vllm 启动时报错：
  torch.cuda.OutOfMemoryError: CUDA out of memory.
  Tried to allocate 2.00 GiB. GPU 0 has a total capacity of 79.13 GiB
  of which 1.50 GiB is free.

nvidia-smi
# 查看当前显存占用

+-----------------------------------------------------------------------------+
| GPU  Name        | Memory-Usage | GPU-Util |
|------------------+--------------+----------|
|   0  A100-80G    |  79000MiB    |   0%     |
+-----------------------------------------------------------------------------+

模型权重（FP16）: 7B × 2 bytes = 14 GB
KV Cache（估算）: 取决于 batch 和 seq_len
其他: ~1-2 GB

如果 KV Cache 配置为 max_num_seqs=256, max_model_len=4096:
  KV Cache ≈ 4-8 GB

总计: 14 + 8 + 2 = 24 GB ← 远低于 80 GB，不应该 OOM

# 清理残留进程
ps aux | grep vllm | grep -v grep | awk '{print $2}' | xargs kill -9

# 重新启动，保守配置
vllm serve model_name \
  --gpu-memory-utilization 0.90 \
  --max-model-len 4096 \
  --max-num-seqs 128