Sean Owen
Sean Owen
There is probably an earlier error about shared libraries not being available. Do you have all the additional CUDA libraries installed, like cublas? You can see some package install lines...
Try torch 1.13.1 ? Not sure if 2.0 is causing issues
I think you don't have all the right CUDA libs installed somehow, hard to say.
Nice. CUDA 11.7 works too (that's what I use) but i suspect something else wasn't compatible in here in the shared libraries. It's tricky.
No, I mean it does work.
It tells you the problem: `#033[93m [WARNING] #033[0m sparse_attn requires a torch version >= 1.5 and < 2.0 but detected 2.0`
I think it actually doesn't work here. You're also using deepspeed 0.9.2, and I know we had problems with >= 0.9.0. It could be that or any other differences in...
I think there are lots of answers here. You haven't said what you are doing or what you tried
Just pass `--input-model EleutherAI/pythia-6.9b`
What do you mean 3 dirs? your model exists in a directory on a file system locally. You pass the path to that dir.