This page summarizes user-visible changes. It separates work on main from
published releases and links to the full engineering ledger for exhaustive
implementation detail.
kiln-v0.5.2 — 2026-08-01
First-class OpenEnv RL, fast stable serving, and the new dashboard
- Added native, reproducible OpenEnv discovery, rollout, replay, GRPO training, held-out evaluation, provenance, artifact verification, CLI, API, telemetry, and dashboard workflows for any protocol-compatible environment.
- Made Stable the normal complete product across CUDA, ROCm, Metal, and Vulkan, including inference, training, adapters, memory management, and qualified graph acceleration. Experimental is no longer required for normal speed or training.
- Added the new embedded ember dashboard with guided training and distillation on-ramps, contextual next actions, training-time hints, loss/run visibility, bundled fonts, and resilient empty/error/retry behavior.
- Vulkan execution derives route legality from device capabilities rather than device names or IDs. The tracked source receipt measures 13.46 decode tok/s at 74.29 ms mean inter-token latency on its recorded workload.
kiln-v0.5.1 — 2026-07-30
Vulkan recovery and a usable documentation path
- Restored the generic Vulkan compute routes that a broad portable fallback had disabled, recovering the tracked single-stream decode workload from 0.142 tok/s to 13.43 tok/s.
- Kept dispatch independent of qualification-machine identity. The
capability-derived policy now on
mainsupersedes the release's global route table. - Reworked the documentation around task-focused guides, a searchable reference library, an evidence-bound benchmark explanation, and a demo that distinguishes live UI from seeded or historical material.
kiln-v0.5.0 — 2026-07-29
Bounded thinking and typed startup policy
- Added token and wall-clock thinking budgets with explicit request, server default, streaming, and telemetry behavior.
- Made public environment overrides mechanically derive from typed configuration fields and rejected removed or conflicting configuration authorities.
- Expanded effective-configuration, runtime-policy, provenance, checkpoint, teacher, and artifact identity contracts.
- Moved CUDA, Metal, ROCm, Vulkan, batching, streaming-prefill, and training decisions toward immutable typed startup policy instead of hot-path environment reads.
kiln-v0.4.1 — 2026-06-12
Multi-turn prefix-cache repair
- Restored reusable prefix snapshots for multi-turn ROCm traffic.
- Added the same split-snapshot behavior to non-batched serving paths.
- Prevented an oversized entry from evicting useful cache contents when that entry could never fit.
kiln-v0.4.0 — 2026-06-11
Managed pi runs and Muon training
- Added server-managed pi runs with queued execution, event feeds, steering, follow-up, abort, persistence, and trace ingestion.
- Added the dashboard's Distill → Agent runs workflow.
- Made Muon the default optimizer for supported training routes, with AdamW and SGD still selectable per request.
kiln-v0.3.5 — 2026-06-09
Task-first dashboard workflows
- Added in-place dataset upload and training flows, post-training eval setup, adapter-focused comparisons, completion notifications, correction capture, and the embedded pi terminal.
- Improved first-run recovery and made adapter, health, and flywheel state agree after swaps or failures.
Older releases and exhaustive detail
Historical entries preserve the terminology and evidence available when each change landed. Use current product guides for current configuration and support claims.