saymrwulf/onnxruntime
1
0
Fork 0
mirror of https://github.com/saymrwulf/onnxruntime.git synced 2026-05-17 21:10:43 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions
onnxruntime/cmake/onnxruntime_optimizer.cmake
pengwa 2f5bf75e51
Optimize computation orders (#13672) 
### Optimize computation orders

In `Roberta/Electra`, when `ClassificationHead` is used, there is
slicing operation on features on sequence_length dimensions, then loss
calculations only depend on this sliced data. This is a slicing at axis
1. Before slicing the shape is [batch, sequence_length, hidden], after
slicing, it becomes [batch , hidden_stage]

We had opportunities to bring this slicing earlier as much as possible,
by passing through simple elementwise ops (like Add/Div), or
Layernorm/Softmax(if their reduce axis is after the slicing axis), or
even MatMul's the left operand (if only it did not affect the last
dims).

For operators like Reshape/Transpose, it is special since they have
either data specified (after slicing we need update), or they have perm
specified, which requires the input rank remain unchanged. So for those
kinds of operators, we can remain the original rank, but just leave the
sliced dim to be 1, after the compute completed, we do a Squeeze.

```
class RobertaClassificationHead(nn.Module):
    """Head for sentence-level classification tasks."""

    def __init__(self, config):
        super().__init__()
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        classifier_dropout = (
            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
        )
        self.dropout = nn.Dropout(classifier_dropout)
        self.out_proj = nn.Linear(config.hidden_size, config.num_labels)

    def forward(self, features, **kwargs):
        x = features[:, 0, :]  # take <s> token (equiv. to [CLS])
        x = self.dropout(x)
        x = self.dense(x)
        x = torch.tanh(x)
        x = self.dropout(x)
        x = self.out_proj(x)
        return x
```

src\transformers\models\roberta\modeling_roberta.py
src\transformers\models\electra\modeling_electra.py

#### Benchmark

A simple benchmark shows Robeta training latency dropped from 208ms ~
199ms. 4.5+% reduction.
More comprehensive tests are on the way.

