memory-debug
Installation
SKILL.md
Memory & data-corruption debug for tt-emule
Distills the diagnostic playbook used across rounds 7-12. Targets the
failure mode "kernel runs, no crash, but output bytes don't match
input". Most fixes land in include/jit_hw/api/dataflow/,
include/jit_hw/llk_pack.h, include/jit_hw/api/compute/common.h,
or tt_metal/impl/emulation/emulated_program_runner.cpp.
For pure JIT-compile failures (missing shim), use /compute-llk-bringup.
For missing API surface, use /implement-mock.
When to invoke
- "ATOL > 1e-3" / "PCC failure" / "max diff X" on a kernel that previously worked or that runs to completion.
- "Output is mostly correct but has wrong regions" / "some rows are wrong" / "some tiles are zero".
- "Sharded → interleaved" (or vice versa) data movement is wrong.
- "Output bytes are zero" where they shouldn't be.