Delivered Triton Debugging Metadata Enhancement for facebookexperimental/triton. Lowered name and type information into LLVM debug metadata to improve debugging with cuda-gdb/rocm-gdb, and added MLIR/LLVM passes to extract variable information and integrate with existing debug workflows. Added tests (Lit, C++, Python) and validated with pre-commit checks. Cherry-picked from Triton OAI (commit fe070acfff668e2fa4a7f29ea10b230e3d674d39) and aligned for integration with PR #7521. These changes reduce debugging friction for GPU kernels and streamline root-cause analysis in production-like environments.

September 2025 monthly summary for the Intel XPU backend for Triton focused on advancing debuggability and traceability of GPU-compiled Triton programs. Implemented LLVM debug metadata support to preserve source-level debug names and types in GPU artifacts (cubin/hsaco), enabling the emission of full debug information via LLVM_EXTRACT_DI_LOCAL_VARIABLES. This allows developers to evaluate values directly in GPU debuggers (cuda-gdb, rocm-gdb) and preserves source-level naming and type information throughout the Triton pipeline. The work was committed as 799d84673a44ba3d417877c9ab6a35f30ea2bdbd with reference to #7633. Impact: Improves debugging efficiency and accuracy for the Intel XPU backend in Triton, enabling faster root-cause analysis and better developer experience when diagnosing GPU-side issues. Continues to strengthen the backend’s maintainability and traceability across GPU artifacts. Overall accomplishments: Enhanced GPU artifact debug metadata pipeline, enabling end-to-end source-level visibility for Triton programs on Intel GPUs; lays groundwork for more robust GPU debugging workflows and cross-layer diagnostics. Technologies/skills demonstrated: LLVM debug metadata, cubin/hsaco generation, LLVM_EXTRACT_DI_LOCAL_VARIABLES, GPU debugging workflows (cuda-gdb/rocm-gdb), Triton backend integration, source-level mapping of debug information.