--set-grid, -g | A grid axis, repeatable; several axes form a cartesian product. e.g. -g policy.lr=1e-5,2e-5 -g grpo.kl=0.01,0.05 (必填 · 可重复) |
--set, -s | A fixed override shared by every variant, repeatable (可重复) |
--profile | Target hardware and resources in one argument: h200 (the registry default shape), h200:4 (4 GPUs), h200:16 (2 full nodes). Repeat it to split pools per role (the heterogeneous extension point): —profile train=h200:8 —profile rollout=h100:2. Comma-separate roles that share a pool: —profile actor,rollout=h200:8 —profile teacher=h200:8 (可重复) |
--method, -m | Method identifier; read from the experiment’s recipe.lock.json when omitted |
--allow-dirty | Allow uncommitted changes in the working tree |
--no-validate | Skip the pre-submission validation |
--dry-run | Only print the variant list, submit nothing |
--early-stop | Enable ASHA early stopping: cut the clearly hopeless variants at a few step rungs |
--stop-metric | Which metric early stopping watches; defaults to the method’s declared primary metric (see sf methods <method>) |
--rungs | Early-stopping rungs (steps), comma separated. Defaults to 100,300,900 |
--keep <int range> | Keep 1/keep of the variants at each rung. 3 is the usual value: cutting harder hurts slow starters (默认 3) |
