* perf(ml): share RKNN weights across NPU cores and halve model RAM
* perf(ml): default rknn threads to npu cores and cache shape setup
* perf(ml): serve rknn runs in arrival order
(cherry picked from commit e659fb62f2bb3f6b1db08ed643d3d054a539c415)
* add immich_model
* graph per shape
* decouple ocr from ort
* prepared graphs
* use f16 for accelerators
* preload all shapes
* warm-up pass after preload
* keep precision for fp32 graphs
* ensure padding for text recognition
* optimize postprocessing for fp16 outputs
* handle raw ctc logits
* use f16 output for rknn
* mmap token embedding
* read device metadata
* fix(ml): pass model_root_dir to OcrOptions for RapidOCR compatibility
Fix a TypeError (Path(None)) when the OCR model is invoked, caused by an upstream change in RapidOCR v3.8.1 (RapidAI/RapidOCR@8ea9626).
RapidOCR now internally calls `Path(cfg.get("model_root_dir"))`. Since `model_root_dir` was missing from `OcrOptions`, it evaluated to `None` and triggered a `TypeError: argument should be a str or an os.PathLike`.
This fix adds the missing `model_root_dir` argument to prevent the error.
Ref: #28331
* fix(ml-test): update OCR tests for RapidOCR schema change
* chore(ml-test): remove unused `cache_dir` parameter from `TextRecognizer`
* Revert "chore(ml-test): remove unused `cache_dir` parameter from `TextRecognizer`"
This reverts commit 007ad7b3f22b0c93cfc237509b460eb6d5c51075.
* fix(ml): use self.cache_dir for model_root_dir in OcrOptions
* fix: stabilize ROCm MIGraphX inference
Serialize MIGraphX session runs so lazy compiles cannot overlap within a worker.
Use a fixed face-recognition batch size for MIGraphX to avoid compiling a new program for each detected face count.
* fix(ml): increase ROCm worker timeout
* fix(ml): narrow MIGraphX compile locking
* docs: format environment variables table
* docs: apply prettier to environment variables table
* fix(ml): handle empty/corrupt images in face detection
When a corrupt or degenerate image with zero-dimension (0 width or 0 height)
reaches the face detection pipeline, insightface's RetinaFace.detect() calls
cv2.resize() with a target size of 0, triggering an OpenCV assertion failure:
error: (-215:Assertion failed) inv_scale_x > 0 in function 'resize'
This crashes the ML worker and returns a 500 error to the server.
Add an early return in FaceDetector._predict() that checks for zero-dimension
images after decoding and returns empty detection results instead of passing
them to the insightface model.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix(ml): move empty image validation to request level
Per review feedback, validate image dimensions in the predict endpoint
(returning 400) rather than in each model's _predict method. This
catches all zero-dimension images before they reach any model task.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix(ml): resolve mypy strict type error in predict endpoint
Use intermediate `decoded` variable so mypy knows `.width` and `.height`
are accessed on `Image`, not on `Image | str`.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* Enable OpenVINO CPU acceleration in Immich
* Remove unnecessary debug log
* Removing checking for device_ids for openvino since cpu will always be available
* Find OpenVINOExecutionProvider index instead of assuming index 0
* Fix openvino tests
* Fix failing test mock. OpenVINO expects provider options, but cuda provide doesn't so use that for mocked tests.
* Support empty provider options in OrtSessions in which case ONNXRuntime will use its own defaults
* Use OpenVINOExecutionProvider for test_sets_provider_options_kwarg
* fix mock
* simplify
* unused variable
---------
Co-authored-by: Aleksander <pejcic@adobe.com>
Co-authored-by: mertalev <101130780+mertalev@users.noreply.github.com>
