Add codeQL workflow

Updated CodeQL workflow to use reusable workflow from Hugging Face and simplified language matrix.

docs/source/en/api/pipelines/ltx_video.md

@@ -136,7 +136,7 @@ export_to_video(video, "output.mp4", fps=24)
   - The recommended dtype for the transformer, VAE, and text encoder is `torch.bfloat16`. The VAE and text encoder can also be `torch.float32` or `torch.float16`.
   - For guidance-distilled variants of LTX-Video, set `guidance_scale` to `1.0`. The `guidance_scale` for any other model should be set higher, like `5.0`, for good generation quality.
   - For timestep-aware VAE variants (LTX-Video 0.9.1 and above), set `decode_timestep` to `0.05` and `image_cond_noise_scale` to `0.025`.
-  - For variants that support interpolation between multiple conditioning images and videos (LTX-Video 0.9.5 and above), use similar images and videos for the best results. Divergence from the conditioning inputs may lead to abrupt transitions in the generated video.
+  - For variants that support interpolation between multiple conditioning images and videos (LTX-Video 0.9.5 and above), use similar images and videos for the best results. Divergence from the conditioning inputs may lead to abrupt transitionts in the generated video.
 
 - LTX-Video 0.9.7 includes a spatial latent upscaler and a 13B parameter transformer. During inference, a low resolution video is quickly generated first and then upscaled and refined.
 
@@ -329,7 +329,7 @@ export_to_video(video, "output.mp4", fps=24)
 
   <details>
   <summary>Show example code</summary>
-
+  
   ```python
   import torch
   from diffusers import LTXConditionPipeline, LTXLatentUpsamplePipeline
@@ -474,12 +474,6 @@ export_to_video(video, "output.mp4", fps=24)
 
   </details>
 
-## LTXI2VLongMultiPromptPipeline
-
-[[autodoc]] LTXI2VLongMultiPromptPipeline
-  - all
-  - __call__
-
 ## LTXPipeline
 
 [[autodoc]] LTXPipeline

examples/dreambooth/README_flux2.md

@@ -98,9 +98,6 @@ Flux.2 uses  Mistral Small 3.1 as text encoder which is quite large and can take
 This way, the text encoder model is not loaded into memory during training.
 > [!NOTE] 
 > to enable remote text encoding you must either be logged in to your HuggingFace account (`hf auth login`) OR pass a token with `--hub_token`.
-### FSDP Text Encoder 
-Flux.2 uses  Mistral Small 3.1 as text encoder which is quite large and can take up a lot of memory. To mitigate this, we can use the `--fsdp_text_encoder` flag to enable distributed computation of the prompt embeddings. 
-This way, it distributes the memory cost across multiple nodes.
 ### CPU Offloading 
 To offload parts of the model to CPU memory, you can use `--offload` flag. This will offload the vae and text encoder to CPU memory and only move them to GPU when needed.
 ### Latent Caching 
@@ -169,26 +166,6 @@ To better track our training experiments, we're using the following flags in the
 > [!NOTE]
 > If you want to train using long prompts with the T5 text encoder, you can use `--max_sequence_length` to set the token limit. The default is 77, but it can be increased to as high as 512. Note that this will use more resources and may slow down the training in some cases.
 
-### FSDP on the transformer
-By setting the accelerate configuration with FSDP, the transformer block will be wrapped automatically. E.g. set the configuration to:
-
-```shell
-distributed_type: FSDP
-fsdp_config:
-  fsdp_version: 2
-  fsdp_offload_params: false
-  fsdp_sharding_strategy: HYBRID_SHARD
-  fsdp_auto_wrap_policy: TRANSFOMER_BASED_WRAP
-  fsdp_transformer_layer_cls_to_wrap: Flux2TransformerBlock, Flux2SingleTransformerBlock
-  fsdp_forward_prefetch: true
-  fsdp_sync_module_states: false
-  fsdp_state_dict_type: FULL_STATE_DICT
-  fsdp_use_orig_params: false
-  fsdp_activation_checkpointing: true
-  fsdp_reshard_after_forward: true
-  fsdp_cpu_ram_efficient_loading: false
-```
-
 ## LoRA + DreamBooth
 
 [LoRA](https://huggingface.co/docs/peft/conceptual_guides/adapter#low-rank-adaptation-lora) is a popular parameter-efficient fine-tuning technique that allows you to achieve full-finetuning like performance but with a fraction of learnable parameters.

examples/dreambooth/train_dreambooth_lora_flux2.py

@@ -44,7 +44,6 @@ import shutil
 import warnings
 from contextlib import nullcontext
 from pathlib import Path
-from typing import Any
 
 import numpy as np
 import torch
@@ -76,16 +75,13 @@ from diffusers import (
 from diffusers.optimization import get_scheduler
 from diffusers.training_utils import (
     _collate_lora_metadata,
-    _to_cpu_contiguous,
     cast_training_params,
     compute_density_for_timestep_sampling,
     compute_loss_weighting_for_sd3,
     find_nearest_bucket,
     free_memory,
-    get_fsdp_kwargs_from_accelerator,
     offload_models,
     parse_buckets_string,
-    wrap_with_fsdp,
 )
 from diffusers.utils import (
     check_min_version,
@@ -97,9 +93,6 @@ from diffusers.utils.import_utils import is_torch_npu_available
 from diffusers.utils.torch_utils import is_compiled_module
 
 
-if getattr(torch, "distributed", None) is not None:
-    import torch.distributed as dist
-
 if is_wandb_available():
     import wandb
 
@@ -729,7 +722,6 @@ def parse_args(input_args=None):
     )
     parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
     parser.add_argument("--enable_npu_flash_attention", action="store_true", help="Enabla Flash Attention for NPU")
-    parser.add_argument("--fsdp_text_encoder", action="store_true", help="Use FSDP for text encoder")
 
     if input_args is not None:
         args = parser.parse_args(input_args)
@@ -1227,11 +1219,7 @@ def main(args):
         if args.bnb_quantization_config_path is not None
         else {"device": accelerator.device, "dtype": weight_dtype}
     )
-
-    is_fsdp = accelerator.state.fsdp_plugin is not None
-    if not is_fsdp:
-        transformer.to(**transformer_to_kwargs)
-
+    transformer.to(**transformer_to_kwargs)
     if args.do_fp8_training:
         convert_to_float8_training(
             transformer, module_filter_fn=module_filter_fn, config=Float8LinearConfig(pad_inner_dim=True)
@@ -1275,42 +1263,17 @@ def main(args):
 
     # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
     def save_model_hook(models, weights, output_dir):
-        transformer_cls = type(unwrap_model(transformer))
-
-        # 1) Validate and pick the transformer model
-        modules_to_save: dict[str, Any] = {}
-        transformer_model = None
-
-        for model in models:
-            if isinstance(unwrap_model(model), transformer_cls):
-                transformer_model = model
-                modules_to_save["transformer"] = model
-            else:
-                raise ValueError(f"unexpected save model: {model.__class__}")
-
-        if transformer_model is None:
-            raise ValueError("No transformer model found in 'models'")
-
-        # 2) Optionally gather FSDP state dict once
-        state_dict = accelerator.get_state_dict(model) if is_fsdp else None
-
-        # 3) Only main process materializes the LoRA state dict
-        transformer_lora_layers_to_save = None
         if accelerator.is_main_process:
-            peft_kwargs = {}
-            if is_fsdp:
-                peft_kwargs["state_dict"] = state_dict
+            transformer_lora_layers_to_save = None
+            modules_to_save = {}
+            for model in models:
+                if isinstance(model, type(unwrap_model(transformer))):
+                    transformer_lora_layers_to_save = get_peft_model_state_dict(model)
+                    modules_to_save["transformer"] = model
+                else:
+                    raise ValueError(f"unexpected save model: {model.__class__}")
 
-            transformer_lora_layers_to_save = get_peft_model_state_dict(
-                unwrap_model(transformer_model) if is_fsdp else transformer_model,
-                **peft_kwargs,
-            )
-
-            if is_fsdp:
-                transformer_lora_layers_to_save = _to_cpu_contiguous(transformer_lora_layers_to_save)
-
-            # make sure to pop weight so that corresponding model is not saved again
-            if weights:
+                # make sure to pop weight so that corresponding model is not saved again
                 weights.pop()
 
             Flux2Pipeline.save_lora_weights(
@@ -1322,20 +1285,13 @@ def main(args):
     def load_model_hook(models, input_dir):
         transformer_ = None
 
-        if not is_fsdp:
-            while len(models) > 0:
-                model = models.pop()
+        while len(models) > 0:
+            model = models.pop()
 
-                if isinstance(unwrap_model(model), type(unwrap_model(transformer))):
-                    transformer_ = unwrap_model(model)
-                else:
-                    raise ValueError(f"unexpected save model: {model.__class__}")
-        else:
-            transformer_ = Flux2Transformer2DModel.from_pretrained(
-                args.pretrained_model_name_or_path,
-                subfolder="transformer",
-            )
-            transformer_.add_adapter(transformer_lora_config)
+            if isinstance(model, type(unwrap_model(transformer))):
+                transformer_ = model
+            else:
+                raise ValueError(f"unexpected save model: {model.__class__}")
 
         lora_state_dict = Flux2Pipeline.lora_state_dict(input_dir)
 
@@ -1551,21 +1507,6 @@ def main(args):
                     args.validation_prompt, text_encoding_pipeline
                 )
 
-    # Init FSDP for text encoder
-    if args.fsdp_text_encoder:
-        fsdp_kwargs = get_fsdp_kwargs_from_accelerator(accelerator)
-        text_encoder_fsdp = wrap_with_fsdp(
-            model=text_encoding_pipeline.text_encoder,
-            device=accelerator.device,
-            offload=args.offload,
-            limit_all_gathers=True,
-            use_orig_params=True,
-            fsdp_kwargs=fsdp_kwargs,
-        )
-
-        text_encoding_pipeline.text_encoder = text_encoder_fsdp
-        dist.barrier()
