TinyTitan 5.17

TinyTitan 5.17 — the katcoder install works again

This release fixes the converter defect that made every fresh KAT-Coder-V2.5-Dev install fail (issue
#19), makes an interrupted conversion stop promptly and cleanly instead of ending in a crash report,
and moves the user documentation into the wiki.

What is new

• tools/install_models.sh katcoder and katcoder both install again. KAT is the one
  checkpoint in this family that ships its routed experts per tensor, and the fusion that stacks
  them guarded against a repeated tensor by the tensor's name alone — while the converter adds one
  tensor once per requested width, and this family is always converted with --bits 4 8. The second
  width therefore looked like a duplicate, and the conversion stopped on the first routed expert
  with duplicate source tensor model.language_model.layers.N.mlp.experts.0.down_proj.weight. The
  guard is keyed by (width, name) now. Checked by converting the pinned checkpoint for real:
  shard model-00000, which carries 2,075 per-expert routed tensors, converts at both widths and
  writes the fused switch_mlp.{gate,up,down}_proj stacks with a 256-expert axis (1.75 GB at 4-bit,
  2.82 GB at 8-bit, 27 tensors each), then continues to the next shard with no error. Installs
  already under models/ were never affected: the defect was in conversion, not in the runtime. The
  full install and the §4 verification of docs/adding-a-model.md remain open as TT-048.
• An interrupted conversion stops when you stop it. Ctrl-C, a pkill or a dying parent during
  the ~69 GB fetch now ends in under a second, with no orphan curl and no crash report. Two
  defects were involved: the fetcher threads were daemon threads left alive at interpreter
  finalization (CPython 3.14 aborts there and files a macOS crash report; 3.13 only printed the
  fatal error), and a chunk retry could sit in a wait of up to two minutes without noticing the
  stop, so even a pool that was joined took ~20 s to die. Checked by
  benchmark/test_prepare_agentworld.py — twelve tests, of which the shutdown test reports three
  live fetcher threads against the previous commit — and by interrupting the real download: the
  converter is gone in 0 s, no curl is left, the log carries no fatal error and no crash report is
  written. A stopped download keeps its shards and resumes from them.
• The user documentation is in the wiki. The README is the front page again: what the repository
  delivers, one worked example, the benchmark table, the supported models and the links. The guides
  — getting started, cookbook, installation and configuration, runtime controls, the OpenAI-
  compatible server, the FAQ — are wiki pages, so each is found in one place instead of two. Wiki
  links to repository files are absolute GitHub URLs, because the wiki is a separate repository.

Also in this release

• The Qwen3.5-MoE converter has a test suite for the first time:
  benchmark/test_prepare_agentworld.py pins the per-expert naming contract, the both-widths
  fusion, expert placement by index rather than by arrival, per-width emission, the incomplete-layer
  refusal and the shutdown behaviour. It runs in the converter gate in CI, from a synthetic shard,
  with no checkpoint and no network.
• docs/t6-reply-check-offline.md records the side-engine reply-check task measured on 417 real
  recorded replies: 66.7% precision and 38.5% recall against a pre-registered 95%/50% gate, so the
  task is not wired. The harness, the hand-audited labels and the tests that pin the label traps
  ship with it as the record.
• The wiki's install page states the breakage honestly: the katcoder command was broken in
  5.4–5.16, there is no installer workaround on those builds (convert one width at a time with
  prepare_agentworld.py --bits 4), and a fresh full install is tracked as TT-048.

Performance

• No performance change and no new timings.
• This release touches the converter and the documentation, not the runtime; the README's benchmark
  table is unchanged and was not re-run.

Verification

• tools/lint.sh — all eleven gates clean.
• swift test --no-parallel — 1,527 tests in 234 suites.
• 7 of the 16 stored golden baselines compared byte-identical, then a clean scratch release
  build with a clean warning scan.
• Not checked, because their install is not under models/ and nothing may be fetched to change
  that: ornith-4, ornith-8, qwen38-8, agentworld-4, agentworld-8, katcoder-4,
  katcoder-8, qwen35-2b-4, qwen35-2b-8.

Checksum

tinytitan-5.17-macos-arm64.tar.gz sha256: e88102c7767fc91cd387f5340463005bfbda540e80c077368bd446c556180a23
tinytitan-5.17-macos-arm64.tar.gz size: 15877987 bytes
tinytitan-lib-5.17-macos-arm64.tar.gz sha256: b4bf4be5c9839c4070317d4973ed15f5a7c21c34232736046156f88d2ee674d2
tinytitan-lib-5.17-macos-arm64.tar.gz size: 17033554 bytes