new handler

handler.py

@@ -1,6 +1,8 @@
-from typing import Dict, List, Any
+from typing import cast, Union
+
+import PIL.Image
 import torch
-from base64 import b64decode
+
 from diffusers import AutoencoderKL
 from diffusers.image_processor import VaeImageProcessor
 
@@ -9,11 +11,7 @@ class EndpointHandler:
     def __init__(self, path=""):
         self.device = "cuda"
         self.dtype = torch.bfloat16
-        self.vae = (
-            AutoencoderKL.from_pretrained(path, torch_dtype=self.dtype)
-            .to(self.device, self.dtype)
-            .eval()
-        )
+        self.vae = cast(AutoencoderKL, AutoencoderKL.from_pretrained(path, torch_dtype=self.dtype).to(self.device, self.dtype).eval())
 
         self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
         self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
@@ -35,47 +33,41 @@ class EndpointHandler:
         return latents
 
     @torch.no_grad()
-    def __call__(self, data: Any) -> List[List[Dict[str, float]]]:
+    def __call__(self, data) -> Union[torch.Tensor, PIL.Image.Image]:
         """
         Args:
             data (:obj:):
                 includes the input data and the parameters for the inference.
         """
-        tensor = data["inputs"]
-        tensor = b64decode(tensor.encode("utf-8"))
-        parameters = data.get("parameters", {})
-        if "shape" not in parameters:
-            raise ValueError("Expected `shape` in parameters.")
-        if "dtype" not in parameters:
-            raise ValueError("Expected `dtype` in parameters.")
-        if "height" not in parameters:
-            raise ValueError("Expected `height` in parameters.")
-        if "width" not in parameters:
-            raise ValueError("Expected `width` in parameters.")
-
-        DTYPE_MAP = {
-            "float16": torch.float16,
-            "float32": torch.float32,
-            "bfloat16": torch.bfloat16,
-        }
-
-        shape = parameters.get("shape")
-        dtype = DTYPE_MAP.get(parameters.get("dtype"))
-        height = parameters.get("height")
-        width = parameters.get("width")
-
-        tensor = torch.frombuffer(bytearray(tensor), dtype=dtype).reshape(shape)
+        tensor = cast(torch.Tensor, data["inputs"])
+        parameters = cast(dict, data.get("parameters", {}))
+        if tensor.ndim == 3 and ("height" not in parameters or "width" not in parameters):
+            raise ValueError("Expected `height` and `width` in parameters.")
+        height = cast(int, parameters.get("height", 0))
+        width = cast(int, parameters.get("width", 0))
+        do_scaling = cast(bool, parameters.get("do_scaling", True))
+        output_type = cast(str, parameters.get("output_type", "pil"))
+        partial_postprocess = cast(bool, parameters.get("partial_postprocess", False))
+        if partial_postprocess and output_type != "pt":
+            output_type = "pt"
 
         tensor = tensor.to(self.device, self.dtype)
+        if tensor.ndim == 3:
+            tensor = self._unpack_latents(tensor, height, width, self.vae_scale_factor)
 
-        tensor = self._unpack_latents(tensor, height, width, self.vae_scale_factor)
-        tensor = (
-            tensor / self.vae.config.scaling_factor
-        ) + self.vae.config.shift_factor
+        if do_scaling:
+            tensor = (
+                tensor / self.vae.config.scaling_factor
+            ) + self.vae.config.shift_factor
 
         with torch.no_grad():
-            image = self.vae.decode(tensor, return_dict=False)[0]
+            image = cast(torch.Tensor, self.vae.decode(tensor, return_dict=False)[0])
 
-        image = self.image_processor.postprocess(image, output_type="pil")
+        if partial_postprocess:
+            image = (image * 0.5 + 0.5).clamp(0, 1)
+            image = image.permute(0, 2, 3, 1).contiguous().float()
+            image = (image * 255).round().to(torch.uint8)
+        elif output_type == "pil":
+            image = cast(PIL.Image.Image, self.image_processor.postprocess(image, output_type="pil")[0])
 
-        return image[0]
+        return image