Fix import with Flax but without PyTorch

src/diffusers/__init__.py

@@ -73,6 +73,7 @@
         FlaxKarrasVeScheduler,
         FlaxLMSDiscreteScheduler,
         FlaxPNDMScheduler,
+        FlaxSchedulerMixin,
         FlaxScoreSdeVeScheduler,
     )
 else:

src/diffusers/modeling_flax_utils.py

@@ -27,8 +27,8 @@
 from huggingface_hub.utils import EntryNotFoundError, RepositoryNotFoundError, RevisionNotFoundError
 from requests import HTTPError
 
+from . import is_torch_available
 from .modeling_flax_pytorch_utils import convert_pytorch_state_dict_to_flax
-from .modeling_utils import load_state_dict
 from .utils import (
     CONFIG_NAME,
     DIFFUSERS_CACHE,
@@ -391,6 +391,14 @@ def from_pretrained(
                 )
 
         if from_pt:
+            if is_torch_available():
+                from .modeling_utils import load_state_dict
+            else:
+                raise EnvironmentError(
+                    "Can't load the model in PyTorch format because PyTorch is not installed. "
+                    "Please, install PyTorch or use native Flax weights."
+                )
+
             # Step 1: Get the pytorch file
             pytorch_model_file = load_state_dict(model_file)
 

src/diffusers/pipeline_flax_utils.py

@@ -30,7 +30,7 @@
 
 from .configuration_utils import ConfigMixin
 from .modeling_flax_utils import FLAX_WEIGHTS_NAME, FlaxModelMixin
-from .schedulers.scheduling_utils import SCHEDULER_CONFIG_NAME, SchedulerMixin
+from .schedulers.scheduling_utils_flax import SCHEDULER_CONFIG_NAME, FlaxSchedulerMixin
 from .utils import CONFIG_NAME, DIFFUSERS_CACHE, BaseOutput, is_transformers_available, logging
 
 
@@ -46,7 +46,7 @@
 LOADABLE_CLASSES = {
     "diffusers": {
         "FlaxModelMixin": ["save_pretrained", "from_pretrained"],
-        "SchedulerMixin": ["save_config", "from_config"],
+        "FlaxSchedulerMixin": ["save_config", "from_config"],
         "FlaxDiffusionPipeline": ["save_pretrained", "from_pretrained"],
     },
     "transformers": {
@@ -436,7 +436,7 @@ def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.P
                     else:
                         loaded_sub_model, loaded_params = load_method(loadable_folder, _do_init=False)
                     params[name] = loaded_params
-                elif issubclass(class_obj, SchedulerMixin):
+                elif issubclass(class_obj, FlaxSchedulerMixin):
                     loaded_sub_model, scheduler_state = load_method(loadable_folder)
                     params[name] = scheduler_state
                 else:

src/diffusers/pipelines/__init__.py

@@ -1,12 +1,16 @@
 from ..utils import is_flax_available, is_onnx_available, is_torch_available, is_transformers_available
-from .ddim import DDIMPipeline
-from .ddpm import DDPMPipeline
-from .latent_diffusion_uncond import LDMPipeline
-from .pndm import PNDMPipeline
-from .score_sde_ve import ScoreSdeVePipeline
-from .stochastic_karras_ve import KarrasVePipeline
 
 
+if is_torch_available():
+    from .ddim import DDIMPipeline
+    from .ddpm import DDPMPipeline
+    from .latent_diffusion_uncond import LDMPipeline
+    from .pndm import PNDMPipeline
+    from .score_sde_ve import ScoreSdeVePipeline
+    from .stochastic_karras_ve import KarrasVePipeline
+else:
+    from ..utils.dummy_pt_objects import *  # noqa F403
+
 if is_torch_available() and is_transformers_available():
     from .latent_diffusion import LDMTextToImagePipeline
     from .stable_diffusion import (

src/diffusers/pipelines/stable_diffusion/__init__.py

@@ -6,7 +6,7 @@
 import PIL
 from PIL import Image
 
-from ...utils import BaseOutput, is_flax_available, is_onnx_available, is_transformers_available
+from ...utils import BaseOutput, is_flax_available, is_onnx_available, is_torch_available, is_transformers_available
 
