* Gelu contrib op & transformer
* Gelu kernels for CPU&cuda
* Merged PR 5034: fix a condition for gelu transformer
The ONNX models doesn't guarantee to assign an unique name to each node, so the previous condition could fail.
(cherry picked from commit e335ef5466444cb0aae45f885ea3a825ed9f1088)
* Fix builds
* remove useless comments
* fix test failure when nocontribp
* Move impelmentation under KMSdomain
* fix comments
* fix linux build
* Fix few comments
* fix linux build
* Update Cast op to use precision of 8 when casting floating point numbers to strings. This matches numpy precision.
Update unit tests to include non-trivial floats in the input.
Update onnx test infrastructure to document why the test cases are disabled
* py fallback initial commit.
* fixes.
* update NGRAPHCustomOp::Initialize() to return Status
* fixes in session.py
* FAIL status to EP_FAIL in ngraph custom op
* disable fallback for backend api
Remove gsl subodule and replace with a local copy of gsl-lite
Refactor for onnxruntime::make_unique
gsl::span size and index are now size_t
Remove lambda auto argument type detection.
Remove constexpr from fail_fast in gsl due to Linux not being happy.
Comment out std::stream support due to MacOS std lib broken.
Move make_unique into include/core/common so it is accessible for server builds.
Relax requirements for onnxruntime/test/providers/cpu/ml/write_scores_test.cc
due to x86 build.
Add ONNXRUNTIME_ROOT to Server Lib includes so gsl is recognized
* Update ops that had strides/dilations documentation updates to default to 1.
Code was already doing this.
Add tests to explicilty test.
* Update optimizers to add opset 11 support where possible
* Fixed a bug of missing tvm in python wheel
* Put Nuphar Python scripts into wheel
* Add note book tutorial
* Some improvements in symbolic shape inference for quantized models
* OpSet 11 Update for Neg Axis:
scan, flatten, compress, concat, gather, slice, split, squeeze, unsqueeze
* fix flatten op test
* Fix flatten and Squeeze
* fix test cases
* add gather neg indices to both cpu and cuda
* Exclude nGraph from neg axis test
* re-enable test cases
* Fix test cases
Fix a bug in Concat when only part of input has sympy_data
Fix a bug in ConstantOfShape when shape is scalar
Add support for GatherElements and ScatterElements
* Added support for Hetero plugin
Signed-off-by: suryasidd <surya.siddharth.pemmaraju@intel.com>
* Fixed spelling error in cmake for hetero plugin
Signed-off-by: suryasidd <surya.siddharth.pemmaraju@intel.com>
* Added listener to print messages from the plugin
Signed-off-by: suryasidd <surya.siddharth.pemmaraju@intel.com>
* Updated Documentation for VAD-F enablement
Signed-off-by: suryasidd <surya.siddharth.pemmaraju@intel.com>
* Added VAD-F option for FPGA
*Disabled unit tests and backed tests because FPGA only accepts NCHW models
Signed-off-by: suryasidd <surya.siddharth.pemmaraju@intel.com>
* Added comment for why tests need to be disabled on VAD-F
Signed-off-by: suryasidd <surya.siddharth.pemmaraju@intel.com>
* call MLAS's pooling function as an external call for Nuphar
Note that at the moment Nuphar provider doesn't handle the cases below:
- symbolic height/weight dimensions
- Indices output of MaxPool
- non-default dilations
* unify the pool interface for mti and mti_x86
* Changed ConvBase into a class member variable
Currently, all Conv-family classes inherit from both ConvBase
and OpKernel. Since what ConvBase provides is all about
processing convolution attributes, it's more natural to move it as
a class member variable.
This change renamed ConvBase to ConvAttributes and moved
it into a separate file conv_attributes by its own. Instead of
inheriting from ConvBase, now each Conv-related class has a
class member variable that is of type ConvAttributes.
Hence, we removed unecessary multiple inheritance and increase
composibility. More importantly, the change made it possible for
some other providers such as Nuphar be able to re-use the functionalities
provided by ConvAttributes class.
Note that we also made similar changes to ConvTransposeBase.
* fixed cuda build issue
* Add CumSum and Round for Opset 11
* add back 1 test
* Add back one broken test
* Add back more broken tests
* activate cumsum, round, dynamicquantizelinear tests
* removed python backend tests
* re-comment out dynamicquantizelinear_* tests. ReduceMin(11) not implemented yet
* re-comment out dynamicquantizelinear_* tests. ReduceMin(11) not implemented yet
* comment out cumsum_1d_reverse_exclusive
* Remove few types for csum. Keep only float, int32, int64
* Added friendly error message
* Added double type to pass ONNX tests.
* Don't return shape for non-const initializer in InferenceContextImpl::getInputType
Don't return initializer for non-const initializer in InferenceContextImpl::getInputData
Update graph_utils to support these scenarios
- fix GetConstantInitializer to make sure a name is for an outer scope value before checking a parent graph, as local name could shadow an outer scope initializer.
* call MLAS's pooling function as an external call for Nuphar
Note that at the moment Nuphar provider doesn't handle the cases below:
- symbolic height/weight dimensions
- Indices output of MaxPool
- non-default dilations
* unify the pool interface for mti and mti_x86
* Address two issues:
Thread-safety issue with LTSM/RNN running lambda in parallel
Propagate lambda exceptions and report them when running in
parallel.
Refactor cpu provider's pool ops by extracting pooling attributes
into a separate helper class PoolAttributes. With this change,
other providers such as Nuphar can re-use the same routines
for processing pooling attributes. This refactorying doesn't
have any functional changes.
