Tips

Small tricks that make gflow workflows faster and safer.

Dependency Shortcuts (@)

Use @ to refer to recent submissions:

• @: most recently submitted job
• @~1: previous submission
• @~2: two submissions ago

gbatch --time 10 python preprocess.py
gbatch --depends-on @ --gpus 1 --time 4:00:00 python train.py
gbatch --depends-on @ --time 10 python evaluate.py

@ also works in lists:

gbatch --depends-on-all @,@~1,@~2 python merge.py

Conda Env Auto-detect

When submitting a command (not a script), if --conda-env is omitted, gbatch will use the current shell’s $CONDA_DEFAULT_ENV (if set).

Example:

conda activate myenv
gbatch python -c 'import os,sys; print("CONDA_DEFAULT_ENV=", os.getenv("CONDA_DEFAULT_ENV")); print("python=", sys.executable)'

Example output:

Submitted batch job 42 (silent-pump-6338)

Verify:

gjob show 42
cat ~/.local/share/gflow/logs/42.log

Example log output:

CONDA_DEFAULT_ENV= myenv
python= /path/to/miniconda/envs/myenv/bin/python

Parameter Sweeps

Use --param to submit multiple jobs from one command by filling {param} placeholders (cartesian product across params).

gbatch --dry-run \
  --param lr=0.001,0.01 \
  --param bs=32,64 \
  --name-template 'lr{lr}_bs{bs}' \
  python train.py --lr {lr} --batch-size {bs}

Example output:

Would submit 4 batch job(s):
  [1] python train.py --lr 0.001 --batch-size 32 (GPUs: 0)
  [2] python train.py --lr 0.001 --batch-size 64 (GPUs: 0)
  [3] python train.py --lr 0.01 --batch-size 32 (GPUs: 0)
  [4] python train.py --lr 0.01 --batch-size 64 (GPUs: 0)

Ranges are supported (no commas): start:stop or start:stop:step (use :step for floats, e.g. 0:1:0.1).

From a CSV (--param-file)

gbatch --param-file params.csv --name-template 'run_{id}' python train.py --id {id}

params.csv must have a header row; each row is one job’s parameter set.

Limit Concurrency (--max-concurrent)

gbatch --param lr=0.001,0.01 --max-concurrent 1 python train.py --lr {lr}

Tree View When Debugging Pipelines

gqueue -t

Preview Before Cancelling

gcancel --dry-run <job_id>
gcancel <job_id>

Redo a Whole Chain

Fix the root cause, then redo the failed job and all dependent jobs:

gjob redo <job_id> --cascade

Restrict GPUs at Runtime

Keep GPUs free for non-gflow workloads by restricting what gflow can allocate:

gctl set-gpus 0,2
gctl show-gpus
gctl set-gpus all

See also: Configuration -> GPU Selection.

GPU Reservations (gctl reserve)

Block out GPUs for a user/time window (e.g. a demo) so other users can't take them while the reservation is active.

gctl reserve create --user alice --gpus 2 --start '2026-01-28 14:00' --duration 2h
gctl reserve list --active
gctl reserve list --timeline --range 48h
gctl reserve cancel <reservation_id>

--start supports ISO8601 (e.g. 2026-01-28T14:00:00Z) or YYYY-MM-DD HH:MM (local time). Times must be on :00 or :30; durations are multiples of 30 minutes.

See Also

• Job Submission
• Job Dependencies
• Time Limits
• Quick Reference