Order averaging (Prime, after the 739aa03 spike):
- A decision may set orderings: rotations|all (all only for <= 4 options). Every
ordering goes to the engine in one shared batch.
- The reply keeps each native result and adds combined {probabilities (log-mean),
top, agreement, spread}.
- Through the service on SemIf's labelled sets (252 rows): 78.6% -> 88.1%
(group-bootstrap 95% CI +5.1..+14.3). Unanimous agreement is 94.5% accurate.
Fast kernels: flash-linear-attention 0.5.2 and causal-conv1d 1.7.0 are now the
default build. A/B on the empty GPU 3:
- parity with upstream went from 142/144 to 144/144;
- a ~2k-token /decide went from 169 to 92 ms server-side;
- short 3-rotation batches cost ~3-6 ms more.
triton builds a C shim at runtime, so the image carries gcc. Without it the
warm-up failed and startup failed closed.
Heid bug-hunt panel (4/4 arms, thread 01M3H3F4RR7XBP90KQ3A39H4SX), folded:
- Startup validation: VRAM cap 0 no longer means uncapped (C1); limits must be
>= 1 (S1); the token must be visible ASCII (S2); the calibration file must
exist and parse, with T in [0.05, 20] (S8, and C3's NaN leg).
- The body limit is checked before a chunk is kept, and a Unicode-digit
Content-Length no longer crashes (C2, S3).
- Failures while building the response now get the 500 envelope (C3).
- 429 busy past SEMIF_MAX_QUEUE requests in progress (C6).
- The engine releases memory on every non-validation failure, unchained after
gc; an empty OOM message is handled; 'out of memory' RuntimeErrors map to 503
(C4, C5, S9).
- The entry point forces HF_HUB_OFFLINE (S10). README wording fixed (S5, S6).
- New guard tests close the gaps the arms' mutation grids exposed: early stop of
the body read, a shared-route lock, calibration pass-through, the gc cycle,
the exact caps, TorchEngine.load's arch and device checks, and the offline
entry point.
86 tests.
Deployed on fv-ml1 GPU 1: parity 144/144, OOM and burst release verified, shared
capacity 63/51/26/16 rows at ~140/520/1960/3900 prefix tokens.
Latency, measured from nh3-dev (3 runs x 20 per condition; network floor 31 ms):
- /decide short: 71 ms end to end, 38 ms server-side;
- /decide with a ~2,000-token state: 210 / 169 ms;
- shared, 3 rotations: 113 / 79 ms;
- shared, 6 orderings: 137 / 99 ms.
Qwen3.5's fast kernels (causal_conv1d, flash-linear-attention) are not installed,
so transformers falls back to its reference PyTorch paths. That is a speed lever,
and using it needs a parity re-check.
Averaging over option orderings, on SemIf authored144 + perturbations108 (252 rows,
72 groups):
- a single ordering scores 78.6%;
- log-mean over the 3 rotations scores 87.7% (+9.1 pts, group-bootstrap 95% CI +4.7
to +13.8);
- all 6 permutations score 88.1%.
Rotations capture nearly all of the gain. Rows where the rotations agree unanimously
(161) are 94.4% accurate; split rows (91) are 75.8%.
vh