Code

Pull in DPO loss#48

Code

Open

saforem2 wants to merge 50 commits intoargonne-lcf:mainargonne-lcf/Megatron-DeepSpeed:mainfrom ramanathanlab:mainramanathanlab/Megatron-DeepSpeed:mainCopy head branch name to clipboard

+8,171-30

Commits

Commits on Mar 20, 2024

• changes for dpo

  gdharuman11

  committed
  e330ae3

  Copy full SHA for e330ae3

• moved dpo_training script

  gdharuman11

  committed
  5df25e7

  Copy full SHA for 5df25e7

Commits on Mar 21, 2024

• commented out assert in load checkpoint func

  gdharuman11

  committed
  461524e

  Copy full SHA for 461524e

• tuple structure removed

  gdharuman11

  committed
  2c66771

  Copy full SHA for 2c66771

• loading correct env

  gdharuman11

  committed
  f983241

  Copy full SHA for f983241

• modified main func

  gdharuman11

  committed
  1aeebca

  Copy full SHA for 1aeebca

• run script with modified exec

  gdharuman11

  committed
  aa9d9b2

  Copy full SHA for aa9d9b2

• added modified load model weights func

  gdharuman11

  committed
  ded92e3

  Copy full SHA for ded92e3

• component wise loss func

  gdharuman11

  committed
  96a1d26

  Copy full SHA for 96a1d26

Commits on Mar 22, 2024

• added arguments and datapaths for pref unpref datasets

  gdharuman11

  committed
  fa587cf

  Copy full SHA for fa587cf

• added arguments for pref unpref datapaths

  gdharuman11

  committed
  3c5b89a

  Copy full SHA for 3c5b89a

• pref unpref dataloaders and iterators

  gdharuman11

  committed
  4ef1df7

  Copy full SHA for 4ef1df7

• testing model forward and loss compute

  gdharuman11

  committed
  d89121e

  Copy full SHA for d89121e

• additional args for pref unpref data files

  gdharuman11

  committed
  e550a43

  Copy full SHA for e550a43

• returning tensor parallel logps

  gdharuman11

  committed
  a40fb14

  Copy full SHA for a40fb14

• tensor parallel logps for pref and unpref batches

  gdharuman11

  committed
  e1f0396

  Copy full SHA for e1f0396

• computing pref and unpref logps from reference model

  gdharuman11

  committed
  c22d090

  Copy full SHA for c22d090

Commits on Mar 23, 2024

• dpo loss func restructured

  gdharuman11

  committed
  5c3db71

  Copy full SHA for 5c3db71

Commits on Mar 25, 2024

• tracing the forward backward func

  gdharuman11

  committed
  8915f2a

  Copy full SHA for 8915f2a

• train step func

  gdharuman11

  committed
  eec8697

  Copy full SHA for eec8697

• training iteration

  gdharuman11

  committed
  09aa61b

  Copy full SHA for 09aa61b

Commits on Mar 27, 2024

• restored return in cross entropy

  gdharuman11

  committed
  6402e62

  Copy full SHA for 6402e62

• post processing cross entropy return

  gdharuman11

  committed
  1a04e84

  Copy full SHA for 1a04e84

• calls a different script

  gdharuman11

  committed
  f956f30

  Copy full SHA for f956f30

• working backprop

  gdharuman11

  committed
  66ae844

  Copy full SHA for 66ae844

• training log func

  gdharuman11

  committed
  de88521

  Copy full SHA for de88521

• loggin training metrics

  gdharuman11

  committed
  b50e140

  Copy full SHA for b50e140

Commits on Apr 2, 2024

• text-sequence data files

  gdharuman11

  committed
  24737c0

  Copy full SHA for 24737c0

Commits on Apr 5, 2024

• generate func post model training

  gdharuman11

  committed
  cbbd2a6

  Copy full SHA for cbbd2a6

• flops profiler enabled

  gdharuman11

  committed
  e1d3e6f

  Copy full SHA for e1d3e6f

• flops profiler enabled

  gdharuman11

  committed
  a976fd7

  Copy full SHA for a976fd7

• removed dtype assertion that throws error

  gdharuman11

  committed
  b964d58

  Copy full SHA for b964d58

• changed to generation after dpo

  gdharuman11

  committed
  ad780cb

  Copy full SHA for ad780cb

• newer preprocess data script

  gdharuman11

  committed
  6a3b03e

  Copy full SHA for 6a3b03e

• batch preprocess

  gdharuman11

  committed
  9cb0089

  Copy full SHA for 9cb0089

• convert to megatron format

  gdharuman11

  committed
  a2aba88

  Copy full SHA for a2aba88

• write megatron data file

  gdharuman11

  committed
  9e13d4e

  Copy full SHA for 9e13d4e

Commits on Apr 11, 2024

• Add @hzheng s example run_hzheng.sh

  saforem2

  committed
  702bb26

  Copy full SHA for 702bb26

• Prefer torch.optim for --optimizer

  saforem2

  committed
  8d048ac

  Copy full SHA for 8d048ac

• Remove global batch from run_hzheng.sh

  saforem2

  committed
  859ec88

  Copy full SHA for 859ec88

Commits on Jul 19, 2024

• minor changes

  

  gdharuman

  committed
  02961dc

  Copy full SHA for 02961dc

• scaling run launch script for dgx cluster

  

  gdharuman

  committed
  984fcf0

  Copy full SHA for 984fcf0

Commits on Jul 22, 2024

• minor changes

  

  gdharuman

  committed
  907d158

  Copy full SHA for 907d158

Commits on Aug 1, 2024

• dsflops log to file

  

  gdharuman

  committed
  f7f7da3

  Copy full SHA for f7f7da3

• all data

  

  gdharuman

  committed
  f91ba8c

  Copy full SHA for f91ba8c

• reference optimizer removed

  

  gdharuman

  committed
  be896de

  Copy full SHA for be896de

• with reference optimizer

  

  gdharuman

  committed
  67ff707

  Copy full SHA for 67ff707

• scaling 3.5B model

  gdharuman11

  committed
  4804eb7

  Copy full SHA for 4804eb7

• scaling 70B model

  gdharuman11

  committed
  14e2bf3

  Copy full SHA for 14e2bf3

• scaling 7B model

  gdharuman11

  committed
  71db627

  Copy full SHA for 71db627