jungfrauandClaude Fable 5.1 20fb1d03fb spark cycle trains several model sizes in turn and times each one the way the server calls it ...

MODELS="n s m l" runs the promoted recipe with yolo26n, s, m and l one after the other;
only the nano run may promote itself, the others are candidates until their speed at the
beamline is known. Each training waits for MEM_FLOOR_GB of free memory, an out-of-memory
run is tried again at batch 8, and vLLM is restarted at the end only if it was running.
tools/bench_latency.py measures the answer time per frame as the live server asks for it
(frame scaled by infer_scale, letterboxed to 640, PyTorch fp16) and writes runs/<id>/latency.json.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

2026-10-22 11:38:01 +02:00

..

__init__.py

training loop: run ledger, studio export importer with an image pool, train_run with gold gate and auto-promotion, SLURM template, CI

2026-09-28 16:06:06 +02:00

bench_latency.py

spark cycle trains several model sizes in turn and times each one the way the server calls it

2026-10-22 11:38:01 +02:00

compare_unseen.py

gate: bootstrap by mount, light states and excluded frames of the gold set

2026-10-02 15:46:31 +02:00

dedup_study.py

dedup study: suppressing nested duplicate outlines on the unseen gold frames

2026-10-05 22:26:18 +02:00

gate_stats.py

gate: bootstrap by mount, light states and excluded frames of the gold set

2026-10-02 15:46:31 +02:00

gold_groups.py

gate: bootstrap by mount, light states and excluded frames of the gold set

2026-10-02 15:46:31 +02:00

import_studio_export.py

training loop: run ledger, studio export importer with an image pool, train_run with gold gate and auto-promotion, SLURM template, CI

2026-09-28 16:06:06 +02:00

ledger.py

training loop: run ledger, studio export importer with an image pool, train_run with gold gate and auto-promotion, SLURM template, CI

2026-09-28 16:06:06 +02:00

measure_weights.py

gate: baseline and candidate measured in the same environment

2026-10-02 13:25:45 +02:00

spark_cycle.sh

spark cycle trains several model sizes in turn and times each one the way the server calls it

2026-10-22 11:38:01 +02:00

train_run.py

gate: bootstrap by mount, light states and excluded frames of the gold set

2026-10-02 15:46:31 +02:00