feat: pipeline-level quantization config

docs/source/en/api/quantization.md

@@ -13,16 +13,17 @@ specific language governing permissions and limitations under the License.
 
 # Quantization
 
-Quantization techniques reduce memory and computational costs by representing weights and activations with lower-precision data types like 8-bit integers (int8). This enables loading larger models you normally wouldn't be able to fit into memory, and speeding up inference. Diffusers supports 8-bit and 4-bit quantization with [bitsandbytes](https://huggingface.co/docs/bitsandbytes/en/index).
-
-Quantization techniques that aren't supported in Transformers can be added with the [`DiffusersQuantizer`] class.
+Quantization techniques reduce memory and computational costs by representing weights and activations with lower-precision data types like 8-bit integers (int8). This enables loading larger models you normally wouldn't be able to fit into memory, and speeding up inference.
 
 <Tip>
 
 Learn how to quantize models in the [Quantization](../quantization/overview) guide.
 
 </Tip>
 
+## PipelineQuantizationConfig
+
+[[autodoc]] quantizers.PipelineQuantizationConfig
 
 ## BitsAndBytesConfig
 

docs/source/en/quantization/overview.md

@@ -39,3 +39,62 @@ Diffusers currently supports the following quantization methods.
 - [Quanto](./quanto.md)
 
 [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.
+
+## Pipeline-level quantization
+
+Diffusers allows users to directly initialize pipelines from checkpoints that may contain quantized models ([example](https://huggingface.co/hf-internal-testing/flux.1-dev-nf4-pkg)). However, users may want to apply
+quantization on-the-fly when initializing a pipeline from a pre-trained and non-quantized checkpoint. You can
+do this with [`~quantizers.PipelineQuantizationConfig`].
+
+Start by defining a `PipelineQuantizationConfig`:
+
+```py
+import torch
+from diffusers import DiffusionPipeline
+from diffusers.quantizers.quantization_config import QuantoConfig
+from diffusers.quantizers import PipelineQuantizationConfig
+from transformers import BitsAndBytesConfig
+
+pipeline_quant_config = PipelineQuantizationConfig(
+    quant_mapping={
+        "transformer": QuantoConfig(weights_dtype="int8"),
+        "text_encoder_2": BitsAndBytesConfig(
+            load_in_4bit=True, compute_dtype=torch.bfloat16
+        ),
+    }
+)
+```
+
+Then pass it to [`~DiffusionPipeline.from_pretrained`] and run inference:
+
+```py
+pipe = DiffusionPipeline.from_pretrained(
+    "black-forest-labs/FLUX.1-dev",
+    quantization_config=pipeline_quant_config,
+    torch_dtype=torch.bfloat16,
+).to("cuda")
+
+image = pipe("photo of a cute dog").images[0]
+```
+
+This method allows for more granular control over the quantization specifications of individual 
+model-level components of a pipeline. It also allows for different quantization backends for
+different components. In the above example, you used a combination of Quanto and BitsandBytes.
+
+The other method is simpler in terms of experience but is
+less-flexible. Start by defining a `PipelineQuantizationConfig` but in a different way:
+
+```py
+pipeline_quant_config = PipelineQuantizationConfig(
+    quant_backend="bitsandbytes_4bit",
+    quant_kwargs={"load_in_4bit": True, "bnb_4bit_quant_type": "nf4", "bnb_4bit_compute_dtype": torch.bfloat16},
+    components_to_quantize=["transformer", "text_encoder_2"],
+)
+```
+
+This `pipeline_quant_config` can now be passed to [`~DiffusionPipeline.from_pretrained`] similar to the above example.
+
+In this case, `quant_kwargs` will be used to initialize the quantization specifications
+of the respective quantization configuration class of `quant_backend`. `components_to_quantize`
+is used to denote the components that will be quantized. For most pipelines, you would want to
+keep `transformer` in the list as that is often the most compute and memory intensive.

