Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Track Modifications of Tensorflow #59

Open
merrymercy opened this issue Nov 24, 2021 · 0 comments
Open

Track Modifications of Tensorflow #59

merrymercy opened this issue Nov 24, 2021 · 0 comments

Comments

@merrymercy
Copy link
Member

merrymercy commented Nov 24, 2021

Incompatible changes

  • tensorflow/compiler/xla/python/xla_compiler.cc. The return value of spmd_output_sharding and spmd_parameters_shardings is not converted to tuple.
  • tensorflow/compiler/xla/service/hlo_verifier.cc. Temporarily disable the verification on fused instruction due to our CommonComputationElimination pass.
  • tensorflow/compiler/xla/service/reduce_scatter_utils.cc. The check of ar->GetModule()->config().use_spmd_partitioning() is removed.
  • tensorflow/compiler/xla/service/gpu/gpu_compiler.cc. Disable DotMerger because it breaks some auto-sharding tests with small dots.
  • tensorflow/compiler/xla/service/gpu/ir_emitter_unnested.cc: Use EmitRngGetAndUpdateStateThunk to generate custom rng thunks.
  • tensorflow/compiler/xla/service/gpu/nccl_all_to_all_thunk.cc. Temporarily disable the CanImplement check on all-to-all. Some valid all-to-alls are classified as invalid because the size-1 dimension are permuted during HLO->MLIR conversion.
  • tensorflow/compiler/xla/service/spmd/dot_handler.cc. Temporarily disable windowed dot.
  • tensorflow/compiler/xla/service/spmd/spmd_partitioner.cc::HandleRng. Use a different strategy for replicated rng instructions. We favor replicated computation over communication.
  • tensorflow/compiler/xla/service/spmd/spmd_partitioner.cc::HandleAllReduce. Use handle elementwise for profiling.
  • tensorflow/compiler/xla/service/optimization_barrier_expander.cc. OptimizationBarrier is converted to bitcast.
  • The commit 96b8219 about AllGather is reverted due to performance regression.

Improvements & Bug Fixes

  • tensorflow/core/platform/env.cc. Add random number to file names to avoid duplication.
@merrymercy merrymercy changed the title All incompatible changes we made Track Modifications of Tensorflow Apr 8, 2022
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

1 participant