You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[Feature][Windows] Enable the static XPU Triton launcher (zebin kernel format + StaticXpuLauncher) on Windows #6167
Feature: make Inductor's static Triton launcher for XPU work on Windows, so torch._C._StaticXpuLauncher is available and CachingAutotuner.can_statically_launch succeeds for XPU kernels.
Motivation: today it is unavailable on Windows for two reasons:
Root dependency: the static launcher requires Triton to compile kernels to zebin, which depends on ocloc. Windows does not have ocloc installed by default.
torch/_inductor/utils.py:496: XPU_KERNEL_FORMAT = "spv" if _IS_WINDOWS else env("zebin"). With spv, can_statically_launch (runtime/triton_heuristics.py:~3123) raises CannotStaticallyLaunchKernel ("Static XPU ... does not support SPIR-V").
torch/csrc/Module.cpp:~2631 guards StaticXpuLauncher_init with #if defined(USE_XPU) && !defined(_WIN32) (added in pytorch #168952), so even a zebin Windows build has no C++ launcher (hasattr(torch._C, "_StaticXpuLauncher") is False).
Kernels silently fall back to TritonCompileResult, which has no recheck_autotune_cache. Consequences:
Activity