{
 "date": "2026-10-23",
 "change": "opt-in FP8 combine replies across NVLink domains on the directional hybrid (AFDEP_HYBRID_FP8_COMBINE=1) and their minimum bucket (AFDEP_HYBRID_FP8_COMBINE_MIN_BUCKET); docs/design.md section 19; off by default",
 "commits": [
  "4c8521d",
  "1e0a1b5",
  "3989bcf",
  "189c0eb"
 ],
 "native_source_sha256": {
  "ungated_gpu_suite_and_screens": "024e98b05de66aecbc2e72da3046d8d083b0ec4776dae3ec927f6330f252b846",
  "gated_final": "2cee2f51ee29cdbeb62ad4c38f77ba6bd5fd006b73e02181f5a9c8590cb4edae",
  "fault_tree": "60cf91b3c74ef7814e8da71ae73ebffb82ffd86cf12a7e1c0f841398e38dd1e9",
  "negative_control_tree": "679c96af7f1ae589ce65483c2a606493e3b31f6452c55a585ff978e0bb6c9124",
  "accuracy_check": "b5078d629964b48d706eba81f1251e294c138c5e08f8d86240f711e1676e7405"
 },
 "setup": "kermit (CoreWeave) H200, NCCL 2.30.7 GIN over InfiniBand; 6 pods x 4 GPUs (one NVLink domain each) and 4 pods x 8 GPUs; vLLM 0.30.0 with the afd-plugin lanes tree (5215c41) for serving; Qwen3-235B-A22B-FP8 shapes (hidden 4096, top-8 of 128 experts, FP8 dispatch) unless stated",
 "decision": "opt-in; recommended with AFDEP_HYBRID_FP8_COMBINE_MIN_BUCKET=64. Round trip -4% to -21% from 64 tokens per Attention GPU, neutral below with the minimum bucket (+3% to +8% without it). Serving at 4,096/2,048 tokens: -1.4% to +0.9% TPOT. Lossy; GSM8K on Qwen3-235B is unchanged within run-to-run noise with FP8 replies at every bucket (95.07% against 95.30% and 95.07% for two BF16 runs); per-layer MoE error and next-token KL are not measured yet.",
 "gpu_suite_ungated": [
  {
   "case": "g14-h4096",
   "result": "PASS",
   "pods_exit0": "6/6"
  },
  {
   "case": "g14-h4096-sparse",
   "result": "PASS",
   "pods_exit0": "6/6"
  },
  {
   "case": "g1-c8",
   "result": "PASS",
   "pods_exit0": "6/6"
  },
  {
   "case": "g1-c64",
   "result": "PASS",
   "pods_exit0": "6/6"
  },
  {
   "case": "g1-c2048",
   "result": "PASS",
   "pods_exit0": "6/6"
  },
  {
   "case": "g2-bf16-c8",
   "result": "PASS",
   "pods_exit0": "6/6"
  },
  {
   "case": "g2-bf16-c64",
   "result": "PASS",
   "pods_exit0": "6/6"
  },
  {
   "case": "g3-h4096-fp8",
   "result": "PASS",
   "pods_exit0": "6/6"
  },
  {
   "case": "g3-h4096-bf16",
   "result": "PASS",
   "pods_exit0": "6/6"
  },
  {
   "case": "g3-h7168-fp8",
   "result": "PASS",
   "pods_exit0": "6/6"
  },
  {
   "case": "g3-h7168-bf16",
   "result": "PASS",
   "pods_exit0": "6/6"
  },
  {
   "case": "g3-h8192-fp8",
   "result": "PASS",
   "pods_exit0": "6/6"
  },
  {
   "case": "g3-h8192-bf16",
   "result": "PASS",
   "pods_exit0": "6/6"
  },
  {
   "case": "g4-topk2",
   "result": "PASS",
   "pods_exit0": "6/6"
  },
  {
   "case": "g5-c8",
   "result": "PASS",
   "pods_exit0": "6/6"
  },
  {
   "case": "g5-c64-stagger16",
   "result": "PASS",
   "pods_exit0": "6/6"
  },
  {
   "case": "g5-c2048",
   "result": "PASS",
   "pods_exit0": "6/6"
  },
  {
   "case": "g5-nostagger",
   "result": "PASS",
   "pods_exit0": "6/6"
  },
  {
   "case": "g6-f8-c8",
   "result": "PASS",
   "pods_exit0": "6/6"
  },
  {
   "case": "g6-f8-c8-sparse",
   "result": "PASS",
   "pods_exit0": "6/6"
  },
  {
   "case": "g6-f8-c64",
   "result": "PASS",
   "pods_exit0": "6/6"
  },
  {
   "case": "g6-f8-c64-sparse",
   "result": "PASS",
   "pods_exit0": "6/6"
  },
  {
   "case": "g7-s1",
   "result": "PASS",
   "pods_exit0": "2/2"
  },
  {
   "case": "g7-s1-sparse",
   "result": "PASS",
