Code

Generate: end-to-end compilation#30788

Code

Merged

gante merged 31 commits intohuggingface:mainhuggingface/transformers:mainfrom gante:end_to_end_mvpgante/transformers:end_to_end_mvpCopy head branch name to clipboardJul 29, 2024

+282-100

Commits

Commits on Jul 27, 2024

• mvp

  gante

  committed
  a984179

• added test (a few models need fixes)

  gante

  committed
  f0336ed

• fix a few test cases

  gante

  committed
  06983d8

• test nits

  gante

  committed
  17580dc

• harder test 😈

  gante

  committed
  d06e999

• revert changes in stablelm

  gante

  committed
  903b81f

• test with improved condition

  gante

  committed
  08cc2c0

• add todo

  gante

  committed
  683f3e7

• tmp commit

  gante

  committed
  e063fc8

• merged with main

  gante

  committed
  0ebc4c7

• nits

  gante

  committed
  86c7170

• add todo

  gante

  committed
  b2b6001

• final corrections

  gante

  committed
  2fcc207

• add docs for generation compilation

  gante

  committed
  e84aedb

• docs nits

  gante

  committed
  d2b45a4

• add tip

  gante

  committed
  64ce18b

• PR suggestions

  gante

  committed
  ef4d419

• add more details to the compilation docs

  gante

  committed
  d5e920d

• fix cache positions

  gante

  committed
  40482d3

• cache is now init in generate; update docs

  gante

  committed
  e3d9c04

• tag test as flaky

  gante

  committed
  139e212

• docs

  gante

  committed
  bc4ad7d

• post rebase make fixup and other nits

  gante

  committed
  54c9eef

• remove unintended changes

  gante

  committed
  3186b14

• whisper (encoder-decoder) not supported

  gante

  committed
  484d922

• move token default updates to ; add tests for token defaults

  gante

  committed
  bf9ef8a

• push changes

  gante

  committed
  f2e2833

• manual rebase

  gante

  committed
  16f92f4

• chameleon doesn't support this

  gante

  committed
  838ba6a

• fix test_static_cache_mha_mqa_gqa (broken in another PR)

  gante

  committed
  795d058

• docs: dynamic is better with end-to-end compilation

  gante

  committed
  d2e423b