[Tianlei Wu (tianleiwu)]

Summary

Allow an empty optional split input edge with the opset-18 num_outputs attribute, checking whether the second input actually exists rather than rejecting any two-slot input list. Also correct the shared native split-size calculation used by CPU fallback so valid empty final outputs work and large dimensions do not lose integer precision.

Key changes

• Preserve optional-input handling across the constructor and provider paths.
• Use integer ceiling division and compute the final chunk from the remaining dimension, including [1, 0] and [2, 2, 2, 2, 2, 0].
• Validate positive output counts, attribute/output-count agreement and non-negative final chunks.
• Add native coverage for zero final outputs, one-element inputs, zero split dimensions and dimension 16777217, using both omitted-optional-input encodings.
• Retain the existing browser regression fixtures and their exact output assertions.

Tests

• Freshly rebuilt common production code, Python binding and actual operator-test binary: all 37 SplitOperatorTest cases pass with no skips, including all four new regressions and existing invalid-input/count/axis cases.
• All 17 additional native graph/transpose/quantization Split tests pass.
• The hash-verified fresh CPU binding matches ONNX reference shapes and values in 200 executions, including 58 zero-final-output cases, both absent-input forms and dimensions above float integer precision.
• Scoped repository linters pass.
• Earlier constructor-change validation included a 13-test CUDA/CPU focused run on A100 and compilation for CPU, built-in CUDA and the CUDA plugin. The new common arithmetic repair is locally validated on CPU; exact-new-head browser/CUDA integration CI remains the final verification gate.

Follow-up to #32970.

[Copilot]

Copilot review overview

🟡 Changes recommended

The JS Split constructor still treats the empty edge as present and rejects the valid model.

Review effort: Balanced
Findings: 1

Open (1)

• JS Split fails to derive sizes for empty optional input · New

What changed in this PR

Fixes Split-18 handling of an omitted optional split edge when num_outputs is specified.

Changes:

• Detects actual input-edge presence across built-in and plugin kernels.
• Adds regression coverage for an explicit empty optional edge.

File	Description
onnxruntime/​core/​providers/​cpu/​tensor/​split.h	Distinguishes empty optional edges from provided inputs.
onnxruntime/​test/​providers/​cpu/​tensor/​split_op_test.cc	Tests even splitting with an empty optional edge.

---

💡 Configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Akshay Sonawane (apsonawane)]

Tianlei Wu (@tianleiwu) 1 webgpu test is failing can you fix that. Otherwise, the PR looks good

[Tianlei Wu (tianleiwu)]

Fixed the required Web CI failures in 379afbd. Split now allocates zero-sized outputs outside the JSEP shader and binds only non-empty chunks, preserving their original output indices (the native WebGPU implementation already uses this approach). Validation: 11 JSEP Split runtime-shape tests passed, Web TypeScript checks passed with current common declarations, and targeted ESLint passed. Merged origin/main; new-head CI is running.

[Tianlei Wu (tianleiwu)]

Maintenance update at 3143158367232f22f681ca39dae872833b5d5102: the main refresh, optional-input handling and Web regressions are retained. The shared CPU/CUDA partition repair now has actual production and plugin proof.

Integer ceiling division and checked final-tail calculation replace floating-point sizing, preserve valid zero-tail outputs, reject negative final sizes and enforce num_outputs/output-count agreement. The protected inline helper is available to CPU and CUDA/plugin preparation without changing provider ABI/layout or CUDA launch, pinned-buffer or CUDA Graph behavior.

The fresh exact314 production CPU/CUDA build passed all 41 SplitOperatorTest cases, including all four original CI-red Split18 cases, the 33-output CUDA case and dimension-product regressions, with no failures/skips. The staged Python binding passed 12 cases without failures/skips, covering both omitted-input encodings and exact count diagnostics; ten profiles prove CUDA-assigned Split with CPU fallback and graph optimization disabled. Exact empty shapes include [0,8388609], [0,8388608] and the valid outer-axis final [0,2] output.

Actual current-head Linux CUDA plugin 42607563127 completed SUCCESS. Its separate real test job129155373975 executed all41 Split cases with 41 OK and zero failed/skipped, plus the broader native provider suite. The plugin Python step ran94 tests with2 environmental skips and overall OK; this is not reported as zero-skip targeted Python proof. Generic ONNX tests in the build wrapper were intentionally skipped; the separate plugin native/Python executions are the evidence, not the noisy update-stage or missing-wheel annotation.

Actual current-head Windows CUDA plugin 42607563171 also completed SUCCESS: real build129127297227 and test129150805592 succeeded, including Run Tests and Run CUDA Plugin EP Python Tests. The CodeQL Java failure was a Gradle8.7 connection timeout before compilation; only failed jobs were retried. CodeQL42607563106 attempt2 and fresh Java job129139564483 Autobuild/Perform CodeQL Analysis both completed SUCCESS.

Actual current-head Linux CUDA42607561560 completed SUCCESS. Its separate real test job129150181469 executed exactly41 SplitOperatorTest cases, with all41 OK and no failures/skips. All three normal-CUDA/Linux-plugin/Windows-plugin hard gates now have actual executed-test proof.

The subsequent Web Debug failure first occurred at ScatterND add int32: GPUBuffer.mapAsync reported a lost device, followed by WASM-abort and stale-session-creation cascades. The relevant native-WebGPU/ScatterND surfaces are unchanged from main; no Split source defect was identified from this log. After guarding the exact current314 and completed provider, a failed-only recovery was requested. Actual attempt2 Debug129178170394 completed SUCCESS on unchanged code: the original failing case passes in all six executed modes, and the formerly failing native-WebGPU GPU-location suite completes1974 tests successfully. Successful sibling start/end timestamps remain from the original attempt despite new provider record IDs; they were preserved.

The fresh complete all-20 audit at 2026-11-09T13:46:34Z verifies this exact current314 head open, non-draft, mergeable, with zero unresolved threads and all current-head checks green. This PR's CI preparation is complete; normal reviewer approval remains external. No test, tolerance, provider fallback or assertion was weakened.