PyTorch lands XNNPACK update f6486e3e1d, fixes Vulkan conv2d tests
After three re-lands, PyTorch's XNNPACK bump finally passes CI with Vulkan fixes.
PyTorch landed PR #191206, which updates XNNPACK to GitHub revision f6486e3e1d. This is the third attempt at this change, hence the "Re-re-re-land" tag, and it finally passed CI. The PR, linked to Executorch PR #21399, also updates the Buck build configuration. A key part of the change fixes Vulkan backend tests that were failing with the XNNPACK bump, specifically the conv2d_gemm_dynamic_test. This test validates dynamic convolution operations using GEMM (general matrix multiplication) on Vulkan, a cross-platform GPU API commonly used on Android devices.
XNNPACK is a highly optimized library for neural network inference on ARM and x86 CPUs, often used in mobile and edge deployments. By updating it, PyTorch ensures better performance and correctness for models running through Executorch, its lightweight runtime for on-device AI. The fix for the Vulkan test is particularly important for developers targeting Android GPUs, as it resolves a potential regression in convolution performance. Overall, this merge tightens the integration between PyTorch's core and its mobile inference stack, promising smoother and faster AI execution on edge devices.
- PyTorch merged PR #191206, updating XNNPACK to revision f6486e3e1d after two previous landing attempts.
- The change fixes Vulkan backend tests, specifically conv2d_gemm_dynamic_test, ensuring stable conv2d operations on Android GPUs.
- Includes Buck build updates and is linked to Executorch PR #21399 for unified mobile inference improvements.
Why It Matters
Stable XNNPACK and Vulkan support mean faster, more reliable on-device AI for mobile developers.