Summary

Fix for ZhaoqiongZ/torch-xpu-ops-exp#346

Issue: [upstream pytorch#3450] [ai_generated] XPU crash on torch.linalg.ldl_solve with invalid pivots instead of raising validation error

Root Cause: torch.linalg.ldl_solve falls back to CPU on XPU via XPUFallback.template. The upstream fix (pytorch#181032) added pivot validation (1 <= |pivot| <= N) inside ldl_solve_kernel in BatchLinearAlgebraKernel.cpp, which only runs on the CPU dispatch path. When XPU is used, the fallback occurs after the IMPL function (linalg_ldl_solve_out in BatchLinearAlgebra.cpp) copies B to result and calls ldl_solve_stub with the XPU device type — but the XPU fallback dispatches down to the CPU kernel without going through the validated path first, triggering a segfault in LAPACK SYTRS before validation runs.

Failed Tests:

• No specific test method identified; reproducer is a standalone script targeting torch.linalg.ldl_solve on XPU with out-of-range pivots.

Diff stat:

aten/src/ATen/native/BatchLinearAlgebra.cpp       | 10 ++++++++++
 aten/src/ATen/native/BatchLinearAlgebraKernel.cpp | 10 ----------
 2 files changed, 10 insertions(+), 10 deletions(-)