-
     # If custom instance prompts are NOT provided (i.e. the instance prompt is used for all images),
     # pack the statically computed variables appropriately here. This is so that we don't
     # have to pass them to the dataloader.
@@ -1595,8 +1536,6 @@ def main(args):
                 if train_dataset.custom_instance_prompts:
                     if args.remote_text_encoder:
                         prompt_embeds, text_ids = compute_remote_text_embeddings(batch["prompts"])
-                    elif args.fsdp_text_encoder:
-                        prompt_embeds, text_ids = compute_text_embeddings(batch["prompts"], text_encoding_pipeline)
                     else:
                         with offload_models(text_encoding_pipeline, device=accelerator.device, offload=args.offload):
                             prompt_embeds, text_ids = compute_text_embeddings(batch["prompts"], text_encoding_pipeline)
@@ -1838,7 +1777,7 @@ def main(args):
                 progress_bar.update(1)
                 global_step += 1
 
-                if accelerator.is_main_process or is_fsdp:
+                if accelerator.is_main_process:
                     if global_step % args.checkpointing_steps == 0:
                         # _before_ saving state, check if this save would set us over the `checkpoints_total_limit`
                         if args.checkpoints_total_limit is not None:
@@ -1897,41 +1836,15 @@ def main(args):
 
     # Save the lora layers
     accelerator.wait_for_everyone()
-
-    if is_fsdp:
-        transformer = unwrap_model(transformer)
-        state_dict = accelerator.get_state_dict(transformer)
     if accelerator.is_main_process:
         modules_to_save = {}
-        if is_fsdp:
-            if args.bnb_quantization_config_path is None:
-                if args.upcast_before_saving:
-                    state_dict = {
-                        k: v.to(torch.float32) if isinstance(v, torch.Tensor) else v for k, v in state_dict.items()
-                    }
-                else:
-                    state_dict = {
-                        k: v.to(weight_dtype) if isinstance(v, torch.Tensor) else v for k, v in state_dict.items()
-                    }
-
-            transformer_lora_layers = get_peft_model_state_dict(
-                transformer,
-                state_dict=state_dict,
-            )
-            transformer_lora_layers = {
-                k: v.detach().cpu().contiguous() if isinstance(v, torch.Tensor) else v
-                for k, v in transformer_lora_layers.items()
-            }
-
-        else:
-            transformer = unwrap_model(transformer)
-            if args.bnb_quantization_config_path is None:
-                if args.upcast_before_saving:
-                    transformer.to(torch.float32)
-                else:
-                    transformer = transformer.to(weight_dtype)
-            transformer_lora_layers = get_peft_model_state_dict(transformer)
-
+        transformer = unwrap_model(transformer)
+        if args.bnb_quantization_config_path is None:
+            if args.upcast_before_saving:
+                transformer.to(torch.float32)
+            else:
+                transformer = transformer.to(weight_dtype)
+        transformer_lora_layers = get_peft_model_state_dict(transformer)
         modules_to_save["transformer"] = transformer
 
         Flux2Pipeline.save_lora_weights(

examples/dreambooth/train_dreambooth_lora_flux2_img2img.py

@@ -43,7 +43,6 @@ import random
 import shutil
 from contextlib import nullcontext
 from pathlib import Path
-from typing import Any
 
 import numpy as np
 import torch
@@ -75,16 +74,13 @@ from diffusers.optimization import get_scheduler
 from diffusers.pipelines.flux2.image_processor import Flux2ImageProcessor
 from diffusers.training_utils import (
     _collate_lora_metadata,
-    _to_cpu_contiguous,
     cast_training_params,
     compute_density_for_timestep_sampling,
     compute_loss_weighting_for_sd3,
     find_nearest_bucket,
     free_memory,
-    get_fsdp_kwargs_from_accelerator,
     offload_models,
     parse_buckets_string,
-    wrap_with_fsdp,
 )
 from diffusers.utils import (
     check_min_version,
@@ -97,9 +93,6 @@ from diffusers.utils.import_utils import is_torch_npu_available
 from diffusers.utils.torch_utils import is_compiled_module
 
 
-if getattr(torch, "distributed", None) is not None:
-    import torch.distributed as dist
-
 if is_wandb_available():
     import wandb
 
@@ -698,7 +691,6 @@ def parse_args(input_args=None):
 
     parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
     parser.add_argument("--enable_npu_flash_attention", action="store_true", help="Enabla Flash Attention for NPU")
-    parser.add_argument("--fsdp_text_encoder", action="store_true", help="Use FSDP for text encoder")
 
     if input_args is not None:
         args = parser.parse_args(input_args)
@@ -1164,11 +1156,7 @@ def main(args):
         if args.bnb_quantization_config_path is not None
         else {"device": accelerator.device, "dtype": weight_dtype}
     )
-
-    is_fsdp = accelerator.state.fsdp_plugin is not None
-    if not is_fsdp:
-        transformer.to(**transformer_to_kwargs)
-
+    transformer.to(**transformer_to_kwargs)
     if args.do_fp8_training:
         convert_to_float8_training(
             transformer, module_filter_fn=module_filter_fn, config=Float8LinearConfig(pad_inner_dim=True)
@@ -1212,42 +1200,17 @@ def main(args):
 
     # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
     def save_model_hook(models, weights, output_dir):
-        transformer_cls = type(unwrap_model(transformer))
-
-        # 1) Validate and pick the transformer model
-        modules_to_save: dict[str, Any] = {}
-        transformer_model = None
-
-        for model in models:
-            if isinstance(unwrap_model(model), transformer_cls):
-                transformer_model = model
-                modules_to_save["transformer"] = model
-            else:
-                raise ValueError(f"unexpected save model: {model.__class__}")
-
-        if transformer_model is None:
-            raise ValueError("No transformer model found in 'models'")
-
-        # 2) Optionally gather FSDP state dict once
-        state_dict = accelerator.get_state_dict(model) if is_fsdp else None
-
-        # 3) Only main process materializes the LoRA state dict
-        transformer_lora_layers_to_save = None
         if accelerator.is_main_process:
-            peft_kwargs = {}
-            if is_fsdp:
-                peft_kwargs["state_dict"] = state_dict
+            transformer_lora_layers_to_save = None
+            modules_to_save = {}
+            for model in models:
+                if isinstance(model, type(unwrap_model(transformer))):
+                    transformer_lora_layers_to_save = get_peft_model_state_dict(model)
+                    modules_to_save["transformer"] = model
+                else:
+                    raise ValueError(f"unexpected save model: {model.__class__}")
 
-            transformer_lora_layers_to_save = get_peft_model_state_dict(
-                unwrap_model(transformer_model) if is_fsdp else transformer_model,
-                **peft_kwargs,
-            )
-
-            if is_fsdp:
