Update python documentation - #8051
Closed
xzhu1900 wants to merge 1 commit into
Closed
Conversation
Contributor
|
Superseded by #10181 |
titaiwangms
added a commit
to titaiwangms/onnxruntime
that referenced
this pull request
Jun 9, 2026
…2 rc2 ONNX 1.22 rc2 (cherry-pick microsoft#8051) turned the silent `return;` early-exits in convTransposeShapeInference into hard fail_shape_inference(...) on output_shape / output_padding size mismatch. ORT's CPU/CUDA ConvTranspose kernel officially supports the rank+2 (full N,C,H,W) output_shape form (conv_transpose_attributes.h 241-258), which is an ORT extension beyond the ONNX spec (the schema doc excludes N,C). Under rc2 those valid ORT models that loaded under rc1 now fail at Graph::Resolve. This restores rc1 behavior via onnx.patch (Option B). Correction to the rc2 source-pin bump commit message: that change is NOT 'no valid-model behavior change' for ConvTranspose -- rc2's microsoft#8051 output_shape / output_padding strictness would regress real ORT models. This patch restores the rank+2 leniency to preserve ORT's documented kernel extension. Patch (onnx.patch + byte-identical binskim.patch), 6 sites across all 3 ConvTranspose opset variants, fail_shape_inference(...) -> `return;`: - defs.cc:1228 (output_shape) / :1240 (output_padding) [ConvTranspose-22] - old.cc:729 / :741 [convTransposeShapeInference_opset11, opset 11-21] - old.cc:3076 / :3088 [convTransposeShapeInference_opset1, opset 1-10] Reverted to `return;` (not deletion): the downstream input_shape.dim(i+2) loop runs output_shape.size() times and would OOB-read on a rank+2 shape if the guard were removed; `return;` short-circuits before the loop (memory-safe, behavior- neutral vs rc1). The new microsoft#8051 negative-value guards and all other hardening are kept. kernel_shape strictness is kept (it has no rank+2 form, so the model is genuinely invalid per spec): ConvTranspose_InvalidKernelShape now expects ONNX's rc2 shape-inference message 'Attribute kernel_shape has incorrect size' instead of the old CPU-kernel message. No other test edits are needed -- the 7 output_shape tests (incl. InvalidBiasShape_1/2) pass unedited because B restores rc1 behavior. Validation: onnxruntime_provider_test ConvTransposeTest/ConvTest/PoolTest/RoiPool -> 120 passed, 0 failed, 9 skipped (CUDA/DML EP unavailable). Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Agent-signed-off: Developer (a478a765) [claude-opus-4.8 via copilot] Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
titaiwangms
added a commit
to titaiwangms/onnxruntime
that referenced
this pull request
Jun 9, 2026
… leniency) ONNX 1.22 (rc2, cherry-pick microsoft#8051) tightened convTransposeShapeInference to fail_shape_inference when an output_shape/output_padding attribute size does not match the number of spatial dimensions. ORT historically also accepted a non-spec rank+2 (full N,C,H,W) output_shape form. This is Option A: instead of patching ONNX to restore the leniency (Option B, commit 031c777), conform ORT's own test models to the spec so the onnx.patch ConvTranspose hunks never land in main. Changes: - conv_transpose_op_test.cc: 10 output_shape attributes -> spatial-only (N,C prefix dropped; Y_shape/expected_vals unchanged). Keep B's InvalidKernelShape expected-message fix (ONNX rejects kernel_shape rank at Graph::Resolve). Add a new InferenceSession-based regression test (ConvTranspose_RankPlus2_OutputShape_DynamicRankInput_Runtime) that feeds an unknown-rank input so the kept rank+2 kernel branch stays exercised. - xnnpack_basic_test.cc: 1 output_shape attribute -> spatial-only. - conv_transpose_attributes.h: keep the rank+2 toleration (runtime-reachable for dynamic-rank inputs) and document why it is retained; no behavior change. - onnx.patch / binskim.patch: unchanged at the 3-hunk base (no ConvTranspose reverts) since this branch is built on the pre-B base. Breaking change: models specifying a rank+2 output_shape AND a statically-known input rank now fail to load under ONNX 1.22 with 'Attribute output_shape has incorrect size'. Migration: use spatial-only output_shape. See PR microsoft#28754 description. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
titaiwangms
added a commit
to titaiwangms/onnxruntime
that referenced
this pull request
Jun 15, 2026
… leniency) ONNX 1.22 (rc2, cherry-pick microsoft#8051) tightened convTransposeShapeInference to fail_shape_inference when an output_shape/output_padding attribute size does not match the number of spatial dimensions. ORT historically also accepted a non-spec rank+2 (full N,C,H,W) output_shape form. This is Option A: instead of patching ONNX to restore the leniency (Option B, commit 031c777), conform ORT's own test models to the spec so the onnx.patch ConvTranspose hunks never land in main. Changes: - conv_transpose_op_test.cc: 10 output_shape attributes -> spatial-only (N,C prefix dropped; Y_shape/expected_vals unchanged). Keep B's InvalidKernelShape expected-message fix (ONNX rejects kernel_shape rank at Graph::Resolve). Add a new InferenceSession-based regression test (ConvTranspose_RankPlus2_OutputShape_DynamicRankInput_Runtime) that feeds an unknown-rank input so the kept rank+2 kernel branch stays exercised. - xnnpack_basic_test.cc: 1 output_shape attribute -> spatial-only. - conv_transpose_attributes.h: keep the rank+2 toleration (runtime-reachable for dynamic-rank inputs) and document why it is retained; no behavior change. - onnx.patch / binskim.patch: unchanged at the 3-hunk base (no ConvTranspose reverts) since this branch is built on the pre-B base. Breaking change: models specifying a rank+2 output_shape AND a statically-known input rank now fail to load under ONNX 1.22 with 'Attribute output_shape has incorrect size'. Migration: use spatial-only output_shape. See PR microsoft#28754 description. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
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
2 participants
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: update python document for 1.8.0 release
Motivation and Context