High performance self-hosted photo and video management solution.
TypeScript backup-toolfluttergoogle-photosgoogle-photos-alternativejavascriptmobile-appnestjsnodejsphoto-galleryphotosphotos-managementself-hostedsveltesveltekittypescriptvideos
fix(ml): stabilize MIGraphX inference (#28444)
Fabian Wimberger53a24783f58e
diff --git a/docs/docs/install/environment-variables.md b/docs/docs/install/environment-variables.mdindex ca22c5ad34..0932fa2855 100644--- a/docs/docs/install/environment-variables.md+++ b/docs/docs/install/environment-variables.md@@ -154,33 +154,33 @@ Redis (Sentinel) URL example JSON before encoding:## Machine Learning-| Variable | Description | Default | Containers |-| :---------------------------------------------------------- | :----------------------------------------------------------------------------------------------------------------------------------------------------------- | :-----------------------------: | :--------------- |-| `MACHINE_LEARNING_MODEL_TTL` | Inactivity time (s) before a model is unloaded (disabled if \<= 0) | `300` | machine learning |-| `MACHINE_LEARNING_MODEL_TTL_POLL_S` | Interval (s) between checks for the model TTL (disabled if \<= 0) | `10` | machine learning |-| `MACHINE_LEARNING_CACHE_FOLDER` | Directory where models are downloaded | `/cache` | machine learning |-| `MACHINE_LEARNING_REQUEST_THREADS`<sup>\*1</sup> | Thread count of the request thread pool (disabled if \<= 0) | number of CPU cores | machine learning |-| `MACHINE_LEARNING_MODEL_INTER_OP_THREADS` | Number of parallel model operations | `1` | machine learning |-| `MACHINE_LEARNING_MODEL_INTRA_OP_THREADS` | Number of threads for each model operation | `2` | machine learning |-| `MACHINE_LEARNING_WORKERS`<sup>\*2</sup> | Number of worker processes to spawn | `1` | machine learning |-| `MACHINE_LEARNING_HTTP_KEEPALIVE_TIMEOUT_S`<sup>\*3</sup> | HTTP Keep-alive time in seconds | `2` | machine learning |-| `MACHINE_LEARNING_WORKER_TIMEOUT` | Maximum time (s) of unresponsiveness before a worker is killed | `120` (`300` if using OpenVINO) | machine learning |-| `MACHINE_LEARNING_PRELOAD__CLIP__TEXTUAL` | Comma-separated list of (textual) CLIP model(s) to preload and cache | | machine learning |-| `MACHINE_LEARNING_PRELOAD__CLIP__VISUAL` | Comma-separated list of (visual) CLIP model(s) to preload and cache | | machine learning |-| `MACHINE_LEARNING_PRELOAD__FACIAL_RECOGNITION__RECOGNITION` | Comma-separated list of (recognition) facial recognition model(s) to preload and cache | | machine learning |-| `MACHINE_LEARNING_PRELOAD__FACIAL_RECOGNITION__DETECTION` | Comma-separated list of (detection) facial recognition model(s) to preload and cache | | machine learning |-| `MACHINE_LEARNING_PRELOAD__OCR__RECOGNITION` | Comma-separated list of (recognition) OCR model(s) to preload and cache | | machine learning |-| `MACHINE_LEARNING_PRELOAD__OCR__DETECTION` | Comma-separated list of (detection) OCR model(s) to preload and cache | | machine learning |-| `MACHINE_LEARNING_ANN` | Enable ARM-NN hardware acceleration if supported | `True` | machine learning |-| `MACHINE_LEARNING_ANN_FP16_TURBO` | Execute operations in FP16 precision: increasing speed, reducing precision (applies only to ARM-NN) | `False` | machine learning |-| `MACHINE_LEARNING_ANN_TUNING_LEVEL` | ARM-NN GPU tuning level (1: rapid, 2: normal, 3: exhaustive) | `2` | machine learning |-| `MACHINE_LEARNING_DEVICE_IDS`<sup>\*4</sup> | Device IDs to use in multi-GPU environments | `0` | machine learning |-| `MACHINE_LEARNING_MAX_BATCH_SIZE__FACIAL_RECOGNITION` | Set