Commit History

  1. feat: vmap optimizer (#166)

    Unverified signature
    b993d27

  1. feat: scan layers + gradient checkpointing (#161)⁠

    Unverified signature
    07a6f9a

  1. feat: better multi-node support (#158)⁠

    Unverified signature
    728a3c3

  1. feat: update shampoo

    9ecdd3f

  1. feat(model): allow bias (#152)

    Unverified signature
    361a994

  1. feat(train): google-cloud-storage is optional

    02b2308
  2. feat(train): rename logged config

    955dc20

  1. feat: add mini_glu config

    a7e5050

  1. feat: update mini config

    d9a16f2

  1. feat: implement transformer variants (#144)⁠

    Unverified signature
    542378c

  1. feat(train): log norm and histograms (#143)⁠

    Unverified signature
    b7b619a

  1. feat(data): super conditioning (#141)⁠

    Unverified signature
    7939874

  1. feat: support pod (#139)

    Unverified signature
    803ccbf

  1. fix: no gradient checkpointing for new model

    2e02683
  2. feat: no gradient checkpointing for params init

    b798ed3
  3. feat: update configs

    79557f9
  4. fix(train): consider schedule offset

    bc4734f
  5. feat(train): local jax cache

    9f5e879
  6. feat: add bucket reference to artifact

    d368fb6

  1. style: lint

    d5d442a
  2. feat: handle gradient checkpointing

    5173ec7
  3. feat: load from bucket

    1c4e839
  4. feat(train): save to bucket

    50498e6
  5. feat: reduce artifact space + offset step

    34cf91c

  1. feat: restore weights on CPU

    5f954fc
  2. feat(train): simplify tokenizer loading

    4cb21dd

  1. feat(train): use compilation cache

    da9367c

  1. fix: style

    d483294

  1. Merge branch 'main' of https://github.com/borisdayma/dalle-mini into main

    0a691de
  2. feat: log num_parameters early

    7cfe576
  3. fix: distributed shampoo class

    696422e
  4. feat: update distributed_shampoo

    5996680

  1. feat(train) - handle multiple nodes (#130)

    Unverified signature
    0952927
  2. feat: handle model parallel

    1bb3269

  1. feat(train): more custom x-axis

    5f28cd2

  1. fix(train): opt_state_shape for distributed_shampoo

    225b6ff

  1. feat(train): split artifact into model/state

    fa5b058

  1. feat(train): another 25% faster

    14abe8c
  2. feat(train): overhead from 70% to 1% 🥳

    2b7f5f1
  3. feat(pjit): follow t5x style

    7b5868f

  1. fix(train): grads spec

    00710bc
  2. feat(train): improve pjit speed

    f254058
  3. fix(train): consider correct batch size

    b7c7458

  1. feat(train): custom start_preconditioning_step

    8149924
  2. feat(train): handle distributed_shampoo in pjit

    032f623
  3. feat: update distributed_shampoo + fix None spec

    8a9e367
  4. feat(train): distributed_shampoo with pjit

    cc34d07

  1. fix style

    f044cb8
  2. feat(train): restore opt_state efficiently

    1bfc1b5

  1. feat(model): clean way to load on cpu

    12f323d