Code

Benchmarks - Add LLaMA-2 Models#668

Code

Merged

abuccts merged 30 commits intomicrosoft:mainmicrosoft/superbenchmark:mainfrom dpower4:feat/llamadpower4/superbenchmark:feat/llamaCopy head branch name to clipboardNov 28, 2024

+408-7

Commits

Commits on Nov 18, 2024

• add llama init template

  dpower4

  committed
  36bbf10

• add llama2 unit test

  dpower4

  committed
  697138a

• fix dims for llama2 unit test

  dpower4

  committed
  9355e22

• update transformers version for LLamaConfig

  dpower4

  committed
  60eae36

• update docs

  dpower4

  committed
  dadb56a

Commits on Nov 19, 2024

• update opset for torch onnx conversion

  dpower4

  committed
  d2731a8

• format and lint

  dpower4

  committed
  3644985

• remove remnant

  dpower4

  committed
  52f4900

Commits on Nov 20, 2024

• lint fix

  dpower4

  committed
  f826676

• replace py 3.6 with 3.10 and update cuda to 12.4 for unit test

  dpower4

  committed
  6a41087

• remove 3.6 from setup, codecov and docs

  dpower4

  committed
  5b816b4

• add llama fp8 unit test for better code coverage

  dpower4

  committed
  f322c98

• llama fp8 precision test only, to reduce memory required

  dpower4

  committed
  b28ee17

• lint fix

  dpower4

  committed
  e6f6be3

Commits on Nov 21, 2024

• remove deprecated NaN usage for numpy>2.0

  dpower4

  committed
  297a229

• fix argparse formatting related test cases failure for 3.10

  dpower4

  committed
  5f72f51

• fix lint

  dpower4

  committed
  8bbe326

• fix lint

  dpower4

  committed
  50452ef

• add llama2 to tensorrt

  dpower4

  committed
  0b1da4f

Commits on Nov 22, 2024

• Merge branch 'main' into feat/llama

  dpower4

  authored
  97b3d72

• add more params to llama config

  

  dpatlolla

  committed
  e423aec

• fix lint

  dpower4

  committed
  d850210

• Merge branch 'main' into feat/llama

  abuccts

  authored
  b08f9e3

Commits on Nov 27, 2024

• Merge branch 'main' into feat/llama

  abuccts

  authored
  54c3e85

• llama test: use fp16 instead of fp8 to relax cuda CC req.

  dpower4

  committed
  670dc76

• fix comment and lint

  dpower4

  committed
  27c788c

• fix precision arg as float16

  dpower4

  committed
  00d09ba

• limit tokenizers version to < 0.20.3 as 0.20.4 doesnt support py3.8

  dpower4

  committed
  bd47fc3

Commits on Nov 28, 2024

• address review comments

  dpower4

  committed
  bed3e01

• Merge branch 'main' into feat/llama

  abuccts

  authored
  1570707