the maximum number of faces that will be processed at once by the facial recognition model | None (`1` if using OpenVINO) | machine learning |-| `MACHINE_LEARNING_MAX_BATCH_SIZE__OCR` | Set the maximum number of boxes that will be processed at once by the OCR model | `6` | machine learning |-| `MACHINE_LEARNING_RKNN` | Enable RKNN hardware acceleration if supported | `True` | machine learning |-| `MACHINE_LEARNING_RKNN_THREADS` | How many threads of RKNN runtime should be spun up while inferencing. | `1` | machine learning |-| `MACHINE_LEARNING_MODEL_ARENA` | Pre-allocates CPU memory to avoid memory fragmentation | true | machine learning |-| `MACHINE_LEARNING_OPENVINO_PRECISION` | If set to FP16, uses half-precision floating-point operations for faster inference with reduced accuracy (one of [`FP16`, `FP32`], applies only to OpenVINO) | `FP32` | machine learning |+| Variable | Description | Default | Containers |+| :---------------------------------------------------------- | :----------------------------------------------------------------------------------------------------------------------------------------------------------- | :--------------------------: | :--------------- |+| `MACHINE_LEARNING_MODEL_TTL` | Inactivity time (s) before a model is unloaded (disabled if \<= 0) | `300` | machine learning |+| `MACHINE_LEARNING_MODEL_TTL_POLL_S` | Interval (s) between checks for the model TTL (disabled if \<= 0) | `10` | machine learning |+| `MACHINE_LEARNING_CACHE_FOLDER` | Directory where models are downloaded | `/cache` | machine learning |+| `MACHINE_LEARNING_REQUEST_THREADS`<sup>\*1</sup> | Thread count of the request thread pool (disabled if \<= 0) | number of CPU cores | machine learning |+| `MACHINE_LEARNING_MODEL_INTER_OP_THREADS` | Number of parallel model operations | `1` | machine learning |+| `MACHINE_LEARNING_MODEL_INTRA_OP_THREADS` | Number of threads for each model operation | `2` | machine learning |+| `MACHINE_LEARNING_WORKERS`<sup>\*2</sup> | Number of worker processes to spawn | `1` | machine learning |+| `MACHINE_LEARNING_HTTP_KEEPALIVE_TIMEOUT_S`<sup>\*3</sup> | HTTP Keep-alive time in seconds | `2` | machine learning |+| `MACHINE_LEARNING_WORKER_TIMEOUT` | Maximum time (s) of unresponsiveness before a worker is killed | `300` (`900` if using ROCm) | machine learning |+| `MACHINE_LEARNING_PRELOAD__CLIP__TEXTUAL` | Comma-separated list of (textual) CLIP model(s) to preload and cache | | machine learning |+| `MACHINE_LEARNING_PRELOAD__CLIP__VISUAL` | Comma-separated list of (visual) CLIP model(s) to preload and cache | | machine learning |+| `MACHINE_LEARNING_PRELOAD__FACIAL_RECOGNITION__RECOGNITION` | Comma-separated list of (recognition) facial recognition model(s) to preload and cache | | machine learning |+| `MACHINE_LEARNING_PRELOAD__FACIAL_RECOGNITION__DETECTION` | Comma-separated list of (detection) facial recognition model(s) to preload and cache | | machine learning |+| `MACHINE_LEARNING_PRELOAD__OCR__RECOGNITION` | Comma-separated list of (recognition) OCR model(s) to preload and cache | | machine learning |+| `MACHINE_LEARNING_PRELOAD__OCR__DETECTION` | Comma-separated list of (detection) OCR model(s) to preload and cache | | machine learning |+| `MACHINE_LEARNING_ANN` | Enable ARM-NN hardware acceleration if supported | `True` | machine learning |+| `MACHINE_LEARNING_ANN_FP16_TURBO` | Execute operations in FP16 precision: increasing speed, reducing precision (applies only to ARM-NN) | `False` | machine learning |+| `MACHINE_LEARNING_ANN_TUNING_LEVEL` | ARM-NN GPU tuning