### Support SCELoss/SCELossGrad run with larger sized input
#### Motivation and Context: Run bigger batch size for Bloom model.
For Bloom560M model, ORT has potential to run bigger batch size from
initialally 6 to now 10. SCELoss/SCELossGrad's input size is Bsz X 1023
* 250680. When Bsz is bigger than 8, totoal element count cannot be
represented by int32_t, which those kernels are using to passing total
elem count. There is silent overflow causing other indirectly
exceptions, or wrong mistake without errors.
#### Changes in this PR
- For SCELossInternal/SCELossGradInternal CUDA kernels, use uint64_t if
total element count is bigger than int32::max() to pass all element
count and element index for the ops mentioned above.
- For SCELossInternal/SCELossGradInternal CPU kernels,
- always use uint64_t to pass the element count.
- update the Eigen functions involved in the two kernels'
implementations, to use `ptrdiff_t` to pass element count instead of
original `int`.
- Parallelize SCELossInternal/SCELossGradInternal CPU kernels,
otherwise, it is super slow when handling so many elements.
- Others changed needed:
- Add `CompareOrtValueNumerals` to compare two OrtValue with different
data types (float or float16), without caller explicitly converting to
the lower-precision data types. The comparison is also done in parallel,
which reduce the comparsion time for the large UT case from 22s to
~1.6s.
- The check of `IsResultCloselyMatch` is buggy for nan/inf cases, so fix
the bugs.
- The cross entropy tests are running CPU base line with float, then the
result is used to compare with float16 results of CUDA runs. But there
is precision issue when we check the results. Because the randomized
input data is represented in float, CPU use it directly, but CUDA use a
float16 version of it, so there is precision diff between the inputs, as
the test data count increases, it make the results fail even on 1e-2.
The fix is: generate data in float16, convert to float for CPU run,
directly use float16 for CUDA runs. When compare the output, cast back
CPU float to float16 then compare with CUDA outputs.
- `RandomValueGenerator ` for the large size take about ~20second, so
`ParallelRandomValueGenerator ` is added to random input in parallel, it
takes about <2s for preparing input data.
#### Non-goals
`SoftmaxCrossEntropyLoss` && `SoftmaxCrossEntropyLossGrad` is not
covered in this PR
### Description
Expose `OrtValue` class API as first-class citizen.
Make it simular with C++ API.
Enable safe direct native memory access.
Make string tensor manipulation more efficient.
Avoid intermediate structures such as `NamedOnnxValue`,
`DisposableNamedOnnxvalue` and etc.
Provide more examples with `IOBinding`, although `OrtValue` API
potentially makes `IOBinding` redundant for most of scenarios, since
`OrtValue` can be created on top of any memory.
Run all the pre-trained models now with `OrtValue` API as well.
Obsolete `OrtExternalMemory class`. Obsolete IOBinding API that takes
`FixedBufferOnnxValue`.
### Motivation and Context
Make the API efficient and uniform with C++.
This aspires to address:
https://github.com/microsoft/onnxruntime/issues/14918https://github.com/microsoft/onnxruntime/issues/15381
Cc: @Craigacp
### Description
Remove AllocatorManager class
### Motivation and Context
After the refactor PR #15833 is in, AllocatorManager class is not
referenced anymore.
### Description
- Implement Objective-C binding for `ORTTrainingSession`
- Add `ORTUtils` utility class to handle conversion between C++ and
Objective-C types
- Add test case for saving checkpoint
- Add unit test cases for `ORTTrainingSession`
### Motivation and Context
This PR is part of implementing Objective-C bindings for training API.
It implements objective-c binding for training session. The objective-C
API closely resembles the C++ API.
---------
Co-authored-by: Edward Chen <18449977+edgchen1@users.noreply.github.com>
### Description
This PR implements a backward-compatible way to define custom operators
with fallible compute functions. The C++ API templated gained an
optional `Fallible` argument. Closes#14287
### Motivation and Context
#14287 contains more context. The gist is that the current C-API defines
compute operations of custom operators as functions returning `void`
rather than an `OrtStatusPtr`. Currently, errors are often propagated
across the C-ABI using C++ exceptions. That is very unsafe and undefined
behavior. Moreover, it is difficult for languages other than C++ to use
this approach even if they wanted to. A C-compliant sound and safe way
to propagate errors allows for non-C++ fallible custom operators.
### An example in action
https://github.com/cbourjau/ort-custom-op/pull/6/files is a
demonstration of how this PR can be used to write safe and fallible
custom operators in Rust.
### Description
<!-- Describe your changes. -->
New logic to share allocators among module, optimizer and eval sessions
for Training scenario
### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->
Previously on device training using shared allocator by sharing EP, now
with new mechanism to share allocator, we need to explicitly register
allocator in the environment.
---------
Co-authored-by: Lei Cao <leca@microsoft.com>
### Description
Make BeamScorer run on the GPU vs the CPU.
