Add SparseAttention cpu implementation.
- [x] Refactoring GQAAttentionBase
- [x] Add SparseAttention implementation
- [x] Add test cases
This is unfused version. Flash attention version will be added later.
### Description
Make current ROCm packaging stages to a single workflow.
Reduce the possibility of all nightly packages can't be generated by one
failed stage
### Motivation and Context
Our plan is to reduce the complexity of the current zip-nuget pipeline
to improve the stability and performance of nightly packages generation.
ROCm packaging stages has no dependencies with other packaging jobs and
it's the most time-consuming route.
After this change, the most used CPU/CUDA/Mobile packaging workflow
duration can be reduced roughly from 3h20m to 2h30m.
### Problem
Currently, the codebase contains some logics pertaining to model
re-export checks and graph_builder reinitialization checks. Ideally,
these operations should function akin to a state machine. However, upon
inspecting the implementation, it becomes apparent that certain states
are checked or set in various scattered locations. This fragmentation
makes it challenging to comprehend when a re-export or re-initialization
will be triggered. For optimal clarity and maintainability, it is
advisable to consolidate these states into a cohesive component, rather
than dispersing them within the current graph execution manager.
Furthermore, the process of model exports and post-export processing for
stage 3 support or memory-efficient gradient management introduces
considerable complexity. To enhance the codebase's structure, it would
be beneficial to extract these intricate functionalities into a
dedicated component, divorcing them from the current graph execution
manager.
As part of the effort to improve the codebase, it's essential to address
inconsistencies in handling input/output flatten/unflatten operations.
Currently, there are several functions performing these operations
recursively, each with slightly different implementations. This
inconsistency leads to varying support for input/output data types and
structures in different parts of the code. To rectify this, the proposed
pull request simplifies these operations into a set of primitive
functions, ensuring uniformity. This not only streamlines the code but
also facilitates the maintenance of consistency when introducing bug
fixes or supporting new data types. One thing to mention here: input
output handling is deeply bound to the graph transition mentioned above,
so it is difficult to make this change separately.
While acknowledging the complexity of these logics, it is reassuring
that the codebase benefits from an extensive suite of unit tests that
cover all possible branches. Despite the intricacies, ensuring the
passage of all tests has been a time-intensive but necessary aspect of
this development effort.
### Design
Introduce `GraphTransitionManager` and put all model export and
post-export processing logics in it.
1. Re-export check
2. Do export
3. Re-post-export process check
4. Do post-export process
5. Return `PostExportProcessedModelInfo`, which contains all the
information we need, to pass to ORT to build gradient graph (currently
we do the same for training or evaluating, but ideally we should not do
it for evaluating, let's keep this behavior as it is now, and make the
change later).
```
# Input names for the pre-gradient-build graph.
# This may be different with the one in ExportedGraph since we may
modify the graph inputs as needed
# for example when memory efficient gradient management is enabled.
self.onnx_graph_input_names: list[str] = onnx_graph_input_names
# A subset of onnx_graph_input_names.
# Input names that require gradients for the pre-gradient-build graph.
self.onnx_graph_input_names_require_grad: list[str] =
onnx_graph_input_names_require_grad
# Create symbolic names for each dimension of the graph input (e.g.
onnx_graph_input_names).
# The key is the input name, the value is a dict of {dim_index:
symbolic_dim_name}
# e.g. {"input1": {0: "input1_dim0", 1: "input1_dim1"}, "input2": {0:
"input2_dim0"}}
self.onnx_graph_input_dynamic_axes_map: dict[str, dict[int, str]] =
onnx_graph_input_dynamic_axes_map
self.buffer_for_ort_runs: dict[str, torch.Tensor] = OrderedDict()
self.onnx_graph_input_names_user_defined = (
onnx_graph_input_names_user_defined # The ONNX graph input names
excluding the parameters, buffers.
