vh
f65e27b08f
The ~6.5-9 s first-call-per-bucket autotune lived in the container's writable layer and died on every recreate. 0.1.3 creates /tmp/triton-cache in the image owned by 10001 so the named volume intern-decision_triton-cache inherits a writable mount point, and compose mounts it. Bucket model PROVEN, not inferred: 2,048-token buckets, 16 up to 32,768. After one warmed call per bucket, 12 random sizes across 8k-32k were all warm (worst 2.09 s); cold entries cost 6.5-9 s. Full cold warm-up 109 s; warm re-run 17 s. scripts/intern-decision-warmup: one noul call per bucket, MAX_TOKENS from /health, two-point live calibration of the tokenizer's linear token model (a single probe overcorrects and the aim oscillates around the bucket edge), per-bucket wall times, non-zero exit on a missed bucket. Run it after an IMAGE CHANGE only; the volume carries ordinary recreates (measured: force-recreate, then a warmed 32k call answered in 2.11 s). Acceptance on 0.1.3: JevBench 202/231, hard 83/111, 0 diffs / 924; warm 32k GPU 1 peak 15,218 MiB (budget 15,220; a COLD autotune touched 15,224 once, README caveat); /decide answers. Artifacts in the acceptance dir.
/v1/systemone acceptance — 0.1.1 on fv-ml1, 2026-09-30
Live service: http://intern-decision.fv.internal:8033, image intern-decision-serve:0.1.1
(built in /opt/docker/src/intern-decision-serve-0.1.1; 0.1.0 kept for rollback).
JevBench (the pass/fail line)
JevBench v1.2.16 (5e95f23), its own typesafe adapter, over the 231 public items:
TYPESAFE_ENDPOINT=http://intern-decision.fv.internal:8033 TYPESAFE_API_KEY=$(secret get intern-decision/api-token) \
PYTHONPATH=/tmp/jevbench python3 -m jevbench.cli run --tasks datasets/public/{easy,original,hard}.jsonl \
--adapter typesafe --results ... --ledger ... --raw-dir ... --price-in-per-m 0 --price-out-per-m 0
- all 202/231, hard 83/111 — matches the expected numbers exactly.
- Row-by-row vs the bench's own runs r1..r4 (
bench-jev-2026-09-30/raw/out/intern-decision-4b-native/): 0 changed answers/probabilities across 924 rows. - Artifacts here: jb-results.jsonl, jb-summary.json, jb-ledger.jsonl, jb-manifest.json.
- The first-run finding: the response
modelmust be a string — the runner hashes it into its manifest and a dict crashed the manifest step. Fixed before the run above.
Controls (live)
- wrong token → 401; 17 questions → 422 ("never chunked");
images→ 422 "images not supported". - token limit at the boundary: 7,167 → 200, 7,169 → 422 (binary search; "one token over").
/decide/sharedpositive control: 2 decisions, 1 call, fields q1/q2 — unchanged.
Memory (GPU 1 is packed beside scriberr)
gpu1-peak.csv: per-process GPU 1 footprint, nvidia-smi at 0.2 s, container PID resolved via
docker inspect. Largest accepted request (16 noul questions, 7,167-token prompt) found by binary
search. Peak 9,866 MiB ≤ 9,876 budget (10 MiB inside).
0.1.2 (audit fixes, same day)
infra-ops audit PASS with two low findings, both fixed in 0.1.2:
- doc-only: the contract's response example showed
modelas an object; the wire returns the"<name>@<revision>"string. Example and prose aligned. images: []/nullnow count as ABSENT (only a non-empty value is 422) — a client that always sends the field must not be rejected for nothing. Live:[]→200,null→200,["a.png"]→422.
Re-run on 0.1.2 (jb2-* files, the 0.1.1 artifacts kept with their version suffix): all 202/231, hard 83/111, 0 changed rows across 924. 124 tests green. Rollback: 0.1.1.
0.1.3 — triton-cache volume + warm-up script (same day, Prime task)
Image change only (Dockerfile creates /tmp/triton-cache owned 10001; compose mounts the named volume there). Measured facts in the stack README "Cold-shape latency": bucket width 2,048 PROVEN, 16 buckets, cold full warm-up 109 s, warm re-run 17 s. Acceptance here:
- warmup-cold.txt / warmup-warm.txt: the two runs; every bucket in-bucket, cold 109.3 s vs 17.2 s.
- random-validation.txt: 12 random sizes 8.3k–31.8k after warm-up, worst wall 2.09 s — no cold call.
- Recreate survival: force-recreate, then a warmed 32k call answered in 2.11 s; volume holds 3,200+ files owned by uid 10001 (new files newer than the image build = the volume is in use).
- Memory: warm 32k peak 15,218 MiB (budget 15,220); a COLD autotune touched 15,224 once, during the pre-deploy 0.1.2-era measurement — see README caveat. gpu1-cold-run.csv, gpu1-32k-warm.csv.
- JevBench on 0.1.3: all 202/231, hard 83/111, 0 diffs / 924 (jb-results-0.1.3.jsonl).
- /decide positive control: answered (yes, 0.969).