Which rounding does the SERVED norm hand the EXL3 quant boundary?
==============================================================================

attn input norm -> q/k/v        (x16 layers)  (G=3, fused_add=False, M=2048)
    norm_fp32                     438 / 10485760 fp16 bits differ  (  0.0042%)  <== SERVED
    norm_bf16                 9174501 / 10485760 fp16 bits differ  ( 87.4949%)
    all 3 consumers receive an identical tensor: True
    NULL CONTROL (no (1+w) fold):   10448619 (99.6458%)

gdn input norm  -> qkv/z        (x48 layers)  (G=2, fused_add=False, M=2048)
    norm_fp32                     438 / 10485760 fp16 bits differ  (  0.0042%)  <== SERVED
    norm_bf16                 9174501 / 10485760 fp16 bits differ  ( 87.4949%)
    all 2 consumers receive an identical tensor: True
    NULL CONTROL (no (1+w) fold):   10448619 (99.6458%)

post-attn norm  -> gate/up      (x64 layers)  (G=2, fused_add=True, M=2048)
    add_bf16/norm_fp32        5739290 / 10485760 fp16 bits differ  ( 54.7341%)
    add_bf16/norm_bf16        9331381 / 10485760 fp16 bits differ  ( 88.9910%)
    add_fp32/norm_fp32            357 / 10485760 fp16 bits differ  (  0.0034%)  <== SERVED
    add_fp32/norm_bf16        9166527 / 10485760 fp16 bits differ  ( 87.4188%)
    residual buffer == bf16(residual + x): True (0 differ)
    all 2 consumers receive an identical tensor: True
    NULL CONTROL (no (1+w) fold):   10439066 (99.5547%)

reference point, the site already built:

a norm with ONE consumer        (control)  (G=1, fused_add=False, M=2048)
    norm_fp32                     438 / 10485760 fp16 bits differ  (  0.0042%)  <== SERVED
    norm_bf16                 9174501 / 10485760 fp16 bits differ  ( 87.4949%)
    all 1 consumers receive an identical tensor: True
    NULL CONTROL (no (1+w) fold):   10448619 (99.6458%)