   "pods_exit0": "2/2"
  },
  {
   "case": "g8-shared-collect",
   "result": "PASS",
   "pods_exit0": "6/6"
  },
  {
   "case": "g9-mismatch",
   "result": "PASS",
   "pods_exit0": "0/6",
   "note": "negative: every rank must reject the mismatch ('same hybrid configuration'); 6/6 rejected"
  },
  {
   "case": "g10-mixed",
   "result": "PASS",
   "pods_exit0": "6/6",
   "note": "off topology: fp8_combine reported false on every rank, BF16 reference"
  },
  {
   "case": "g10-union",
   "result": "PASS",
   "pods_exit0": "6/6",
   "note": "off topology: fp8_combine reported false on every rank, BF16 reference"
  },
  {
   "case": "g10-innode",
   "result": "PASS",
   "pods_exit0": "1/1",
   "note": "off topology: fp8_combine reported false on every rank, BF16 reference"
  },
  {
   "case": "g10-innode-off",
   "result": "PASS",
   "pods_exit0": "1/1",
   "note": "off topology: fp8_combine reported false on every rank, BF16 reference"
  },
  {
   "case": "g11-route",
   "result": "PASS",
   "pods_exit0": "0/6",
   "note": "negative: an invalid route must stop at the in-kernel assertion; it fired on 6/6 pods"
  }
 ],
 "gpu_suite_gated": [
  {
   "case": "gate-c8-mb4",
   "result": "PASS",
   "pods_exit0": "4/4"
  },
  {
   "case": "gate-c8-mb4-sparse",
   "result": "PASS",
   "pods_exit0": "4/4"
  },
  {
   "case": "gate-c64-mb32",
   "result": "PASS",
   "pods_exit0": "4/4"
  },
  {
   "case": "gate-c64-mb32-sparse",
   "result": "PASS",
   "pods_exit0": "4/4"
  },
  {
   "case": "gate-c64-mb32-bf16",
   "result": "PASS",
   "pods_exit0": "4/4"
  },
  {
   "case": "gate-f8-c64-mb32",
   "result": "PASS",
   "pods_exit0": "4/4"
  },
  {
   "case": "gate-f8-c64-mb32-sparse",
   "result": "PASS",
   "pods_exit0": "4/4"
  },
  {
   "case": "gate-s1-c64-mb32",
   "result": "PASS",
   "pods_exit0": "2/2"
  },
  {
   "case": "gate-s1-c64-mb32-sparse",
   "result": "PASS",
   "pods_exit0": "2/2"
  },
  {
   "case": "gate-c64-mb4096",
   "result": "PASS",
   "pods_exit0": "4/4"
  }
 ],
 "screens": {
  "4gpu_domains_4F_4A": {
   "pods": "2 pods (koc) x 4 GPUs, InfiniBand, GIN",
   "args": "--payload fp8 --routing random --hidden 4096 --experts 128 --top-k 8 --tokens 1 8 32 64 128 256 512 --capacity 2048 --hybrid on --ffn-ranks 4 --attention-ranks 4 --warmup 100 --iterations 200 --blocks 5",
   "digest": [
    "024e98b05de66aecbc2e72da3046d8d083b0ec4776dae3ec927f6330f252b846"
   ],
   "median_us": [
    {
     "tokens": 1,
     "default": 79.1,
     "default_fp8": 81.3,
     "default_fp8_paired_pct": 2.8,
     "sparse": 74.5,
     "sparse_paired_pct": -5.8,
     "sparse_fp8": 77.3,
     "sparse_fp8_paired_pct": -2.2
    },
    {
     "tokens": 8,
     "default": 87.7,
     "default_fp8": 90.4,
     "default_fp8_paired_pct": 3.0,
     "sparse": 86.1,
     "sparse_paired_pct": -1.8,
     "sparse_fp8": 88.8,
     "sparse_fp8_paired_pct": 1.2
    },
    {
     "tokens": 32,
     "default": 94.3,
     "default_fp8": 96.0,
     "default_fp8_paired_pct": 1.9,
     "sparse": 93.8,
     "sparse_paired_pct": -0.4,
     "sparse_fp8": 96.3,
     "sparse_fp8_paired_pct": 2.2
    },
    {
     "tokens": 64,
     "default": 103.6,
     "default_fp8": 103.4,
     "default_fp8_paired_pct": -0.2,
     "sparse": 100.2,
     "sparse_paired_pct": -3.3,
     "sparse_fp8": 102.5,
     "sparse_fp8_paired_pct": -1.1
    },
    {
     "tokens": 128,
     "default": 124.1,
     "default_fp8": 119.8,
