libffcv/ffcv

doubt about mutli-gpu train when use imagenet 4 gpus

daixiangzi opened this issue · 1 comments

daixiangzi commented

Why are 4 processes started on GPU card 0 instead of 1 process?
0 N/A N/A 2579940 C .../miniconda3/envs/ffcv/bin/python3.9 13724MiB |
| 0 N/A N/A 2579941 C .../miniconda3/envs/ffcv/bin/python3.9 414MiB |
| 0 N/A N/A 2579942 C .../miniconda3/envs/ffcv/bin/python3.9 414MiB |
| 0 N/A N/A 2579943 C .../miniconda3/envs/ffcv/bin/python3.9 414MiB |
| 1 N/A N/A 2579941 C .../miniconda3/envs/ffcv/bin/python3.9 13868MiB |
| 2 N/A N/A 2579942 C .../miniconda3/envs/ffcv/bin/python3.9 13908MiB |
| 3 N/A N/A 2579943 C .../miniconda3/envs/ffcv/bin/python3.9 13958MiB |

souravdalai18 commented

when you load pretrained checkpoints states , use map_location to 'cpu'