src/diffusers/pipelines/pipeline_loading_utils.py

@@ -675,8 +675,10 @@ def load_sub_model(
     use_safetensors: bool,
     dduf_entries: Optional[Dict[str, DDUFEntry]],
     provider_options: Any,
+    quantization_config: Optional[Any] = None,
 ):
     """Helper method to load the module `name` from `library_name` and `class_name`"""
+    from ..quantizers import PipelineQuantizationConfig
 
     # retrieve class candidates
 
@@ -769,6 +771,17 @@ def load_sub_model(
         else:
             loading_kwargs["low_cpu_mem_usage"] = False
 
+    if (
+        quantization_config is not None
+        and isinstance(quantization_config, PipelineQuantizationConfig)
+        and issubclass(class_obj, torch.nn.Module)
+    ):
+        model_quant_config = quantization_config._resolve_quant_config(
+            is_diffusers=is_diffusers_model, module_name=name
+        )
+        if model_quant_config is not None:
+            loading_kwargs["quantization_config"] = model_quant_config
+
     # check if the module is in a subdirectory
     if dduf_entries:
         loading_kwargs["dduf_entries"] = dduf_entries

src/diffusers/pipelines/pipeline_utils.py

@@ -47,6 +47,7 @@
 from ..models import AutoencoderKL
 from ..models.attention_processor import FusedAttnProcessor2_0
 from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, ModelMixin
+from ..quantizers import PipelineQuantizationConfig
 from ..quantizers.bitsandbytes.utils import _check_bnb_status
 from ..schedulers.scheduling_utils import SCHEDULER_CONFIG_NAME
 from ..utils import (
@@ -725,6 +726,7 @@ def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.P
         use_safetensors = kwargs.pop("use_safetensors", None)
         use_onnx = kwargs.pop("use_onnx", None)
         load_connected_pipeline = kwargs.pop("load_connected_pipeline", False)
+        quantization_config = kwargs.pop("quantization_config", None)
 
         if torch_dtype is not None and not isinstance(torch_dtype, dict) and not isinstance(torch_dtype, torch.dtype):
             torch_dtype = torch.float32
@@ -741,6 +743,9 @@ def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.P
                 " install accelerate\n```\n."
             )
 
+        if quantization_config is not None and not isinstance(quantization_config, PipelineQuantizationConfig):
+            raise ValueError("`quantization_config` must be an instance of `PipelineQuantizationConfig`.")
+
         if low_cpu_mem_usage is True and not is_torch_version(">=", "1.9.0"):
             raise NotImplementedError(
                 "Low memory initialization requires torch >= 1.9.0. Please either update your PyTorch version or set"
@@ -1001,6 +1006,7 @@ def load_module(name, value):
                     use_safetensors=use_safetensors,
                     dduf_entries=dduf_entries,
                     provider_options=provider_options,
+                    quantization_config=quantization_config,
                 )
                 logger.info(
                     f"Loaded {name} as {class_name} from `{name}` subfolder of {pretrained_model_name_or_path}."

