Code

Add ONNX exporter#26

Code

Draft

santhnm2 wants to merge 63 commits intomainmicrosoft/dist-ir:mainfrom export_to_onnxmicrosoft/dist-ir:export_to_onnxCopy head branch name to clipboard

+1,271-108

Commits

Commits on Apr 19, 2021

• Projector: DistIR function -> per-rank function

  siddharth-krishna

  committed
  8ac2c50

  Copy full SHA for 8ac2c50

• Fix bug: constructing variadic ops when specifying output_values

  siddharth-krishna

  committed
  193dc84

  Copy full SHA for 193dc84

• Fix sequential executor docstring

  siddharth-krishna

  committed
  c5bfbd9

  Copy full SHA for c5bfbd9

Commits on Apr 25, 2021

• A distributed PyTorch backend

  siddharth-krishna

  committed
  d2f9e0a

  Copy full SHA for d2f9e0a

• Upgrade PyTorch version

  siddharth-krishna

  committed
  d9cc787

  Copy full SHA for d9cc787

• Add test: one-weird-trick (matmul version)

  siddharth-krishna

  committed
  a2a4cc2

  Copy full SHA for a2a4cc2

• Refactor run_multiprocess

  siddharth-krishna

  committed
  7ab770d

  Copy full SHA for 7ab770d

• Make run_multiprocess take Functions not nn.Modules

  siddharth-krishna

  committed
  4e37ce3

  Copy full SHA for 4e37ce3

• End-of-file newlines

  siddharth-krishna

  committed
  9576811

  Copy full SHA for 9576811

Commits on Apr 27, 2021

• Black

  siddharth-krishna

  committed
  c4ca1a0

  Copy full SHA for c4ca1a0

• Parametrize pytest test_owt

  siddharth-krishna

  committed
  a095665

  Copy full SHA for a095665

• Revert "Fix bug: constructing variadic ops when specifying output_values"

