  in_proj_qkv    M=256    k=5120   n=10240  rel(legB vs fp32)=2.186e-03  rel(int8 vs fp32)=1.413e-02  rel(int8 gemm vs rotated fp32)=1.418e-02
  in_proj_qkv    M=1024   k=5120   n=10240  rel(legB vs fp32)=3.709e-03  rel(int8 vs fp32)=1.408e-02  rel(int8 gemm vs rotated fp32)=1.415e-02
  in_proj_z      M=256    k=5120   n=6144   rel(legB vs fp32)=2.647e-03  rel(int8 vs fp32)=1.353e-02  rel(int8 gemm vs rotated fp32)=1.352e-02
  in_proj_z      M=1024   k=5120   n=6144   rel(legB vs fp32)=2.644e-03  rel(int8 vs fp32)=1.363e-02  rel(int8 gemm vs rotated fp32)=1.363e-02
  out_proj       M=256    k=6144   n=5120   rel(legB vs fp32)=2.392e-03  rel(int8 vs fp32)=1.410e-02  rel(int8 gemm vs rotated fp32)=1.368e-02
  out_proj       M=1024   k=6144   n=5120   rel(legB vs fp32)=2.897e-03  rel(int8 vs fp32)=1.408e-02  rel(int8 gemm vs rotated fp32)=1.375e-02
  down_proj      M=256    k=17408  n=5120   rel(legB vs fp32)=2.842e-03  rel(int8 vs fp32)=1.418e-02  rel(int8 gemm vs rotated fp32)=1.415e-02
  down_proj      M=1024   k=17408  n=5120   rel(legB vs fp32)=3.453e-03  rel(int8 vs fp32)=1.392e-02  rel(int8 gemm vs rotated fp32)=1.390e-02
  gate_proj      M=256    k=5120   n=17408  rel(legB vs fp32)=2.642e-03  rel(int8 vs fp32)=1.372e-02  rel(int8 gemm vs rotated fp32)=1.373e-02
  gate_proj      M=1024   k=5120   n=17408  rel(legB vs fp32)=3.725e-03  rel(int8 vs fp32)=1.359e-02  rel(int8 gemm vs rotated fp32)=1.360e-02
  up_proj        M=256    k=5120   n=17408  rel(legB vs fp32)=2.642e-03  rel(int8 vs fp32)=1.330e-02  rel(int8 gemm vs rotated fp32)=1.330e-02
  up_proj        M=1024   k=5120   n=17408  rel(legB vs fp32)=3.724e-03  rel(int8 vs fp32)=1.338e-02  rel(int8 gemm vs rotated fp32)=1.338e-02
GATE 2b: PASS
