Kiln Desktop desktop-v0.2.16

desktop-v0.2.16 — 2026-05-15

Aligned cut with kiln-v0.2.16 server. No desktop source changes since
desktop-v0.2.15; this release re-pins the bundled server payload to the
latest server cut so installers ship the new eval system, the Phase 2
Vulkan training hardening, the full Qwen3.5 stochastic sampler, and the
Playground overhaul.

Server payload

This cut ships with the kiln-v0.2.16 server binary. Highlights:

• Phase 11 — Evals as a first-class peer of training: end-to-end eval
  system with pluggable scorers (exact_match, contains, regex,
  json_validity, multiple_choice, numeric_tolerance, tool_call,
  code, llm_judge, all / any composites), dataset → suite synthesis,
  post-training auto-eval, head-to-head adapter compare, and an A/B
  JudgmentStore. The embedded /ui gains a Cmd-K command palette, eval /
  training / adapter detail modals with live progress, loss curves with
  downsampled history, the A/B compare playground with keyboard shortcuts,
  a VRAM donut, and a tok/s sparkline. New /v1/eval/* and /v1/judgments/*
  HTTP surface, kiln-eval CLI binary, [eval] config section.
• Phase 2 — Vulkan training hardening (Strix Halo / unified-memory APUs):
  on-device AdamW (adamw_step_{f32,bf16}.comp, decoupled weight decay)
  alongside the BF16 SGD path; lazy candle-CPU-storage sync removes the
  per-step GPU→CPU readback; VulkanLinearOp (CustomOp1) with autograd
  parity and in-op chunking via KILN_VULKAN_LINEAR_MAX_GFLOP;
  sdpa_prefill_f32.comp replaces the old flash_attn.comp placeholder;
  FLCE auto-engages at active_count ≥ 16 and KILN_USE_FLCE,
  KILN_VULKAN_LINEAR, and KILN_VULKAN_SDPA default ON; unified-memory
  APU VRAM-budget correction; HTTP 413 rejection with actionable hints for
  oversized training submissions.
• Sampling — full Qwen3.5 stochastic sampler: fused on Vulkan, Metal,
  and CUDA (zero vocab readback on Vulkan), new penalty fields, and
  Qwen3.5 recommended defaults everywhere.
• Playground overhaul: compare mode that streams both sides with
  reasoning, markdown rendering, inline edit-and-resend on user turns,
  per-turn badges + scroll lock + retry + transcript export, system-prompt
  and advanced-sampling controls, live <think> drain.
• Reliability: near-instant Ctrl+C shutdown; SSE parser hardened with a
  first-token >5 s hint; uptime cap only applies after the drain signal;
  default batching engine ON with KILN_BATCHING_ENGINE=0 opt-out.

See CHANGELOG.md for the full server changelog.