### Description
Pause caching the docker images in pipeline cache in Linux Aten
Pipeline.
### Motivation and Context
We need to work out a better way to reduce the storage.
### Description
- Adds support for newer opset of Reduction operators (ReduceSum,
ReduceMax, ReduceMin, ReduceMean, ReduceProd) with axes as an
initializer input.
- Adds tests for HTP and CPU backends.
### Motivation and Context
Newer opset versions changed the `axes` attribute into an optional
input. This PR adds support for these newer reduction operators as long
as the axes input is defined as an initializer. The goal is to enable more models on QNN.
### Description
Improve Slice to support Onnx opset9 which has starts, ends & axes in
node attributes.
### Motivation and Context
To unblock some models.
### Description
`lintrunner` is a linter runner successfully used by pytorch, onnx and
onnx-script. It provides a uniform experience running linters locally
and in CI. It supports all major dev systems: Windows, Linux and MacOs.
The checks are enforced by the `Python format` workflow.
This PR adopts `lintrunner` to onnxruntime and fixed ~2000 flake8 errors
in Python code. `lintrunner` now runs all required python lints
including `ruff`(replacing `flake8`), `black` and `isort`. Future lints
like `clang-format` can be added.
Most errors are auto-fixed by `ruff` and the fixes should be considered
robust.
Lints that are more complicated to fix are applied `# noqa` for now and
should be fixed in follow up PRs.
### Notable changes
1. This PR **removed some suboptimal patterns**:
- `not xxx in` -> `xxx not in` membership checks
- bare excepts (`except:` -> `except Exception`)
- unused imports
The follow up PR will remove:
- `import *`
- mutable values as default in function definitions (`def func(a=[])`)
- more unused imports
- unused local variables
2. Use `ruff` to replace `flake8`. `ruff` is much (40x) faster than
flake8 and is more robust. We are using it successfully in onnx and
onnx-script. It also supports auto-fixing many flake8 errors.
3. Removed the legacy flake8 ci flow and updated docs.
4. The added workflow supports SARIF code scanning reports on github,
example snapshot:
5. Removed `onnxruntime-python-checks-ci-pipeline` as redundant
### 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. -->
Unified linting experience in CI and local.
Replacing https://github.com/microsoft/onnxruntime/pull/14306
---------
Signed-off-by: Justin Chu <justinchu@microsoft.com>
### Description
<!-- Describe your changes. -->
Remove `deletion` of copy functions from `OrtApi` as its initialization
no longer compiles in C++20.
Introduce a non-copyable member to implicitly delete copy ctor.
### 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. -->
Inspired by https://github.com/microsoft/onnxruntime/pull/14901
Solution credits: @RyanUnderhill
Cc: @georgthegreat
### Description
Apply the DML graph fusion transformer optimization independently of ORT
graph optimization level.
### Motivation and Context
The DML graph fusion transformer is not a graph optimizer in the normal
sense: it isn't optimizing the ONNX graph structure, but rather fusing
nodes into what will later become a single IDMLCompiledOperator (using
IDMLDevice1::CompileGraph). This transformer can't be done ahead of time
(hence why it's disabled if saving an optimized model), but it's also
gated by the ORT graph optimization level; this makes it impossible to
preoptimize ONNX models ("offline mode") and then later disable graph
optimizations for better startup performance ("online mode") while
benefiting from DML graph fusion.
- Add compute type for Skiplayernorm to fix ROCm CI and get more
accurate results.
SkipLayerNorm:
type T: input, skip, bias
type U: epsilon, compute result
type V: output, beta, gamma
- refactor the usage of aligned_vector, reduce the usage of
`reinterpret_cast`.
### Description
Use build.sourceversion in docker image cache key.
### Motivation and Context
We used filpath as the cache key in #14496.
In most cases, the docker base image tag is latest.
So, the hash of the files couldn't be aware of the change of base image.
As the result, the docker image restored, but the image will still be
rebuilt .
The maintenance cost would be huge if we pin image hash in docker file.
For example,
https://quay.io/repository/pypa/manylinux2014_x86_64?tab=tags&tag=latest,
it's updated almost every week.
So far, the build.sourceversion is the right way to keep cache is
updated and valid.
### Description
<!-- Describe your changes. -->
Currently we bail on constant folding if QDQ is enabled and we hit a DQ
node. However, if we have a simple DQ -> X -> Q node unit where the DQ
and X do not produce graph outputs, their output only has one consumer,
and X is deterministic, we can constant fold all three nodes.
Add support for this simple scenario primarily to constant fold a QDQ
model that has had initializers updated by layout transformation, which
results in patterns like `initializer -> DQ -> Transpose -> Q` or
`initializer- > DQ -> Unsqueeze -> Q -> DQ -> Transpose -> Q` if the
initializer is broadcast.
### 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. -->
Improve end result of layout transformation on a QDQ model.
