| name | gpu_pressure |
| description | GPU memory and utilization headroom
|
| category | memory |
| tables | ["gpu.utilization","gpu.devices","python.torch_trace"] |
| tags | ["gpu","VRAM","utilization","显存","利用率"] |
| keywords | {"en":["GPU memory","VRAM","GPU utilization","GPU idle"],"zh":["显存不够","GPU 利用率","VRAM","显存占用","GPU 空闲"]} |
| parameters | {"sample_limit":{"type":"integer","default":20}} |
GPU memory and utilization pressure
查看 gpu.utilization 采样与 python.torch_trace 中的 allocated 是否一致,
判断是「真 OOM 风险」还是「利用率低 / 内存碎片」。
Parameters
sample_limit (integer, default 20):
Related skills
- 显存持续上涨 → skill: memory_leak
- MPS Mac → torch.mps.current_allocated_memory 已在 torch_trace 中
- 启用更细 profiling → probing.torch.profiling=on,random:0.1