Merge branch 'main' into device-map-direct

docs: improve docstring scheduling_ddim_inverse.py

src/diffusers/models/transformers/transformer_longcat_image.py

@@ -406,6 +406,7 @@ class LongCatImageTransformer2DModel(
     """
 
     _supports_gradient_checkpointing = True
+    _repeated_blocks = ["LongCatImageTransformerBlock", "LongCatImageSingleTransformerBlock"]
 
     @register_to_config
     def __init__(

src/diffusers/pipelines/pipeline_utils.py

@@ -111,7 +111,7 @@ LIBRARIES = []
 for library in LOADABLE_CLASSES:
     LIBRARIES.append(library)
 
-SUPPORTED_DEVICE_MAP = ["balanced"] + [get_device()]
+SUPPORTED_DEVICE_MAP = ["balanced"] + [get_device(), "cpu"]
 
 logger = logging.get_logger(__name__)
 
@@ -467,8 +467,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
         pipeline_is_sequentially_offloaded = any(
             module_is_sequentially_offloaded(module) for _, module in self.components.items()
         )
-
-        is_pipeline_device_mapped = self.hf_device_map is not None and len(self.hf_device_map) > 1
+        is_pipeline_device_mapped = self._is_pipeline_device_mapped()
         if is_pipeline_device_mapped:
             raise ValueError(
                 "It seems like you have activated a device mapping strategy on the pipeline which doesn't allow explicit device placement using `to()`. You can call `reset_device_map()` to remove the existing device map from the pipeline."
@@ -1187,7 +1186,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
         """
         self._maybe_raise_error_if_group_offload_active(raise_error=True)
 
-        is_pipeline_device_mapped = self.hf_device_map is not None and len(self.hf_device_map) > 1
+        is_pipeline_device_mapped = self._is_pipeline_device_mapped()
         if is_pipeline_device_mapped:
             raise ValueError(
                 "It seems like you have activated a device mapping strategy on the pipeline so calling `enable_model_cpu_offload() isn't allowed. You can call `reset_device_map()` first and then call `enable_model_cpu_offload()`."
@@ -1311,7 +1310,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
             raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
         self.remove_all_hooks()
 
-        is_pipeline_device_mapped = self.hf_device_map is not None and len(self.hf_device_map) > 1
+        is_pipeline_device_mapped = self._is_pipeline_device_mapped()
         if is_pipeline_device_mapped:
             raise ValueError(
                 "It seems like you have activated a device mapping strategy on the pipeline so calling `enable_sequential_cpu_offload() isn't allowed. You can call `reset_device_map()` first and then call `enable_sequential_cpu_offload()`."
@@ -2200,6 +2199,21 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
                 return True
         return False
 
+    def _is_pipeline_device_mapped(self):
+        # We support passing `device_map="cuda"`, for example. This is helpful, in case
+        # users want to pass `device_map="cpu"` when initializing a pipeline. This explicit declaration is desirable
+        # in limited VRAM environments because quantized models often initialize directly on the accelerator.
+        device_map = self.hf_device_map
+        is_device_type_map = False
+        if isinstance(device_map, str):
+            try:
+                torch.device(device_map)
+                is_device_type_map = True
+            except RuntimeError:
+                pass
+
+        return not is_device_type_map and isinstance(device_map, dict) and len(device_map) > 1
+
 
 class StableDiffusionMixin:
     r"""

src/diffusers/schedulers/scheduling_ddim_flax.py

@@ -22,6 +22,7 @@ import flax
 import jax.numpy as jnp
 
 from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import logging
 from .scheduling_utils_flax import (
     CommonSchedulerState,
     FlaxKarrasDiffusionSchedulers,
@@ -32,6 +33,9 @@ from .scheduling_utils_flax import (
 )
 
 
+logger = logging.get_logger(__name__)
+
+
 @flax.struct.dataclass
 class DDIMSchedulerState:
     common: CommonSchedulerState
@@ -125,6 +129,10 @@ class FlaxDDIMScheduler(FlaxSchedulerMixin, ConfigMixin):
         prediction_type: str = "epsilon",
         dtype: jnp.dtype = jnp.float32,
     ):
+        logger.warning(
+            "Flax classes are deprecated and will be removed in Diffusers v1.0.0. We "
+            "recommend migrating to PyTorch classes or pinning your version of Diffusers."
+        )
         self.dtype = dtype
 
     def create_state(self, common: Optional[CommonSchedulerState] = None) -> DDIMSchedulerState:
@@ -152,7 +160,10 @@ class FlaxDDIMScheduler(FlaxSchedulerMixin, ConfigMixin):
         )
 
