Manage quantization of models within the loader

Author: brandonrising

invokeai/app/invocations/fields.py

@@ -126,6 +126,7 @@ class FieldDescriptions:
     negative_cond = "Negative conditioning tensor"
     noise = "Noise tensor"
     clip = "CLIP (tokenizer, text encoder, LoRAs) and skipped layer count"
+    t5Encoder = "T5 tokenizer and text encoder"
     unet = "UNet (scheduler, LoRAs)"
     transformer = "Transformer"
     vae = "VAE"

invokeai/app/invocations/flux_text_to_image.py

@@ -6,14 +6,14 @@
 from diffusers import AutoencoderKL, FlowMatchEulerDiscreteScheduler
 from diffusers.models.transformers.transformer_flux import FluxTransformer2DModel
 from diffusers.pipelines.flux.pipeline_flux import FluxPipeline
-from invokeai.app.invocations.model import ModelIdentifierField
+from invokeai.app.invocations.model import TransformerField, CLIPField, T5EncoderField, VAEField
 from optimum.quanto import qfloat8
 from PIL import Image
 from transformers import CLIPTextModel, CLIPTokenizer, T5EncoderModel, T5TokenizerFast
 from transformers.models.auto import AutoModelForTextEncoding
 
 from invokeai.app.invocations.baseinvocation import BaseInvocation, invocation
-from invokeai.app.invocations.fields import InputField, WithBoard, WithMetadata, UIType, Input
+from invokeai.app.invocations.fields import InputField, WithBoard, WithMetadata, Input, FieldDescriptions
 from invokeai.app.invocations.primitives import ImageOutput
 from invokeai.app.services.shared.invocation_context import InvocationContext
 from invokeai.backend.quantization.fast_quantized_diffusion_model import FastQuantizedDiffusersModel
@@ -42,14 +42,24 @@ class QuantizedModelForTextEncoding(FastQuantizedTransformersModel):
 class FluxTextToImageInvocation(BaseInvocation, WithMetadata, WithBoard):
     """Text-to-image generation using a FLUX model."""
 
-    flux_model: ModelIdentifierField = InputField(
-        description="The Flux model",
-        input=Input.Any,
-        ui_type=UIType.FluxMainModel
+    transformer: TransformerField = InputField(
+        description=FieldDescriptions.unet,
+        input=Input.Connection,
+        title="Transformer",
     )
-    model: TFluxModelKeys = InputField(description="The FLUX model to use for text-to-image generation.")
-    use_8bit: bool = InputField(
-        default=False, description="Whether to quantize the transformer model to 8-bit precision."
+    clip: CLIPField = InputField(
+        title="CLIP",
+        description=FieldDescriptions.clip,
+        input=Input.Connection,
+    )
+    t5Encoder: T5EncoderField = InputField(
+        title="T5EncoderField",
+        description=FieldDescriptions.t5Encoder,
+        input=Input.Connection,
+    )
+    vae: VAEField = InputField(
+        description=FieldDescriptions.vae,
+        input=Input.Connection,
     )
     positive_prompt: str = InputField(description="Positive prompt for text-to-image generation.")
     width: int = InputField(default=1024, multiple_of=16, description="Width of the generated image.")
@@ -63,45 +73,40 @@ class FluxTextToImageInvocation(BaseInvocation, WithMetadata, WithBoard):
 
     @torch.no_grad()
     def invoke(self, context: InvocationContext) -> ImageOutput:
-        model_path = context.models.download_and_cache_model(FLUX_MODELS[self.model])
-
-        t5_embeddings, clip_embeddings = self._encode_prompt(context, model_path)
-        latents = self._run_diffusion(context, model_path, clip_embeddings, t5_embeddings)
-        image = self._run_vae_decoding(context, model_path, latents)
+        t5_embeddings, clip_embeddings = self._encode_prompt(context)
+        latents = self._run_diffusion(context, clip_embeddings, t5_embeddings)
+        image = self._run_vae_decoding(context, latents)
         image_dto = context.images.save(image=image)
         return ImageOutput.build(image_dto)
 
