DeepLink-org / DeepLinkExt

BSD 3-Clause "New" or "Revised" License
11 stars 0 forks source link

feat: reimpl rotary embedding for npu #127

Closed jingguo-st closed 1 month ago

jingguo-st commented 1 month ago
  1. reimpl rotary embedding with npu_rotary_mul in torch_npu
  2. using combined rms_norm and rotary_embedding temporarily for accuracy.