Releases: lloyal-ai/liblloyal
v1.5.4
Full Changelog: v1.5.3...v1.5.4
v1.5.3
feat(legal): Apache 2 license everything
v1.5.2
Full Changelog: v1.5.1...v1.5.2
v1.5.1
What's Changed
- chore(deps): bump llama.cpp b8087 -> b8369 by @lloyal-research in #29
Full Changelog: v1.5.0...v1.5.1
v1.5.0
What's Changed
- chore(deps): bump llama.cpp b7899 -> b8087 by @lloyal-research in #23
- feat(api): replace raw pointer based sampler and grammar state with h… by @lloyal-research in #24
- refactor(branch): rename decode primitives to match intent by @lloyal-research in #25
- feat(api): support cache type configuration by @lloyal-research in #26
- Feat/lazy grammar by @lloyal-research in #27
- feat(branching): add fork head tracking by @lloyal-research in #28
Full Changelog: v1.4.0...v1.5.0
v1.4.0
What's Changed
- feat(decode): refactor decoding abstraction to include new algorithms by @lloyal-research in #20
- feat(branch): kv::tenancy — seq_id vacancy manager, topology, RESTRIC… by @lloyal-research in #21
- Feat/kv tenancy by @lloyal-research in #22
Full Changelog: v1.2.0...v1.4.0
v1.2.0
What's Changed
- feat(test): share model weights between tests by @lloyal-research in #15
- feat(tokenizer): add turn seperator support by @lloyal-research in #16
- fix(upstream): drop baggage use llama.cpp/common directly by @lloyal-research in #17
- feat(templates): migrate to llama.cpp common chat template by @lloyal-research in #18
- feat(chat): new chat api by @lloyal-research in #19
Full Changelog: v1.0.5...v1.2.0
v1.0.5
What's Changed
- chore(deps): bump llama.cpp b6870 -> b7899 by @lloyal-research in #14
Full Changelog: v1.0.4...v1.0.5
v1.0.4
What's New
The Branch API is a new high-level primitive that consolidates KV cache, sampler chain, grammar, metrics, logits, and logit bias into a single forkable handle. Create a branch, fork it to explore alternatives, compare by perplexity, prune losers — one API for the entire tree search loop.
Core API
-
Handle-based branch pooling — Branches are opaque uint32_t handles backed by a freelist pool. Generation counters prevent ABA bugs on slot reuse. No malloc churn during search.
-
Fork semantics — fork() clones KV cache sequence, sampler chain (penalties, PRNG, filters), grammar state, metrics trackers, logits snapshot, and logit bias. Each branch is fully independent after fork. steer_fn is intentionally not cloned (captures references, unsafe to copy).
-
RAII wrapper — Branch class provides move-only scoped ownership. Branches auto-prune on scope exit. release_kv() and release_full() extract owned state before destruction.
v1.0.3-alpha
fix(windows): use wrapper headers for cross-platform llama/ namespace
Fixes symbol redefinition errors on all platforms.
Previous approach (copying headers) caused duplicate definitions when
llama target exported its include paths to consumers.
New approach (wrapper headers):
- Template files in cmake/templates/*.in
- Generated at configure time via configure_file()
- Forward-include the originals without duplication
- BEFORE ensures wrappers are found first
Works on Windows (no symlinks), Unix (no conflicts), everywhere.
Full Changelog: v1.0.3-alpha...v1.0.3-alpha