flashinfer.fused_moe.prepare_nvfp4_w1_gate_up_scales

flashinfer.fused_moe.prepare_nvfp4_w1_gate_up_scales(prepared_scales, original_weight_shape)

Pair existing prepared W1 scale bytes with adjacent gate/up data.

prepared_scales is the uint8 output of prepare_nvfp4_w1_scales(); original_weight_shape is the raw W1 shape [E,N,K/2]. The result is [E*(N/256)*(K/256),32,128] and preserves every E4M3 byte. Reuse it with data prepared from the same device-local weight load.