Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
19 changes: 19 additions & 0 deletions modelopt/onnx/op_types.py
Original file line number Diff line number Diff line change
Expand Up @@ -303,3 +303,22 @@ def is_data_dependent_shape_op(op_type: str):
"NonZero",
"RoiAlign",
]


def get_symmetric_ops():
"""Returns set of commutative/symmetric operations where operand order doesn't matter."""
return {
"Add",
"Mul",
"And",
"Or",
"Xor",
"Equal",
"Max",
"Min",
"Sum",
"Mean",
"BitwiseAnd",
"BitwiseOr",
"BitwiseXor",
}
95 changes: 95 additions & 0 deletions modelopt/onnx/quantization/autotune/__init__.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,95 @@
# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

"""Pattern-Based Q/DQ Autotuning for ONNX Models.

This package provides automated optimization of Quantize/Dequantize (Q/DQ) node placement
in ONNX computation graphs to minimize TensorRT inference latency. It uses pattern-based
region analysis to efficiently explore and optimize Q/DQ insertion strategies.

**Core Components:**

Autotuner Classes:
- QDQAutotuner: Main autotuner with automatic hierarchical region discovery
- QDQAutotunerBase: Base class for custom region identification strategies

Region Management:
- Region: Hierarchical subgraph representation (nodes + children)
- RegionType: Enumeration (LEAF, COMPOSITE, ROOT)
- CombinedRegionSearch: Two-phase region discovery (partitioning + refinement)
- RegionPattern: Structural pattern analysis and matching for region grouping

Q/DQ Insertion Points:
- InsertionScheme: Collection of Q/DQ insertion points for a region pattern
- NodeInputInsertionPoint: Q/DQ insertion at specific node inputs
- ChildRegionInputInsertionPoint: Q/DQ insertion at child region input boundaries
- RegionOutputInsertionPoint: Q/DQ insertion at region output boundaries

Configuration & State:
- Config: Autotuning parameters (quant type, thresholds, verbosity)
- PatternCache: Top-performing schemes indexed by pattern (warm-start)
- PatternSchemes: Scheme collection and measurement results for a pattern

Benchmarking:
- Benchmark: Abstract base class for model benchmarking
- TensorRTPyBenchmark: Benchmark using TensorRT Python API (recommended)
- TrtExecBenchmark: Benchmark using trtexec command-line tool (legacy)

**See Also:**

- workflows.region_pattern_autotuning_workflow: Complete end-to-end optimization
- QDQAutotuner: Main autotuner class documentation
- RegionPattern: Pattern matching and signature computation
"""

# Core data structures
from .common import (
AutotunerError,
AutotunerNotInitializedError,
Config,
InsertionScheme,
InvalidSchemeError,
PatternCache,
PatternSchemes,
Region,
RegionType,
)
from .insertion_points import (
ChildRegionInputInsertionPoint,
NodeInputInsertionPoint,
RegionOutputInsertionPoint,
ResolvedInsertionPoint,
)
from .region_pattern import RegionPattern
from .region_search import CombinedRegionSearch

__all__ = [
"AutotunerError",
"AutotunerNotInitializedError",
"ChildRegionInputInsertionPoint",
"CombinedRegionSearch",
"Config",
"InsertionScheme",
"InvalidSchemeError",
"NodeInputInsertionPoint",
"PatternCache",
"PatternSchemes",
"Region",
"RegionError",
"RegionOutputInsertionPoint",
"RegionPattern",
"RegionType",
"ResolvedInsertionPoint",
]
Loading