slowlyc/agent-gpu-skills/triton-skill

Triton and Gluon development

Use the local Triton checkout as the primary source for APIs and implementation patterns. Prefer current tutorials and source over remembered signatures because Triton and Gluon evolve quickly.

Locate the checkout

Resolve the directory containing this SKILL.md, then use its repos/triton/ child. The installer links that path to agent-gpu-skills/third_party/triton/ or to the checkout supplied through TRITON_REPO.

In commands below, replace TRITON_REPO with the resolved absolute path:

TRITON_REPO=/absolute/path/to/triton-skill/repos/triton

If the checkout is missing, run this from the agent-gpu-skills repository and reinstall the Skill:

bash update-repos.sh triton
bash install.sh --skill triton-skill

Choose the source surface

Read quick-reference.md when choosing a tutorial, a complete example, or a production-kernel implementation.

Query workflow

1. Identify whether the task is Triton, Gluon, or compiler internals.
2. Find the closest current tutorial or implementation for the operation and architecture.
3. Verify every API used against its definition or another current call site.
4. Preserve the target workload's shape, dtype, stride, layout, masking, and numerical contract.
5. Establish correctness before changing launch parameters or optimization strategy.

Discover current examples before relying on a remembered filename:

find "$TRITON_REPO/python/tutorials" -maxdepth 2 -type f | sort
find "$TRITON_REPO/python/examples" -type f | sort

Query Triton language usage and definitions:

rg -n 'tl\.dot|tl\.dot_scaled' "$TRITON_REPO/python/tutorials"
rg -n '@triton\.autotune' "$TRITON_REPO/python/tutorials"
rg -n '^def (load|store|dot|dot_scaled)' \
  "$TRITON_REPO/python/triton/language"

Query Gluon architecture patterns:

rg -n '@gluon\.jit' "$TRITON_REPO/python/tutorials/gluon"
rg -n 'wgmma|tcgen05|mbarrier|tma' \
  "$TRITON_REPO/python/tutorials/gluon" \
  "$TRITON_REPO/python/examples"

Trace production kernels:

rg -n 'persistent|TensorDescriptor' \
  "$TRITON_REPO/python/triton_kernels/triton_kernels/matmul_details"
 
rg -n 'mxfp|flexpoint' \
  "$TRITON_REPO/python/triton_kernels/triton_kernels/numerics_details"

Trace compiler definitions and lowering:

rg -n 'def.*Op' "$TRITON_REPO/include/triton/Dialect/Triton/IR"
rg -n 'Encoding' "$TRITON_REPO/include/triton/Dialect/TritonGPU/IR"
rg -n 'wgmma|tma|tcgen05' \
  "$TRITON_REPO/include/triton/Dialect/TritonNvidiaGPU"
rg -n 'Pattern|Rewrite' "$TRITON_REPO/lib/Conversion/TritonGPUToLLVM"

Implementation discipline

Keep these layers separate during diagnosis:

Python kernel and launch metadata
  → Triton/Gluon IR and compiler transforms
  → generated GPU code on the selected target

A source-level pattern does not prove that the compiled kernel uses the intended instruction or memory path. Inspect compiler output or profile data when that distinction matters. Add cuda-skill for PTX semantics, compute capability, Nsight, or Compute Sanitizer details.

For correctness work:

• compare against an independent reference over representative and boundary shapes;
• test masked tails, non-power-of-two dimensions, strides and dtype conversions;
• separate compilation failures from runtime correctness and numerical tolerance;
• reproduce the original dispatch and launch metadata before reducing the case.

For performance work:

• freeze the benchmark shape set and measurement method;
• confirm autotune keys cover every dimension that changes the best configuration;
• change one tile, warp, stage, persistence, or specialization choice at a time;
• verify the generated path before attributing a result to TMA, WGMMA, tcgen05, or warp specialization.

Updating the source

From the agent-gpu-skills repository:

bash update-repos.sh triton
python3 scripts/validate_repo.py --require-sources

The checkout follows Triton main, while third_party/UPSTREAMS.toml records the commit last accepted by this Skill. Review source-map drift before updating that record.