should be faster

server.py

@@ -41,6 +41,9 @@ class ChatCompletionRequest(BaseModel):
     messages: List[Dict[str, Any]]
     max_tokens: int = 8192
 
+MAX_IMAGE_WIDTH = 512
+MAX_IMAGE_HEIGHT = 388
+
 def _decode_base64_image(data_url: str) -> Image.Image:
     try:
         is_data_url = data_url.startswith("data:")
@@ -52,6 +55,21 @@ def _decode_base64_image(data_url: str) -> Image.Image:
             logger.debug("Decoding image from raw base64 string; length=%d", len(b64data))
         img_bytes = base64.b64decode(b64data)
         img = Image.open(io.BytesIO(img_bytes)).convert("RGB")
+        orig_w, orig_h = img.width, img.height
+        # Downscale if larger than bounds, preserving aspect ratio
+        if orig_w > MAX_IMAGE_WIDTH or orig_h > MAX_IMAGE_HEIGHT:
+            target = (MAX_IMAGE_WIDTH, MAX_IMAGE_HEIGHT)
+            img = img.copy()
+            img.thumbnail(target, Image.LANCZOS)
+            logger.debug(
+                "Resized image from %sx%s to %sx%s (bounds %sx%s)",
+                orig_w,
+                orig_h,
+                img.width,
+                img.height,
+                MAX_IMAGE_WIDTH,
+                MAX_IMAGE_HEIGHT,
+            )
         try:
             logger.debug("Decoded image: size=%sx%s mode=%s", img.width, img.height, img.mode)
         except Exception: