{
 "date": "2026-10-21",
 "decision": "rejected",
 "change": "Attention ranks skip their send-free main combine kernel; every reduce-epilogue CTA waits for the FFN arrival counters itself (scratch build; AFDEP_SCRATCH_MERGE=1 arm)",
 "native_source_sha256": "302198fbc0e13ca8d1f9182fee4ccf7033ab16223c02963682ba9e85f9ada7af",
 "scope": "4 x H100, 2A2F, H2048 random, BF16, graphs, 24 SMs; three fresh-process repeats",
 "correctness": "BF16/FP8 small and large GPU cases passed with the merge enabled",
 "processes": [
  {
   "order": [
    "afd",
    "afd_merge"
   ],
   "median_us": {
    "afd": {
     "1": 22.63271927833557,
     "8": 26.186000108718872,
     "32": 28.18511962890625,
     "128": 36.96936011314392,
     "512": 72.12655782699585,
     "2048": 194.11416053771973
    },
    "afd_merge": {
     "1": 22.445839643478394,
     "8": 27.0687997341156,
     "32": 28.48599910736084,
     "128": 36.362799406051636,
     "512": 72.10839986801147,
     "2048": 194.42527770996094
    }
   }
  },
  {
   "order": [
    "afd_merge",
    "afd"
   ],
   "median_us": {
    "afd_merge": {
     "1": 22.451200485229492,
     "8": 26.557281017303467,
     "32": 27.820159196853638,
     "128": 36.28319978713989,
     "512": 72.22055912017822,
     "2048": 194.10847663879395
    },
    "afd": {
     "1": 22.67359972000122,
     "8": 26.603440046310425,
     "32": 27.731279134750366,
     "128": 37.09432005882263,
     "512": 72.0307993888855,
     "2048": 193.86968612670898
    }
   }
  },
  {
   "order": [
    "afd",
    "afd_merge"
   ],
   "median_us": {
    "afd": {
     "1": 22.785359621047974,
     "8": 27.041919231414795,
     "32": 28.1112003326416,
     "128": 36.75920009613037,
     "512": 72.02936172485352,
     "2048": 194.05184745788574
    },
    "afd_merge": {
     "1": 22.2816801071167,
     "8": 26.550800800323486,
     "32": 28.025760650634766,
     "128": 36.23416066169739,
     "512": 72.37488031387329,
     "2048": 194.20543670654297
    }
   }
  }
 ],
 "table": [
  {
   "tokens": 1,
   "afd": {
    "median_us": 22.674,
    "geomean_change_pct": 0.0,
    "repeats_faster": 0
   },
   "afd_merge": {
    "median_us": 22.446,
    "geomean_change_pct": -1.341,
    "repeats_faster": 3
   }
  },
  {
   "tokens": 8,
   "afd": {
    "median_us": 26.603,
    "geomean_change_pct": 0.0,
    "repeats_faster": 0
   },
   "afd_merge": {
    "median_us": 26.557,
    "geomean_change_pct": 0.437,
    "repeats_faster": 2
   }
  },
  {
   "tokens": 32,
   "afd": {
    "median_us": 28.111,
    "geomean_change_pct": 0.0,
    "repeats_faster": 0
   },
   "afd_merge": {
    "median_us": 28.026,
    "geomean_change_pct": 0.36,
    "repeats_faster": 1
   }
  },
  {
   "tokens": 128,
   "afd": {
    "median_us": 36.969,
    "geomean_change_pct": 0.0,
    "repeats_faster": 0
   },
   "afd_merge": {
    "median_us": 36.283,
    "geomean_change_pct": -1.752,
    "repeats_faster": 3
   }
  },
  {
   "tokens": 512,
   "afd": {
    "median_us": 72.031,
    "geomean_change_pct": 0.0,
    "repeats_faster": 0
   },
   "afd_merge": {
    "median_us": 72.221,
    "geomean_change_pct": 0.239,
    "repeats_faster": 1
   }
  },
  {
   "tokens": 2048,
   "afd": {
    "median_us": 194.052,
    "geomean_change_pct": 0.0,
    "repeats_faster": 0
   },
   "afd_merge": {
    "median_us": 194.205,
    "geomean_change_pct": 0.121,
    "repeats_faster": 0
   }
  }
 ],
 "conclusion": "At most -1.8% (0.2-0.7 us), and only by occupying every Attention SM with polling CTAs for the whole FFN round trip; the kernel handoff is not where the time goes. Not shipped."
}
