Brandon/flux model loading

.github/workflows/python-tests.yml

@@ -60,7 +60,7 @@ jobs:
             extra-index-url: 'https://download.pytorch.org/whl/cpu'
             github-env: $GITHUB_ENV
           - platform: macos-default
-            os: macOS-12
+            os: macOS-14
             github-env: $GITHUB_ENV
           - platform: windows-cpu
             os: windows-2022

invokeai/app/invocations/fields.py

@@ -40,6 +40,7 @@ class UIType(str, Enum, metaclass=MetaEnum):
 
     # region Model Field Types
     MainModel = "MainModelField"
+    FluxMainModel = "FluxMainModelField"
     SDXLMainModel = "SDXLMainModelField"
     SDXLRefinerModel = "SDXLRefinerModelField"
     ONNXModel = "ONNXModelField"
@@ -48,6 +49,7 @@ class UIType(str, Enum, metaclass=MetaEnum):
     ControlNetModel = "ControlNetModelField"
     IPAdapterModel = "IPAdapterModelField"
     T2IAdapterModel = "T2IAdapterModelField"
+    T5EncoderModel = "T5EncoderModelField"
     SpandrelImageToImageModel = "SpandrelImageToImageModelField"
     # endregion
 
@@ -125,13 +127,16 @@ class FieldDescriptions:
     negative_cond = "Negative conditioning tensor"
     noise = "Noise tensor"
     clip = "CLIP (tokenizer, text encoder, LoRAs) and skipped layer count"
+    t5_encoder = "T5 tokenizer and text encoder"
     unet = "UNet (scheduler, LoRAs)"
+    transformer = "Transformer"
     vae = "VAE"
     cond = "Conditioning tensor"
     controlnet_model = "ControlNet model to load"
     vae_model = "VAE model to load"
     lora_model = "LoRA model to load"
     main_model = "Main model (UNet, VAE, CLIP) to load"
+    flux_model = "Flux model (Transformer) to load"
     sdxl_main_model = "SDXL Main model (UNet, VAE, CLIP1, CLIP2) to load"
     sdxl_refiner_model = "SDXL Refiner Main Modde (UNet, VAE, CLIP2) to load"
     onnx_main_model = "ONNX Main model (UNet, VAE, CLIP) to load"
@@ -231,6 +236,12 @@ def tuple(self) -> Tuple[int, int, int, int]:
         return (self.r, self.g, self.b, self.a)
 
 
+class FluxConditioningField(BaseModel):
+    """A conditioning tensor primitive value"""
+
+    conditioning_name: str = Field(description="The name of conditioning tensor")
+
+
 class ConditioningField(BaseModel):
     """A conditioning tensor primitive value"""
 