### Description
<!-- Describe your changes. -->
As synced offline, rename this op and will create another op for mha
that supports both self and cross attention.
### 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. -->
---------
Co-authored-by: Ubuntu <wy@v100-2.0cdb2e52twzevn1i4fi45bylyg.jx.internal.cloudapp.net>
### Description
<!-- Describe your changes. -->
1. upgrade cutlass to 3.0 that containing attn_bias support.
2. extend Attention/MHA to use memory efficient attention when
rel_pos_bias with [1, num_head, s, s*] and 1d mask with [2 * batch_size
+ 1] are present.
new mask format introduction:
MASK_1D_KEY_SEQ_LEN_START,
[3 * batch_size + 2] with [key_len[0], ..., key_len[batch_size - 1],
query_start[0], ..., query_start[batch_size - 1], query_end[batch_size -
1], key_start[0], ..., key_start[batch_size - 1], key_end[batch_size -
1]]
e.g
2D mask with [[1, 1, 1, 0, 0, 0], [1, 1, 1, 1, 1, 0]] converts to this
1D mask is [3, 5, 0, 6, 12, 0, 6, 12]
### 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. -->
It potentially benefits tnlrv6 and t5(encoder)
---------
Co-authored-by: Ubuntu <wy@v100-2.0cdb2e52twzevn1i4fi45bylyg.jx.internal.cloudapp.net>
Co-authored-by: Kunal Vaishnavi <kvaishnavi@microsoft.com>
Co-authored-by: Kunal Vaishnavi <kvaishnavi@microsoft.com@orttrainingdev7.d32nl1ml4oruzj4qz3bqlggovf.px.internal.cloudapp.net>
* Current SkipLayernorm did not using stable algo and cause correctness
issue.
* Enrich existing layernorm kernel to accept bias and residual.
* Tune standard layernorm threads.y according to elements and device
property.
* Remove existing skiplayernorm cuda implementation.
Fix two issues: (1) GPT attention fusion: get_parent could return None when the input is
initializer, add a check (2) ONNX node could have optional inputs and outputs. During
prune_graph, we shall exclude empty inputs/outputs. Here we exclude ""
from output_name_to_node and input_name_to_nodes.
Add an option allow_remove_graph_inputs in prune_graph
### Description
Fix issue relating to unused inputs of model-local functions. ORT
creates a schema for all such functions. The creation of this schema
does not handle unused function-inputs. The schema-creation relies on
the use of the function-inputs to infer type-constraints for the input,
and the code ends up creating an erroneous input-descriptor when there
is no use of the function-input.
The fix is to create an input with the given name, with a
type-constraint that allows all types.
### Motivation and Context
Fix https://github.com/microsoft/onnxruntime/issues/15046
Fix https://github.com/microsoft/onnx-script/issues/524
---------
Signed-off-by: Ganesan Ramalingam <grama@microsoft.com>
Co-authored-by: Justin Chu <justinchuby@users.noreply.github.com>
Co-authored-by: Scott McKay <skottmckay@gmail.com>
### Statistics tool for ORTModule convergence parity
As ORTModule get more and more validated, it is pretty fast to
intergrade PyTorch based model with ORT.
The same time, we need make sure once there is convergence issue, we
don't spend months of time to investigate. As part of this efforts, this
PR is introducing a tool to dump activation statistics without much
involvement from users. The dumping results contains only some statistic
numbers plus sampled data, which is not big, compared with dumping all
the tensors, it is much faster and space efficient.
For us to use it, two single lines are needed before wrapping ORTModule.
For baseline run, need also apply the same trick.
```
+ from onnxruntime.training.utils.hooks import SubscriberManager, StatisticsSubscriber
+ SubscriberManager.subscribe(model, [StatisticsSubscriber("pt_out", override_output_dir=True)])
```
Once you run the steps, following command can be used to merge result
into per-step-summary respectively for ORT and baseline runs.
```bash
python -m onnxruntime.training.utils.hooks.merge_activation_summary --pt_dir pt_out --ort_dir ort_out --output_dir /tmp/output
```
Docs is added here as part of this PR [convergence investigation
notes](https://github.com/microsoft/onnxruntime/blob/pengwa/conv_tool/docs/ORTModule_Convergence_Notes.md)
Based on the generated merged files, we can compare them with tools.
### Design and Implementation
This PR introduced a common mechanism registering custom logic for
nn.Module's post forward hooks. And statistics for activation
(StatisticsSubscriber) is one of the implementations. If there is other
needs, we can define another XXSubscriber to do the customized things.
### Description
<!-- Describe your changes. -->
As title.
- Only support constant mode Pad for CoreML EP for now.
- Enable Pad tests for CoreML with inputs as initializer types.
CoreML Spec for reference:
https://apple.github.io/coremltools/mlmodel/Format/NeuralNetwork.html#paddinglayerparams
### 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. -->
Fill operator gaps for ClipChamp models.
