PyTorch reverts XPU sync PR to fix SYCL Graph bug on Intel Xe2 GPUs
Intel's new Xe2 GPUs expose a tricky synchronization bug in PyTorch's XPU backend
PyTorch has reverted a previous commit that introduced XPU device-wide synchronization, addressing a bug that only manifests on Intel's newest Xe2 GPU architecture (codenamed BMG). The original PR (#182630) aimed to improve synchronization for XPU devices but inadvertently caused a SYCL Graph synchronization issue on invalid queues via `ext_oneapi_wait_and_throw`.
This bug (issue #187277) went undetected because PyTorch's CI runs on Data Center GPU (Xe) rather than the newer BMG hardware. The revert (PR #187306) was approved by EikanWang and atalman, and must be cherry-picked to the release branch. This highlights the challenge of maintaining hardware-specific backends when CI coverage lags behind new architectures.
- Bug only triggers on Intel Xe2 (BMG) GPUs, not on older Xe hardware used in CI
- Root cause: ext_oneapi_wait_and_throw introduces SYCL Graph sync on invalid queue
- Revert must be cherry-picked to the release branch to fix production Xe2 deployments
Why It Matters
Ensures PyTorch stability on new Intel Xe2 GPUs, critical for AI workloads on upcoming hardware.