Code

feat(train): support deepspeed#1849

Code

Merged

robin1001 merged 17 commits intomainwenet-e2e/wenet:mainfrom xcsong-deepspeedwenet-e2e/wenet:xcsong-deepspeedCopy head branch name to clipboardMay 27, 2023

+481-63

Commits

Commits on May 10, 2023

• feat(train): support deepspeed

  xingchensong

  committed
  7c19311

  Copy full SHA for 7c19311

• feat(train): fix bug

  xingchensong

  committed
  79a5bc2

  Copy full SHA for 79a5bc2

• feat(train): enable deepspeed in run.sh

  xingchensong

  committed
  2a434ef

  Copy full SHA for 2a434ef

• feat(train): fix step bug for tensorboard logging

  xingchensong

  committed
  0c3d6d2

  Copy full SHA for 0c3d6d2

• feat(train): recover cv log tensorboard logging

  xingchensong

  committed
  d07c8c4

  Copy full SHA for d07c8c4

Commits on May 11, 2023

• feat(train): make save_states configurable

  xingchensong

  committed
  06ee7ba

  Copy full SHA for 06ee7ba

Commits on May 15, 2023

• feat(deepspeed): Support fp16/bf16 and deepspedCPUadam+customLRscheduler

  xingchensong

  committed
  4850e7f

  Copy full SHA for 4850e7f

• feat(deepspeed): fix lint

  xingchensong

  committed
  0c85d58

  Copy full SHA for 0c85d58

• feat(deepspeed): fix lint

  xingchensong

  committed
  3499b5c

  Copy full SHA for 3499b5c

• feat(deepspeed): update stage2 config

  xingchensong

  committed
  5350a7a

  Copy full SHA for 5350a7a

Commits on May 16, 2023

• feat(deepspeed): avoid re-generate filtered list if exists

  xingchensong

  committed
  761088a

  Copy full SHA for 761088a

Commits on May 17, 2023

• feat(deepspeed): add 1.8B model

  xingchensong

  committed
  8e60b06

  Copy full SHA for 8e60b06

Commits on May 19, 2023

• feat(deepspeed): make workers&prefetch configurable

  xingchensong

  committed
  46a3195

  Copy full SHA for 46a3195

Commits on May 25, 2023

• feat(deepspeed): refine comment

  xingchensong

  committed
  7ef383e

  Copy full SHA for 7ef383e

Commits on May 26, 2023

• feat(deepspeed): fix saving yaml

  xingchensong

  committed
  f4e3823

  Copy full SHA for f4e3823

• feat(deepspeed): refine if-else

  xingchensong

  committed
  0926c30

  Copy full SHA for 0926c30

• feat(deepspeed): refine if-else

  xingchensong

  committed
  63e8bd6

  Copy full SHA for 63e8bd6