onnxruntime

dc894603 - [WebGPU] Handle empty-rank MatMul and DequantizeLinear correctly (#33171)

[WebGPU] Handle empty-rank MatMul and DequantizeLinear correctly (#33171) ### Change and motivation Models with optional quantized LoRA adapters can represent an inactive adapter with zero-rank tensors. These valid inputs produce empty dequantization results and MatMuls with a zero inner dimension. WebGPU must handle them without binding null input buffers or issuing zero-sized dispatches. The adapter tensors are overridable initializer-backed inputs, so their zero-rank defaults cannot be treated as permanently constant when adapters can be activated or removed at runtime. This change returns correctly shaped empty outputs without GPU work and initializes nonempty zero-K MatMul outputs with zero, bias and any applicable activation. Empty weight transposes used by Conv are handled before dispatch. DequantizeLinear validates scale and zero-point shapes before returning for empty input, with safe handling of zero dimensions during block-size inference. The MatMul shape helper checks leading batch dimensions directly when selecting its flattening shortcut, preserving batch broadcasting when K is zero. Empty-input handling precedes the current MatMul execution planner, while explicitly forced algorithms still undergo zero-K prerequisite validation. General batch broadcasting preserves zero dimensions using the same rule as the common broadcast helper; offset generation reuses the validated output batch dimensions. Regressions cover zero batches on either operand, eager and deferred offset generation, and rejection of incompatible zero/non-unit batches. The CPU FP32 and FP64 zero-K kernels initialize the complete broadcast output tensor, not just one M-by-N slice. CPU-only regressions poison preallocated outputs and verify every element across repeated runs: 120 elements for broadcast MatMul and six elements for FusedMatMul with transposed A=[2,0,3] and vector B=[0], in both FP32 and FP64. Regression tests cover FP32/FP16, vectors, broadcasting, empty outputs, Conv bias/activation and malformed quantization inputs. Capture tests retain stable GPU-bound tensors, assert capture, and reset output buffers between runs to verify that replay computes the expected result. CPU fallback is disabled in GPU regressions. ### Validation - Native ORT 1.30 builds and targeted source checks passed on Windows, including contrib-enabled and contrib-disabled configurations; repository-pinned lint passed. The shared MatMul helper also passes 32-bit x86 compile checks with and without NDEBUG and with implicit 64-to-32 narrowing treated as an error; a static assertion verifies 32-bit ptrdiff_t. - Actual compact adapters passed base -> rewrite -> webpage summary -> base on native CPU FP32 and an RTX 4070 browser FP16 run with full WebGPU validation and capture enabled. Repeated responses matched within each backend, and all eight replayed CPU responses matched the GPU responses exactly. - CPU prefill/decode logits matched the reference exactly for base, both active adapters and restored base states. - Native CPU FP32 broadcast and zero-batch regressions passed twice per shape, including the sentinel-filled 120-element output and zero batches with a nonzero inner dimension. Affected CPU/WebGPU sources and tests compile with and without NDEBUG. - [macOS ARM64 CPU Debug CI](https://github.com/microsoft/onnxruntime/actions/runs/37890352409/job/113689652274) executed and passed the FP64 broadcast/zero-batch regressions and the shared-helper eager/deferred-offset and incompatible-batch tests on correctness commit 938581b989. - [macOS ARM64 WebGPU Release CI](https://github.com/microsoft/onnxruntime/actions/runs/37890352409/job/113689652248) executed and passed the FP32/FP16 zero-K matrix/batch, vector/scalar and empty-output tests on that correctness commit. The shared capture-test helper now resolves its graph using the normal OpTester resolve options before serialization. Both consumers compile with and without NDEBUG, and pinned lint passes; the corrected capture setup has also passed targeted Windows Debug execution as described below. - The expanded native Windows Debug suite, with the full CPU registry and bundled Dawn/D3D12, executed all 20 focused regressions in 13.131 seconds, including the FP32/FP64 transposed zero-K right-vector cases added in 3b55d1471b. Coverage also includes shared-helper offsets, WebGPU FP32/FP16 zero-K/empty outputs, graph capture/replay, Conv bias/activation/empty outputs, and empty DequantizeLinear validation. All 20 passed; none were skipped or missing. Capture assertions, stable bindings, replay sentinels and disabled CPU fallback remained enabled. - After merging current main and resolving the MatMul dispatcher/test conflicts, the rebuilt native Windows Debug runner passed all 20 focused regressions without skips, plus 10 additional upstream MatMul execution/algorithm tests. Six upstream subgroup, subgroup-matrix and split-K tests were skipped because the local adapter lacks their required features/configuration; these remain covered by CI. Forced naive zero-K execution, forced packed zero-K rejection, and dynamic-shape algorithm reselection passed locally. Fresh cross-platform CI, including the corrected WebGPU capture regressions, must pass before merge. --------- Co-authored-by: Mustapha Jaber <mustjab+microsoft@microsoft.com> Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>