Datasets:

---

humanfia-lab

/

inference-kernels

Like 0

Follow

Humanfia 20

Curate the dataset while keeping each model shape as a separate definition. The final dataset has 21 definitions, 1,104 workloads, 5,506 tensor paths and 5,124 unique snapshots.

• Remove the two image input/output projection GEMMs and the standalone half-width gate/up slice; retain only M >= 4096 workloads for the three remaining dense GEMMs.
• Replace the old GDN tasks with the unchanged MLSys 2026 FlashInfer contest prefill/decode definitions, 154 workloads and FlashInfer baselines. These are Qwen3-Next TP=4 contest contracts, not a drop-in conversion of the original SGLang GDN calls; README documents the state-pool and normalization boundaries.
• Remove the all-visible LTX attention mask from definitions, workload inputs and unreferenced tensor files. Update both per-shape FA4 solutions to the mask-free signature, and make future captures verify visibility before omitting the mask.
• Add 19 pinned SGLang baseline wrappers and 19 KDA acceptance configs, including tolerances, dev workloads and timing. Keep their required interface adaptations explicit.

Per-shape definitions, axes, tensor directories, workloads, solutions, configs and figures remain separate. No shape-merging change is included in the final result.

Validation:

• Whole-repository metadata: all definitions, solution signatures, config references, workload/capture UUIDs and tensor references pass.
• Native FlashInfer-Bench input validation: all 21 definitions, 1,104 workloads and 5,506 tensors pass shape, dtype, provenance and complete content-hash checks; zero errors.
• LTX capture contract: 14 cases pass, covering both head dimensions, absent/all-visible masks and rejection of hidden positions or additive bias.
• FA4 GPU validation: both per-shape mask-free solutions pass all 20 dev workloads on B300, comparing every output element to the independent reference.

The existing SGLang capture-output replay records are retained. Imported contest GDN workloads have no published engine outputs; this change does not claim an in-engine replacement test or a serving speedup.