Fix FLUX output image clamping. And a few other minor fixes to make inference work with the full bfloat16 FLUX transformer model.

invokeai/app/invocations/flux_text_to_image.py

@@ -17,9 +17,9 @@ 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, unpack
+from invokeai.backend.model_manager.config import CheckpointConfigBase
 from invokeai.backend.stable_diffusion.diffusion.conditioning_data import FLUXConditioningInfo
 from invokeai.backend.util.devices import TorchDevice
-from invokeai.backend.model_manager.config import CheckpointConfigBase
 
 
 @invocation(
@@ -90,7 +90,11 @@ class FluxTextToImageInvocation(BaseInvocation, WithMetadata, WithBoard):
         img, img_ids = self._prepare_latent_img_patches(x)
 
         # HACK(ryand): Find a better way to determine if this is a schnell model or not.
-        is_schnell = "schnell" in transformer_info.config.config_path if transformer_info.config and isinstance(transformer_info.config, CheckpointConfigBase) else ""
+        is_schnell = (
+            "schnell" in transformer_info.config.config_path
+            if transformer_info.config and isinstance(transformer_info.config, CheckpointConfigBase)
+            else ""
+        )
         timesteps = get_schedule(
             num_steps=self.num_steps,
             image_seq_len=img.shape[1],
@@ -161,7 +165,7 @@ class FluxTextToImageInvocation(BaseInvocation, WithMetadata, WithBoard):
             latents.to(torch.float32)
             img = vae.decode(latents)
 
-        img.clamp(-1, 1)
+        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())
 

invokeai/backend/flux/sampling.py

@@ -104,9 +104,18 @@ def denoise(
     timesteps: list[float],
     guidance: float = 4.0,
 ):
+    dtype = model.txt_in.bias.dtype
+
+    # TODO(ryand): This shouldn't be necessary if we manage the dtypes properly in the caller.
+    img = img.to(dtype=dtype)
+    img_ids = img_ids.to(dtype=dtype)
+    txt = txt.to(dtype=dtype)
+    txt_ids = txt_ids.to(dtype=dtype)
+    vec = vec.to(dtype=dtype)
+
     # this is ignored for schnell
     guidance_vec = torch.full((img.shape[0],), guidance, device=img.device, dtype=img.dtype)
-    for t_curr, t_prev in zip(timesteps[:-1], timesteps[1:], strict=False):
+    for t_curr, t_prev in zip(timesteps[:-1], timesteps[1:], strict=True):
         t_vec = torch.full((img.shape[0],), t_curr, dtype=img.dtype, device=img.device)
         pred = model(
             img=img,

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

@@ -1,12 +1,12 @@
 # Copyright (c) 2024, Brandon W. Rising and the InvokeAI Development Team
 """Class for Flux model loading in InvokeAI."""
 
-import accelerate
-import torch
 from dataclasses import fields
 from pathlib import Path
 from typing import Any, Optional
 
+import accelerate
+import torch
 import yaml
 from safetensors.torch import load_file
 from transformers import CLIPTextModel, CLIPTokenizer, T5EncoderModel, T5Tokenizer
@@ -25,15 +25,15 @@ from invokeai.backend.model_manager import (
 from invokeai.backend.model_manager.config import (
     CheckpointConfigBase,
     CLIPEmbedDiffusersConfig,
-    MainCheckpointConfig,
     MainBnbQuantized4bCheckpointConfig,
+    MainCheckpointConfig,
     T5EncoderConfig,
     VAECheckpointConfig,
 )
 from invokeai.backend.model_manager.load.model_loader_registry import ModelLoaderRegistry
 from invokeai.backend.model_manager.load.model_loaders.generic_diffusers import GenericDiffusersLoader
-from invokeai.backend.util.silence_warnings import SilenceWarnings
 from invokeai.backend.quantization.bnb_nf4 import quantize_model_nf4
+from invokeai.backend.util.silence_warnings import SilenceWarnings
 
 app_config = get_config()
 
@@ -109,7 +109,9 @@ class T5EncoderCheckpointModel(GenericDiffusersLoader):
             case SubModelType.Tokenizer2:
                 return T5Tokenizer.from_pretrained(Path(config.path) / "tokenizer_2", max_length=512)
             case SubModelType.TextEncoder2:
-                return T5EncoderModel.from_pretrained(Path(config.path) / "text_encoder_2") #TODO: Fix hf subfolder install
+                return T5EncoderModel.from_pretrained(
+                    Path(config.path) / "text_encoder_2"
+                )  # TODO: Fix hf subfolder install
 
         raise Exception("Only Checkpoint Flux models are currently supported.")
 
@@ -153,7 +155,7 @@ class FluxCheckpointModel(GenericDiffusersLoader):
         params = FluxParams(**filtered_data)
 
         with SilenceWarnings():
-            model = load_class(params).to(self._torch_dtype)
+            model = load_class(params)
             sd = load_file(model_path)
             model.load_state_dict(sd, strict=False, assign=True)
         return model