freezing layers have differenct behaves for different models

[hjc3613]

System Info

Dockerfile:

Information

• The official example scripts
• My own modified scripts

🐛 Describe the bug

when freezing the top n layers, it has different effect for different models:
for qwen, the freezed layers has no gradients, so the gpu ram becomes lower contrast to none freezing.
but for qwen1.5 and wizard-8x22B, the freezed layers seemes to have gradients, so the gpu ram still high, even when I freeze top n-1 layers and fintune the last one layer, it gives me OOM when trained on 3 nodes(total 24* A800).

Error logs

cuda out of memoy

Expected behavior

can freeze top layers for any models

[HamidShojanazeri]

@hjc3613 sorry for the inconvenience, this feature is not well tested thats why we didn't mention it much. If you are interested, would love to work with you and make a PR to fix the issues.

[hjc3613]

@hjc3613 sorry for the inconvenience, this feature is not well tested thats why we didn't mention it much. If you are interested, would love to work with you and make a PR to fix the issues.

I think this feature is necessary, it can save much gpu memory and get the same results as full parameter training in some practical. I have test in my task that when freeze top 40 layers(total 80 layers), the test results between none freeze are same！！but only use 1 nodes(8*80G)