Changelog

v0.8.0 (unreleased)

Deprecated

  • Plugin system API (PluginDiscovery / PluginServer / PluginEndpoint) is deprecated and scheduled for removal in v0.8.0. The platform does not ship the /run/aipc/plugins discovery mechanism; the SDK keeps the imports working for now, emits a DeprecationWarning, and the docs page has been taken offline. Speak up if you depend on it.

Documentation

  • Fixed the video-stream examples across docs and README: dropped references to MediaClient (which never shipped), standardized on FdMediaClient; stream IDs corrected from cam0_main / cam0_sub to the device-exposed main / sub; removed usages of the nonexistent get_stream_info() / get_raw_stream(); corrected the get_encoded_stream() return semantics and the flattened frame.data misuse; added API docs for EncodedStreamClient / EncodedFrame

  • Fixed missing and misleading imports in the application examples: added import time to the multi-model example, import json to the GenAI example; dropped from grpc import RpcError from the error-handling example (the SDK raises RuntimeError); removed unused numpy / sys imports

  • Synced the Chinese inference page with the English one: translated the seven Usage sections that had been English-only since v0.3.0 (segmentation, OCR, CLIP image embeddings, CLIP text encoding, depth estimation, runtime postprocess-config update, GenAI) and added the SegmentationMask / OcrLine / Embedding / DepthMap data-type entries

v0.7.4 (2026-09-08)

New Features

  • Convenience-layer hardware routing (sdk-hardware-routing P5): color.rgb_to_nv12 / color.nv12_to_rgb and draw.draw_detections (NV12 arrays) ride the accel router’s DSP legs by default, degrading to the software impls with accounting when unavailable; Frame.to_jpeg_bytes is hardware-first for keep-fd frames (zero-copy camera-daemon EncodeImage import) while in-memory frames keep the direct CPU encode so streaming hot loops pay no extra copy or RPC; software legs bind private impls only — the routing public functions cannot re-enter the router

  • Frame.resize’s DSP fast path now respects the router policy: SOFTWARE_ONLY skips the attempt, HARDWARE_ONLY raises on failure (refusing the silent CPU fallback), PREFER_HARDWARE reports via the new AccelRouter.note_degradation() (plus a typed policy property; externally reported fallbacks on unregistered ops stay visible in health() and fire on_degradation)

Bug Fixes

  • _encode_jpeg_hw hardcoded fmt="rgb24", so every NV12 keep-fd frame submit was rejected by _resolve_source as a format mismatch; frame-like sources now carry their handle’s format

  • dsp_format’s software leg re-entered the router through the routing public functions (cross-op coupling); it now binds the color impls directly

  • unified the color legs’ ValueError contract (ndim / declared dims / even-size checks run before any daemon submit — caller errors no longer land in the degradation counters)

v0.7.3 (2026-09-07)

New Features

  • Hardware-first accel router (accel module): AccelRouter / RoutePolicy / health() degradation accounting + on_degradation hook / get_default_router(); the default router registers DSP legs for resize_nv12 / rgb_to_nv12 / nv12_to_rgb / encode_jpeg / draw_detections plus the nms software leg

  • DSP annotation compositing (dsp-offload P1): DspClient.blend_hw composites ARGB32 overlays onto NV12 in place on the DSP; draw.render_overlay_rgba renders boxes+captions as a minimal straight-alpha canvas reproducing the software raster’s output

  • Async DSP jobs (dsp-offload P2): every *_hw op takes wait=False with PendingDspJob exposing wait() / done() / release(); blend_hw zero-copy keep-fd chain (refused by default — a state-dependent fatal path on current firmware); encode_jpeg_hw src_buffer_id passthrough skips the pixel read-back; keep-fd frame sources skip the ascontiguousarray pre-copy in the resize/convert legs; polygon and track shapes

  • Overlay annotation: annotate() / annotate_result() push detections straight to camera-daemon’s renderer over the event bus

  • Color conversion and NMS: BT.601 limited-range RGB↔NV12 (numpy path when cv2 is missing); cross-class-aware postprocess.nms

Bug Fixes

  • Finished the inference codec refactor: stray self in _tensor_to_numpy / _parse_post_result broke the respective parse paths; removed the four full method implementations the thin codec delegates silently shadowed; added AST structural meta-tests to catch this class of drift

v0.7.2 (2026-09-03)

Bug Fixes

  • Stray self in the _parse_infer_response codec signature: after the 0.7.0 refactor every infer() / infer_batch() call raised TypeError; added a regression test driving a real InferResponse

v0.7.1 (2026-09-03)

  • Version aligned with setup.py; no API changes

v0.7.0 (2026-09-02)

New Features

  • App developer toolkit (developed as 0.6.0, never released separately, shipped with 0.7.0):

    • Frame.crop() / Frame.resize() / Frame.to_jpeg_bytes() — crop, three resize modes (stretch/letterbox/crop), JPEG encoding; falls back to numpy/PIL when cv2 is missing

