onnxruntime

c1722108 - webgpu: Skip MatMulNBitsMlpFusion when kernel unavailable (#29089)

webgpu: Skip MatMulNBitsMlpFusion when kernel unavailable (#29089) ## Summary - Add kernel availability check at transformer registration time - Query WebGPU EP kernel registry for MatMulNBitsMlp before enabling fusion - Skip fusion gracefully when kernel is not registered - Add unit test for kernel-unavailable case ## Motivation MatMulNBitsMlpFusion is a WebGPU-specific optimization that fuses gated MLP patterns into a single MatMulNBitsMlp kernel operation. However, it was previously registered unconditionally without verifying that the MatMulNBitsMlp kernel actually exists at runtime. This created a mismatch scenario: **The fusion optimizer could be compiled into ORT, but the kernel might not be registered on the WebGPU EP**, leading to "kernel not found" failures at inference time. ### Realistic Scenario: Version Mismatch Consider a common deployment scenario: - **ORT** is compiled with MatMulNBitsMlpFusion and MatMulNBitsMlp kernel support in WebGPU - **WebGPU plugin** (deployed separately) is an older version that predates MatMulNBitsMlp kernel implementation - At inference time: - The ORT optimizer successfully runs MatMulNBitsMlpFusion and creates fused nodes - The WebGPU plugin loads, but its kernel registry does not have MatMulNBitsMlp - Inference fails: "kernel 'MatMulNBitsMlp' not found on WebGpuExecutionProvider" This can occur when optimizer versions and EP plugin versions diverge in production deployments. ### The Fix Rather than unconditionally assuming the kernel exists, the fix queries the WebGPU EP's kernel registry at graph optimization time (session initialization). If the kernel is not found, fusion is skipped entirely, preventing "kernel not found" errors and allowing graceful fallback to unfused (but functional) operators. ## Test plan - [x] Build ORT with WebGPU EP succeeds - [x] Model loads and infers correctly (phi4-graph-prune verified) - [x] Fusion enabled when kernel available (40 MatMulNBitsMlp nodes fused) - [x] New test \MatMulNBitsMlpFusionSkipsWhenKernelUnavailable\ verifies graceful skip when kernel unavailable - [x] Lintrunner reports no issues