tmp_path).to(torch_device)
pipe.load_components(torch_dtype=torch.float32)
pipe.load_components(dtype=torch.float32)
pipe.to(torch_device)
pipe.image_processor = VaeImageProcessor(vae_scale_factor=2)
@@ -216,7 +216,7 @@ class TestFluxKontextModularPipelineFast(ModularPipelineTesterMixin):
base_pipe.save_pretrained(str(tmp_path))
pipe = ModularPipeline.from_pretrained(tmp_path).to(torch_device)
pipe.load_components(torch_dtype=torch.float32)
pipe.load_components(dtype=torch.float32)
pipe.to(torch_device)
pipe.image_processor = VaeImageProcessor(vae_scale_factor=2)

16
tests/modular_pipelines/stable_diffusion_3/test_modular_pipeline_stable_diffusion_3.py
View File

@@ -67,8 +67,8 @@ class TestStableDiffusion3ModularPipelineFast(ModularPipelineTesterMixin):
"output_type": "pt",
}
def get_pipeline(self, components_manager=None, torch_dtype=torch.float32):
return super().get_pipeline(components_manager, torch_dtype)
def get_pipeline(self, components_manager=None, dtype=torch.float32):
return super().get_pipeline(components_manager, dtype)
def test_save_from_pretrained(self, tmp_path):
pipes = []
@@ -77,7 +77,7 @@ class TestStableDiffusion3ModularPipelineFast(ModularPipelineTesterMixin):
base_pipe.save_pretrained(str(tmp_path))
pipe = self.pipeline_class.from_pretrained(tmp_path).to(torch_device)
pipe.load_components(torch_dtype=torch.float32)
pipe.load_components(dtype=torch.float32)
pipe.to(torch_device)
pipes.append(pipe)
@@ -94,7 +94,7 @@ class TestStableDiffusion3ModularPipelineFast(ModularPipelineTesterMixin):
base_pipe.save_pretrained(str(tmp_path))
pipe = self.pipeline_class.from_pretrained(tmp_path)
pipe.load_components(torch_dtype=torch.float32)
pipe.load_components(dtype=torch.float32)
assert set(base_pipe.components.keys()) == set(pipe.components.keys())
@@ -135,8 +135,8 @@ class TestStableDiffusion3Img2ImgModularPipelineFast(ModularPipelineTesterMixin)
# (guidance_scale) which are intentionally omitted from pipeline inputs.
pass
def get_pipeline(self, components_manager=None, torch_dtype=torch.float32):
pipeline = super().get_pipeline(components_manager, torch_dtype)
def get_pipeline(self, components_manager=None, dtype=torch.float32):
pipeline = super().get_pipeline(components_manager, dtype)
pipeline.image_processor = VaeImageProcessor(vae_scale_factor=8)
return pipeline
@@ -165,7 +165,7 @@ class TestStableDiffusion3Img2ImgModularPipelineFast(ModularPipelineTesterMixin)
base_pipe.save_pretrained(str(tmp_path))
pipe = self.pipeline_class.from_pretrained(tmp_path).to(torch_device)
pipe.load_components(torch_dtype=torch.float32)
pipe.load_components(dtype=torch.float32)
pipe.to(torch_device)
pipe.image_processor = VaeImageProcessor(vae_scale_factor=8)
pipes.append(pipe)
@@ -183,7 +183,7 @@ class TestStableDiffusion3Img2ImgModularPipelineFast(ModularPipelineTesterMixin)
base_pipe.save_pretrained(str(tmp_path))
pipe = self.pipeline_class.from_pretrained(tmp_path)
pipe.load_components(torch_dtype=torch.float32)
pipe.load_components(dtype=torch.float32)
assert set(base_pipe.components.keys()) == set(pipe.components.keys())

2
tests/modular_pipelines/stable_diffusion_xl/test_modular_pipeline_stable_diffusion_xl.py
View File

@@ -106,7 +106,7 @@ class SDXLModularIPAdapterTesterMixin:
blocks = self.pipeline_blocks_class()
_ = blocks.sub_blocks.pop("ip_adapter")
pipe = blocks.init_pipeline(self.pretrained_model_name_or_path)
pipe.load_components(torch_dtype=torch.float32)
pipe.load_components(dtype=torch.float32)
pipe = pipe.to(torch_device)
cross_attention_dim = pipe.unet.config.get("cross_attention_dim")

