flashinfer.fused_moe.prepare_nvfp4_w2_scales¶
- flashinfer.fused_moe.prepare_nvfp4_w2_scales(w2_scale)¶
Return native CP-layout U8 panels
[E*(K/128)*B,8,128].Raw scales are contiguous E4M3
[E,K,8*B], B=N/256. At row r=32*g+8*c+d, word q, byte b, the 1024-byte panel’s destination is((2*c+q)*8+d)*16+4*g+b. Original raw scales remain the oracle input.