-    def _encode_prompt(self, context: InvocationContext, flux_model_dir: Path) -> tuple[torch.Tensor, torch.Tensor]:
-        # Determine the T5 max sequence length based on the model.
-        if self.model == "flux-schnell":
-            max_seq_len = 256
-        # elif self.model == "flux-dev":
-        #     max_seq_len = 512
-        else:
-            raise ValueError(f"Unknown model: {self.model}")
+    def _encode_prompt(self, context: InvocationContext) -> tuple[torch.Tensor, torch.Tensor]:
+        # TODO: Determine the T5 max sequence length based on the model.
+        # if self.model == "flux-schnell":
+        max_seq_len = 256
+        # # elif self.model == "flux-dev":
+        # #     max_seq_len = 512
+        # else:
+        #     raise ValueError(f"Unknown model: {self.model}")
 
-        # Load the CLIP tokenizer.
-        clip_tokenizer_path = flux_model_dir / "tokenizer"
-        clip_tokenizer = CLIPTokenizer.from_pretrained(clip_tokenizer_path, local_files_only=True)
-        assert isinstance(clip_tokenizer, CLIPTokenizer)
+        # Load CLIP.
+        clip_tokenizer_info = context.models.load(self.clip.tokenizer)
+        clip_text_encoder_info = context.models.load(self.clip.text_encoder)
 
-        # Load the T5 tokenizer.
-        t5_tokenizer_path = flux_model_dir / "tokenizer_2"
-        t5_tokenizer = T5TokenizerFast.from_pretrained(t5_tokenizer_path, local_files_only=True)
-        assert isinstance(t5_tokenizer, T5TokenizerFast)
+        # Load T5.
+        t5_tokenizer_info = context.models.load(self.t5Encoder.tokenizer)
+        t5_text_encoder_info = context.models.load(self.t5Encoder.text_encoder)
 
-        clip_text_encoder_path = flux_model_dir / "text_encoder"
-        t5_text_encoder_path = flux_model_dir / "text_encoder_2"
         with (
-            context.models.load_local_model(
-                model_path=clip_text_encoder_path, loader=self._load_flux_text_encoder
-            ) as clip_text_encoder,
-            context.models.load_local_model(
-                model_path=t5_text_encoder_path, loader=self._load_flux_text_encoder_2
-            ) as t5_text_encoder,
+            clip_text_encoder_info as clip_text_encoder,
+            t5_text_encoder_info as t5_text_encoder,
+            clip_tokenizer_info as clip_tokenizer,
+            t5_tokenizer_info as t5_tokenizer,
         ):
             assert isinstance(clip_text_encoder, CLIPTextModel)
             assert isinstance(t5_text_encoder, T5EncoderModel)
+            assert isinstance(clip_tokenizer, CLIPTokenizer)
+            assert isinstance(t5_tokenizer, T5TokenizerFast)
+
             pipeline = FluxPipeline(
                 scheduler=None,
                 vae=None,
@@ -114,7 +119,7 @@ def _encode_prompt(self, context: InvocationContext, flux_model_dir: Path) -> tu
 
             # prompt_embeds: T5 embeddings
             # pooled_prompt_embeds: CLIP embeddings
-            prompt_embeds, pooled_prompt_embeds, text_ids = pipeline.encode_prompt(
+            prompt_embeds, pooled_prompt_embeds, _ = pipeline.encode_prompt(
                 prompt=self.positive_prompt,
                 prompt_2=self.positive_prompt,
                 device=TorchDevice.choose_torch_device(),
@@ -128,22 +133,23 @@ def _encode_prompt(self, context: InvocationContext, flux_model_dir: Path) -> tu
     def _run_diffusion(
         self,
         context: InvocationContext,
-        flux_model_dir: Path,
         clip_embeddings: torch.Tensor,
         t5_embeddings: torch.Tensor,
     ):
-        scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(flux_model_dir / "scheduler", local_files_only=True)
+        scheduler_info = context.models.load(self.transformer.scheduler)
+        transformer_info = context.models.load(self.transformer.transformer)
 
