Code

fp8 bwd#108

Code

Draft

micmelesse wants to merge 13 commits intomain_perfROCm/flash-attention:main_perffrom micmelesse/fp8ROCm/flash-attention:micmelesse/fp8Copy head branch name to clipboard

+1,001-62

Commits

Commits on Dec 9, 2024

• Alex fp8 work

  Show description for 94b2da3
  

  alexkranias-amd

  authored and

  micmelesse

  committed
  94b2da3

• fix mismatches

  micmelesse

  committed
  7434112

• no navi for now

  micmelesse

  committed
  2ea54b1

• fix: ref uses scaling + added ENV VAR to enable/disable quantization scaling

  alexkranias-amd

  committed
  89d3d7d

• fix: fp8 ref matches kernel

  alexkranias-amd

  committed
  c65af82

• misc: added note about p_scale

  alexkranias-amd

  committed
  9297d78

• feat: added precision error test for various triton ops

  alexkranias-amd

  committed
  f92ca5b

Commits on Dec 10, 2024

• save

  alexkranias-amd

  committed
  9ed1d00

• feat: added benchmark for fp8 flash attention

  alexkranias-amd

  committed
  1c3f756

• fix: quantization scaling in fp8 benchmark

  alexkranias-amd

  committed
  c4ca789

• checkpoint

  alexkranias-amd

  committed
  937e814

Commits on Dec 11, 2024

• feat: added fp8 to precision test

  alexkranias-amd

  committed
  fd342f7

Commits on Dec 13, 2024

• fix: refactor fp32 for torch, moved scaling of fp8 to out of kernel

  alexkranias-amd

  committed
  543736b