Code

Fp8 integration#1086

Code

Merged

sgugger merged 34 commits intomainhuggingface/accelerate:mainfrom fp8_integrationhuggingface/accelerate:fp8_integrationCopy head branch name to clipboardMar 7, 2023

+334-54

Commits

Commits on Feb 15, 2023

• Draft of FP8 support

  sgugger

  committed
  68b98fc

  Copy full SHA for 68b98fc

• Missing import

  sgugger

  committed
  f10a4b0

  Copy full SHA for f10a4b0

• Fix names

  sgugger

  committed
  23b0e04

  Copy full SHA for 23b0e04

• Conversion is inplace

  sgugger

  committed
  12ac9bf

  Copy full SHA for 12ac9bf

• Enable fp8 in examples

  sgugger

  committed
  ab878ac

  Copy full SHA for ab878ac

• Customization point for Recipe

  sgugger

  committed
  b5578ea

  Copy full SHA for b5578ea

• Auto-enable FP8 depending on compute capability

  sgugger

  committed
  b818aee

  Copy full SHA for b818aee

• Fix typo

  sgugger

  committed
  9e7157e

  Copy full SHA for 9e7157e

• Put back mixed precision arg

  sgugger

  committed
  a1a86d1

  Copy full SHA for a1a86d1

• Add debug script

  sgugger

  committed
  94d7101

  Copy full SHA for 94d7101

• Add more tests in debug

  sgugger

  committed
  13deee8

  Copy full SHA for 13deee8

• Add more stuff to debug

  sgugger

  committed
  097d3eb

  Copy full SHA for 097d3eb

• Don't forget train

  sgugger

  committed
  23d188f

  Copy full SHA for 23d188f

• Put the train in the right place

  sgugger

  committed
  4656b23

  Copy full SHA for 4656b23

• Add options for selective conversion

  sgugger

  committed
  17d0d38

  Copy full SHA for 17d0d38

• Fix typo

  sgugger

  committed
  aa3c0c2

  Copy full SHA for aa3c0c2

• Properly recurse

  sgugger

  committed
  8c69cd4

  Copy full SHA for 8c69cd4

• Add more debug utils

  sgugger

  committed
  4df14d0

  Copy full SHA for 4df14d0

• Typo and init

  sgugger

  committed
  978ed82

  Copy full SHA for 978ed82

• Last choice

  sgugger

  committed
  a2efc93

  Copy full SHA for a2efc93

• More fixes

  sgugger

  committed
  583c5c5

  Copy full SHA for 583c5c5

• More options in example

  sgugger

  committed
  cb2b2f1

  Copy full SHA for cb2b2f1

• Remove debug scripts

  sgugger

  committed
  7deee02

  Copy full SHA for 7deee02

• Clean up debug and new names

  sgugger

  committed
  e3e2ee9

  Copy full SHA for e3e2ee9

• Add torch.no_grad for conversion

  sgugger

  committed
  659e1dc

  Copy full SHA for 659e1dc

• Optimizer is deconnected from model?

  sgugger

  committed
  9c715b0

  Copy full SHA for 9c715b0

• Re-attach model parameters to optimizer

  sgugger

  committed
  b3f997f

  Copy full SHA for b3f997f

• Fix extract

  sgugger

  committed
  2787834

  Copy full SHA for 2787834

• Style

  sgugger

  committed
  e2070df

  Copy full SHA for e2070df

• Cleanup post-rebase

  sgugger

  committed
  6f21320

  Copy full SHA for 6f21320

• Deal with apdding

  sgugger

  committed
  063d33e

  Copy full SHA for 063d33e

• fix examples

  sgugger

  committed
  9f6b40f

  Copy full SHA for 9f6b40f

Commits on Mar 2, 2023

• Update src/accelerate/accelerator.py

  Show description for 73e06ac
  

  sgugger

  and

  pacman100

  authored
  73e06ac

  Copy full SHA for 73e06ac

• Address comments

  sgugger

  committed
  10f879a

  Copy full SHA for 10f879a