level (1: rapid, 2: normal, 3: exhaustive) | `2` | machine learning |+| `MACHINE_LEARNING_DEVICE_IDS`<sup>\*4</sup> | Device IDs to use in multi-GPU environments | `0` | machine learning |+| `MACHINE_LEARNING_MAX_BATCH_SIZE__FACIAL_RECOGNITION` | Set the maximum number of faces that will be processed at once by the facial recognition model | None (`1` if using OpenVINO) | machine learning |+| `MACHINE_LEARNING_MAX_BATCH_SIZE__OCR` | Set the maximum number of boxes that will be processed at once by the OCR model | `6` | machine learning |+| `MACHINE_LEARNING_RKNN` | Enable RKNN hardware acceleration if supported | `True` | machine learning |+| `MACHINE_LEARNING_RKNN_THREADS` | How many threads of RKNN runtime should be spun up while inferencing. | `1` | machine learning |+| `MACHINE_LEARNING_MODEL_ARENA` | Pre-allocates CPU memory to avoid memory fragmentation | true | machine learning |+| `MACHINE_LEARNING_OPENVINO_PRECISION` | If set to FP16, uses half-precision floating-point operations for faster inference with reduced accuracy (one of [`FP16`, `FP32`], applies only to OpenVINO) | `FP32` | machine learning |\*1: It is recommended to begin with this parameter when changing the concurrency levels of the machine learning service and then tune the other ones.diff --git a/machine-learning/immich_ml/config.py b/machine-learning/immich_ml/config.pyindex c5ba0bdf0a..d1e44fd70a 100644--- a/machine-learning/immich_ml/config.py+++ b/machine-learning/immich_ml/config.py@@ -6,7 +6,7 @@ from pathlib import Pathfrom socket import socketfrom gunicorn.arbiter import Arbiter-from pydantic import BaseModel+from pydantic import BaseModel, Fieldfrom pydantic_settings import BaseSettings, SettingsConfigDictfrom rich.console import Consolefrom rich.logging import RichHandler@@ -42,6 +42,10 @@ class MaxBatchSize(BaseModel):ocr: int | None = None+def default_worker_timeout() -> int:+ return 900 if os.environ.get("DEVICE") == "rocm" else 300++class Settings(BaseSettings):model_config = SettingsConfigDict(env_prefix="MACHINE_LEARNING_",@@ -54,7 +58,7 @@ class Settings(BaseSettings):model_ttl: int = 300model_ttl_poll_s: int = 10workers: int = 1- worker_timeout: int = 300+ worker_timeout: int = Field(default_factory=default_worker_timeout)http_keepalive_timeout_s: int = 2test_full: bool = Falserequest_threads: int = os.cpu_count() or 4diff --git a/machine-learning/immich_ml/models/facial_recognition/recognition.py b/machine-learning/immich_ml/models/facial_recognition/recognition.pyindex ed1897c9f9..c144bd3eaa 100644--- a/machine-learning/immich_ml/models/facial_recognition/recognition.py+++ b/machine-learning/immich_ml/models/facial_recognition/recognition.py@@ -89,4 +89,10 @@ class FaceRecognizer(InferenceModel):@propertydef _batch_size_default(self) -> int | None:providers = ort.get_available_providers()- return None if self.model_format == ModelFormat.ONNX and "OpenVINOExecutionProvider" not in providers else 1+ if (+ self.model_format == ModelFormat.ONNX+ and "MIGraphXExecutionProvider" not in providers+ and "OpenVINOExecutionProvider" not in providers+ ):+ return None+ return 1diff --git a/machine-learning/immich_ml/sessions/ort.py b/machine-learning/immich_ml/sessions/ort.pyindex bebd235970..579cafc1cd 100644--- a/machine-learning/immich_ml/sessions/ort.py+++ b/machine-learning/immich_ml/sessions/ort.py@@ -1,6 +1,7 @@from __future__ import annotationsfrom pathlib import Path+from threading import Lockfrom typing import Anyimport numpy as np@@ -12,6 +13,37 @@ from immich_ml.schemas import ModelPrecision, SessionNodefrom ..config import log, settings+MigraphxInputSignature = tuple[tuple[str, str, tuple[int, ...]], ...]