v0.3.28
feyninc/pulpiev0.3.28Oct 5, 2026by github-actions[bot]
Upgrading?
AI Summary
Introduces foundational training infrastructure with the new laya.train module and CLI, along with a router registry for custom checkpoints. It also updates the Docker base image and includes fixes for packaging, serving, and various language SDKs.
Key Highlights
- New laya.train module and CLI for fine-tuning shared across notebooks and scripts
- Router registry for custom checkpoints alongside built-ins
- Docker base image updated to Debian 13 (trixie)
- Feature: predict_tournament for choice questions past the option budget
- Packaging improvements including inference backends in wheels
New Features
- laya.train module and CLI
- Router registry for checkpoints
- predict_tournament feature
- Debian 13 base image
Full Release Notes
## What's Changed * fix(serve): refuse unpublished path-like model ids instead of auto-routing by @sathariels in https://github.com/NandhaKishorM/laya/pull/930 * fix(finetune): the Apple Silicon and research scripts fit temperatures outside the runtime clamp by @Bruce-Yii in https://github.com/NandhaKishorM/laya/pull/851 * feat(train): add laya.train, one fine-tuning loop for the notebook and scripts to share by @GuilhermeFusari in https://github.com/NandhaKishorM/laya/pull/899 * docs(router): explain why predict_batch composes default hooks by @sathariels in https://github.com/NandhaKishorM/laya/pull/929 * feat(train): add laya-train CLI and CSV/expected dataset loaders (#887) by @Swatantra-66 in https://github.com/NandhaKishorM/laya/pull/931 * fix(train): warn when the fine-tune's calibration rests on too little by @GuilhermeFusari in https://github.com/NandhaKishorM/laya/pull/933 * fix(train): skip questions whose options run past max_len instead of crashing the batch by @GuilhermeFusari in https://github.com/NandhaKishorM/laya/pull/934 * fix(ts): forward per-call hook options through predictBatch by @sathariels in https://github.com/NandhaKishorM/laya/pull/936 * feat(router): a registry of checkpoints beside the built-ins by @alandefreitas in https://github.com/NandhaKishorM/laya/pull/937 * fix(ts): clean French mail like Python by @kevin9327 in https://github.com/NandhaKishorM/laya/pull/939 * fix(packaging): include inference backends in wheels by @NikitaaRamesh in https://github.com/NandhaKishorM/laya/pull/940 * fix(train): normalize partial accumulation windows by @NikitaaRamesh in https://github.com/NandhaKishorM/laya/pull/941 * fix(java): keep empty action probabilities finite by @NikitaaRamesh in https://github.com/NandhaKishorM/laya/pull/942 * fix(common): support torch 2.0-2.3 in the action-head dtype check by @thread3d in https://github.com/NandhaKishorM/laya/pull/945 * test(compile): guard the duck-shaping checks on a torch without the config by @thread3d in https://github.com/NandhaKishorM/laya/pull/946 * feat(setup): a macOS/Intel checkout setup and verification path by @thread3d in https://github.com/NandhaKishorM/laya/pull/948 * test(model): skip the bf16 legs on a Windows CPU, which SIGILLs by @thread3d in https://github.com/NandhaKishorM/laya/pull/949 * feat(shortlist): add predict_tournament for choice questions past the option budget by @Rish-it in https://github.com/NandhaKishorM/laya/pull/950 * feat(common): parallel option layout for order-invariant decisions by @sharath-sms in https://github.com/NandhaKishorM/laya/pull/951 * build(docker): default the base image to Debian 13 (trixie) by @somtri in https://github.com/NandhaKishorM/laya/pull/953 * fix(evals): refuse a NaN --min-accuracy or --max-ece like the --min pairs by @sameedkhan17 in https://github.com/NandhaKishorM/laya/pull/954 * fix(ts): count an English/foreign collision word once, like Python by @sameedkhan17 in https://github.com/NandhaKishorM/laya/pull/955 * fix(dotnet): match French device footers with normal spacing by @sameedkhan17 in https://github.com/NandhaKishorM/laya/pull/956 * fix(calibrate): fit abstention cuts on the 4-decimal confidences the gate reads by @sameedkhan17 in https://github.com/NandhaKishorM/laya/pull/957 * fix(mcp): keep and validate a question's option_order by @sameedkhan17 in https://github.com/NandhaKishorM/laya/pull/958 * fix(onnx): catch an in-place question rewrite in the scan-budget guard by @sameedkhan17 in https://github.com/NandhaKishorM/laya/pull/959 * fix(structured): resolve local $defs refs so pydantic Enum fields become choices by @sameedkhan17 in https://github.com/NandhaKishorM/laya/pull/960 * fix(ci): ignore false positive sha256 checksums in gitleaks (#961) by @Swatantra-66 in https://github.com/NandhaKishorM/laya/pull/962 ## New Contributors * @GuilhermeFusari made their first contribution in https://github.com/NandhaKishorM/laya/pull/899 * @alandefreitas made their first contribution in https://github.com/NandhaKishorM/laya/pull/937 * @kevin9327 made their first contribution in https://github.com/NandhaKishorM/laya/pull/939 * @sharath-sms made their first contribution in https://github.com/NandhaKishorM/laya/pull/951 * @sameedkhan17 made their first contribution in https://github.com/NandhaKishorM/laya/pull/954 **Full Changelog**: https://github.com/NandhaKishorM/laya/compare/v0.3.27...v0.3.28
Related Repositories
f.k.a. Awesome ChatGPT Prompts. Share, discover, and collect prompts from the community. Free and open source — self-host for your organization with complete privacy.
Open-source foundation of ToolJet AI - the enterprise app generation platform for internal tools, dashboards, business applications, workflows and AI agents. Build visually, from a prompt, or from Claude Code, Codex and Cursor over MCP 🚀
1,324-exercise fitness dataset — animation GIFs, 180×180 thumbnails, muscle-group & equipment data, and step-by-step instructions in 6 languages. The exercise data layer behind the LogPress app.
A complete web-based remote monitoring and management web site. Once setup you can install agents and perform remote desktop session to devices on the local network or over the Internet.
A general purpose HTTP reverse proxy and forwarding tool. Now written in Go!