try with lower diffusers patch

Author: ErikKaum

setup.py

@@ -32,7 +32,7 @@
 extras = {}
 
 extras["st"] = ["sentence_transformers==3.4.1"]
-extras["diffusers"] = ["diffusers==0.32.2", "accelerate==1.4.0"]
+extras["diffusers"] = ["diffusers==0.32.1", "accelerate==1.4.0"]
 # Includes `peft` as PEFT requires `torch` so having `peft` as a core dependency
 # means that `torch` will be installed even if the `torch` extra is not specified.
 extras["torch"] = ["torch==2.3.1", "torchvision", "torchaudio", "peft==0.14.0"]

src/huggingface_inference_toolkit/diffusers_utils.py

@@ -73,11 +73,7 @@ def __call__(
             logger.warning("The `output_type` cannot be modified, and PIL will be used by default instead.")
 
         # Call pipeline with parameters
-        if self.pipeline.device.type == "cuda":
-            with torch.autocast("cuda", dtype=torch.bfloat16):
-                out = self.pipeline(prompt, num_images_per_prompt=1, **kwargs)
-        else:
-            out = self.pipeline(prompt, num_images_per_prompt=1, **kwargs)
+        out = self.pipeline(prompt, num_images_per_prompt=1, **kwargs)
 
         return out.images[0]