fix handler

handler.py

@@ -1,4 +1,4 @@
-# handler.py
+# HANDLER_MARKER_2025-08-29
 import io, base64
 from PIL import Image
 import torch
@@ -6,30 +6,17 @@ from diffusers import StableDiffusionXLInpaintPipeline
 
 class EndpointHandler:
     def __init__(self, path="."):
-        print("handler v2: loading without variant arg")  # <-- shows in logs
+        print("HANDLER_MARKER_2025-08-29: loading WITHOUT variant arg")
         model_id = "andro-flock/LUSTIFY-SDXL-NSFW-checkpoint-v2-0-INPAINTING"
-
-        # No `variant=` here. Start with fp16, fall back if needed.
-        try:
-            self.pipe = StableDiffusionXLInpaintPipeline.from_pretrained(
-                model_id, torch_dtype=torch.float16, use_safetensors=True
-            ).to("cuda")
-        except Exception as e:
-            print("fp16 failed -> trying bf16:", e)
-            try:
-                self.pipe = StableDiffusionXLInpaintPipeline.from_pretrained(
-                    model_id, torch_dtype=torch.bfloat16, use_safetensors=True
-                ).to("cuda")
-            except Exception as e2:
-                print("bf16 failed -> using fp32:", e2)
-                self.pipe = StableDiffusionXLInpaintPipeline.from_pretrained(
-                    model_id, torch_dtype=torch.float32, use_safetensors=True
-                ).to("cuda")
-
+        # NO variant= here
+        self.pipe = StableDiffusionXLInpaintPipeline.from_pretrained(
+            model_id, torch_dtype=torch.float16, use_safetensors=True
+        ).to("cuda")
         self.pipe.enable_attention_slicing()
 
     def _to_pil(self, data, mode):
         if isinstance(data, str):
+            import base64, io
             data = base64.b64decode(data)
         return Image.open(io.BytesIO(data)).convert(mode)
 
@@ -37,19 +24,10 @@ class EndpointHandler:
         prompt   = data.get("prompt", "")
         init_img = self._to_pil(data["image"], "RGB")
         mask_img = self._to_pil(data["mask"],  "L")  # white=repaint, black=keep
-
         steps    = int(data.get("num_inference_steps", 30))
         guidance = float(data.get("guidance_scale", 7.0))
         strength = float(data.get("strength", 0.85))
-
-        out = self.pipe(
-            prompt=prompt,
-            image=init_img,
-            mask_image=mask_img,
-            num_inference_steps=steps,
-            guidance_scale=guidance,
-            strength=strength
-        ).images[0]
-
+        out = self.pipe(prompt=prompt, image=init_img, mask_image=mask_img,
+                        num_inference_steps=steps, guidance_scale=guidance, strength=strength).images[0]
         buf = io.BytesIO(); out.save(buf, format="PNG")
         return {"image_base64": base64.b64encode(buf.getvalue()).decode()}