Brief overview:
Adds a CUDA 'CudaBeamSearchScorer' implementation of IBeamScorer
Instead of a 'done' flag per beam, there is one single 'not done'
variable that is copied to the CPU every iteration
Removes some of the extra CPU side buffers and parameters that are no
longer needed
Remaining future optimizations:
CPU copied beam indices is still used in the non
DecoderMaskedSelfAttention case. An extra kernel can be written to avoid
PickGptPasteState needing CPU copied beam indices (called from
UpdateGptFeeds).
### Motivation and Context
It's faster to keep the work on the GPU to avoid GPU->CPU->GPU copies of
data.
### Description
Fixes a typo that prevents skipping a test that targets the QNN HTP
backend on Windows x64.
### Motivation and Context
- Windows x64 machines cannot load/run the QNN HTP backend. Therefore,
we need to skip such tests on Windows x64.
- Fixes the QNN_Nuget_Windows pipeline.
### Description
<!-- Describe your changes. -->
As title.
### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->
Works with local onnxruntime-c pod in js/rn/e2e test.
A couple of places in onnxruntime used `float_t` data type alias as an
alternative to `float`. However, this is not entirely correct, since
`float_t` is an implementation-defined type alias, which may be `float`,
`double`, `long double` or some other implementation-defined data type,
depending on the state of the internal `FLT_EVAL_METHOD` macro:
https://en.cppreference.com/w/c/numeric/math/float_t
On most major platforms and compilers (clang, GCC, MSVC) this is only a
cosmetic change and will not lead to any changes. However, icpx compiler
(and legacy icc) tends to substitute `float_t` with `long double`,
resulting in a linker error (unresolved reference) to the base onnx
library, that only contains the `ParseData` function for `float` and
`double` as in
[here](9264e09367/onnx/defs/tensor_proto_util.cc (L133-L134)).
Overall, this PR cleans up the implementation-defined behaviour and
enables building onnxruntime with icpx.
The image for the onnxruntime-CI-nightly-ort-pipeline is too old.
The ort package in the image is older than latest test code in nightly
ci. This causes the nightly ci failed.
### Manage ORTModule options
Move all env vars that used for feature ON/OFF into runtime options for
consistent managements.
Be noted: the features' switch are assigned in 2 phases: default values,
overwritten by env vars (if specified by users). So env vars take the
highest priority when all 2 phases both given value explicitly for one
feature.
### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->
### Use PadAndUnflatten to replace GatherGrad for restore
### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->
### Description
Fix the nullptr check so that it would check the actual existence of
engine/context
(Currently, it checks the address of unique_ptr, which is always not
null. Thx @jslhcl for pointing that out)
> A quick recall of struct
[trt_state](https://github.com/microsoft/onnxruntime/blob/main/onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.h#L104):
> ```
> std::unique_ptr<nvinfer1::ICudaEngine>* engine = nullptr;
>std::unique_ptr<nvinfer1::IExecutionContext>* context = nullptr;
>```
### Motivation and Context
https://github.com/microsoft/onnxruntime/issues/15982
The incorrect check couldn't stop TRT EP from loading incompatible
engine cache on purpose, which invokes unhandled exception
### Description
Eliminate Cast operator if Shape is the next one.
### Motivation and Context
#### Cast
When working with onnx opset 15 and above, the shape operator now
accepts all types of variables.
This change is documented in the [onnx
Changelog](https://github.com/onnx/onnx/blob/main/docs/Changelog.md#Shape-15).
As a result, casting variables right before the shape operation becomes
unnecessary.
Removing these unnecessary casts will improve the graph and potentially
provide better performance gains.
## Results
On :
torchrun examples/onnxruntime/training/language-modeling/run_clm.py
--model_name_or_path gpt2 --do_train --overwrite_output_dir --output_dir
./outputs/ --seed 1337 --fp16 True --per_device_train_batch_size 4
--num_train_epochs 1 --dataset_name wikitext --dataset_config_name
wikitext-2-raw-v1 --learning_rate 2e-5 --report_to none --optim
adamw_ort_fused
without changes:
***** train metrics *****
epoch = 1.0
train_loss = 3.2981
train_runtime = 0:02:13.29
train_samples = 2318
train_samples_per_second = 17.39
train_steps_per_second = 4.351
With my changes:
***** train metrics *****
epoch = 1.0
train_loss = 3.2981
train_runtime = 0:02:08.98
train_samples = 2318
train_samples_per_second = 17.971
train_steps_per_second = 4.497
We see around 3% gain.
---------
Co-authored-by: Adam Louly <adamlouly@microsoft.com@orttrainingdev9.d32nl1ml4oruzj4qz3bqlggovf.px.internal.cloudapp.net>
### Description
<!-- Describe your changes. -->
Update file list to adjust for recent changes to test infra.
### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->
Some CI jobs may interrupted unexpectedly and didn't execute umount data
step. The data left in host device will cause `device or resource busy`
and make subsequent CI jobs fail.
