Add LabelEncoder CUDA execution provider for numeric types#28045
Merged
Add LabelEncoder CUDA execution provider for numeric types#28045
Conversation
Implements LabelEncoder for the CUDA execution provider supporting numeric types (int64, float, double). Uses sorted arrays and binary search on GPU for efficient O(log n) per-element lookup. Supports: - Opset 2-3: int64↔float, int64↔int64, float↔float - Opset 4+: above plus double↔double, double↔int64, int64↔double String types remain CPU-only as they cannot run on GPU. Agent-Logs-Url: https://github.com/microsoft/onnxruntime/sessions/d17c0a15-3bf2-4ac4-bc57-255876153271 Co-authored-by: tianleiwu <30328909+tianleiwu@users.noreply.github.com>
Add tests for: - Float NaN keys to int64 values (opset 4) - Float NaN keys to float values (opset 4) - Double NaN keys to int64 values (opset 4) - Int64 to double conversion (opset 4) - Double to double conversion (opset 4) These tests exercise the CUDA binary search with NaN handling and double type support. Agent-Logs-Url: https://github.com/microsoft/onnxruntime/sessions/d17c0a15-3bf2-4ac4-bc57-255876153271 Co-authored-by: tianleiwu <30328909+tianleiwu@users.noreply.github.com>
… test intent Agent-Logs-Url: https://github.com/microsoft/onnxruntime/sessions/d17c0a15-3bf2-4ac4-bc57-255876153271 Co-authored-by: tianleiwu <30328909+tianleiwu@users.noreply.github.com>
Copilot
AI
changed the title
[WIP] Add LabelEncoder support for CUDA provider
Add LabelEncoder CUDA execution provider for numeric types
Apr 13, 2026
tianleiwu
reviewed
Apr 19, 2026
Contributor
tianleiwu
left a comment
tianleiwu
left a comment
There was a problem hiding this comment.
Review Summary
This PR adds the first-ever ML domain (ai.onnx.ml) kernel registration for the CUDA EP, implementing LabelEncoder for opset 2-3 and 4+ with numeric types only (int64, float, double). The implementation uses a GPU-appropriate sorted-array + binary search approach instead of hash maps, with proper NaN handling. The code is well-structured, closely mirrors CPU patterns, and all CI checks pass.
Overall: Implementation is correct and well-structured. The concerns below are minor (style/convention adherence and test coverage) and do not affect correctness or safety.
Findings Summary
| # | Severity | Component | Issue |
|---|---|---|---|
| 1 | Nitpick | label_encoder.cc |
Unused #include <filesystem> |
| 2 | Suggestion | label_encoder.cc CopyToGpu |
Size arithmetic should use SafeInt per project conventions |
| 3 | Suggestion | label_encoder_test.cc |
No empty-input-tensor edge case test |
…y-tensor tests - Remove unused #include <filesystem> - Use SafeInt<size_t> for cudaMemcpy size arithmetic in CopyToGpu - Add empty input tensor tests for opset 2 and opset 4 LabelEncoder
Contributor
There was a problem hiding this comment.
Pull request overview
Adds a CUDA Execution Provider implementation of ai.onnx.ml.LabelEncoder for numeric key/value types, aiming to avoid CPU round-trips for large label maps.
Changes:
- Introduces CUDA kernel + host-side op implementations for LabelEncoder (opset 2–3 and 4+ numeric variants).
- Registers the new ML-domain CUDA kernels and adds the ML domain constant for shared-library/provider API usage.
- Expands existing LabelEncoder tests (NaN handling, double variants, empty-input) and updates CUDA operator kernel documentation.
Reviewed changes
Copilot reviewed 8 out of 8 changed files in this pull request and generated 2 comments.
