Merge branch 'main' into fix-cogvideox-pos-embedding-device

Author: a-r-r-o-w

.github/workflows/nightly_tests.yml

@@ -357,6 +357,8 @@ jobs:
         config:
           - backend: "bitsandbytes"
             test_location: "bnb"
+          - backend: "gguf"
+            test_location: "gguf"
     runs-on:
       group: aws-g6e-xlarge-plus
     container:

docs/source/en/_toctree.yml

@@ -157,6 +157,8 @@
     title: Getting Started
   - local: quantization/bitsandbytes
     title: bitsandbytes
+  - local: quantization/gguf
+    title: gguf
   - local: quantization/torchao
     title: torchao
   title: Quantization Methods

docs/source/en/api/quantization.md

@@ -28,6 +28,9 @@ Learn how to quantize models in the [Quantization](../quantization/overview) gui
 
 [[autodoc]] BitsAndBytesConfig
 
+## GGUFQuantizationConfig
+
+[[autodoc]] GGUFQuantizationConfig
 ## TorchAoConfig
 
 [[autodoc]] TorchAoConfig

docs/source/en/quantization/gguf.md

@@ -0,0 +1,70 @@
+<!--Copyright 2024 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.
+
+-->
+
+# GGUF
+
+The GGUF file format is typically used to store models for inference with [GGML](https://github.com/ggerganov/ggml) and supports a variety of block wise quantization options. Diffusers supports loading checkpoints prequantized and saved in the GGUF format via `from_single_file` loading with Model classes. Loading GGUF checkpoints via Pipelines is currently not supported.
+
+The following example will load the [FLUX.1 DEV](https://huggingface.co/black-forest-labs/FLUX.1-dev) transformer model using the GGUF Q2_K quantization variant.
+
+Before starting please install gguf in your environment
+
+```shell
+pip install -U gguf
+```
+
+Since GGUF is a single file format, use [`~FromSingleFileMixin.from_single_file`] to load the model and pass in the [`GGUFQuantizationConfig`].
+
+When using GGUF checkpoints, the quantized weights remain in a low memory `dtype`(typically `torch.unint8`) and are dynamically dequantized and cast to the configured `compute_dtype` during each module's forward pass through the model. The `GGUFQuantizationConfig` allows you to set the `compute_dtype`. 
+
+The functions used for dynamic dequantizatation are based on the great work done by [city96](https://github.com/city96/ComfyUI-GGUF), who created the Pytorch ports of the original (`numpy`)[https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/gguf/quants.py] implementation by [compilade](https://github.com/compilade).
+
+```python
+import torch
+
+from diffusers import FluxPipeline, FluxTransformer2DModel, GGUFQuantizationConfig
+
+ckpt_path = (
+    "https://huggingface.co/city96/FLUX.1-dev-gguf/blob/main/flux1-dev-Q2_K.gguf"
+)
+transformer = FluxTransformer2DModel.from_single_file(
+    ckpt_path,
+    quantization_config=GGUFQuantizationConfig(compute_dtype=torch.bfloat16),
+    torch_dtype=torch.bfloat16,
+)
+pipe = FluxPipeline.from_pretrained(
+    "black-forest-labs/FLUX.1-dev",
+    transformer=transformer,
+    generator=torch.manual_seed(0),
+    torch_dtype=torch.bfloat16,
+)
+pipe.enable_model_cpu_offload()
+prompt = "A cat holding a sign that says hello world"
+image = pipe(prompt).images[0]
+image.save("flux-gguf.png")
+```
+
+## Supported Quantization Types
+
+- BF16
+- Q4_0
+- Q4_1
+- Q5_0
+- Q5_1
+- Q8_0
+- Q2_K
+- Q3_K
+- Q4_K
+- Q5_K
+- Q6_K
+

docs/source/en/quantization/overview.md

@@ -17,7 +17,7 @@ Quantization techniques focus on representing data with less information while a
 
 <Tip>
 
-Interested in adding a new quantization method to Transformers? Refer to the [Contribute new quantization method guide](https://huggingface.co/docs/transformers/main/en/quantization/contribute) to learn more about adding a new quantization method.
+Interested in adding a new quantization method to Diffusers? Refer to the [Contribute new quantization method guide](https://huggingface.co/docs/transformers/main/en/quantization/contribute) to learn more about adding a new quantization method.
 
