Code

support ctrcvr sharding mode#70

Code

Merged

qingshui merged 34 commits intojiaoxuewu:paddleboxjiaoxuewu/PaddleBox:paddleboxfrom qingshui:paddleboxqingshui/Paddle:paddleboxCopy head branch name to clipboardDec 1, 2023

+1,538-683

Commits

Commits on Sep 5, 2023

• return count of nan and inf (#90)

  Show description for 52950fc
  

  gumplh

  and

  lihui53

  authored
  52950fc

  Copy full SHA for 52950fc

Commits on Sep 15, 2023

• paddlebox sharding (#91)

  Show description for 14f97ae
  

  gumplh

  and

  lihui53

  authored
  14f97ae

  Copy full SHA for 14f97ae

• fix sharding var error

  

  humingqing

  authored and

  humingqing

  committed
  4106e30

  Copy full SHA for 4106e30

• fix sharding var error

  

  humingqing

  authored and

  humingqing

  committed
  ba25686

  Copy full SHA for ba25686

Commits on Sep 18, 2023

• fix block bug

  

  humingqing

  authored and

  humingqing

  committed
  2e3ff0a

  Copy full SHA for 2e3ff0a

• format

  

  humingqing

  authored and

  humingqing

  committed
  6e6cc23

  Copy full SHA for 6e6cc23

Commits on Sep 19, 2023

• load param by thread

  

  humingqing

  authored and

  humingqing

  committed
  d34f370

  Copy full SHA for d34f370

• load param by multi thread

  

  humingqing

  authored and

  humingqing

  committed
  6271a06

  Copy full SHA for 6271a06

Commits on Sep 21, 2023

• support sharded dump, and optimize dump performance, reformat code

  

  humingqing

  authored and

  humingqing

  committed
  0748e8a

  Copy full SHA for 0748e8a

• param sharding new split

  

  humingqing

  authored and

  humingqing

  committed
  3824727

  Copy full SHA for 3824727

• fix paddlebox scale

  

  humingqing

  authored and

  humingqing

  committed
  640ab78

  Copy full SHA for 640ab78

Commits on Sep 26, 2023

• fix large model blocking and add sharding stage1 mode

  

  humingqing

  authored and

  humingqing

  committed
  bff41ed

  Copy full SHA for bff41ed

• fix pdc endpoint more than node num

  

  humingqing

  authored and

  humingqing

  committed
  97f271f

  Copy full SHA for 97f271f

Commits on Sep 27, 2023

• fix gc bug, add reformat create scope

  

  humingqing

  authored and

  humingqing

  committed
  98285d0

  Copy full SHA for 98285d0

• fix gc

  

  humingqing

  authored and

  humingqing

  committed
  d66472f

  Copy full SHA for d66472f

Commits on Sep 28, 2023

• fix rank_attention_op rank_offset gc bug, rollback paddle gc helper

  

  humingqing

  authored and

  humingqing

  committed
  a4ab2f7

  Copy full SHA for a4ab2f7

• format

  

  humingqing

  authored and

  humingqing

  committed
  86b0169

  Copy full SHA for 86b0169

Commits on Oct 8, 2023

• fix hbm memory explosion in saved model

  

  humingqing

  authored and

  humingqing

  committed
  c8f12a0

  Copy full SHA for c8f12a0

Commits on Oct 16, 2023

• add print sharding size and rebuild save op

  

  humingqing

  authored and

  humingqing

  committed
  1e87518

  Copy full SHA for 1e87518

• remove no used code

  

  humingqing

  authored and

  humingqing

  committed
  87f6af8

  Copy full SHA for 87f6af8

Commits on Oct 18, 2023

• fix infer matmul split diff

  

  humingqing

  authored and

  humingqing

  committed
  afaafb6

  Copy full SHA for afaafb6

• fix matmul infer

  

  humingqing

  authored and

  humingqing

  committed
  8654e0a

  Copy full SHA for 8654e0a

• fix matmul infer

  

  humingqing

  authored and

  humingqing

  committed
  7a59736

  Copy full SHA for 7a59736

Commits on Oct 20, 2023

• fix cuda11.8 complier

  

  humingqing

  authored and

  humingqing

  committed
  e2f55b5

  Copy full SHA for e2f55b5

Commits on Oct 23, 2023

• remove no used code

  

  humingqing

  authored and

  humingqing

  committed
  d156685

  Copy full SHA for d156685

• rollback matmul

  

  humingqing

  authored and

  humingqing

  committed
  077af71

  Copy full SHA for 077af71

• add infer enable_cublas_tf32_op_math gflags

  

  humingqing

  authored and

  humingqing

  committed
  fa6efd4

  Copy full SHA for fa6efd4

Commits on Oct 31, 2023

• add tensor copy_from interface support sharding, delete some gpu context code

  

  humingqing

  authored and

  humingqing

  committed
  363b555

  Copy full SHA for 363b555

• expand slot's feasign for cvr model in fused_seqpool_cvm_with_conv_op (#94)

  Show description for 76eecf9
  

  qingshui

  and

  dingyuandong

  authored
  76eecf9

  Copy full SHA for 76eecf9

• sharding mode, erase other device param (#93)

  gumplh

  authored
  1d44c42

  Copy full SHA for 1d44c42

Commits on Nov 6, 2023

• sharding mode support avg_weight (#95)

  Show description for 0bd927a
  gumplh

  authored
  0bd927a

  Copy full SHA for 0bd927a

• expand slot's feasign for cvr model in fused_seqpool_cvm_with_conv_op (#96)

  Show description for b72d1bd
  

  qingshui

  and

  dingyuandong

  authored
  b72d1bd

  Copy full SHA for b72d1bd

• merge from master

  

  humingqing

  authored and

  humingqing

  committed
  01c48f6

  Copy full SHA for 01c48f6

Commits on Dec 1, 2023

• merge

  

  humingqing

  authored and

  humingqing

  committed
  d32628e

  Copy full SHA for d32628e