add custom mesh support

docs: improve docstring scheduling_ddim_inverse.py

docker/diffusers-pytorch-cuda/Dockerfile

@@ -2,7 +2,7 @@ FROM nvidia/cuda:12.1.0-runtime-ubuntu20.04
 LABEL maintainer="Hugging Face"
 LABEL repository="diffusers"
 
-ARG PYTHON_VERSION=3.12
+ARG PYTHON_VERSION=3.11
 ENV DEBIAN_FRONTEND=noninteractive
 
 RUN apt-get -y update \
@@ -32,10 +32,12 @@ RUN uv venv --python ${PYTHON_VERSION} --seed ${VIRTUAL_ENV}
 ENV PATH="$VIRTUAL_ENV/bin:$PATH"
 
 # pre-install the heavy dependencies (these can later be overridden by the deps from setup.py)
+# Install torch, torchvision, and torchaudio together to ensure compatibility
 RUN uv pip install --no-cache-dir \
     torch \
     torchvision \
-    torchaudio
+    torchaudio \
+    --index-url https://download.pytorch.org/whl/cu121
 
 RUN uv pip install --no-cache-dir "git+https://github.com/huggingface/diffusers.git@main#egg=diffusers[test]"
 

docker/diffusers-pytorch-xformers-cuda/Dockerfile

@@ -2,7 +2,7 @@ FROM nvidia/cuda:12.1.0-runtime-ubuntu20.04
 LABEL maintainer="Hugging Face"
 LABEL repository="diffusers"
 
-ARG PYTHON_VERSION=3.12
+ARG PYTHON_VERSION=3.11
 ENV DEBIAN_FRONTEND=noninteractive
 
 RUN apt-get -y update \
@@ -32,10 +32,12 @@ RUN uv venv --python ${PYTHON_VERSION} --seed ${VIRTUAL_ENV}
 ENV PATH="$VIRTUAL_ENV/bin:$PATH"
 
 # pre-install the heavy dependencies (these can later be overridden by the deps from setup.py)
+# Install torch, torchvision, and torchaudio together to ensure compatibility
 RUN uv pip install --no-cache-dir \
     torch \
     torchvision \
-    torchaudio
+    torchaudio \
+    --index-url https://download.pytorch.org/whl/cu121
 
 RUN uv pip install --no-cache-dir "git+https://github.com/huggingface/diffusers.git@main#egg=diffusers[test]"
 

src/diffusers/models/_modeling_parallel.py

@@ -59,6 +59,12 @@ class ContextParallelConfig:
         rotate_method (`str`, *optional*, defaults to `"allgather"`):
             Method to use for rotating key/value states across devices in ring attention. Currently, only `"allgather"`
             is supported.
+        mesh (`torch.distributed.device_mesh.DeviceMesh`, *optional*):
+            A custom device mesh to use for context parallelism. If provided, this mesh will be used instead of
+            creating a new one. This is useful when combining context parallelism with other parallelism strategies
+            (e.g., FSDP, tensor parallelism) that share the same device mesh. The mesh must have both "ring" and
+            "ulysses" dimensions. Use size 1 for dimensions not being used (e.g., `mesh_shape=(2, 1, 4)` with
+            `mesh_dim_names=("ring", "ulysses", "fsdp")` for ring attention only with FSDP).
 
     """
 
@@ -67,6 +73,7 @@ class ContextParallelConfig:
     convert_to_fp32: bool = True
     # TODO: support alltoall
     rotate_method: Literal["allgather", "alltoall"] = "allgather"
+    mesh: Optional[torch.distributed.device_mesh.DeviceMesh] = None
 
     _rank: int = None
     _world_size: int = None
@@ -115,7 +122,7 @@ class ContextParallelConfig:
                 f"The product of `ring_degree` ({self.ring_degree}) and `ulysses_degree` ({self.ulysses_degree}) must not exceed the world size ({world_size})."
             )
 
-        self._flattened_mesh = self._mesh._flatten()
+        self._flattened_mesh = self._mesh["ring", "ulysses"]._flatten()
         self._ring_mesh = self._mesh["ring"]
         self._ulysses_mesh = self._mesh["ulysses"]
         self._ring_local_rank = self._ring_mesh.get_local_rank()

src/diffusers/models/modeling_utils.py

@@ -1569,7 +1569,7 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
         mesh = None
         if config.context_parallel_config is not None:
             cp_config = config.context_parallel_config
-            mesh = torch.distributed.device_mesh.init_device_mesh(
+            mesh = cp_config.mesh or torch.distributed.device_mesh.init_device_mesh(
                 device_type=device_type,
                 mesh_shape=cp_config.mesh_shape,
                 mesh_dim_names=cp_config.mesh_dim_names,

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/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