 </Tip>
 
@@ -32,4 +32,9 @@ If you are new to the quantization field, we recommend you to check out these be
 
 ## When to use what?
 
-Diffusers supports [bitsandbytes](https://huggingface.co/docs/bitsandbytes/main/en/index) and [torchao](https://github.com/pytorch/ao). Refer to this [table](https://huggingface.co/docs/transformers/main/en/quantization/overview#when-to-use-what) to help you determine which quantization backend to use.
+Diffusers currently supports the following quantization methods.
+- [BitsandBytes]()
+- [TorchAO]()
+- [GGUF]()
+
+[This resource](https://huggingface.co/docs/transformers/main/en/quantization/overview#when-to-use-what) provides a good overview of the pros and cons of different quantization techniques.

docs/source/en/tutorials/using_peft_for_inference.md

@@ -56,7 +56,7 @@ image
 
 With the `adapter_name` parameter, it is really easy to use another adapter for inference! Load the [nerijs/pixel-art-xl](https://huggingface.co/nerijs/pixel-art-xl) adapter that has been fine-tuned to generate pixel art images and call it `"pixel"`.
 
-The pipeline automatically sets the first loaded adapter (`"toy"`) as the active adapter, but you can activate the `"pixel"` adapter with the [`~diffusers.loaders.UNet2DConditionLoadersMixin.set_adapters`] method:
+The pipeline automatically sets the first loaded adapter (`"toy"`) as the active adapter, but you can activate the `"pixel"` adapter with the [`~PeftAdapterMixin.set_adapters`] method:
 
 ```python
 pipe.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
@@ -85,7 +85,7 @@ By default, if the most up-to-date versions of PEFT and Transformers are detecte
 
 You can also merge different adapter checkpoints for inference to blend their styles together.
 
-Once again, use the [`~diffusers.loaders.UNet2DConditionLoadersMixin.set_adapters`] method to activate the `pixel` and `toy` adapters and specify the weights for how they should be merged.
+Once again, use the [`~PeftAdapterMixin.set_adapters`] method to activate the `pixel` and `toy` adapters and specify the weights for how they should be merged.
 
 ```python
 pipe.set_adapters(["pixel", "toy"], adapter_weights=[0.5, 1.0])
@@ -114,7 +114,7 @@ Impressive! As you can see, the model generated an image that mixed the characte
 > [!TIP]
 > Through its PEFT integration, Diffusers also offers more efficient merging methods which you can learn about in the [Merge LoRAs](../using-diffusers/merge_loras) guide!
 
-To return to only using one adapter, use the [`~diffusers.loaders.UNet2DConditionLoadersMixin.set_adapters`] method to activate the `"toy"` adapter:
+To return to only using one adapter, use the [`~PeftAdapterMixin.set_adapters`] method to activate the `"toy"` adapter:
 
 ```python
 pipe.set_adapters("toy")
@@ -127,7 +127,7 @@ image = pipe(
 image
 ```
 
-Or to disable all adapters entirely, use the [`~diffusers.loaders.UNet2DConditionLoadersMixin.disable_lora`] method to return the base model.
+Or to disable all adapters entirely, use the [`~PeftAdapterMixin.disable_lora`] method to return the base model.
 
 ```python
 pipe.disable_lora()
@@ -140,7 +140,8 @@ image
 ![no-lora](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/peft_integration/diffusers_peft_lora_inference_20_1.png)
 
 ### Customize adapters strength
-For even more customization, you can control how strongly the adapter affects each part of the pipeline. For this, pass a dictionary with the control strengths (called "scales") to [`~diffusers.loaders.UNet2DConditionLoadersMixin.set_adapters`].
+
+For even more customization, you can control how strongly the adapter affects each part of the pipeline. For this, pass a dictionary with the control strengths (called "scales") to [`~PeftAdapterMixin.set_adapters`].
 
 For example, here's how you can turn on the adapter for the `down` parts, but turn it off for the `mid` and `up` parts:
 ```python
@@ -195,7 +196,7 @@ image
 
