how much vram could LLaMa 3 400B model require to be trained for chinese llama type training ?

[StephennFernandes]

Hi there just wanted to know the estimates on how much VRAM could be needed to train chinese llama type training on the 400B llama model ? as extending the tokenizer also extends the vocab_size in the model parameters would be needing to account for those values as well.

i currently have 8x A6000's wanted to know if these could suffice ? additionally can i load the model in 4 bit and train the training script.