vllm.model_executor.layers.fused_moe.experts.flashinfer_moe_ep
¶
FlashInfer MoE-EP megakernels as a modular-kernel experts implementation.
The megakernel routes on vLLM's top-k output and then dispatches, computes and combines in one launch, so it is paired with a pass-through prepare/finalize.
Classes:
Functions:
-
epilogue_from_quant_config–Per-expert epilogue constants from the canonical quant config.
-
sm107_nvfp4_epilogue–NVFP4 epilogue of the SM107 megakernel from the checkpoint's activation
FlashInferMoeEpExperts
¶
Bases: FusedMoEExpertsModular
Methods:
-
process_weights_after_loading–Build the megakernel from the layer's canonical weights.
Source code in vllm/model_executor/layers/fused_moe/experts/flashinfer_moe_ep.py
56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 | |
process_weights_after_loading(layer)
¶
Build the megakernel from the layer's canonical weights.
The oracle has already put the weights in kernel format; the per-expert
weight global scales arrive through the quant config and become the
epilogue alphas. On SM107 the checkpoint's activation input scales also
set the activation and FC1-output quantization (sm107_nvfp4_epilogue).
Source code in vllm/model_executor/layers/fused_moe/experts/flashinfer_moe_ep.py
epilogue_from_quant_config(quant_config)
¶
Per-expert epilogue constants from the canonical quant config.
The weight global scales (g1_alphas/g2_alphas) become the fc1/fc2
alphas. Activations are quantized dynamically inside the megakernel, so on
SM100/SM103 the checkpoint's input scales drop out; SM107 uses them
(sm107_nvfp4_epilogue). MXFP4 checkpoints carry no global scales
and keep the default epilogue. The alphas alias the layer's parameters, so
EPLB permutations reach the kernel.
Source code in vllm/model_executor/layers/fused_moe/experts/flashinfer_moe_ep.py
sm107_nvfp4_epilogue(epilogue, w13_input_scale, w2_input_scale)
¶
NVFP4 epilogue of the SM107 megakernel from the checkpoint's activation scales.
With unit norm constants small activations can underflow the E4M3 block scales. Tokens are quantized on their source rank, so the activation grid is the layer maximum over every EP rank. The alphas are scaled in place and keep aliasing the layer's weight global scales. Collective over the EP group: every EP rank must call it for the same MoE layers in the same order.