pi0 multi gps train

[ximiluuuu]

if i have multi 4090, how to modify to train pi0?

only 1 4090 just error

[BruceGeLi]

same question + 1

[ximiluuuu]

same question + 1

sad

[YanJiaHuan]

Also like to know. Another question is how to freeze siglip and gemma, and only update the para for gemma expert?

[ximiluuuu]

Also like to know.  也想知道。 Another question is how to freeze siglip and gemma, and only update the para for gemma expert?另一个问题是如何冻结 siglip 和 gemma，只更新 gemma expert 的段落？

+1 i have same question

[Jack-liu1998]

+1 not everyone have h100

[YanJiaHuan]

they seems to freeze already, only train action expert, but it still taks up 29G on GPU.

[ximiluuuu]

they seems to freeze already, only train action expert, but it still taks up 29G on GPU.他们似乎已经冻结了，只训练 Action Expert，但它在 GPU 上仍然占用 29G。

sad

[chenkang455]

Also like to know. Another question is how to freeze siglip and gemma, and only update the para for gemma expert?

Try setting train_expert_only = True, this will enable you to train your model with a single GPU with 24 G.

[zttiannnn]

前辈们成功使用多卡训练了吗？😟

[YanJiaHuan]

huggingface trainer is easy to use distributed training, just wonder why they do not use it

[pkooij]

Hi we update our pi0 implementation and released pi05 support in this PR: https://github.com/huggingface/lerobot/pull/1910 🚀 . If you still have this issue please reopen this thread, but I am closing it for now. We also will introduce multi gpu training with accelerate soon.

[qinglin-ai]

前辈们成功使用多卡训练了吗？😟

前辈们成功了嘛