=== box arbi-whitebox-105, RTX 4090 sole tenant, clock pin cleared (Applications Clocks Setting: Not Active) ===
=== commit 3f319e199 on feat/int8-prologue-fusion, merged base feat/exl3-int8-gemm b3b03020b ===
=== Qwen3.8-27B-exl3-4.0bpw, 6 prompts x 16384 tok, shape 27, chunk 16384 ===

--- v1 (fresh output alloc per layer) ---
  [0] code           REF     16384tok  wall=  5.76s  finish=length
  [0] code           INT8F   16384tok  wall=  4.07s  finish=length
  [0] code           INT8P   16384tok  wall=  4.18s  finish=length
  [1] code           REF     16384tok  wall=  5.51s  finish=stop
  [1] code           INT8F   16384tok  wall=  3.83s  finish=stop
  [1] code           INT8P   16384tok  wall=  3.95s  finish=stop
  [2] wiki           REF     16384tok  wall=  5.75s  finish=length
  [2] wiki           INT8F   16384tok  wall=  4.05s  finish=length
  [2] wiki           INT8P   16384tok  wall=  4.19s  finish=length
  [3] math           REF     16384tok  wall=  5.76s  finish=length
  [3] math           INT8F   16384tok  wall=  4.07s  finish=length
  [3] math           INT8P   16384tok  wall=  4.19s  finish=length
  [4] longctx        REF     16384tok  wall=  5.76s  finish=length
  [4] longctx        INT8F   16384tok  wall=  4.18s  finish=length
  [4] longctx        INT8P   16384tok  wall=  4.19s  finish=length
  [5] sci            REF     16384tok  wall=  5.77s  finish=length
  [5] sci            INT8F   16384tok  wall=  4.07s  finish=length
  [5] sci            INT8P   16384tok  wall=  4.19s  finish=length

--- v2 (output buffer cached: allocation hypothesis REFUTED, gap unchanged) ---
  [0] code           REF     16384tok  wall=  5.50s  finish=length
  [0] code           INT8F   16384tok  wall=  4.01s  finish=length
  [0] code           INT8P   16384tok  wall=  4.13s  finish=length
  [1] code           REF     16384tok  wall=  5.29s  finish=stop
  [1] code           INT8F   16384tok  wall=  3.80s  finish=stop
  [1] code           INT8P   16384tok  wall=  3.94s  finish=stop
  [2] wiki           REF     16384tok  wall=  5.52s  finish=length
  [2] wiki           INT8F   16384tok  wall=  4.03s  finish=length
  [2] wiki           INT8P   16384tok  wall=  4.15s  finish=length
  [3] math           REF     16384tok  wall=  5.53s  finish=length
  [3] math           INT8F   16384tok  wall=  4.02s  finish=length
  [3] math           INT8P   16384tok  wall=  4.14s  finish=length
  [4] longctx        REF     16384tok  wall=  5.55s  finish=length
  [4] longctx        INT8F   16384tok  wall=  4.02s  finish=length
  [4] longctx        INT8P   16384tok  wall=  4.14s  finish=length
  [5] sci            REF     16384tok  wall=  5.56s  finish=length
  [5] sci            INT8F   16384tok  wall=  4.03s  finish=length
  [5] sci            INT8P   16384tok  wall=  4.15s  finish=length

[i8p] prologue calls=6784  fallbacks: not_serveable=0 lora=0 pad=0 act=0 noncontig=0
[i8p] dynamo unique graphs at exit: 14

--- full-vocab KL (v1) ---
arm         first-prefill-position KL       median          max
REF                         0.000e+00    0.000e+00    0.000e+00
INT8F                       2.031e-03    9.847e-04    8.533e-03
INT8P                       1.633e-03    9.952e-04    6.091e-03

[i8p] prologue calls=7104  fallbacks: not_serveable=0 lora=0 pad=0 act=0 noncontig=0