32
tests/modular_pipelines/test_modular_pipelines_common.py
View File

@@ -158,11 +158,11 @@ class ModularPipelineTesterMixin:
gc.collect()
backend_empty_cache(torch_device)
def get_pipeline(self, components_manager=None, torch_dtype=torch.float32):
def get_pipeline(self, components_manager=None, dtype=torch.float32):
pipeline = self.pipeline_blocks_class().init_pipeline(
self.pretrained_model_name_or_path, components_manager=components_manager
)
pipeline.load_components(torch_dtype=torch_dtype)
pipeline.load_components(dtype=dtype)
pipeline.set_progress_bar_config(disable=None)
return pipeline
@@ -378,7 +378,7 @@ class ModularPipelineTesterMixin:
base_pipe.save_pretrained(str(tmp_path))
pipe = ModularPipeline.from_pretrained(tmp_path).to(torch_device)
pipe.load_components(torch_dtype=torch.float32)
pipe.load_components(dtype=torch.float32)
pipe.to(torch_device)
pipes.append(pipe)
@@ -412,7 +412,7 @@ class ModularPipelineTesterMixin:
expected = _get_specified_components(save_dir)
loaded_pipe = ModularPipeline.from_pretrained(save_dir)
loaded_pipe.load_components(torch_dtype=torch.float32)
loaded_pipe.load_components(dtype=torch.float32)
actual = {
name
@@ -723,7 +723,7 @@ class TestAutoModelLoadIdTagging:
def test_automodel_update_components(self):
pipe = ModularPipeline.from_pretrained("hf-internal-testing/tiny-stable-diffusion-xl-pipe")
pipe.load_components(torch_dtype=torch.float32)
pipe.load_components(dtype=torch.float32)
auto_model = AutoModel.from_pretrained("hf-internal-testing/tiny-stable-diffusion-xl-pipe", subfolder="unet")
@@ -760,7 +760,7 @@ class TestAutoModelLoadIdTagging:
class TestLoadComponentsSkipBehavior:
def test_load_components_skips_already_loaded(self):
pipe = ModularPipeline.from_pretrained("hf-internal-testing/tiny-stable-diffusion-xl-pipe")
pipe.load_components(torch_dtype=torch.float32)
pipe.load_components(dtype=torch.float32)
original_unet = pipe.unet
@@ -772,7 +772,7 @@ class TestLoadComponentsSkipBehavior:
def test_load_components_selective_loading(self):
pipe = ModularPipeline.from_pretrained("hf-internal-testing/tiny-stable-diffusion-xl-pipe")
pipe.load_components(names="unet", torch_dtype=torch.float32)
pipe.load_components(names="unet", dtype=torch.float32)
# Verify only requested component was loaded.
assert hasattr(pipe, "unet")
@@ -783,8 +783,8 @@ class TestLoadComponentsSkipBehavior:
"""Loading a subset of components should not affect already-loaded components."""
pipe = ModularPipeline.from_pretrained("hf-internal-testing/tiny-stable-diffusion-xl-pipe")
pipe.load_components(names="unet", torch_dtype=torch.float32)
pipe.load_components(names="text_encoder", torch_dtype=torch.float32)
pipe.load_components(names="unet", dtype=torch.float32)
pipe.load_components(names="text_encoder", dtype=torch.float32)
assert hasattr(pipe, "unet")
assert pipe.unet is not None
@@ -800,7 +800,7 @@ class TestLoadComponentsSkipBehavior:
pretrained_model_name_or_path=None,
default_creation_method="from_pretrained",
)
pipe.load_components(torch_dtype=torch.float32)
pipe.load_components(dtype=torch.float32)
# Verify test_component was not loaded
assert not hasattr(pipe, "test_component") or pipe.test_component is None
@@ -813,7 +813,7 @@ class TestCustomModelSavePretrained:
import json
pipe = ModularPipeline.from_pretrained("hf-internal-testing/tiny-stable-diffusion-xl-pipe")
pipe.load_components(torch_dtype=torch.float32)
pipe.load_components(dtype=torch.float32)
pipe.unet._diffusers_load_id = "null"
@@ -833,7 +833,7 @@ class TestCustomModelSavePretrained:
def test_save_pretrained_roundtrip_with_local_model(self, tmp_path):
"""A pipeline with a custom/local model should be saveable and re-loadable with identical outputs."""
pipe = ModularPipeline.from_pretrained("hf-internal-testing/tiny-stable-diffusion-xl-pipe")
pipe.load_components(torch_dtype=torch.float32)
pipe.load_components(dtype=torch.float32)
pipe.unet._diffusers_load_id = "null"
@@ -843,7 +843,7 @@ class TestCustomModelSavePretrained:
pipe.save_pretrained(save_dir)
loaded_pipe = ModularPipeline.from_pretrained(save_dir)
loaded_pipe.load_components(torch_dtype=torch.float32)
loaded_pipe.load_components(dtype=torch.float32)
assert loaded_pipe.unet is not None
assert loaded_pipe.unet.__class__.__name__ == pipe.unet.__class__.__name__
@@ -861,7 +861,7 @@ class TestCustomModelSavePretrained:
from diffusers import UNet2DConditionModel
pipe = ModularPipeline.from_pretrained("hf-internal-testing/tiny-stable-diffusion-xl-pipe")
pipe.load_components(torch_dtype=torch.float32)
pipe.load_components(dtype=torch.float32)
unet = UNet2DConditionModel.from_pretrained(
"hf-internal-testing/tiny-stable-diffusion-xl-pipe", subfolder="unet"
@@ -888,7 +888,7 @@ class TestCustomModelSavePretrained:
import json
pipe = ModularPipeline.from_pretrained("hf-internal-testing/tiny-stable-diffusion-xl-pipe")
pipe.load_components(torch_dtype=torch.float32)
pipe.load_components(dtype=torch.float32)
save_dir = str(tmp_path / "my-pipeline")
pipe.save_pretrained(save_dir, overwrite_modular_index=True)
@@ -906,7 +906,7 @@ class TestCustomModelSavePretrained:
assert spec["subfolder"] == component_name
loaded_pipe = ModularPipeline.from_pretrained(save_dir)
loaded_pipe.load_components(torch_dtype=torch.float32)
loaded_pipe.load_components(dtype=torch.float32)
assert loaded_pipe.unet is not None
assert loaded_pipe.vae is not None

