Commits

Branch selector

 master

User selector

All users

All time

Commit History

Commits on May 13, 2023

• Update distributed optimizer with new coalescing manager API in PyTorch (#1663)

  Show description for 8b7a1ff
  

  timmoon10

  and

  crcrpar

  authored
  8b7a1ff

  Copy full SHA for 8b7a1ff

Commits on May 12, 2023

• allow for custom directory of xml reports (#1661)

  Show description for 2d876b8
  crcrpar

  authored
  2d876b8

  Copy full SHA for 2d876b8

Commits on Apr 27, 2023

• Using regular layernorm for ONNX and TS export (#1634)

  Show description for 85e9edd
  

  borisfom

  and

  crcrpar

  authored
  85e9edd

  Copy full SHA for 85e9edd

• [Transformer] Update p2p communication routine (#1650)

  Aidyn-A

  authored
  f2d6f29

  Copy full SHA for f2d6f29

Commits on Apr 26, 2023

• [Transformer][test/L0] Increase max world_size for BERT and GPT (#1651)

  Aidyn-A

  authored
  5b5d410

  Copy full SHA for 5b5d410

Commits on Apr 21, 2023

• Mention `Ninja` for faster build (#1607)

  Show description for b3e3bab
  crcrpar

  authored
  b3e3bab

  Copy full SHA for b3e3bab

Commits on Apr 19, 2023

• [Transformer] Manipulate NCCL_NET for hybrid IB/Socket setups (#1620)

  Show description for 817e818
  Aidyn-A

  authored
  817e818

  Copy full SHA for 817e818

Commits on Apr 18, 2023

• use nvcr 23.03 (#1644)

  ZhiyuanChen

  authored
  eae1ddf

  Copy full SHA for eae1ddf

• add bf16 support for FusedDense (#1627)

  Agoniii

  authored
  1ba89ee

  Copy full SHA for 1ba89ee

• include stdint.h & stdio.h in fast_layer_norm/ln.h (#1637)

  Show description for 4e1ae43
  ZhiyuanChen

  authored
  4e1ae43

  Copy full SHA for 4e1ae43

Commits on Apr 14, 2023

• Add support for FP32 master weights in FusedAdam optimizer (#1625)

  Show description for 6bd01c4
  

  4 peopleauthored
  6bd01c4

  Copy full SHA for 6bd01c4

Commits on Mar 31, 2023

• [apex.contrib.sparsity] Grouped conv permutations (#1628)

  Show description for 89cc215
  jpool-nv

  authored
  89cc215

  Copy full SHA for 89cc215

Commits on Mar 29, 2023

• create mpi proc group: (#1626)

  erhoo82

  authored
  57057e2

  Copy full SHA for 57057e2

Commits on Mar 23, 2023

• Overlap first param sync with optimizer step in distributed Adam (#1621)

  Show description for 7150e20
  timmoon10

  authored
  7150e20

  Copy full SHA for 7150e20

• check for empty params in group (#1596)

  Show description for 0420735
  

  3 peopleauthored
  0420735

  Copy full SHA for 0420735

Commits on Mar 22, 2023

• Revert "Add support for FP32 master weights in FusedAdam optimizer (#1623)" (#1624)

  Show description for 4060aef
  crcrpar

  authored
  4060aef

  Copy full SHA for 4060aef

• Add support for FP32 master weights in FusedAdam optimizer (#1623)

  Show description for 6952004
  

  3 peopleauthored
  6952004

  Copy full SHA for 6952004

Commits on Mar 20, 2023

• p2p communication overlap at interleaved pipelining (#1616)

  Show description for d8643ef
  erhoo82

  authored
  d8643ef

  Copy full SHA for d8643ef

Commits on Mar 17, 2023

• Align distopt communication in interleaved pipeline parallelism (#1611)

  Show description for 27de66c
  

  timmoon10

  and

  crcrpar

  authored
  27de66c

  Copy full SHA for 27de66c

• [fmha] Use 8bit comparison to generate dropout mask (#1614)

  vasunvidia

  authored
  07b2935

  Copy full SHA for 07b2935

Commits on Mar 13, 2023

• ConvFrozenScaleBiasReLU fusion (#1608)

  Show description for 623315a
  minitu

  authored
  623315a

  Copy full SHA for 623315a

Commits on Mar 7, 2023

• [FMHA] Remove zero fill for softmax output (#1565)

  Show description for b1c7600
  vasunvidia

  authored
  b1c7600

  Copy full SHA for b1c7600

• s/maximimize/maximize/ (#1604)

  Show description for be819ea
  EIFY

  authored
  be819ea

  Copy full SHA for be819ea

Commits on Mar 6, 2023

• gracefully skipping permutations when torch.fx fails tracing (#1595)

  jpool-nv

  authored
  e29fb45

  Copy full SHA for e29fb45

Commits on Feb 28, 2023

• Update layers.py to ensure grad_output is contiguous (#1601)

  Show description for 95cea8e
  markelsanz14

  authored
  95cea8e

  Copy full SHA for 95cea8e

Commits on Feb 25, 2023

• use a single communicator for amax reduction in both TP and DP domains (#1597)

  erhoo82

  authored
  eec7250

  Copy full SHA for eec7250

• Decouple optimizer state and grad dtypes in distributed Adam optimizer (#1575)

  Show description for 03c9d80
  timmoon10

  authored
  03c9d80

  Copy full SHA for 03c9d80

Commits on Feb 16, 2023

• Use a separate communicator for AMAX reduction (#1585)

  Show description for 0c8400a
  erhoo82

  authored
  0c8400a

  Copy full SHA for 0c8400a

• use `torch.tensor` to create a tensor with initializer values (#1588)

  Show description for 6943fd2
  crcrpar

  authored
  6943fd2

  Copy full SHA for 6943fd2

Commits on Feb 15, 2023

• torch.empty() must have args (#1584)

  NouamaneTazi

  authored
  93bb36a

  Copy full SHA for 93bb36a

Commits on Feb 13, 2023

• replace torch.Tensor with torch.empty (#1578)

  Show description for ba027dd
  NouamaneTazi

  authored
  ba027dd

  Copy full SHA for ba027dd

Commits on Feb 9, 2023

• decrease warnings & remove append_nvcc_threads from setup.py (#1581)

  Show description for 0145d69
  crcrpar

  authored
  0145d69

  Copy full SHA for 0145d69

Commits on Feb 7, 2023

• return none-list input as as: pass the data loader iterator to training function (#1580)

  Show description for a32d7a6
  erhoo82

  authored
  a32d7a6

  Copy full SHA for a32d7a6

• Distributed optimizer overlaps forward compute and param all-gather (#1559)

  Show description for c0a0b0f
  timmoon10

  authored
  c0a0b0f

  Copy full SHA for c0a0b0f

Commits on Feb 3, 2023

• [fused softmax] integer overflow and non-contiguous memory (#1568)

  Show description for 5c9625c
  cyanguwa

  authored
  5c9625c

  Copy full SHA for 5c9625c