DeepSpeed
Samyamr/full precision for ZeRO Stage2 and Stage3
#1004
Merged

Commits
  • Adding tf32 and fp32 support for ZeRO Stage 3
    samyam committed 4 years ago
  • Changing to location of self.dtype assignment
    samyam committed 4 years ago
  • Exhaustive setting of self.dtype
    samyam committed 4 years ago
  • Adding fp32 and tf32 support for ZeRO Stage 2
    samyam committed 4 years ago
  • fix loss scale value for static loss scale
    samyam committed 4 years ago
  • Adding tf32 and fp32 support for ZeRO Stage 3
    samyam committed 4 years ago
  • Changing to location of self.dtype assignment
    samyam committed 4 years ago
  • Exhaustive setting of self.dtype
    samyam committed 4 years ago
  • Adding fp32 and tf32 support for ZeRO Stage 2
    samyam committed 4 years ago
  • fix loss scale value for static loss scale
    samyam committed 4 years ago
  • adding documentation for dtype in zero init
    samyam committed 4 years ago
  • fixing conflict
    samyam committed 4 years ago
  • Merge branch 'master' into samyamr/full-precision-for-stage3
    jeffra committed 4 years ago
  • Merge branch 'master' into samyamr/full-precision-for-stage3
    tjruwase committed 4 years ago
  • Disable cpu-adam update_copy api for fp32
    tjruwase committed 4 years ago
  • Disable gradient clipping in engine for ZeRO
    tjruwase committed 4 years ago
  • fp16 mode init required for ZeRO-3
    tjruwase committed 4 years ago
  • Update engine.py
    samyam committed 4 years ago
  • Formatting fix
    tjruwase committed 4 years ago
  • bump DSE
    jeffra committed 4 years ago
  • Merge branch 'master' into samyamr/full-precision-for-stage3
    jeffra committed 4 years ago
Loading