Fix VAE dtype mismatch (fp16)

app.py

@@ -223,14 +223,37 @@ def get_pipe_and_device() -> Tuple[StableDiffusionXLControlNetImg2ImgPipeline, s
         cn_kwargs["variant"] = "fp16"
 
     controlnet = ControlNetModel.from_pretrained(CONTROLNET_ID, **cn_kwargs).to(device)
-    pipe = StableDiffusionXLControlNetImg2ImgPipeline.from_pretrained(
-        BASE_MODEL_ID,
+
+    # ---------------------------------------------------------
+    # ✅ Fix 1) Force a consistent dtype for VAE to avoid:
+    # RuntimeError: Input type (c10::Half) and bias type (float) should be the same
+    # ---------------------------------------------------------
+    pipe_kwargs = dict(
         controlnet=controlnet,
         use_safetensors=True,
         torch_dtype=dtype,
         add_watermarker=False,
+    )
+    if dtype == torch.float16:
+        pipe_kwargs["variant"] = "fp16"
+
+    pipe = StableDiffusionXLControlNetImg2ImgPipeline.from_pretrained(
+        BASE_MODEL_ID,
+        **pipe_kwargs,
     ).to(device)
 
+    # Force VAE params/bias dtype to match the pipeline dtype
+    if device == "cuda":
+        try:
+            pipe.vae.to(dtype=dtype)
+            # Some pipelines keep VAE in fp32 on purpose; your custom pipeline doesn't
+            # auto-cast inputs to vae.dtype before encode, so disable upcast if present.
+            if hasattr(pipe.vae, "config") and hasattr(pipe.vae.config, "force_upcast"):
+                pipe.vae.config.force_upcast = False
+            print(f"[PIPE] VAE casted to {dtype}. force_upcast set to False (if supported).", flush=True)
+        except Exception as e:
+            print("[PIPE] VAE dtype cast failed:", repr(e), flush=True)
+
     pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
     pipe.enable_attention_slicing()
     try: