Validate FeatureVectorizer batch sizes

[Akshay Sonawane (apsonawane)]

This pull request improves the robustness of the FeatureVectorizer operator by validating runtime input shapes before allocating or writing the output.

FeatureVectorizer input validation:

• Enforces that all input tensors have the same batch size.
• Rejects scalar (0-D) inputs with a clean Status error instead of indexing an empty shape.
• Returns validation failures as Status values so exception-disabled/minimal builds do not abort the host process.

Behavior change:

Rank-1 inputs are treated as a single-row batch, as required by the ai.onnx.ml specification. Mixing a rank-1 input with a rank-2 input whose batch size is greater than one now fails instead of silently leaving additional output rows zero-filled.

Unit test improvements:

• Verifies mismatched batch sizes are rejected.
• Verifies scalar input is rejected for both the first input and a later variadic input.

[Copilot]

Pull request overview

This PR strengthens the CPU ML FeatureVectorizer kernel by validating that all variadic inputs share a consistent batch size (first dimension), and adds a unit test that exercises the new failure path.

Changes:

• Added runtime validation in FeatureVectorizer::Compute to reject mismatched batch sizes across inputs.
• Added a unit test RejectMismatchedBatchSizes to ensure mismatched batches fail with an appropriate error.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
onnxruntime/core/providers/cpu/ml/feature_vectorizer.cc	Adds per-input batch-size validation in the kernel implementation.
onnxruntime/test/providers/cpu/ml/feature_vectorizer_test.cc	Adds a regression test covering mismatched batch sizes.

[Tianlei Wu (tianleiwu)]

Summary

The validation is placed correctly — it runs before context->Output(0, {N, total_dimensions_}), so a bad input never allocates an output or writes partial rows. The rank check is also a genuine bug fix: a 0-D input previously reached x_dims[0] / input_dims[0] on an empty dims span, which is an out-of-bounds read rather than a clean error. And input_dims.size() == 1 ? 1 : input_dims[0] matches both the ai.onnx.ml spec ("1-D tensors are treated as [1,C]") and the row-count logic already in VectorizeTensor, so validation and execution agree on what "batch" means.

One robustness point and a couple of nits inline. Verdict: COMMENT.

Cross-cutting notes

Behavior change worth documenting. Previously a rank-1 input mixed with a rank-2 input of batch > 1 was silently accepted: VectorizeTensor copied a single row and the remaining output rows stayed zero-filled. Those models now hard-fail. This is spec-conformant and the old output was almost certainly wrong, but it is still a behavior change for previously loadable models — worth calling out in the PR description.

Description drift. The description mentions one new test; the diff adds two (RejectMismatchedBatchSizes, RejectScalarInput) plus a minimum-rank check that is not described.

Existing coverage is unaffected. BasicFunctionality, HandleInputDimensionMismatch, Batch, and BatchWith3DInput all use a consistent batch across inputs, so they remain valid under the stricter rule.

[Tianlei Wu (tianleiwu)]

Re-reviewed the current head after the follow-up fixes. The prior concerns are addressed: rank and batch validation now return Status instead of aborting exception-disabled builds, missing inputs have useful diagnostics, and scalar input 0 has dedicated coverage. Validation still completes before output allocation, and the batch interpretation matches VectorizeTensor. Relevant CPU, minimal-build, and lint checks are green; the two failing CUDA plugin checks are both the unrelated GatherBlockQuantized plugin test.