28
tests/modular_pipelines/test_modular_pipelines_custom_blocks.py
View File

@@ -19,6 +19,7 @@ from collections import deque
from typing import List
import numpy as np
import pytest
import torch
from diffusers import FluxTransformer2DModel
@@ -379,8 +380,8 @@ class TestModularCustomBlocks:
loaded_pipe.update_components(custom_model=custom_model)
assert getattr(loaded_pipe, "custom_model", None) is not None
def test_automodel_type_hint_preserves_torch_dtype(self, tmp_path):
"""Regression test for #13271: torch_dtype was incorrectly removed when type_hint is AutoModel."""
def test_automodel_type_hint_preserves_dtype(self, tmp_path):
"""Regression test for #13271: dtype was incorrectly removed when type_hint is AutoModel."""
from diffusers import AutoModel
model_dir = str(tmp_path / "model")
@@ -412,10 +413,27 @@ class TestModularCustomBlocks:
block = DtypeTestBlock()
pipe = block.init_pipeline()
pipe.load_components(torch_dtype=torch.float16, trust_remote_code=True)
pipe.load_components(dtype=torch.float16, trust_remote_code=True)
assert pipe.model.dtype == torch.float16
def test_component_spec_load_torch_dtype_is_deprecated(self, tmp_path):
from diffusers import AutoModel
model_dir = str(tmp_path / "model")
os.makedirs(model_dir)
_create_tiny_model_dir(model_dir)
spec = ComponentSpec("model", AutoModel, pretrained_model_name_or_path=model_dir)
with pytest.warns(FutureWarning, match="torch_dtype"):
component = spec.load(torch_dtype=torch.float16, trust_remote_code=True)
# The deprecated argument is still honoured until it is removed.
assert component.dtype == torch.float16
with pytest.raises(ValueError, match="passed both `dtype` and `torch_dtype`"):
spec.load(dtype=torch.float16, torch_dtype=torch.float16, trust_remote_code=True)
@require_torch_accelerator
def test_automodel_type_hint_preserves_device(self, tmp_path):
"""Test that ComponentSpec with AutoModel type_hint correctly passes device_map."""
@@ -707,7 +725,7 @@ class TestKreaCustomBlocksIntegration:
pipe.load_components(
trust_remote_code=True,
device_map="cuda",
torch_dtype={"default": torch.bfloat16, "vae": torch.float16},
dtype={"default": torch.bfloat16, "vae": torch.float16},
)
assert len(pipe.components) == 7
assert sorted(pipe.components) == sorted(
@@ -720,7 +738,7 @@ class TestKreaCustomBlocksIntegration:
pipe.load_components(
trust_remote_code=True,
device_map="cuda",
torch_dtype={"default": torch.bfloat16, "vae": torch.float16},
dtype={"default": torch.bfloat16, "vae": torch.float16},
)
num_frames_per_block = 2

14
tests/pipelines/bria/test_pipeline_bria.py
View File

@@ -170,7 +170,7 @@ class BriaPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
with tempfile.TemporaryDirectory() as tmpdir:
pipe.save_pretrained(tmpdir)
pipe_loaded = self.pipeline_class.from_pretrained(tmpdir, torch_dtype=torch.float16)
pipe_loaded = self.pipeline_class.from_pretrained(tmpdir, dtype=torch.float16)
for component in pipe_loaded.components.values():
if hasattr(component, "set_default_attn_processor"):
component.set_default_attn_processor()
@@ -215,14 +215,14 @@ class BriaPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
model_dtypes = [component.dtype for component in components.values() if hasattr(component, "dtype")]
self.assertTrue([dtype == torch.float32 for dtype in model_dtypes] == [True, True, True])
def test_torch_dtype_dict(self):
def test_dtype_dict(self):
components = self.get_dummy_components()
pipe = self.pipeline_class(**components)
with tempfile.TemporaryDirectory() as tmpdirname:
pipe.save_pretrained(tmpdirname)
torch_dtype_dict = {"transformer": torch.bfloat16, "default": torch.float16}
loaded_pipe = self.pipeline_class.from_pretrained(tmpdirname, torch_dtype=torch_dtype_dict)
dtype_dict = {"transformer": torch.bfloat16, "default": torch.float16}
loaded_pipe = self.pipeline_class.from_pretrained(tmpdirname, dtype=dtype_dict)
self.assertEqual(loaded_pipe.transformer.dtype, torch.bfloat16)
self.assertEqual(loaded_pipe.text_encoder.dtype, torch.float16)
@@ -230,8 +230,8 @@ class BriaPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
with tempfile.TemporaryDirectory() as tmpdirname:
pipe.save_pretrained(tmpdirname)
torch_dtype_dict = {"default": torch.float16}
loaded_pipe = self.pipeline_class.from_pretrained(tmpdirname, torch_dtype=torch_dtype_dict)
dtype_dict = {"default": torch.float16}
loaded_pipe = self.pipeline_class.from_pretrained(tmpdirname, dtype=dtype_dict)
self.assertEqual(loaded_pipe.transformer.dtype, torch.float16)
self.assertEqual(loaded_pipe.text_encoder.dtype, torch.float16)
@@ -272,7 +272,7 @@ class BriaPipelineSlowTests(unittest.TestCase):
def test_bria_inference_bf16(self):
pipe = self.pipeline_class.from_pretrained(
self.repo_id, torch_dtype=torch.bfloat16, text_encoder=None, tokenizer=None
self.repo_id, dtype=torch.bfloat16, text_encoder=None, tokenizer=None
)
pipe.to(torch_device)

