ucla-labx/distbelief

Issues

• NCCL doesn't support send and recv

  #29 opened 3 years ago by bkpcoding
  0

• run on OpenMPI

  #26 opened 6 years ago by jcaip
  0

• Investigate different update frequencies

  #17 opened 6 years ago by jcaip
  1

• decentralized async sgd

  #25 opened 6 years ago by jcaip
  0

• Ability to initialize arbitrary number of clients

  #7 opened 6 years ago by rohan-varma
  1

• Abstractions to specify which model a client should train

  #11 opened 6 years ago by rohan-varma
  1

• Benchmark performance against a singular training instance

  #8 opened 6 years ago by jcaip
  1

• Send only raw gradients between client/server

  #16 opened 6 years ago by rohan-varma
  0

• don't require payload in send_message

  #15 opened 6 years ago by rohan-varma
  2

• Add support for different types of optimizers

  #6 opened 6 years ago by rohan-varma
  1

• Gracefully stop servers + client once training is completeed

  #5 opened 6 years ago by rohan-varma
  1

• profile code

  #18 opened 6 years ago by jcaip
  0

• segfaults occasionally while model is training

  #12 opened 6 years ago by rohan-varma
  1