update

docs/source/en/_toctree.yml

@@ -22,8 +22,6 @@
     title: Reproducibility
   - local: using-diffusers/schedulers
     title: Schedulers
-  - local: using-diffusers/guiders
-    title: Guiders
   - local: using-diffusers/automodel
     title: AutoModel
   - local: using-diffusers/other-formats
@@ -112,6 +110,8 @@
     title: ModularPipeline
   - local: modular_diffusers/components_manager
     title: ComponentsManager
+  - local: modular_diffusers/guiders
+    title: Guiders
   - local: modular_diffusers/custom_blocks
     title: Building Custom Blocks
   - local: modular_diffusers/mellon

docs/source/en/api/pipelines/hunyuan_video15.md

@@ -99,7 +99,7 @@ To update guider configuration, you can run `pipe.guider = pipe.guider.new(...)`
 pipe.guider = pipe.guider.new(guidance_scale=5.0)
 ```
 
-Read more on Guider [here](../../using-diffusers/guiders).
+Read more on Guider [here](../../modular_diffusers/guiders).
 
 
 

docs/source/en/api/pipelines/hunyuanimage21.md

@@ -30,7 +30,7 @@ HunyuanImage-2.1 comes in the following variants:
 
 ## HunyuanImage-2.1
 
-HunyuanImage-2.1 applies [Adaptive Projected Guidance (APG)](https://huggingface.co/papers/2410.02416) combined with Classifier-Free Guidance (CFG) in the denoising loop. `HunyuanImagePipeline` has a `guider` component (read more about [Guider](../../using-diffusers/guiders)) and does not take a `guidance_scale` parameter at runtime. To change guider-related parameters, e.g., `guidance_scale`, you can update the `guider` configuration instead.
+HunyuanImage-2.1 applies [Adaptive Projected Guidance (APG)](https://huggingface.co/papers/2410.02416) combined with Classifier-Free Guidance (CFG) in the denoising loop. `HunyuanImagePipeline` has a `guider` component (read more about [Guider](../modular_diffusers/guiders.md)) and does not take a `guidance_scale` parameter at runtime. To change guider-related parameters, e.g., `guidance_scale`, you can update the `guider` configuration instead.
 
 ```python
 import torch

docs/source/en/modular_diffusers/modular_pipeline.md

@@ -338,7 +338,7 @@ guider = ClassifierFreeGuidance(guidance_scale=5.0)
 pipeline.update_components(guider=guider)
 ```
 
-See the [Guiders](../using-diffusers/guiders) guide for more details on available guiders and how to configure them.
+See the [Guiders](./guiders) guide for more details on available guiders and how to configure them.
 
 ## Splitting a pipeline into stages
 

docs/source/en/modular_diffusers/overview.md

@@ -39,7 +39,7 @@ The Modular Diffusers docs are organized as shown below.
 
 - [ModularPipeline](./modular_pipeline) shows you how to create and convert pipeline blocks into an executable [`ModularPipeline`].
 - [ComponentsManager](./components_manager) shows you how to manage and reuse components across multiple pipelines.
-- [Guiders](../using-diffusers/guiders) shows you how to use different guidance methods in the pipeline.
+- [Guiders](./guiders) shows you how to use different guidance methods in the pipeline.
 
 ## Mellon Integration
 

docs/source/en/optimization/memory.md

@@ -482,6 +482,144 @@ print(
 )  # (2880, 1, 960, 320) having a stride of 1 for the 2nd dimension proves that it works
 ```
 
