pytorch
276deb6d - Update on "[DDP] Disable reducer hooks from running outside of DDP backwards."

Commit

3 years ago

Update on "[DDP] Disable reducer hooks from running outside of DDP backwards." Sometimes local modules can fire hooks (such as when user calls backward after using `ddp_module.module` explicitly). This isn't supported behavior and can cause issues with various state and gradient reduction we run in DDP, so it's best to disable this entirely. Differential Revision: [D29435737](https://our.internmc.facebook.com/intern/diff/D29435737/) **NOTE FOR REVIEWERS**: This PR has internal Facebook specific changes or comments, please review them on [Phabricator](https://our.internmc.facebook.com/intern/diff/D29435737/)! [ghstack-poisoned]

Author

rohan-varma

Parents

4e53e87e

db8d3110

Files29

.circleci/docker/common
- install_conda.sh
.github
- scale-config.yml
- workflows
  - lint.yml
.jenkins/pytorch
- macos-test.sh
- win-test-helpers
  - build_pytorch.bat
  - setup_pytorch_env.bat
caffe2
- perfkernels
  - lstm_unit_cpu-impl.h
- predictor
  - transforms.cc
- utils
  - conversions.h
  - cpuid.h
  - eigen_utils.h
  - string_utils.h
mypy.ini
test
- benchmark_utils
  - test_benchmark_utils.py
- quantization/fx
  - test_quantize_fx.py
- run_test.py
- test_expecttest.py
tools
- build_variables.bzl
torch
- autograd
  - function.py
- csrc
  - distributed/c10d
    - reducer.cpp
    - reducer.hpp
  - jit/runtime/static
    - ops.cpp
- fx
  - _symbolic_trace.py
- jit
  - _script.py
- quantization/fx
  - prepare.py
- testing/_internal
  - common_quantization.py
  - common_utils.py
  - distributed
    - distributed_test.py
  - expecttest.py

pytorch 276deb6d - Update on "[DDP] Disable reducer hooks from running outside of DDP backwards."

pytorch
276deb6d - Update on "[DDP] Disable reducer hooks from running outside of DDP backwards."