GitLab

* Add FP8 support for Ada
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Fixes
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* better message
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* lint fixes
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Address review comments
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* better message for no fp8
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* same thing for onnx test
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* fix
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Fix CI and review
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

---------
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Bugfix - compute scale_inv when loading checkpoint
Signed-off-by: Gal Hubara Agam <ghubaraagam@nvidia.com>

* Save inverse scale in extra state tensor + minor CR fixes
Signed-off-by: Gal Hubara Agam <ghubaraagam@nvidia.com>

* Fix lint
Co-authored-by: Gal Hubara Agam <ghubaraagam@nvidia.com>
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

---------
Signed-off-by: Gal Hubara Agam <ghubaraagam@nvidia.com>
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Change FP8 recipe defaults
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Increase default amax history length
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Always check history size
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* no amax history for onnx export
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* revert onnx export test changes
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Fix indices in onnx test
Co-authored-by: Neta Zmora <nzmora@nvidia.com>
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

---------
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Co-authored-by: Neta Zmora <nzmora@nvidia.com>

Signed-off-by: Tal Cherckez <tcherckez@nvidia.com>
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Fix GELU ONNX export

* Wrap GELU export with cast to/from FP32 to achieve same compute precision as TE.
* Increase GELU export test thresholds.
* Change export to ONNX opset 17 for smaller representation of LN (single node instead of subgraph).
* Remove the need for LN work-around for ORT
Signed-off-by: Neta Zmora <nzmora@nvidia.com>

* Add docstring to te_onnx_extensions.py::compute_in_fp32
Signed-off-by: Neta Zmora <nzmora@nvidia.com>

* Tune threshold for GELU ONNX export

Ran 8K test instances to verify the threshold.
Allow 2 coefficients to escape threshold. Two wrong coefficients
are not a failure.
Signed-off-by: Neta Zmora <nzmora@nvidia.com>

---------
Signed-off-by: Neta Zmora <nzmora@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

Add an option to serialize test i/o to file

Small refactoring of the inferencing code.
Change the default directory where generated ONNX files are stored.
Use the temp directory to avoid clogging the file system.
Add an option to serialize test input/output tensors to a
Polygraphy RunResults object.
Signed-off-by: Neta Zmora <nzmora@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* add layernorm1p fp8 test
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* combine tests for easy maintenance
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* using torch.autocast for AMP and check grad types
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Add test for wgrad accumulation fusion
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* rename file
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Setup numerical tests + SAR
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Add test for full activation recompute
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Add tests for checkpoint load/store
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* TE vs framework numerical tests
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* fix ci
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* relax thresholds
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

---------
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Add a temporary workaround to layernorm export

Seems like ORT is performing template-matching for LN and incorrectly concludes
that it doesn't have a kernel for FP32 LN. The work-around adds the addition of
fake_zero which is meant to prevent the template matching while keeping the graph
virtually unchanged. This also requires `do_constant_folding=False` in
`torch.onnx.export`.
Signed-off-by: Neta Zmora <nzmora@nvidia.com>

* Adjust test threshold
Signed-off-by: Neta Zmora <nzmora@nvidia.com>

* Opened an ORT bug and added the link for tracking
Signed-off-by: Neta Zmora <nzmora@nvidia.com>

* Fix Python linter errors
Signed-off-by: Neta Zmora <nzmora@nvidia.com>

* Simplify the LN workaround solution (ONNX export)

After discussing https://github.com/microsoft/onnxruntime/issues/15021


with Microsoft engineers, replaced the LN workaround with a simpler
implementation.

In addition:
* To make test more robust add `allow_cnt_errors` to `validate_result`
* Add more documentation to clarify the purpose and methodology of the
ONNX export tests
Signed-off-by: Neta Zmora <nzmora@nvidia.com>

* Fix unused import
Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

* Fix unused import
Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

* Fix unused import
Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

---------
Signed-off-by: Neta Zmora <nzmora@nvidia.com>
Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

* deprecate qk layer scaling and fp32 softmax args
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* apply QK layer scaling for fp16 training
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* address review comments
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

---------
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* ignore self attention mask for causal type
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* further relax checks to run FA, update docs
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* fix pytorch softmax path
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* fixes
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* minimum ampere requirement for fa
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

---------
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* move TE/PyTorch UT to tests/pytorch

1. move tests/* files to tests/pytorch/
2. adjust UT paths in qa/L0_unittest/test.sh
Signed-off-by: Ryan Jeng <rjeng@nvidia.com>

* update build.yml
Signed-off-by: Ryan Jeng <rjeng@nvidia.com>

---------
Signed-off-by: Ryan Jeng <rjeng@nvidia.com>