Update app.py

app.py

@@ -60,12 +60,17 @@ def fill_image(prompt, image, model_selection, paste_back):
     
     pipe = get_pipeline(model_selection)
     
+    # Encode the prompt and ensure the embeddings are in float16
     (
         prompt_embeds,
         negative_prompt_embeds,
         pooled_prompt_embeds,
         negative_pooled_prompt_embeds,
     ) = pipe.encode_prompt(prompt, "cuda", True)
+    prompt_embeds = prompt_embeds.to(torch.float16)
+    negative_prompt_embeds = negative_prompt_embeds.to(torch.float16)
+    pooled_prompt_embeds = pooled_prompt_embeds.to(torch.float16)
+    negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.to(torch.float16)
     
     source = image["background"]
     mask = image["layers"][0]
@@ -251,12 +256,17 @@ def infer(image, width, height, overlap_percentage, num_inference_steps, resize_
     final_prompt = f"{prompt_input}, high quality, 4k"
     
     pipe = get_pipeline(model_selection)
+    # Encode the prompt and ensure the embeddings are in float16
     (
         prompt_embeds,
         negative_prompt_embeds,
         pooled_prompt_embeds,
         negative_pooled_prompt_embeds,
     ) = pipe.encode_prompt(final_prompt, "cuda", True)
+    prompt_embeds = prompt_embeds.to(torch.float16)
+    negative_prompt_embeds = negative_prompt_embeds.to(torch.float16)
+    pooled_prompt_embeds = pooled_prompt_embeds.to(torch.float16)
+    negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.to(torch.float16)
     
     for output_image in pipe(
         prompt_embeds=prompt_embeds,