Example bundles
The repository examples are development fixtures for specific parts of the component ABI. They are not a catalog of active crowns, and their comments or metadata are not validator decisions.
Examples that omit an explicit contribution target rely on legacy singleton
resolution. Before adapting one for submission, add the appropriate
[competition] table and revalidate it against the registered target catalog.
Positive learning examples
| Example | What it demonstrates | Limits |
|---|---|---|
miner_silu_torch | smallest CPU-importable entry(x, out) bundle | correctness/packaging only; not expected to beat a tuned incumbent |
miner_silu_triton | Triton activation implementation and architecture constraints | needs matching GPU/Triton environment |
miner_rmsnorm_triton | pure RMSNorm output ownership in Triton | local example, not crown evidence |
miner_attention_torch | attention.sdpa block ABI and GQA/MQA reference shape | slow Torch diagnostic implementation |
miner_attention_decode_torch | attention.decode ABI with seq_lens | slow eager diagnostic implementation |
miner_moe_fused_experts_torch | MoE prepare plus serving entry | dense reference-style code, not a quantized fast path |
miner_moe_fused_experts_reduce_torch | distributed MoE entry that owns its trailing reduction | needs multi-rank verification for the real contract |
miner_allreduce_torch | simplest collective ABI using the supplied group | correctness example, not a competitive collective algorithm |
Start with miner_silu_torch for the workflow in
Your first component bundle. For a new target, use the
signature in Kernel ABI, not a superficially similar example.
Negative and adversarial examples
These bundles are meant to fail or expose a gate:
| Example | Intended lesson |
|---|---|
miner_silu_broken_torch | wrong activation math fails CPU correctness |
miner_silu_broken | a GPU implementation cannot win by skipping required work |
miner_silu_sparse | sparse corruption can evade naive averages, so tail/disagreement and end-to-end gates matter |
miner_rmsnorm_broken | an incorrect normalization is rejected despite plausible output shape |
miner_setup_demo | legacy engine-wide setup surface for isolation tests |
miner_setup_demo is not a registered component template. No registered
target permits setup; a cross-cutting engine change is not a valid
submission — widening the catalog is a reviewed validator-side change.
Identity fixtures are not kernels
Two test fixtures are useful for inspecting modern manifest structure:
stack_msa_singletonshows explicit singleton competition identity and capability metadata;stack_fused_epilogue_atomicshows explicit atomic identity, member rows, declared CUDA source, dependency patch, and reviewed rebuild steps.
They test intake, identity, and publication machinery. Their callable/native bodies are deliberately minimal and may not implement the live slot ABI. Do not copy them as performance kernels.
A safe way to reuse an example
- Copy only a committed source example whose ABI matches your target.
- Change
bundle_idand add explicit[competition]identity. - Replace descriptive metadata with an honest capability domain.
- Remove files and declarations your implementation does not use.
- Run
scanandverifyon the matching target environment. - Inspect the packaged archive before hosting it.
Do not copy caches, generated binaries, local result directories, machine paths, wallet material, or performance claims into a proposal. The submission must be self-contained source and declarations that the validator can reproduce.
Finding an improvement
The useful question is not “can this kernel beat a reference implementation?” It is “does this exact registered delta improve the current incumbent serving stack on the published arena workload, while preserving requi…
Submit on-chain
Production submission is a hotkey-signed, timelock commit-reveal containing a content hash and an HTTPS fetch URL. The chain carries a reference, not the archive bytes.