-                transformer_lora_layers_to_save = _to_cpu_contiguous(transformer_lora_layers_to_save)
-
-            # make sure to pop weight so that corresponding model is not saved again
-            if weights:
+                # make sure to pop weight so that corresponding model is not saved again
                 weights.pop()
 
             Flux2Pipeline.save_lora_weights(
@@ -1259,20 +1222,13 @@ def main(args):
     def load_model_hook(models, input_dir):
         transformer_ = None
 
-        if not is_fsdp:
-            while len(models) > 0:
-                model = models.pop()
+        while len(models) > 0:
+            model = models.pop()
 
-                if isinstance(unwrap_model(model), type(unwrap_model(transformer))):
-                    transformer_ = unwrap_model(model)
-                else:
-                    raise ValueError(f"unexpected save model: {model.__class__}")
-        else:
-            transformer_ = Flux2Transformer2DModel.from_pretrained(
-                args.pretrained_model_name_or_path,
-                subfolder="transformer",
-            )
-            transformer_.add_adapter(transformer_lora_config)
+            if isinstance(model, type(unwrap_model(transformer))):
+                transformer_ = model
+            else:
+                raise ValueError(f"unexpected save model: {model.__class__}")
 
         lora_state_dict = Flux2Pipeline.lora_state_dict(input_dir)
 
@@ -1474,21 +1430,6 @@ def main(args):
                     args.validation_prompt, text_encoding_pipeline
                 )
 
-    # Init FSDP for text encoder
-    if args.fsdp_text_encoder:
-        fsdp_kwargs = get_fsdp_kwargs_from_accelerator(accelerator)
-        text_encoder_fsdp = wrap_with_fsdp(
-            model=text_encoding_pipeline.text_encoder,
-            device=accelerator.device,
-            offload=args.offload,
-            limit_all_gathers=True,
-            use_orig_params=True,
-            fsdp_kwargs=fsdp_kwargs,
-        )
-
-        text_encoding_pipeline.text_encoder = text_encoder_fsdp
-        dist.barrier()
-
     # If custom instance prompts are NOT provided (i.e. the instance prompt is used for all images),
     # pack the statically computed variables appropriately here. This is so that we don't
     # have to pass them to the dataloader.
@@ -1520,8 +1461,6 @@ def main(args):
                 if train_dataset.custom_instance_prompts:
                     if args.remote_text_encoder:
                         prompt_embeds, text_ids = compute_remote_text_embeddings(batch["prompts"])
-                    elif args.fsdp_text_encoder:
-                        prompt_embeds, text_ids = compute_text_embeddings(batch["prompts"], text_encoding_pipeline)
                     else:
                         with offload_models(text_encoding_pipeline, device=accelerator.device, offload=args.offload):
                             prompt_embeds, text_ids = compute_text_embeddings(batch["prompts"], text_encoding_pipeline)
@@ -1761,7 +1700,7 @@ def main(args):
                 progress_bar.update(1)
                 global_step += 1
 
-                if accelerator.is_main_process or is_fsdp:
+                if accelerator.is_main_process:
                     if global_step % args.checkpointing_steps == 0:
                         # _before_ saving state, check if this save would set us over the `checkpoints_total_limit`
                         if args.checkpoints_total_limit is not None:
@@ -1820,41 +1759,15 @@ def main(args):
 
     # Save the lora layers
     accelerator.wait_for_everyone()
-
-    if is_fsdp:
-        transformer = unwrap_model(transformer)
-        state_dict = accelerator.get_state_dict(transformer)
     if accelerator.is_main_process:
         modules_to_save = {}
-        if is_fsdp:
-            if args.bnb_quantization_config_path is None:
-                if args.upcast_before_saving:
-                    state_dict = {
-                        k: v.to(torch.float32) if isinstance(v, torch.Tensor) else v for k, v in state_dict.items()
-                    }
-                else:
-                    state_dict = {
-                        k: v.to(weight_dtype) if isinstance(v, torch.Tensor) else v for k, v in state_dict.items()
-                    }
-
-            transformer_lora_layers = get_peft_model_state_dict(
-                transformer,
-                state_dict=state_dict,
-            )
-            transformer_lora_layers = {
-                k: v.detach().cpu().contiguous() if isinstance(v, torch.Tensor) else v
-                for k, v in transformer_lora_layers.items()
-            }
-
-        else:
-            transformer = unwrap_model(transformer)
-            if args.bnb_quantization_config_path is None:
-                if args.upcast_before_saving:
-                    transformer.to(torch.float32)
-                else:
-                    transformer = transformer.to(weight_dtype)
-            transformer_lora_layers = get_peft_model_state_dict(transformer)
-
+        transformer = unwrap_model(transformer)
+        if args.bnb_quantization_config_path is None:
+            if args.upcast_before_saving:
+                transformer.to(torch.float32)
+            else:
+                transformer = transformer.to(weight_dtype)
+        transformer_lora_layers = get_peft_model_state_dict(transformer)
         modules_to_save["transformer"] = transformer
 
         Flux2Pipeline.save_lora_weights(

src/diffusers/__init__.py

@@ -353,7 +353,6 @@ else:
             "KDPM2AncestralDiscreteScheduler",
             "KDPM2DiscreteScheduler",
             "LCMScheduler",
-            "LTXEulerAncestralRFScheduler",
             "PNDMScheduler",
             "RePaintScheduler",
             "SASolverScheduler",
@@ -539,7 +538,6 @@ else:
             "LongCatImageEditPipeline",
             "LongCatImagePipeline",
             "LTXConditionPipeline",
-            "LTXI2VLongMultiPromptPipeline",
             "LTXImageToVideoPipeline",
             "LTXLatentUpsamplePipeline",
             "LTXPipeline",
@@ -1090,7 +1088,6 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             KDPM2AncestralDiscreteScheduler,
             KDPM2DiscreteScheduler,
             LCMScheduler,
-            LTXEulerAncestralRFScheduler,
             PNDMScheduler,
             RePaintScheduler,
             SASolverScheduler,
@@ -1255,7 +1252,6 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             LongCatImageEditPipeline,
             LongCatImagePipeline,
             LTXConditionPipeline,
-            LTXI2VLongMultiPromptPipeline,
             LTXImageToVideoPipeline,
             LTXLatentUpsamplePipeline,
             LTXPipeline,

src/diffusers/models/attention_dispatch.py

@@ -1420,7 +1420,6 @@ def _flash_attention(
     query: torch.Tensor,
     key: torch.Tensor,
     value: torch.Tensor,
-    attn_mask: Optional[torch.Tensor] = None,
     dropout_p: float = 0.0,
     is_causal: bool = False,
     scale: Optional[float] = None,
@@ -1428,9 +1427,6 @@ def _flash_attention(
     _parallel_config: Optional["ParallelConfig"] = None,
 ) -> torch.Tensor:
     lse = None
-    if attn_mask is not None:
-        raise ValueError("`attn_mask` is not supported for flash-attn 2.")
-
     if _parallel_config is None:
         out = flash_attn_func(
             q=query,
@@ -1473,7 +1469,6 @@ def _flash_attention_hub(
     query: torch.Tensor,
     key: torch.Tensor,
     value: torch.Tensor,
-    attn_mask: Optional[torch.Tensor] = None,
     dropout_p: float = 0.0,
     is_causal: bool = False,
     scale: Optional[float] = None,
@@ -1481,9 +1476,6 @@ def _flash_attention_hub(
     _parallel_config: Optional["ParallelConfig"] = None,
 ) -> torch.Tensor:
     lse = None
-    if attn_mask is not None:
-        raise ValueError("`attn_mask` is not supported for flash-attn 2.")
-
     func = _HUB_KERNELS_REGISTRY[AttentionBackendName.FLASH_HUB].kernel_fn
     out = func(
         q=query,
@@ -1620,15 +1612,11 @@ def _flash_attention_3(
     query: torch.Tensor,
     key: torch.Tensor,
     value: torch.Tensor,
-    attn_mask: Optional[torch.Tensor] = None,
     scale: Optional[float] = None,
     is_causal: bool = False,
     return_lse: bool = False,
     _parallel_config: Optional["ParallelConfig"] = None,
 ) -> torch.Tensor:
-    if attn_mask is not None:
-        raise ValueError("`attn_mask` is not supported for flash-attn 3.")
-
     out, lse = _wrapped_flash_attn_3(
         q=query,
         k=key,
@@ -1648,7 +1636,6 @@ def _flash_attention_3_hub(
     query: torch.Tensor,
     key: torch.Tensor,
     value: torch.Tensor,
-    attn_mask: Optional[torch.Tensor] = None,
     scale: Optional[float] = None,
     is_causal: bool = False,
     window_size: Tuple[int, int] = (-1, -1),
@@ -1659,8 +1646,6 @@ def _flash_attention_3_hub(
 ) -> torch.Tensor:
     if _parallel_config:
         raise NotImplementedError(f"{AttentionBackendName._FLASH_3_HUB.value} is not implemented for parallelism yet.")
-    if attn_mask is not None:
-        raise ValueError("`attn_mask` is not supported for flash-attn 3.")
 
     func = _HUB_KERNELS_REGISTRY[AttentionBackendName._FLASH_3_HUB].kernel_fn
     out = func(
@@ -1800,16 +1785,12 @@ def _aiter_flash_attention(
     query: torch.Tensor,
     key: torch.Tensor,
     value: torch.Tensor,
-    attn_mask: Optional[torch.Tensor] = None,
     dropout_p: float = 0.0,
     is_causal: bool = False,
     scale: Optional[float] = None,
     return_lse: bool = False,
     _parallel_config: Optional["ParallelConfig"] = None,
 ) -> torch.Tensor:
-    if attn_mask is not None:
-        raise ValueError("`attn_mask` is not supported for aiter attention")
-
     if not return_lse and torch.is_grad_enabled():
         # aiter requires return_lse=True by assertion when gradients are enabled.
         out, lse, *_ = aiter_flash_attn_func(
@@ -2047,7 +2028,6 @@ def _native_flash_attention(
     query: torch.Tensor,
     key: torch.Tensor,
     value: torch.Tensor,
-    attn_mask: Optional[torch.Tensor] = None,
     dropout_p: float = 0.0,
     is_causal: bool = False,
     scale: Optional[float] = None,
@@ -2055,9 +2035,6 @@ def _native_flash_attention(
     return_lse: bool = False,
     _parallel_config: Optional["ParallelConfig"] = None,
 ) -> torch.Tensor:
-    if attn_mask is not None:
-        raise ValueError("`attn_mask` is not supported for aiter attention")
-
     lse = None
     if _parallel_config is None and not return_lse:
         query, key, value = (x.permute(0, 2, 1, 3) for x in (query, key, value))
@@ -2136,14 +2113,11 @@ def _native_npu_attention(
     query: torch.Tensor,
     key: torch.Tensor,
     value: torch.Tensor,
-    attn_mask: Optional[torch.Tensor] = None,
     dropout_p: float = 0.0,
     scale: Optional[float] = None,
     return_lse: bool = False,
     _parallel_config: Optional["ParallelConfig"] = None,
 ) -> torch.Tensor:
-    if attn_mask is not None:
-        raise ValueError("`attn_mask` is not supported for NPU attention")
     if return_lse:
         raise ValueError("NPU attention backend does not support setting `return_lse=True`.")
     query, key, value = (x.transpose(1, 2).contiguous() for x in (query, key, value))
@@ -2174,13 +2148,10 @@ def _native_xla_attention(
     query: torch.Tensor,
     key: torch.Tensor,
     value: torch.Tensor,
-    attn_mask: Optional[torch.Tensor] = None,
     is_causal: bool = False,
     return_lse: bool = False,
     _parallel_config: Optional["ParallelConfig"] = None,
 ) -> torch.Tensor:
-    if attn_mask is not None:
-        raise ValueError("`attn_mask` is not supported for XLA attention")
     if return_lse:
         raise ValueError("XLA attention backend does not support setting `return_lse=True`.")
     query, key, value = (x.permute(0, 2, 1, 3) for x in (query, key, value))
@@ -2204,14 +2175,11 @@ def _sage_attention(
     query: torch.Tensor,
     key: torch.Tensor,
     value: torch.Tensor,
-    attn_mask: Optional[torch.Tensor] = None,
     is_causal: bool = False,
     scale: Optional[float] = None,
     return_lse: bool = False,
     _parallel_config: Optional["ParallelConfig"] = None,
 ) -> torch.Tensor:
-    if attn_mask is not None:
-        raise ValueError("`attn_mask` is not supported for sage attention")
     lse = None
     if _parallel_config is None:
         out = sageattn(
@@ -2255,14 +2223,11 @@ def _sage_attention_hub(
     query: torch.Tensor,
     key: torch.Tensor,
     value: torch.Tensor,
-    attn_mask: Optional[torch.Tensor] = None,
     is_causal: bool = False,
     scale: Optional[float] = None,
     return_lse: bool = False,
     _parallel_config: Optional["ParallelConfig"] = None,
 ) -> torch.Tensor:
-    if attn_mask is not None:
-        raise ValueError("`attn_mask` is not supported for sage attention")
     lse = None
     func = _HUB_KERNELS_REGISTRY[AttentionBackendName.SAGE_HUB].kernel_fn
     if _parallel_config is None:
@@ -2344,14 +2309,11 @@ def _sage_qk_int8_pv_fp8_cuda_attention(
     query: torch.Tensor,
     key: torch.Tensor,
     value: torch.Tensor,
-    attn_mask: Optional[torch.Tensor] = None,
     is_causal: bool = False,
     scale: Optional[float] = None,
     return_lse: bool = False,
     _parallel_config: Optional["ParallelConfig"] = None,
 ) -> torch.Tensor:
-    if attn_mask is not None:
-        raise ValueError("`attn_mask` is not supported for sage attention")
     return sageattn_qk_int8_pv_fp8_cuda(
         q=query,
         k=key,
@@ -2371,14 +2333,11 @@ def _sage_qk_int8_pv_fp8_cuda_sm90_attention(
     query: torch.Tensor,
     key: torch.Tensor,
     value: torch.Tensor,
-    attn_mask: Optional[torch.Tensor] = None,
     is_causal: bool = False,
     scale: Optional[float] = None,
     return_lse: bool = False,
     _parallel_config: Optional["ParallelConfig"] = None,
 ) -> torch.Tensor:
-    if attn_mask is not None:
-        raise ValueError("`attn_mask` is not supported for sage attention")
     return sageattn_qk_int8_pv_fp8_cuda_sm90(
         q=query,
         k=key,
@@ -2398,14 +2357,11 @@ def _sage_qk_int8_pv_fp16_cuda_attention(
     query: torch.Tensor,
     key: torch.Tensor,
     value: torch.Tensor,
-    attn_mask: Optional[torch.Tensor] = None,
     is_causal: bool = False,
     scale: Optional[float] = None,
     return_lse: bool = False,
     _parallel_config: Optional["ParallelConfig"] = None,
 ) -> torch.Tensor:
-    if attn_mask is not None:
-        raise ValueError("`attn_mask` is not supported for sage attention")
     return sageattn_qk_int8_pv_fp16_cuda(
         q=query,
         k=key,
@@ -2425,14 +2381,11 @@ def _sage_qk_int8_pv_fp16_triton_attention(
     query: torch.Tensor,
     key: torch.Tensor,
     value: torch.Tensor,
-    attn_mask: Optional[torch.Tensor] = None,
     is_causal: bool = False,
     scale: Optional[float] = None,
     return_lse: bool = False,
     _parallel_config: Optional["ParallelConfig"] = None,
 ) -> torch.Tensor:
-    if attn_mask is not None:
-        raise ValueError("`attn_mask` is not supported for sage attention")
     return sageattn_qk_int8_pv_fp16_triton(
         q=query,
         k=key,

src/diffusers/models/transformers/transformer_flux.py

@@ -22,7 +22,7 @@ import torch.nn.functional as F
 
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...loaders import FluxTransformer2DLoadersMixin, FromOriginalModelMixin, PeftAdapterMixin
-from ...utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
+from ...utils import USE_PEFT_BACKEND, is_torch_npu_available, logging, scale_lora_layers, unscale_lora_layers
 from ...utils.torch_utils import maybe_allow_in_graph
 from .._modeling_parallel import ContextParallelInput, ContextParallelOutput
 from ..attention import AttentionMixin, AttentionModuleMixin, FeedForward
@@ -717,7 +717,11 @@ class FluxTransformer2DModel(
             img_ids = img_ids[0]
 
         ids = torch.cat((txt_ids, img_ids), dim=0)
-        image_rotary_emb = self.pos_embed(ids)
+        if is_torch_npu_available():
+            freqs_cos, freqs_sin = self.pos_embed(ids.cpu())
+            image_rotary_emb = (freqs_cos.npu(), freqs_sin.npu())
+        else:
+            image_rotary_emb = self.pos_embed(ids)
 
         if joint_attention_kwargs is not None and "ip_adapter_image_embeds" in joint_attention_kwargs:
             ip_adapter_image_embeds = joint_attention_kwargs.pop("ip_adapter_image_embeds")

src/diffusers/models/transformers/transformer_flux2.py

@@ -21,7 +21,7 @@ import torch.nn.functional as F
 
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...loaders import FluxTransformer2DLoadersMixin, FromOriginalModelMixin, PeftAdapterMixin
-from ...utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
+from ...utils import USE_PEFT_BACKEND, is_torch_npu_available, logging, scale_lora_layers, unscale_lora_layers
 from .._modeling_parallel import ContextParallelInput, ContextParallelOutput
 from ..attention import AttentionMixin, AttentionModuleMixin
 from ..attention_dispatch import dispatch_attention_fn
@@ -835,8 +835,14 @@ class Flux2Transformer2DModel(
         if txt_ids.ndim == 3:
             txt_ids = txt_ids[0]
 
-        image_rotary_emb = self.pos_embed(img_ids)
-        text_rotary_emb = self.pos_embed(txt_ids)
+        if is_torch_npu_available():
+            freqs_cos_image, freqs_sin_image = self.pos_embed(img_ids.cpu())
+            image_rotary_emb = (freqs_cos_image.npu(), freqs_sin_image.npu())
+            freqs_cos_text, freqs_sin_text = self.pos_embed(txt_ids.cpu())
+            text_rotary_emb = (freqs_cos_text.npu(), freqs_sin_text.npu())
+        else:
+            image_rotary_emb = self.pos_embed(img_ids)
+            text_rotary_emb = self.pos_embed(txt_ids)
         concat_rotary_emb = (
             torch.cat([text_rotary_emb[0], image_rotary_emb[0]], dim=0),
             torch.cat([text_rotary_emb[1], image_rotary_emb[1]], dim=0),

src/diffusers/models/transformers/transformer_longcat_image.py

@@ -21,7 +21,7 @@ import torch.nn.functional as F
 
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...loaders import FromOriginalModelMixin, PeftAdapterMixin
-from ...utils import logging
+from ...utils import is_torch_npu_available, logging
 from ...utils.torch_utils import maybe_allow_in_graph
 from ..attention import AttentionModuleMixin, FeedForward
 from ..attention_dispatch import dispatch_attention_fn
@@ -499,7 +499,11 @@ class LongCatImageTransformer2DModel(
         encoder_hidden_states = self.context_embedder(encoder_hidden_states)
 
         ids = torch.cat((txt_ids, img_ids), dim=0)
-        image_rotary_emb = self.pos_embed(ids)
+        if is_torch_npu_available():
+            freqs_cos, freqs_sin = self.pos_embed(ids.cpu())
+            image_rotary_emb = (freqs_cos.npu(), freqs_sin.npu())
+        else:
+            image_rotary_emb = self.pos_embed(ids)
 
         for index_block, block in enumerate(self.transformer_blocks):
             if torch.is_grad_enabled() and self.gradient_checkpointing and self.use_checkpoint[index_block]:

src/diffusers/models/transformers/transformer_ovis_image.py

@@ -21,7 +21,7 @@ import torch.nn.functional as F
 
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...loaders import FromOriginalModelMixin, PeftAdapterMixin
-from ...utils import logging
+from ...utils import is_torch_npu_available, logging
 from ...utils.torch_utils import maybe_allow_in_graph
 from ..attention import AttentionModuleMixin, FeedForward
 from ..attention_dispatch import dispatch_attention_fn
@@ -530,7 +530,11 @@ class OvisImageTransformer2DModel(
             img_ids = img_ids[0]
 
         ids = torch.cat((txt_ids, img_ids), dim=0)
-        image_rotary_emb = self.pos_embed(ids)
+        if is_torch_npu_available():
+            freqs_cos, freqs_sin = self.pos_embed(ids.cpu())
+            image_rotary_emb = (freqs_cos.npu(), freqs_sin.npu())
+        else:
+            image_rotary_emb = self.pos_embed(ids)
 
         for index_block, block in enumerate(self.transformer_blocks):
             if torch.is_grad_enabled() and self.gradient_checkpointing:

src/diffusers/pipelines/__init__.py

@@ -288,7 +288,6 @@ else:
         "LTXImageToVideoPipeline",
         "LTXConditionPipeline",
         "LTXLatentUpsamplePipeline",
-        "LTXI2VLongMultiPromptPipeline",
     ]
     _import_structure["lumina"] = ["LuminaPipeline", "LuminaText2ImgPipeline"]
     _import_structure["lumina2"] = ["Lumina2Pipeline", "Lumina2Text2ImgPipeline"]
@@ -730,13 +729,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             LEditsPPPipelineStableDiffusionXL,
         )
         from .longcat_image import LongCatImageEditPipeline, LongCatImagePipeline
-        from .ltx import (
-            LTXConditionPipeline,
-            LTXI2VLongMultiPromptPipeline,
-            LTXImageToVideoPipeline,
-            LTXLatentUpsamplePipeline,
-            LTXPipeline,
-        )
+        from .ltx import LTXConditionPipeline, LTXImageToVideoPipeline, LTXLatentUpsamplePipeline, LTXPipeline
         from .lucy import LucyEditPipeline
         from .lumina import LuminaPipeline, LuminaText2ImgPipeline
         from .lumina2 import Lumina2Pipeline, Lumina2Text2ImgPipeline

src/diffusers/pipelines/ltx/__init__.py

@@ -25,7 +25,6 @@ else:
     _import_structure["modeling_latent_upsampler"] = ["LTXLatentUpsamplerModel"]
     _import_structure["pipeline_ltx"] = ["LTXPipeline"]
     _import_structure["pipeline_ltx_condition"] = ["LTXConditionPipeline"]
-    _import_structure["pipeline_ltx_i2v_long_multi_prompt"] = ["LTXI2VLongMultiPromptPipeline"]
     _import_structure["pipeline_ltx_image2video"] = ["LTXImageToVideoPipeline"]
     _import_structure["pipeline_ltx_latent_upsample"] = ["LTXLatentUpsamplePipeline"]
 
@@ -40,7 +39,6 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .modeling_latent_upsampler import LTXLatentUpsamplerModel
         from .pipeline_ltx import LTXPipeline
         from .pipeline_ltx_condition import LTXConditionPipeline
-        from .pipeline_ltx_i2v_long_multi_prompt import LTXI2VLongMultiPromptPipeline
         from .pipeline_ltx_image2video import LTXImageToVideoPipeline
         from .pipeline_ltx_latent_upsample import LTXLatentUpsamplePipeline
 

src/diffusers/schedulers/__init__.py

@@ -66,7 +66,6 @@ else:
     _import_structure["scheduling_k_dpm_2_ancestral_discrete"] = ["KDPM2AncestralDiscreteScheduler"]
     _import_structure["scheduling_k_dpm_2_discrete"] = ["KDPM2DiscreteScheduler"]
     _import_structure["scheduling_lcm"] = ["LCMScheduler"]
-    _import_structure["scheduling_ltx_euler_ancestral_rf"] = ["LTXEulerAncestralRFScheduler"]
     _import_structure["scheduling_pndm"] = ["PNDMScheduler"]
     _import_structure["scheduling_repaint"] = ["RePaintScheduler"]
     _import_structure["scheduling_sasolver"] = ["SASolverScheduler"]
@@ -169,7 +168,6 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .scheduling_k_dpm_2_ancestral_discrete import KDPM2AncestralDiscreteScheduler
         from .scheduling_k_dpm_2_discrete import KDPM2DiscreteScheduler
         from .scheduling_lcm import LCMScheduler
-        from .scheduling_ltx_euler_ancestral_rf import LTXEulerAncestralRFScheduler
         from .scheduling_pndm import PNDMScheduler
         from .scheduling_repaint import RePaintScheduler
         from .scheduling_sasolver import SASolverScheduler

src/diffusers/schedulers/scheduling_ltx_euler_ancestral_rf.py

@@ -1,386 +0,0 @@
-# Copyright 2025 Lightricks and The HuggingFace Team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-"""
-LTXEulerAncestralRFScheduler
-
-This scheduler implements a K-diffusion style Euler-Ancestral sampler specialized for flow / CONST parameterization,
-closely mirroring ComfyUI's `sample_euler_ancestral_RF` implementation used for LTX-Video.
