Repository navigation
Extend ONNX runtime operator support - #8
Merged
Merged
Conversation
nvtw
marked this pull request as draft
August 24, 2026 11:30
nvtw
force-pushed
the
dev/tw/extend_onnx_runtime
branch
from
August 24, 2026 11:47
9ff9ee9 to
fbaf25c
Compare
nvtw
marked this pull request as ready for review
August 24, 2026 11:53
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
This MR extends the lightweight ONNX policy runtime with Add, Sub, Mul, Div, BatchNormalization, Relu, Tanh, Sqrt, and ReduceMean. It also uses Warp-NN's tiled GEMM for inference while preserving the differentiable GEMM path.
Common Gemm/BatchNormalization/Relu and RMS-normalization sequences are fused for efficient, graph-capturable execution. Fusion is limited to supported operator variants and static scale parameters, while dynamic scaling remains on the regular operator path.
Tests cover numerical correctness, fused and unfused execution, dynamic RMS scaling, unsupported variants, and deterministic CUDA graph replay. All 9 focused CUDA tests and pre-commit hooks pass.