Code

DeepSeek Support FP8#9956

Code

Open

ming1753 wants to merge 14 commits intoPaddlePaddle:developPaddlePaddle/PaddleNLP:developfrom ming1753:deepseek-v3-mla-fp8ming1753/PaddleNLP:deepseek-v3-mla-fp8Copy head branch name to clipboard

+6,935-36

Commits

Commits on Feb 27, 2025

• DeekSeek suport FP8

  ming1753

  committed
  ef90990

  Copy full SHA for ef90990

• fix weight only

  ming1753

  committed
  56bd113

  Copy full SHA for 56bd113

• fix bug

  ming1753

  committed
  28c86d7

  Copy full SHA for 28c86d7

• modify api

  ming1753

  committed
  8df1723

  Copy full SHA for 8df1723

Commits on Feb 28, 2025

• support fp8 block gemm and group quant

  ckl117

  committed
  03be9e1

  Copy full SHA for 03be9e1

• support fp8 gemm with tensor scale tune on sm90

  ckl117

  committed
  3e77662

  Copy full SHA for 3e77662

• support read weight_block_size from model config.json

  ckl117

  committed
  04b7864

  Copy full SHA for 04b7864

• check Copyright

  ckl117

  committed
  43cbaa2

  Copy full SHA for 43cbaa2

• Merge pull request #1 from ckl117/dsv3-fp8

  Show description for 7c23af9
  ming1753

  authored
  7c23af9

  Copy full SHA for 7c23af9

• check

  ckl117

  committed
  63e0062

  Copy full SHA for 63e0062

• move tune_fp8_gemm.sh to csrc

  ckl117

  committed
  8c8bed0

  Copy full SHA for 8c8bed0

• commit by zkk

  zhoutianzi666

  committed
  e1a415d

  Copy full SHA for e1a415d

• commit by zkk

  zhoutianzi666

  committed
  ebd1d70

  Copy full SHA for ebd1d70

• Merge pull request #7 from zhoutianzi666/add_moe_fp81

  Show description for 4e168e5
  ming1753

  authored
  4e168e5

  Copy full SHA for 4e168e5