Debug AOTInductor errors, segfaults, and CUDA illegal memory access issues.
npx skills add https://github.com/Billkst/Causal-TabDiff --skill aoti-debug
Please help me install this Agent Skill. Skill: aoti-debug Source: https://github.com/Billkst/Causal-TabDiff/tree/main/.agents/skills/aoti-debug Command: npx skills add https://github.com/Billkst/Causal-TabDiff --skill aoti-debug
This Skill helps diagnose and resolve errors and crashes encountered when using AOTInductor (AOTI), a compilation toolchain for PyTorch.
aot_load() fails with a segfault or runtime errors occur during model inference due to device mismatches, this skill provides the necessary steps to identify and correct the configuration.Use the aoti-debug skill to help resolve a segfault encountered during aot_load().
High-intent search queries and answers about installing and using this skill.
A segfault during aot_load often stems from device and shape mismatches between AOTInductor compilation and runtime inputs. You must systematically verify that runtime tensor configurations match the compiled model's expected inputs.
Debug a CUDA illegal memory access (IMA) by enabling specific debugging flags and environment variables. This approach systematically pinpoints the exact memory access violation occurring during AOTI model inference.
Runtime input device mismatches cause crashes because the loaded model expects tensors on a specific device. You must ensure the runtime device type exactly matches the device configuration used during the AOTI compilation phase.
Constant loading failures in PyTorch AOTInductor typically arise from incorrect compilation configurations. Diagnosing this requires verifying that all constant parameters are correctly aligned with the runtime environment during the aot_load process.
Debugging AOTInductor crashes without matching input shapes is severely limited because shape mismatches directly cause runtime errors and segfaults. You must verify shape consistency between compilation and runtime to resolve these crashes.
SnakeORefactor PyTorch code with modular nn.Module, torch.compile, AMP, and optimized DataLoader.
stelemanDiagnose and fix AOTInductor errors with structured device and shape validation.
mlc-aiAdd memory profiling instrumentation to distributed MoE training workflows.
tenstorrentConvert PyTorch tensors and execute tensor operations on Tenstorrent MeshDevice topologies.