====== 强行占用20G的显存 ======

因为某些特定的原因，我们必须占用20G的显存

<code python use_gpu.py>
import torch
import argparse
from math import ceil

def parse_memory_size(memory_str):
    memory_str = memory_str.upper()
    if memory_str.endswith('G'):
        gb = float(memory_str[:-1])
        return int(gb * 1024**3)
    elif memory_str.endswith('M'):
        mb = float(memory_str[:-1])
        return int(mb * 1024**2)
    else:
        raise ValueError("Unsupported unit! Use 'G' (GB) or 'M' (MB), e.g., '20G' or '512M'")

def allocate_vram(memory_str="20G", gpu_id=0):
    if not torch.cuda.is_available():
        raise RuntimeError("No available GPU for VRAM allocation!")
    
    target_bytes = parse_memory_size(memory_str)
    target_gb = target_bytes / (1024**3)
    print(f"Target VRAM: {target_gb:.2f} GB")
    
    gpu_props = torch.cuda.get_device_properties(gpu_id)
    total_vram_gb = gpu_props.total_memory / (1024**3)
    print(f"Current GPU: {gpu_props.name} (Total VRAM: {total_vram_gb:.2f} GB)")
    
    if target_gb > total_vram_gb:
        raise RuntimeError(f"Insufficient GPU VRAM! Required: {target_gb:.2f} GB, Available: {total_vram_gb:.2f} GB")
    
    element_size = 4
    total_elements = ceil(target_bytes / element_size)
    print(f"Required float32 elements: {total_elements:,} (≈ {target_gb:.2f} GB)")
    
    chunk_size = 10**9
    tensors = []
    allocated_elements = 0
    
    while allocated_elements < total_elements:
        current_chunk = min(chunk_size, total_elements - allocated_elements)
        tensor = torch.randn(
            current_chunk,
            dtype=torch.float32,
            device=f'cuda:{gpu_id}',
            requires_grad=False
        )
        tensors.append(tensor)
        allocated_elements += current_chunk
        
        allocated_gb = (allocated_elements * element_size) / (1024**3)
        print(f"Allocated: {allocated_gb:.2f} GB")
    
    used_vram = torch.cuda.memory_allocated(gpu_id) / (1024**3)
    cached_vram = torch.cuda.memory_reserved(gpu_id) / (1024**3)
    print(f"\nAllocation completed!")
    print(f"Used VRAM (allocated): {used_vram:.2f} GB")
    print(f"Cached VRAM (reserved): {cached_vram:.2f} GB")
    
    input("\nPress Enter to release VRAM...")
    del tensors
    torch.cuda.empty_cache()
    print("VRAM released")

if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="GPU VRAM Allocator (default: 20GB)")
    parser.add_argument(
        "--memory",
        default="20G",
        help="VRAM size (e.g., '10G' or '512M', default: 20G)"
    )
    parser.add_argument(
        "--gpu",
        type=int,
        default=0,
        help="GPU device ID (default: 0)"
    )
    args = parser.parse_args()
    allocate_vram(args.memory, args.gpu)
</code>