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

docker/rocm/Dockerfile

@@ -15,14 +15,14 @@ ARG AMDGPU
 
 RUN apt update -qq && \
     apt install -y wget gpg && \
-    wget -O rocm.deb https://repo.radeon.com/amdgpu-install/6.4.1/ubuntu/jammy/amdgpu-install_6.4.60401-1_all.deb && \
+    wget -O rocm.deb https://repo.radeon.com/amdgpu-install/7.0.1/ubuntu/jammy/amdgpu-install_7.0.1.70001-1_all.deb && \
     apt install -y ./rocm.deb && \
     apt update && \
     apt install -qq -y rocm
 
 RUN mkdir -p /opt/rocm-dist/opt/rocm-$ROCM/lib
 RUN cd /opt/rocm-$ROCM/lib && \
-    cp -dpr libMIOpen*.so* libamd*.so* libhip*.so* libhsa*.so* libmigraphx*.so* librocm*.so* librocblas*.so* libroctracer*.so* librocsolver*.so* librocfft*.so* librocprofiler*.so* libroctx*.so* /opt/rocm-dist/opt/rocm-$ROCM/lib/ && \
+    cp -dpr libMIOpen*.so* libamd*.so* libhip*.so* libhsa*.so* libmigraphx*.so* librocm*.so* librocblas*.so* libroctracer*.so* librocsolver*.so* librocfft*.so* librocprofiler*.so* libroctx*.so*  librocroller.so* /opt/rocm-dist/opt/rocm-$ROCM/lib/ && \
     mkdir -p /opt/rocm-dist/opt/rocm-$ROCM/lib/migraphx/lib && \
     cp -dpr migraphx/lib/* /opt/rocm-dist/opt/rocm-$ROCM/lib/migraphx/lib
 RUN cd /opt/rocm-dist/opt/ && ln -s rocm-$ROCM rocm
@@ -64,11 +64,10 @@ COPY --from=rocm /opt/rocm-dist/ /
 
 #######################################################################
 FROM deps-prelim AS rocm-prelim-hsa-override0
-ENV HSA_ENABLE_SDMA=0
-ENV TF_ROCM_USE_IMMEDIATE_MODE=1
-
-# avoid kernel crashes
-ENV HIP_FORCE_DEV_KERNARG=1
+ENV MIGRAPHX_DISABLE_MIOPEN_FUSION=1
+ENV MIGRAPHX_DISABLE_SCHEDULE_PASS=1
+ENV MIGRAPHX_DISABLE_REDUCE_FUSION=1
+ENV MIGRAPHX_ENABLE_HIPRTC_WORKAROUNDS=1
 
 COPY --from=rocm-dist / /
 

docker/rocm/requirements-wheels-rocm.txt

@@ -1 +1 @@
-onnxruntime-rocm @ https://github.com/NickM-27/frigate-onnxruntime-rocm/releases/download/v6.4.1/onnxruntime_rocm-1.21.1-cp311-cp311-linux_x86_64.whl
+onnxruntime-migraphx @ https://github.com/NickM-27/frigate-onnxruntime-rocm/releases/download/v7.0.1/onnxruntime_migraphx-1.23.0-cp311-cp311-linux_x86_64.whl

docker/rocm/rocm.hcl

@@ -2,7 +2,7 @@ variable "AMDGPU" {
   default = "gfx900"
 }
 variable "ROCM" {
-  default = "6.4.1"
+  default = "7.0.1"
 }
 variable "HSA_OVERRIDE_GFX_VERSION" {
   default = ""

docs/docs/configuration/object_detectors.md

@@ -555,6 +555,17 @@ $ docker exec -it frigate /bin/bash -c '(unset HSA_OVERRIDE_GFX_VERSION && /opt/
 
 ### ROCm Supported Models
 
+:::tip
+
+The AMD GPU kernel is known problematic especially when converting models to mxr format. The recommended approach is:
+
+1. Disable object detection in the config.
+2. Startup Frigate with the onnx detector configured, the main object detection model will be converted to mxr format and cached in the config directory.
+3. Once this is finished as indicated by the logs, enable object detection in the UI and confirm that it is working correctly.
+4. Re-enable object detection in the config.
+
+:::
+
 See [ONNX supported models](#supported-models) for supported models, there are some caveats:
 
 - D-FINE models are not supported

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":
+            migraphx_cache_dir = os.path.join(MODEL_CACHE_DIR, "migraphx")
+            os.makedirs(migraphx_cache_dir, exist_ok=True)
+
             providers.append(provider)
-                options.append({})
-            else:
-                continue
+            options.append(
+                {
+                    "migraphx_model_cache_dir": migraphx_cache_dir,
+                }
+            )
         elif provider == "CPUExecutionProvider":
             providers.append(provider)
             options.append(