+## torch.jit.trace
+
+[torch.jit.trace](https://pytorch.org/docs/stable/generated/torch.jit.trace.html) records the operations a model performs on a sample input and creates a new, optimized representation of the model based on the recorded execution path. During tracing, the model is optimized to reduce overhead from Python and dynamic control flows and operations are fused together for more efficiency. The returned executable or [ScriptFunction](https://pytorch.org/docs/stable/generated/torch.jit.ScriptFunction.html) can be compiled.
+
+```py
+import time
+import torch
+from diffusers import StableDiffusionPipeline
+import functools
+
+# torch disable grad
+torch.set_grad_enabled(False)
+
+# set variables
+n_experiments = 2
+unet_runs_per_experiment = 50
+
+# load sample inputs
+def generate_inputs():
+    sample = torch.randn((2, 4, 64, 64), device="cuda", dtype=torch.float16)
+    timestep = torch.rand(1, device="cuda", dtype=torch.float16) * 999
+    encoder_hidden_states = torch.randn((2, 77, 768), device="cuda", dtype=torch.float16)
+    return sample, timestep, encoder_hidden_states
+
+
+pipeline = StableDiffusionPipeline.from_pretrained(
+    "stable-diffusion-v1-5/stable-diffusion-v1-5",
+    torch_dtype=torch.float16,
+    use_safetensors=True,
+).to("cuda")
+unet = pipeline.unet
+unet.eval()
+unet.to(memory_format=torch.channels_last)  # use channels_last memory format
+unet.forward = functools.partial(unet.forward, return_dict=False)  # set return_dict=False as default
+
+# warmup
+for _ in range(3):
+    with torch.inference_mode():
+        inputs = generate_inputs()
+        orig_output = unet(*inputs)
+
+# trace
+print("tracing..")
+unet_traced = torch.jit.trace(unet, inputs)
+unet_traced.eval()
+print("done tracing")
+
+# warmup and optimize graph
+for _ in range(5):
+    with torch.inference_mode():
+        inputs = generate_inputs()
+        orig_output = unet_traced(*inputs)
+
+# benchmarking
+with torch.inference_mode():
+    for _ in range(n_experiments):
+        torch.cuda.synchronize()
+        start_time = time.time()
+        for _ in range(unet_runs_per_experiment):
+            orig_output = unet_traced(*inputs)
+        torch.cuda.synchronize()
+        print(f"unet traced inference took {time.time() - start_time:.2f} seconds")
+    for _ in range(n_experiments):
+        torch.cuda.synchronize()
+        start_time = time.time()
+        for _ in range(unet_runs_per_experiment):
+            orig_output = unet(*inputs)
+        torch.cuda.synchronize()
+        print(f"unet inference took {time.time() - start_time:.2f} seconds")
+
+# save the model
+unet_traced.save("unet_traced.pt")
+```
+
+Replace the pipeline's UNet with the traced version.
+
+```py
+import torch
+from diffusers import StableDiffusionPipeline
+from dataclasses import dataclass
+
+@dataclass
+class UNet2DConditionOutput:
+    sample: torch.Tensor
+
+pipeline = StableDiffusionPipeline.from_pretrained(
+    "stable-diffusion-v1-5/stable-diffusion-v1-5",
+    torch_dtype=torch.float16,
+    use_safetensors=True,
+).to("cuda")
+
+# use jitted unet
+unet_traced = torch.jit.load("unet_traced.pt")
+
+# del pipeline.unet
+class TracedUNet(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.in_channels = pipe.unet.config.in_channels
+        self.device = pipe.unet.device
+
+    def forward(self, latent_model_input, t, encoder_hidden_states):
+        sample = unet_traced(latent_model_input, t, encoder_hidden_states)[0]
+        return UNet2DConditionOutput(sample=sample)
+
+pipeline.unet = TracedUNet()
+
+with torch.inference_mode():
+    image = pipe([prompt] * 1, num_inference_steps=50).images[0]
+```
+
 ## Memory-efficient attention
 
-Diffusers supports multiple memory-efficient attention backends (FlashAttention, xFormers, SageAttention, and more) through [`~ModelMixin.set_attention_backend`]. Refer to the [Attention backends](./attention_backends) guide to learn how to switch between them.
+> [!TIP]
+> Memory-efficient attention optimizes for memory usage *and* [inference speed](./fp16#scaled-dot-product-attention)!
+
+The Transformers attention mechanism is memory-intensive, especially for long sequences, so you can try using different and more memory-efficient attention types.
+
+By default, if PyTorch >= 2.0 is installed, [scaled dot-product attention (SDPA)](https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html) is used. You don't need to make any additional changes to your code.
+
+SDPA supports [FlashAttention](https://github.com/Dao-AILab/flash-attention) and [xFormers](https://github.com/facebookresearch/xformers) as well as a native C++ PyTorch implementation. It automatically selects the most optimal implementation based on your input.
+
+You can explicitly use xFormers with the [`~ModelMixin.enable_xformers_memory_efficient_attention`] method.
+
+```py
+# pip install xformers
+import torch
+from diffusers import StableDiffusionXLPipeline
+
+pipeline = StableDiffusionXLPipeline.from_pretrained(
+    "stabilityai/stable-diffusion-xl-base-1.0",
+    torch_dtype=torch.float16,
+).to("cuda")
+pipeline.enable_xformers_memory_efficient_attention()
+```
+
+Call [`~ModelMixin.disable_xformers_memory_efficient_attention`] to disable it.
+
+```py
+pipeline.disable_xformers_memory_efficient_attention()
+```