10
tests/pipelines/cosmos/test_cosmos.py
View File

@@ -321,7 +321,7 @@ class CosmosTextToWorldPipelineFastTests(PipelineTesterMixin, unittest.TestCase)
is_folder = os.path.isdir(folder_path) and subfolder in config
assert is_folder and any(p.split(".")[1].startswith(variant) for p in os.listdir(folder_path))
def test_torch_dtype_dict(self):
def test_dtype_dict(self):
components = self.get_dummy_components()
if not components:
self.skipTest("No dummy components defined.")
@@ -332,16 +332,16 @@ class CosmosTextToWorldPipelineFastTests(PipelineTesterMixin, unittest.TestCase)
with tempfile.TemporaryDirectory(ignore_cleanup_errors=True) as tmpdirname:
pipe.save_pretrained(tmpdirname, safe_serialization=False)
torch_dtype_dict = {specified_key: torch.bfloat16, "default": torch.float16}
dtype_dict = {specified_key: torch.bfloat16, "default": torch.float16}
loaded_pipe = self.pipeline_class.from_pretrained(
tmpdirname, safety_checker=DummyCosmosSafetyChecker(), torch_dtype=torch_dtype_dict
tmpdirname, safety_checker=DummyCosmosSafetyChecker(), dtype=dtype_dict
)
for name, component in loaded_pipe.components.items():
if name == "safety_checker":
continue
if isinstance(component, torch.nn.Module) and hasattr(component, "dtype"):
expected_dtype = torch_dtype_dict.get(name, torch_dtype_dict.get("default", torch.float32))
expected_dtype = dtype_dict.get(name, dtype_dict.get("default", torch.float32))
self.assertEqual(
component.dtype,
expected_dtype,
@@ -349,7 +349,7 @@ class CosmosTextToWorldPipelineFastTests(PipelineTesterMixin, unittest.TestCase)
)
def test_dtype_alias(self):
# `dtype` is an alias for `torch_dtype` in `from_pretrained`.
# `torch_dtype` is deprecated in favor of `dtype` in `from_pretrained`.
components = self.get_dummy_components()
pipe = self.pipeline_class(**components)

16
tests/pipelines/cosmos/test_cosmos2_5_predict.py
View File

@@ -44,9 +44,9 @@ class Cosmos2_5_PredictBaseWrapper(Cosmos2_5_PredictBasePipeline):
if "safety_checker" not in kwargs or kwargs["safety_checker"] is None:
safety_checker = DummyCosmosSafetyChecker()
device_map = kwargs.get("device_map", "cpu")
torch_dtype = kwargs.get("torch_dtype")
if device_map is not None or torch_dtype is not None:
safety_checker = safety_checker.to(device_map, dtype=torch_dtype)
dtype = kwargs.get("dtype")
if device_map is not None or dtype is not None:
safety_checker = safety_checker.to(device_map, dtype=dtype)
kwargs["safety_checker"] = safety_checker
return Cosmos2_5_PredictBasePipeline.from_pretrained(*args, **kwargs)
@@ -300,7 +300,7 @@ class Cosmos2_5_PredictPipelineFastTests(PipelineTesterMixin, unittest.TestCase)
is_folder = os.path.isdir(folder_path) and subfolder in config
assert is_folder and any(p.split(".")[1].startswith(variant) for p in os.listdir(folder_path))
def test_torch_dtype_dict(self):
def test_dtype_dict(self):
components = self.get_dummy_components()
if not components:
self.skipTest("No dummy components defined.")
@@ -311,16 +311,16 @@ class Cosmos2_5_PredictPipelineFastTests(PipelineTesterMixin, unittest.TestCase)
with tempfile.TemporaryDirectory(ignore_cleanup_errors=True) as tmpdirname:
pipe.save_pretrained(tmpdirname, safe_serialization=False)
torch_dtype_dict = {specified_key: torch.bfloat16, "default": torch.float16}
dtype_dict = {specified_key: torch.bfloat16, "default": torch.float16}
loaded_pipe = self.pipeline_class.from_pretrained(
tmpdirname, safety_checker=DummyCosmosSafetyChecker(), torch_dtype=torch_dtype_dict
tmpdirname, safety_checker=DummyCosmosSafetyChecker(), dtype=dtype_dict
)
for name, component in loaded_pipe.components.items():
if name == "safety_checker":
continue
if isinstance(component, torch.nn.Module) and hasattr(component, "dtype"):
expected_dtype = torch_dtype_dict.get(name, torch_dtype_dict.get("default", torch.float32))
expected_dtype = dtype_dict.get(name, dtype_dict.get("default", torch.float32))
self.assertEqual(
component.dtype,
expected_dtype,
@@ -328,7 +328,7 @@ class Cosmos2_5_PredictPipelineFastTests(PipelineTesterMixin, unittest.TestCase)
)
def test_dtype_alias(self):
# `dtype` is an alias for `torch_dtype` in `from_pretrained`.
# `torch_dtype` is deprecated in favor of `dtype` in `from_pretrained`.
components = self.get_dummy_components()
pipe = self.pipeline_class(**components)