     "default_fp8_paired_pct": -3.5,
     "sparse": 119.3,
     "sparse_paired_pct": -3.8,
     "sparse_fp8": 115.1,
     "sparse_fp8_paired_pct": -7.2
    },
    {
     "tokens": 256,
     "default": 163.9,
     "default_fp8": 158.2,
     "default_fp8_paired_pct": -3.5,
     "sparse": 166.8,
     "sparse_paired_pct": 1.8,
     "sparse_fp8": 162.0,
     "sparse_fp8_paired_pct": -1.2
    },
    {
     "tokens": 512,
     "default": 270.3,
     "default_fp8": 239.1,
     "default_fp8_paired_pct": -11.5,
     "sparse": 268.8,
     "sparse_paired_pct": -0.5,
     "sparse_fp8": 239.9,
     "sparse_fp8_paired_pct": -11.2
    }
   ]
  },
  "4gpu_domains_4F_12A": {
   "pods": "4 pods (koc) x 4 GPUs, InfiniBand, GIN",
   "args": "--payload fp8 --routing random --hidden 4096 --experts 128 --top-k 8 --tokens 1 8 32 64 128 256 512 --capacity 2048 --hybrid on --ffn-ranks 4 --attention-ranks 12 --warmup 100 --iterations 200 --blocks 5",
   "digest": [
    "024e98b05de66aecbc2e72da3046d8d083b0ec4776dae3ec927f6330f252b846"
   ],
   "median_us": [
    {
     "tokens": 1,
     "default": 119.6,
     "default_fp8": 125.8,
     "default_fp8_paired_pct": 5.2,
     "sparse": 77.0,
     "sparse_paired_pct": -35.6,
     "sparse_fp8": 81.1,
     "sparse_fp8_paired_pct": -32.2
    },
    {
     "tokens": 8,
     "default": 151.6,
     "default_fp8": 156.9,
     "default_fp8_paired_pct": 3.5,
     "sparse": 92.5,
     "sparse_paired_pct": -39.0,
     "sparse_fp8": 96.3,
     "sparse_fp8_paired_pct": -36.5
    },
    {
     "tokens": 32,
     "default": 160.7,
     "default_fp8": 166.7,
     "default_fp8_paired_pct": 3.8,
     "sparse": 106.3,
     "sparse_paired_pct": -33.9,
     "sparse_fp8": 106.8,
     "sparse_fp8_paired_pct": -33.6
    },
    {
     "tokens": 64,
     "default": 177.6,
     "default_fp8": 176.8,
     "default_fp8_paired_pct": -0.5,
     "sparse": 134.4,
     "sparse_paired_pct": -24.3,
     "sparse_fp8": 127.8,
     "sparse_fp8_paired_pct": -28.0
    },
    {
     "tokens": 128,
     "default": 233.4,
     "default_fp8": 220.5,
     "default_fp8_paired_pct": -5.5,
     "sparse": 215.7,
     "sparse_paired_pct": -7.6,
     "sparse_fp8": 200.1,
     "sparse_fp8_paired_pct": -14.3
    },
    {
     "tokens": 256,
     "default": 361.4,
     "default_fp8": 327.2,
     "default_fp8_paired_pct": -9.5,
     "sparse": 364.4,
     "sparse_paired_pct": 0.8,
     "sparse_fp8": 321.8,
     "sparse_fp8_paired_pct": -10.9
    },
    {
     "tokens": 512,
     "default": 642.3,
     "default_fp8": 531.5,
     "default_fp8_paired_pct": -17.2,
     "sparse": 643.2,
     "sparse_paired_pct": 0.1,
     "sparse_fp8": 533.8,
     "sparse_fp8_paired_pct": -16.9
    }
   ]
  },
  "4gpu_domains_4F_20A": {
   "pods": "6 pods (koc) x 4 GPUs, InfiniBand, GIN",
   "args": "--payload fp8 --routing random --hidden 4096 --experts 128 --top-k 8 --tokens 1 8 32 64 128 256 512 --capacity 2048 --hybrid on --ffn-ranks 4 --attention-ranks 20 --warmup 100 --iterations 200 --blocks 5",
   "digest": [
    "024e98b05de66aecbc2e72da3046d8d083b0ec4776dae3ec927f6330f252b846"
   ],
   "median_us": [
    {
     "tokens": 1,
     "default": 163.3,
     "default_fp8": 172.1,
     "default_fp8_paired_pct": 5.4,
     "sparse": 79.0,
     "sparse_paired_pct": -51.7,
     "sparse_fp8": 84.6,
     "sparse_fp8_paired_pct": -48.2
    },
    {
     "tokens": 8,
     "default": 197.5,
     "default_fp8": 205.2,
     "default_fp8_paired_pct": 3.9,
     "sparse": 94.0,
     "sparse_paired_pct": -52.4,
     "sparse_fp8": 98.6,
     "sparse_fp8_paired_pct": -50.1
    },
    {
     "tokens": 32,
     "default": 213.3,
     "default_fp8": 219.9,
     "default_fp8_paired_pct": 3.1,
     "sparse": 124.0,
     "sparse_paired_pct": -41.9,
     "sparse_fp8": 120.2,
     "sparse_fp8_paired_pct": -43.6
    },
    {
     "tokens": 64,
     "default": 245.6,
     "default_fp8": 245.6,
     "default_fp8_paired_pct": -0.0,
     "sparse": 183.4,
     "sparse_paired_pct": -25.3,
     "sparse_fp8": 177.1,
     "sparse_fp8_paired_pct": -27.9
    },
    {
     "tokens": 128,
     "default": 336.0,
     "default_fp8": 311.0,
     "default_fp8_paired_pct": -7.4,
     "sparse": 324.5,
     "sparse_paired_pct": -3.4,
     "sparse_fp8": 297.2,
     "sparse_fp8_paired_pct": -11.5
    },
    {
     "tokens": 256,
     "default": 542.8,
     "default_fp8": 483.1,
     "default_fp8_paired_pct": -11.0,
     "sparse": 540.5,
     "sparse_paired_pct": -0.4,
     "sparse_fp8": 478.8,
     "sparse_fp8_paired_pct": -11.8
    },
    {
     "tokens": 512,
     "default": 994.0,
     "default_fp8": 811.1,
     "default_fp8_paired_pct": -18.4,
     "sparse": 997.4,
     "sparse_paired_pct": 0.3,
     "sparse_fp8": 813.6,
     "sparse_fp8_paired_pct": -18.1
    }
   ]
  },
  "4gpu_domains_8F_16A": {
   "pods": "6 pods (koc) x 4 GPUs, InfiniBand, GIN",
   "args": "--payload fp8 --routing random --hidden 4096 --experts 128 --top-k 8 --tokens 1 8 32 64 128 256 512 --capacity 2048 --hybrid on --ffn-ranks 8 --attention-ranks 16 --warmup 100 --iterations 200 --blocks 5",
   "digest": [
    "024e98b05de66aecbc2e72da3046d8d083b0ec4776dae3ec927f6330f252b846"
   ],
   "median_us": [
    {
     "tokens": 1,
     "default": 126.0,
     "default_fp8": 135.0,
     "default_fp8_paired_pct": 7.2,
     "sparse": 71.4,
     "sparse_paired_pct": -43.3,
     "sparse_fp8": 77.3,
     "sparse_fp8_paired_pct": -38.6
    },
    {
     "tokens": 8,
     "default": 143.1,
     "default_fp8": 150.9,
     "default_fp8_paired_pct": 5.5,
     "sparse": 85.0,
     "sparse_paired_pct": -40.6,
     "sparse_fp8": 90.5,
     "sparse_fp8_paired_pct": -36.7
    },
    {
     "tokens": 32,
     "default": 156.7,
     "default_fp8": 164.8,
     "default_fp8_paired_pct": 5.2,
     "sparse": 106.2,
     "sparse_paired_pct": -32.2,
     "sparse_fp8": 104.7,
     "sparse_fp8_paired_pct": -33.2
    },
    {
     "tokens": 64,
     "default": 182.1,
     "default_fp8": 180.7,
     "default_fp8_paired_pct": -0.8,
     "sparse": 149.8,
     "sparse_paired_pct": -17.8,
     "sparse_fp8": 142.7,
     "sparse_fp8_paired_pct": -21.7
    },
    {
     "tokens": 128,
     "default": 250.1,
     "default_fp8": 229.9,
     "default_fp8_paired_pct": -8.1,
     "sparse": 242.4,
     "sparse_paired_pct": -3.0,
     "sparse_fp8": 221.4,
     "sparse_fp8_paired_pct": -11.4
    },
    {
     "tokens": 256,
     "default": 415.6,
     "default_fp8": 362.3,
     "default_fp8_paired_pct": -12.8,
     "sparse": 418.0,
     "sparse_paired_pct": 0.6,
     "sparse_fp8": 359.2,
     "sparse_fp8_paired_pct": -13.6
    },
    {
     "tokens": 512,
     "default": 763.6,
     "default_fp8": 603.4,
     "default_fp8_paired_pct": -21.0,
     "sparse": 765.1,
     "sparse_paired_pct": 0.2,
     "sparse_fp8": 608.5,
     "sparse_fp8_paired_pct": -20.3
    }
   ]
  },
  "4gpu_domains_4F_20A_bf16": {
