Commit History

Autor SHA1 Mensaxe Data
  Hamid Shojanazeri a955ed1999 added checks for dist barrier and commented cuda exapnadable segements and dist_dbug hai 1 ano
  Hamid Shojanazeri a2403c7c1a clean up hai 1 ano
  Hamid Shojanazeri e9559d2669 fixing the train/eval_loss calcualtion hai 1 ano
  Hamid Shojanazeri 4ba4400a75 adding dist barrier before and after checkpointing hai 1 ano
  Hamid Shojanazeri a49a2c2804 adding PT cuda allocation expand flag hai 1 ano
  Hamid Shojanazeri 442c1ccf7c adding barrier to end of trainer loop hai 1 ano
  Hamid Shojanazeri f74d57dc08 printing scores based on fsdp usage or single gpu hai 1 ano
  Hamid Shojanazeri 3d887ea483 update with active memory and removing rank0 for eval score hai 1 ano
  Hamid Shojanazeri bedb96b78a fixing the full state path in checkpoint handler hai 1 ano
  Hamid Shojanazeri bd01f64cbd Merge branch 'main' into fix-cuda_id hai 1 ano
  Andrew Gu 71fdc4920a Save memory and fix typos hai 1 ano
  Hamid Shojanazeri a7156dfb5d fixing the cuda id hai 1 ano
  Hamid Shojanazeri 707af7ea24 adding cuda:0 for non-fsdp situations hai 1 ano
  Hamid Shojanazeri 6678be75ad fixing identation hai 1 ano
  Hamid Shojanazeri 6a84e9e4d5 fixing scaler for both fsdp and non fsdp hai 1 ano
  Hamid Shojanazeri 065ddaa77b fixing the condition for moving to cuda hai 1 ano
  Hamid Shojanazeri 20b061e01c modify to steping the lr scheduler each epoch hai 1 ano
  chauhang 4767f09ecd Initial commit hai 1 ano