fix(env_config): graceful CPU fallback when ROCm PyTorch unavailable

- load_optimized() now catches device='cuda' failures on ROCm systems
  where PyTorch-ROCm is not installed, degrades to CPU gracefully
- deploy.sh removes CPU-only onnxruntime before installing onnxruntime-rocm
  to prevent the shadowing bug

Author: solderzzc

skills/detection/yolo-detection-2026/deploy.sh

@@ -160,6 +160,14 @@ fi
 log "Installing dependencies from $REQ_FILE ..."
 emit "{\"event\": \"progress\", \"stage\": \"install\", \"message\": \"Installing $BACKEND dependencies...\"}"
 
+# ROCm: remove CPU-only onnxruntime if present (it shadows onnxruntime-rocm)
+if [ "$BACKEND" = "rocm" ]; then
+    if "$PIP" show onnxruntime &>/dev/null 2>&1; then
+        log "Removing CPU-only onnxruntime to avoid shadowing onnxruntime-rocm..."
+        "$PIP" uninstall -y onnxruntime -q 2>&1 || true
+    fi
+fi
+
 "$PIP" install -r "$REQ_FILE" -q 2>&1 | tail -5 >&2
 
 # ─── Step 5: Pre-convert model to optimized format ───────────────────────────

skills/detection/yolo-detection-2026/scripts/env_config.py

@@ -467,13 +467,23 @@ def load_optimized(self, model_name: str, use_optimized: bool = True):
 
             # Fallback: use the PT model we already loaded
             _log("Falling back to PyTorch model")
-            pt_model.to(self.device)
+            try:
+                pt_model.to(self.device)
+            except Exception as e:
+                _log(f"Device {self.device} unavailable ({e}), falling back to CPU")
+                self.device = "cpu"
+                pt_model.to("cpu")
             self.load_ms = (time.perf_counter() - t0) * 1000
             return pt_model, "pytorch"
 
         # No optimization requested or framework missing
         model = YOLO(f"{model_name}.pt")
-        model.to(self.device)
+        try:
+            model.to(self.device)
+        except Exception as e:
+            _log(f"Device {self.device} unavailable ({e}), falling back to CPU")
+            self.device = "cpu"
+            model.to("cpu")
         self.load_ms = (time.perf_counter() - t0) * 1000
         return model, "pytorch"
 

skills/lib/env_config.py

@@ -467,13 +467,23 @@ def load_optimized(self, model_name: str, use_optimized: bool = True):
 
             # Fallback: use the PT model we already loaded
             _log("Falling back to PyTorch model")
-            pt_model.to(self.device)
+            try:
+                pt_model.to(self.device)
+            except Exception as e:
+                _log(f"Device {self.device} unavailable ({e}), falling back to CPU")
+                self.device = "cpu"
+                pt_model.to("cpu")
             self.load_ms = (time.perf_counter() - t0) * 1000
             return pt_model, "pytorch"
 
         # No optimization requested or framework missing
         model = YOLO(f"{model_name}.pt")
-        model.to(self.device)
+        try:
+            model.to(self.device)
+        except Exception as e:
+            _log(f"Device {self.device} unavailable ({e}), falling back to CPU")
+            self.device = "cpu"
+            model.to("cpu")
         self.load_ms = (time.perf_counter() - t0) * 1000
         return model, "pytorch"