)
# The ONNX graph input names excluding the parameters, buffers.
self.onnx_graph_input_names_require_grad_user_defined =
onnx_graph_input_names_require_grad_user_defined
self._post_export_processed_model: onnx.ModelProto | None =
post_export_processed_model
# A function to access the input data from the args and kwargs.
# If it is not None, the length is same as onnx_graph_input_names.
# For i-th input name, we can use the i-th function to get the input
data from args and kwargs.
self.data_accessor: list[callable] | None = data_accessor
# Used for unflattening the outputs from the ORT forward run.
self.module_forward_output_schema: ORTModelInputOutputSchemaType | None
= module_forward_output_schema```
The `GraphTransitionManager` instance is a property of
`GraphExecutionManager` (e.g. `TrainingManager` or ``InferenceManager),
1. Use
'self._graph_transition_manager.use_cache_or_reconstruct_post_processed_model(inputs,
kwargs)' to check whether the PyTorch module need a re-export or
re-post-export-process.
2. Use
`self._graph_transition_manager._post_export_processed_model_info.construct_inputs`
to construct the list of inputs used for ORT runs.
3. Use
`self._graph_transition_manager._post_export_processed_model_info.restore_outputs(user_outputs)`
to restore the outputs in original PyTorch output structure.
### 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
Add some macro to help print data to console for debugging purpose.
Example usage:
```
int input_id;
vector<int> some_vector;
DUMP_CPU_TENSOR_INIT()
DUMP_CPU_TENSOR("some vector", some_vector);
DUMP_STRING("input_id=", input_id);
```
- To enable dump thread id, set environment variable
`ORT_DUMP_THREAD_ID=0`.
- User can disable dumping by environment variable
`ORT_ENABLE_CPU_DUMP=0`.
### 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
It's the prerequisite step of reducing complexity of current zip-nuget
pipeline.
Some packaging tasks could be cut from the most complex nuget pipline
and easily be published
### 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
Disable using CoreML ML Program for a matmul where one of the inputs is
1D as the CoreML implementation appears to be broken. See
https://github.com/apple/coremltools/issues/2263
Add some debugging notes.
### Motivation and Context
Fix failing test on macos-14.
### Description
<!-- Describe your changes. -->
-It is an initial PR for VSINPU execution provider
### 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. -->
- For support VeriSilicon hardware
- TIM-VX(Tensor Interface Module)
(https://github.com/VeriSilicon/TIM-VX) is an integrated software
solution by Verisilicon for our hardware(A311D/i.MX 8M Plus etc.)
design, it is easy to use Verisilicon’s hardware by simply connecting
onnxruntime with the TIM-VX API by this VSINPU execution provider.
### Description
Update upstream packaging pipeline name to make it more meaningful.
### Motivation and Context
The upstream pipeline used to only building Nuget packages, but now it
also builds Zip and Java. So change the name will make it more
meaningful.
### Description
1. Update the functions in tensorprotoutils.h to use
std::filesystem::path instead of onnxruntime::Path. Eventually we can
remove the whole onnxruntime::Path class, but to this PR small I am not
doing that.
2. Remove the _SILENCE_EXPERIMENTAL_FILESYSTEM_DEPRECATION_WARNING macro
def when TensorRT EP is enabled.
### Description
Vitis AI EP synchronously supports the TensorProto data types supported
by ONNX 1.16.
Add error message show when graph resolve fail for troubleshooting.
### Motivation and Context
ONNX 1.15 & 1.16 add support some new TensorProto DataType , such as
- FLOAT8E4M3FN
- FLOAT8E4M3FNUZ
- FLOAT8E5M2
- FLOAT8E5M2FNUZ
- UINT4
- INT4
---------
Co-authored-by: liumingyue <mingyue@xilinx.com>
### Description
Provide user level options to control the fallback on CPU for models not
supported on Intel's NPU hardware.
### Motivation and Context
- Current workflow of OVEP allows safe fallback from OV NPU to OV CPU on
compilation failures. Also supports MLAS CPU fallback in presence of
unsupported custom ops.
- The PR provides a build-time option to disable fallback from OV NPU to
OV CPU.
- The session Option "kOrtSessionOptionsDisableCPUEPFallback" disables
OV CPU and MLAS CPU fallback.
- Also has bug fix for proto creation.
---------
Co-authored-by: jatinwadhwa921 <jatin.wadhwa@intel.com>
Co-authored-by: ankitm3k <ankit.maheshkar@intel.com>
### Description
1. Add QNN UTs for QNN Pad Op with FP16 data on HTP backend
2. Improve Pad op builder to handle invalid optional input
3. Add UT for ReduceSum for FP16 precision with 5D for issue reproduce
### Description
<!-- Describe your changes. -->
### 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
As suggested by SciPy's doc, we will
`Build against NumPy 2.0.0, then it will work for all NumPy versions
with the same major version number (NumPy does maintain backwards ABI
compatibility), and as far back as NumPy 1.19 series at the time of
writing`
I think it works because in
[numpyconfig.h#L64](https://github.com/numpy/numpy/blob/main/numpy/_core/include/numpy/numpyconfig.h#L64)
there is a macro NPY_FEATURE_VERSION. By default it is set to
NPY_1_19_API_VERSION. And the NPY_FEATURE_VERSION macro controls ABI.
This PR only upgrade the build time dependency; When a user installs
ONNX Runtime, they still can use numpy 1.x.
### Motivation and Context
Recently numpy published a new version, 2.0.0, which is incompatible with the latest ONNX Runtime release.
### Description
### Motivation and Context
The pipeline is green even Llama2 parity_check fails.
The PR should be merged after the below exception is solved.
'''
2024-06-25 03:49:43.621298481 [E:onnxruntime:,
sequential_executor.cc:514 ExecuteKernel] Non-zero status code returned
while running Expand node. Name:'/model/Expand' Status Message:
/model/Expand: left operand cannot broadcast on dim 3 LeftShape:
{1,1,9,9}, RightShape: {2,1,9,17}
An error occurred while verifying parity: Error in execution: Non-zero
status code returned while running Expand node. Name:'/model/Expand'
Status Message: /model/Expand: left operand cannot broadcast on dim 3
LeftShape: {1,1,9,9}, RightShape: {2,1,9,17}
Traceback (most recent call last):
File
"/workspace/onnxruntime/python/tools/transformers/models/llama/convert_to_onnx.py",
line 1043, in main
parity_check(parity_cmd)
File
"/workspace/onnxruntime/python/tools/transformers/models/llama/llama_parity.py",
line 298, in main
verify_parity(args, location, use_auth_token, kv_cache_ortvalues,
pytorch_model=llama, config=config)
File
"/workspace/onnxruntime/python/tools/transformers/models/llama/llama_parity.py",
line 137, in verify_parity
ort_model.run_with_iobinding(io_binding)
File
"/home/onnxruntimedev/.local/lib/python3.8/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py",
line 331, in run_with_iobinding
self._sess.run_with_iobinding(iobinding._iobinding, run_options)
RuntimeError: Error in execution: Non-zero status code returned while
running Expand node. Name:'/model/Expand' Status Message: /model/Expand:
left operand cannot broadcast on dim 3 LeftShape: {1,1,9,9}, RightShape:
{2,1,9,17}
'''
The exception looks caused by #19832
ONNX's Expand supports bidirectionally broadcast, while WebNN's expand
op only supports unidirectionally broadcast. Thus we should calculate
the output shape for 'newShape' input of WebNN's expand op.
### Description
<!-- Describe your changes. -->
The split op is using pin_memory when split on different sizes.
But pin_memory is not capable for using cudagraph.
Add a new implementation for only transformer scenarios, it split the
qkv_proj into q, k, v, not using pin_memory.
### 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
<!-- Describe your changes. -->
Setting the log level after environment creation is too late in some
cases.
If the DML EP is enabled, it will create a composite sink with the
original logger using the creation time log severity, as well as
additional ETW sink. As it saves the current severity levels for each
sink inside the composite sink that prevents being able to get verbose
log output to stdout even if you set that at the session level.