 
 @dataclass
@@ -27,7 +27,7 @@ class StableDiffusionPipelineOutput(BaseOutput):
     nsfw_content_detected: List[bool]
 
 
-if is_transformers_available():
+if is_transformers_available() and is_torch_available():
     from .pipeline_stable_diffusion import StableDiffusionPipeline
     from .pipeline_stable_diffusion_img2img import StableDiffusionImg2ImgPipeline
     from .pipeline_stable_diffusion_inpaint import StableDiffusionInpaintPipeline

src/diffusers/schedulers/__init__.py

@@ -34,10 +34,12 @@
     from .scheduling_lms_discrete_flax import FlaxLMSDiscreteScheduler
     from .scheduling_pndm_flax import FlaxPNDMScheduler
     from .scheduling_sde_ve_flax import FlaxScoreSdeVeScheduler
+    from .scheduling_utils_flax import FlaxSchedulerMixin
 else:
     from ..utils.dummy_flax_objects import *  # noqa F403
 
-if is_scipy_available():
+
+if is_scipy_available() and is_torch_available():
     from .scheduling_lms_discrete import LMSDiscreteScheduler
 else:
     from ..utils.dummy_torch_and_scipy_objects import *  # noqa F403

src/diffusers/schedulers/scheduling_ddim_flax.py

@@ -23,7 +23,7 @@
 import jax.numpy as jnp
 
 from ..configuration_utils import ConfigMixin, register_to_config
-from .scheduling_utils import SchedulerMixin, SchedulerOutput
+from .scheduling_utils_flax import FlaxSchedulerMixin, FlaxSchedulerOutput
 
 
 def betas_for_alpha_bar(num_diffusion_timesteps, max_beta=0.999) -> jnp.ndarray:
@@ -68,11 +68,11 @@ def create(cls, num_train_timesteps: int, alphas_cumprod: jnp.ndarray):
 
 
 @dataclass
-class FlaxSchedulerOutput(SchedulerOutput):
+class FlaxDDIMSchedulerOutput(FlaxSchedulerOutput):
     state: DDIMSchedulerState
 
 
-class FlaxDDIMScheduler(SchedulerMixin, ConfigMixin):
+class FlaxDDIMScheduler(FlaxSchedulerMixin, ConfigMixin):
     """
     Denoising diffusion implicit models is a scheduler that extends the denoising procedure introduced in denoising
     diffusion probabilistic models (DDPMs) with non-Markovian guidance.
@@ -183,7 +183,7 @@ def step(
         timestep: int,
         sample: jnp.ndarray,
         return_dict: bool = True,
-    ) -> Union[FlaxSchedulerOutput, Tuple]:
+    ) -> Union[FlaxDDIMSchedulerOutput, Tuple]:
         """
         Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion
         process from the learned model outputs (most often the predicted noise).
@@ -197,11 +197,11 @@ def step(
             key (`random.KeyArray`): a PRNG key.
             eta (`float`): weight of noise for added noise in diffusion step.
             use_clipped_model_output (`bool`): TODO
-            return_dict (`bool`): option for returning tuple rather than SchedulerOutput class
+            return_dict (`bool`): option for returning tuple rather than FlaxDDIMSchedulerOutput class
 
         Returns:
-            [`FlaxSchedulerOutput`] or `tuple`: [`FlaxSchedulerOutput`] if `return_dict` is True, otherwise a `tuple`.
-            When returning a tuple, the first element is the sample tensor.
+            [`FlaxDDIMSchedulerOutput`] or `tuple`: [`FlaxDDIMSchedulerOutput`] if `return_dict` is True, otherwise a
+            `tuple`. When returning a tuple, the first element is the sample tensor.
 
         """
         if state.num_inference_steps is None:
@@ -252,7 +252,7 @@ def step(
         if not return_dict:
             return (prev_sample, state)
 
-        return FlaxSchedulerOutput(prev_sample=prev_sample, state=state)
+        return FlaxDDIMSchedulerOutput(prev_sample=prev_sample, state=state)
 
     def add_noise(
         self,

src/diffusers/schedulers/scheduling_ddpm_flax.py

@@ -23,7 +23,7 @@
 from jax import random
 
 from ..configuration_utils import ConfigMixin, register_to_config
-from .scheduling_utils import SchedulerMixin, SchedulerOutput
+from .scheduling_utils_flax import FlaxSchedulerMixin, FlaxSchedulerOutput
 
 
 def betas_for_alpha_bar(num_diffusion_timesteps, max_beta=0.999) -> jnp.ndarray:
@@ -67,11 +67,11 @@ def create(cls, num_train_timesteps: int):
 
 
 @dataclass
-class FlaxSchedulerOutput(SchedulerOutput):
+class FlaxDDPMSchedulerOutput(FlaxSchedulerOutput):
     state: DDPMSchedulerState
 
 
-class FlaxDDPMScheduler(SchedulerMixin, ConfigMixin):
+class FlaxDDPMScheduler(FlaxSchedulerMixin, ConfigMixin):
     """
     Denoising diffusion probabilistic models (DDPMs) explores the connections between denoising score matching and
     Langevin dynamics sampling.
@@ -191,7 +191,7 @@ def step(
         key: random.KeyArray,
         predict_epsilon: bool = True,
         return_dict: bool = True,
-    ) -> Union[FlaxSchedulerOutput, Tuple]:
+    ) -> Union[FlaxDDPMSchedulerOutput, Tuple]:
         """
         Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion
         process from the learned model outputs (most often the predicted noise).
@@ -205,11 +205,11 @@ def step(
             key (`random.KeyArray`): a PRNG key.
             predict_epsilon (`bool`):
                 optional flag to use when model predicts the samples directly instead of the noise, epsilon.
-            return_dict (`bool`): option for returning tuple rather than SchedulerOutput class
+            return_dict (`bool`): option for returning tuple rather than FlaxDDPMSchedulerOutput class
 
         Returns:
-            [`FlaxSchedulerOutput`] or `tuple`: [`FlaxSchedulerOutput`] if `return_dict` is True, otherwise a `tuple`.
-            When returning a tuple, the first element is the sample tensor.
+            [`FlaxDDPMSchedulerOutput`] or `tuple`: [`FlaxDDPMSchedulerOutput`] if `return_dict` is True, otherwise a
+            `tuple`. When returning a tuple, the first element is the sample tensor.
 
         """
         t = timestep
@@ -257,7 +257,7 @@ def step(
         if not return_dict:
             return (pred_prev_sample, state)
 
-        return FlaxSchedulerOutput(prev_sample=pred_prev_sample, state=state)
+        return FlaxDDPMSchedulerOutput(prev_sample=pred_prev_sample, state=state)
 
     def add_noise(
         self,

src/diffusers/schedulers/scheduling_karras_ve_flax.py

@@ -22,7 +22,7 @@
 
 from ..configuration_utils import ConfigMixin, register_to_config
 from ..utils import BaseOutput
-from .scheduling_utils import SchedulerMixin
+from .scheduling_utils_flax import FlaxSchedulerMixin
 
 
 @flax.struct.dataclass
@@ -56,7 +56,7 @@ class FlaxKarrasVeOutput(BaseOutput):
     state: KarrasVeSchedulerState
 
 
-class FlaxKarrasVeScheduler(SchedulerMixin, ConfigMixin):
+class FlaxKarrasVeScheduler(FlaxSchedulerMixin, ConfigMixin):
     """
     Stochastic sampling from Karras et al. [1] tailored to the Variance-Expanding (VE) models [2]. Use Algorithm 2 and
     the VE column of Table 1 from [1] for reference.
@@ -170,7 +170,7 @@ def step(
             sigma_hat (`float`): TODO
             sigma_prev (`float`): TODO
             sample_hat (`torch.FloatTensor` or `np.ndarray`): TODO
-            return_dict (`bool`): option for returning tuple rather than SchedulerOutput class
+            return_dict (`bool`): option for returning tuple rather than FlaxKarrasVeOutput class
 
         Returns:
             [`~schedulers.scheduling_karras_ve_flax.FlaxKarrasVeOutput`] or `tuple`: Updated sample in the diffusion
@@ -209,7 +209,7 @@ def step_correct(
             sample_hat (`torch.FloatTensor` or `np.ndarray`): TODO
             sample_prev (`torch.FloatTensor` or `np.ndarray`): TODO
             derivative (`torch.FloatTensor` or `np.ndarray`): TODO
-            return_dict (`bool`): option for returning tuple rather than SchedulerOutput class
+            return_dict (`bool`): option for returning tuple rather than FlaxKarrasVeOutput class
 
         Returns:
             prev_sample (TODO): updated sample in the diffusion chain. derivative (TODO): TODO

src/diffusers/schedulers/scheduling_lms_discrete_flax.py

@@ -20,7 +20,7 @@
 from scipy import integrate
 
 from ..configuration_utils import ConfigMixin, register_to_config
-from .scheduling_utils import SchedulerMixin, SchedulerOutput
+from .scheduling_utils_flax import FlaxSchedulerMixin, FlaxSchedulerOutput
 
 
 @flax.struct.dataclass
@@ -37,11 +37,11 @@ def create(cls, num_train_timesteps: int, sigmas: jnp.ndarray):
 
 
 @dataclass
-class FlaxSchedulerOutput(SchedulerOutput):
+class FlaxLMSSchedulerOutput(FlaxSchedulerOutput):
     state: LMSDiscreteSchedulerState
 
 
-class FlaxLMSDiscreteScheduler(SchedulerMixin, ConfigMixin):
+class FlaxLMSDiscreteScheduler(FlaxSchedulerMixin, ConfigMixin):
     """
     Linear Multistep Scheduler for discrete beta schedules. Based on the original k-diffusion implementation by
     Katherine Crowson:
@@ -145,7 +145,7 @@ def step(
         sample: jnp.ndarray,
         order: int = 4,
         return_dict: bool = True,
-    ) -> Union[SchedulerOutput, Tuple]:
+    ) -> Union[FlaxLMSSchedulerOutput, Tuple]:
         """
         Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion
         process from the learned model outputs (most often the predicted noise).
@@ -157,11 +157,11 @@ def step(
             sample (`jnp.ndarray`):
                 current instance of sample being created by diffusion process.
             order: coefficient for multi-step inference.
-            return_dict (`bool`): option for returning tuple rather than SchedulerOutput class
+            return_dict (`bool`): option for returning tuple rather than FlaxLMSSchedulerOutput class
 
         Returns:
-            [`FlaxSchedulerOutput`] or `tuple`: [`FlaxSchedulerOutput`] if `return_dict` is True, otherwise a `tuple`.
-            When returning a tuple, the first element is the sample tensor.
+            [`FlaxLMSSchedulerOutput`] or `tuple`: [`FlaxLMSSchedulerOutput`] if `return_dict` is True, otherwise a
+            `tuple`. When returning a tuple, the first element is the sample tensor.
 
         """
         sigma = state.sigmas[timestep]
@@ -187,7 +187,7 @@ def step(
         if not return_dict:
             return (prev_sample, state)
 
-        return FlaxSchedulerOutput(prev_sample=prev_sample, state=state)
+        return FlaxLMSSchedulerOutput(prev_sample=prev_sample, state=state)
 
     def add_noise(
         self,