Merge branch 'main' into gguf-compile-offload

Author: sayakpaul

docs/source/en/_toctree.yml

@@ -180,6 +180,8 @@
     title: Caching
   - local: optimization/memory
     title: Reduce memory usage
+  - local: optimization/speed-memory-optims
+    title: Compile and offloading quantized models
   - local: optimization/pruna
     title: Pruna
   - local: optimization/xformers

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

@@ -27,9 +27,36 @@ Chroma can use all the same optimizations as Flux.
 
 </Tip>
 
-## Inference (Single File)
+## Inference
 
-The `ChromaTransformer2DModel` supports loading checkpoints in the original format. This is also useful when trying to load finetunes or quantized versions of the models that have been published by the community.
+The Diffusers version of Chroma is based on the [`unlocked-v37`](https://huggingface.co/lodestones/Chroma/blob/main/chroma-unlocked-v37.safetensors) version of the original model, which is available in the [Chroma repository](https://huggingface.co/lodestones/Chroma).
+
+```python
+import torch
+from diffusers import ChromaPipeline
+
+pipe = ChromaPipeline.from_pretrained("lodestones/Chroma", torch_dtype=torch.bfloat16)
+pipe.enabe_model_cpu_offload()
+
+prompt = [
+    "A high-fashion close-up portrait of a blonde woman in clear sunglasses. The image uses a bold teal and red color split for dramatic lighting. The background is a simple teal-green. The photo is sharp and well-composed, and is designed for viewing with anaglyph 3D glasses for optimal effect. It looks professionally done."
+]
+negative_prompt =  ["low quality, ugly, unfinished, out of focus, deformed, disfigure, blurry, smudged, restricted palette, flat colors"]
+
+image = pipe(
+    prompt=prompt,
+    negative_prompt=negative_prompt,
+    generator=torch.Generator("cpu").manual_seed(433),
+    num_inference_steps=40,
+    guidance_scale=3.0,
+    num_images_per_prompt=1,
+).images[0]
+image.save("chroma.png")
+```
+
+## Loading from a single file
+
+To use updated model checkpoints that are not in the Diffusers format, you can use the `ChromaTransformer2DModel` class to load the model from a single file in the original format. This is also useful when trying to load finetunes or quantized versions of the models that have been published by the community.
 
 The following example demonstrates how to run Chroma from a single file.
 
@@ -38,34 +65,39 @@ Then run the following example
 ```python
 import torch
 from diffusers import ChromaTransformer2DModel, ChromaPipeline
-from transformers import T5EncoderModel
 
-bfl_repo = "black-forest-labs/FLUX.1-dev"
+model_id = "lodestones/Chroma"
 dtype = torch.bfloat16
 
-transformer = ChromaTransformer2DModel.from_single_file("https://huggingface.co/lodestones/Chroma/blob/main/chroma-unlocked-v35.safetensors", torch_dtype=dtype)
-
-text_encoder = T5EncoderModel.from_pretrained(bfl_repo, subfolder="text_encoder_2", torch_dtype=dtype)
-tokenizer = T5Tokenizer.from_pretrained(bfl_repo, subfolder="tokenizer_2", torch_dtype=dtype)
-
-pipe = ChromaPipeline.from_pretrained(bfl_repo, transformer=transformer, text_encoder=text_encoder, tokenizer=tokenizer, torch_dtype=dtype)
+transformer = ChromaTransformer2DModel.from_single_file("https://huggingface.co/lodestones/Chroma/blob/main/chroma-unlocked-v37.safetensors", torch_dtype=dtype)
 
+pipe = ChromaPipeline.from_pretrained(model_id, transformer=transformer, torch_dtype=dtype)
 pipe.enable_model_cpu_offload()
 
-prompt = "A cat holding a sign that says hello world"
+prompt = [
+    "A high-fashion close-up portrait of a blonde woman in clear sunglasses. The image uses a bold teal and red color split for dramatic lighting. The background is a simple teal-green. The photo is sharp and well-composed, and is designed for viewing with anaglyph 3D glasses for optimal effect. It looks professionally done."
+]
+negative_prompt =  ["low quality, ugly, unfinished, out of focus, deformed, disfigure, blurry, smudged, restricted palette, flat colors"]
+
 image = pipe(
-    prompt,
-    guidance_scale=4.0,
-    output_type="pil",
-    num_inference_steps=26,
-    generator=torch.Generator("cpu").manual_seed(0)
+    prompt=prompt,
+    negative_prompt=negative_prompt,
+    generator=torch.Generator("cpu").manual_seed(433),
+    num_inference_steps=40,
+    guidance_scale=3.0,
 ).images[0]
 
-image.save("image.png")
+image.save("chroma-single-file.png")
 ```
 
 ## ChromaPipeline
 
 [[autodoc]] ChromaPipeline
 	- all
 	- __call__
+
+## ChromaImg2ImgPipeline
+
+[[autodoc]] ChromaImg2ImgPipeline
+	- all
+	- __call__

docs/source/en/optimization/memory.md

@@ -17,7 +17,7 @@ Modern diffusion models like [Flux](../api/pipelines/flux) and [Wan](../api/pipe
 This guide will show you how to reduce your memory usage. 
 
