OOM 问题, 显卡是A00 40G

[gongye19]

用llama factory进行sft可以使用deepspeed zero2 微调llama3-8B的模型，但这个框架就算batch设为1，用deepspeed zero2也会报OOM。
用zero3训练会变得很慢，出现这个问题：

2 pytorch allocator cache flushes since last step. this happens when there is high memory pressure and is detrimental to performance. if this is happening frequently consider adjusting settings to reduce memory consumption. If you are unable to make the cache flushes go away consider adding get_accelerator().empty_cache() calls in your training loop to ensure that all ranks flush their caches at the same time

[fe1ixxu]

OOM issue could be the reason that llama3 has 128K vocab size, while llama2 is 32K.

[gongye19]

OOM issue could be the reason that llama3 has 128K vocab size, while llama2 is 32K.

I tried deepseek-7B， same question

[fe1ixxu]

The deepseek vocab size is also large -- 100K. The memory I used for training is 64GB for llama-2 with 8/16 GPUs. Maybe you want to try using fsdp.

[gongye19]

The deepseek vocab size is also large -- 100K. The memory I used for training is 64GB for llama-2 with 8/16 GPUs. Maybe you want to try using fsdp.

谢谢，我现在是先用llama factory进行sft，再到你的框架上进行cpo。这样可以吗？

[moore3930]

Technically, 1 GPU (80G) should be fine to fine-tune LLaMA 7B with Lora, but it seems always OOM under your codebase if we do not use 8 GPUs. I am just wondering why it costs a lot.