ant-scheduler-control

name

ant-scheduler-control

description

Use when the user wants the coding agent to interact with ANT Scheduler, the GPU job scheduler at github.com/gjamesgoenawan/ant-scheduler, to launch, monitor, and inspect research ML or deep-learning runs. Use it for GPU training jobs, no-GPU jobs, project default runner setup in .ant-scheduler/env, queue or running status checks, limited log reads, and full log path lookup through the bundled helper scripts.

ANT Mission Control

Use this skill to interact with ANT Scheduler, the GPU job scheduler at github.com/gjamesgoenawan/ant-scheduler.

ANT is primarily used to schedule research experiments and deep-learning or ML training runs that need GPUs, but it also supports jobs with no GPU assignment.

Use this skill for:

• project-level ANT default setup
• launching ANT tasks
• monitoring queued, ongoing, and completed tasks
• task control operations such as queue removal, restart, and termination
• worker control operations such as enabling or disabling GPUs
• runner health and version checks
• fetching limited logs
• locating full log files

Project defaults

Before using this skill on a project, define these defaults in .ant-scheduler/env:

• ANT_SCHEDULER_URL='http://localhost:5000'
• ANT_CONDA_ENV=''
• ANT_CONDA_PATH='conda'

This file is the project default source of truth unless the user explicitly overrides values for a specific task. If the user asks to change the project defaults, update .ant-scheduler/env.

Load behavior:

• if ANT_SCHEDULER_URL, ANT_CONDA_ENV, or ANT_CONDA_PATH are already defined in the environment, use those values
• otherwise, load them from .ant-scheduler/env
• if .ant-scheduler/env does not exist yet, ask the user for all three project defaults first, then create the file
• if the user has no project-specific preference, initialize the file with:
  • ANT_SCHEDULER_URL='http://localhost:5000'
  • ANT_CONDA_ENV=''
  • ANT_CONDA_PATH='conda'

ANT_WD should not be stored in .ant-scheduler/env. Infer it from the repository context and the user request whenever possible.

When setting up or changing project defaults:

• always ask the user about all three values first:
  • ANT_SCHEDULER_URL
  • ANT_CONDA_ENV
  • ANT_CONDA_PATH
• after the user confirms them, write the file with scripts/init_project_env.sh
• do not silently create .ant-scheduler/env without user confirmation
• if the user says to use the standard defaults, still write those explicit values into .ant-scheduler/env

Other assumptions:

• queue_mode is always "single"
• For randomized ports or seeds inside a runner command, prefer ANT's built-in RNG syntax such as {rand int 20000 40000} instead of shell $RANDOM.

Workflow

1. Check runner state before launching:
   • bash skills/ant-mission-control/scripts/status.sh health
   • bash skills/ant-mission-control/scripts/status.sh gpus
   • bash skills/ant-mission-control/scripts/status.sh ongoing
   • bash skills/ant-mission-control/scripts/status.sh queue
2. If enough GPUs are allowed and available, submit the task with scripts/create_task.sh.
3. Monitor the task with scripts/status.sh and scripts/log.sh.
4. Use scripts/task.sh or scripts/gpu.sh for runner control operations.
5. When needed, resolve the full log file with scripts/find_log_path.sh.

JSON shape

The /vis endpoint returns:

• .data.monitor
• .data.task_completed
• .data.task_ongoing
• .data.task_queue

Useful task fields:

• task_id
• command
• n_gpus
• gpu_ids
• gpu_idx
• log_file_name
• envar.ant_conda_env
• envar.ant_conda_path
• envar.ant_wd
• time.start
• time.runtime

Scripts

Use these bundled scripts instead of rewriting curl payloads:

• scripts/common.sh
• scripts/init_project_env.sh
• scripts/create_task.sh
• scripts/status.sh
• scripts/task.sh
• scripts/gpu.sh
• scripts/log.sh
• scripts/find_log_path.sh

Run --help on a script first if the exact arguments are not obvious.

Submission rules

• Always use a unique task_id.
• Keep the full training command inside the runner command field.
• Do not prepend cd when ant_wd is set through runner environment variables.
• Prefer torchrun commands exactly as used in this repo.
• If the user specifies GPUs, pass that count as n_gpus.
• If the user specifies ANT_SCHEDULER_URL, ANT_CONDA_ENV, ANT_CONDA_PATH, or ANT_WD, pass them through the helper script via environment variables.
• When a command needs a randomized rendezvous port, use ANT interpolation directly inside the command, for example:
  • --rdzv-endpoint=localhost:{rand int 20000 40000}
  • PORT={rand int 20000 40000}
• Do not use shell randomization like $((20000 + RANDOM % 40000)) inside runner-submitted commands unless the user explicitly asks for shell-side randomness.

Monitoring rules

• Use status.sh health before first use when you need to confirm the runner is alive.
• A healthy runner root response looks like:
  • {"message":"Ant-Scheduler Backend (commit: commit_id)","status":"success"}
• Use status.sh version if you only need the runner version string.
• For status, prefer summarizing whether the task is queued, running, or completed.
• For logs, use the limited /get_log endpoint first.
• If the user wants to read the complete log contents, use log.sh --full <task_id>.
• Only resolve the full log path when the user explicitly asks for it or the limited log is insufficient.

Examples

Check GPUs:

bash skills/ant-mission-control/scripts/status.sh gpus

Check runner health:

bash skills/ant-mission-control/scripts/status.sh health

Check runner version:

bash skills/ant-mission-control/scripts/status.sh version

Initialize project defaults:

bash skills/ant-mission-control/scripts/init_project_env.sh \
  'http://localhost:5000' \
  '' \
  'conda' \
  .

Create a task:

ANT_CONDA_ENV=my_env ANT_CONDA_PATH=/home/anaconda/bin/conda \
bash skills/ant-mission-control/scripts/create_task.sh \
  relu_vit-tiny_imagenet_mymethod_rc6 \
  2 \
  'torchrun --nproc_per_node=2 --rdzv-endpoint=localhost:{rand int 20000 40000} tasks/setting_1.py --model_type vit_tiny --dataset imagenet --attn_type parallel_relu --load_strat force --ckpt pretrained --teacher_ckpt pretrained --save_last --train_v --conversion_steps 20000 --conversion_strat mymethod_rc6 --comment mymethod_rc6'

Check one task:

bash skills/ant-mission-control/scripts/status.sh task relu_vit-tiny_imagenet_mymethod_rc6

Fetch limited log:

bash skills/ant-mission-control/scripts/log.sh relu_vit-tiny_imagenet_mymethod_rc6

Fetch full log text:

bash skills/ant-mission-control/scripts/log.sh --full relu_vit-tiny_imagenet_mymethod_rc6

Get full log path:

bash skills/ant-mission-control/scripts/find_log_path.sh relu_vit-tiny_imagenet_mymethod_rc6

Remove queued task:

bash skills/ant-mission-control/scripts/task.sh remove-queue relu_vit-tiny_imagenet_mymethod_rc6

Terminate running task:

bash skills/ant-mission-control/scripts/task.sh kill relu_vit-tiny_imagenet_mymethod_rc6

Restart task:

bash skills/ant-mission-control/scripts/task.sh restart relu_vit-tiny_imagenet_mymethod_rc6

Disable GPU 0:

bash skills/ant-mission-control/scripts/gpu.sh disable 0

Enable GPU 0:

bash skills/ant-mission-control/scripts/gpu.sh enable 0