Move the mount data step into docker container, the host machine will
not be occupied when CI jobs exit incorrectly.
enhance cast-propagation for "softmax can be put at fp16 when data flow
is cast-to-fp32 > softmax > cast-to-fp16"
this optimization can save gpu memory and have performance gain
### Description
Before transformers 4.27, the causal mask uses uint8 data type, so there
is extra Cast node to convert it to bool. This adds a pattern that
without Cast node to support attention fusion for GPT-2 models exported
with transformers >= 4.27.
### Motivation and Context
https://github.com/microsoft/onnxruntime/issues/16453
Python script to modify Onnx model to make it aligned with converted QNN
model
### Description
Onnxruntime QNN EP can support context binary file generated by QNN tool chain. However QNN generated context binary file uses channel last and 8 bits or 16 bits for input and output. This script get the QNN model input & output information from QNN converted model_net.json file, and insert Cast, Transpose nodes to Onnx model if required.
🛠️ __Changes in this pull request:__
This pull request introduces two significant changes to the project:
- Changing on device training checkpoint format: The current
implementation stores the on device training checkpoint as a sequence of
tensors in multiple files inside a checkpoint folder, which can be
inefficient in terms of storage and performance. In this PR, I have
modified the checkpoint format to utilize the flatbuffer table to save
the checkpoint to a single file, providing a more compact and efficient
representation. The changes around this are twofold:
- Add the checkpoint flatbuffer schema that will generate the necessary
checkpoint source files.
- Update the checkpoint saving and loading functionality to use the new
format.
- Adding support for onnxruntime minimal build: To support scenarios
where binary size is a constraint, I made changes to ensure that the
training build can work well with the minimal build.
🔍 __Open Issues:__
- In order to extract the optimizer type, the existing implementation
re-loaded the onnx optimizer model and parsed it. This is no longer
possible, since the model format can either be onnx or ort. One idea is
to do the same for ort format optimizer model. This needs some
investigation.
- Changes to the offline tooling to generate ort format training
artifacts.
- End-to-end training example showcasing the use of the minimal training
build.
- Add support for export model for inferencing in a minimal build.
### Description
<!-- Describe your changes. -->
protobuf CopyFrom doesn't work for model > 2GB for version 4.23. This PR
removes the copy for Calibrator.
Currently Calibrator copies the ModelProto to avoid changing it. The
reason is that: quantize_static passes a ModelProto to Calibrator to
calibrate quantitation parameters, and then use it for quantization. If
calibrator changes the ModelProto, quantizaiton won't work.
This PR changes quantize_static to pass in a model path to Calibrator
instead of a ModelProto, and make Calibrator only take in model path as
input, which is how it is used in most cases.
This PR also remove the optimization from quantization. User needs to
call pre-process to optimize the model
### Description
This PR fixes a typo with assigning the `repetition_penalty` input in
the Whisper export with beam search model. It is a follow-up to the
[export stabilization
PR](https://github.com/microsoft/onnxruntime/pull/16297).
### Motivation and Context
The `repetition_penalty` input should be set to `repetition_penalty`
instead of `input_features`.
### Description
Revert docker base image to
nvidia/cuda:11.8.0-cudnn8-devel-ubuntu20.04@sha256:b754c43fe9d62e88862d168c4ab9282618a376dbc54871467870366cacfa456e
### Motivation and Context
The default img env of nvidia/cuda:11.8.0-cudnn8-devel-ubuntu20.04 has
minor upgrade, which make Linux MultiGPU TensorRT CI (NV12 instance with
Maxwell GPU) fail on three CApiTestGlobalThreadPoolsWithProvider
tests (these three tests have higher error which are above the tolerance)
That minor upgrade includes cudnn 8.7.0->8.9.0, which might be a factor
that make maxwell GPU generator higher error. CIs with T4 GPU are not
affected.
### Description
Modify the creating of webgl context.
Previous behavior:
STEP.1 - create canvas (document.createElement), if failed, goto step.2
else step.3
STEP.2 - create offscreenCanvas, if failed abort
STEP.3 - use the canvas created in step.1 or 2 to create webgl context.
if successful return context else abort
Now bahavior:
STEP.1 create offscreenCanvas, if failed goto step.3
STEP.2 use it to create webgl context. if successful, return context
STEP.3 create canvas (document.createElement). if failed, abort
STEP.4 use it to create webgl context. if successful, return context
else abort
Motivation:
we found in some environment, normalCanvas.getContext() returns null but
offscreenCanvas.getContext() returns the context object. and when
offscreenCanvas is available it is good idea to always prefer to use it.
### Description
DORT support for custom ops
### Motivation and Context
Custom ops registered via custom_ops shared_library cannot be run using
DORT atm. This PR enables it using:
1. registering custom_ops supported in DORT
2. plumbing down session_options from OrtBackend when creating the
InferenceSession, that were used to register the custom_ops shared
library using
`session_options.register_custom_ops_library(shared_library)`
