Livt packages

Livt.ML

Version 1.6.1

Fixed-size quantized inference, tensor storage, attention, and transformer building blocks for Livt.

CHANGELOG.md

Changelog

All notable changes to this package are documented here.

1.6.1 - 2026-11-08

Added

  • Publish package discovery metadata, the README, the changelog, and the MIT license with this version.

Changed

  • Update dependencies to Livt.Math 1.7.2 and Livt.IO 1.8.1. Existing component APIs remain unchanged from 1.6.0.

1.6.0 - 2026-11-07

Added

  • Linear.Int8MatrixAccumulator32<ACTIVATION_BITS, PRODUCT_STAGES> for 64-column INT8 matrix chunks with signed int32 activations and up to 1,024 retained BRAM-backed int64 sums. Supports bias seeding and checked updates.
  • Linear.Int8MatrixAccumulator32Stream for ordered ready/valid row batches over the same native core. Eight multiply lanes overlap row feeding, reduction and BRAM commits without a second 512-bit weight-row buffer.
  • Checked activation ranges from 2 to 32 bits and one- or two-stage product pipelines with identical arithmetic. Overflow and source failures invalidate retained sums until explicit recovery.
  • Public Livt regressions and a matrix contract documenting packing, ownership, acceptance, completion, cancellation, reset and failure recovery.

Compatibility

  • Existing APIs and defaults are unchanged. Dependencies remain Livt.Math 1.7.0 and Livt.IO 1.3.0.
  • Serialize scheduled calls. A successful batch returns only after every row has committed. The producer must retain each valid row until its transfer edge and report unavailable sources through streamError; there is no implicit timeout.
  • The native core uses synchronous active-high reset. Reset producer and consumer together across a rising clock edge, then call Begin before use. RAM payload is not reset. Model layout, memory transport, scaling and rounding remain caller responsibilities.

1.5.0 - 2026-11-06

Added

  • Attention.AttentionAccumulator32<PRODUCT_STAGES> for exact 32-element Q.K scores and retained probability-times-V numerators using the checked Math core.
  • Attention.AttentionAccumulator32Port<PRODUCT_STAGES> for live 512-bit input rows, avoiding full-row copies through scheduled argument frames.
  • Ordered block-pair validation, query invalidation, partial-score checks and poisoned weighted-sum recovery. Results carry the checked int64 sum and a failure flag together; reset cancels pending work.
  • One- and two-stage product configurations with identical arithmetic, and a generated-wrapper regression for the paired-row API.

Changed

  • Update the Livt.Math dependency from 1.3.0 to 1.7.0 for the checked block arithmetic primitive. Livt.IO remains pinned to 1.3.0.

Compatibility

  • Existing INT8/16-bit kernels and their defaults are unchanged. The new attention components require exactly 32 signed int32 values, supplied as two little-endian sixteen-lane blocks.
  • Score scaling, softmax, final division, fixed-point rounding and external memory layout remain the caller's responsibility. Serialize operations; live-row inputs must remain stable until the scheduled call returns.

Earlier releases

Earlier release history remains available in Git; the preceding tagged version is 1.4.1.