Regarding the initialization on a single V100-32G card, the weights and configurations that should be used.

[txq-hub]

When I perform single-card pre-training on the V100-32G using the open_clip_pytorch_model.bin and model.safetensors weights, I always run out of memory. I understand that you have already conducted single-card testing on the V100-32G and would like to inquire about the specific graphics card configuration you used for pre-training. Is it possible to initialize without using the open_clip_pytorch_model.bin weights?