Hackathon #1: NPU Kernel Challenges
Six Ascend 910C operator challenges. The competition has ended; these instructions and final results are preserved for reference.
- 1. VectorAdd — One-dimensional vector addition with FP32 arithmetic and a final cast to the input dtype
- 2. BF16MatMul — BF16 matrix multiplication with FP32 accumulation and a final BF16 output
- 3. Fused Linear Layer with SwiGLU — Fused Linear Layer with SwiGLU: two bias-free linear projections followed by standard SiLU gating
- 4. MoE Gating (Linear + Softmax + TopK) — MoE Gating: bias-free linear router, full-expert stable softmax and top-k expert selection
- 5. LocalGroupedGEMM — Single-card grouped GEMM with runtime row offsets and support for empty groups
- 6. IndexedSparseAttention — Single-card indexed sparse attention: repeated indices count separately; all-invalid rows produce zeros