++_migraphx_registry_lock = Lock()+_migraphx_model_locks: dict[str, Lock] = {}+_migraphx_compiled_inputs: set[tuple[str, MigraphxInputSignature]] = set()+++def _migraphx_get_model_lock(model_key: str) -> Lock:+ with _migraphx_registry_lock:+ lock = _migraphx_model_locks.get(model_key)+ if lock is None:+ lock = Lock()+ _migraphx_model_locks[model_key] = lock+ return lock+++def _migraphx_has_compiled_input(key: tuple[str, MigraphxInputSignature]) -> bool:+ with _migraphx_registry_lock:+ return key in _migraphx_compiled_inputs+++def _migraphx_mark_compiled_input(key: tuple[str, MigraphxInputSignature]) -> None:+ with _migraphx_registry_lock:+ _migraphx_compiled_inputs.add(key)+++def _migraphx_input_signature(+ input_feed: dict[str, NDArray[np.float32]] | dict[str, NDArray[np.int32]],+) -> MigraphxInputSignature:+ return tuple((name, str(value.dtype), tuple(value.shape)) for name, value in sorted(input_feed.items()))+class OrtSession:session: ort.InferenceSession@@ -48,7 +80,21 @@ class OrtSession:input_feed: dict[str, NDArray[np.float32]] | dict[str, NDArray[np.int32]],run_options: Any = None,) -> list[NDArray[np.float32]]:- outputs: list[NDArray[np.float32]] = self.session.run(output_names, input_feed, run_options)+ if "MIGraphXExecutionProvider" in self.providers:+ model_key = self.model_path.resolve().as_posix()+ input_key = (model_key, _migraphx_input_signature(input_feed))+ if not _migraphx_has_compiled_input(input_key):+ model_lock = _migraphx_get_model_lock(model_key)+ with model_lock:+ if not _migraphx_has_compiled_input(input_key):+ outputs: list[NDArray[np.float32]] = self.session.run(output_names, input_feed, run_options)+ _migraphx_mark_compiled_input(input_key)+ return outputs++ outputs = self.session.run(output_names, input_feed, run_options)+ return outputs++ outputs = self.session.run(output_names, input_feed, run_options)return outputs@propertydiff --git a/machine-learning/test_main.py b/machine-learning/test_main.pyindex cce334e40e..b281c0d417 100644--- a/machine-learning/test_main.py+++ b/machine-learning/test_main.py@@ -35,7 +35,37 @@ from immich_ml.sessions.ort import OrtSessionfrom immich_ml.sessions.rknn import RknnSession, run_inference+class FakeLock:+ def __init__(self) -> None:+ self.enter = mock.Mock()+ self.exit = mock.Mock()++ def __enter__(self) -> None:+ self.enter()++ def __exit__(self, *args: object) -> None:+ self.exit(*args)++class TestBase:+ def test_sets_default_worker_timeout(self, monkeypatch: MonkeyPatch) -> None:+ monkeypatch.delenv("DEVICE", raising=False)+ monkeypatch.delenv("MACHINE_LEARNING_WORKER_TIMEOUT", raising=False)++ assert Settings().worker_timeout == 300++ def test_sets_rocm_default_worker_timeout(self, monkeypatch: MonkeyPatch) -> None:+ monkeypatch.setenv("DEVICE", "rocm")+ monkeypatch.delenv("MACHINE_LEARNING_WORKER_TIMEOUT", raising=False)++ assert Settings().worker_timeout == 900++ def test_worker_timeout_env_override(self, monkeypatch: MonkeyPatch) -> None:+ monkeypatch.setenv("DEVICE", "rocm")+ monkeypatch.setenv("MACHINE_LEARNING_WORKER_TIMEOUT", "1200")++ assert Settings().worker_timeout == 1200+def test_sets_default_cache_dir(self) -> None:encoder = OpenClipTextualEncoder("ViT-B-32__openai")@@ -413,6 +443,52 @@ class TestOrtSession:assert sess_options is session.sess_options+ def test_serializes_rocm_first_run_for_new_input_signature(self, mocker: MockerFixture) -> None:+ lock = FakeLock()+ get_model_lock = mocker.patch("immich_ml.sessions.ort._migraphx_get_model_lock", return_value=lock)+ mocker.patch("immich_ml.sessions.ort._migraphx_compiled_inputs", set())+ mocker.patch("immich_ml.sessions.ort.Path.mkdir")+ session = OrtSession("/cache/ViT-B-32__openai/model.onnx", providers=["MIGraphXExecutionProvider"])+ input_feed = {"input": np.random.rand(1, 3, 224, 224).astype(np.float32)}++ session.run(None, input_feed)+ session.run(None, input_feed)++ lock.enter.assert_called_once()+ lock.exit.assert_called_once()+ get_model_lock.assert_called_once()+ session.session.run.assert_has_calls([mock.call(None, input_feed, None), mock.call(None, input_feed, None)])++ def test_serializes_rocm_run_for_each_new_input_signature(self, mocker: MockerFixture) -> None:+ lock = FakeLock()+ mocker.patch("immich_ml.sessions.ort._migraphx_get_model_lock", return_value=lock)+ mocker.patch("immich_ml.sessions.ort._migraphx_compiled_inputs", set())+ mocker.patch("immich_ml.sessions.ort.Path.mkdir")+ session = OrtSession("/cache/ViT-B-32__openai/model.onnx", providers=["MIGraphXExecutionProvider"])+ input_feed = {"input": np.random.rand(1, 3, 224, 224).astype(np.float32)}+ new_shape_input_feed = {"input": np.random.rand(2, 3, 224, 224).astype(np.float32)}++ session.run(None, input_feed)+ session.run(None, new_shape_input_feed)++ assert lock.enter.call_count == 2+ assert lock.exit.call_count == 2+ session.session.run.assert_has_calls(+ [mock.call(None, input_feed, None), mock.call(None, new_shape_input_feed, None)]+ )++ def test_does_not_serialize_non_rocm_run(self, mocker: MockerFixture) -> None:+ lock = FakeLock()+ get_model_lock = mocker.patch("immich_ml.sessions.ort._migraphx_get_model_lock", return_value=lock)+ session = OrtSession("/cache/ViT-B-32__openai/model.onnx", providers=["CPUExecutionProvider"])+ input_feed = {"input": np.random.rand(1, 3, 224, 224).astype(np.float32)}++ session.run(None, input_feed)++ get_model_lock.assert_not_called()+ lock.enter.assert_not_called()+ session.session.run.assert_called_once_with(None, input_feed, None)+class TestAnnSession:def test_creates_ann_session(self, ann_session: mock.Mock, info: mock.Mock) -> None:@@ -883,6 +959,34 @@ class TestFaceRecognition:onnx.load.assert_not_called()onnx.save.assert_not_called()+ def test_recognition_does_not_add_batch_axis_for_migraphx(+ self, ort_session: mock.Mock, path: mock.Mock, mocker: MockerFixture+ ) -> None:+ onnx = mocker.patch("immich_ml.models.facial_recognition.recognition.onnx", autospec=True)+ update_dims = mocker.patch(+ "immich_ml.models.facial_recognition.recognition.update_inputs_outputs_dims", autospec=True+ )+ mocker.patch("immich_ml.models.base.InferenceModel.download")+ mocker.patch("immich_ml.models.facial_recognition.recognition.ArcFaceONNX")+ mocker.patch(+ "immich_ml.models.facial_recognition.recognition.ort.get_available_providers",+ return_value=["MIGraphXExecutionProvider", "CPUExecutionProvider"],+ )+ path.return_value.__truediv__.return_value.__truediv__.return_value.suffix = ".onnx"++ inputs = [SimpleNamespace(name="input.1", shape=(1, 3, 224, 224))]+ outputs = [SimpleNamespace(name="output.1", shape=(1, 800))]+ ort_session.return_value.get_inputs.return_value = inputs+ ort_session.return_value.get_outputs.return_value = outputs++ face_recognizer = FaceRecognizer("buffalo_s", cache_dir=path)+ face_recognizer.load()++ assert face_recognizer.batch_size == 1+ update_dims.assert_not_called()+ onnx.load.assert_not_called()+ onnx.save.assert_not_called()+def test_set_custom_max_batch_size(self, mocker: MockerFixture) -> None:mocker.patch.object(settings, "max_batch_size", MaxBatchSize(facial_recognition=2))