   "pods": "6 pods (koc) x 4 GPUs, InfiniBand, GIN",
   "args": "--payload bf16 --routing random --hidden 4096 --experts 128 --top-k 8 --tokens 64 128 512 --capacity 2048 --hybrid on --ffn-ranks 4 --attention-ranks 20 --warmup 100 --iterations 200 --blocks 5",
   "digest": [
    "024e98b05de66aecbc2e72da3046d8d083b0ec4776dae3ec927f6330f252b846"
   ],
   "median_us": [
    {
     "tokens": 64,
     "default": 283.9,
     "default_fp8": 284.2,
     "default_fp8_paired_pct": 0.1,
     "sparse": 221.0,
     "sparse_paired_pct": -22.1,
     "sparse_fp8": 215.0,
     "sparse_fp8_paired_pct": -24.3
    },
    {
     "tokens": 128,
     "default": 409.8,
     "default_fp8": 384.9,
     "default_fp8_paired_pct": -6.1,
     "sparse": 398.9,
     "sparse_paired_pct": -2.7,
     "sparse_fp8": 372.4,
     "sparse_fp8_paired_pct": -9.1
    },
    {
     "tokens": 512,
     "default": 1283.3,
     "default_fp8": 1099.8,
     "default_fp8_paired_pct": -14.3,
     "sparse": 1292.0,
     "sparse_paired_pct": 0.7,
     "sparse_fp8": 1108.3,
     "sparse_fp8_paired_pct": -13.6
    }
   ]
  },
  "gated_4F_12A": {
   "pods": "4 pods (koc) x 4 GPUs, InfiniBand, GIN",
   "args": "--payload fp8 --routing random --hidden 4096 --experts 128 --top-k 8 --tokens 1 8 32 64 128 256 512 --capacity 2048 --hybrid on --ffn-ranks 4 --attention-ranks 12 --warmup 100 --iterations 200 --blocks 5",
   "digest": [
    "2cee2f51ee29cdbeb62ad4c38f77ba6bd5fd006b73e02181f5a9c8590cb4edae"
   ],
   "median_us": [
    {
     "tokens": 1,
     "sparse": 76.4,
     "sparse_fp8": 81.1,
     "sparse_fp8_paired_pct": 6.1,
     "sparse_fp8_gate64": 76.9,
     "sparse_fp8_gate64_paired_pct": 0.6
    },
    {
     "tokens": 8,
     "sparse": 90.1,
     "sparse_fp8": 94.4,
     "sparse_fp8_paired_pct": 4.7,
     "sparse_fp8_gate64": 90.3,
     "sparse_fp8_gate64_paired_pct": 0.2
    },
    {
     "tokens": 32,
     "sparse": 102.1,
     "sparse_fp8": 104.8,
     "sparse_fp8_paired_pct": 2.7,
     "sparse_fp8_gate64": 102.2,
     "sparse_fp8_gate64_paired_pct": 0.1
    },
    {
     "tokens": 64,
     "sparse": 129.9,
     "sparse_fp8": 123.8,
     "sparse_fp8_paired_pct": -4.6,
     "sparse_fp8_gate64": 123.6,
     "sparse_fp8_gate64_paired_pct": -4.8
    },
    {
     "tokens": 128,
     "sparse": 213.0,
     "sparse_fp8": 197.3,
     "sparse_fp8_paired_pct": -7.4,
     "sparse_fp8_gate64": 196.7,
     "sparse_fp8_gate64_paired_pct": -7.6
    },
    {
     "tokens": 256,
     "sparse": 361.5,
     "sparse_fp8": 319.8,
     "sparse_fp8_paired_pct": -11.5,
     "sparse_fp8_gate64": 319.6,
     "sparse_fp8_gate64_paired_pct": -11.6
    },
    {
     "tokens": 512,
     "sparse": 639.8,
     "sparse_fp8": 529.6,
     "sparse_fp8_paired_pct": -17.2,
     "sparse_fp8_gate64": 530.4,
     "sparse_fp8_gate64_paired_pct": -17.1
    }
   ]
  },
  "8gpu_domains_8F_8A": {
   "pods": "2 pods (koc) x 8 GPUs, InfiniBand, GIN",
   "args": "--payload fp8 --routing random --hidden 4096 --experts 128 --top-k 8 --tokens 1 8 32 64 128 256 512 --capacity 2048 --hybrid on --ffn-ranks 8 --attention-ranks 8 --warmup 100 --iterations 200 --blocks 5",
   "digest": [
    "2cee2f51ee29cdbeb62ad4c38f77ba6bd5fd006b73e02181f5a9c8590cb4edae"
   ],
   "median_us": [
    {
     "tokens": 1,
     "default": 81.6,
     "sparse": 76.8,
     "sparse_paired_pct": -5.9,
     "sparse_fp8_gate64": 77.1,
     "sparse_fp8_gate64_paired_pct": -5.5
    },
    {
     "tokens": 8,
