Code

[WIP] Enable Real Varlen Attention#1065

Code

Draft

jiqing-feng wants to merge 15 commits intohuggingface:mainhuggingface/optimum-intel:mainfrom jiqing-feng:varlenjiqing-feng/optimum-intel:varlenCopy head branch name to clipboard

+150-38

Commits

Commits on Dec 11, 2024

• use real varlen attn

  Show description for 6d21075
  jiqing-feng

  committed
  6d21075

Commits on Dec 12, 2024

• optimize gpt2 by using linear instead of conv1D

  Show description for b792875
  jiqing-feng

  committed
  b792875

• Merge branch 'huggingface:main' into varlen

  jiqing-feng

  authored
  422134f

• fix usage without pkv

  Show description for 36884cb
  jiqing-feng

  committed
  36884cb

• use sdpa for no cache forward

  Show description for d061e69
  jiqing-feng

  committed
  d061e69

• fix format

  Show description for 31c635a
  jiqing-feng

  committed
  31c635a

• fix sdpa

  Show description for 73a5ef7
  jiqing-feng

  committed
  73a5ef7

• revert shape for sdpa

  Show description for f9c021b
  jiqing-feng

  committed
  f9c021b

• fix sdpa precision, still have error

  Show description for d069407
  jiqing-feng

  committed
  d069407

Commits on Dec 13, 2024

• fix sdpa shape

  Show description for 2c54045
  jiqing-feng

  committed
  2c54045

• upgrad minimum torch version to 2.5

  Show description for bce9aa9
  jiqing-feng

  committed
  bce9aa9

• rm pdb

  Show description for 72ac9e6
  jiqing-feng

  committed
  72ac9e6

Commits on Dec 16, 2024

• fix non patch path

  Show description for 3fdb3a5
  jiqing-feng

  committed
  3fdb3a5

Commits on Dec 18, 2024

• Merge branch 'main' into varlen

  jiqing-feng

  authored
  7e20b86

Commits on Dec 25, 2024

• Merge branch 'huggingface:main' into varlen

  jiqing-feng

  authored
  c1bd7f7