---------
Co-authored-by: rachguo <rachguo@rachguos-Mac-mini.local>
Co-authored-by: rachguo <rachguo@rachguos-Mini.attlocal.net>
### Fix reference count for autograd
When PythonOp kernel initialized, `AddPointerScalarArgs` creates
`const_args_` which put all non-tensor references (including
ProcessGroup, string, or other user types) in it.
In kernel's destructor, all ref cnt got decreased for `const_args_`.
```
void PythonOpBase::Clear() {
for (auto ptr : const_args_) {
auto obj = reinterpret_cast<PyObject*>(ptr);
Py_DECREF(obj);
}
}
```
It means, we did not increase cnt, but just decrease cnt. Running the
unit, segmentation fault will be thrown. The simple fix is to remove the
Py_DECREF for those pointer-type constant inputs triggered by kernel
destructor.
NONTENSOR_OBJECT_POINTER_STORE is the place we increase the reference
during export, then the reference will remain until the python program
terminates.
Additionally tunings:
1. Move some logs into verbose instead of warning in case of flooding
training logs.
2. Move pointer type ref holding from python side
(NONTENSOR_OBJECT_POINTER_STORE) to
orttraining/orttraining/core/framework/torch/custom_function_register.h.
Then we use a consistent approach to manage all PythonOp related python
object/methonds ref count increasing and decreasing.
### Description
This PR includes the following changes:
- upgrade js dependencies
- enable STRICT mode for web assembly build.
- corresponding fix for cmake-js upgrade
- corresponsing fix for linter upgrade
- upgrade default typescript compile option of:
- `moduleResolution`: from `node` to `node16`
- `target`: from `es2017` to `es2020`
- fix ESM module import in commonJS source file
## change explanation
### changes to onnxruntime_webassembly.cmake
`-s WASM=1` and `-s LLD_REPORT_UNDEFINED` in latest version is
by-default and deprecated.
### changes to onnxruntime_node.cmake
The npm package `cmake-js` updated its way to find file `node.lib`.
previously it downloads this file from Node.js public release channel,
and now it generates it from a definition file.
The node.js release channel does not contain a windows/arm64 version, so
previously cmake-js will fail to download `node.lib` for that platform.
this is why we made special handling to download the unofficial binary
to build. now this is no longer needed so we removed that from the cmake
file.
### changes to tsconfig.json
`node16` module resolution supports async import and `es2020` as target
supports top level await.
### Description
<!-- Describe your changes. -->
AMX isn't supportted until assembler 2.40 even though the GCC frontend
supports it.
### 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. -->
(1) Allow model to be path, and use infer_shapes_path to fix
https://github.com/microsoft/onnxruntime/issues/15063
(2) Add some logging for float data truncation
(3) Add RandomUniformLike to default op_block_list
(4) Some minor changes to use f string.
### Description
<!-- Describe your changes. -->
Transformer models can handle batch of inputs at once. However,
sequences in a batch usually have different length. Then we have to pad
the short one to have same length as the longest. This is not efficient
especially for large batch with high variance.
This PR introduces a PackedAttention operator which can take in packed
sequences (no padding) and also produces output in packing mode.
There will be another PR to use the PackedAttention to implement the
encoder in packing mode.
### 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
Convolution for fp16 datatype. Use NHWC for computation. For NCHW input,
it rearranges the input tensor to NHWC format before computing the
result.
Support two optional fusion:
1. Activation
2. Add (not yet implemented)
### Motivation and Context
Accelerating fp16 inference
### Description
Check the Mac x86_64 packages installation.
### Motivation and Context
To avoid installation error, add packages smoking test before release.
### Description
Patch https://github.com/microsoft/onnxruntime/pull/14767 in order to
make two provider options `force_timing_cache` and `detailed_build_log`
can be updated. Otherwise, they only use default value.
`timing_cache_enable` is good.
### Description
Remove the use of `eval` in test code so we don't (1) use eval and (2)
create "unused" local vars that ruff will remove. Predecessor to #15085
### Description
<!-- Describe your changes. -->
We are introducing the FasterTransfomer model-level integration using
ORT [custom op runtime
wrapper](https://github.com/microsoft/onnxruntime/pull/13427).
In order to make the FT wrapper/integration work, two things need to be
done:
- New API `KernelInfoGetConstantInput_tensor`. (Done in this PR)
During custom op kernel initialization, it needs to get the model
weights (saved as node's constant inputs) ready for FT's weights
instantiation. What's why we need to add this new API to make kernel
info capable of getting constant inputs.
- Custom op and custom op kernel to wrap FT model. (Will provide in
onnxruntime extensions or inference examples)
During custom op kernel initialization, it can fetch attributes from
kernel info to determine which kind of FT model instance create. During
custom op kernel compute/inference, it can get input/output from kernel
context and then assign input/output buffers for model instance to run.
