Merge branch 'main' into fix-vae-scaling-in-training

Author: asomoza

docs/source/en/optimization/fp16.md

@@ -174,39 +174,36 @@ Feel free to open an issue if dynamic compilation doesn't work as expected for a
 
 ### Regional compilation
 
+[Regional compilation](https://docs.pytorch.org/tutorials/recipes/regional_compilation.html) trims cold-start latency by only compiling the *small and frequently-repeated block(s)* of a model - typically a transformer layer - and enables reusing compiled artifacts for every subsequent occurrence.
+For many diffusion architectures, this delivers the same runtime speedups as full-graph compilation and reduces compile time by 8–10x.
 
-[Regional compilation](https://docs.pytorch.org/tutorials/recipes/regional_compilation.html) trims cold-start latency by compiling **only the small, frequently-repeated block(s)** of a model, typically a Transformer layer, enabling reuse of compiled artifacts for every subsequent occurrence.
-For many diffusion architectures this delivers the *same* runtime speed-ups as full-graph compilation yet cuts compile time by **8–10 ×**.
-
-To make this effortless, [`ModelMixin`] exposes [`ModelMixin.compile_repeated_blocks`] API, a helper that wraps `torch.compile` around any sub-modules you designate as repeatable:
+Use the [`~ModelMixin.compile_repeated_blocks`] method, a helper that wraps `torch.compile`, on any component such as the transformer model as shown below.
 
 ```py
 # pip install -U diffusers
 import torch
 from diffusers import StableDiffusionXLPipeline
 
-pipe = StableDiffusionXLPipeline.from_pretrained(
+pipeline = StableDiffusionXLPipeline.from_pretrained(
     "stabilityai/stable-diffusion-xl-base-1.0",
     torch_dtype=torch.float16,
 ).to("cuda")
 
-# Compile only the repeated Transformer layers inside the UNet
-pipe.unet.compile_repeated_blocks(fullgraph=True)
+# compile only the repeated transformer layers inside the UNet
+pipeline.unet.compile_repeated_blocks(fullgraph=True)
 ```
 
-To enable a new model with regional compilation, add a `_repeated_blocks` attribute to your model class containing the class names (as strings) of the blocks you want compiled:
-
+To enable regional compilation for a new model, add a `_repeated_blocks` attribute to a model class containing the class names (as strings) of the blocks you want to compile.
 
 ```py
 class MyUNet(ModelMixin):
     _repeated_blocks = ("Transformer2DModel",)  # ← compiled by default
 ```
 
-For more examples, see the reference [PR](https://github.com/huggingface/diffusers/pull/11705).
-
-**Relation to Accelerate compile_regions** There is also a separate API in [accelerate](https://huggingface.co/docs/accelerate/index) - [compile_regions](https://github.com/huggingface/accelerate/blob/273799c85d849a1954a4f2e65767216eb37fa089/src/accelerate/utils/other.py#L78). It takes a fully automatic approach: it walks the module, picks candidate blocks, then compiles the remaining graph separately. That hands-off experience is handy for quick experiments, but it also leaves fewer knobs when you want to fine-tune which blocks are compiled or adjust compilation flags.
-
+> [!TIP]
+> For more regional compilation examples, see the reference [PR](https://github.com/huggingface/diffusers/pull/11705).
 
+There is also a [compile_regions](https://github.com/huggingface/accelerate/blob/273799c85d849a1954a4f2e65767216eb37fa089/src/accelerate/utils/other.py#L78) method in [Accelerate](https://huggingface.co/docs/accelerate/index) that automatically selects candidate blocks in a model to compile. The remaining graph is compiled separately. This is useful for quick experiments because there aren't as many options for you to set which blocks to compile or adjust compilation flags.
 
 ```py
 # pip install -U accelerate
@@ -219,8 +216,8 @@ pipeline = StableDiffusionXLPipeline.from_pretrained(
 ).to("cuda")
 pipeline.unet = compile_regions(pipeline.unet, mode="reduce-overhead", fullgraph=True)
 ```
-`compile_repeated_blocks`, by contrast, is intentionally explicit. You list the repeated blocks once (via `_repeated_blocks`) and the helper compiles exactly those, nothing more. In practice this small dose of control hits a sweet spot for diffusion models: predictable behavior, easy reasoning about cache reuse, and still a one-liner for users.
 
+[`~ModelMixin.compile_repeated_blocks`] is intentionally explicit. List the blocks to repeat in `_repeated_blocks` and the helper only compiles those blocks. It offers predictable behavior and easy reasoning about cache reuse in one line of code.
 
