Code

[ Misc ] fbgemm checkpoints#6559

Code

Merged

simon-mo merged 39 commits intovllm-project:mainvllm-project/vllm:mainfrom neuralmagic:fbgemm-checkpointsneuralmagic/nm-vllm:fbgemm-checkpointsCopy head branch name to clipboardJul 20, 2024

+234-47

Commits

Commits on Jul 18, 2024

• stash

  robertgshaw2-neuralmagic

  committed
  2426e29

• format

  robertgshaw2-neuralmagic

  committed
  7665b7b

• tweak arg name

  robertgshaw2-neuralmagic

  committed
  ef27613

• fix test

  robertgshaw2-neuralmagic

  committed
  2f96157

• format

  robertgshaw2-neuralmagic

  committed
  e748554

Commits on Jul 19, 2024

• working e2e with our cutlass kernels

  robertgshaw2-neuralmagic

  committed
  3ef571b

• added fp8 gemm

  robertgshaw2-neuralmagic

  committed
  ad83666

• remove

  robertgshaw2-neuralmagic

  committed
  eb7d48c

• format

  robertgshaw2-neuralmagic

  committed
  90bd839

• Merge branch 'main' into turn-on-fp8-dyn-per-token

  robertgshaw2-neuralmagic

  committed
  15cc823

• stash

  robertgshaw2-neuralmagic

  committed
  d064dd7

• dynamic per token

  robertgshaw2-neuralmagic

  committed
  6aa37e5

• format

  robertgshaw2-neuralmagic

  committed
  c9d819a

• reenable cutlass

  robertgshaw2-neuralmagic

  committed
  08cbaf7

• cleanup comment

  robertgshaw2-neuralmagic

  committed
  f4cdda1

• format

  robertgshaw2-neuralmagic

  committed
  2971f4d

• added dynamic per token test case

  robertgshaw2-neuralmagic

  committed
  b601033

• Merge branch 'turn-on-fp8-dyn-per-token' into fbgemm-checkpoints

  robertgshaw2-neuralmagic

  committed
  5d8edf9

• added use per token

  robertgshaw2-neuralmagic

  committed
  8b5d638

• format

  

  Varun Sundar Rabindranth

  committed
  006ccf0

• format

  

  Varun Sundar Rabindranth

  committed
  1884acf

• Make optional ubs none

  

  Varun Sundar Rabindranth

  committed
  fe14072

• format

  

  Varun Sundar Rabindranth

  committed
  254dcff

• Merge branch 'fp8-dpt-fpgemm' into fbgemm-checkpoints

  robertgshaw2-neuralmagic

  committed
  919d866

• hook up end to end with varun's ub quant kernel

  robertgshaw2-neuralmagic

  committed
  227a277

• formatted

  robertgshaw2-neuralmagic

  committed
  951834a

• updated for nonuniform

  robertgshaw2-neuralmagic

  committed
  9aa66d3

• formatting after passing prefix around

  robertgshaw2-neuralmagic

  committed
  458a410

• Merge branch 'main' into fbgemm-checkpoints

  robertgshaw2-neuralmagic

  committed
  278f6d6

• fixed bad merge

  robertgshaw2-neuralmagic

  committed
  3e4aaad

Commits on Jul 20, 2024

• updated message

  robertgshaw2-neuralmagic

  committed
  de2a764

• Merge branch 'main' into fbgemm-checkpoints

  robertgshaw2-neuralmagic

  committed
  268fe94

• merged varun's pr

  robertgshaw2-neuralmagic

  committed
  c88fe34

• fixed

  robertgshaw2-neuralmagic

  committed
  bb02a3f

• cleanup pr

  robertgshaw2-neuralmagic

  committed
  1c8f71c

• Update config.py

  robertgshaw2-neuralmagic

  authored
  6970e50

• fixed config

  robertgshaw2-neuralmagic

  committed
  94617f0

• updated for new ckpt format, turned on ada lovelace, and added test case

  robertgshaw2-neuralmagic

  committed
  f9d569c

• format

  robertgshaw2-neuralmagic

  committed
  ae45615