when using multiple gpu, it works weirdly. More safetensor files are created than the set number.

my expectation is,
epochs : 10
there are 10 safetensor files for output.

because there are files as same as i set before with 1 gpu.

but when using multiple gpu, More safetensor files are created than the set number.
in my case, 10 epochs with 4 gpus, it makes me 33 safe tensor files.
30 epochs with 4 gpus, 100 safe tensor files.

is it work properly ?

Reproduce the report with 10 epochs on 1 GPU and 4 GPUs, then compare the generated safetensor file counts; the issue reports 10 files on 1 GPU versus 33 on 4 GPUs. Trace the multi-GPU checkpoint-saving behavior from the training entry point, and consider the issue done when the output count matches the configured epochs or the behavior is clearly documented.

Written by the indexing model from the issue text.