-
-Reference implementation (ComfyUI):
-    comfy.k_diffusion.sampling.sample_euler_ancestral_RF
-"""
-
-from dataclasses import dataclass
-from typing import List, Optional, Tuple, Union
-
-import torch
-
-from ..configuration_utils import ConfigMixin, register_to_config
-from ..utils import BaseOutput, logging
-from ..utils.torch_utils import randn_tensor
-from .scheduling_utils import SchedulerMixin
-
-
-logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
-
-
-@dataclass
-class LTXEulerAncestralRFSchedulerOutput(BaseOutput):
-    """
-    Output class for the scheduler's `step` function output.
-
-    Args:
-        prev_sample (`torch.FloatTensor`):
-            Updated sample for the next step in the denoising process.
-    """
-
-    prev_sample: torch.FloatTensor
-
-
-class LTXEulerAncestralRFScheduler(SchedulerMixin, ConfigMixin):
-    """
-    Euler-Ancestral scheduler for LTX-Video (RF / CONST parametrization).
-
-    This scheduler is intended for models where the network is trained with a CONST-like parameterization (as in LTXV /
-    FLUX). It approximates ComfyUI's `sample_euler_ancestral_RF` sampler and is useful when reproducing ComfyUI
-    workflows inside diffusers.
-
-    The scheduler can either:
-    - reuse the [`FlowMatchEulerDiscreteScheduler`] sigma / timestep logic when only `num_inference_steps` is provided
-      (default diffusers-style usage), or
-    - follow an explicit ComfyUI-style sigma schedule when `sigmas` (or `timesteps`) are passed to [`set_timesteps`].
-
-    Args:
-        num_train_timesteps (`int`, defaults to 1000):
-            Included for config compatibility; not used to build the schedule.
-        eta (`float`, defaults to 1.0):
-            Stochasticity parameter. `eta=0.0` yields deterministic DDIM-like sampling; `eta=1.0` matches ComfyUI's
-            default RF behavior.
-        s_noise (`float`, defaults to 1.0):
-            Global scaling factor for the stochastic noise term.
-    """
-
-    # Allow config migration from the flow-match scheduler and back.
-    _compatibles = ["FlowMatchEulerDiscreteScheduler"]
-    order = 1
-
-    @register_to_config
-    def __init__(
-        self,
-        num_train_timesteps: int = 1000,
-        eta: float = 1.0,
-        s_noise: float = 1.0,
-    ):
-        # Note: num_train_timesteps is kept only for config compatibility.
-        self.num_inference_steps: Optional[int] = None
-        self.sigmas: Optional[torch.Tensor] = None
-        self.timesteps: Optional[torch.Tensor] = None
-        self._step_index: Optional[int] = None
-        self._begin_index: Optional[int] = None
-
-    @property
-    def step_index(self) -> Optional[int]:
-        return self._step_index
-
-    @property
-    def begin_index(self) -> Optional[int]:
-        """
-        The index for the first timestep. It can be set from a pipeline with `set_begin_index` to support
-        image-to-image like workflows that start denoising part-way through the schedule.
-        """
-        return self._begin_index
-
-    def set_begin_index(self, begin_index: int = 0):
-        """
-        Included for API compatibility; not strictly needed here but kept to allow pipelines that call
-        `set_begin_index`.
-        """
-        self._begin_index = begin_index
-
-    def index_for_timestep(
-        self, timestep: Union[float, torch.Tensor], schedule_timesteps: Optional[torch.Tensor] = None
-    ) -> int:
-        """
-        Map a (continuous) `timestep` value to an index into `self.timesteps`.
-
-        This follows the convention used in other discrete schedulers: if the same timestep value appears multiple
-        times in the schedule (which can happen when starting in the middle of the schedule), the *second* occurrence
-        is used for the first `step` call so that no sigma is accidentally skipped.
-        """
-        if schedule_timesteps is None:
-            if self.timesteps is None:
-                raise ValueError("Timesteps have not been set. Call `set_timesteps` first.")
-            schedule_timesteps = self.timesteps
-
-        if isinstance(timestep, torch.Tensor):
-            timestep = timestep.to(schedule_timesteps.device)
-
-        indices = (schedule_timesteps == timestep).nonzero()
-
-        # The sigma index that is taken for the **very** first `step`
-        # is always the second index (or the last index if there is only 1)
-        # This way we can ensure we don't accidentally skip a sigma in
-        # case we start in the middle of the denoising schedule (e.g. for image-to-image)
-        pos = 1 if len(indices) > 1 else 0
-
-        if len(indices) == 0:
-            raise ValueError(
-                "Passed `timestep` is not in `self.timesteps`. Make sure to use values from `scheduler.timesteps`."
-            )
-
-        return indices[pos].item()
-
-    def _init_step_index(self, timestep: Union[float, torch.Tensor]):
-        """
-        Initialize the internal step index based on a given timestep.
-        """
-        if self.timesteps is None:
-            raise ValueError("Timesteps have not been set. Call `set_timesteps` first.")
-
-        if self.begin_index is None:
-            if isinstance(timestep, torch.Tensor):
-                timestep = timestep.to(self.timesteps.device)
-            self._step_index = self.index_for_timestep(timestep)
-        else:
-            self._step_index = self._begin_index
-
-    def set_timesteps(
-        self,
-        num_inference_steps: Optional[int] = None,
-        device: Union[str, torch.device, None] = None,
-        sigmas: Optional[Union[List[float], torch.Tensor]] = None,
-        timesteps: Optional[Union[List[float], torch.Tensor]] = None,
-        mu: Optional[float] = None,
-        **kwargs,
-    ):
-        """
-        Set the sigma / timestep schedule for sampling.