     "default": 96.2,
     "sparse": 91.2,
     "sparse_paired_pct": -5.2,
     "sparse_fp8_gate64": 90.9,
     "sparse_fp8_gate64_paired_pct": -5.6
    },
    {
     "tokens": 32,
     "default": 99.5,
     "sparse": 98.0,
     "sparse_paired_pct": -1.5,
     "sparse_fp8_gate64": 97.9,
     "sparse_fp8_gate64_paired_pct": -1.6
    },
    {
     "tokens": 64,
     "default": 108.2,
     "sparse": 105.7,
     "sparse_paired_pct": -2.3,
     "sparse_fp8_gate64": 107.3,
     "sparse_fp8_gate64_paired_pct": -0.8
    },
    {
     "tokens": 128,
     "default": 129.4,
     "sparse": 125.5,
     "sparse_paired_pct": -3.0,
     "sparse_fp8_gate64": 121.6,
     "sparse_fp8_gate64_paired_pct": -6.0
    },
    {
     "tokens": 256,
     "default": 183.9,
     "sparse": 183.9,
     "sparse_paired_pct": 0.0,
     "sparse_fp8_gate64": 181.7,
     "sparse_fp8_gate64_paired_pct": -1.2
    },
    {
     "tokens": 512,
     "default": 303.3,
     "sparse": 303.0,
     "sparse_paired_pct": -0.1,
     "sparse_fp8_gate64": 274.3,
     "sparse_fp8_gate64_paired_pct": -9.5
    }
   ]
  },
  "8gpu_domains_8F_16A": {
   "pods": "3 pods (koc) x 8 GPUs, InfiniBand, GIN",
   "args": "--payload fp8 --routing random --hidden 4096 --experts 128 --top-k 8 --tokens 1 8 32 64 128 256 512 --capacity 2048 --hybrid on --ffn-ranks 8 --attention-ranks 16 --warmup 100 --iterations 200 --blocks 5",
   "digest": [
    "2cee2f51ee29cdbeb62ad4c38f77ba6bd5fd006b73e02181f5a9c8590cb4edae"
   ],
   "median_us": [
    {
     "tokens": 1,
     "default": 103.5,
     "sparse": 77.9,
     "sparse_paired_pct": -24.7,
     "sparse_fp8_gate64": 78.0,
     "sparse_fp8_gate64_paired_pct": -24.6
    },
    {
     "tokens": 8,
     "default": 122.5,
     "sparse": 93.0,
     "sparse_paired_pct": -24.1,
     "sparse_fp8_gate64": 93.1,
     "sparse_fp8_gate64_paired_pct": -24.0
    },
    {
     "tokens": 32,
     "default": 132.5,
     "sparse": 101.7,
     "sparse_paired_pct": -23.3,
     "sparse_fp8_gate64": 101.8,
     "sparse_fp8_gate64_paired_pct": -23.1
    },
    {
     "tokens": 64,
     "default": 145.2,
     "sparse": 120.0,
     "sparse_paired_pct": -17.3,
     "sparse_fp8_gate64": 116.2,
     "sparse_fp8_gate64_paired_pct": -20.0
    },
    {
     "tokens": 128,
     "default": 186.4,
     "sparse": 176.1,
     "sparse_paired_pct": -5.5,
     "sparse_fp8_gate64": 179.8,
     "sparse_fp8_gate64_paired_pct": -3.5
    },
    {
     "tokens": 256,
     "default": 287.3,
     "sparse": 294.2,
     "sparse_paired_pct": 2.4,
     "sparse_fp8_gate64": 273.7,
     "sparse_fp8_gate64_paired_pct": -4.7
    },
    {
     "tokens": 512,
     "default": 523.6,
     "sparse": 528.8,
     "sparse_paired_pct": 1.0,
     "sparse_fp8_gate64": 458.9,
     "sparse_fp8_gate64_paired_pct": -12.4
    }
   ]
  },
  "8gpu_domains_8F_24A": {
   "pods": "4 pods (koc) x 8 GPUs, InfiniBand, GIN",
   "args": "--payload fp8 --routing random --hidden 4096 --experts 128 --top-k 8 --tokens 1 8 32 64 128 256 512 --capacity 2048 --hybrid on --ffn-ranks 8 --attention-ranks 24 --warmup 100 --iterations 200 --blocks 5",
   "digest": [
    "2cee2f51ee29cdbeb62ad4c38f77ba6bd5fd006b73e02181f5a9c8590cb4edae"
   ],
   "median_us": [
    {
     "tokens": 1,
     "default": 137.4,
     "sparse": 80.4,
     "sparse_paired_pct": -41.5,
     "sparse_fp8_gate64": 81.8,
     "sparse_fp8_gate64_paired_pct": -40.5
    },
    {
     "tokens": 8,
     "default": 159.4,
