Commits
  • NCCL-based 1-bit Adam + Code Refactor for Comm. Backends (#594)
    awan-10 committed 5 years ago
  • Merge branch 'master' into staging-1bit-nccl-v2
    conglongli committed 5 years ago
  • Revert "Merge branch 'master' into staging-1bit-nccl-v2"
    conglongli committed 5 years ago
  • Revert "Revert "Merge branch 'master' into staging-1bit-nccl-v2""
    conglongli committed 5 years ago
  • Merge branch 'master' into staging-1bit-nccl-v2
    conglongli committed 5 years ago
  • comm optimization + 1-bit lamb
    conglongli committed 5 years ago
  • Saving/debugging commit.
    awan-10 committed 5 years ago
  • finalizing 1-bit lamb
    conglongli committed 5 years ago
  • finalizing 1-bit lamb
    conglongli committed 5 years ago
  • add momentum mask and chkpt handling for 1-bit adam
    conglongli committed 5 years ago
  • Merge remote-tracking branch 'origin/staging-1bit-nccl-v2' into staging-1bit-nccl-v2
    awan-10 committed 5 years ago
  • Cleanup and modify nccl test to be runnable with deepspeed launcher.
    awan-10 committed 5 years ago
  • Merge branch 'master' into staging-1bit-nccl-v2
    awan-10 committed 5 years ago
  • Fix format.
    awan-10 committed 5 years ago
  • fix formatting again.
    awan-10 committed 5 years ago
  • make test runnable without mpi4py
    awan-10 committed 5 years ago
  • Add dist.alltoall and dist.allgather instead of custom functions.
    awan-10 committed 5 years ago
  • remove debug prints.
    awan-10 committed 5 years ago
  • formatting and renaming
    conglongli committed 5 years ago
  • renaming
    conglongli committed 5 years ago
  • renaming
    conglongli committed 5 years ago
  • add unit test, fix existing tests
    conglongli committed 5 years ago
  • Merge branch 'master' into staging-1bit-adam-v2
    awan-10 committed 5 years ago
  • skip unit test when torch < 1.8
    conglongli committed 5 years ago
  • revert 1-bit lamb
    conglongli committed 5 years ago
  • flatten momentum when dimension is more than 1
    conglongli committed 5 years ago
  • add warning message for 1-bit adam under fp32
    conglongli committed 5 years ago
  • improve version check
    conglongli committed 5 years ago
  • add fp32 test
    conglongli committed 5 years ago
  • Merge remote-tracking branch 'origin' into staging-1bit-adam-v2
    conglongli committed 5 years ago
  • 1-bit adam doc
    conglongli committed 5 years ago
  • fix file name
    conglongli committed 5 years ago
  • doc fix
    conglongli committed 5 years ago
  • torch 1.8 is released
    conglongli committed 5 years ago
  • doc fix
    conglongli committed 5 years ago
  • fix tests
    conglongli committed 5 years ago
  • Merge branch 'master' into staging-1bit-adam-v2
    conglongli committed 5 years ago
  • Merge branch 'master' into staging-1bit-adam-v2
    conglongli committed 5 years ago
  • update news
    conglongli committed 5 years ago
  • Merge branch 'master' into staging-1bit-adam-v2
    conglongli committed 5 years ago
  • add doc for momentum mask
    conglongli committed 5 years ago
  • Merge branch 'staging-1bit-adam-v2' of github.com:microsoft/DeepSpeed into staging-1bit-adam-v2
    conglongli committed 5 years ago
  • fix checkpoing handling, add unit test
    conglongli committed 5 years ago
  • checkpoint handling doc
    conglongli committed 5 years ago
  • doc final cleanup
    conglongli committed 5 years ago
  • Merge branch 'master' into staging-1bit-adam-v2
    conglongli committed 5 years ago
  • Merge branch 'master' into staging-1bit-adam-v2
    jeffra committed 5 years ago
  • bump dates
    conglongli committed 5 years ago
  • Merge branch 'staging-1bit-adam-v2' of github.com:microsoft/DeepSpeed into staging-1bit-adam-v2
    conglongli committed 5 years ago
  • update tests
    conglongli committed 5 years ago
  • url change
    conglongli committed 5 years ago
  • doc fix
    conglongli committed 5 years ago
  • fix test
    conglongli committed 5 years ago
  • doc update
    conglongli committed 5 years ago
  • Merge branch 'master' into staging-1bit-adam-v2
    conglongli committed 5 years ago
Loading