-
-        When `sigmas` or `timesteps` are provided explicitly, they are used as the RF sigma schedule (ComfyUI-style)
-        and are expected to include the terminal 0.0. When both are `None`, the scheduler reuses the
-        [`FlowMatchEulerDiscreteScheduler`] logic to generate sigmas from `num_inference_steps` and the stored config
-        (including any resolution-dependent shifting, Karras/beta schedules, etc.).
-
-        Args:
-            num_inference_steps (`int`, *optional*):
-                Number of denoising steps. If provided together with explicit `sigmas`/`timesteps`, they are expected
-                to be consistent and are otherwise ignored with a warning.
-            device (`str` or `torch.device`, *optional*):
-                Device to move the internal tensors to.
-            sigmas (`List[float]` or `torch.Tensor`, *optional*):
-                Explicit sigma schedule, e.g. `[1.0, 0.99, ..., 0.0]`.
-            timesteps (`List[float]` or `torch.Tensor`, *optional*):
-                Optional alias for `sigmas`. If `sigmas` is None and `timesteps` is provided, timesteps are treated as
-                sigmas.
-            mu (`float`, *optional*):
-                Optional shift parameter used when delegating to [`FlowMatchEulerDiscreteScheduler.set_timesteps`] and
-                `config.use_dynamic_shifting` is `True`.
-        """
-        # 1. Auto-generate schedule (FlowMatch-style) when no explicit sigmas/timesteps are given
-        if sigmas is None and timesteps is None:
-            if num_inference_steps is None:
-                raise ValueError(
-                    "LTXEulerAncestralRFScheduler.set_timesteps requires either explicit `sigmas`/`timesteps` "
-                    "or a `num_inference_steps` value."
-                )
-
-            # We reuse FlowMatchEulerDiscreteScheduler to construct a sigma schedule that is
-            # consistent with the original LTX training setup (including optional time shifting,
-            # Karras / exponential / beta schedules, etc.).
-            from .scheduling_flow_match_euler_discrete import FlowMatchEulerDiscreteScheduler
-
-            base_scheduler = FlowMatchEulerDiscreteScheduler.from_config(self.config)
-            base_scheduler.set_timesteps(
-                num_inference_steps=num_inference_steps,
-                device=device,
-                sigmas=None,
-                mu=mu,
-                timesteps=None,
-            )
-
-            self.num_inference_steps = base_scheduler.num_inference_steps
-            # Keep sigmas / timesteps on the requested device so step() can operate on-device without
-            # extra transfers.
-            self.sigmas = base_scheduler.sigmas.to(device=device)
-            self.timesteps = base_scheduler.timesteps.to(device=device)
-            self._step_index = None
-            self._begin_index = None
-            return
-
-        # 2. Explicit sigma schedule (ComfyUI-style path)
-        if sigmas is None:
-            # `timesteps` is treated as sigmas in RF / flow-matching setups.
-            sigmas = timesteps
-
-        if isinstance(sigmas, list):
-            sigmas_tensor = torch.tensor(sigmas, dtype=torch.float32)
-        elif isinstance(sigmas, torch.Tensor):
-            sigmas_tensor = sigmas.to(dtype=torch.float32)
-        else:
-            raise TypeError(f"`sigmas` must be a list or torch.Tensor, got {type(sigmas)}.")
-
-        if sigmas_tensor.ndim != 1:
-            raise ValueError(f"`sigmas` must be a 1D tensor, got shape {tuple(sigmas_tensor.shape)}.")
-
-        if sigmas_tensor[-1].abs().item() > 1e-6:
-            logger.warning(
-                "The last sigma in the schedule is not zero (%.6f). "
-                "For best compatibility with ComfyUI's RF sampler, the terminal sigma "
-                "should be 0.0.",
-                sigmas_tensor[-1].item(),
-            )
-
-        # Move to device once, then derive timesteps.
-        if device is not None:
-            sigmas_tensor = sigmas_tensor.to(device)
-
-        # Internal sigma schedule stays in [0, 1] (as provided).
-        self.sigmas = sigmas_tensor
-        # Timesteps are scaled to match the training setup of LTX (FlowMatch-style),
-        # where the network expects timesteps on [0, num_train_timesteps].
-        # This keeps the transformer conditioning in the expected range while the RF
-        # scheduler still operates on the raw sigma values.
-        num_train = float(getattr(self.config, "num_train_timesteps", 1000))
-        self.timesteps = sigmas_tensor * num_train
-
-        if num_inference_steps is not None and num_inference_steps != len(sigmas) - 1:
-            logger.warning(
-                "Provided `num_inference_steps=%d` does not match `len(sigmas)-1=%d`. "
-                "Overriding `num_inference_steps` with `len(sigmas)-1`.",
-                num_inference_steps,
-                len(sigmas) - 1,
-            )
-
-        self.num_inference_steps = len(sigmas) - 1
-        self._step_index = None
-        self._begin_index = None
-
-    def _sigma_broadcast(self, sigma: torch.Tensor, sample: torch.Tensor) -> torch.Tensor:
-        """
-        Helper to broadcast a scalar sigma to the shape of `sample`.
-        """
-        while sigma.ndim < sample.ndim:
-            sigma = sigma.view(*sigma.shape, 1)
-        return sigma
-
-    def step(
-        self,
-        model_output: torch.FloatTensor,
-        timestep: Union[float, torch.Tensor],
-        sample: torch.FloatTensor,
-        generator: Optional[torch.Generator] = None,
-        return_dict: bool = True,
-    ) -> Union[LTXEulerAncestralRFSchedulerOutput, Tuple[torch.FloatTensor]]:
-        """
-        Perform a single Euler-Ancestral RF update step.
-
-        Args:
-            model_output (`torch.FloatTensor`):
-                Raw model output at the current step. Interpreted under the CONST parametrization as `v_t`, with
-                denoised state reconstructed as `x0 = x_t - sigma_t * v_t`.
-            timestep (`float` or `torch.Tensor`):
-                The current sigma value (must match one entry in `self.timesteps`).
-            sample (`torch.FloatTensor`):
-                Current latent sample `x_t`.
-            generator (`torch.Generator`, *optional*):
-                Optional generator for reproducible noise.
-            return_dict (`bool`):
-                If `True`, return a `LTXEulerAncestralRFSchedulerOutput`; otherwise return a tuple where the first
-                element is the updated sample.
-        """
-
-        if isinstance(timestep, (int, torch.IntTensor, torch.LongTensor)):
-            raise ValueError(
-                (
-                    "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to"
-                    " `LTXEulerAncestralRFScheduler.step()` is not supported. Make sure to pass"
-                    " one of the `scheduler.timesteps` values as `timestep`."
-                ),
-            )
-
-        if self.sigmas is None or self.timesteps is None:
-            raise ValueError("Scheduler has not been initialized. Call `set_timesteps` before `step`.")
-
-        if self._step_index is None:
-            self._init_step_index(timestep)
-
-        i = self._step_index
-        if i >= len(self.sigmas) - 1:
-            # Already at the end; simply return the current sample.