     "sparse": 95.3,
     "sparse_paired_pct": -40.2,
     "sparse_fp8_gate64": 95.5,
     "sparse_fp8_gate64_paired_pct": -40.1
    },
    {
     "tokens": 32,
     "default": 171.9,
     "sparse": 109.2,
     "sparse_paired_pct": -36.5,
     "sparse_fp8_gate64": 109.1,
     "sparse_fp8_gate64_paired_pct": -36.5
    },
    {
     "tokens": 64,
     "default": 186.4,
     "sparse": 145.4,
     "sparse_paired_pct": -22.0,
     "sparse_fp8_gate64": 139.3,
     "sparse_fp8_gate64_paired_pct": -25.3
    },
    {
     "tokens": 128,
     "default": 249.7,
     "sparse": 239.3,
     "sparse_paired_pct": -4.1,
     "sparse_fp8_gate64": 224.2,
     "sparse_fp8_gate64_paired_pct": -10.2
    },
    {
     "tokens": 256,
     "default": 399.7,
     "sparse": 418.6,
     "sparse_paired_pct": 4.7,
     "sparse_fp8_gate64": 383.3,
     "sparse_fp8_gate64_paired_pct": -4.1
    },
    {
     "tokens": 512,
     "default": 728.2,
     "sparse": 748.6,
     "sparse_paired_pct": 2.8,
     "sparse_fp8_gate64": 642.6,
     "sparse_fp8_gate64_paired_pct": -11.7
    }
   ]
  }
 },
 "kernel_trace_4F_20A": {
  "method": "nsys, graph replays, median over devices; sparse vs sparse + FP8 on 024e98b0",
  "attention_reduce_epilogue_us": {
   "8_tokens": [
    3.9,
    7.87
   ],
   "64_tokens": [
    4.03,
    7.9
   ],
   "512_tokens": [
    4.74,
    8.32
   ]
  },
  "attention_collect_wait_us": {
   "8_tokens": [
    72.46,
    72.19
   ],
   "64_tokens": [
    161.31,
    151.14
   ],
   "512_tokens": [
    967.55,
    779.98
   ]
  },
  "ffn_combine_kernel_us": {
   "8_tokens": [
    46.62,
    47.86
   ],
   "64_tokens": [
    80.53,
    81.65
   ],
   "512_tokens": [
    318.85,
    320.89
   ]
  },
  "registers": "afdep_ffn_fp8_hybrid_combine and its sparse copy 120, afdep_fp8_combine_reduce_epilogue 72; no spills"
 },
 "fault_injection": {
  "method": "the section 11 deferred-put patches ported to the FP8 (and BF16) FFN combines and the Attention dispatch copies; replies to the last Attention domain held 300 us, dispatch to the last FFN domain 3 ms; device counters mark each deferral",
  "positive": "17 runs pass on 4F:12A and 8F:8A (hidden 2048/4096, capacity 64/2048, FP8, FP8 + sparse switches, minimum bucket 32 so formats switch under the fault) with 16k-65k deferrals per FFN rank",
  "negative": "dropping the deferred FP8 replies fails 2/2 runs: the last Attention domain's outputs mismatch (100%)",
  "limits": "a pass shows no failure under this delay, not that the delay reaches the section 11 overlap window here; BF16 dispatch and the section 17 read delay were not fault-tested"
 },
 "serving_16A_8F_4k": {
  "setup": "six 4-GPU pods: FFN DP8 over two pods, Attention DP16 over four; sparse switches; 4,096 in / 2,048 out; tree 2cee2f51; one arm after another",
  "fp8_off": [
   {
    "concurrency": 64,
    "requests_per_attention_gpu": 4.0,
    "tok_s_per_gpu": 107.0,
    "tpot_ms": 24.93
   },
   {
    "concurrency": 256,
    "requests_per_attention_gpu": 16.0,
    "tok_s_per_gpu": 346.7,
    "tpot_ms": 30.77
   },
   {
    "concurrency": 1536,
    "requests_per_attention_gpu": 96.0,
    "tok_s_per_gpu": 958.3,
    "tpot_ms": 66.79
   }
  ],
  "fp8_always": [
   {
    "concurrency": 64,
    "requests_per_attention_gpu": 4.0,
    "tok_s_per_gpu": 103.9,
    "tpot_ms": 25.66
   },
   {
    "concurrency": 256,
    "requests_per_attention_gpu": 16.0,
    "tok_s_per_gpu": 339.4,
    "tpot_ms": 31.42
