For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).

Mojo package

fused_swiglu

SM100 Fused GEMM+SwiGLU kernel: BF16 matmul with SwiGLU in the epilogue.

The caller pre-permutes weight W on its N axis so adjacent output columns (2i, 2i+1) carry (gate, up) pairs. The epilogue computes silu(gate)*up in FP32 registers, writes BF16 results to double-buffered SMEM, and TMA-stores to GMEM at half-N positions.

Modules

• ​config: SM100 configuration for fused GEMM+SwiGLU.
• ​dispatch: Dispatch for fused GEMM+SwiGLU on SM100.
• ​matmul_swiglu: CPU entry point for fused GEMM+SwiGLU on SM100.
• ​matmul_swiglu_kernels: SM100 fused GEMM+SwiGLU kernel (monolithic warp-specialized).
• ​tuning_configs: Tuning configurations for fused GEMM+SwiGLU on SM100.

Was this page helpful?

Thank you! We'll create more content like this.

Thank you for helping us improve!

😔 What went wrong?

Some code doesn’t work

It includes inaccurate information

It's missing information I need

It was difficult to understand

Other

Submit