Code

Make training on the MPS device work#131

Code

Open

dirkgr wants to merge 19 commits intomainallenai/OLMo-core:mainfrom dirkg/MPSallenai/OLMo-core:dirkg/MPSCopy head branch name to clipboard

+52-52

Commits

Commits on Jan 2, 2025

• New check

  dirkgr

  committed
  76e3f82

  Copy full SHA for 76e3f82

• Build RNG on device

  dirkgr

  committed
  637c676

  Copy full SHA for 637c676

• World size is 1 when we don't have a process group

  dirkgr

  committed
  1bea8c0

  Copy full SHA for 1bea8c0

• Only use the GPU manager when CUDA is available

  dirkgr

  committed
  ca78850

  Copy full SHA for ca78850

• Set default device when there is no process group

  dirkgr

  committed
  9fccb4d

  Copy full SHA for 9fccb4d

• Set the MPS device as default if available

  dirkgr

  committed
  49c0213

  Copy full SHA for 49c0213

• Turn off compile when there is no CUDA

  dirkgr

  committed
  5050c70

  Copy full SHA for 5050c70

• If there is no cluster, use GS for storage.

  dirkgr

  committed
  50353e7

  Copy full SHA for 50353e7

• Command to train on a single device

  dirkgr

  committed
  8083774

  Copy full SHA for 8083774

• Adds help for train_single

  dirkgr

  committed
  51a9294

  Copy full SHA for 51a9294

• CUDA specific code again

  dirkgr

  committed
  fb4466c

  Copy full SHA for fb4466c

• Don't have the dataset cache in git

  dirkgr

  committed
  0dc158b

  Copy full SHA for 0dc158b

Commits on Jan 17, 2025

• Merge branch 'main' into dirkg/MPS

  dirkgr

  authored
  2d676bd

  Copy full SHA for 2d676bd

• Fix config for train_single

  dirkgr

  committed
  5715ef3

  Copy full SHA for 5715ef3

• Those evals don't exist.

  dirkgr

  committed
  64fc957

  Copy full SHA for 64fc957

• Changelog

  dirkgr

  committed
  c824ca5

  Copy full SHA for c824ca5

• Unused import

  dirkgr

  committed
  86b218e

  Copy full SHA for 86b218e

• Formatting some stuff I touched

  dirkgr

  committed
  a622aa4

  Copy full SHA for a622aa4

Commits on Jan 18, 2025

• Merge branch 'main' into dirkg/MPS

  dirkgr

  authored
  313a234

  Copy full SHA for 313a234