Commits

Branch selector

 dev

User selector

All users

Datepicker

All time

Commit History

Commits on Jan 4, 2026

• tiny fixes

  ngc92

  committed
  2adae39

  Copy full SHA for 2adae39

Commits on Dec 29, 2025

• update readme descriptions

  ngc92

  committed
  453c54b

  Copy full SHA for 453c54b

Commits on Dec 19, 2025

• small text fixes and added profiling commands; 8xH100 results

  ngc92

  committed
  dcad498

  Copy full SHA for dcad498

• don't call reduce_scatter in single worker setups

  ngc92

  committed
  6b730c2

  Copy full SHA for 6b730c2

Commits on Dec 18, 2025

• update expected losses/norms after scale fix

  ngc92

  committed
  0d911e2

  Copy full SHA for 0d911e2

• formatting

  ngc92

  committed
  bcceafa

  Copy full SHA for bcceafa

• bugfix: missing abs

  ngc92

  committed
  218fd7d

  Copy full SHA for 218fd7d

• add outlier monitoring

  ngc92

  committed
  161502f

  Copy full SHA for 161502f

• yet more movement to common base

  ngc92

  committed
  f8003a2

  Copy full SHA for f8003a2

• moved debug timing to logger

  ngc92

  committed
  5afca1d

  Copy full SHA for 5afca1d

• further renaming

  ngc92

  committed
  9bee58b

  Copy full SHA for 9bee58b

• rename llama_config to transformer_config

  ngc92

  committed
  31258fe

  Copy full SHA for 31258fe

• move more elements into the generic run state

  ngc92

  committed
  747ceb3

  Copy full SHA for 747ceb3

• generic model state

  ngc92

  committed
  5c37b3a

  Copy full SHA for 5c37b3a

• trigger logging

  ngc92

  committed
  23d6cbb

  Copy full SHA for 23d6cbb

• wandb improvements

  ngc92

  committed
  6bd1171

  Copy full SHA for 6bd1171

• add logging of abs-maxes used for quantization

  ngc92

  committed
  ebb2a45

  Copy full SHA for ebb2a45

Commits on Dec 17, 2025

• codegen separate kernels for with/without abs-max

  ngc92

  committed
  7bc3068

  Copy full SHA for 7bc3068

• bugfix

  ngc92

  committed
  04ad980

  Copy full SHA for 04ad980

• fix ldgsts performance problem in swiglu

  ngc92

  committed
  74fc910

  Copy full SHA for 74fc910

• swiglu: remove divisions

  ngc92

  committed
  5a7a65e

  Copy full SHA for 5a7a65e

• cache hints for vector store

  ngc92

  committed
  aeb2db5

  Copy full SHA for aeb2db5

• add a test file for classifier

  ngc92

  committed
  63c159f

  Copy full SHA for 63c159f

Commits on Dec 16, 2025

• fix potential infinite loop in fused_classifier

  ngc92

  committed
  baff7b6

  Copy full SHA for baff7b6

• allow continuing from specific checkpoint

  ngc92

  committed
  679b300

  Copy full SHA for 679b300

• fix wandb export

  ngc92

  committed
  604d393

  Copy full SHA for 604d393

Commits on Dec 15, 2025

• fix json logging

  ngc92

  committed
  14ed420

  Copy full SHA for 14ed420

• avoid deadlock in case of GPU oversubscription

  ngc92

  committed
  229ba2f

  Copy full SHA for 229ba2f

• add lr cooldown option

  ngc92

  committed
  d6e84cf

  Copy full SHA for d6e84cf

• update README.md with new results and reference to benchmark folders

  ngc92

  committed
  7f895bd

  Copy full SHA for 7f895bd

• detailed performance baselines for 4090

  ngc92

  committed
  bc5f776

  Copy full SHA for bc5f776

• better logging

  ngc92

  committed
  6c570e2

  Copy full SHA for 6c570e2

• allow specifying data-loader seeds

  ngc92

  committed
  3dace66

  Copy full SHA for 3dace66

• make an attempt to copy basic config files to have the generated model work out of the box with transformers/vllm

  ngc92

  committed
  061e83e

  Copy full SHA for 061e83e

• allow specifying number of epochs instead of number of steps

  ngc92

  committed
  767c763

  Copy full SHA for 767c763

Pagination

Previous

Next