llvm
/
torch-mlir
mirror of https://github.com/llvm/torch-mlir
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity
torch-mlir/build_tools/torchscript_e2e_heavydep_tests
History
Prashant Kumar 10c8e3c593 Add simple neural_net and bert_training scripts. 
1. With the help of `make_fx` we are able to get the full training graph
   with weight updates.
2. NeuralNet_training passes. Bert_training passes after cherry-picking
   https://github.com/llvm/torch-mlir/pull/844.
3. TODO: Remove the functorch's dependency after make_fx moves to
   pytorch core.
 		2022-05-19 06:18:42 +05:30
..
README.md Remove the last mentions of npcomp from torch-mlir 2021-10-05 20:17:23 +00:00
generate_serialized_tests.sh Add simple neural_net and bert_training scripts. 2022-05-19 06:18:42 +05:30
hf_sequence_classification.py Remove heavy_deps models that don't get serialized. 2022-04-29 17:21:25 +05:30
main.py Add simple neural_net and bert_training scripts. 2022-05-19 06:18:42 +05:30
train_models.py Add simple neural_net and bert_training scripts. 2022-05-19 06:18:42 +05:30
vision_models.py Remove heavy_deps models that don't get serialized. 2022-04-29 17:21:25 +05:30

README.md

Additional end-to-end tests with heavy dependencies (heavydep tests)

Some end-to-end tests require additional dependencies which don't make sense to include as part of the default torch-mlir setup. Additionally, these dependencies often don't work with the same HEAD PyTorch version that torch-mlir builds against at the C++ level.

We have a self-contained script that generates all the needed artifacts from a self-contained virtual environment. It can be used like so:

# Build the virtual environment in the specified directory and generate the
# serialized test artifacts in the other specified directory.
# This command is safe to re-run if you have already built the virtual
# environment and just changed the tests.
build_tools/torchscript_e2e_heavydep_tests/generate_serialized_tests.sh \
  path/to/heavydep_venv \
  path/to/heavydep_serialized_tests

# Add the --serialized-test-dir flag to point at the directory containing the
# serialized tests. All other functionality is the same as the normal invocation
# of torchscript_e2e_test.sh, but the serialized tests will be available.
tools/torchscript_e2e_test.sh --serialized-test-dir=path/to/heavydep_serialized_tests

The tests use the same (pure-Python) test framework as the normal torchscript_e2e_test.sh, but the tests are added in build_tools/torchscript_e2e_heavydep_tests instead of frontends/pytorch/e2e_testing/torchscript.

We rely critically on serialized TorchScript compatibility across PyTorch versions to transport the tests + pure-Python compatibility of the torch API, which has worked well so far.