A queue service for quickly developing scripts that use all your GPUs efficiently

fastgpu provides a single command, fastgpu_poll, which polls a directory to check for scripts to run, and then runs them on the first available GPU. If no GPUs are available, it waits until one is. If more than one GPU is available, multiple scripts are run in parallel, one per GPU.

An API is also provided for polling programmatically, which is extensible for assigning other resources to processes besides GPUs. For details on the API, see the docs for core.

pip install fastgpu

--help provides command help:

$ fastgpu_poll --help
usage: fastgpu_poll [-h] [--path PATH] [--exit EXIT]

Poll `path` for scripts using `ResourcePoolGPU.poll_scripts`

optional arguments:
  -h, --help   show this help message and exit
  --path PATH  Path containing `to_run` directory
  --exit EXIT  Exit when `to_run` is empty

path defaults to the current directory. The path should contain a subdirectory to_run containing executable scripts you wish to run. It should not contain any other files, although it can contain subdirectories (which are ignored).

fastgpu_poll will run each script in to_run in sorted order. Each script will be assigned to one GPU. The CUDA_VISIBLE_DEVICES environment variable will be set to the ID of this GPU in the script's subprocess. In addition, the FASTGPU_ID environment variable will also be set to this ID.

Once a script is selected to be run, it is moved into a directory called running. Once it's finished, it's moved into complete or fail as appropriate. stdout and stderr are captured to files with the same name as the script, plus stdout or stderr appended.

If exit is 1 (which is the default), then once all scripts are run, fastgpu_poll will exit. If it is 0 then fastgpu_poll will continue running until it is killed; it will keep polling for any new scripts that are added to to_run.

To limit the GPUs available to fastgpu, set CUDA_VISIBLE_DEVICES before polling, e.g.:

CUDA_VISIBLE_DEVICES=2,3 fastgpu_poll script_dir