Ir Override Ablation

IR Override Ablation

Use this skill to turn a noisy TTGIR reproducer into a smaller, evidence-backed set of IR variants. The goal is not only to simplify the IR; it is to preserve the debugging signal and keep the test oracle honest when edits intentionally change kernel semantics.

Setup

1. Identify the target signal before editing:
   • compile failure, pass assertion, crash, hang, wrong result, or diagnostic
   • the exact command and environment that reproduces it
   • the minimum architecture requirement, for example sm90 or sm100
2. Get a baseline TTGIR:
   • Use the provided .ttgir directly when available.
   • Otherwise dump it with:

     TRITON_ALWAYS_COMPILE=1 TRITON_KERNEL_DUMP=1 TRITON_DUMP_DIR=<dump_dir> <run_command>
     

3. Decide the override mechanism:
   • Prefer triton.Config(..., ir_override="<variant.ttgir>") when the harness already uses configs/autotune or can be cheaply adapted.
   • Use TRITON_KERNEL_OVERRIDE=1 TRITON_OVERRIDE_DIR=<dump_dir> when testing against the dumped override directory layout is simpler.

Ablation Loop

1. Build or accept a list of ideas. Each idea should state the hypothesis and the expected effect on the target signal.
2. Copy the baseline TTGIR to a named variant file before editing.
3. Classify the test oracle before running the variant:
   • preserve: the original numeric/reference output must still match.
   • replace: the edit intentionally changes output values; update the expected output to match the new semantics.
   • relax: success is reproducing a compile failure, crash, hang, or diagnostic; numeric equality is not the primary oracle.
   • observe: temporary exploration only; do not treat this as final evidence.
4. Update the harness when the oracle is not preserve.
   • Example: if a store operand is replaced by a constant tensor, expected output must become that constant over the same mask/shape, not the original reference computation.
   • If an edit removes a store, expected output may need to assert unchanged sentinel values instead of comparing to the old result.
5. Run exactly the command recorded for the variant with TRITON_ALWAYS_COMPILE=1 to avoid cache reuse.
6. Record whether the target signal survives, weakens, disappears, or changes into a different failure.
7. Prefer single-concept edits. When multiple edits are needed, stage them as separate variants so the final reproducer has a defensible chain of evidence.

Harness Rules

• Keep baseline and variant runs side by side until the final reproducer is chosen.
• Make oracle-changing behavior explicit in code or comments near the assertion.
• For wrong-result bugs, keep at least one shape that fails and one nearby shape that passes when practical.
• For hangs, set a short external timeout and run third_party/tlx/killgpu.sh if the GPU process survives longer than expected.
• Do not run performance benchmarks unless the user explicitly asks.

Reporting

Return a compact result matrix:

variant | hypothesis | oracle | result | target signal | kept?

Then state the smallest surviving reproducer, the concepts still present in the IR, and the concepts successfully removed.

References

• Use references/variant-manifest-template.md when tracking more than a few ideas or when handing work to another agent.
• Use references/harness-template.md when creating or adapting the Python harness around ir_override.