 > [!TIP]
-> Keep in mind these techniques may need to be adjusted depending on the model! For example, a transformer-based diffusion model may not benefit equally from these inference speed optimizations as a UNet-based model.
+> Keep in mind these techniques may need to be adjusted depending on the model. For example, a transformer-based diffusion model may not benefit equally from these memory optimizations as a UNet-based model.
 
 ## Multiple GPUs
 
@@ -63,7 +63,12 @@ pipeline = StableDiffusionXLPipeline.from_pretrained(
 > [!WARNING]
 > Device placement is an experimental feature and the API may change. Only the `balanced` strategy is supported at the moment. We plan to support additional mapping strategies in the future.
 
-The `device_map` parameter controls how the model components in a pipeline are distributed across devices. The `balanced` device placement strategy evenly splits the pipeline across all available devices.
+The `device_map` parameter controls how the model components in a pipeline or the layers in an individual model are distributed across devices. 
+
+<hfoptions id="device-map">
+<hfoption id="pipeline level">
+
+The `balanced` device placement strategy evenly splits the pipeline across all available devices.
 
 ```py
 import torch
@@ -83,7 +88,10 @@ print(pipeline.hf_device_map)
 {'unet': 1, 'vae': 1, 'safety_checker': 0, 'text_encoder': 0}
 ```
 
-The `device_map` parameter also works on the model-level. This is useful for loading large models, such as the Flux diffusion transformer which has 12.5B parameters. Instead of `balanced`, set it to `"auto"` to automatically distribute a model across the fastest device first before moving to slower devices. Refer to the [Model sharding](../training/distributed_inference#model-sharding) docs for more details.
+</hfoption>
+<hfoption id="model level">
+
+The `device_map` is useful for loading large models, such as the Flux diffusion transformer which has 12.5B parameters. Set it to `"auto"` to automatically distribute a model across the fastest device first before moving to slower devices. Refer to the [Model sharding](../training/distributed_inference#model-sharding) docs for more details.
 
 ```py
 import torch
@@ -97,7 +105,43 @@ transformer = AutoModel.from_pretrained(
 )
 ```
 