 ### Graph breaks
 
@@ -296,3 +293,9 @@ An input is projected into three subspaces, represented by the projection matric
 ```py
 pipeline.fuse_qkv_projections()
 ```
+
+## Resources
+
+- Read the [Presenting Flux Fast: Making Flux go brrr on H100s](https://pytorch.org/blog/presenting-flux-fast-making-flux-go-brrr-on-h100s/) blog post to learn more about how you can combine all of these optimizations with [TorchInductor](https://docs.pytorch.org/docs/stable/torch.compiler.html) and [AOTInductor](https://docs.pytorch.org/docs/stable/torch.compiler_aot_inductor.html) for a ~2.5x speedup using recipes from [flux-fast](https://github.com/huggingface/flux-fast).
+
+    These recipes support AMD hardware and [Flux.1 Kontext Dev](https://huggingface.co/black-forest-labs/FLUX.1-Kontext-dev).

docs/source/en/optimization/speed-memory-optims.md

@@ -14,6 +14,9 @@ specific language governing permissions and limitations under the License.
 
 Optimizing models often involves trade-offs between [inference speed](./fp16) and [memory-usage](./memory). For instance, while [caching](./cache) can boost inference speed, it also increases memory consumption since it needs to store the outputs of intermediate attention layers. A more balanced optimization strategy combines quantizing a model, [torch.compile](./fp16#torchcompile) and various [offloading methods](./memory#offloading).
 
+> [!TIP]
+> Check the [torch.compile](./fp16#torchcompile) guide to learn more about compilation and how they can be applied here. For example, regional compilation can significantly reduce compilation time without giving up any speedups. 
+
 For image generation, combining quantization and [model offloading](./memory#model-offloading) can often give the best trade-off between quality, speed, and memory. Group offloading is not as effective for image generation because it is usually not possible to *fully* overlap data transfer if the compute kernel finishes faster. This results in some communication overhead between the CPU and GPU.
 
 For video generation, combining quantization and [group-offloading](./memory#group-offloading) tends to be better because video models are more compute-bound. 
@@ -25,7 +28,7 @@ The table below provides a comparison of optimization strategy combinations and
 | quantization  | 32.602 | 14.9453 |
 | quantization, torch.compile  | 25.847 | 14.9448 |
 | quantization, torch.compile, model CPU offloading | 32.312 | 12.2369 |
-<small>These results are benchmarked on Flux with a RTX 4090. The transformer and text_encoder components are quantized. Refer to the <a href="https://gist.github.com/sayakpaul/0db9d8eeeb3d2a0e5ed7cf0d9ca19b7d" benchmarking script</a> if you're interested in evaluating your own model.</small>
+<small>These results are benchmarked on Flux with a RTX 4090. The transformer and text_encoder components are quantized. Refer to the [benchmarking script](https://gist.github.com/sayakpaul/0db9d8eeeb3d2a0e5ed7cf0d9ca19b7d) if you're interested in evaluating your own model.</small>
 
 This guide will show you how to compile and offload a quantized model with [bitsandbytes](../quantization/bitsandbytes#torchcompile). Make sure you are using [PyTorch nightly](https://pytorch.org/get-started/locally/) and the latest version of bitsandbytes.
 

src/diffusers/loaders/single_file_model.py

@@ -24,6 +24,7 @@
 from .. import __version__
 from ..quantizers import DiffusersAutoQuantizer
 from ..utils import deprecate, is_accelerate_available, logging
+from ..utils.torch_utils import device_synchronize, empty_device_cache
 from .single_file_utils import (
     SingleFileComponentError,
     convert_animatediff_checkpoint_to_diffusers,
@@ -430,6 +431,10 @@ def from_single_file(cls, pretrained_model_link_or_path_or_dict: Optional[str] =
                 keep_in_fp32_modules=keep_in_fp32_modules,
                 unexpected_keys=unexpected_keys,
             )
+            # Ensure tensors are correctly placed on device by synchronizing before returning control to user. This is
+            # required because we move tensors with non_blocking=True, which is slightly faster for model loading.
+            empty_device_cache()
+            device_synchronize()
         else:
             _, unexpected_keys = model.load_state_dict(diffusers_format_checkpoint, strict=False)
 

