Repository navigation
Release v0.6.1
Neat Core 0.6.1 is a maintenance release for the C++ and Python application framework. It improves tensor transfer memory use, fixes model input and output handling, and allows asynchronous runs to resume after an idle period.
Tensor transfers and memory
Host-to-device transfers reuse DMA buffers, including nearby tensor sizes, instead of repeatedly allocating storage. The shared transfer cache is limited to 256 MiB and 128 one-buffer pools. Eviction releases idle cached storage while tensors still held by applications remain valid. Busy pools do not block other transfers; requests that do not fit the cache use uncached storage.
The update also fixes ownership and cleanup when a transfer fails or throws an exception. The cache limit applies to retained cache capacity, not to all live memory an application can hold. See #1274 and #1322.
Model input and output fixes
- Typed MLA outputs: affected model packs now return their native tensor views, including dtype, shape, strides, and offset, instead of one raw
UInt8carrier. This corrects cases where packed signed INT8 outputs were exposed as unsigned bytes. Views can share storage and be non-contiguous. Applications that manually split the old carrier must consume the returned views instead. In Python,to_numpy(copy=True)can produce a separate array when needed. See #1285. - Complete input batches: native MLA input validation now checks the full compiled batch span rather than only its first image. Each model input must still contain the complete batch; separate per-image pushes are not a substitute. See #1296.
- Legacy decoded YOLO heads: model packs without a decoder contract can explicitly select the existing
BoxDecodeType::Yoloprofile for already-decoded coordinates. Raw YOLOv6 heads continue to useYoloV6. Existing decoded packs do not need recompilation for this selection fix. See #1281. - Diagnostics and preprocessing guidance: execution reports retain native BoxDecode settings. Documentation clarifies that preprocessing
Offpreserves required model Cast/Quant/Tess adapters and the public input contract. See #1282 and #1286.
Idle runs and timeout recovery
An application-fed asynchronous Run can pause after completing its work and accept more input later. Output silence alone no longer permanently fails the Run. A finite pull() timeout ends that wait without stopping the Run; pull(-1) remains unlimited.
Failed synchronous convenience calls discard cached runners so a late result from a timed-out call cannot satisfy a later call. Backend errors and existing teardown protections remain active. Applications that need silent-stall detection should apply a deadline while requests remain pending; a timeout does not prove that device work has retired. See #1288.
Video and offline installation
- Raw
VideoSenderinput no longer has Core's fixed 3840 × 2160 ceiling. The codec backend still determines supported resolution and frame-rate combinations; this change alone does not qualify full-resolution camera encoding. See #1299. SimaEncoderejects unsupported H.265 level 4.2 before execution. H.264 level 4.2 remains accepted. See #1295.- Offline bundles rebuild when their dependency inputs change, even if the Core source commit is unchanged. New resource archive names include a checksum suffix; metadata-driven downloads and older archive names remain supported. See #1270.
Compatibility and validation
The package version is 0.6.1; Core remains on ABI 6 (libsima_neat.so.6). Install matching runtime, development, Python, and lower dependency packages. Review the typed-output and timeout behavior changes above when updating applications.
The documented pairing is SDK and Model Compiler 3.0.0.1 with DevKit software 3.0.0. SDK hosts require Docker Engine 20.10.10 or later. See the compatibility guide.
The release branch build was still running when this draft was prepared. The new tag build must complete before its result can be reported. The linked fixes contain focused host and device evidence, which does not establish exhaustive accuracy, performance, or hardware qualification for the final release package set.