-            prev_sample = sample
-        else:
-            # Work in float32 for numerical stability
-            sample_f = sample.to(torch.float32)
-            model_output_f = model_output.to(torch.float32)
-
-            sigma = self.sigmas[i]
-            sigma_next = self.sigmas[i + 1]
-
-            sigma_b = self._sigma_broadcast(sigma.view(1), sample_f)
-            sigma_next_b = self._sigma_broadcast(sigma_next.view(1), sample_f)
-
-            # Approximate denoised x0 under CONST parametrization:
-            #   x0 = x_t - sigma_t * v_t
-            denoised = sample_f - sigma_b * model_output_f
-
-            if sigma_next.abs().item() < 1e-8:
-                # Final denoising step
-                x = denoised
-            else:
-                eta = float(self.config.eta)
-                s_noise = float(self.config.s_noise)
-
-                # Downstep computation (ComfyUI RF variant)
-                downstep_ratio = 1.0 + (sigma_next / sigma - 1.0) * eta
-                sigma_down = sigma_next * downstep_ratio
-
-                alpha_ip1 = 1.0 - sigma_next
-                alpha_down = 1.0 - sigma_down
-
-                # Deterministic part (Euler step in (x, x0)-space)
-                sigma_down_b = self._sigma_broadcast(sigma_down.view(1), sample_f)
-                alpha_ip1_b = self._sigma_broadcast(alpha_ip1.view(1), sample_f)
-                alpha_down_b = self._sigma_broadcast(alpha_down.view(1), sample_f)
-
-                sigma_ratio = sigma_down_b / sigma_b
-                x = sigma_ratio * sample_f + (1.0 - sigma_ratio) * denoised
-
-                # Stochastic ancestral noise
-                if eta > 0.0 and s_noise > 0.0:
-                    renoise_coeff = (
-                        (sigma_next_b**2 - sigma_down_b**2 * alpha_ip1_b**2 / (alpha_down_b**2 + 1e-12))
-                        .clamp(min=0.0)
-                        .sqrt()
-                    )
-
-                    noise = randn_tensor(
-                        sample_f.shape, generator=generator, device=sample_f.device, dtype=sample_f.dtype
-                    )
-                    x = (alpha_ip1_b / (alpha_down_b + 1e-12)) * x + noise * renoise_coeff * s_noise
-
-            prev_sample = x.to(sample.dtype)
-
-        # Advance internal step index
-        self._step_index = min(self._step_index + 1, len(self.sigmas) - 1)
-
-        if not return_dict:
-            return (prev_sample,)
-
-        return LTXEulerAncestralRFSchedulerOutput(prev_sample=prev_sample)
-
-    def __len__(self) -> int:
-        # For compatibility with other schedulers; used e.g. in some training
-        # utilities to infer the maximum number of training timesteps.
-        return int(getattr(self.config, "num_train_timesteps", 1000))

src/diffusers/training_utils.py

@@ -6,18 +6,11 @@ import random
 import re
 import warnings
 from contextlib import contextmanager
-from functools import partial
-from typing import Any, Dict, Iterable, List, Optional, Set, Tuple, Type, Union
+from typing import Any, Dict, Iterable, List, Optional, Tuple, Union
 
 import numpy as np
 import torch
 
-
-if getattr(torch, "distributed", None) is not None:
-    from torch.distributed.fsdp import CPUOffload, ShardingStrategy
-    from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
-    from torch.distributed.fsdp.wrap import transformer_auto_wrap_policy
-
 from .models import UNet2DConditionModel
 from .pipelines import DiffusionPipeline
 from .schedulers import SchedulerMixin
@@ -25,7 +18,6 @@ from .utils import (
     convert_state_dict_to_diffusers,
     convert_state_dict_to_peft,
     deprecate,
-    is_accelerate_available,
     is_peft_available,
     is_torch_npu_available,
     is_torchvision_available,
@@ -39,9 +31,6 @@ if is_transformers_available():
     if transformers.integrations.deepspeed.is_deepspeed_zero3_enabled():
         import deepspeed
 
-if is_accelerate_available():
-    from accelerate.logging import get_logger
-
 if is_peft_available():
     from peft import set_peft_model_state_dict
 
@@ -405,86 +394,6 @@ def find_nearest_bucket(h, w, bucket_options):
     return best_bucket_idx
 
 
-def _to_cpu_contiguous(state_dicts) -> dict:
-    return {k: v.detach().cpu().contiguous() if isinstance(v, torch.Tensor) else v for k, v in state_dicts.items()}
-
-
-def get_fsdp_kwargs_from_accelerator(accelerator) -> dict:
-    """
-    Extract and convert FSDP config from Accelerator into PyTorch FSDP kwargs.
-    """
-
-    kwargs = {}
-    fsdp_state = getattr(accelerator.state, "fsdp_plugin", None)
-
-    if fsdp_state is None:
-        raise ValueError("Accelerate isn't configured to handle FSDP. Please update your installation.")
-
-    fsdp_plugin = accelerator.state.fsdp_plugin
-
-    if fsdp_plugin is None:
-        # FSDP not enabled in Accelerator
-        kwargs["sharding_strategy"] = ShardingStrategy.FULL_SHARD
-    else:
-        # FSDP is enabled → use plugin's strategy, or default if None
-        kwargs["sharding_strategy"] = fsdp_plugin.sharding_strategy or ShardingStrategy.FULL_SHARD
-
-    return kwargs
-
-
-def wrap_with_fsdp(
-    model: torch.nn.Module,
-    device: Union[str, torch.device],
-    offload: bool = True,
-    use_orig_params: bool = True,
-    limit_all_gathers: bool = True,
-    fsdp_kwargs: Optional[Dict[str, Any]] = None,
-    transformer_layer_cls: Optional[Set[Type[torch.nn.Module]]] = None,
-) -> FSDP:
-    """
-    Wrap a model with FSDP using common defaults and optional transformer auto-wrapping.
-
-    Args:
-        model: Model to wrap
-        device: Target device (e.g., accelerator.device)
-        offload: Whether to enable CPU parameter offloading
-        use_orig_params: Whether to use original parameters
-        limit_all_gathers: Whether to limit all gathers
-        fsdp_kwargs: FSDP arguments (sharding_strategy, etc.) — usually from Accelerate config
-        transformer_layer_cls: Classes for auto-wrapping (if not using policy from fsdp_kwargs)
-
-    Returns:
-        FSDP-wrapped model
-    """
-
-    logger = get_logger(__name__)
-
-    if transformer_layer_cls is None:
-        # Set the default layers if transformer_layer_cls is not provided
-        transformer_layer_cls = type(model.model.language_model.layers[0])
-        logger.info(f"transformer_layer_cls is not provided, auto-inferred as {transformer_layer_cls.__name__}")
-
-    # Add auto-wrap policy if transformer layers specified
-    auto_wrap_policy = partial(
-        transformer_auto_wrap_policy,
-        transformer_layer_cls={transformer_layer_cls},
-    )
-
-    config = {
-        "device_id": device,
-        "cpu_offload": CPUOffload(offload_params=offload) if offload else None,
-        "use_orig_params": use_orig_params,
-        "limit_all_gathers": limit_all_gathers,
-        "auto_wrap_policy": auto_wrap_policy,
-    }
-
-    if fsdp_kwargs:
-        config.update(fsdp_kwargs)
-
-    fsdp_model = FSDP(model, **config)
-    return fsdp_model
-
-
 # Adapted from torch-ema https://github.com/fadel/pytorch_ema/blob/master/torch_ema/ema.py#L14
 class EMAModel:
     """

src/diffusers/utils/dummy_pt_objects.py

@@ -2634,21 +2634,6 @@ class LCMScheduler(metaclass=DummyObject):
         requires_backends(cls, ["torch"])
 
 
-class LTXEulerAncestralRFScheduler(metaclass=DummyObject):
-    _backends = ["torch"]
-
-    def __init__(self, *args, **kwargs):
-        requires_backends(self, ["torch"])
-
-    @classmethod
-    def from_config(cls, *args, **kwargs):
-        requires_backends(cls, ["torch"])
-
-    @classmethod
-    def from_pretrained(cls, *args, **kwargs):
-        requires_backends(cls, ["torch"])
-
-
 class PNDMScheduler(metaclass=DummyObject):
     _backends = ["torch"]
 

src/diffusers/utils/dummy_torch_and_transformers_objects.py

@@ -1892,21 +1892,6 @@ class LTXConditionPipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
-class LTXI2VLongMultiPromptPipeline(metaclass=DummyObject):
-    _backends = ["torch", "transformers"]
-
-    def __init__(self, *args, **kwargs):
-        requires_backends(self, ["torch", "transformers"])
-
-    @classmethod
-    def from_config(cls, *args, **kwargs):
-        requires_backends(cls, ["torch", "transformers"])
-
-    @classmethod
-    def from_pretrained(cls, *args, **kwargs):
-        requires_backends(cls, ["torch", "transformers"])
-
-
 class LTXImageToVideoPipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]