search

young-geng / EasyLM

Large language models (LLMs) made easy, EasyLM is a one stop solution for pre-training, finetuning, evaluating and serving LLMs in JAX/Flax.
Apache License 2.0
2.33k stars 247 forks source link

how to train llama2-7b in a100 80G gpu? #87

Open brewswang opened 11 months ago

brewswang commented 11 months ago

how to train llama2-7b in a100 80G gpu?

young-geng commented 11 months ago

It would be difficult to do that without some parameter efficient finetuning methods such as LoRA. For the full 7B model, the parameters and optimizer states would take 84GB of memory, exceeding the 80GB of a single A100.

bet0x commented 10 months ago

@young-geng Got a table or something similar with those requirements by model/size?

kuangdao commented 1 month ago

dose not support zero of jax？