Repository navigation
[upstream_ut] test_cdist_backward_large_index GPU segfault: int32 overflow in XPU cdist backward buffer offset #5660
Description
🐛 Describe the bug
test_cdist_backward_large_index (test/test_torch.py, added upstream alongside the CUDA fix pytorch/pytorch#128791) causes a GPU segfault that aborts the process on XPU for both parametrizations:
test_cdist_backward_large_index_b_2_r_1024_m_2048_xputest_cdist_backward_large_index_b_32_r_8192_m_1_xpu
import torch
b, r, m = 2, 1024, 2048 # also fails with 32, 8192, 1
dev = "xpu"
x1 = torch.randn(b, r, m, device=dev, requires_grad=True)
x2 = torch.randn(b, r, m, device=dev)
grad = torch.randint(-4, 5, (b, r, r), device=dev, dtype=torch.float)
(actual,) = torch.autograd.grad(torch.cdist(x1, x2, p=1), x1, grad)Segmentation fault from GPU at 0xff0000020b800000, ctx_id: 1 (CCS) type: 0 (NotPresent), level: 2 (PDP), access: 1 (Write), banned: 1, aborting.
Abort was called at 288 line in file:
./shared/source/os_interface/linux/drm_neo.cpp
Root cause: in src/ATen/native/xpu/sycl/DistanceKernels.cpp, cdist_backward_kernel_impl_func:
const int l_size = r_size * m; // int32
...
scalar_t* buff_i = buff_data + l * l_size + (r1 * j + i) * m + init;l_size and l are declared int. For (b=2, r=1024, m=2048), r_size = r1*r2 = 2^20, so l_size = r_size*m = 2^31 overflows int32, and l * l_size wraps, producing an out-of-bounds GPU write. The (32, 8192, 1) case overflows dist.numel() == 2^31 similarly. This is the same int32 indexing bug fixed for CUDA in pytorch/pytorch#128791.
Fix: widen l_size (and l/k/y) to int64_t, mirroring the CUDA fix. (r1 * j + i) * m is already int64.
Because the failure is a process-aborting GPU segfault (not a catchable exception), the upstream test is skipped on XPU via @skipIfXpu referencing this issue.
Versions
torch 2.15.0.dev20260923+xpu
GPU: Intel(R) Data Center GPU Max 1550
Activity
daisyden commentedon Nov 10, 2026
label-issue: #5660
Root cause: int32
l = y / r_size,l_size = r_size * moverflow incdist_backward_kernel_impl_func(src/ATen/native/xpu/sycl/DistanceKernels.cpp:468,471,484), giving an OOB GPU write; owner torch-xpu-ops (CUDA fixed in pytorch#128791).
typeBugissue_typeempty; GPU segfault aborts the process.testtest: uttest_cdist_backward_large_indexin test/test_torch.py; no e2e/oob signal.modulemodule: opspriorityMediumfunctionalitytrainingtorch.autograd.gradof cdist).Test cases (2):
Applied to this issue. Generated by a local
label-issuerun.Use 64-bit indices in cdist backward kernel to fix large-input OOB write