def multi_gpu_launcher()

in domainbed/command_launchers.py [0:0]

• 17 lines of code
• 8 McCabe index (conditional complexity)

def multi_gpu_launcher(commands):
    """
    Launch commands on the local machine, using all GPUs in parallel.
    """
    print('WARNING: using experimental multi_gpu_launcher.')
    n_gpus = torch.cuda.device_count()
    procs_by_gpu = [None]*n_gpus

    while len(commands) > 0:
        for gpu_idx in range(n_gpus):
            proc = procs_by_gpu[gpu_idx]
            if (proc is None) or (proc.poll() is not None):
                # Nothing is running on this GPU; launch a command.
                cmd = commands.pop(0)
                new_proc = subprocess.Popen(
                    f'CUDA_VISIBLE_DEVICES={gpu_idx} {cmd}', shell=True)
                procs_by_gpu[gpu_idx] = new_proc
                break
        time.sleep(1)

    # Wait for the last few tasks to finish before returning
    for p in procs_by_gpu:
        if p is not None:
            p.wait()