src/diffusers/quantizers/__init__.py

@@ -12,5 +12,183 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import inspect
+from typing import Dict, List, Optional, Union
+
+from ..utils import is_transformers_available, logging
 from .auto import DiffusersAutoQuantizer
 from .base import DiffusersQuantizer
+from .quantization_config import QuantizationConfigMixin as DiffQuantConfigMixin
+
+
+try:
+    from transformers.utils.quantization_config import QuantizationConfigMixin as TransformersQuantConfigMixin
+except ImportError:
+
+    class TransformersQuantConfigMixin:
+        pass
+
+
+logger = logging.get_logger(__name__)
+
+
+class PipelineQuantizationConfig:
+    """
+    Configuration class to be used when applying quantization on-the-fly to [`~DiffusionPipeline.from_pretrained`].
+
+    Args:
+        quant_backend (`str`): Quantization backend to be used. When using this option, we assume that the backend
+            is available to both `diffusers` and `transformers`.
+        quant_kwargs (`dict`): Params to initialize the quantization backend class.
+        components_to_quantize (`list`): Components of a pipeline to be quantized.
+        quant_mapping (`dict`): Mapping defining the quantization specs to be used for the pipeline
+            components. When using this argument, users are not expected to provide `quant_backend`, `quant_kawargs`,
+            and `components_to_quantize`.
+    """
+
+    def __init__(
+        self,
+        quant_backend: str = None,
+        quant_kwargs: Dict[str, Union[str, float, int, dict]] = None,
+        components_to_quantize: Optional[List[str]] = None,
+        quant_mapping: Dict[str, Union[DiffQuantConfigMixin, "TransformersQuantConfigMixin"]] = None,
+    ):
+        self.quant_backend = quant_backend
+        # Initialize kwargs to be {} to set to the defaults.
+        self.quant_kwargs = quant_kwargs or {}
+        self.components_to_quantize = components_to_quantize
+        self.quant_mapping = quant_mapping
+
+        self.post_init()
+
+    def post_init(self):
+        quant_mapping = self.quant_mapping
+        self.is_granular = True if quant_mapping is not None else False
+
+        self._validate_init_args()
+
+    def _validate_init_args(self):
+        if self.quant_backend and self.quant_mapping:
+            raise ValueError("Both `quant_backend` and `quant_mapping` cannot be specified at the same time.")
+
+        if not self.quant_mapping and not self.quant_backend:
+            raise ValueError("Must provide a `quant_backend` when not providing a `quant_mapping`.")
+
+        if not self.quant_kwargs and not self.quant_mapping:
+            raise ValueError("Both `quant_kwargs` and `quant_mapping` cannot be None.")
+
+        if self.quant_backend is not None:
+            self._validate_init_kwargs_in_backends()
+
+        if self.quant_mapping is not None:
+            self._validate_quant_mapping_args()
+
+    def _validate_init_kwargs_in_backends(self):
+        quant_backend = self.quant_backend
+
+        self._check_backend_availability(quant_backend)
+
+        quant_config_mapping_transformers, quant_config_mapping_diffusers = self._get_quant_config_list()
+
+        if quant_config_mapping_transformers is not None:
+            init_kwargs_transformers = inspect.signature(quant_config_mapping_transformers[quant_backend].__init__)
+            init_kwargs_transformers = {name for name in init_kwargs_transformers.parameters if name != "self"}
+        else:
+            init_kwargs_transformers = None
+
+        init_kwargs_diffusers = inspect.signature(quant_config_mapping_diffusers[quant_backend].__init__)
+        init_kwargs_diffusers = {name for name in init_kwargs_diffusers.parameters if name != "self"}
+
+        if init_kwargs_transformers != init_kwargs_diffusers:
+            raise ValueError(
+                "The signatures of the __init__ methods of the quantization config classes in `diffusers` and `transformers` don't match. "
+                f"Please provide a `quant_mapping` instead, in the {self.__class__.__name__} class. Refer to the docs to learn more about how "
+                "this mapping would look like: TODO."
+            )
+
+    def _validate_quant_mapping_args(self):
+        quant_mapping = self.quant_mapping
+        transformers_map, diffusers_map = self._get_quant_config_list()
+
+        available_transformers = list(transformers_map.values()) if transformers_map else None
+        available_diffusers = list(diffusers_map.values())
+
+        for module_name, config in quant_mapping.items():
+            if any(isinstance(config, cfg) for cfg in available_diffusers):
+                continue
+
+            if available_transformers and any(isinstance(config, cfg) for cfg in available_transformers):
+                continue
+
+            if available_transformers:
+                raise ValueError(
+                    f"Provided config for module_name={module_name} could not be found. "
+                    f"Available diffusers configs: {available_diffusers}; "
+                    f"Available transformers configs: {available_transformers}."
+                )
+            else:
+                raise ValueError(
+                    f"Provided config for module_name={module_name} could not be found. "
+                    f"Available diffusers configs: {available_diffusers}."
+                )
+
+    def _check_backend_availability(self, quant_backend: str):
+        quant_config_mapping_transformers, quant_config_mapping_diffusers = self._get_quant_config_list()
+
+        available_backends_transformers = (
+            list(quant_config_mapping_transformers.keys()) if quant_config_mapping_transformers else None
+        )
+        available_backends_diffusers = list(quant_config_mapping_diffusers.keys())
+
+        if (
+            available_backends_transformers and quant_backend not in available_backends_transformers
+        ) or quant_backend not in quant_config_mapping_diffusers:
+            error_message = f"Provided quant_backend={quant_backend} was not found."
+            if available_backends_transformers:
+                error_message += f"\nAvailable ones (transformers): {available_backends_transformers}."
+            error_message += f"\nAvailable ones (diffusers): {available_backends_diffusers}."
+            raise ValueError(error_message)
+
+    def _resolve_quant_config(self, is_diffusers: bool = True, module_name: str = None):
+        quant_config_mapping_transformers, quant_config_mapping_diffusers = self._get_quant_config_list()
+
+        quant_mapping = self.quant_mapping
+        components_to_quantize = self.components_to_quantize
+
+        # Granular case
+        if self.is_granular and module_name in quant_mapping:
+            logger.debug(f"Initializing quantization config class for {module_name}.")
+            config = quant_mapping[module_name]
+            return config
+
+        # Global config case
+        else:
+            should_quantize = False
+            # Only quantize the modules requested for.
+            if components_to_quantize and module_name in components_to_quantize:
+                should_quantize = True
+            # No specification for `components_to_quantize` means all modules should be quantized.
+            elif not self.is_granular and not components_to_quantize:
+                should_quantize = True
+
+            if should_quantize:
+                logger.debug(f"Initializing quantization config class for {module_name}.")
+                mapping_to_use = quant_config_mapping_diffusers if is_diffusers else quant_config_mapping_transformers
+                quant_config_cls = mapping_to_use[self.quant_backend]
+                quant_kwargs = self.quant_kwargs
+                return quant_config_cls(**quant_kwargs)
+
+        # Fallback: no applicable configuration found.
+        return None
+
+    def _get_quant_config_list(self):
+        if is_transformers_available():
+            from transformers.quantizers.auto import (
+                AUTO_QUANTIZATION_CONFIG_MAPPING as quant_config_mapping_transformers,
+            )
+        else:
+            quant_config_mapping_transformers = None
+
+        from ..quantizers.auto import AUTO_QUANTIZATION_CONFIG_MAPPING as quant_config_mapping_diffusers
+
+        return quant_config_mapping_transformers, quant_config_mapping_diffusers

