DeepSpeed just doesn't install properly on Databricks about deepspeed HOT 5 CLOSED

Hi @vikram71198 - I assume we can close this issue. If not, please comment and we can re-open.

from deepspeed.

Hi @vikram71198 - it looks like DeepSpeed is installed, what you are seeing is that you have not pre-compiled any ops. That's fine, you don't need to, the ops can be JIT compiled just fine. You probably don't need to pre-compile, but you can read more about that here and decide if you need to. If you do, determine what ops you will need and you can pre-compile those. Some ops have other dependencies, async_io, cutlass kernels, etc, that's why you see some envs with those disabled.

from deepspeed.

Gotcha. I explicitly pip install torch == 2.2.1+cu118 (torch == 2.2.2+cu121 is the default torch which I attempt to override), so another part of ds_report that I find confounding is this:

DeepSpeed general environment info:
torch install path ............... ['/databricks/python3/lib/python3.10/site-packages/torch']
torch version .................... 2.2.2+cu121
deepspeed install path ........... ['/databricks/python3/lib/python3.10/site-packages/deepspeed']
deepspeed info ................... 0.14.2, unknown, unknown
torch cuda version ............... 12.1
torch hip version ................ None
nvcc version ..................... 11.8
deepspeed wheel compiled w. ...... torch 2.2, cuda 12.1
shared memory (/dev/shm) size .... 560.90 GB

Why do torch version, torch cuda version & deepspeed compiled wheel all indicate torch == 2.2.2+cu121 & not 2.2.1+cu118?

The Databricks Cluster Runtime I'm currently using has CUDA == 11.8.

And yes, I run the torch installation before the DeepSpeed installation.

from deepspeed.

Okay, I fixed this myself. Nvm.

from deepspeed.

So we can close this issue?

from deepspeed.