[bri-prism]

What

• Opt-in fusions for the MoE layer epilogue, selected by the bit mask GGML_CUDA_MOE_EP_FUSION (default 0, off):
  • bit 1: fold the per-expert scale gather (REPEAT + GET_ROWS of the down-expert scales by the selected ids) into the weighted expert reduction, which reads scale_table[ids] directly;
  • bit 2: fuse the router RMS_NORM -> SCALE -> MUL(weight) into one kernel.
• GGML_CUDA_MOE_EP_VERIFY=1|2 compares the bit-1 fusion against the existing fused reduction or the fully unfused nodes.

Why

• MoE decode runs these as several small serial kernels per layer.

How

• Bit 1 reads the same float values, so the reduction is unchanged. Bit 2 follows the reduction path the unfused RMS_NORM takes (scalar, or float4 with one float4 per thread) and does not fuse rows wider than that path covers, so it stays bit-identical.

Testing

• Greedy output is byte-identical with the switch off, the switch on and the base build, on two MoE models (Q2_0 experts and rotated PQ2_0 experts).
• GGML_CUDA_MOE_EP_VERIFY 1 and 2: 0 mismatches in decode and in 512-token batches.
• test-backend-ops (MUL_MAT_ID, MOE_WEIGHTED_REDUCE, RMS_NORM, TOP_K, GET_ROWS, ADD, MUL, SOFT_MAX, ARGSORT) identical to the base.
• H200 tg128, switch 3 vs 0, alternating runs: +3.7% (rotated PQ2_0 experts) and +4.2% (Q2_0 experts). Switch off is within noise of the base.