vllm.models.minimax_m3.nvidia.ops

Modules:

• dense_gemm1_mxfp8 –

  MXFP8 dense GEMM1 (gate_up) with fused clamped-SwiGLU + MXFP8 quant epilogue.

• finalize_allreduce_norm –

  MoE top-k finalize + TP all-reduce + Gemma RMSNorm + MXFP8 quantization.

• gather_nvfp4_index_k –

  Gather NVFP4 index-K pages into compact FP8 pages for the Q8KV8 prefill indexer.

• index_decode_score –

  CuteDSL MiniMax M3 index decode score kernel.

• norm_mxfp8 –

  Gemma RMSNorm (optionally fused with the residual add) in one Triton pass.

• qkv_index_fused_epilogue –

  MXFP8 QKV+indexer GEMM with a fused QK-norm/RoPE + KV-insert epilogue.

• sparse_prefill –

  Hopper query-tiled sparse prefill attention for MiniMax M3.