Diagnose TileLang GPU kernel correctness failures across configurations and shapes.
npx skills add https://github.com/tile-ai/TileOPs --skill kernel-debug
Please help me install this Agent Skill. Skill: kernel-debug Source: https://github.com/tile-ai/TileOPs/tree/main/.claude/skills/kernel-debug Command: npx skills add https://github.com/tile-ai/TileOPs --skill kernel-debug
This Skill provides a structured SOP for diagnosing correctness failures in TileLang kernels, guiding engineers through triage signals, reproducibility checks, and wiki references to reduce debugging time.
Run the TileOps kernel debugging SOP on a failing kernel to triage errors by checking configurations, examining index formulas, and clearing the JIT cache.
High-intent search queries and answers about installing and using this skill.
To debug TileLang GPU kernel failures, apply a structured SOP that runs tests across configurations, inspects index formulas and boundary handling, clears the JIT cache, and compares numerical paths like fp16 versus fp32.
An unexpected max_err on one configuration often indicates a config-level issue. You can isolate the offending path by running tests across multiple shapes and autotuning states to distinguish the root cause.
You distinguish shape-level bugs from config-level issues by applying the debugging SOP across multiple configurations and autotuning states, isolating whether the failure stems from explicit shapes or specific tuning parameters.
To diagnose TileOps kernel failures, follow the SOP steps: run tests across explicit configs, inspect index formulas and boundary handling, clear the JIT cache, and compare numerical paths such as fp16 versus fp32.
Yes, clearing the JIT cache is a required SOP step when debugging TileLang kernels. It ensures reproducibility checks are accurate by removing stale compiled artifacts before re-running tests.
Yes, comparing fp16 versus fp32 numerical paths is a core SOP step to fix TileLang kernel bugs. This comparison identifies precision-related failures and guides targeted fixes by validating against a reference.
marin-communityImplement, benchmark, and autotune TPU/GPU kernels with JAX Pallas.
mattmirelesValidate CoreML model numerical accuracy against PyTorch reference models.
tile-aiMigrate operators into TileOPs with a phased kernel-to-op workflow.
LancerLabTune and validate Python JIT GPU kernels using CuTe DSL.
