DeepSpeed
Support AutoEP with ZeRO-3 zero.Init source modules
#8060
Merged

Commits
  • Enable AutoEP ZeRO-3 placement groups
    Masahiro Tanaka committed 83 days ago
  • Fix ZeRO-3 multi-gather wait forwarding
    Masahiro Tanaka committed 83 days ago
  • Gather AutoEP source ZeRO params
    Masahiro Tanaka committed 83 days ago
  • Document constrained AutoEP ZeRO-3 support
    Masahiro Tanaka committed 83 days ago
  • Fix AutoEP zero.Init unit test import
    Masahiro Tanaka committed 83 days ago
  • Add 8 GPU AutoEP zero.Init validation
    Masahiro Tanaka committed 83 days ago
  • Fix AutoEP ZeRO-3 expert gradient averaging
    Masahiro Tanaka committed 83 days ago
  • Add partition-native AutoEP ZeRO-3 checkpoints
    Masahiro Tanaka committed 83 days ago
  • Address review findings for partition-native AutoEP checkpoints
    Masahiro Tanaka committed 82 days ago
  • Fail fast on pre-partitioned AutoEP expert params with mismatched groups
    Masahiro Tanaka committed 82 days ago
  • Update AutoEP checkpoint docs for ZeRO-3 partition-native support
    Masahiro Tanaka committed 82 days ago
  • Validate AutoEP universal topology loads
    Masahiro Tanaka committed 82 days ago
  • Fix AutoEP topology fixture registration
    Masahiro Tanaka committed 82 days ago
  • Register AutoEP topology baseline as an explicit pytest fixture
    Masahiro Tanaka committed 82 days ago
  • Fix AutoEP topology universal router check
    Masahiro Tanaka committed 82 days ago
  • Update config-json AutoEP ZeRO-3 universal topology-change note
    Masahiro Tanaka committed 82 days ago
  • Support AutoEP universal module-only loads
    Masahiro Tanaka committed 82 days ago
  • Preserve fp32 masters for AutoEP module-only loads
    Masahiro Tanaka committed 81 days ago
  • Address AutoEP ZeRO-3 review comments
    Masahiro Tanaka committed 71 days ago
  • Merge upstream master into AutoEP ZeRO-3 PR
    Masahiro Tanaka committed 71 days ago
  • Merge branch 'master' into tohtana/autoep-zero3-zero-init
    delock committed 69 days ago
Loading