I don't know enough about the setup that combines ETW with the original
sink to say whether we should also be updating the severity of
individual sinks in the combined sink, so this change is limited to
making the unit tests behave in the expected manner when the default log
severity is set in the background and not directly controlled.
### 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. -->
Make it possible to get verbose output to stdout when the DML EP is
enabled.
### Description
Skip softmax BF16 test for ROCm, because BFloat16 is unsupported by
MIOpen, and `torch.cuda.is_available()` also returns `True` for ROCm.
### Problem
Newer models using more novel equations (e.g. `bhwc,hkc->bhwk` in
Segment Anything's encoder or `bqc,bchw->bqhw`) cause fallback from DML
to CPU, yielding performance issues. The EP had some pattern matching to
map more common equations to existing DML operators, but the number of
permutations was prohibitive and could not catch them all.
### Solution
So, ditch the static mapping, and instead handle any 1-input or 2-input
cases via remapped strides and a mini-graph of elementwise
multiplication & sum reduction (as if DML had a
`DML_OPERATOR_DOT_PRODUCT` that took `axes`). A subset of mappings still
exist for performance (GEMM, pure reduction, transpose...), but they are
identified generally rather than via a pattern table. Also...
- Diagonals are supported now (e.g. iji->i).
- Removes any remaining DML-specific EinSum `GTEST_SKIP` statements.
- Handles any cases up to 8 unique labels (DML dimension limit is 8D).
- \>= 3 inputs and arbitrary size inputs via ellipsis are not handled,
but we have yet to come across a model.
### Description
Alternative design from #20942
Allow users to pass in a model path for the generate_artifacts API.
### Motivation and Context
- ONNX API calls such as the onnx checker + shape inference fail when
given a model > 2GB, but work if a path to a model >2GB is passed in.
Context and motivation:
When quantizing large transformer models, we faced OOM issue when the
number of calibration samples goes up. To resolve this, in the PR we
want to add support for reading quantization data in chunck, calculating
ranges for intermediate tensors, then accumulating results for the final
ranges.
### Description
This reverts commit 1d7bf56947 because it
broken the AMD GPU CI pipeline. Sorry when I reviewed the PR I forgot to
run the AMD GPU CI pipeline.
Will revert the PR first then ask the author to fix the issue.
### Description
Update protobuf_cmake.patch to allow extra disablements. ORT repo
already patches protobuf to not disable the warning 4996.
### Motivation and Context
To meet SDL requirements, Microsoft repos have to fail build if there is
warning 4996
Binskim also gives errors if warning 4996 is disabled.
We can suppress the Binskim issues, but we need a way to disable the
warnings for the minimal set of code that has them.
Right now, WindowsAI disables 4996 for entirety of ORT, but it should
only be disabled for protobuf.
### Description
<!-- Describe your changes. -->
### 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
Under certain conditions with enabling & disabling ETW continuously, we
got a crash report.
Allows ETW callbacks to be de-registered upon class destructor.
Related to #20537
### Motivation and Context
Fixes crash
### Callstack
We see it crash in
[0x0]
onnxruntime!<lambda_967a738fca8512372f170fcaf2d094d4>::operator()+0x34
0x12941ff570 0x7ffa994f0a04
[0x1] onnxruntime!std::_Func_class<void,_GUID const *,unsigned
long,unsigned char,unsigned __int64,unsigned
__int64,_EVENT_FILTER_DESCRIPTOR *,void *>::operator()+0x54 0x12941ff7b0
0x7ffa994f0d64
[0x2]
onnxruntime!onnxruntime::logging::EtwRegistrationManager::InvokeCallbacks+0xcc
0x12941ff7b0 0x7ffa994f0d64
[0x3]
onnxruntime!onnxruntime::logging::EtwRegistrationManager::ORT_TL_EtwEnableCallback+0x94
0x12941ff860 0x7ffa98d19628
and seems to us that the this pointer captured in
etwRegistrationManager.RegisterInternalCallback(
[&etwRegistrationManager, this](
...
is no longer valid when the callback is called.