src/diffusers/loaders/single_file_utils.py

@@ -46,6 +46,7 @@
 )
 from ..utils.constants import DIFFUSERS_REQUEST_TIMEOUT
 from ..utils.hub_utils import _get_model_file
+from ..utils.torch_utils import device_synchronize, empty_device_cache
 
 
 if is_transformers_available():
@@ -1689,6 +1690,10 @@ def create_diffusers_clip_model_from_ldm(
 
     if is_accelerate_available():
         load_model_dict_into_meta(model, diffusers_format_checkpoint, dtype=torch_dtype)
+        # Ensure tensors are correctly placed on device by synchronizing before returning control to user. This is
+        # required because we move tensors with non_blocking=True, which is slightly faster for model loading.
+        empty_device_cache()
+        device_synchronize()
     else:
         model.load_state_dict(diffusers_format_checkpoint, strict=False)
 
@@ -2148,6 +2153,10 @@ def create_diffusers_t5_model_from_checkpoint(
 
     if is_accelerate_available():
         load_model_dict_into_meta(model, diffusers_format_checkpoint, dtype=torch_dtype)
+        # Ensure tensors are correctly placed on device by synchronizing before returning control to user. This is
+        # required because we move tensors with non_blocking=True, which is slightly faster for model loading.
+        empty_device_cache()
+        device_synchronize()
     else:
         model.load_state_dict(diffusers_format_checkpoint)
 

src/diffusers/loaders/transformer_flux.py

@@ -18,11 +18,8 @@
     MultiIPAdapterImageProjection,
 )
 from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, load_model_dict_into_meta
-from ..utils import (
-    is_accelerate_available,
-    is_torch_version,
-    logging,
-)
+from ..utils import is_accelerate_available, is_torch_version, logging
+from ..utils.torch_utils import device_synchronize, empty_device_cache
 
 
 if is_accelerate_available():
@@ -84,6 +81,8 @@ def _convert_ip_adapter_image_proj_to_diffusers(self, state_dict, low_cpu_mem_us
         else:
             device_map = {"": self.device}
             load_model_dict_into_meta(image_projection, updated_state_dict, device_map=device_map, dtype=self.dtype)
+            empty_device_cache()
+            device_synchronize()
 
         return image_projection
 
@@ -158,6 +157,9 @@ def _convert_ip_adapter_attn_to_diffusers(self, state_dicts, low_cpu_mem_usage=_
 
                 key_id += 1
 
+        empty_device_cache()
+        device_synchronize()
+
         return attn_procs
 
     def _load_ip_adapter_weights(self, state_dicts, low_cpu_mem_usage=_LOW_CPU_MEM_USAGE_DEFAULT):

src/diffusers/loaders/transformer_sd3.py

@@ -18,6 +18,7 @@
 from ..models.embeddings import IPAdapterTimeImageProjection
 from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, load_model_dict_into_meta
 from ..utils import is_accelerate_available, is_torch_version, logging
+from ..utils.torch_utils import device_synchronize, empty_device_cache
 
 
 logger = logging.get_logger(__name__)
@@ -80,6 +81,9 @@ def _convert_ip_adapter_attn_to_diffusers(
                     attn_procs[name], layer_state_dict[idx], device_map=device_map, dtype=self.dtype
                 )
 
+        empty_device_cache()
+        device_synchronize()
+
         return attn_procs
 
     def _convert_ip_adapter_image_proj_to_diffusers(
@@ -147,6 +151,8 @@ def _convert_ip_adapter_image_proj_to_diffusers(
         else:
             device_map = {"": self.device}
             load_model_dict_into_meta(image_proj, updated_state_dict, device_map=device_map, dtype=self.dtype)
+            empty_device_cache()
+            device_synchronize()
 
         return image_proj
 

src/diffusers/loaders/unet.py

@@ -43,6 +43,7 @@
     is_torch_version,
     logging,
 )
+from ..utils.torch_utils import device_synchronize, empty_device_cache
 from .lora_base import _func_optionally_disable_offloading
 from .lora_pipeline import LORA_WEIGHT_NAME, LORA_WEIGHT_NAME_SAFE, TEXT_ENCODER_NAME, UNET_NAME
 from .utils import AttnProcsLayers
@@ -753,6 +754,8 @@ def _convert_ip_adapter_image_proj_to_diffusers(self, state_dict, low_cpu_mem_us
         else:
             device_map = {"": self.device}
             load_model_dict_into_meta(image_projection, updated_state_dict, device_map=device_map, dtype=self.dtype)
+            empty_device_cache()
+            device_synchronize()
 