docs/source/en/optimization/xformers.md

@@ -23,7 +23,7 @@ pip install xformers
 > [!TIP]
 > The xFormers `pip` package requires the latest version of PyTorch. If you need to use a previous version of PyTorch, then we recommend [installing xFormers from the source](https://github.com/facebookresearch/xformers#installing-xformers).
 
-After xFormers is installed, you can use it with [`~ModelMixin.set_attention_backend`] as shown in the [Attention backends](./attention_backends) guide.
+After xFormers is installed, you can use `enable_xformers_memory_efficient_attention()` for faster inference and reduced memory consumption as shown in this [section](memory#memory-efficient-attention).
 
 > [!WARNING]
 > According to this [issue](https://github.com/huggingface/diffusers/issues/2234#issuecomment-1416931212), xFormers `v0.0.16` cannot be used for training (fine-tune or DreamBooth) in some GPUs. If you observe this problem, please install a development version as indicated in the issue comments.

docs/source/zh/_toctree.yml

@@ -14,8 +14,6 @@
   sections:
   - local: using-diffusers/schedulers
     title: Load schedulers and models
-  - local: using-diffusers/guiders
-    title: Guiders
 
 - title: Inference
   isExpanded: false
@@ -82,6 +80,8 @@
     title: ModularPipeline
   - local: modular_diffusers/components_manager
     title: ComponentsManager
+  - local: modular_diffusers/guiders
+    title: Guiders
 
 - title: Training
   isExpanded: false

tests/models/testing_utils/parallelism.py

@@ -26,17 +26,9 @@ from diffusers.models._modeling_parallel import ContextParallelConfig
 from ...testing_utils import (
     is_context_parallel,
     require_torch_multi_accelerator,
-    torch_device,
 )
 
 
-# Device configuration mapping
-DEVICE_CONFIG = {
-    "cuda": {"backend": "nccl", "module": torch.cuda},
-    "xpu": {"backend": "xccl", "module": torch.xpu},
-}
-
-
 def _find_free_port():
     """Find a free port on localhost."""
     with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
@@ -55,17 +47,12 @@ def _context_parallel_worker(rank, world_size, master_port, model_class, init_di
         os.environ["RANK"] = str(rank)
         os.environ["WORLD_SIZE"] = str(world_size)
 
-        # Get device configuration
-        device_config = DEVICE_CONFIG.get(torch_device, DEVICE_CONFIG["cuda"])
-        backend = device_config["backend"]
-        device_module = device_config["module"]
-
         # Initialize process group
-        dist.init_process_group(backend=backend, rank=rank, world_size=world_size)
+        dist.init_process_group(backend="nccl", rank=rank, world_size=world_size)
 
         # Set device for this process
-        device_module.set_device(rank)
-        device = torch.device(f"{torch_device}:{rank}")
+        torch.cuda.set_device(rank)
+        device = torch.device(f"cuda:{rank}")
 
         # Create model
         model = model_class(**init_dict)
@@ -116,16 +103,10 @@ def _custom_mesh_worker(
         os.environ["RANK"] = str(rank)
         os.environ["WORLD_SIZE"] = str(world_size)
 
-        # Get device configuration
-        device_config = DEVICE_CONFIG.get(torch_device, DEVICE_CONFIG["cuda"])
-        backend = device_config["backend"]
-        device_module = device_config["module"]
+        dist.init_process_group(backend="nccl", rank=rank, world_size=world_size)
 
-        dist.init_process_group(backend=backend, rank=rank, world_size=world_size)
-
-        # Set device for this process
-        device_module.set_device(rank)
-        device = torch.device(f"{torch_device}:{rank}")
+        torch.cuda.set_device(rank)
+        device = torch.device(f"cuda:{rank}")
 
         model = model_class(**init_dict)
         model.to(device)
@@ -135,7 +116,7 @@ def _custom_mesh_worker(
 
