- Keep only kernels real inference uses; add SGLang baselines and KDA configs (5f5a41cf9e5fe452168acb5fb624af5cf884647a)
- Merge per-shape definitions into one definition per operator (9bebaaa1e3e3badfc9f1d887c8a00527c6bbfe84)
- Keep per-shape definitions and complete mask-free attention curation (f08a80540afe4ba2e7b26f06dedda04beecd81da)
Co-authored-by: Dongyun Zou <DongyunZou@users.noreply.huggingface.co>
3 days ago
Use model-independent names for generic kernel definitions
Rename 20 generic operator definitions and keep model names for the four
QSA, router and mixed-MLA interfaces with model-specific rules or layouts.
Update workload/capture references, tensor directories, baseline names,
capture hooks and shape charts together. Preserve workload UUIDs, all
tensor bytes, reference mathematics and baseline implementations.
Validation: all 24 definitions, 1383 workloads and 6763 tensor paths pass
the native schema, shape, dtype, provenance and complete content-hash checks.
24 unchanged definitions and references, 1383 workloads with matching captures, 24 baseline solutions, and all raw safetensors grouped by definition. Include only capture, validation, model preparation and plotting tooling.