         return image_projection
 
@@ -850,6 +853,9 @@ def _convert_ip_adapter_attn_to_diffusers(self, state_dicts, low_cpu_mem_usage=_
 
                 key_id += 2
 
+        empty_device_cache()
+        device_synchronize()
+
         return attn_procs
 
     def _load_ip_adapter_weights(self, state_dicts, low_cpu_mem_usage=_LOW_CPU_MEM_USAGE_DEFAULT):

src/diffusers/models/model_loading_utils.py

@@ -16,9 +16,10 @@
 
 import importlib
 import inspect
+import math
 import os
 from array import array
-from collections import OrderedDict
+from collections import OrderedDict, defaultdict
 from pathlib import Path
 from typing import Dict, List, Optional, Union
 from zipfile import is_zipfile
@@ -38,6 +39,7 @@
     _get_model_file,
     deprecate,
     is_accelerate_available,
+    is_accelerate_version,
     is_gguf_available,
     is_torch_available,
     is_torch_version,
@@ -252,6 +254,10 @@ def load_model_dict_into_meta(
                 param = param.to(dtype)
                 set_module_kwargs["dtype"] = dtype
 
+        if is_accelerate_version(">", "1.8.1"):
+            set_module_kwargs["non_blocking"] = True
+            set_module_kwargs["clear_cache"] = False
+
         # For compatibility with PyTorch load_state_dict which converts state dict dtype to existing dtype in model, and which
         # uses `param.copy_(input_param)` that preserves the contiguity of the parameter in the model.
         # Reference: https://github.com/pytorch/pytorch/blob/db79ceb110f6646523019a59bbd7b838f43d4a86/torch/nn/modules/module.py#L2040C29-L2040C29
@@ -520,3 +526,60 @@ def load_gguf_checkpoint(gguf_checkpoint_path, return_tensors=False):
         parsed_parameters[name] = GGUFParameter(weights, quant_type=quant_type) if is_gguf_quant else weights
 
     return parsed_parameters
+
+
+def _find_mismatched_keys(state_dict, model_state_dict, loaded_keys, ignore_mismatched_sizes):
+    mismatched_keys = []
+    if not ignore_mismatched_sizes:
+        return mismatched_keys
+    for checkpoint_key in loaded_keys:
+        model_key = checkpoint_key
+        # If the checkpoint is sharded, we may not have the key here.
+        if checkpoint_key not in state_dict:
+            continue
+
+        if model_key in model_state_dict and state_dict[checkpoint_key].shape != model_state_dict[model_key].shape:
+            mismatched_keys.append(
+                (checkpoint_key, state_dict[checkpoint_key].shape, model_state_dict[model_key].shape)
+            )
+            del state_dict[checkpoint_key]
+    return mismatched_keys
+
+
+def _expand_device_map(device_map, param_names):
+    """
+    Expand a device map to return the correspondence parameter name to device.
+    """
+    new_device_map = {}
+    for module, device in device_map.items():
+        new_device_map.update(
+            {p: device for p in param_names if p == module or p.startswith(f"{module}.") or module == ""}
+        )
+    return new_device_map
+
+
+# Adapted from: https://github.com/huggingface/transformers/blob/0687d481e2c71544501ef9cb3eef795a6e79b1de/src/transformers/modeling_utils.py#L5859
+def _caching_allocator_warmup(model, expanded_device_map: Dict[str, torch.device], dtype: torch.dtype) -> None:
+    """
+    This function warm-ups the caching allocator based on the size of the model tensors that will reside on each
+    device. It allows to have one large call to Malloc, instead of recursively calling it later when loading the model,
+    which is actually the loading speed bottleneck. Calling this function allows to cut the model loading time by a
+    very large margin.
+    """
+    # Remove disk and cpu devices, and cast to proper torch.device
+    accelerator_device_map = {
+        param: torch.device(device)
+        for param, device in expanded_device_map.items()
+        if str(device) not in ["cpu", "disk"]
+    }
+    parameter_count = defaultdict(lambda: 0)
+    for param_name, device in accelerator_device_map.items():
+        try:
+            param = model.get_parameter(param_name)
+        except AttributeError:
+            param = model.get_buffer(param_name)
+        parameter_count[device] += math.prod(param.shape)
+
+    # This will kick off the caching allocator to avoid having to Malloc afterwards
+    for device, param_count in parameter_count.items():
+        _ = torch.empty(param_count, dtype=dtype, device=device, requires_grad=False)