         # DeviceMesh must be created after init_process_group, inside each worker process.
         mesh = torch.distributed.device_mesh.init_device_mesh(
-            torch_device, mesh_shape=mesh_shape, mesh_dim_names=mesh_dim_names
+            "cuda", mesh_shape=mesh_shape, mesh_dim_names=mesh_dim_names
         )
         cp_config = ContextParallelConfig(**cp_dict, mesh=mesh)
         model.enable_parallelism(config=cp_config)

tests/pipelines/hunyuan_video/test_hunyuan_image2video.py

@@ -139,9 +139,7 @@ class HunyuanVideoImageToVideoPipelineFastTests(
             num_hidden_layers=2,
             image_size=224,
         )
-        llava_text_encoder_config = LlavaConfig(
-            vision_config=vision_config, text_config=text_config, pad_token_id=100, image_token_index=101
-        )
+        llava_text_encoder_config = LlavaConfig(vision_config, text_config, pad_token_id=100, image_token_index=101)
 
         clip_text_encoder_config = CLIPTextConfig(
             bos_token_id=0,

utils/make_tiny_model.py

@@ -1,210 +0,0 @@
-# /// script
-# requires-python = ">=3.10"
-# dependencies = [
-#     "diffusers",
-#     "torch",
-#     "huggingface_hub",
-#     "accelerate",
-#     "transformers",
-#     "sentencepiece",
-#     "protobuf",
-# ]
-# ///
-
-# Copyright 2025 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.
-
-"""
-Utility script to create tiny versions of diffusers models by reducing layer counts.
-
-Can be run locally or submitted as an HF Job via `--launch`.
-
-Usage:
-    # Run locally
-    python make_tiny_model.py --model_repo_id <model_repo_id> --output_repo_id <output_repo_id> [--subfolder transformer] [--num_layers 2]
-
-    # Push to Hub
-    python make_tiny_model.py --model_repo_id <model_repo_id> --output_repo_id <output_repo_id> --push_to_hub --token $HF_TOKEN
-
-    # Submit as an HF Job
-    python make_tiny_model.py --model_repo_id <model_repo_id> --output_repo_id <output_repo_id> --launch [--flavor cpu-basic]
-"""
-
-import argparse
-import os
-import re
-
-
-LAYER_PARAM_PATTERN = re.compile(r"^(num_.*layers?|n_layers|n_refiner_layers)$")
-
-DIM_PARAM_PATTERNS = {
-    re.compile(r"^num_attention_heads$"): 2,
-    re.compile(r"^num_.*attention_heads$"): 2,
-    re.compile(r"^num_key_value_heads$"): 2,
-    re.compile(r"^num_kv_heads$"): 1,
-    re.compile(r"^n_heads$"): 2,
-    re.compile(r"^n_kv_heads$"): 2,
-    re.compile(r"^attention_head_dim$"): 8,
-    re.compile(r"^.*attention_head_dim$"): 4,
-    re.compile(r"^cross_attention_dim.*$"): 8,
-    re.compile(r"^joint_attention_dim$"): 32,
-    re.compile(r"^pooled_projection_dim$"): 32,
-    re.compile(r"^caption_projection_dim$"): 32,
-    re.compile(r"^caption_channels$"): 8,
-    re.compile(r"^cap_feat_dim$"): 16,
-    re.compile(r"^hidden_size$"): 16,
-    re.compile(r"^dim$"): 16,
-    re.compile(r"^.*embed_dim$"): 16,
-    re.compile(r"^.*embed_.*dim$"): 16,
-    re.compile(r"^text_dim$"): 16,
-    re.compile(r"^time_embed_dim$"): 4,
-    re.compile(r"^ffn_dim$"): 32,
-    re.compile(r"^intermediate_size$"): 32,
-    re.compile(r"^sample_size$"): 32,
-}
-
-
-def parse_args():
-    parser = argparse.ArgumentParser(description="Create a tiny version of a diffusers model.")
-    parser.add_argument("--model_repo_id", type=str, required=True, help="HuggingFace repo ID of the source model.")
-    parser.add_argument(
-        "--output_repo_id",
-        type=str,
-        required=True,
-        help="HuggingFace repo ID or local path to save the tiny model to.",
-    )
-    parser.add_argument("--subfolder", type=str, default=None, help="Subfolder within the model repo.")
-    parser.add_argument("--num_layers", type=int, default=2, help="Number of layers to use for the tiny model.")
-    parser.add_argument(
-        "--shrink_dims",
-        action="store_true",
-        help="Also reduce dimension parameters (attention heads, hidden size, embedding dims, etc.).",
-    )
