Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
15 changes: 15 additions & 0 deletions docs/hw-gate.md
Original file line number Diff line number Diff line change
Expand Up @@ -386,6 +386,21 @@ What the goldens record at the release pin:
- On `varied-1620w`, the cached path answers differently from a fresh request at a near-tie: top-2 margin 0.19, a WARN.
- ZAYA1-8B and Qwen3.8-27B answer every case.

The `qwen3.8-27b` row was re-recorded on 2026-11-16, off the release pin by orchestrator decision (UNSLOTH-UD31).
The reason is that the box moved to the file `models/pins.json` has named since 2026-11-12: unsloth's UD3.1
`Qwen3.8-27B-UD-Q4_K_XL.gguf`, sha256 `2a5f3565…`, 17,659,407,264 B. The row was previously recorded on the
pre-UD3.1 file, `3f227079…`, 17,559,178,144 B.

It was recorded at llama.cpp 148e7f329 / hrx-system 7f423692, with the private kernels at gpu-kernels aa5f0e90b. A
check run straight after it gives:

- kat 9/9 with 0 warnings, and every answer's text identical to the old file's goldens (only the logprobs moved);
- KLD mean 0.000001, p99 0.00004, same top 100 %.

The A/B measured the new file against the old base: mean KLD 0.0064, p99 0.065, same top 96.2 %. That is inside
the row's own `limits` in `manifest.json` (`max_mean_kld` 0.01, `max_kld_p99` 0.2, `min_same_top_p` 95.0), and it
is the size of the re-quantization, not a kernel change.

The MoE and IQ rows (`qwen3-coder-30b-a3b`, `qwen3-4b-iq4xs`) were recorded on 2026-10-20 at
llama.cpp 00f2a68d / hrx-system 82804197 (`manifest.json` `recorded_at`; a submodule whose pin could not be read is
recorded as `null` and named on stderr, never as the string "unknown"). Both answer every kat case.
Expand Down
Loading
Loading