| name: 06_sonic_moe_swiglu | |
| display_name: "Sonic-MoE up-projection (Grouped GEMM + SwiGLU)" | |
| precision: bf16 | |
| regime: compute | |
| hardware: [RTX_PRO_6000] | |
| peak_tflops_key: bf16 | |
| peak_bandwidth_key: dram | |
| # Dense-equivalent FLOPs: gate GEMM + up GEMM + SwiGLU (negligible elementwise). | |
| # Per token: 2*H*I FMAs for gate, 2*H*I for up => 2 * T_total * H * (2*I). | |
| # (Each of T_total tokens visits K experts, but total work = T_total * K * (2*H*I*2) | |
| # only if you count routing. The standard MoE FLOPs convention counts only the | |
| # active per-token compute: T_total tokens * 2 * (2*I) * H. We follow that.) | |
| flops_formula: "2 * T_total * H * (2 * I)" | |
| # Bytes moved (approximate, lower bound): | |
| # read hidden (T_perm = T_total*K rows of H bf16) + read 2 weight matrices per | |
| # expert (E * H * 2*I bf16) + write output (T_perm rows of I bf16). | |
| bytes_formula: "T_total*K*H*2 + E*H*(2*I)*2 + T_total*K*I*2" | |
| tolerance: | |
| bfloat16: 0.02 | |
| # Forbidden ops -- agent must write the grouped GEMM + fused SwiGLU themselves. | |
| # - torch.matmul / torch.bmm / F.linear: cuBLAS dispatch, defeats the point. | |
| # - sonic_moe imports: vendor-call cheating; the SOTA is graded separately. | |
| forbidden: | |
| - "torch.matmul" | |
| - "torch.bmm" | |
| - "torch.nn.functional.linear" | |
| - "F.linear" | |
| - "from sonic_moe" | |
| - "import sonic_moe" | |
| sota: | |
| name: "Sonic-MoE up-projection (Tri Dao)" | |
| url: "https://github.com/Dao-AILab/sonic-moe" | |
| function: "sonic_moe.fused_moe_up" | |
| deps: | |
| - "sonic-moe>=0.1.2" # requires Python>=3.12, sm_120 support in-progress | |
| - "quack-kernels" # CuTeDSL grouped GEMM that sonic-moe dispatches to | |
| # Documented H100 paper number for this configuration (informational, not graded | |
| # live on SM120). Sonic-MoE reports 1.87-4.04x over ScatterMoE/MoMoE on H100. | |
| reference_throughput_tflops_h100: 480 | |
| num_correct_trials: 3 | |
| num_perf_trials: 20 | |