|
Download notes/simplify-reports/W9-speed-knobs.md from pablovela5620/zipdepth-promptda: direct link, hf CLI and curl.
- Browser
- Download file 2.23 kB
-
https://huggingface.co/pablovela5620/zipdepth-promptda/resolve/main/notes/simplify-reports/W9-speed-knobs.md
- Command line
-
hf download hf://pablovela5620/zipdepth-promptda/notes/simplify-reports/W9-speed-knobs.md
-
curl -L -o W9-speed-knobs.md https://huggingface.co/pablovela5620/zipdepth-promptda/resolve/main/notes/simplify-reports/W9-speed-knobs.md
2.23 kB
W9 — expose the training-speed knobs the hill-climb needs, as Tyro flags on zipdepth-train-catalog, on branch zdp/8-train-fast
(already created from zdp/7-eval-tools and checked out; the idiomatic dataloader from W8 lands later on this branch — do not touch
zipdepth/catalog/dataset.py or the dataset construction beyond what a flag needs). Worktree:
/home/pablo/.paseo/worktrees/26sr02x6/dirty-swan. REPORT: /tmp/zdp-recut/apply/W9-REPORT.md. Rules: python-conventions
(PEP 526, jaxtyping, no Any, names never carry dtype/shape), AGENTS.md, no vendored edits, loss/architecture/eval unchanged.
APPLY (each default-off or default = today's behaviour so the baseline recipe is bit-for-bit today's):
--compile-mode {off,default,reduce-overhead,max-autotune}(torch.compile of the student for training; today's behaviour stays the default; document that beartype dev mode disables it),--channels-last(model + batches to channels_last memory format),--fused-adamw(torch.optim.AdamW(fused=True)),--amp-dtype {bfloat16,float16}if not already a flag.--max-lralready exists; add--warmup-pctand--final-div-factorpassthroughs to the OneCycle schedule, and--schedule {onecycle,constant-cooldown}where constant-cooldown = constant lr then linear cooldown over the last--cooldown-pctof total_steps.- Progressive resolution:
--stage-schedule "384x512:0.3,768x1024:0.7"(fractions of total_steps per resolution; the loader rebuilds at a stage boundary; checkpoints carry the stage). Skip if the idiomatic dataset cannot change resolution mid-run without a restart — say so; the orchestrator can then run stages as separate resumed arms. --preset {v4,fast}: named recipes that set the above (v4 = today's; fast = filled in after the hill-climb) — the winning recipe lands here later.- Tests: each flag has a unit test proving it changes the constructed model/optimizer/scheduler and that the default equals
today's objects (bit-equal state_dict / same class + kwargs). Gates: zipdepth-catalog-dev lint typecheck deadcode tests;
the live smoke
zipdepth-catalog-train-smokemust pass with--compile-mode default --channels-last --fused-adamw.