DeepSpeed
Fix for fragmented linear inputs in ZeRO 3 Linear layers where reshap…
#881
Merged

Commits
  • Fix for fragmented linear inputs in ZeRO 3 Linear layers where reshape is needed instead of view. Should not affect performance for cases which only requires view since reshape will just do view when possible
    samyam committed 5 years ago
  • Enable memory efficient linear when ZeRO 3 model is initialized in Stage 3 initialize instead of using deepspeed.init
    samyam committed 5 years ago
  • Merge branch 'master' into samyamr/fix-for-fragmented-linear-inputs
    tjruwase committed 5 years ago
  • Merge branch 'master' into samyamr/fix-for-fragmented-linear-inputs
    jeffra committed 5 years ago
  • Merge branch 'master' into samyamr/fix-for-fragmented-linear-inputs
    jeffra committed 5 years ago
  • Merge branch 'master' into samyamr/fix-for-fragmented-linear-inputs
    jeffra committed 5 years ago
Loading