Show a summary per file
| File | Description |
|---|---|
| onnxruntime/core/providers/cuda/ml/label_encoder_impl.h | Declares CUDA kernel launcher for sorted-key binary-search LabelEncoder. |
| onnxruntime/core/providers/cuda/ml/label_encoder_impl.cu | Implements per-element binary search kernel with NaN short-circuit for float/double keys. |
| onnxruntime/core/providers/cuda/ml/label_encoder.h | Declares CUDA LabelEncoder kernel classes for opset 2–3 and opset 4+. |
| onnxruntime/core/providers/cuda/ml/label_encoder.cc | Implements attribute loading (list/tensor), key sorting, GPU copy, and kernel registrations for supported type pairs. |
| onnxruntime/core/providers/cuda/cuda_execution_provider.cc | Adds kernel-class forward decls + registration routine for ML-domain LabelEncoder variants and wires registration into registry init. |
| onnxruntime/core/providers/shared_library/provider_api.h | Adds missing kMLDomain constant for shared-library/provider builds. |
| onnxruntime/test/providers/cpu/ml/label_encoder_test.cc | Adds numeric NaN-key tests, double type-combination tests, and empty-input tests (intended to exercise CUDA EP when available). |
| docs/OperatorKernels.md | Documents CUDA provider support for ai.onnx.ml::LabelEncoder. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
…tics SortKeysValues() now uses std::stable_sort and deduplicates sorted keys, keeping only the first occurrence of each key (including NaN). This matches the CPU LabelEncoder's map_.emplace() first-occurrence-wins behavior. num_keys_ is now set after SortKeysValues since dedup may shrink the arrays.
Contributor
|
/azp run Win_TRT_Minimal_CUDA_Test_CI, Windows GPU Doc Gen CI Pipeline |
|
Azure Pipelines successfully started running 2 pipeline(s). |
Contributor
|
/azp run Win_TRT_Minimal_CUDA_Test_CI, Windows GPU Doc Gen CI Pipeline |
|
Azure Pipelines successfully started running 2 pipeline(s). |
kunal-vaishnavi
previously approved these changes
Apr 21, 2026
Add the CUDA LabelEncoder sources to the minimal provider build so Windows TRT minimal CI links the new kernel registrations. Also factor the non-plugin TensorProto construction into a helper so the shared-provider path only needs one conditional block.
Move the default_tensor read path in LabelEncoder behind a helper so the plugin and shared-provider TensorProto handling stays in one place. This keeps GetDefaultValue focused on fallback selection instead of attribute transport details.
kunal-vaishnavi
approved these changes
Apr 22, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Description
Implements
ai.onnx.ml.LabelEncoderon the CUDA execution provider for numeric key/value types using sorted arrays + binary search (O(log n) per element).New files (
onnxruntime/core/providers/cuda/ml/):label_encoder_impl.cu/.h— CUDA kernel: per-thread binary search on sorted keys, NaN-aware for float/doublelabel_encoder.cc/.h— Host-side op classes (CudaLabelEncoderfor opset 2-3,CudaLabelEncoder_4for opset 4+). Constructor sorts keys, copies to GPU;ComputeInternallaunches kernel.Modified files:
cuda_execution_provider.cc— Register 11 kernel variants (4 versioned opset 2-3, 7 opset 4+)provider_api.h— Add missingkMLDomainconstant (first ML-domain op on CUDA EP)docs/OperatorKernels.md— Addai.onnx.mlsection to CUDA provider tableSupported type combinations:
int64↔float,int64↔int64,float↔floatdouble↔double,double↔int64,int64↔doubleString types remain CPU-only. NaN keys are placed at end of sorted array and short-circuited before binary search.
Tests: 5 new test cases covering NaN-key-to-numeric-value mappings and double type combinations. Existing numeric tests (
FloatToInt64Opset2,Int64ToFloatOpset2, etc.) will automatically run on CUDA viaOpTester::Run().Motivation and Context
Models with large LabelEncoder nodes (>100k entries) force a CPU round-trip when all other nodes run on GPU. This adds the CUDA implementation to eliminate that data transfer bottleneck.