Code

PTQ for generate_v2#1866

Code

Open

joecummings wants to merge 20 commits intopytorch:mainpytorch/torchtune:mainfrom joecummings:add-quantize-generate-v2joecummings/torchtune:add-quantize-generate-v2Copy head branch name to clipboard

+114-33

Commits

Commits on Oct 18, 2024

• [WIP] Quantization for generate_v2

  joecummings

  committed
  f89fdd4

• Update config

  joecummings

  committed
  86b7784

• Add initial test for quantization

  joecummings

  committed
  e006f78

• Remove annoying logging errors due to atexit usage in _dynamo

  joecummings

  committed
  eafd3b2

Commits on Oct 26, 2024

• Merge remote-tracking branch 'upstream/main' into add-quantize-generate-v2

  joecummings

  committed
  e09f1a1

• Update comment for what is not supported

  joecummings

  committed
  0575b67

Commits on Oct 29, 2024

• Don't claim that HQQ is better

  joecummings

  committed
  f318412

• Debugging code for CI

  joecummings

  committed
  2faf50c

• Try smaller but more powerful, newer G5 runner

  joecummings

  committed
  80bb4e3

• Fix linting dummy

  joecummings

  committed
  ff2ffba

• Specify 2.5.1 in runer

  joecummings

  committed
  322c802

• Dumb typo

  joecummings

  committed
  24508fb

• Hopefully fix formatting to pick up GPU tests

  joecummings

  committed
  67718b9

• Switch order of these things b/c apparantly torchvision installs stable

  joecummings

  committed
  a864ef9

Commits on Oct 30, 2024

• 2.5.1 was released so we're all good

  joecummings

  committed
  31f64b5

• Test reseting compile before test

  joecummings

  committed
  fc501ee

Commits on Oct 31, 2024

• Comment out LoRA just for testing

  joecummings

  committed
  3594697

• TORCHINDUCTOR DISABLE CACHES

  joecummings

  committed
  d03db9f

• Set backend to eager

  joecummings

  committed
  ae2cf5b

• Fix linting

  joecummings

  committed
  23c9bb7