    • draw module — detection visualization (draw_boxes / draw_text / draw_detections)

    • recording module — pure-Python TS muxing, HLS segments, event pre-record buffer (TsWriter / HlsWriter / PrerollBuffer), no ffmpeg required

    • web module — MJPEG streaming (MjpegServer / MjpegStream / mjpeg_wsgi_app)

    • DeviceClient native autofocus group of 6 methods (start_oneshot_af / start_zoom_follow / get_autofocus_status / cancel_autofocus / set_af_windows / get_af_measurement)

    • CameraClient imaging/infrared/privacy-mask/OSD/config group of 12 methods (incl. IR preset management, get_osd symmetric with set_osd)

  • Fixed InferenceClient.subscribe() silently dropping failed frames

Refactoring

  • Internal layering: extracted the shared _transport primitives; split media into frame/encoded/fd_client, dsp into dsp_wire/dsp_format, inference into types/codec/genai. Facades keep every historical import path — zero public API removals

  • DSP CPU fallback now raises a UserWarning instead of a silent log

  • 2D-array grayscale inference input emits a DeprecationWarning (pass fmt= explicitly)

  • EncodedStreamClient() default socket path is now /run/aipc/encoded/{stream}.sock (ENCODED_SOCK_DIR env override; explicit paths still win)

Fixes

  • Fixed an import crash on Python 3.8/3.9 (from __future__ import annotations everywhere)

Other

  • Added py.typed marker to the package; test suite grown to 270 green tests

v0.5.0 (2026-08-20)

Refactoring

  • Rebrand: hailo_ipc_sdk → neoruntime_ipc_sdk (breaking, no shim; functional names such as hailort/ne503/aipc are preserved)

Fixes

  • Fixed dual-layout decoding of audio_capture video-layout headers

Other

  • CI gained TestPyPI/PyPI trusted-publishing dual release flow

v0.4.0 (2026-07-14)

New Features

  • DeviceClient — 7 new methods: - set_lens_limits(zoom_limit, focus_limit) — Set lens axis position limits - oneshot_autofocus(timeout) — One-shot autofocus (composite: enable → wait for convergence → disable) - set_wiegand_out(channel, enable) — Wiegand output control - get_wiegand_out(channel) — Wiegand output state query - rs485_init(baudrate, config) — RS-485 serial initialization - rs485_deinit() — RS-485 serial deinitialization - rs485_tx(data) — RS-485 data transmission

  • AppClient — 1 new method: - restart_app(app_id, timeout_seconds) — Restart application (stop + start)

Improvements

  • Updated API docs with DeviceClient lens limits, Wiegand, and RS-485 examples

  • Updated API docs with complete AppClient usage examples

  • Synchronized Chinese and English documentation

v0.3.0 (2026-05-18)

New Features

  • Added segmentation mask support (SegmentationMask with RLE decoding via to_numpy_mask())

  • Added OCR result support (OcrLine with text, confidence, bounding box)

  • Added embedding vector support (Embedding for CLIP and similar models)

  • Added depth map support (DepthMap with float32 numpy array)

  • Added CLIP text encoding via encode_text() — encode text strings to NPU embeddings

  • Added runtime postprocess config update via update_postprocess_config()

  • Added GenAI (LLM/VLM) support: - genai_create_session() — create LLM or VLM sessions with HEF models - genai_generate() — stream generated tokens with sampling parameters - genai_abort() — abort ongoing generation - genai_destroy_session() — clean up session resources

  • Added LoRA adapter support for GenAI sessions (lora_name parameter)

  • Added memory optimization option for GenAI tokenization (optimize_memory)

Improvements

  • Enhanced subscribe() with raw_output_only parameter for raw tensor streaming

  • Enhanced register_model() with model_type, model_variant, and explicit inputs/outputs specs

  • Improved tensor type handling with full dtype mapping (uint8, int8, uint16, int16, float16, float32, int32, uint32)

Bug Fixes

  • Fixed tensor shape mismatch when output shape does not match data size

  • Fixed InferenceResult missing new result types (masks, ocr, embeddings, depth)

v0.2.0 (2026-03-02)

New Features

  • Added plugin system support (PluginDiscovery, PluginServer)

  • Support for plugin capability discovery and gRPC service invocation

  • Added video stream access (now provided by FdMediaClient / EncodedStreamClient)

  • Support for raw and encoded video stream retrieval

Improvements

  • Optimized InferenceClient performance

  • Improved event bus wildcard matching

  • Enhanced error handling and logging

  • Updated protobuf to 4.21.0

Bug Fixes

  • Fixed memory leak in EventClient subscription

  • Fixed DeviceClient GPIO control issue

  • Fixed connection pool issue in multi-threaded environments

v0.1.0 (2025-12-15)

Initial Release

  • InferenceClient: AI inference client

  • EventClient: Event bus client

  • DeviceClient: Device control client

  • Config: Configuration management

  • Support for Python 3.8+

  • gRPC-based communication