Kernels

Update autotune configuration to avoid crash on AMD devices

#2
by ror HF Staff - opened

When running on an AMD device, trying to autotune with 32 warps causes a crash with RuntimeError: Triton Error [HIP]: Code: 1, Messsage: invalid argument. Thus we are removing that configuration when the device name contains "AMD", which is the case for MI250, MI300 and MI355. Tested on MI300.

deleted
•
This comment has been hidden
deleted
•
This comment has been hidden
Cannot merge
This branch has merge conflicts in the following files:
  • torch-ext/triton_layer_norm/layer_norm.py

Sign up or log in to comment