16
tests/pipelines/cosmos/test_cosmos2_5_transfer.py
View File

@@ -45,9 +45,9 @@ class Cosmos2_5_TransferWrapper(Cosmos2_5_TransferPipeline):
if "safety_checker" not in kwargs or kwargs["safety_checker"] is None:
safety_checker = DummyCosmosSafetyChecker()
device_map = kwargs.get("device_map", "cpu")
torch_dtype = kwargs.get("torch_dtype")
if device_map is not None or torch_dtype is not None:
safety_checker = safety_checker.to(device_map, dtype=torch_dtype)
dtype = kwargs.get("dtype")
if device_map is not None or dtype is not None:
safety_checker = safety_checker.to(device_map, dtype=dtype)
kwargs["safety_checker"] = safety_checker
return Cosmos2_5_TransferPipeline.from_pretrained(*args, **kwargs)
@@ -397,7 +397,7 @@ class Cosmos2_5_TransferPipelineFastTests(PipelineTesterMixin, unittest.TestCase
is_folder = os.path.isdir(folder_path) and subfolder in config
assert is_folder and any(p.split(".")[1].startswith(variant) for p in os.listdir(folder_path))
def test_torch_dtype_dict(self):
def test_dtype_dict(self):
components = self.get_dummy_components()
if not components:
self.skipTest("No dummy components defined.")
@@ -408,9 +408,9 @@ class Cosmos2_5_TransferPipelineFastTests(PipelineTesterMixin, unittest.TestCase
with tempfile.TemporaryDirectory(ignore_cleanup_errors=True) as tmpdirname:
pipe.save_pretrained(tmpdirname, safe_serialization=False)
torch_dtype_dict = {specified_key: torch.bfloat16, "default": torch.float16}
dtype_dict = {specified_key: torch.bfloat16, "default": torch.float16}
loaded_pipe = self.pipeline_class.from_pretrained(
tmpdirname, safety_checker=DummyCosmosSafetyChecker(), torch_dtype=torch_dtype_dict
tmpdirname, safety_checker=DummyCosmosSafetyChecker(), dtype=dtype_dict
)
for name, component in loaded_pipe.components.items():
@@ -418,7 +418,7 @@ class Cosmos2_5_TransferPipelineFastTests(PipelineTesterMixin, unittest.TestCase
if name == "safety_checker":
continue
if isinstance(component, torch.nn.Module) and hasattr(component, "dtype"):
expected_dtype = torch_dtype_dict.get(name, torch_dtype_dict.get("default", torch.float32))
expected_dtype = dtype_dict.get(name, dtype_dict.get("default", torch.float32))
self.assertEqual(
component.dtype,
expected_dtype,
@@ -426,7 +426,7 @@ class Cosmos2_5_TransferPipelineFastTests(PipelineTesterMixin, unittest.TestCase
)
def test_dtype_alias(self):
# `dtype` is an alias for `torch_dtype` in `from_pretrained`.
# `torch_dtype` is deprecated in favor of `dtype` in `from_pretrained`.
components = self.get_dummy_components()
pipe = self.pipeline_class(**components)

10
tests/pipelines/cosmos/test_cosmos2_text2image.py
View File

@@ -305,7 +305,7 @@ class Cosmos2TextToImagePipelineFastTests(PipelineTesterMixin, unittest.TestCase
is_folder = os.path.isdir(folder_path) and subfolder in config
assert is_folder and any(p.split(".")[1].startswith(variant) for p in os.listdir(folder_path))
def test_torch_dtype_dict(self):
def test_dtype_dict(self):
components = self.get_dummy_components()
if not components:
self.skipTest("No dummy components defined.")
@@ -316,16 +316,16 @@ class Cosmos2TextToImagePipelineFastTests(PipelineTesterMixin, unittest.TestCase
with tempfile.TemporaryDirectory(ignore_cleanup_errors=True) as tmpdirname:
pipe.save_pretrained(tmpdirname, safe_serialization=False)
torch_dtype_dict = {specified_key: torch.bfloat16, "default": torch.float16}
dtype_dict = {specified_key: torch.bfloat16, "default": torch.float16}
loaded_pipe = self.pipeline_class.from_pretrained(
tmpdirname, safety_checker=DummyCosmosSafetyChecker(), torch_dtype=torch_dtype_dict
tmpdirname, safety_checker=DummyCosmosSafetyChecker(), dtype=dtype_dict
)
for name, component in loaded_pipe.components.items():
if name == "safety_checker":
continue
if isinstance(component, torch.nn.Module) and hasattr(component, "dtype"):
expected_dtype = torch_dtype_dict.get(name, torch_dtype_dict.get("default", torch.float32))
expected_dtype = dtype_dict.get(name, dtype_dict.get("default", torch.float32))
self.assertEqual(
component.dtype,
expected_dtype,
@@ -333,7 +333,7 @@ class Cosmos2TextToImagePipelineFastTests(PipelineTesterMixin, unittest.TestCase
)
def test_dtype_alias(self):
# `dtype` is an alias for `torch_dtype` in `from_pretrained`.
# `torch_dtype` is deprecated in favor of `dtype` in `from_pretrained`.
components = self.get_dummy_components()
pipe = self.pipeline_class(**components)

