Run a parameter sweep

Start several detached training runs with different settings on Hugging Face, wait for all of them, and collect their results.

Detached runs make sweeps simple: start one run per setting, then wait for all of them. This uses the train.py from Train with uv scripts and Hugging Face, which supports --detach.

Every run bills separately. Five a10g-small runs cost five times $1.00 per hour while they run. Start with a short --max and a CPU flavor.

Start the runs

for epochs in 50 100 200 400; do
  hi compute run --gpu a10g-small --max 30m --detach --yes \
    --name "sweep-e$epochs" train.py -- --epochs "$epochs"
done

--yes confirms each start, because the loop has no terminal to ask in; decide on the cost before running it. Each run prints its job URL.

Watch them

hi compute ls

NAME         PROVIDER  HARDWARE    UP  STOPS IN
sweep-e100   hf        a10g-small  1m  29m
sweep-e200   hf        a10g-small  1m  29m
sweep-e400   hf        a10g-small  1m  29m
sweep-e50    hf        a10g-small  1m  29m

hi compute logs sweep-e400 --follow

Wait for all of them

hi compute wait takes several names and exits non-zero if any run failed:

hi compute wait sweep-e50 sweep-e100 sweep-e200 sweep-e400 && echo "all succeeded"

Collect the results

Logs stay available after the runs end:

for epochs in 50 100 200 400; do
  printf '%s\t' "$epochs"
  hi compute logs "sweep-e$epochs" -n 5 | grep 'loss' | tail -1
done

Stop early

hi compute stop sweep-e400    # one
hi compute stop --all         # everything, after asking