         # HACK(ryand): Manually empty the cache. Currently we don't check the size of the model before loading it from
         # disk. Since the transformer model is large (24GB), there's a good chance that it will OOM on 32GB RAM systems
         # if the cache is not empty.
-        context.models._services.model_manager.load.ram_cache.make_room(24 * 2**30)
+        # context.models._services.model_manager.load.ram_cache.make_room(24 * 2**30)
 
-        transformer_path = flux_model_dir / "transformer"
-        with context.models.load_local_model(
-            model_path=transformer_path, loader=self._load_flux_transformer
-        ) as transformer:
+        with (
+            transformer_info as transformer,
+            scheduler_info as scheduler
+        ):
             assert isinstance(transformer, FluxTransformer2DModel)
+            assert isinstance(scheduler, FlowMatchEulerDiscreteScheduler)
 
             flux_pipeline_with_transformer = FluxPipeline(
                 scheduler=scheduler,
@@ -176,11 +182,10 @@ def _run_diffusion(
     def _run_vae_decoding(
         self,
         context: InvocationContext,
-        flux_model_dir: Path,
         latents: torch.Tensor,
     ) -> Image.Image:
-        vae_path = flux_model_dir / "vae"
-        with context.models.load_local_model(model_path=vae_path, loader=self._load_flux_vae) as vae:
+        vae_info = context.models.load(self.vae.vae)
+        with vae_info as vae:
             assert isinstance(vae, AutoencoderKL)
 
             flux_pipeline_with_vae = FluxPipeline(
@@ -205,81 +210,3 @@ def _run_vae_decoding(
 
         assert isinstance(image, Image.Image)
         return image
-
-    @staticmethod
-    def _load_flux_text_encoder(path: Path) -> CLIPTextModel:
-        model = CLIPTextModel.from_pretrained(path, local_files_only=True)
-        assert isinstance(model, CLIPTextModel)
-        return model
-
-    def _load_flux_text_encoder_2(self, path: Path) -> T5EncoderModel:
-        if self.use_8bit:
-            model_8bit_path = path / "quantized"
-            if model_8bit_path.exists():
-                # The quantized model exists, load it.
-                # TODO(ryand): The requantize(...) operation in from_pretrained(...) is very slow. This seems like
-                # something that we should be able to make much faster.
-                q_model = QuantizedModelForTextEncoding.from_pretrained(model_8bit_path)
-
-                # Access the underlying wrapped model.
-                # We access the wrapped model, even though it is private, because it simplifies the type checking by
-                # always returning a T5EncoderModel from this function.
-                model = q_model._wrapped
-            else:
-                # The quantized model does not exist yet, quantize and save it.
-                # TODO(ryand): dtype?
-                model = T5EncoderModel.from_pretrained(path, local_files_only=True)
-                assert isinstance(model, T5EncoderModel)
-
-                q_model = QuantizedModelForTextEncoding.quantize(model, weights=qfloat8)
-
-                model_8bit_path.mkdir(parents=True, exist_ok=True)
-                q_model.save_pretrained(model_8bit_path)
-
-                # (See earlier comment about accessing the wrapped model.)
-                model = q_model._wrapped
-        else:
-            model = T5EncoderModel.from_pretrained(path, local_files_only=True)
-
-        assert isinstance(model, T5EncoderModel)
-        return model
-
-    def _load_flux_transformer(self, path: Path) -> FluxTransformer2DModel:
-        if self.use_8bit:
-            model_8bit_path = path / "quantized"
-            if model_8bit_path.exists():
-                # The quantized model exists, load it.
-                # TODO(ryand): The requantize(...) operation in from_pretrained(...) is very slow. This seems like
-                # something that we should be able to make much faster.
-                q_model = QuantizedFluxTransformer2DModel.from_pretrained(model_8bit_path)
-
-                # Access the underlying wrapped model.
-                # We access the wrapped model, even though it is private, because it simplifies the type checking by
-                # always returning a FluxTransformer2DModel from this function.
-                model = q_model._wrapped
-            else:
-                # The quantized model does not exist yet, quantize and save it.
-                # TODO(ryand): Loading in float16 and then quantizing seems to result in NaNs. In order to run this on
-                # GPUs that don't support bfloat16, we would need to host the quantized model instead of generating it
-                # here.
-                model = FluxTransformer2DModel.from_pretrained(path, local_files_only=True, torch_dtype=torch.bfloat16)
-                assert isinstance(model, FluxTransformer2DModel)
-
-                q_model = QuantizedFluxTransformer2DModel.quantize(model, weights=qfloat8)
-
-                model_8bit_path.mkdir(parents=True, exist_ok=True)
-                q_model.save_pretrained(model_8bit_path)
-
-                # (See earlier comment about accessing the wrapped model.)
-                model = q_model._wrapped
-        else:
-            model = FluxTransformer2DModel.from_pretrained(path, local_files_only=True, torch_dtype=torch.bfloat16)
-
-        assert isinstance(model, FluxTransformer2DModel)
-        return model
-
-    @staticmethod
-    def _load_flux_vae(path: Path) -> AutoencoderKL:
-        model = AutoencoderKL.from_pretrained(path, local_files_only=True)
-        assert isinstance(model, AutoencoderKL)
-        return model