 ![block-lora-mixed](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/peft_integration/diffusers_peft_lora_inference_block_mixed.png)
 
-## Manage active adapters
+## Manage adapters
 
 You have attached multiple adapters in this tutorial, and if you're feeling a bit lost on what adapters have been attached to the pipeline's components, use the [`~diffusers.loaders.StableDiffusionLoraLoaderMixin.get_active_adapters`] method to check the list of active adapters:
 
@@ -212,3 +213,11 @@ list_adapters_component_wise = pipe.get_list_adapters()
 list_adapters_component_wise
 {"text_encoder": ["toy", "pixel"], "unet": ["toy", "pixel"], "text_encoder_2": ["toy", "pixel"]}
 ```
+
+The [`~PeftAdapterMixin.delete_adapters`] function completely removes an adapter and their LoRA layers from a model.
+
+```py
+pipe.delete_adapters("toy")
+pipe.get_active_adapters()
+["pixel"]
+```

src/diffusers/__init__.py

@@ -31,7 +31,7 @@
     "loaders": ["FromOriginalModelMixin"],
     "models": [],
     "pipelines": [],
-    "quantizers.quantization_config": ["BitsAndBytesConfig", "TorchAoConfig"],
+    "quantizers.quantization_config": ["BitsAndBytesConfig", "GGUFQuantizationConfig", "TorchAoConfig"],
     "schedulers": [],
     "utils": [
         "OptionalDependencyNotAvailable",
@@ -569,7 +569,7 @@
 
 if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     from .configuration_utils import ConfigMixin
-    from .quantizers.quantization_config import BitsAndBytesConfig, TorchAoConfig
+    from .quantizers.quantization_config import BitsAndBytesConfig, GGUFQuantizationConfig, TorchAoConfig
 
     try:
         if not is_onnx_available():

src/diffusers/loaders/single_file_model.py

@@ -17,8 +17,10 @@
 from contextlib import nullcontext
 from typing import Optional
 
+import torch
 from huggingface_hub.utils import validate_hf_hub_args
 
+from ..quantizers import DiffusersAutoQuantizer
 from ..utils import deprecate, is_accelerate_available, logging
 from .single_file_utils import (
     SingleFileComponentError,
@@ -214,6 +216,8 @@ def from_single_file(cls, pretrained_model_link_or_path_or_dict: Optional[str] =
         subfolder = kwargs.pop("subfolder", None)
         revision = kwargs.pop("revision", None)
         torch_dtype = kwargs.pop("torch_dtype", None)
+        quantization_config = kwargs.pop("quantization_config", None)
+        device = kwargs.pop("device", None)
 
         if isinstance(pretrained_model_link_or_path_or_dict, dict):
             checkpoint = pretrained_model_link_or_path_or_dict
@@ -227,6 +231,12 @@ def from_single_file(cls, pretrained_model_link_or_path_or_dict: Optional[str] =
                 local_files_only=local_files_only,
                 revision=revision,
             )
+        if quantization_config is not None:
+            hf_quantizer = DiffusersAutoQuantizer.from_config(quantization_config)
+            hf_quantizer.validate_environment()
+
+        else:
+            hf_quantizer = None
 
         mapping_functions = SINGLE_FILE_LOADABLE_CLASSES[mapping_class_name]
 
@@ -309,8 +319,36 @@ def from_single_file(cls, pretrained_model_link_or_path_or_dict: Optional[str] =
         with ctx():
             model = cls.from_config(diffusers_model_config)
 
+        # Check if `_keep_in_fp32_modules` is not None
+        use_keep_in_fp32_modules = (cls._keep_in_fp32_modules is not None) and (
+            (torch_dtype == torch.float16) or hasattr(hf_quantizer, "use_keep_in_fp32_modules")
+        )
+        if use_keep_in_fp32_modules:
+            keep_in_fp32_modules = cls._keep_in_fp32_modules
+            if not isinstance(keep_in_fp32_modules, list):
+                keep_in_fp32_modules = [keep_in_fp32_modules]
+
+        else:
+            keep_in_fp32_modules = []
+
+        if hf_quantizer is not None:
+            hf_quantizer.preprocess_model(
+                model=model,
+                device_map=None,
+                state_dict=diffusers_format_checkpoint,
+                keep_in_fp32_modules=keep_in_fp32_modules,
+            )
+
         if is_accelerate_available():
-            unexpected_keys = load_model_dict_into_meta(model, diffusers_format_checkpoint, dtype=torch_dtype)
+            param_device = torch.device(device) if device else torch.device("cpu")
+            unexpected_keys = load_model_dict_into_meta(
+                model,
+                diffusers_format_checkpoint,
+                dtype=torch_dtype,
+                device=param_device,
+                hf_quantizer=hf_quantizer,
+                keep_in_fp32_modules=keep_in_fp32_modules,
+            )
 
         else:
             _, unexpected_keys = model.load_state_dict(diffusers_format_checkpoint, strict=False)
@@ -324,7 +362,11 @@ def from_single_file(cls, pretrained_model_link_or_path_or_dict: Optional[str] =
                 f"Some weights of the model checkpoint were not used when initializing {cls.__name__}: \n {[', '.join(unexpected_keys)]}"
             )
 
-        if torch_dtype is not None:
+        if hf_quantizer is not None:
+            hf_quantizer.postprocess_model(model)
+            model.hf_quantizer = hf_quantizer
+
+        if torch_dtype is not None and hf_quantizer is None:
             model.to(torch_dtype)
 
         model.eval()

src/diffusers/loaders/single_file_utils.py

@@ -81,8 +81,14 @@
     "open_clip_sd3": "text_encoders.clip_g.transformer.text_model.embeddings.position_embedding.weight",
     "stable_cascade_stage_b": "down_blocks.1.0.channelwise.0.weight",
     "stable_cascade_stage_c": "clip_txt_mapper.weight",
-    "sd3": "model.diffusion_model.joint_blocks.0.context_block.adaLN_modulation.1.bias",
-    "sd35_large": "model.diffusion_model.joint_blocks.37.x_block.mlp.fc1.weight",
+    "sd3": [
+        "joint_blocks.0.context_block.adaLN_modulation.1.bias",
+        "model.diffusion_model.joint_blocks.0.context_block.adaLN_modulation.1.bias",
+    ],
+    "sd35_large": [
+        "joint_blocks.37.x_block.mlp.fc1.weight",
+        "model.diffusion_model.joint_blocks.37.x_block.mlp.fc1.weight",
+    ],
     "animatediff": "down_blocks.0.motion_modules.0.temporal_transformer.transformer_blocks.0.attention_blocks.0.pos_encoder.pe",
     "animatediff_v2": "mid_block.motion_modules.0.temporal_transformer.norm.bias",
     "animatediff_sdxl_beta": "up_blocks.2.motion_modules.0.temporal_transformer.norm.weight",
@@ -542,13 +548,20 @@ def infer_diffusers_model_type(checkpoint):
     ):
         model_type = "stable_cascade_stage_b"
 
-    elif CHECKPOINT_KEY_NAMES["sd3"] in checkpoint and checkpoint[CHECKPOINT_KEY_NAMES["sd3"]].shape[-1] == 9216:
-        if checkpoint["model.diffusion_model.pos_embed"].shape[1] == 36864:
+    elif any(key in checkpoint for key in CHECKPOINT_KEY_NAMES["sd3"]) and any(
+        checkpoint[key].shape[-1] == 9216 if key in checkpoint else False for key in CHECKPOINT_KEY_NAMES["sd3"]
+    ):
+        if "model.diffusion_model.pos_embed" in checkpoint:
+            key = "model.diffusion_model.pos_embed"
+        else:
+            key = "pos_embed"
+
+        if checkpoint[key].shape[1] == 36864:
             model_type = "sd3"
-        elif checkpoint["model.diffusion_model.pos_embed"].shape[1] == 147456:
+        elif checkpoint[key].shape[1] == 147456:
             model_type = "sd35_medium"
 
-    elif CHECKPOINT_KEY_NAMES["sd35_large"] in checkpoint:
+    elif any(key in checkpoint for key in CHECKPOINT_KEY_NAMES["sd35_large"]):
         model_type = "sd35_large"
 
     elif CHECKPOINT_KEY_NAMES["animatediff"] in checkpoint: