This PR also includes:
* Remove defaults from named tuples to support python 3.6
* Allows model which takes dicts as input
* Adapts BERT finetuning example to run on the new frontend
* Match numbers for BERT fine tuning model
Co-authored-by: liqun <liqun@OrtTrainingDev4.af05slrtruoetgaxwwjv5nsq5e.px.internal.cloudapp.net>
Co-authored-by: Thiago Crepaldi <thiago.crepaldi@microsoft.com>
* Port legacy checkpoint API into new front-end
This PR also fixes:
* Warnings on ORTTrainer for improper tensor copies
* Inaccurate LRScheduler tests using wrong LR
* Stale DeepSpeed documentation
* Minor code refactoring for Toy BERT tests
* Move experimental state_dict() and load_state_dict() into checkpoint ns
* adding generic configurations for session options
* fix a build break on linux
* fix training ci build break
* fix training ci build break
* addressed CR comments
* fix traning ci build break
* move config_key from enum to string
* add c# api
* add python api
* fix build break
* move prepacking from 2 new api entries to session options configs
* fix traning ci build break
* add python test, update some comments, move const key definition to avoid build break
* addressed comments
* move definitions of keys to common.h
* move api to version 5
* remove accidental change in build.py
* remove pragma to avoid build break
* addressed CR comments
* fix the python build break, and move location of config keys definition
* small typo changes
`LossScaler.update()` was not being properly called due to the incorrect TrainStepInfo.all_finite assignment.
Additionally to this fix, _ORTTrainerModelDesc.is_finite was renamed to _ORTTrainerModelDesc.all_finite to make it more uniform with TrainStepInfo
* Check status of BinaryElementwise::Prepare().
* Add additional status checks for BinaryElementwise::Prepare() and UnaryElementwise::Prepare().
* Add status checks for BinaryElementwisePreparation::BinaryElementwiseBroadcastPrepareHelper().
* Add ORTTrainerOptions class for the new pytorch frontend (#4382)
Add ORTTrainerOptions class and some placeholders
* Add _ORTTrainerModelDesc to perform validation for model description (#4416)
* Add Loss Scaler classes to the new frontend (#4306)
* Add TrainStepInfo used on the new frontend API (#4256)
* Add Optimizer classes to the new frontend (#4280)
* Add LRScheduler implementation (#4357)
* Add basic ORTTrainer API (#4435)
This PR presents the public API for ORTTrainer for the short term
development.
It also validates and saves input parameters, which will be used in the
next stages, such as building ONNX model, post processing the model and
configuring the training session
* Add opset_version into ORTTrainerOptions and change type of ORTTrainer.loss_fn (#4592)
* Update ModelDescription and minor fix on ORTTrainer ctor (#4605)
* Update ModelDescription and minor fix on ORTTrainer/ORTTrainerOptions
This PR keeps the public API intact, but changes how model description is stored on the backend
Currently, users creates a dict with two lists of tuples.
One list called 'inputs' and each tuple has the following format tuple(name, shape).
The second list is called 'outputs' and each tuple can be either tuple(name, shape) or tuple(name, shape, is_loss).
With this PR, when this dict is passed in to ORTTrainer, it is fully validated as usual.
However, tuples are internally replaced by namedtuples and all output tuples will have
tuple(name, shape, is_loss) format instead of is_loss being optionally present.
Additionally to that normalization in the internal representation (which eases coding),
two internal methods were created to replace a namedtuple(name, shape) to namedtuple(name, shape, dtype)
or namedtuple(name, shape, is_loss, dtype) dependeing whether the tuple is an input or output.
This is necessary as ORTTRainer finds out data types of each input/output during model export to onnx.
Finally, a minor fix was done on ORTTrainer. It could initialize ORTTrainerOptions incorrectly when options=None
* Rename input name for test
* Add ONNX Model Export to New Frontend (#4612)
Co-authored-by: Rayan Krishnan <t-rakr@OrtDevTest2v100.af05slrtruoetgaxwwjv5nsq5e.px.internal.cloudapp.net>
Co-authored-by: Thiago Crepaldi <thiago.crepaldi@microsoft.com>
* Create training session + minor improvements (#4668)
Co-authored-by: Rayan Krishnan <t-rakr@OrtDevTest2v100.af05slrtruoetgaxwwjv5nsq5e.px.internal.cloudapp.net>
* Save ONNX model in file (#4671)
Co-authored-by: Rayan Krishnan <t-rakr@OrtDevTest2v100.af05slrtruoetgaxwwjv5nsq5e.px.internal.cloudapp.net>
* Add eval step (#4674)
Co-authored-by: Rayan Krishnan <t-rakr@OrtDevTest2v100.af05slrtruoetgaxwwjv5nsq5e.px.internal.cloudapp.net>
* Add train_step (#4677)
Co-authored-by: Rayan Krishnan <t-rakr@OrtDevTest2v100.af05slrtruoetgaxwwjv5nsq5e.px.internal.cloudapp.net>
* Add LR Scheduler (#4694)
Co-authored-by: Rayan Krishnan <t-rakr@OrtDevTest2v100.af05slrtruoetgaxwwjv5nsq5e.px.internal.cloudapp.net>
Co-authored-by: Thiago Crepaldi <thiago.crepaldi@microsoft.com>
* Add deterministic compute tests (#4716)
Co-authored-by: Rayan Krishnan <t-rakr@OrtDevTest2v100.af05slrtruoetgaxwwjv5nsq5e.px.internal.cloudapp.net>
Co-authored-by: Thiago Crepaldi <thiago.crepaldi@microsoft.com>
* Add legacy vs experimental ORTTrainer accuracy comparison (#4727)
Co-authored-by: Rayan Krishnan <t-rakr@OrtDevTest2v100.af05slrtruoetgaxwwjv5nsq5e.px.internal.cloudapp.net>
Co-authored-by: Thiago Crepaldi <thiago.crepaldi@microsoft.com>
* Add Mixed precision/LossScaler + several fixes (#4739)
Additionally to the mixed precision/loss scaler code, this PR includes:
* Fix CUDA training
* Add optimization_step into TrainStepInfo class
* Refactor LRSCheduler to use optimization_step instead of step
* Updated several default values at ORTTrainerOptions
* Add initial Gradient Accumulation supported. Untested
* Fix ONNX model post processing
* Refactor unit tests
* Add ONNX BERT example + minor fixes (#4757)
* Fix training issue when passing ONNX file into ORTTrainer
Co-authored-by: Thiago Crepaldi <thiago.crepaldi@microsoft.com>
Co-authored-by: Rayan Krishnan <t-rakr@OrtDevTest2v100.af05slrtruoetgaxwwjv5nsq5e.px.internal.cloudapp.net>
* Add Dynamic Shape support (#4758)
* Update DeepSpeed Zero Stage option to a separate option group (#4772)
* Add support to fetches (#4777)
* Add Gradient Accumulation Steps support (#4793)
* Fix Dynamic Axes feature and add unit test (#4795)
* Add frozen weights test (#4807)
* Move new pytorch front-end to 'experimental' namespace (#4814)
* Fix build
Co-authored-by: Rayan-Krishnan <rayankrishnan@live.com>
Co-authored-by: Rayan Krishnan <t-rakr@OrtDevTest2v100.af05slrtruoetgaxwwjv5nsq5e.px.internal.cloudapp.net>
* bump onnx to support bfloat16
* sign test code
* fix ut failures
* add bfloat type in gradient schema
* add bfloat16 to gathernd
* add bfloat16 into grad op defs
* temp disable gpu fusing transformers
* bfloat16 support fix
* more fix to bfloat
* bug ifx
* add bfloat16 to transpose matmul
* fix sce loss
* fix cast opset13 and other missing part of bfloat16
* Revert "temp disable gpu fusing transformers"
This reverts commit b627bc9019.
* add SCEloss back
* fix build break
* fix gpu failure due to missing kernel in opset13
* add tile opset 13 kernel
* Revert "fix gpu failure due to missing kernel in opset13"
This reverts commit 661d63d0599029757f240d29afd64b197b76b880.
* fix comments in pr
* fix cuda break due to opset13
* fix missing msdomain
* add nll loss tests into android build's broken list; disable bfloat16 cast tests due to the wrong type saved in onnx test data, will fix it in onnx first
Co-authored-by: Cheng Tang <chenta@microsoft.com>
Fix Transpose MatMul fusion handling of existing TransposeScaleMatMul node's attributes and enable support for missing Transpose perm attribute.
Update expected test data to account for floating point calculation differences resulting from the fusion.
* Reshape optimization
* Refactor the Reshape optimization to be more generic
Co-authored-by: Ke Deng <kedeng@OrtTrainingDev1.af05slrtruoetgaxwwjv5nsq5e.px.internal.cloudapp.net>
* Gelu Activation Recompute Draft
* Prototype for localized recompute
* Introduce localized_recompute rewriter
* Command line args for enabling recompute
* Add logger to Gradient Graph Builder
* use const when possible
Update TransposeMatMul to support scaling of the matrix product by a constant scalar value (analogous to the GEMM alpha parameter). Rename TransposeMatMul to TransposeScaleMatMul.
Fuse MatMul with surrounding Mul/Div with constant scalar into TransposeScaleMatMul.
* Add ability to retrieve inferred shapes when executing a kernel.
This ability helps Recv to know its output shapes without doing
actual cummunication. Of course, if the output shapes cannot be
inferred, Recv still needs to do communication to get shapes from
Send.
* Avoid communicating shape information when it can be inferred statically
* Replace unordered_map with thread-safe wrapper.
We don't want to have racing condition and undefined behavior
when using parallel executor.y
* Remove cout
* Add missing file
* Address comments
* Check dim_value. -1 means missing
* lock properly
* Address comments (remove thread-safe map)
* Remove poc header
* Replace Stream with DeferredReleaseCPUPtr
* added reducesumlogexp gradient
added test
fixed type mismatch when calling cudnnreduce kernel
fixed python frontend to remove redundant states to match pytorch state dict
* Adding CPU implementation of BroadcastGradientArgs op
Modify to take shape as input instead of tensor
Cleanup
Correct schema
Corrected kernel, added tests, addressed review comments.
Initial change, to add ReduceSumTraining cpu op
cpu support
Initial changes to gradient builder
Non-empty reduction case passing.
Added exception,test for invalid broadcast,addresed review comments.
Initial change, to add ReduceSumTraining cpu op
cpu support
cuda support + more UTs
on comments + UT
no op support for {} axes with new attr - noop_with_empty_axes
Add noop attribute to ReduceSumTraining use
Add testing for no-shape graph, modify AddSub grad builder, logging.:
MulGrad support
Div support
Expand support
Gemm support
MatMul grad change
Transpose Grad change
BiasGeluGrad change.
Fixes after squash
* Remove logging, add specific exception for shape inference error
* fix build
* Review comments
* Review comments
* Fix windows build
Co-authored-by: Ethan Tao <ettao@microsoft.com>
* Adjust indentation of statement, without this fix GCC 7.5 errors
out with:
"this ‘if’ clause does not guard this statement, but the
latter is misleadingly indented as if it were guarded by the ‘if’"
* Add braces around the if-statement for improved clarity.
Co-authored-by: Alberto Magni <alberto.magni@microsoft.com>
* Working changes for ConcatTraining op
* Refactor to move changes to orttraining
* Fix segfault
* Support -ve axis for shape inferencing
* fix build
Co-authored-by: Ethan Tao <ettao@microsoft.com>
* Adding CPU implementation of BroadcastGradientArgs op
* Modify to take shape as input instead of tensor
* Cleanup
* Correct schema
* Corrected kernel, added tests, addressed review comments.
* Added exception,test for invalid broadcast,addresed review comments.
* Fix mac build error.
* Initial change, to add ReduceSumTraining cpu op
* cpu support
* cuda support + more UTs
* on comments + UT
* no op support for {} axes with new attr - noop_with_empty_axes
* on comments
* fix build
* on comments
Co-authored-by: aishwarya bhandare <aibhanda@microsoft.com>
Co-authored-by: Ethan Tao <ettao@microsoft.com>
* Deprecate TrainableDropout.
* Add Dropout(12) back into Megatron transformer.
* Remove TrainableDropout from front-end test models.
* Update baseline for front-end tests after converting test models to opset-12.
* Update baseline for front-end tests after converting test models to opset-12.
* Revise pipeline schedule to consider communication ops
* Add test
* Fix warning
* inline some short functions
* Fix warnings
* Rename a class
* Add comment for test
* op renamed to task
* Fix NVTX wrapper's bug
* concat
* add path_utils
* address feedback
* use string in test
* convert wstring to sting in windows
* address feedback
* address feedback
* fix comment
* Replace loss function in BERT_LOSS with SoftmaxCrossEntropyLoss.
* Update BERT loss function with correct logit shapes for softmax cross entropy loss.
* fix test and PR comments.
* support bert partition with shared initializer
* address feedback
* address feedback
* address feedback
* add more test
* remove bert-tiny model
* address feedback
* address function comment
* move CreateNodeArg to graph_utils
* rename function name
* rename function name
* fix windows build
* fix windows type conversion warning
* add function comment
- make size_ and data_ data members private
- rename GetCapacity() to Capacity() to be consistent (e.g., with Size())
- add static_assert for trivially copyable T because it is copied with memcpy
