Code

Add FP16Act-FP6Weight Linear#223

Code

Merged

msaroufim merged 44 commits intopytorch:mainpytorch/ao:mainfrom gau-nernst:fp6gau-nernst/ao:fp6Copy head branch name to clipboardMay 14, 2024

+1,882-2

Commits

Commits on May 7, 2024

• add files from fp6_llm

  gau-nernst

  committed
  c742a0e

  Copy full SHA for c742a0e

• Merge branch 'pytorch:main' into fp6

  gau-nernst

  authored
  963bfa6

  Copy full SHA for 963bfa6

• try to port weight packing first

  gau-nernst

  committed
  4eb8be6

  Copy full SHA for 4eb8be6

• rename

  gau-nernst

  committed
  8608664

  Copy full SHA for 8608664

• rename fp6 weight packing

  gau-nernst

  committed
  b7c7b28

  Copy full SHA for b7c7b28

• add fp16act_fp6weight_linear

  gau-nernst

  committed
  3c9aac7

  Copy full SHA for 3c9aac7

• fix function def

  gau-nernst

  committed
  031379a

  Copy full SHA for 031379a

• Merge branch 'main' into fp6

  msaroufim

  authored
  75ca602

  Copy full SHA for 75ca602

Commits on May 8, 2024

• delete duplicate file

  gau-nernst

  committed
  c436c43

  Copy full SHA for c436c43

• move weight quant file

  gau-nernst

  committed
  12823fe

  Copy full SHA for 12823fe

• rename

  gau-nernst

  committed
  9180fef

  Copy full SHA for 9180fef

• add pytorch interface for fp6 weight dequant

  gau-nernst

  committed
  1b24424

  Copy full SHA for 1b24424

• add fake_fp6 to fp6

  gau-nernst

  committed
  2671c9c

  Copy full SHA for 2671c9c

• move weight_quant to csrc/cuda due to cuda_fp16.h dependency

  gau-nernst

  committed
  e61be51

  Copy full SHA for e61be51

• add fake_fp6_to_fp6 test

  gau-nernst

  committed
  21acfd1

  Copy full SHA for 21acfd1

• add test for fp16act_fp6weight_linear

  gau-nernst

  committed
  67fd6f8

  Copy full SHA for 67fd6f8

• add test for fp6_weight_dequant

  gau-nernst

  committed
  084b7e4

  Copy full SHA for 084b7e4

• Fp6WeightOnlyQuantizedLinearWeight (not working yet)

  gau-nernst

  committed
  6d2fc3e

  Copy full SHA for 6d2fc3e

• skip some tests, since the functions are not built w/o CUDA

  gau-nernst

  committed
  68f2415

  Copy full SHA for 68f2415

Commits on May 9, 2024

• Merge branch 'main' into fp6

  gau-nernst

  authored
  0c78635

  Copy full SHA for 0c78635

• add the original test

  gau-nernst

  committed
  5989599

  Copy full SHA for 5989599

• implement transpose and clone so that F.linear will work

  gau-nernst

  committed
  92dfde4

  Copy full SHA for 92dfde4

• remove print

  gau-nernst

  committed
  da1421b

  Copy full SHA for da1421b

• Merge branch 'main' into fp6

  gau-nernst

  authored
  fecd0cc

  Copy full SHA for fecd0cc

• remove dequantize

  gau-nernst

  committed
  a0a53a0

  Copy full SHA for a0a53a0

• add notes and some rename

  gau-nernst

  committed
  079e16b

  Copy full SHA for 079e16b

• typo

  gau-nernst

  committed
  06e8438

  Copy full SHA for 06e8438

• small cleanup

  gau-nernst

  committed
  ca45274

  Copy full SHA for ca45274

• improve tensor subclass and add test (which is failing for torch-compile)

  gau-nernst

  committed
  7a0f6e2

  Copy full SHA for 7a0f6e2

• add note

  gau-nernst

  committed
  320827e

  Copy full SHA for 320827e

• add note

  gau-nernst

  committed
  74b8094

  Copy full SHA for 74b8094

• add qtorch as dev requirement

  gau-nernst

  committed
  c8d47c3

  Copy full SHA for c8d47c3

• update error message

  gau-nernst

  committed
  e08ba6a

  Copy full SHA for e08ba6a

• add __repr__ and fix transposed issue

  gau-nernst

  committed
  b090b4b

  Copy full SHA for b090b4b

• add fp6 perplexity test

  gau-nernst

  committed
  f6f93c3

  Copy full SHA for f6f93c3

Commits on May 10, 2024

• rename variables

  gau-nernst

  committed
  b857645

  Copy full SHA for b857645

• Merge branch 'main' into fp6

  gau-nernst

  committed
  e69e6db

  Copy full SHA for e69e6db

• remove subclass

  gau-nernst

  committed
  f0eba1a

  Copy full SHA for f0eba1a

• add correctness test

  gau-nernst

  committed
  8f1ef8d

  Copy full SHA for 8f1ef8d

• remove unwanted changes

  gau-nernst

  committed
  cb05d30

  Copy full SHA for cb05d30

• add apache 2.0 notice

  gau-nernst

  committed
  56aefc6

  Copy full SHA for 56aefc6

• add benchmark script

  gau-nernst

  committed
  7d3a5b1

  Copy full SHA for 7d3a5b1

Commits on May 14, 2024

• add note about FP6 kernel

  gau-nernst

  committed
  08a95ac

  Copy full SHA for 08a95ac

• relax tolerance

  gau-nernst

  committed
  a8b4dd3

  Copy full SHA for a8b4dd3