flashinfer.fused_moe.prepare_nvfp4_w1_gate_up_data

flashinfer.fused_moe.prepare_nvfp4_w1_gate_up_data(w1)

Return byte-preserving adjacent gate/up panels for the routed entry.

Contiguous uint8 W1 [E,N,K/2] requires N % 256 == 0 and K % 256 == 0. The result is [E*(N/256)*(K/256),256,128] on the same device. Prepare after weight loading, before warmup or graph capture; keep raw weights and derived panels alive and immutable.