invokeai/app/invocations/model.py

@@ -65,6 +65,10 @@ class TransformerField(BaseModel):
     transformer: ModelIdentifierField = Field(description="Info to load Transformer submodel")
     scheduler: ModelIdentifierField = Field(description="Info to load scheduler submodel")
 
+class T5EncoderField(BaseModel):
+    tokenizer: ModelIdentifierField = Field(description="Info to load tokenizer submodel")
+    text_encoder: ModelIdentifierField = Field(description="Info to load text_encoder submodel")
+
 
 class VAEField(BaseModel):
     vae: ModelIdentifierField = Field(description="Info to load vae submodel")
@@ -133,8 +137,8 @@ class FluxModelLoaderOutput(BaseInvocationOutput):
     """Flux base model loader output"""
 
     transformer: TransformerField = OutputField(description=FieldDescriptions.transformer, title="Transformer")
-    clip: CLIPField = OutputField(description=FieldDescriptions.clip, title="CLIP 1")
-    clip2: CLIPField = OutputField(description=FieldDescriptions.clip, title="CLIP 2")
+    clip: CLIPField = OutputField(description=FieldDescriptions.clip, title="CLIP")
+    t5Encoder: T5EncoderField = OutputField(description=FieldDescriptions.t5Encoder, title="T5 Encoder")
     vae: VAEField = OutputField(description=FieldDescriptions.vae, title="VAE")
 
 
@@ -166,7 +170,7 @@ def invoke(self, context: InvocationContext) -> FluxModelLoaderOutput:
         return FluxModelLoaderOutput(
             transformer=TransformerField(transformer=transformer, scheduler=scheduler),
             clip=CLIPField(tokenizer=tokenizer, text_encoder=text_encoder, loras=[], skipped_layers=0),
-            clip2=CLIPField(tokenizer=tokenizer2, text_encoder=text_encoder2, loras=[], skipped_layers=0),
+            t5Encoder=T5EncoderField(tokenizer=tokenizer2, text_encoder=text_encoder2),
             vae=VAEField(vae=vae),
         )
 

invokeai/backend/model_manager/load/model_loaders/generic_diffusers.py

@@ -78,7 +78,12 @@ def get_hf_load_class(self, model_path: Path, submodel_type: Optional[SubModelTy
 
