Code

[automodel] re-enable FSDP2 tests#12325

Code

Draft

akoumpa wants to merge 19 commits intomainNVIDIA/NeMo:mainfrom akoumparouli/fix_testNVIDIA/NeMo:akoumparouli/fix_testCopy head branch name to clipboard

+170-74

Commits

Commits on Feb 22, 2025

• use /tmp for HF_HOME

  Show description for 29b9237
  akoumpa

  committed
  29b9237

  Copy full SHA for 29b9237

• re-enable tests

  Show description for 86f71e8
  akoumpa

  committed
  86f71e8

  Copy full SHA for 86f71e8

• fix

  Show description for 02a0b35
  akoumpa

  committed
  02a0b35

  Copy full SHA for 02a0b35

• fix fsdp2 strat

  Show description for 2069c9e
  akoumpa

  committed
  2069c9e

  Copy full SHA for 2069c9e

• use _init_model_parallel & _setup_optimizers to bypass PTL init for PEFT

  Show description for c570f78
  akoumpa

  committed
  c570f78

  Copy full SHA for c570f78

• guard setup_optim with trainer.state

  Show description for 10dadef
  akoumpa

  committed
  10dadef

  Copy full SHA for 10dadef

• improve logs

  Show description for 5bdb43d
  akoumpa

  committed
  5bdb43d

  Copy full SHA for 5bdb43d

• use fsdp2 in vlm test

  Show description for 3b43423
  akoumpa

  committed
  3b43423

  Copy full SHA for 3b43423

• Apply isort and black reformatting

  Show description for c3db35b
  akoumpa

  committed
  c3db35b

  Copy full SHA for c3db35b

• Update hf_auto_model_for_image_text_to_text.py

  Show description for c7e77ef
  akoumpa

  authored
  c7e77ef

  Copy full SHA for c7e77ef

• Update fsdp2_strategy.py

  Show description for fc15f27
  akoumpa

  authored
  fc15f27

  Copy full SHA for fc15f27

• add missing import

  Show description for cde4250
  akoumpa

  committed
  cde4250

  Copy full SHA for cde4250

• add auto-resume for peft test

  Show description for 0bb8c08
  akoumpa

  committed
  0bb8c08

  Copy full SHA for 0bb8c08

Commits on Feb 23, 2025

• update save_pretraned

  Show description for 60e1828
  akoumpa

  committed
  60e1828

  Copy full SHA for 60e1828

• make test mandatory

  Show description for 356481c
  akoumpa

  committed
  356481c

  Copy full SHA for 356481c

• reduce accumulate_grad_batches 10 -> 2

  Show description for 2c48ade
  akoumpa

  committed
  2c48ade

  Copy full SHA for 2c48ade

• add missing import

  Show description for e153e7c
  akoumpa

  committed
  e153e7c

  Copy full SHA for e153e7c

• fix test

  Show description for a6b5309
  akoumpa

  committed
  a6b5309

  Copy full SHA for a6b5309

• Apply isort and black reformatting

  Show description for 8cc3b7c
  akoumpa

  committed
  8cc3b7c

  Copy full SHA for 8cc3b7c