Releases: lloyal-ai/lloyal.node
Releases · lloyal-ai/lloyal.node
v1.1.0
Full Changelog: v1.0.9...v1.1.0
v1.0.9
Full Changelog: v1.0.8...v1.0.9
v1.0.8
Synced with upstream: liblloyal and llama.cpp
Full Changelog: v1.0.7...v1.0.8
v1.0.7
lloyal.node v1.0.7 — First Stable Release
The first stable release of @lloyal-labs/lloyal.node. Ships the Branch API for covalent generation and prebuilt binaries across 13 platform/GPU combinations.
Branch API
New Branch class for forkable inference. Each branch owns a KV cache sequence, sampler chain, logits snapshot, and perplexity tracker. Under unified KV, forking is a metadata-only operation — no KV tensor buffers are copied.
Key methods:
- Branch.create() / fork() — create root or fork from any branch (root or intermediate) to build trees
- produce() / commit() — two-phase protocol: sample without KV write, then commit to advance
- prune() — discard loser branch and its divergent KV entries
- destroy() — release handle but keep KV (for winners continuing with raw ops)
- reseedSampler() — unique PRNG per fork for stochastic diversity
- perplexity — rolling PPL per branch for quality-based selection
Updated Examples
- best-of-n — Rewritten with Branch API. Parallel streaming generation with perplexity selection.
- speculative — Rewritten with Branch API. Draft/verify with fork and prune.
- streaming-summary — New. Infinite context generation via BlinkKV reseeding with sidecar summarization model and ghostwritten progress sinks.
Other Changes
- liblloyal upgraded to v1.0.4 (Branch API, DEFAULT_N_BATCH fix, slot reuse leak fix)\
- Removed vestigial APIs from public surface
Platforms
Prebuilt binaries for macOS (arm64 Metal, x64), Linux (x64/arm64 × CPU/CUDA/Vulkan), Windows (x64 × CPU/CUDA/Vulkan, arm64 × CPU/Vulkan).