CI: deploy Docker/PDM Space

deploy-development.yaml

@@ -70,7 +70,6 @@ steps:
               "training_model/*"
           ]
       )
-      PY
     displayName: "Deploy to Hugging Face Space (Docker/PDM)"
     env:
       HF_TOKEN: $(HF_TOKEN_DEV) # Add this as a secret variable in Pipeline settings

deploy-main.yaml

@@ -37,8 +37,8 @@ steps:
       import os
       from huggingface_hub import HfApi, upload_folder
 
-      token = os.environ["HF_TOKEN_PROD"]           # provided via Pipeline variable (Secret)
-      space_id = os.environ["HF_SPACE_ID_PROD"]     # from variables above
+      token = os.environ["HF_TOKEN"]           # provided via Pipeline variable (Secret)
+      space_id = os.environ["HF_SPACE_ID"]     # from variables above
 
       api = HfApi(token=token)
 
@@ -61,7 +61,6 @@ steps:
           ignore_patterns=[
               ".git/*",
               "__pycache__/*",
-              "*.mp4", "*.avi", "*.mov",
               "*.zip", "*.tar", "*.tar.gz",
               "*.ipynb", "*.ipynb_checkpoints/*",
               "venv/*", ".venv/*",
@@ -71,7 +70,6 @@ steps:
               "training_model/*"
           ]
       )
-      PY
     displayName: "Deploy to Hugging Face Space (Docker/PDM)"
     env:
       HF_TOKEN: $(HF_TOKEN_PROD) # Add this as a secret variable in Pipeline settings

services/app_service/deim_model.py

@@ -64,7 +64,28 @@ class DeimHgnetV2MDrone(BaseModel):
                 + (", ".join(sorted(set(available))) or "<none>")
             )
 
-        self.device = "cuda" if device == "gpu" and torch.cuda.is_available() else "cpu"
+        # Check if CUDA is available and if we have CUDA model weights
+        cuda_available = torch.cuda.is_available()
+        cuda_model_exists = any(
+            "cuda" in str(p) for p in candidate_paths if p and p.exists()
+        )
+
+        if device == "gpu" and cuda_available and cuda_model_exists:
+            self.device = "cuda"
+            print(f"Using GPU device: {self.device}")
+        else:
+            self.device = "cpu"
+            if device == "gpu":
+                if not cuda_available:
+                    print("Warning: GPU requested but CUDA not available, using CPU")
+                elif not cuda_model_exists:
+                    print(
+                        "Warning: GPU requested but CUDA model not found, using CPU model"
+                    )
+                else:
+                    print("Warning: GPU requested but falling back to CPU")
+            else:
+                print(f"Using CPU device: {self.device}")
         cfg_path = weights_path.with_suffix(".json")
         if not cfg_path.exists():
             raise FileNotFoundError(
@@ -88,7 +109,12 @@ class DeimHgnetV2MDrone(BaseModel):
             )
 
         self.cfg = json.load(open(cfg_path, "r"))
+        print(f"Loading model from: {weights_path}")
+        print(f"Model device: {self.device}")
         self.model = torch.jit.load(weights_path, map_location=self.device).eval()
+        # Ensure model is on the correct device
+        self.model = self.model.to(self.device)
+        print(f"Model loaded successfully on device: {self.device}")
 
     def _preprocess_image(self, image: Image):
         transforms = T.Compose(
@@ -173,8 +199,21 @@ class DeimHgnetV2MDrone(BaseModel):
 
     def predict_image(self, image: Image, min_confidence: float) -> Image:
         tensor = self._preprocess_image(image.copy())
-        with torch.no_grad():
-            labels, bboxes = self.model(tensor)
+        try:
+            with torch.no_grad():
+                labels, bboxes = self.model(tensor)
+        except RuntimeError as e:
+            if "device" in str(e).lower():
+                print(f"Device error: {e}")
+                print(
+                    f"Tensor device: {tensor.device}, Model device: {next(self.model.parameters()).device}"
+                )
+                # Try to move tensor to model's device
+                tensor = tensor.to(next(self.model.parameters()).device)
+                with torch.no_grad():
+                    labels, bboxes = self.model(tensor)
+            else:
+                raise
         dets = self._postprocess_detections(labels, bboxes, min_confidence, image.size)
         dets = self._nms(dets)
         image_np: np.ndarray = np.array(image)