scripts/reserve_gpu.py

"""Transparent, idle VRAM reservation; SIGTERM releases CUDA allocations."""
import os, signal, time
import torch
assert os.environ['CUDA_VISIBLE_DEVICES']=='GPU-2a9c799b-ab9e-c677-db0e-379f480cb00f'
assert torch.cuda.device_count()==1
assert 'H200' in torch.cuda.get_device_name(0)
free,total=torch.cuda.mem_get_info()
assert free > 125*1024**3, f'GPU1 no longer idle: {free=}'
buf=torch.empty(110*1024**3,dtype=torch.uint8,device='cuda')
buf.fill_(0)
print(f'Latent rubric GPU1 idle reservation: {buf.numel()/1024**3:.0f} GiB; PID {os.getpid()}',flush=True)
signal.signal(signal.SIGTERM, lambda *_: exit(0))
while True: time.sleep(10)