3 ms·
MLX reports peak memory usage at the end of the response. Otherwise I'll use Activity Monitor.
by simonw 2y ago
MLX reports peak memory usage at the end of the response. Otherwise I'll use Activity Monitor.
- aukejw 2y agoI'm also trusting `get_peak_memory` + some small buffer for now. Still, it reports accurate peak memory usage for tensors living on GPU, but seems to miss some of the non-Metal overhead, however small (https://github.com/aukejw/mlx_transformers_benchmark/issues/1 https://github.com/aukejw/mlx_transformers_benchmark/issues/...).