more app fix

app.py

@@ -124,6 +124,7 @@ def run_inference_via_server(
     image_path: str,
     instruction: str,
     server_url: str,
+    use_qwenvl: bool = False,
 ) -> Tuple[str, Optional[str]]:
     """Run inference via trace eval server. Returns (prediction, overlay_path)."""
     with open(image_path, "rb") as f:
@@ -131,7 +132,11 @@ def run_inference_via_server(
     headers = {"ngrok-skip-browser-warning": "true"} if "ngrok" in server_url else {}
     r = requests.post(
         f"{server_url.rstrip('/')}/predict",
-        json={"image_base64": image_b64, "instruction": instruction},
+        json={
+            "image_base64": image_b64,
+            "instruction": instruction,
+            "use_qwenvl": use_qwenvl,
+        },
         timeout=120.0,
         headers=headers,
     )
@@ -339,10 +344,14 @@ with demo:
             )
 
         if server_url:
-            prompt = build_prompt(instruction)
+            prompt = (
+                build_franka_prompt(instruction or "predict the trace")
+                if use_qwenvl
+                else build_prompt(instruction)
+            )
             prompt_md = f"**Prompt sent to model:**\n\n```\n{prompt}\n```"
             pred, overlay_path = run_inference_via_server(
-                image_path, instruction, server_url
+                image_path, instruction, server_url, use_qwenvl
             )
         elif use_qwenvl:
             prompt = build_franka_prompt(instruction or "predict the trace")

eval_server.py

@@ -33,6 +33,7 @@ from trace_inference import (
     build_prompt,
     load_model,
     run_inference,
+    run_inference_qwenvl,
 )
 from trace_inference import _model_state as _trace_model_state
 from trajectory_viz import extract_trajectory_from_text
@@ -68,11 +69,13 @@ class TraceEvalServer:
         image_path: Optional[str] = None,
         image_base64: Optional[str] = None,
         instruction: str = "",
+        use_qwenvl: bool = False,
     ) -> Dict[str, Any]:
         """
         Run inference on a single image.
 
         Provide either image_path (file path) or image_base64 (base64-encoded image).
+        If use_qwenvl=True, uses run_inference_qwenvl (Franka-style, requires qwenvl).
         """
         if image_path is None and image_base64 is None:
             return {"error": "Provide image_path or image_base64"}
@@ -99,8 +102,15 @@ class TraceEvalServer:
                 return {"error": f"Invalid image data: {e}"}
 
         try:
-            prompt = build_prompt(instruction)
-            prediction, overlay_path, _ = run_inference(image_path, prompt, self.model_id)
+            if use_qwenvl:
+                output_dict, prediction, _, trace_text = run_inference_qwenvl(
+                    image_path, instruction or "predict the trace", self.model_id
+                )
+                if not output_dict and trace_text and "qwenvl package not found" in trace_text:
+                    return {"error": trace_text}
+            else:
+                prompt = build_prompt(instruction)
+                prediction, _, _ = run_inference(image_path, prompt, self.model_id)
         finally:
             if temp_file_path and os.path.exists(temp_file_path):
                 try:
@@ -112,10 +122,13 @@ class TraceEvalServer:
             return {"error": prediction}
 
         trajectory = extract_trajectory_from_text(prediction)
-        return {
+        result: Dict[str, Any] = {
             "prediction": prediction,
             "trajectory": trajectory,
         }
+        if use_qwenvl and output_dict:
+            result["output_dict"] = output_dict
+        return result
 
     def predict_batch(
         self,
@@ -133,6 +146,7 @@ class TraceEvalServer:
                 image_path=sample.get("image_path"),
                 image_base64=sample.get("image_base64"),
                 instruction=sample.get("instruction", ""),
+                use_qwenvl=sample.get("use_qwenvl", False),
             )
             elapsed = time.time() - start
 
@@ -200,12 +214,14 @@ def create_app(
             - image_path: (optional) path to image file
             - image_base64: (optional) base64-encoded image
             - instruction: natural language task description
+            - use_qwenvl: (optional) if true, use Franka/qwenvl inference (requires qwenvl)
         """
         body = await request.json()
         return trace_server.predict_one(
             image_path=body.get("image_path"),
             image_base64=body.get("image_base64"),
             instruction=body.get("instruction", ""),
+            use_qwenvl=body.get("use_qwenvl", False),
         )
 
     @app.post("/predict_batch")