10
tests/pipelines/cosmos/test_cosmos2_video2world.py
View File

@@ -319,7 +319,7 @@ class Cosmos2VideoToWorldPipelineFastTests(PipelineTesterMixin, unittest.TestCas
is_folder = os.path.isdir(folder_path) and subfolder in config
assert is_folder and any(p.split(".")[1].startswith(variant) for p in os.listdir(folder_path))
def test_torch_dtype_dict(self):
def test_dtype_dict(self):
components = self.get_dummy_components()
if not components:
self.skipTest("No dummy components defined.")
@@ -330,16 +330,16 @@ class Cosmos2VideoToWorldPipelineFastTests(PipelineTesterMixin, unittest.TestCas
with tempfile.TemporaryDirectory(ignore_cleanup_errors=True) as tmpdirname:
pipe.save_pretrained(tmpdirname, safe_serialization=False)
torch_dtype_dict = {specified_key: torch.bfloat16, "default": torch.float16}
dtype_dict = {specified_key: torch.bfloat16, "default": torch.float16}
loaded_pipe = self.pipeline_class.from_pretrained(
tmpdirname, safety_checker=DummyCosmosSafetyChecker(), torch_dtype=torch_dtype_dict
tmpdirname, safety_checker=DummyCosmosSafetyChecker(), dtype=dtype_dict
)
for name, component in loaded_pipe.components.items():
if name == "safety_checker":
continue
if isinstance(component, torch.nn.Module) and hasattr(component, "dtype"):
expected_dtype = torch_dtype_dict.get(name, torch_dtype_dict.get("default", torch.float32))
expected_dtype = dtype_dict.get(name, dtype_dict.get("default", torch.float32))
self.assertEqual(
component.dtype,
expected_dtype,
@@ -347,7 +347,7 @@ class Cosmos2VideoToWorldPipelineFastTests(PipelineTesterMixin, unittest.TestCas
)
def test_dtype_alias(self):
# `dtype` is an alias for `torch_dtype` in `from_pretrained`.
# `torch_dtype` is deprecated in favor of `dtype` in `from_pretrained`.
components = self.get_dummy_components()
pipe = self.pipeline_class(**components)

10
tests/pipelines/cosmos/test_cosmos_video2world.py
View File

@@ -334,7 +334,7 @@ class CosmosVideoToWorldPipelineFastTests(PipelineTesterMixin, unittest.TestCase
is_folder = os.path.isdir(folder_path) and subfolder in config
assert is_folder and any(p.split(".")[1].startswith(variant) for p in os.listdir(folder_path))
def test_torch_dtype_dict(self):
def test_dtype_dict(self):
components = self.get_dummy_components()
if not components:
self.skipTest("No dummy components defined.")
@@ -345,16 +345,16 @@ class CosmosVideoToWorldPipelineFastTests(PipelineTesterMixin, unittest.TestCase
with tempfile.TemporaryDirectory(ignore_cleanup_errors=True) as tmpdirname:
pipe.save_pretrained(tmpdirname, safe_serialization=False)
torch_dtype_dict = {specified_key: torch.bfloat16, "default": torch.float16}
dtype_dict = {specified_key: torch.bfloat16, "default": torch.float16}
loaded_pipe = self.pipeline_class.from_pretrained(
tmpdirname, safety_checker=DummyCosmosSafetyChecker(), torch_dtype=torch_dtype_dict
tmpdirname, safety_checker=DummyCosmosSafetyChecker(), dtype=dtype_dict
)
for name, component in loaded_pipe.components.items():
if name == "safety_checker":
continue
if isinstance(component, torch.nn.Module) and hasattr(component, "dtype"):
expected_dtype = torch_dtype_dict.get(name, torch_dtype_dict.get("default", torch.float32))
expected_dtype = dtype_dict.get(name, dtype_dict.get("default", torch.float32))
self.assertEqual(
component.dtype,
expected_dtype,
@@ -362,7 +362,7 @@ class CosmosVideoToWorldPipelineFastTests(PipelineTesterMixin, unittest.TestCase
)
def test_dtype_alias(self):
# `dtype` is an alias for `torch_dtype` in `from_pretrained`.
# `torch_dtype` is deprecated in favor of `dtype` in `from_pretrained`.
components = self.get_dummy_components()
pipe = self.pipeline_class(**components)

2
tests/pipelines/prx/test_pipeline_prx.py
View File

@@ -274,7 +274,7 @@ class PRXPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
pass
@unittest.skip("Custom T5GemmaEncoder not compatible with transformers v5.")
def test_torch_dtype_dict(self):
def test_dtype_dict(self):
pass
@unittest.skip("Custom T5GemmaEncoder not compatible with transformers v5.")

87
tests/pipelines/test_pipelines.py
View File

@@ -1342,7 +1342,7 @@ class CustomPipelineTests(unittest.TestCase):
custom_pipeline="clip_guided_stable_diffusion",
clip_model=clip_model,
feature_extractor=feature_extractor,
torch_dtype=torch.float16,
dtype=torch.float16,
)
pipeline.enable_attention_slicing()
pipeline = pipeline.to(torch_device)
@@ -1938,7 +1938,6 @@ class PipelineFastTests(unittest.TestCase):
sd4 = sd.to(device=device_type)
sd5 = sd.to(torch_device=device_type)
sd6 = sd.to(device_type, dtype=torch.float32)
sd7 = sd.to(device_type, torch_dtype=torch.float32)
assert sd1.device.type == device_type
assert sd2.device.type == device_type
@@ -1946,33 +1945,27 @@ class PipelineFastTests(unittest.TestCase):
assert sd4.device.type == device_type
assert sd5.device.type == device_type
assert sd6.device.type == device_type
assert sd7.device.type == device_type
sd1 = sd.to(torch.float16)
sd2 = sd.to(None, torch.float16)
sd3 = sd.to(dtype=torch.float16)
sd4 = sd.to(dtype=torch.float16)
sd5 = sd.to(None, dtype=torch.float16)
sd6 = sd.to(None, torch_dtype=torch.float16)
assert sd1.dtype == torch.float16
assert sd2.dtype == torch.float16
assert sd3.dtype == torch.float16
assert sd4.dtype == torch.float16
assert sd5.dtype == torch.float16
assert sd6.dtype == torch.float16
sd1 = sd.to(device=device_type, dtype=torch.float16)
sd2 = sd.to(torch_device=device_type, torch_dtype=torch.float16)
sd3 = sd.to(device_type, torch.float16)
sd2 = sd.to(device_type, torch.float16)
assert sd1.dtype == torch.float16
assert sd2.dtype == torch.float16
assert sd3.dtype == torch.float16
assert sd1.device.type == device_type
assert sd2.device.type == device_type
assert sd3.device.type == device_type
def test_pipe_same_device_id_offload(self):
unet = self.dummy_cond_unet()
@@ -2011,6 +2004,82 @@ class PipelineFastTests(unittest.TestCase):
)
assert "dduf_file" in str(warning_ctx.warning)
def test_torch_dtype_is_deprecated(self):
sd = StableDiffusionPipeline(
unet=self.dummy_cond_unet(),
scheduler=PNDMScheduler(skip_prk_steps=True),
vae=self.dummy_vae,
text_encoder=self.dummy_text_encoder,
tokenizer=CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip"),
safety_checker=None,
feature_extractor=self.dummy_extractor,
)
with tempfile.TemporaryDirectory() as tmpdirname:
sd.save_pretrained(tmpdirname)
with self.assertWarns(FutureWarning) as warning_ctx:
loaded_sd = StableDiffusionPipeline.from_pretrained(tmpdirname, torch_dtype=torch.float16)
assert "torch_dtype" in str(warning_ctx.warning)
assert "Please use `dtype` instead." in str(warning_ctx.warning)
# The deprecated argument is still honoured until it is removed.
assert loaded_sd.unet.dtype == torch.float16
def test_dtype_does_not_warn(self):
sd = StableDiffusionPipeline(
unet=self.dummy_cond_unet(),
scheduler=PNDMScheduler(skip_prk_steps=True),
vae=self.dummy_vae,
text_encoder=self.dummy_text_encoder,
tokenizer=CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip"),
safety_checker=None,
feature_extractor=self.dummy_extractor,
)
with tempfile.TemporaryDirectory() as tmpdirname:
sd.save_pretrained(tmpdirname)
with warnings.catch_warnings(record=True) as caught:
warnings.simplefilter("always")
loaded_sd = StableDiffusionPipeline.from_pretrained(tmpdirname, dtype=torch.float16)
assert not [w for w in caught if issubclass(w.category, FutureWarning) and "torch_dtype" in str(w.message)]
assert loaded_sd.unet.dtype == torch.float16
def test_torch_dtype_and_dtype_raises(self):
sd = StableDiffusionPipeline(
unet=self.dummy_cond_unet(),
scheduler=PNDMScheduler(skip_prk_steps=True),
vae=self.dummy_vae,
text_encoder=self.dummy_text_encoder,
tokenizer=CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip"),
safety_checker=None,
feature_extractor=self.dummy_extractor,
)
with tempfile.TemporaryDirectory() as tmpdirname:
sd.save_pretrained(tmpdirname)
with self.assertRaises(ValueError) as error_context:
_ = StableDiffusionPipeline.from_pretrained(tmpdirname, dtype=torch.float16, torch_dtype=torch.float16)
assert "passed both `dtype` and `torch_dtype`" in str(error_context.exception)
def test_from_pipe_torch_dtype_is_deprecated(self):
sd = StableDiffusionPipeline(
unet=self.dummy_cond_unet(),
scheduler=PNDMScheduler(skip_prk_steps=True),
vae=self.dummy_vae,
text_encoder=self.dummy_text_encoder,
tokenizer=CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip"),
safety_checker=None,
feature_extractor=self.dummy_extractor,
)
with self.assertWarns(FutureWarning) as warning_ctx:
img2img = StableDiffusionImg2ImgPipeline.from_pipe(sd, torch_dtype=torch.float16)
assert "torch_dtype" in str(warning_ctx.warning)
assert img2img.unet.dtype == torch.float16
@pytest.mark.xfail(condition=is_transformers_version(">", "4.56.2"), reason="Some import error", strict=False)
def test_wrong_model(self):
tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

18
tests/pipelines/test_pipelines_auto.py
View File

@@ -467,9 +467,7 @@ class AutoPipelineIntegrationTest(unittest.TestCase):
def test_pipe_auto(self):
for model_name, model_repo in PRETRAINED_MODEL_REPO_MAPPING.items():
# test txt2img
pipe_txt2img = AutoPipelineForText2Image.from_pretrained(
model_repo, variant="fp16", torch_dtype=torch.float16
)
pipe_txt2img = AutoPipelineForText2Image.from_pretrained(model_repo, variant="fp16", dtype=torch.float16)
self.assertIsInstance(pipe_txt2img, AUTO_TEXT2IMAGE_PIPELINES_MAPPING[model_name])
pipe_to = AutoPipelineForText2Image.from_pipe(pipe_txt2img)
@@ -487,9 +485,7 @@ class AutoPipelineIntegrationTest(unittest.TestCase):
# test img2img
pipe_img2img = AutoPipelineForImage2Image.from_pretrained(
model_repo, variant="fp16", torch_dtype=torch.float16
)
pipe_img2img = AutoPipelineForImage2Image.from_pretrained(model_repo, variant="fp16", dtype=torch.float16)
self.assertIsInstance(pipe_img2img, AUTO_IMAGE2IMAGE_PIPELINES_MAPPING[model_name])
pipe_to = AutoPipelineForText2Image.from_pipe(pipe_img2img)
@@ -509,7 +505,7 @@ class AutoPipelineIntegrationTest(unittest.TestCase):
if "kandinsky" not in model_name:
pipe_inpaint = AutoPipelineForInpainting.from_pretrained(
model_repo, variant="fp16", torch_dtype=torch.float16
model_repo, variant="fp16", dtype=torch.float16
)
self.assertIsInstance(pipe_inpaint, AUTO_INPAINT_PIPELINES_MAPPING[model_name])
@@ -534,7 +530,7 @@ class AutoPipelineIntegrationTest(unittest.TestCase):
# test from_pretrained
for pipe_from_class in auto_pipes:
pipe_from = pipe_from_class.from_pretrained(model_repo, variant="fp16", torch_dtype=torch.float16)
pipe_from = pipe_from_class.from_pretrained(model_repo, variant="fp16", dtype=torch.float16)
pipe_from_config = dict(pipe_from.config)
for pipe_to_class in auto_pipes:
@@ -552,17 +548,17 @@ class AutoPipelineIntegrationTest(unittest.TestCase):
controlnet = ControlNetModel.from_pretrained(controlnet_repo, torch_dtype=torch.float16)
pipe_txt2img = AutoPipelineForText2Image.from_pretrained(
model_repo, controlnet=controlnet, torch_dtype=torch.float16
model_repo, controlnet=controlnet, dtype=torch.float16
)
self.assertIsInstance(pipe_txt2img, AUTO_TEXT2IMAGE_PIPELINES_MAPPING["stable-diffusion-controlnet"])
pipe_img2img = AutoPipelineForImage2Image.from_pretrained(
model_repo, controlnet=controlnet, torch_dtype=torch.float16
model_repo, controlnet=controlnet, dtype=torch.float16
)
self.assertIsInstance(pipe_img2img, AUTO_IMAGE2IMAGE_PIPELINES_MAPPING["stable-diffusion-controlnet"])
pipe_inpaint = AutoPipelineForInpainting.from_pretrained(
model_repo, controlnet=controlnet, torch_dtype=torch.float16
model_repo, controlnet=controlnet, dtype=torch.float16
)
self.assertIsInstance(pipe_inpaint, AUTO_INPAINT_PIPELINES_MAPPING["stable-diffusion-controlnet"])

12
tests/pipelines/test_pipelines_common.py
View File

@@ -1464,7 +1464,7 @@ class PipelineTesterMixin:
with tempfile.TemporaryDirectory() as tmpdir:
pipe.save_pretrained(tmpdir)
pipe_loaded = self.pipeline_class.from_pretrained(tmpdir, torch_dtype=torch.float16)
pipe_loaded = self.pipeline_class.from_pretrained(tmpdir, dtype=torch.float16)
for component in pipe_loaded.components.values():
if hasattr(component, "set_default_attn_processor"):
component.set_default_attn_processor()
@@ -2341,7 +2341,7 @@ class PipelineTesterMixin:
self.assertTrue(np.allclose(output_without_group_offloading, output_with_group_offloading1, atol=1e-4))
self.assertTrue(np.allclose(output_without_group_offloading, output_with_group_offloading2, atol=1e-4))
def test_torch_dtype_dict(self):
def test_dtype_dict(self):
components = self.get_dummy_components()
if not components:
self.skipTest("No dummy components defined.")
@@ -2351,12 +2351,12 @@ class PipelineTesterMixin:
with tempfile.TemporaryDirectory(ignore_cleanup_errors=True) as tmpdirname:
pipe.save_pretrained(tmpdirname, safe_serialization=False)
torch_dtype_dict = {specified_key: torch.bfloat16, "default": torch.float16}
loaded_pipe = self.pipeline_class.from_pretrained(tmpdirname, torch_dtype=torch_dtype_dict)
dtype_dict = {specified_key: torch.bfloat16, "default": torch.float16}
loaded_pipe = self.pipeline_class.from_pretrained(tmpdirname, dtype=dtype_dict)
for name, component in loaded_pipe.components.items():
if isinstance(component, torch.nn.Module) and hasattr(component, "dtype"):
expected_dtype = torch_dtype_dict.get(name, torch_dtype_dict.get("default", torch.float32))
expected_dtype = dtype_dict.get(name, dtype_dict.get("default", torch.float32))
self.assertEqual(
component.dtype,
expected_dtype,
@@ -2364,7 +2364,7 @@ class PipelineTesterMixin:
)
def test_dtype_alias(self):
# `dtype` is an alias for `torch_dtype` in `from_pretrained`.
# `torch_dtype` is deprecated in favor of `dtype` in `from_pretrained`.
components = self.get_dummy_components()
pipe = self.pipeline_class(**components)