invokeai/app/invocations/flux_text_encoder.py

@@ -0,0 +1,86 @@
+from typing import Literal
+
+import torch
+from transformers import CLIPTextModel, CLIPTokenizer, T5EncoderModel, T5Tokenizer
+
+from invokeai.app.invocations.baseinvocation import BaseInvocation, Classification, invocation
+from invokeai.app.invocations.fields import FieldDescriptions, Input, InputField
+from invokeai.app.invocations.model import CLIPField, T5EncoderField
+from invokeai.app.invocations.primitives import FluxConditioningOutput
+from invokeai.app.services.shared.invocation_context import InvocationContext
+from invokeai.backend.flux.modules.conditioner import HFEncoder
+from invokeai.backend.stable_diffusion.diffusion.conditioning_data import ConditioningFieldData, FLUXConditioningInfo
+
+
+@invocation(
+    "flux_text_encoder",
+    title="FLUX Text Encoding",
+    tags=["prompt", "conditioning", "flux"],
+    category="conditioning",
+    version="1.0.0",
+    classification=Classification.Prototype,
+)
+class FluxTextEncoderInvocation(BaseInvocation):
+    """Encodes and preps a prompt for a flux image."""
+
+    clip: CLIPField = InputField(
+        title="CLIP",
+        description=FieldDescriptions.clip,
+        input=Input.Connection,
+    )
+    t5_encoder: T5EncoderField = InputField(
+        title="T5Encoder",
+        description=FieldDescriptions.t5_encoder,
+        input=Input.Connection,
+    )
+    t5_max_seq_len: Literal[256, 512] = InputField(
+        description="Max sequence length for the T5 encoder. Expected to be 256 for FLUX schnell models and 512 for FLUX dev models."
+    )
+    prompt: str = InputField(description="Text prompt to encode.")
+
+    @torch.no_grad()
+    def invoke(self, context: InvocationContext) -> FluxConditioningOutput:
+        t5_embeddings, clip_embeddings = self._encode_prompt(context)
+        conditioning_data = ConditioningFieldData(
+            conditionings=[FLUXConditioningInfo(clip_embeds=clip_embeddings, t5_embeds=t5_embeddings)]
+        )
+
+        conditioning_name = context.conditioning.save(conditioning_data)
+        return FluxConditioningOutput.build(conditioning_name)
+
+    def _encode_prompt(self, context: InvocationContext) -> tuple[torch.Tensor, torch.Tensor]:
+        # Load CLIP.
+        clip_tokenizer_info = context.models.load(self.clip.tokenizer)
+        clip_text_encoder_info = context.models.load(self.clip.text_encoder)
+
+        # Load T5.
+        t5_tokenizer_info = context.models.load(self.t5_encoder.tokenizer)
+        t5_text_encoder_info = context.models.load(self.t5_encoder.text_encoder)
+
+        prompt = [self.prompt]
+
+        with (
+            t5_text_encoder_info as t5_text_encoder,
+            t5_tokenizer_info as t5_tokenizer,
+        ):
+            assert isinstance(t5_text_encoder, T5EncoderModel)
+            assert isinstance(t5_tokenizer, T5Tokenizer)
+
+            t5_encoder = HFEncoder(t5_text_encoder, t5_tokenizer, False, self.t5_max_seq_len)
+
+            prompt_embeds = t5_encoder(prompt)
+
+        with (
+            clip_text_encoder_info as clip_text_encoder,
+            clip_tokenizer_info as clip_tokenizer,
+        ):
+            assert isinstance(clip_text_encoder, CLIPTextModel)
+            assert isinstance(clip_tokenizer, CLIPTokenizer)
+
+            clip_encoder = HFEncoder(clip_text_encoder, clip_tokenizer, True, 77)
+
+            pooled_prompt_embeds = clip_encoder(prompt)
+
+        assert isinstance(prompt_embeds, torch.Tensor)
+        assert isinstance(pooled_prompt_embeds, torch.Tensor)
+        return prompt_embeds, pooled_prompt_embeds

invokeai/app/invocations/flux_text_to_image.py

@@ -0,0 +1,172 @@
+import torch
+from einops import rearrange
+from PIL import Image
+
+from invokeai.app.invocations.baseinvocation import BaseInvocation, Classification, invocation
+from invokeai.app.invocations.fields import (
+    FieldDescriptions,
+    FluxConditioningField,
+    Input,
+    InputField,
+    WithBoard,
+    WithMetadata,
+)
+from invokeai.app.invocations.model import TransformerField, VAEField
+from invokeai.app.invocations.primitives import ImageOutput
+from invokeai.app.services.session_processor.session_processor_common import CanceledException
+from invokeai.app.services.shared.invocation_context import InvocationContext
+from invokeai.backend.flux.model import Flux
+from invokeai.backend.flux.modules.autoencoder import AutoEncoder
+from invokeai.backend.flux.sampling import denoise, get_noise, get_schedule, prepare_latent_img_patches, unpack
+from invokeai.backend.stable_diffusion.diffusion.conditioning_data import FLUXConditioningInfo
+from invokeai.backend.util.devices import TorchDevice
+
+
+@invocation(
+    "flux_text_to_image",
+    title="FLUX Text to Image",
+    tags=["image", "flux"],
+    category="image",
+    version="1.0.0",
+    classification=Classification.Prototype,
+)
+class FluxTextToImageInvocation(BaseInvocation, WithMetadata, WithBoard):
+    """Text-to-image generation using a FLUX model."""
+
+    transformer: TransformerField = InputField(
+        description=FieldDescriptions.flux_model,
+        input=Input.Connection,
+        title="Transformer",
+    )
+    vae: VAEField = InputField(
+        description=FieldDescriptions.vae,
+        input=Input.Connection,
+    )
+    positive_text_conditioning: FluxConditioningField = InputField(
+        description=FieldDescriptions.positive_cond, input=Input.Connection
+    )
+    width: int = InputField(default=1024, multiple_of=16, description="Width of the generated image.")
+    height: int = InputField(default=1024, multiple_of=16, description="Height of the generated image.")
+    num_steps: int = InputField(
+        default=4, description="Number of diffusion steps. Recommend values are schnell: 4, dev: 50."
+    )
+    guidance: float = InputField(
+        default=4.0,
+        description="The guidance strength. Higher values adhere more strictly to the prompt, and will produce less diverse images. FLUX dev only, ignored for schnell.",
+    )
+    seed: int = InputField(default=0, description="Randomness seed for reproducibility.")
+
+    @torch.no_grad()
+    def invoke(self, context: InvocationContext) -> ImageOutput:
+        # Load the conditioning data.
+        cond_data = context.conditioning.load(self.positive_text_conditioning.conditioning_name)
+        assert len(cond_data.conditionings) == 1
+        flux_conditioning = cond_data.conditionings[0]
+        assert isinstance(flux_conditioning, FLUXConditioningInfo)
+
+        latents = self._run_diffusion(context, flux_conditioning.clip_embeds, flux_conditioning.t5_embeds)
+        image = self._run_vae_decoding(context, latents)
+        image_dto = context.images.save(image=image)
+        return ImageOutput.build(image_dto)
+
+    def _run_diffusion(
+        self,
+        context: InvocationContext,
+        clip_embeddings: torch.Tensor,
+        t5_embeddings: torch.Tensor,
+    ):
+        transformer_info = context.models.load(self.transformer.transformer)
+        inference_dtype = torch.bfloat16
+
+        # Prepare input noise.
+        x = get_noise(
+            num_samples=1,
+            height=self.height,
+            width=self.width,
+            device=TorchDevice.choose_torch_device(),
+            dtype=inference_dtype,
+            seed=self.seed,
+        )
+
+        img, img_ids = prepare_latent_img_patches(x)
+
+        is_schnell = "schnell" in transformer_info.config.config_path
+
+        timesteps = get_schedule(
+            num_steps=self.num_steps,
+            image_seq_len=img.shape[1],
+            shift=not is_schnell,
+        )
+
+        bs, t5_seq_len, _ = t5_embeddings.shape
+        txt_ids = torch.zeros(bs, t5_seq_len, 3, dtype=inference_dtype, device=TorchDevice.choose_torch_device())
+
+        # 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)
+
+        with transformer_info as transformer:
+            assert isinstance(transformer, Flux)
+
+            def step_callback() -> None:
+                if context.util.is_canceled():
+                    raise CanceledException
+
+                # TODO: Make this look like the image before re-enabling
+                # latent_image = unpack(img.float(), self.height, self.width)
+                # latent_image = latent_image.squeeze()  # Remove unnecessary dimensions
+                # flattened_tensor = latent_image.reshape(-1)  # Flatten to shape [48*128*128]
+
+                # # Create a new tensor of the required shape [255, 255, 3]
+                # latent_image = flattened_tensor[: 255 * 255 * 3].reshape(255, 255, 3)  # Reshape to RGB format
+
+                # # Convert to a NumPy array and then to a PIL Image
+                # image = Image.fromarray(latent_image.cpu().numpy().astype(np.uint8))
+
+                # (width, height) = image.size
+                # width *= 8
+                # height *= 8
+
+                # dataURL = image_to_dataURL(image, image_format="JPEG")
+
+                # # TODO: move this whole function to invocation context to properly reference these variables
+                # context._services.events.emit_invocation_denoise_progress(
+                #     context._data.queue_item,
+                #     context._data.invocation,
+                #     state,
+                #     ProgressImage(dataURL=dataURL, width=width, height=height),
+                # )
+
+            x = denoise(
+                model=transformer,
+                img=img,
+                img_ids=img_ids,
+                txt=t5_embeddings,
+                txt_ids=txt_ids,
+                vec=clip_embeddings,
+                timesteps=timesteps,
+                step_callback=step_callback,
+                guidance=self.guidance,
+            )
+
+        x = unpack(x.float(), self.height, self.width)
+
+        return x
+
+    def _run_vae_decoding(
+        self,
+        context: InvocationContext,
+        latents: torch.Tensor,
+    ) -> Image.Image:
+        vae_info = context.models.load(self.vae.vae)
+        with vae_info as vae:
+            assert isinstance(vae, AutoEncoder)
+            latents = latents.to(dtype=TorchDevice.choose_torch_dtype())
+            img = vae.decode(latents)
+
+        img = img.clamp(-1, 1)
+        img = rearrange(img[0], "c h w -> h w c")
+        img_pil = Image.fromarray((127.5 * (img + 1.0)).byte().cpu().numpy())
+
+        return img_pil