Commits

Branch selector

 master

User selector

All users

Datepicker

All time

Commit History

Commits on Mar 1, 2022

• [transformer] Update build_model function to support encoder&decoder model (NVIDIA#1307)

  Show description for 59978d5
  crcrpar

  authored
  59978d5

  Copy full SHA for 59978d5

Commits on Feb 27, 2022

• build fused grad accum w/ wgrad only if cuda>10 (NVIDIA#1312)

  crcrpar

  authored
  47c269b

  Copy full SHA for 47c269b

Commits on Feb 26, 2022

• [transformer] Fuse grad accumulation with wgrad (NVIDIA#1297)

  Show description for ddc0803
  

  crcrpar

  and

  erhoo82

  authored
  ddc0803

  Copy full SHA for ddc0803

Commits on Feb 25, 2022

• add setter of pipeline model parallel split rank (NVIDIA#1306)

  crcrpar

  authored
  e95c3b9

  Copy full SHA for e95c3b9

• [transformer] use logger in microbatches module (NVIDIA#1302)

  crcrpar

  authored
  17e1a1f

  Copy full SHA for 17e1a1f

• skip FastLayerNorm (NVIDIA#1305)

  crcrpar

  authored
  4506a68

  Copy full SHA for 4506a68

Commits on Feb 23, 2022

• be more flexible (NVIDIA#1299)

  crcrpar

  authored
  199fa83

  Copy full SHA for 199fa83

• access to pipeline_model_parallel_split_rank (NVIDIA#1300)

  crcrpar

  authored
  069ff33

  Copy full SHA for 069ff33

• Merge pull request NVIDIA#1301 from NVIDIA/bug_fix_in_fast_bottleneck

  Show description for ab1a93a
  thorjohnsen

  authored
  ab1a93a

  Copy full SHA for ab1a93a

• Change data type for virtual tensors to float

  thorjohnsen

  committed
  51e8131

  Copy full SHA for 51e8131

Commits on Feb 15, 2022

• taking channels last 3d into account (NVIDIA#1284)

  crcrpar

  authored
  39fc7cc

  Copy full SHA for 39fc7cc

Commits on Feb 12, 2022

• cast for -Wc++11-narrowing (NVIDIA#1288)

  crcrpar

  authored
  1e21874

  Copy full SHA for 1e21874

Commits on Feb 11, 2022

• [FusedRMSNorm doc] document where epsilon is added (NVIDIA#1295)

  Show description for c8c00ef
  stas00

  authored
  c8c00ef

  Copy full SHA for c8c00ef

Commits on Feb 10, 2022

• 8.6 requires CUDA 11.1 (NVIDIA#1289)

  crcrpar

  authored
  e1aa1fc

  Copy full SHA for e1aa1fc

Commits on Feb 7, 2022

• fix and generate docs for FusedRMSNorm (NVIDIA#1285)

  eqy

  authored
  a786ca0

  Copy full SHA for a786ca0

Commits on Feb 4, 2022

• FusedRMSNorm/"T5LayerNorm" based on FusedLayerNorm (NVIDIA#1274)

  Show description for 684c473
  

  eqy

  and

  crcrpar

  authored
  684c473

  Copy full SHA for 684c473

Commits on Feb 1, 2022

• Add the permutation related support as the extension for asp lib. (NVIDIA#1194)

  Show description for 89edb81
  ChongyuNVIDIA

  authored
  89edb81

  Copy full SHA for 89edb81

• transformer: Allows for custom sync context in no pipelining forward backward function (NVIDIA#1281)

  Show description for 79c0187
  crcrpar

  authored
  79c0187

  Copy full SHA for 79c0187

Commits on Jan 31, 2022

• T5 pipeline parallel changes (NVIDIA#1279)

  Show description for 0da60e1
  crcrpar

  authored
  0da60e1

  Copy full SHA for 0da60e1

• fix group range to compute l2_norm (NVIDIA#1266)

  Show description for a47d1a7
  

  chochowski

  and

  seryilmaz

  authored
  a47d1a7

  Copy full SHA for a47d1a7

Commits on Jan 29, 2022

• add inline asm 128-bit counter (NVIDIA#1265)

  seryilmaz

  authored
  2eafdb3

  Copy full SHA for 2eafdb3

Commits on Jan 28, 2022

• small changes in test and logger format (NVIDIA#1278)

  Show description for b1c75f6
  crcrpar

  authored
  b1c75f6

  Copy full SHA for b1c75f6

• allow for None batch (NVIDIA#1280)

  Show description for a960fe8
  crcrpar

  authored
  a960fe8

  Copy full SHA for a960fe8

Commits on Jan 21, 2022

• Grad scaler (NVIDIA#1277)

  Show description for 2a4ab17
  crcrpar

  authored
  2a4ab17

  Copy full SHA for 2a4ab17

• Fix missing dtype in recv_forward (NVIDIA#1276)

  Show description for 45cd100
  eqy

  authored
  45cd100

  Copy full SHA for 45cd100

Commits on Jan 19, 2022

• pass flags to transducer joint kernel (NVIDIA#1273)

  crcrpar

  authored
  c4e85f7

  Copy full SHA for c4e85f7

Commits on Jan 13, 2022

• support new path to CUDAGeneratorImpl.h (NVIDIA#1267)

  shintaro-iwasaki

  authored
  b2fdf9c

  Copy full SHA for b2fdf9c

Commits on Dec 17, 2021

• Add an argument of dtype to forward_backward functions to specify the dtype used in p2p comm (NVIDIA#1249)

  Show description for b88c507
  crcrpar

  authored
  b88c507

  Copy full SHA for b88c507

Commits on Dec 16, 2021

• version guard (NVIDIA#1253)

  crcrpar

  authored
  e847382

  Copy full SHA for e847382

• Reduce OOM potential and report it if it happens in BERT test (NVIDIA#1250)

  Show description for e0f5ea8
  

  eqy

  and

  crcrpar

  authored
  e0f5ea8

  Copy full SHA for e0f5ea8

Commits on Dec 15, 2021

• Add --threads 4 to extra_compile_args["nvcc"] (NVIDIA#1251)

  Show description for f63dac8
  crcrpar

  authored
  f63dac8

  Copy full SHA for f63dac8

• check model_parallel is initialized in build_model (NVIDIA#1248)

  crcrpar

  authored
  1cd1181

  Copy full SHA for 1cd1181

Commits on Dec 14, 2021

• Faster --fast_multihead_attn build (NVIDIA#1245)

  Show description for 7ec8ed6
  crcrpar

  authored
  7ec8ed6

  Copy full SHA for 7ec8ed6

• check size in kth microbatch (NVIDIA#1247)

  eqy

  authored
  ed94d0b

  Copy full SHA for ed94d0b

Commits on Dec 10, 2021

• Cherry-pick Megatron-LM's changes in pipeline model parallel for T5 (NVIDIA#1232)

  Show description for 0e25fcc
  crcrpar

  authored
  0e25fcc

  Copy full SHA for 0e25fcc

Pagination

Previous

Next