  Show description for 3279ca1
  siddharth-krishna

  committed
  3279ca1

  Copy full SHA for 3279ca1

• Fix Op constructor: better handling of variadic ops/pre-created outputs

  siddharth-krishna

  committed
  c84c61d

  Copy full SHA for c84c61d

Commits on Apr 28, 2021

• Add support for Relu

  siddharth-krishna

  committed
  bb604f6

  Copy full SHA for bb604f6

Commits on Apr 29, 2021

• Add DP test

  siddharth-krishna

  committed
  8d8bbda

  Copy full SHA for 8d8bbda

Commits on Apr 30, 2021

• Backend: run and time on GPU

  siddharth-krishna

  committed
  42751da

  Copy full SHA for 42751da

• Refactor grid_search for interactive use

  siddharth-krishna

  committed
  f73d33b

  Copy full SHA for f73d33b

Commits on May 2, 2021

• Timing code for CPUs

  siddharth-krishna

  committed
  20bae75

  Copy full SHA for 20bae75

• Handle ops with multiple outputs

  siddharth-krishna

  committed
  8e8513d

  Copy full SHA for 8e8513d

• Add support for all MLP training ops

  siddharth-krishna

  committed
  6f3ae59

  Copy full SHA for 6f3ae59

• Type inference: fix function name bug

  siddharth-krishna

  committed
  49d385c

  Copy full SHA for 49d385c

• Prettyprint: support for printing FunctionMakers

  siddharth-krishna

  committed
  2cce7c3

  Copy full SHA for 2cce7c3

• Fix backend op implementations

  siddharth-krishna

  committed
  91f5ecf

  Copy full SHA for 91f5ecf

• Convert per-rank fns to Modules inside each thread

  siddharth-krishna

  committed
  172a5db

  Copy full SHA for 172a5db

• Per-rank projector: remove types

  siddharth-krishna

  committed
  2ed3cf7

  Copy full SHA for 2ed3cf7

• Add some more tests

  siddharth-krishna

  committed
  7eee9f9

  Copy full SHA for 7eee9f9

• Default number of repetitions = 1

  siddharth-krishna

  committed
  78da37d

  Copy full SHA for 78da37d

• DHP transform: return separate init_fn and transformed fn

  siddharth-krishna

  committed
  56e05dc

  Copy full SHA for 56e05dc

• Run pytest with root dir included in PYTHONPATH

  siddharth-krishna

  committed
  78bc1e9

  Copy full SHA for 78bc1e9

• Grid search: remove unused imports

  siddharth-krishna

  committed
  8a288b6

  Copy full SHA for 8a288b6

• Revert unintended changes

  siddharth-krishna

  committed
  4b2b301

  Copy full SHA for 4b2b301

• Move run_pytorch to backend.torch

  siddharth-krishna

  committed
  75ec41a

  Copy full SHA for 75ec41a

Commits on May 5, 2021

• Interpret Function instead of creating fx.Graph

  siddharth-krishna

  committed
  6dbf57c

  Copy full SHA for 6dbf57c

• Prettyprint attributes as Python-style kwargs

  siddharth-krishna

  committed
  4a77383

  Copy full SHA for 4a77383

• Use broadcast with pairwise groups for send/recv on GPUs

  siddharth-krishna

  committed
  e8735c3

  Copy full SHA for e8735c3

• Move new tensors to GPU in each op, outputs back to CPU

  siddharth-krishna

  committed
  fd2e7b1

  Copy full SHA for fd2e7b1

Commits on May 6, 2021

• Add a mock multiprocess backend for debugging

  siddharth-krishna

  committed
  31031cc

  Copy full SHA for 31031cc

• Use spawn start method for multiprocessing

  siddharth-krishna

  committed
  5d99a62

  Copy full SHA for 5d99a62

• Fix MLP DHP tests

  siddharth-krishna

  committed
  2c8852a

  Copy full SHA for 2c8852a

• Revert "Fix MLP DHP tests"

  Show description for 1d54fea
  siddharth-krishna

  committed
  1d54fea

  Copy full SHA for 1d54fea

• Fix MLP DHP tests for real

  siddharth-krishna

  committed
  18eaa08

  Copy full SHA for 18eaa08

Commits on May 7, 2021

• Add code to plot grid search results

  siddharth-krishna

  committed
  0ce2fdb

  Copy full SHA for 0ce2fdb

• Don't use globals while multiprocessing

  siddharth-krishna

  committed
  5588b87

  Copy full SHA for 5588b87

• Fix mock backend, use_gpu=False by default

  siddharth-krishna

  committed
  7d714d4

  Copy full SHA for 7d714d4

• Partial grid search on 4 devices

  siddharth-krishna

  committed
  8646842

  Copy full SHA for 8646842

Commits on May 10, 2021

• Support collectives between a subset of ranks

  siddharth-krishna

  committed
  1db0736

  Copy full SHA for 1db0736

Commits on May 11, 2021

• Bug fixes for distributed process groups (#24)

  Show description for a79ab67
  santhnm2

  authored
  a79ab67

  Copy full SHA for a79ab67

Commits on May 13, 2021

• Debugging MLP deadlock

  siddharth-krishna

  committed
  0c0f7f5

  Copy full SHA for 0c0f7f5

• Enable grid search test again

  siddharth-krishna

  committed
  a5c8b34

  Copy full SHA for a5c8b34

Commits on May 14, 2021

• Fix collective projector

  santhnm2

  committed
  33cefce

  Copy full SHA for 33cefce

Commits on May 20, 2021

• Some comments, debugging code, cuda sync earlier

  siddharth-krishna

  committed
  ca68ec4

  Copy full SHA for ca68ec4

• Projector for gather

  siddharth-krishna

  committed
  b31209d

  Copy full SHA for b31209d

Commits on May 25, 2021

• Don't save input/outputs to file in torch backend

  siddharth-krishna

  committed
  54ce0d2

  Copy full SHA for 54ce0d2

• Remove unnecessary global

  siddharth-krishna

  committed
  5d1b63f

  Copy full SHA for 5d1b63f

• Free tensors after use

  siddharth-krishna

  committed
  3466627

  Copy full SHA for 3466627

Commits on May 26, 2021

• Map DistIR devices to pytorch backend ranks

  siddharth-krishna

  committed
  7d16c90

  Copy full SHA for 7d16c90

• Fix tests

  siddharth-krishna

  committed
  8c4f5e1

  Copy full SHA for 8c4f5e1

• Some documentation and cleanup

  siddharth-krishna

  committed
  79d4a65

  Copy full SHA for 79d4a65

• Fix comment

  siddharth-krishna

  committed
  da7ff5d

  Copy full SHA for da7ff5d

• Remove experiment code and dead code

  siddharth-krishna

  committed
  72589b2

  Copy full SHA for 72589b2

Commits on Jun 1, 2021

• Add ONNX exporter

  santhnm2

  committed
  849ab4e

  Copy full SHA for 849ab4e

• Remove singleton types

  santhnm2

  committed
  054ce0a

  Copy full SHA for 054ce0a

Commits on Jun 2, 2021

• Add test to verify that out-of-opset ops throw error

  santhnm2

  committed
  fab0020

  Copy full SHA for fab0020