-For more fine-grained control, pass a dictionary to enforce the maximum GPU memory to use on each device. If a device is not in `max_memory`, it is ignored and pipeline components won't be distributed to it.
+You can inspect a model's device map with `hf_device_map`.
+
+```py
+print(transformer.hf_device_map)
+```
+
+</hfoption>
+</hfoptions>
+
+When designing your own `device_map`, it should be a dictionary of a model's specific module name or layer and a device identifier (an integer for GPUs, `cpu` for CPUs, and `disk` for disk).
+
+Call `hf_device_map` on a model to see how model layers are distributed and then design your own.
+
+```py
+print(transformer.hf_device_map)
+{'pos_embed': 0, 'time_text_embed': 0, 'context_embedder': 0, 'x_embedder': 0, 'transformer_blocks': 0, 'single_transformer_blocks.0': 0, 'single_transformer_blocks.1': 0, 'single_transformer_blocks.2': 0, 'single_transformer_blocks.3': 0, 'single_transformer_blocks.4': 0, 'single_transformer_blocks.5': 0, 'single_transformer_blocks.6': 0, 'single_transformer_blocks.7': 0, 'single_transformer_blocks.8': 0, 'single_transformer_blocks.9': 0, 'single_transformer_blocks.10': 'cpu', 'single_transformer_blocks.11': 'cpu', 'single_transformer_blocks.12': 'cpu', 'single_transformer_blocks.13': 'cpu', 'single_transformer_blocks.14': 'cpu', 'single_transformer_blocks.15': 'cpu', 'single_transformer_blocks.16': 'cpu', 'single_transformer_blocks.17': 'cpu', 'single_transformer_blocks.18': 'cpu', 'single_transformer_blocks.19': 'cpu', 'single_transformer_blocks.20': 'cpu', 'single_transformer_blocks.21': 'cpu', 'single_transformer_blocks.22': 'cpu', 'single_transformer_blocks.23': 'cpu', 'single_transformer_blocks.24': 'cpu', 'single_transformer_blocks.25': 'cpu', 'single_transformer_blocks.26': 'cpu', 'single_transformer_blocks.27': 'cpu', 'single_transformer_blocks.28': 'cpu', 'single_transformer_blocks.29': 'cpu', 'single_transformer_blocks.30': 'cpu', 'single_transformer_blocks.31': 'cpu', 'single_transformer_blocks.32': 'cpu', 'single_transformer_blocks.33': 'cpu', 'single_transformer_blocks.34': 'cpu', 'single_transformer_blocks.35': 'cpu', 'single_transformer_blocks.36': 'cpu', 'single_transformer_blocks.37': 'cpu', 'norm_out': 'cpu', 'proj_out': 'cpu'}
+```
+
+For example, the `device_map` below places `single_transformer_blocks.10` through `single_transformer_blocks.20` on a second GPU (`1`).
+
+```py
+import torch
+from diffusers import AutoModel
+
+device_map = {
+    'pos_embed': 0, 'time_text_embed': 0, 'context_embedder': 0, 'x_embedder': 0, 'transformer_blocks': 0, 'single_transformer_blocks.0': 0, 'single_transformer_blocks.1': 0, 'single_transformer_blocks.2': 0, 'single_transformer_blocks.3': 0, 'single_transformer_blocks.4': 0, 'single_transformer_blocks.5': 0, 'single_transformer_blocks.6': 0, 'single_transformer_blocks.7': 0, 'single_transformer_blocks.8': 0, 'single_transformer_blocks.9': 0, 'single_transformer_blocks.10': 1, 'single_transformer_blocks.11': 1, 'single_transformer_blocks.12': 1, 'single_transformer_blocks.13': 1, 'single_transformer_blocks.14': 1, 'single_transformer_blocks.15': 1, 'single_transformer_blocks.16': 1, 'single_transformer_blocks.17': 1, 'single_transformer_blocks.18': 1, 'single_transformer_blocks.19': 1, 'single_transformer_blocks.20': 1, 'single_transformer_blocks.21': 'cpu', 'single_transformer_blocks.22': 'cpu', 'single_transformer_blocks.23': 'cpu', 'single_transformer_blocks.24': 'cpu', 'single_transformer_blocks.25': 'cpu', 'single_transformer_blocks.26': 'cpu', 'single_transformer_blocks.27': 'cpu', 'single_transformer_blocks.28': 'cpu', 'single_transformer_blocks.29': 'cpu', 'single_transformer_blocks.30': 'cpu', 'single_transformer_blocks.31': 'cpu', 'single_transformer_blocks.32': 'cpu', 'single_transformer_blocks.33': 'cpu', 'single_transformer_blocks.34': 'cpu', 'single_transformer_blocks.35': 'cpu', 'single_transformer_blocks.36': 'cpu', 'single_transformer_blocks.37': 'cpu', 'norm_out': 'cpu', 'proj_out': 'cpu'
+}
+
+transformer = AutoModel.from_pretrained(
+    "black-forest-labs/FLUX.1-dev", 
+    subfolder="transformer",
+    device_map=device_map,
+    torch_dtype=torch.bfloat16
+)
+```
+
+Pass a dictionary mapping maximum memory usage to each device to enforce a limit. If a device is not in `max_memory`, it is ignored and pipeline components won't be distributed to it.
 
 ```py
 import torch
@@ -145,7 +189,7 @@ print(f"Max memory reserved: {torch.cuda.max_memory_allocated() / 1024**3:.2f} G
 ```
 
 > [!WARNING]
-> [`AutoencoderKLWan`] and [`AsymmetricAutoencoderKL`] don't support slicing.
+> The [`AutoencoderKLWan`] and [`AsymmetricAutoencoderKL`] classes don't support slicing.
 
 ## VAE tiling
 
@@ -172,7 +216,13 @@ print(f"Max memory reserved: {torch.cuda.max_memory_allocated() / 1024**3:.2f} G
 > [!WARNING]
 > [`AutoencoderKLWan`] and [`AsymmetricAutoencoderKL`] don't support tiling.
 
-## CPU offloading
+## Offloading
+
+Offloading strategies move not currently active layers or models to the CPU to avoid increasing GPU memory. These strategies can be combined with quantization and torch.compile to balance inference speed and memory usage.
+
+Refer to the [Compile and offloading quantized models](./speed-memory-optims) guide for more details.
+
+### CPU offloading
 
 CPU offloading selectively moves weights from the GPU to the CPU. When a component is required, it is transferred to the GPU and when it isn't required, it is moved to the CPU. This method works on submodules rather than whole models. It saves memory by avoiding storing the entire model on the GPU.
 
@@ -203,7 +253,7 @@ pipeline(
 print(f"Max memory reserved: {torch.cuda.max_memory_allocated() / 1024**3:.2f} GB")
 ```
 
-## Model offloading
+### Model offloading
 
 Model offloading moves entire models to the GPU instead of selectively moving *some* layers or model components. One of the main pipeline models, usually the text encoder, UNet, and VAE, is placed on the GPU while the other components are held on the CPU. Components like the UNet that run multiple times stays on the GPU until its completely finished and no longer needed. This eliminates the communication overhead of [CPU offloading](#cpu-offloading) and makes model offloading a faster alternative. The tradeoff is memory savings won't be as large.
 
@@ -219,7 +269,7 @@ from diffusers import DiffusionPipeline
 pipeline = DiffusionPipeline.from_pretrained(
     "black-forest-labs/FLUX.1-schnell", torch_dtype=torch.bfloat16
 )
-pipline.enable_model_cpu_offload()
+pipeline.enable_model_cpu_offload()
 
 pipeline(
     prompt="An astronaut riding a horse on Mars",
@@ -234,7 +284,7 @@ print(f"Max memory reserved: {torch.cuda.max_memory_allocated() / 1024**3:.2f} G
 
 [`~DiffusionPipeline.enable_model_cpu_offload`] also helps when you're using the [`~StableDiffusionXLPipeline.encode_prompt`] method on its own to generate the text encoders hidden state.
 
-## Group offloading
+### Group offloading
 
 Group offloading moves groups of internal layers ([torch.nn.ModuleList](https://pytorch.org/docs/stable/generated/torch.nn.ModuleList.html) or [torch.nn.Sequential](https://pytorch.org/docs/stable/generated/torch.nn.Sequential.html)) to the CPU. It uses less memory than [model offloading](#model-offloading) and it is faster than [CPU offloading](#cpu-offloading) because it reduces communication overhead.
 
@@ -278,7 +328,7 @@ print(f"Max memory reserved: {torch.cuda.max_memory_allocated() / 1024**3:.2f} G
 export_to_video(video, "output.mp4", fps=8)
 ```
 
-### CUDA stream
+#### CUDA stream
 
 The `use_stream` parameter can be activated for CUDA devices that support asynchronous data transfer streams to reduce overall execution time compared to [CPU offloading](#cpu-offloading). It overlaps data transfer and computation by using layer prefetching. The next layer to be executed is loaded onto the GPU while the current layer is still being executed. It can increase CPU memory significantly so ensure you have 2x the amount of memory as the model size.
 
@@ -295,15 +345,25 @@ pipeline.transformer.enable_group_offload(onload_device=onload_device, offload_d
 
 The `low_cpu_mem_usage` parameter can be set to `True` to reduce CPU memory usage when using streams during group offloading. It is best for `leaf_level` offloading and when CPU memory is bottlenecked. Memory is saved by creating pinned tensors on the fly instead of pre-pinning them. However, this may increase overall execution time.
 
-<Tip>
+#### Offloading to disk
 
-The offloading strategies can be combined with [quantization](../quantization/overview.md) to enable further memory savings. For image generation, combining [quantization and model offloading](#model-offloading) can often give the best trade-off between quality, speed, and memory. However, for video generation, as the models are more
-compute-bound, [group-offloading](#group-offloading) tends to be better. Group offloading provides considerable benefits when weight transfers can be overlapped with computation (must use streams). When applying group offloading with quantization on image generation models at typical resolutions (1024x1024, for example), it is usually not possible to *fully* overlap weight transfers if the compute kernel finishes faster, making it communication bound between CPU/GPU (due to device synchronizations).
+Group offloading can consume significant system memory depending on the model size. On systems with limited memory, try group offloading onto the disk as a secondary memory.
 
-</Tip>
+Set the `offload_to_disk_path` argument in either [`~ModelMixin.enable_group_offload`] or [`~hooks.apply_group_offloading`] to offload the model to the disk.
+
+```py
+pipeline.transformer.enable_group_offload(onload_device=onload_device, offload_device=offload_device, offload_type="leaf_level", offload_to_disk_path="path/to/disk")
+
+apply_group_offloading(pipeline.text_encoder, onload_device=onload_device, offload_type="block_level", num_blocks_per_group=2, offload_to_disk_path="path/to/disk")
+```
+
+Refer to these [two](https://github.com/huggingface/diffusers/pull/11682#issue-3129365363) [tables](https://github.com/huggingface/diffusers/pull/11682#issuecomment-2955715126) to compare the speed and memory trade-offs.
 
 ## Layerwise casting
 
+> [!TIP]
+> Combine layerwise casting with [group offloading](#group-offloading) for even more memory savings.
+
 Layerwise casting stores weights in a smaller data format (for example, `torch.float8_e4m3fn` and `torch.float8_e5m2`) to use less memory and upcasts those weights to a higher precision like `torch.float16` or `torch.bfloat16` for computation. Certain layers (normalization and modulation related weights) are skipped because storing them in fp8 can degrade generation quality.
 
 > [!WARNING]
@@ -493,7 +553,7 @@ with torch.inference_mode():
 ## Memory-efficient attention
 
 > [!TIP]
-> Memory-efficient attention optimizes for memory usage *and* [inference speed](./fp16#scaled-dot-product-attention!
+> 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.