   },
   {
    "concurrency": 1536,
    "requests_per_attention_gpu": 96.0,
    "tok_s_per_gpu": 968.2,
    "tpot_ms": 65.82
   }
  ],
  "fp8_min_bucket_64": [
   {
    "concurrency": 64,
    "requests_per_attention_gpu": 4.0,
    "tok_s_per_gpu": 106.0,
    "tpot_ms": 25.15
   },
   {
    "concurrency": 256,
    "requests_per_attention_gpu": 16.0,
    "tok_s_per_gpu": 345.4,
    "tpot_ms": 30.88
   },
   {
    "concurrency": 1536,
    "requests_per_attention_gpu": 96.0,
    "tok_s_per_gpu": 970.8,
    "tpot_ms": 65.88
   }
  ],
  "note": "a first A/B on 024e98b0 showed +36% TPOT at 4 and 16 requests per Attention GPU and +15% at 96, 0% at 32. It ran while wide EP32 prefilled 16K prompts on the same InfiniBand fabric and did not reproduce in the repeat above; its control's 64-request point also moved from 51.4 to 58.5 ms between runs"
 },
 "accuracy_gsm8k_16A_8F": {
  "setup": "vLLM 0.30.0 with the afd-plugin lanes tree (5215c41), Qwen3-235B-A22B-FP8, AFD 16A:8F on six 4-GPU pods (FFN DP8 over two pods, Attention DP16 over four, so replies cross NVLink domains), sparse switches, real prefill; tree b5078d62, whose native changes since 2cee2f51 are comments only; arms one after another on fresh servers",
  "method": "GSM8K test, 1,319 problems (openai/grade-school-math test.jsonl, sha256 3730d312f6e34405...); zero-shot chat with thinking off, question plus 'Please reason step by step, and put your final answer within \\boxed{}.'; greedy (temperature 0), at most 1,024 tokens, 256 requests in flight; the answer is the last \\boxed{} number, else the last number",
  "arms": [
   {
    "arm": "bf16_run1",
    "env": "AFDEP_HYBRID_FP8_COMBINE=0",
    "correct": 1257,
    "accuracy": 0.953,
    "truncated": 0
   },
   {
    "arm": "fp8_every_bucket",
    "env": "AFDEP_HYBRID_FP8_COMBINE=1 (minimum bucket 0)",
    "correct": 1254,
    "accuracy": 0.9507,
    "truncated": 1
   },
   {
    "arm": "bf16_run2",
    "env": "AFDEP_HYBRID_FP8_COMBINE=0",
    "correct": 1254,
    "accuracy": 0.9507,
    "truncated": 0
   }
  ],
  "paired_answers": {
   "bf16_run1_vs_run2": {
    "run1_only_correct": 19,
    "run2_only_correct": 16
   },
   "fp8_vs_bf16_run1": {
    "bf16_only_correct": 21,
    "fp8_only_correct": 18,
    "mcnemar_exact_p": 0.75
   },
   "fp8_vs_bf16_run2": {
    "bf16_only_correct": 20,
    "fp8_only_correct": 20,
    "mcnemar_exact_p": 1.0
   },
   "fp8_loss_95pct_interval_points": [
    -0.7,
    1.2
   ]
  },
  "greedy_text": {
   "identical_outputs": {
    "bf16_run1_vs_run2": 69,
    "fp8_vs_bf16_run1": 21,
    "fp8_vs_bf16_run2": 13
   },
   "median_first_divergence_chars": {
    "bf16_run1_vs_run2": 144,
    "fp8_vs_bf16_run1": 68,
    "fp8_vs_bf16_run2": 66
   },
   "note": "batching makes greedy decoding differ between runs, so text match is read against the BF16 run-to-run floor; the FP8 arm diverging earlier than that floor shows the switch changed the numerics"
  },
  "truncated_case": "problem 186 under FP8 ran out of 1,024 tokens in a trial-and-error search over candidate values (no repetition loop); both BF16 runs solved it",
  "result": "task accuracy unchanged within run-to-run noise; FP8 makes outputs diverge earlier than batching alone"
 },
 "scope": "communication-only correctness with synthetic expert outputs, transport timing, serving TPOT, and one task-level accuracy check (GSM8K on Qwen3-235B); no per-layer MoE error or next-token KL yet"
}