### 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. -->
2022-12-22 15:12:52 +08:00

113 lines
6.4 KiB
CMake
Raw Blame History

# Copyright (c) Microsoft Corporation. All rights reserved.
# Licensed under the MIT License.
set(onnxruntime_optimizer_src_patterns)
if (onnxruntime_MINIMAL_BUILD)
# we include a couple of files so a library is produced and we minimize other changes to the build setup.
# if the transformer base class is unused it will be excluded from the final binary size
list(APPEND onnxruntime_optimizer_src_patterns
"${ONNXRUNTIME_INCLUDE_DIR}/core/optimizer/graph_transformer.h"
"${ONNXRUNTIME_ROOT}/core/optimizer/graph_transformer.cc"
)
if (onnxruntime_EXTENDED_MINIMAL_BUILD)
list(APPEND onnxruntime_optimizer_src_patterns
"${ONNXRUNTIME_INCLUDE_DIR}/core/optimizer/graph_transformer_utils.h"
"${ONNXRUNTIME_ROOT}/core/optimizer/conv_activation_fusion.cc"
"${ONNXRUNTIME_ROOT}/core/optimizer/conv_activation_fusion.h"
"${ONNXRUNTIME_ROOT}/core/optimizer/graph_transformer_utils.cc"
"${ONNXRUNTIME_ROOT}/core/optimizer/initializer.cc"
"${ONNXRUNTIME_ROOT}/core/optimizer/initializer.h"
"${ONNXRUNTIME_ROOT}/core/optimizer/nhwc_transformer.cc"
"${ONNXRUNTIME_ROOT}/core/optimizer/nhwc_transformer.h"
"${ONNXRUNTIME_ROOT}/core/optimizer/qdq_transformer/qdq_final_cleanup.cc"
"${ONNXRUNTIME_ROOT}/core/optimizer/qdq_transformer/qdq_final_cleanup.h"
"${ONNXRUNTIME_ROOT}/core/optimizer/qdq_transformer/qdq_util.cc"
"${ONNXRUNTIME_ROOT}/core/optimizer/qdq_transformer/qdq_util.h"
"${ONNXRUNTIME_ROOT}/core/optimizer/qdq_transformer/selectors_actions/qdq_actions.cc"
"${ONNXRUNTIME_ROOT}/core/optimizer/qdq_transformer/selectors_actions/qdq_actions.h"
"${ONNXRUNTIME_ROOT}/core/optimizer/qdq_transformer/selectors_actions/qdq_selector_action_transformer.cc"
"${ONNXRUNTIME_ROOT}/core/optimizer/qdq_transformer/selectors_actions/qdq_selector_action_transformer.h"
"${ONNXRUNTIME_ROOT}/core/optimizer/qdq_transformer/selectors_actions/qdq_selectors.cc"
"${ONNXRUNTIME_ROOT}/core/optimizer/qdq_transformer/selectors_actions/qdq_selectors.h"
"${ONNXRUNTIME_ROOT}/core/optimizer/qdq_transformer/selectors_actions/shared/utils.cc"
"${ONNXRUNTIME_ROOT}/core/optimizer/qdq_transformer/selectors_actions/shared/utils.h"
"${ONNXRUNTIME_ROOT}/core/optimizer/selectors_actions/actions.cc"
"${ONNXRUNTIME_ROOT}/core/optimizer/selectors_actions/actions.h"
"${ONNXRUNTIME_ROOT}/core/optimizer/selectors_actions/helpers.cc"
"${ONNXRUNTIME_ROOT}/core/optimizer/selectors_actions/helpers.cc"
"${ONNXRUNTIME_ROOT}/core/optimizer/selectors_actions/helpers.h"
"${ONNXRUNTIME_ROOT}/core/optimizer/selectors_actions/helpers.h"
"${ONNXRUNTIME_ROOT}/core/optimizer/selectors_actions/selector_action_transformer_apply_contexts.h"
"${ONNXRUNTIME_ROOT}/core/optimizer/selectors_actions/selector_action_transformer.cc"
"${ONNXRUNTIME_ROOT}/core/optimizer/selectors_actions/selector_action_transformer.h"
"${ONNXRUNTIME_ROOT}/core/optimizer/transpose_optimizer/optimizer_api_impl.cc"
"${ONNXRUNTIME_ROOT}/core/optimizer/transpose_optimizer/optimizer_api.h"
"${ONNXRUNTIME_ROOT}/core/optimizer/transpose_optimizer/optimizer_utils.h"
"${ONNXRUNTIME_ROOT}/core/optimizer/transpose_optimizer/ort_transpose_optimizer.cc"
"${ONNXRUNTIME_ROOT}/core/optimizer/transpose_optimizer/ort_transpose_optimizer.h"
"${ONNXRUNTIME_ROOT}/core/optimizer/transpose_optimizer/transpose_optimizer.cc"
"${ONNXRUNTIME_ROOT}/core/optimizer/utils.cc"
"${ONNXRUNTIME_ROOT}/core/optimizer/utils.h"
)
endif()
else()
list(APPEND onnxruntime_optimizer_src_patterns
"${ONNXRUNTIME_INCLUDE_DIR}/core/optimizer/*.h"
"${ONNXRUNTIME_ROOT}/core/optimizer/*.h"
"${ONNXRUNTIME_ROOT}/core/optimizer/*.cc"
"${ONNXRUNTIME_ROOT}/core/optimizer/compute_optimizer/*.h"
"${ONNXRUNTIME_ROOT}/core/optimizer/compute_optimizer/*.cc"
"${ONNXRUNTIME_ROOT}/core/optimizer/qdq_transformer/*.h"
"${ONNXRUNTIME_ROOT}/core/optimizer/qdq_transformer/*.cc"
"${ONNXRUNTIME_ROOT}/core/optimizer/qdq_transformer/selectors_actions/*.h"
"${ONNXRUNTIME_ROOT}/core/optimizer/qdq_transformer/selectors_actions/*.cc"
"${ONNXRUNTIME_ROOT}/core/optimizer/qdq_transformer/selectors_actions/shared/utils.h"
"${ONNXRUNTIME_ROOT}/core/optimizer/qdq_transformer/selectors_actions/shared/utils.cc"
"${ONNXRUNTIME_ROOT}/core/optimizer/selectors_actions/*.h"
"${ONNXRUNTIME_ROOT}/core/optimizer/selectors_actions/*.cc"
"${ONNXRUNTIME_ROOT}/core/optimizer/transpose_optimizer/*.h"
"${ONNXRUNTIME_ROOT}/core/optimizer/transpose_optimizer/*.cc"
)
endif()
if (onnxruntime_ENABLE_TRAINING)
list(APPEND onnxruntime_optimizer_src_patterns
"${ORTTRAINING_SOURCE_DIR}/core/optimizer/*.h"
"${ORTTRAINING_SOURCE_DIR}/core/optimizer/*.cc"
)
endif()
file(GLOB onnxruntime_optimizer_srcs CONFIGURE_DEPENDS ${onnxruntime_optimizer_src_patterns})
source_group(TREE ${REPO_ROOT} FILES ${onnxruntime_optimizer_srcs})
if (onnxruntime_EXTERNAL_TRANSFORMER_SRC_PATH)
set(onnxruntime_external_transformer_src_patterns)
list(APPEND onnxruntime_external_transformer_src_patterns
"${onnxruntime_EXTERNAL_TRANSFORMER_SRC_PATH}/*.cc"
"${onnxruntime_EXTERNAL_TRANSFORMER_SRC_PATH}/*.cpp"
)
file(GLOB onnxruntime_external_transformer_src ${onnxruntime_external_transformer_src_patterns})
list(APPEND onnxruntime_optimizer_srcs ${onnxruntime_external_transformer_src})
endif()
onnxruntime_add_static_library(onnxruntime_optimizer ${onnxruntime_optimizer_srcs})
install(DIRECTORY ${PROJECT_SOURCE_DIR}/../include/onnxruntime/core/optimizer DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/onnxruntime/core)
onnxruntime_add_include_to_target(onnxruntime_optimizer onnxruntime_common onnxruntime_framework onnx onnx_proto ${PROTOBUF_LIB} flatbuffers::flatbuffers Boost::mp11 safeint_interface)
target_include_directories(onnxruntime_optimizer PRIVATE ${ONNXRUNTIME_ROOT})
if (onnxruntime_ENABLE_TRAINING)
target_include_directories(onnxruntime_optimizer PRIVATE ${ORTTRAINING_ROOT})
endif()
add_dependencies(onnxruntime_optimizer ${onnxruntime_EXTERNAL_DEPENDENCIES})
set_target_properties(onnxruntime_optimizer PROPERTIES FOLDER "ONNXRuntime")
if (NOT onnxruntime_BUILD_SHARED_LIB)
install(TARGETS onnxruntime_optimizer
ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR}
LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR}
RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR}
FRAMEWORK DESTINATION ${CMAKE_INSTALL_BINDIR})
endif()
Reference in a new issue View git blame Copy permalink