Datasets:

---

Duplicated from  flashinfer-ai/flashinfer-trace

MengYuNV

/

flashinfer-trace

like 0

Summary

Add dynamic BF16 dense GEMM coverage for all four matrix layout contracts: NN, NT, TN, and TT. M, N, and K remain workload axes so the same definitions can benchmark arbitrary supported shapes.

Added assets

• Four BF16 GEMM definitions and default workloads.
• PyTorch mathematical baselines for NN, NT, TN, and TT.
• DeepGEMM solutions mapped to bf16_gemm_nn, bf16_gemm_nt, bf16_gemm_tn, and bf16_gemm_tt.
• FlashInfer mm_bf16 solution for the NT contract.
• FlashInfer implementation-config matrix for explicit backend and pdl variants.

Validation

Validated on NVIDIA H20 with FlashInfer Bench kernel-arena commit 7c9b1a6 using 10 warmups, 50 timed iterations, and 3 trials.

• Direct DeepGEMM NN, NT, TN, and TT calls matched their mathematical references.
• Kernel Arena batch KAB-20260812T132511Z-fc2be5 completed all four definitions.
• Definition references, all PyTorch baselines, and all DeepGEMM solutions passed correctness.
• FlashInfer NT variants report backend and PDL-specific H20 support independently; unsupported combinations remain explicit runtime results.

This PR is BF16-only and does not add FP8 or MXFP4 coverage.