Diagnose and fix 32-bit vs 64-bit index math overflows in PyTorch CUDA kernels.
npx skills add https://github.com/pytorch/pytorch --skill cuda-index-width
Please help me install this Agent Skill. Skill: cuda-index-width Source: https://github.com/pytorch/pytorch/tree/main/.claude/skills/cuda-index-width Command: npx skills add https://github.com/pytorch/pytorch --skill cuda-index-width
PyTorch CUDA kernels can silently overflow when tensor sizes exceed 2^31 elements, producing wrong results or memory errors. This Skill classifies the overflow site and applies the correct fix pattern without unnecessarily inflating binary size or slowing hot loops.
canUse32BitIndexMath, AT_DISPATCH_INDEX_TYPES, and CUDA_KERNEL_LOOP_TYPE instead of ad hoc checks.TORCH_CHECK failure for size-limited ops.index_t versus a single local cast, with commands to measure object/library size and symbol multiplication.@largeTensorTest and validates with compute-sanitizer.Review the failing CUDA kernel, identify the exact expression that overflows 32-bit indexing, and apply the matching fix pattern from this skill.
High-intent search queries and answers about installing and using this skill.
Identify the exact expression that exceeds 2^31, then apply one of three patterns: a local int64_t cast for setup offsets, AT_DISPATCH_INDEX_TYPES templating for hot per-element loops, or TORCH_CHECK with canUse32BitIndexMath for size-limited operators.
Use a local int64_t cast when only a base-offset multiply overflows outside the hot loop. Use templated index_t with AT_DISPATCH_INDEX_TYPES when per-element linear indexing can exceed 32 bits, since 64-bit division in hot loops is measurable.
Yes. at::native::canUse32BitIndexMath(tensor, INT_MAX) verifies both numel and the maximum storage offset, so a small numel tensor with large strided views can still require 64-bit indexing.
The kernel uses int arithmetic for an offset or loop index that overflows at INT_MAX. Either template the kernel on index_t, cast the specific overflowing expression to int64_t, or add a TORCH_CHECK guard if the operator has a documented size limit.
Each index_t dispatch duplicates the kernel for every scalar dtype and memory-format specialization. Measure impact with stat on the .cu.o file and nm -S --size-sort on libtorch_cuda.so before templating many kernels.
spacecat2002Diagnose CUDA crashes in SGLang by logging tensors at kernel API boundaries.
mazicimertReproduce bugs with failing tests and guide a minimal TDD fix workflow.
sirius-dbDetect race conditions and deadlocks in concurrent Sirius database code.
mindspore-aiOptimize backward fusion of 3D SwiGLU-like operators on Ascend hardware.