Audit SGLang interfaces and refresh validated kernel solutions
The current definitions include capture-only attention arguments, legacy GDN wrappers, and adapters whose standalone success does not establish native SGLang call compatibility. This PR audits every one of the 21 definitions, simplifies the boundaries that can be aligned from existing evidence, and updates measured implementation choices.
Changes:
- Remove invariant attention inputs from four definitions and 202 workloads; align cached-prefix tensor argument order with SGLang. Replay all changed contracts against saved engine outputs, retain UUIDs and surviving tensor values, and remove 13 unreferenced tensors.
- Replace legacy GDN wrappers with maintained FlashInfer 0.7.2 APIs. Honor runtime scale and keep decode from changing the input state. Preserve the strict FP32 failure evidence alongside the existing contest acceptance policy.
- Use BF16 FA4/tiny native SDPA for audio attention; combine native projection/fused routing with one output-ordering kernel; provide the current native TRT-LLM sparse-GQA alternative; optimize QSA index threshold discovery while preserving exact lower-ID ties. Restore the SGLang neighborhood wrappers' native NATTEN branch.
- Select native linear on three measured large-M B300 GEMM shapes, align tested CUTLASS dependencies, remove an unreachable small-M dispatch, and add B300 targets for exercised solutions.
- Add a reproducible validator with exact task policies, input immutability checks, separate eager/prepared-graph measurements, a preferred-solution list, pinned upstream evidence, and per-workload results.
Validation on B300 / PyTorch 2.14.1:
- One current preferred solution for each definition passes all 1,104 workloads, including complete outputs, dtype/shape/device/finite checks and byte-level input preservation.
- Native schema/input/hash/provenance validation passes all 21 definitions and 5,493 tensor paths; all 202 changed attention contracts pass original engine-output replay.
- Exact QSA checks cover all 192 real workloads plus tied/empty/short/padded edge cases. The old solution also passes all 192 with its declared TileLang 0.1.15; its mismatches on SGLang's 0.1.12 are explicitly recorded as dependency incompatibility.
- Both CUDA 13.1 and 13.3 FlashMLA builds pass all 72 workloads. All 45 neighborhood workloads pass default NATTEN and SGLang/NATTEN paths. Acceptance-policy regression tests and native-builder GDN checks pass.
- Every task's tolerance, dev UUID set, timing parameters and grouping axis remain unchanged. No workload or input tensor values are replaced to obtain passes.
Remaining limits are explicit in docs/interface-audit/: seven definitions match specialized leaf tensor dimensions, three GEMMs retain leading-dimension reshaping, and eleven require structural adapters. Native strides/storage offsets/aliasing are not preserved in the snapshots. No full serving substitution is claimed. Pool side effects, per-rank buffers and newer MegaAttention fusion inputs require native recapture/integration work.
The maintained GDN prefill still fails the separate historical strict FP32 diagnostic: 3,909 / 131,072 state elements fail on the six-token case (max error 0.00121519). Passing the unchanged contest policy is not described as resolving that strict-precision issue.
Performance claims are scoped to the recorded B300 standalone graph comparisons, not global optimality or end-to-end serving throughput. See docs/interface-audit/summary.json, results.jsonl, interfaces.json, and the audit README for sources, exact coverage, comparison scopes and build details.