Skip to content

Releases: chandar-lab/tgi-for-mila

Release v1

01 Sep 17:45
aafa4ba
Compare
Choose a tag to compare

The inital release.

Previous versions used a lot of hacks to avoid CUDA 11.8 on Mila.
However, as Mila have now upgraded their drivers that is no longer required.

Previous versions also didn't support the quantize feature. bnb and
accelerate was also not correctly installed.

  • TGI version: 1.0.2
  • enabled features: [bnb, accelerate, quantize]
  • Flash-attention version: 2.0.8