     # TO DO: Add exception handling
     def _hf_definition_to_type(self, module: str, class_name: str) -> ModelMixin:  # fix with correct type
-        if module in ["diffusers", "transformers"]:
+        if module in [
+            "diffusers",
+            "transformers",
+            "invokeai.backend.quantization.fast_quantized_transformers_model",
+            "invokeai.backend.quantization.fast_quantized_diffusion_model",
+        ]:
             res_type = sys.modules[module]
         else:
             res_type = sys.modules["diffusers"].pipelines

invokeai/backend/model_manager/load/model_util.py

@@ -9,7 +9,7 @@
 import torch
 from diffusers.pipelines.pipeline_utils import DiffusionPipeline
 from diffusers.schedulers.scheduling_utils import SchedulerMixin
-from transformers import CLIPTokenizer
+from transformers import CLIPTokenizer, T5TokenizerFast
 
 from invokeai.backend.image_util.grounding_dino.grounding_dino_pipeline import GroundingDinoPipeline
 from invokeai.backend.image_util.segment_anything.segment_anything_pipeline import SegmentAnythingPipeline
@@ -48,6 +48,13 @@ def calc_model_size_by_data(logger: logging.Logger, model: AnyModel) -> int:
         ),
     ):
         return model.calc_size()
+    elif isinstance(
+        model,
+        (
+            T5TokenizerFast,
+        ),
+    ):
+        return len(model)
     else:
         # TODO(ryand): Promote this from a log to an exception once we are confident that we are handling all of the
         # supported model types.

invokeai/backend/quantization/fast_quantized_diffusion_model.py

@@ -12,15 +12,17 @@
 )
 from optimum.quanto.models import QuantizedDiffusersModel
 from optimum.quanto.models.shared_dict import ShardedStateDict
+from diffusers.models.transformers.transformer_flux import FluxTransformer2DModel
 
 from invokeai.backend.requantize import requantize
 
 
 class FastQuantizedDiffusersModel(QuantizedDiffusersModel):
     @classmethod
-    def from_pretrained(cls, model_name_or_path: Union[str, os.PathLike]):
+    def from_pretrained(cls, model_name_or_path: Union[str, os.PathLike], base_class = FluxTransformer2DModel, **kwargs):
         """We override the `from_pretrained()` method in order to use our custom `requantize()` implementation."""
-        if cls.base_class is None:
+        base_class = base_class or cls.base_class
+        if base_class is None:
             raise ValueError("The `base_class` attribute needs to be configured.")
 
         if not is_accelerate_available():
@@ -43,16 +45,16 @@ def from_pretrained(cls, model_name_or_path: Union[str, os.PathLike]):
 
             with open(model_config_path, "r", encoding="utf-8") as f:
                 original_model_cls_name = json.load(f)["_class_name"]
-            configured_cls_name = cls.base_class.__name__
+            configured_cls_name = base_class.__name__
             if configured_cls_name != original_model_cls_name:
                 raise ValueError(
                     f"Configured base class ({configured_cls_name}) differs from what was derived from the provided configuration ({original_model_cls_name})."
                 )
 
             # Create an empty model
-            config = cls.base_class.load_config(model_name_or_path)
+            config = base_class.load_config(model_name_or_path)
             with init_empty_weights():
-                model = cls.base_class.from_config(config)
+                model = base_class.from_config(config)
 
             # Look for the index of a sharded checkpoint
             checkpoint_file = os.path.join(model_name_or_path, SAFE_WEIGHTS_INDEX_NAME)
@@ -72,6 +74,6 @@ def from_pretrained(cls, model_name_or_path: Union[str, os.PathLike]):
             # Requantize and load quantized weights from state_dict
             requantize(model, state_dict=state_dict, quantization_map=qmap)
             model.eval()
-            return cls(model)
+            return cls(model)._wrapped
         else:
             raise NotImplementedError("Reloading quantized models directly from the hub is not supported yet.")