Update src/ai_processor.py

src/ai_processor.py

@@ -29,7 +29,7 @@ def _log_kv(prefix: str, kv: Dict):
 # --- Spaces GPU decorator (REQUIRED) ---
 from spaces import GPU as _SPACES_GPU
 
-@_SPACES_GPU(enable_queue=False)
+@_SPACES_GPU(enable_queue=True)
 def smartheal_gpu_stub(ping: int = 0) -> str:
     return "ready"
 
@@ -152,7 +152,7 @@ def _vlm_infer_gpu(messages, model_id: str, max_new_tokens: int, token: Optional
     pipe = pipeline(
         task="image-text-to-text",
         model=model_id,
-        device_map="auto",            # CUDA init happens here, safely in GPU worker
+        device_map="cpu",            # CUDA init happens here, safely in GPU worker
         token=token,
         trust_remote_code=True,
         model_kwargs={"low_cpu_mem_usage": True},
@@ -217,7 +217,7 @@ def load_yolo_model():
 
 def load_segmentation_model():
     load_model = _import_tf_loader()
-    return load_model(SEG_MODEL_PATH, compile=False)
+    return load_model(SEG_MODEL_PATH, compile=False, custom_objects={'InputLayer': tf.keras.layers.InputLayer})
 
 def load_classification_pipeline():
     pipe = _import_hf_cls()