     def scale_model_input(
-        self, state: DDIMSchedulerState, sample: jnp.ndarray, timestep: Optional[int] = None
+        self,
+        state: DDIMSchedulerState,
+        sample: jnp.ndarray,
+        timestep: Optional[int] = None,
     ) -> jnp.ndarray:
         """
         Args:
@@ -190,7 +201,9 @@ class FlaxDDIMScheduler(FlaxSchedulerMixin, ConfigMixin):
     def _get_variance(self, state: DDIMSchedulerState, timestep, prev_timestep):
         alpha_prod_t = state.common.alphas_cumprod[timestep]
         alpha_prod_t_prev = jnp.where(
-            prev_timestep >= 0, state.common.alphas_cumprod[prev_timestep], state.final_alpha_cumprod
+            prev_timestep >= 0,
+            state.common.alphas_cumprod[prev_timestep],
+            state.final_alpha_cumprod,
         )
         beta_prod_t = 1 - alpha_prod_t
         beta_prod_t_prev = 1 - alpha_prod_t_prev

src/diffusers/schedulers/scheduling_ddim_inverse.py

@@ -99,7 +99,7 @@ def betas_for_alpha_bar(
 
 
 # Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr
-def rescale_zero_terminal_snr(betas):
+def rescale_zero_terminal_snr(betas: torch.Tensor) -> torch.Tensor:
     """
     Rescales betas to have zero terminal SNR Based on https://huggingface.co/papers/2305.08891 (Algorithm 1)
 
@@ -187,14 +187,14 @@ class DDIMInverseScheduler(SchedulerMixin, ConfigMixin):
         num_train_timesteps: int = 1000,
         beta_start: float = 0.0001,
         beta_end: float = 0.02,
-        beta_schedule: str = "linear",
+        beta_schedule: Literal["linear", "scaled_linear", "squaredcos_cap_v2"] = "linear",
         trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
         clip_sample: bool = True,
         set_alpha_to_one: bool = True,
         steps_offset: int = 0,
-        prediction_type: str = "epsilon",
+        prediction_type: Literal["epsilon", "sample", "v_prediction"] = "epsilon",
         clip_sample_range: float = 1.0,
-        timestep_spacing: str = "leading",
+        timestep_spacing: Literal["leading", "trailing"] = "leading",
         rescale_betas_zero_snr: bool = False,
         **kwargs,
     ):
@@ -210,7 +210,15 @@ class DDIMInverseScheduler(SchedulerMixin, ConfigMixin):
             self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
         elif beta_schedule == "scaled_linear":
             # this schedule is very specific to the latent diffusion model.
-            self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
+            self.betas = (
+                torch.linspace(
+                    beta_start**0.5,
+                    beta_end**0.5,
+                    num_train_timesteps,
+                    dtype=torch.float32,
+                )
+                ** 2
+            )
         elif beta_schedule == "squaredcos_cap_v2":
             # Glide cosine schedule
             self.betas = betas_for_alpha_bar(num_train_timesteps)
@@ -256,7 +264,11 @@ class DDIMInverseScheduler(SchedulerMixin, ConfigMixin):
         """
         return sample
 
-    def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None):
+    def set_timesteps(
+        self,
+        num_inference_steps: int,
+        device: Optional[Union[str, torch.device]] = None,
+    ) -> None:
         """
         Sets the discrete timesteps used for the diffusion chain (to be run before inference).
 
@@ -308,20 +320,10 @@ class DDIMInverseScheduler(SchedulerMixin, ConfigMixin):
         Args:
             model_output (`torch.Tensor`):
                 The direct output from learned diffusion model.
-            timestep (`float`):
+            timestep (`int`):
                 The current discrete timestep in the diffusion chain.
             sample (`torch.Tensor`):
                 A current instance of a sample created by the diffusion process.
-            eta (`float`):
-                The weight of noise for added noise in diffusion step.
-            use_clipped_model_output (`bool`, defaults to `False`):
-                If `True`, computes "corrected" `model_output` from the clipped predicted original sample. Necessary
-                because predicted original sample is clipped to [-1, 1] when `self.config.clip_sample` is `True`. If no
-                clipping has happened, "corrected" `model_output` would coincide with the one provided as input and
-                `use_clipped_model_output` has no effect.
-            variance_noise (`torch.Tensor`):
-                Alternative to generating noise with `generator` by directly providing the noise for the variance
-                itself. Useful for methods such as [`CycleDiffusion`].
             return_dict (`bool`, *optional*, defaults to `True`):
                 Whether or not to return a [`~schedulers.scheduling_ddim_inverse.DDIMInverseSchedulerOutput`] or
                 `tuple`.
@@ -335,7 +337,8 @@ class DDIMInverseScheduler(SchedulerMixin, ConfigMixin):
         # 1. get previous step value (=t+1)
         prev_timestep = timestep
         timestep = min(
-            timestep - self.config.num_train_timesteps // self.num_inference_steps, self.config.num_train_timesteps - 1
+            timestep - self.config.num_train_timesteps // self.num_inference_steps,
+            self.config.num_train_timesteps - 1,
         )
 
         # 2. compute alphas, betas
@@ -378,5 +381,5 @@ class DDIMInverseScheduler(SchedulerMixin, ConfigMixin):
             return (prev_sample, pred_original_sample)
         return DDIMSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample)
 
-    def __len__(self):
+    def __len__(self) -> int:
         return self.config.num_train_timesteps