Installation requirement fails: Transformer-engine

[ddofer]

transformer-engine-torch=2.3.0 is unavailable with linux (Aarm64/DGX spark), preventing installation. Is there another solution or workaround to get evo2 working on a gpu device? (I'm fine with running it cpu only if there's no other solution)

conda install -c nvidia cuda-nvcc cuda-cudart-dev -y;
conda install -c conda-forge transformer-engine-torch=2.3.0 -y;
2 channel Terms of Service accepted
Retrieving notices: done
Channels:
 - nvidia
 - defaults
Platform: linux-aarch64
Collecting package metadata (repodata.json): done
Solving environment: done

# All requested packages already installed.

2 channel Terms of Service accepted
Channels:
 - conda-forge
 - defaults
Platform: linux-aarch64
Collecting package metadata (repodata.json): done
Solving environment: failed

PackagesNotFoundError: The following packages are not available from current channels:

  - transformer-engine-torch=2.3.0

[mdanner93]

I encountered the same problem, maybe this helps:

#201 (comment)

[garykbrixi]

Hi, this requirement for Nvidia GPU with hopper capability is due to the fp8 requirement used by our models. The 1B and 40B are numerically sensitive, so we enforce this requirement. The 7B should be possible to run on other hardware and removing transformer engine; please use the tests as shown in README if deploying on non supported hardware or trying different version requirements