Yateng Hong
0978d2bfbe
Fix CUDA test hang: ( #6138 )
...
- Make condition check in `CUDAAllocatorTest` to ensure CUDA device is present.
2020-12-16 16:32:56 +10:00
Guoyu Wang
b648bf641f
nnapi add min max support ( #6117 )
2020-12-15 22:31:28 -08:00
Vincent Wang
b8c8fe91f5
ort's to_dlpack.
2020-12-15 21:11:14 -08:00
George Nash
939cc9b410
Enable running the mnist_training sample without cuda ( #6085 )
...
Signed-off-by: George Nash <george.nash@intel.com>
2020-12-15 17:06:54 -08:00
Ryan Hill
ac62cf8058
Unify IExecutionProvider and IExecutionProviderFactory interfaces ( #6108 )
...
* Remove Provider_IExecutionProvider and make the internal IExecutionProvider usable by shared providers
* Change Provider_IExecutionProviderFactory to be the core version.
2020-12-15 16:45:53 -08:00
Cecilia Liu
980a93c164
Model Fusion For Bart ( #6105 )
...
Fusion fix for Bart models
2020-12-15 14:30:15 -08:00
George Wu
297c824807
remove dnnl_dll_path from post build copy ( #6142 )
2020-12-15 13:47:39 -08:00
Edward Chen
64709b1335
Deprecate Python global configuration functions [Part 1] ( #5923 )
...
Enable options to be set via execution provider (EP)-specific options and log deprecation warning from current global configuration functions.
2020-12-15 11:32:43 -08:00
Vincent Wang
c4f827bee1
remove initializers from original graph
2020-12-15 09:03:08 -08:00
Thiago Crepaldi
f7f435fc27
Improve dynamic axes to work without data descriptors
2020-12-15 09:03:08 -08:00
Vincent Wang
7729bb3c8d
Add initial dynamic axes support
2020-12-15 09:03:08 -08:00
Thiago Crepaldi
004632ff8d
TEMP: Add support to measure method execution time for perf improvement
2020-12-15 09:03:08 -08:00
Thiago Crepaldi
e5fdb455ed
Improve performance by running ApplyTransformers on gradient graph
2020-12-15 09:03:08 -08:00
Thiago Crepaldi
e986ae5f86
Remove dead code
2020-12-15 09:03:08 -08:00
Thiago Crepaldi
41b88ce91d
Remove initializers from forward ONNX graph
2020-12-15 09:03:08 -08:00
Thiago Crepaldi
07f5ae95e5
Refactor MNIST and BERT classifier to add time measures
2020-12-15 09:03:08 -08:00
Thiago Crepaldi
395e082bc3
Remove (unnecessary) gradient graph from frontend
2020-12-15 09:03:08 -08:00
Thiago Crepaldi
4d9267e102
Add IO binding support, which allows CUDA training
2020-12-15 09:03:08 -08:00
Thiago Crepaldi
f13c2a61d5
Refactor IObinding
2020-12-15 09:03:08 -08:00
Vincent Wang
39ac95b2fc
add io binding
2020-12-15 09:03:08 -08:00
Thiago Crepaldi
ff79e8743f
Add support to BERT fine tuning (MVP 3)
...
Additional changes include major refactoring to use new backend API
2020-12-15 09:03:08 -08:00
Thiago Crepaldi
78831d009b
Add list of initializer gradients to the backend training graph spliter
2020-12-15 09:03:08 -08:00
Vincent Wang
60b6e2683f
bugfix
2020-12-15 09:03:08 -08:00
Vincent Wang
f6a8d2aa5f
split graphs info
2020-12-15 09:03:08 -08:00
Vincent Wang
cfd57c0136
fix input order, and input grad.
2020-12-15 09:03:08 -08:00
Vincent Wang
e759da178d
bugfix for graph inputs and outputs.
2020-12-15 09:03:08 -08:00
Thiago Crepaldi
b7564d0732
Refactor after Vincent work on splitting on backend
2020-12-15 09:03:08 -08:00
Vincent Wang
6d8fde8324
sample code change.
2020-12-15 09:03:08 -08:00
Vincent Wang
934feb0c99
gradient graph split in backend.
2020-12-15 09:03:08 -08:00
Thiago Crepaldi
ea5871ac15
Change DropouGrad.input[1].input_type and del logits_grad from backward graph
2020-12-15 09:03:08 -08:00
Thiago Crepaldi
f1dc6e4007
Refactor BERT classifier fine tune for better debugging
2020-12-15 09:03:08 -08:00
Thiago Crepaldi
d4917f2d65
Hard-code input types for DropoutGrad on BERT
2020-12-15 09:03:08 -08:00
Thiago Crepaldi
3b267d1d60
Add BERT classifier example
2020-12-15 09:03:08 -08:00
Thiago Crepaldi
30042b6e0e
Update InferenceSession usage to match master
2020-12-15 09:03:08 -08:00
Thiago Crepaldi
8b0ade0e83
Integrate automatic graph split into ORTModule
2020-12-15 09:03:08 -08:00
Vincent Wang
c36c8e14a7
refactor
2020-12-15 09:03:08 -08:00
Vincent Wang
26e6d6d004
module transformer
2020-12-15 09:03:08 -08:00
Thiago Crepaldi
3524fb04e8
Add working example for MNIST (MVP)
2020-12-15 09:03:08 -08:00
Thiago Crepaldi
f1b5c25b2d
Improve example to display grads before and after optim step
2020-12-15 09:03:07 -08:00
Thiago Crepaldi
f06cafdebd
Fix path on test script
2020-12-15 09:03:07 -08:00
Thiago Crepaldi
56ca4ab05b
Add flag to allow pytorch-only or ORT flexible api runs
2020-12-15 09:03:07 -08:00
Thiago Crepaldi
d4449d86b9
Add script to run Flexible API MVP PoC
2020-12-15 09:03:07 -08:00
Thiago Crepaldi
e71e08851a
Basic plumbing for backward pass. Not fully working
2020-12-15 09:03:07 -08:00
Thiago Crepaldi
77cefcd6c2
Perform forward pass using training graph with intermediate outputs
2020-12-15 09:03:07 -08:00
Thiago Crepaldi
11b69f141e
Forward pass using InferenceSession on exported ONNX
...
Although forward pass works, this has the limitation of not working for
backward pass due to the lack of intermediate tensors needed for
gradient.
Next step is to export a training graph and split it manually
2020-12-15 09:03:07 -08:00
Jesse Benson
a8d549e181
Minor changes to AMD element-wise kernels to converge with CUDA element-wise kernels.
2020-12-15 08:46:36 -08:00
Pranav Sharma
a9548283d0
Don't mark issues that are marked as enhancement as stale ( #6134 )
2020-12-14 18:57:40 -08:00
Edward Chen
9810b9e02b
Reduce amount of compiled CUDA device code ( #6118 )
...
Move CudaKernel from cuda_common.h to a new separate header, cuda_kernel.h. Update include sites to use cuda_kernel.h instead if they need CudaKernel. Inclusions of cuda_common.h are now more lightweight.
Make corresponding changes for ROCM execution provider code.
Other minor cleanup.
2020-12-14 15:27:40 -08:00
Sheil Kumar
a6a23db130
Enable C# .NET5 for WinML ( #6120 )
...
* build for .net5
* only reference cswinrt for .net5
* remove netstandard2.0 references
* upgrade language version
* net5
* remove extra comment closure
* add targetframework
* set target framework
* remove net*
* pep8 errors
* make test project build with .net windows SDK projection
* disable c# builds for non-x64 builds
* fix pep8 errors
* disable for store build
* fix tests
* remove cswinrt and sdk references from package
* bump cswinrt down to 1.0.1
* fix bin path
Co-authored-by: Sheil Kumar <sheilk@microsoft.com>
2020-12-14 15:05:15 -08:00
Sherlock
eb5c1f0fcc
Unify activation and initializer alignment value ( #6109 )
...
* Unify activation and initializer alignment value
* Fix VerifyInputTensorsAllocatedContiguously
2020-12-14 13:13:41 -08:00
