Refactor AMD GPU support (#20239)

* Update ROCm to 7.0.1

* Update ONNXRuntime

* Add back in

* Get basic detection working

* Use env vars

* Handle complex migraphx models

* Enable model caching

* Remove unused

* Add tip to docs

frigate/detectors/detection_runners.py

@@ -78,6 +78,21 @@ class BaseModelRunner(ABC):
 class ONNXModelRunner(BaseModelRunner):
     """Run ONNX models using ONNX Runtime."""
 
+    @staticmethod
+    def is_migraphx_complex_model(model_type: str) -> bool:
+        # Import here to avoid circular imports
+        from frigate.detectors.detector_config import ModelTypeEnum
+        from frigate.embeddings.types import EnrichmentModelTypeEnum
+
+        return model_type in [
+            EnrichmentModelTypeEnum.paddleocr.value,
+            EnrichmentModelTypeEnum.jina_v1.value,
+            EnrichmentModelTypeEnum.jina_v2.value,
+            EnrichmentModelTypeEnum.facenet.value,
+            ModelTypeEnum.rfdetr.value,
+            ModelTypeEnum.dfine.value,
+        ]
+
     def __init__(self, ort: ort.InferenceSession):
         self.ort = ort
 
@@ -441,6 +456,15 @@ def get_optimized_runner(
             options[0]["device_id"],
         )
 
+    if (
+        providers
+        and providers[0] == "MIGraphXExecutionProvider"
+        and ONNXModelRunner.is_migraphx_complex_model(model_type)
+    ):
+        # Don't use MIGraphX for models that are not supported
+        providers.pop(0)
+        options.pop(0)
+
     return ONNXModelRunner(
         ort.InferenceSession(
             model_path,

frigate/util/model.py

@@ -284,7 +284,9 @@ def post_process_yolox(
 
 
 def get_ort_providers(
-    force_cpu: bool = False, device: str | None = "AUTO", requires_fp16: bool = False
+    force_cpu: bool = False,
+    device: str | None = "AUTO",
+    requires_fp16: bool = False,
 ) -> tuple[list[str], list[dict[str, Any]]]:
     if force_cpu:
         return (
@@ -351,12 +353,15 @@ def get_ort_providers(
                     }
                 )
         elif provider == "MIGraphXExecutionProvider":
-            # MIGraphX uses more CPU than ROCM, while also being the same speed
-            if device == "MIGraphX":
-                providers.append(provider)
-                options.append({})
-            else:
-                continue
+            migraphx_cache_dir = os.path.join(MODEL_CACHE_DIR, "migraphx")
+            os.makedirs(migraphx_cache_dir, exist_ok=True)
+
+            providers.append(provider)
+            options.append(
+                {
+                    "migraphx_model_cache_dir": migraphx_cache_dir,
+                }
+            )
         elif provider == "CPUExecutionProvider":
             providers.append(provider)
             options.append(