src/diffusers/quantizers/quantization_config.py

@@ -75,7 +75,7 @@ def from_dict(cls, config_dict, return_unused_kwargs=False, **kwargs):
         Args:
             config_dict (`Dict[str, Any]`):
                 Dictionary that will be used to instantiate the configuration object.
-            return_unused_kwargs (`bool`,*optional*, defaults to `False`):
+            return_unused_kwargs (`bool`, *optional*, defaults to `False`):
                 Whether or not to return a list of unused keyword arguments. Used for `from_pretrained` method in
                 `PreTrainedModel`.
             kwargs (`Dict[str, Any]`):

src/diffusers/utils/testing_utils.py

@@ -38,6 +38,7 @@
     is_note_seq_available,
     is_onnx_available,
     is_opencv_available,
+    is_optimum_quanto_available,
     is_peft_available,
     is_timm_available,
     is_torch_available,
@@ -486,6 +487,13 @@ def require_bitsandbytes(test_case):
     return unittest.skipUnless(is_bitsandbytes_available(), "test requires bitsandbytes")(test_case)
 
 
+def require_quanto(test_case):
+    """
+    Decorator marking a test that requires quanto. These tests are skipped when quanto isn't installed.
+    """
+    return unittest.skipUnless(is_optimum_quanto_available(), "test requires quanto")(test_case)
+
+
 def require_accelerate(test_case):
     """
     Decorator marking a test that requires accelerate. These tests are skipped when accelerate isn't installed.