{
 "date": "2026-10-21",
 "decision": "rejected",
 "change": "streaming FFN expand: sources release per-slot ready flags after each TMA store completes; the FFN dispatch kernel releases its PDL copy epilogue at entry and the epilogue waits for prefix sums, then per slot",
 "scope": "4 x H100, 2 Attention + 2 FFN, BF16, H2048 random, graphs, 24 SMs; max over Attention ranks of per-process block medians",
 "screens": [
  {
   "name": "release flags (candidate)",
   "native_source_sha256": "4a0b4a71062535abbb835294947cdb0360dabce8859626be8c2e043a111ed024",
   "table": [
    {
     "tokens": 1,
     "masked": {
      "median_us": 31.27,
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "afd": {
      "median_us": 22.89,
      "geomean_change_pct": -26.91,
      "repeats_faster": 3
     },
     "afd_nostream": {
      "median_us": 19.59,
      "geomean_change_pct": -37.22,
      "repeats_faster": 3
     }
    },
    {
     "tokens": 32,
     "masked": {
      "median_us": 35.52,
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "afd": {
      "median_us": 26.51,
      "geomean_change_pct": -26.28,
      "repeats_faster": 3
     },
     "afd_nostream": {
      "median_us": 22.01,
      "geomean_change_pct": -38.61,
      "repeats_faster": 3
     }
    },
    {
     "tokens": 128,
     "masked": {
      "median_us": 44.56,
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "afd": {
      "median_us": 35.38,
      "geomean_change_pct": -21.02,
      "repeats_faster": 3
     },
     "afd_nostream": {
      "median_us": 30.28,
      "geomean_change_pct": -32.52,
      "repeats_faster": 3
     }
    },
    {
     "tokens": 512,
     "masked": {
      "median_us": 78.53,
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "afd": {
      "median_us": 72.73,
      "geomean_change_pct": -7.35,
      "repeats_faster": 3
     },
     "afd_nostream": {
      "median_us": 67.94,
      "geomean_change_pct": -13.39,
      "repeats_faster": 3
     }
    },
    {
     "tokens": 2048,
     "masked": {
      "median_us": 200.97,
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "afd": {
      "median_us": 193.29,
      "geomean_change_pct": -3.76,
      "repeats_faster": 3
     },
     "afd_nostream": {
      "median_us": 185.66,
      "geomean_change_pct": -7.73,
      "repeats_faster": 3
     }
    }
   ]
  },
  {
   "name": "relaxed flags + backoff polling (scratch timing variant, unsafe ordering)",
   "native_source_sha256": "d8784f5d66291014181b4fdfae011590311bdce56980ce4194765cc029da5414",
   "repeats": [
    {
     "tokens": 1,
     "variant_us": [
      25.83,
      25.81
     ],
     "control_us": [
      19.61,
      19.67
     ]
    },
    {
     "tokens": 128,
     "variant_us": [
      37.75,
      38.23
     ],
     "control_us": [
      30.41,
      30.49
     ]
    },
    {
     "tokens": 512,
     "variant_us": [
      75.97,
      76.34
     ],
     "control_us": [
      67.55,
      67.66
     ]
    },
    {
     "tokens": 2048,
     "variant_us": [
      196.89,
      197.36
     ],
     "control_us": [
      185.75,
      186.11
     ]
    }
   ]
  },
  {
   "name": "relaxed flags, acquire polling (scratch timing variant, unsafe ordering; per-process JSON not retained)",
   "summary": {
    "1": [
     21.73,
     19.59
    ],
    "128": [
     32.26,
     30.3
    ],
    "512": [
     69.57,
     67.69
    ],
    "2048": [
     188.13,
     185.8
    ]
   },
   "columns": [
    "variant median us",
    "control median us"
   ]
  }
 ],
 "conclusion": "No variant beats its control at any token count; even without a release fence the early-launched epilogue does not shorten the round trip. Not shipped."
}