-    parser.add_argument("--push_to_hub", action="store_true", help="Push the tiny model to the HuggingFace Hub.")
-    parser.add_argument(
-        "--token", type=str, default=None, help="HuggingFace token. Defaults to $HF_TOKEN env var if not provided."
-    )
-
-    launch_group = parser.add_argument_group("HF Jobs launch options")
-    launch_group.add_argument("--launch", action="store_true", help="Submit as an HF Job instead of running locally.")
-    launch_group.add_argument("--flavor", type=str, default="cpu-basic", help="HF Jobs hardware flavor.")
-    launch_group.add_argument("--timeout", type=str, default="30m", help="HF Jobs timeout.")
-
-    args = parser.parse_args()
-    if args.token is None:
-        args.token = os.environ.get("HF_TOKEN")
-    return args
-
-
-def launch_job(args):
-    from huggingface_hub import run_uv_job
-
-    script_args = [
-        "--model_repo_id",
-        args.model_repo_id,
-        "--output_repo_id",
-        args.output_repo_id,
-        "--num_layers",
-        str(args.num_layers),
-    ]
-    if args.subfolder:
-        script_args.extend(["--subfolder", args.subfolder])
-    if args.shrink_dims:
-        script_args.append("--shrink_dims")
-    if args.push_to_hub:
-        script_args.append("--push_to_hub")
-
-    job = run_uv_job(
-        __file__,
-        script_args=script_args,
-        flavor=args.flavor,
-        timeout=args.timeout,
-        secrets={"HF_TOKEN": args.token} if args.token else {},
-    )
-    print(f"Job submitted: {job.url}")
-    print(f"Job ID: {job.id}")
-    return job
-
-
-def make_tiny_model(
-    model_repo_id, output_repo_id, subfolder=None, num_layers=2, shrink_dims=False, push_to_hub=False, token=None
-):
-    from diffusers import AutoModel
-
-    config_kwargs = {}
-    if token:
-        config_kwargs["token"] = token
-
-    config = AutoModel.load_config(model_repo_id, subfolder=subfolder, **config_kwargs)
-
-    modified_keys = {}
-    for key, value in config.items():
-        if LAYER_PARAM_PATTERN.match(key) and isinstance(value, int) and value > num_layers:
-            modified_keys[key] = (value, num_layers)
-            config[key] = num_layers
-
-    if shrink_dims:
-        for key, value in config.items():
-            if not isinstance(value, int) or key.startswith("_"):
-                continue
-            for pattern, tiny_value in DIM_PARAM_PATTERNS.items():
-                if pattern.match(key) and value > tiny_value:
-                    modified_keys[key] = (value, tiny_value)
-                    config[key] = tiny_value
-                    break
-
-    if not modified_keys:
-        print("WARNING: No config parameters were modified.")
-        print(f"Config keys: {[k for k in config if not k.startswith('_')]}")
-        return
-
-    print("Modified config parameters:")
-    for key, (old, new) in modified_keys.items():
-        print(f"  {key}: {old} -> {new}")
-
-    model = AutoModel.from_config(config)
-    total_params = sum(p.numel() for p in model.parameters())
-    print(f"Tiny model created with {total_params:,} parameters.")
-
-    save_kwargs = {}
-    if token:
-        save_kwargs["token"] = token
-    if push_to_hub:
-        save_kwargs["repo_id"] = output_repo_id
-    model.save_pretrained(output_repo_id, push_to_hub=push_to_hub, **save_kwargs)
-    if push_to_hub:
-        print(f"Model pushed to https://huggingface.co/{output_repo_id}")
-    else:
-        print(f"Model saved to {output_repo_id}")
-
-
-def main():
-    args = parse_args()
-
-    if args.launch:
-        launch_job(args)
-    else:
-        make_tiny_model(
-            model_repo_id=args.model_repo_id,
-            output_repo_id=args.output_repo_id,
-            subfolder=args.subfolder,
-            num_layers=args.num_layers,
-            shrink_dims=args.shrink_dims,
-            push_to_hub=args.push_to_hub,
-            token=args.token,
-        )
-
-
-if __name__ == "__main__":
-    main()