{
  "date": "2026-10-08",
  "source_commits": {
    "experts4bit-qlora": "72e6dfeb36dfefbfdb84d71b77b1399c64709581",
    "grouped-nf4-gemm": "6ee2e10408161a9d3c874975c9191a7f2957e6f4",
    "loggetta": "c30ad5a8f5a7bdafd981aa961a917378a8a65aa6"
  },
  "publication": "Dated research after the pinned runtime/kernel releases; does not change release support.",
  "evidence_urls": {
    "qwen_gpu_time": "https://github.com/pjordanandrsn/experts4bit-qlora/blob/72e6dfeb36dfefbfdb84d71b77b1399c64709581/bench/h2h-2026-10-02/tc1/RESULTS-tc1-pos69.md",
    "packed_4k_gpu_time": "https://github.com/pjordanandrsn/experts4bit-qlora/blob/72e6dfeb36dfefbfdb84d71b77b1399c64709581/bench/h2h-2026-10-02/tc1/RESULTS-tc1-pos68.md",
    "wide_decode_quality": "https://github.com/pjordanandrsn/experts4bit-qlora/blob/72e6dfeb36dfefbfdb84d71b77b1399c64709581/bench/p117/RESULTS-p117.md"
  },
  "previous_source_commits": {
    "experts4bit-qlora": "a41c857dd7d5c68f8735af16634ce08426e01271",
    "grouped-nf4-gemm": "c73ff5e0c35a0b0814376ae4631b35602dba319c",
    "loggetta": "ca0383dd9e45849cedc5a3f528ddea31a7930a31"
  },
  "revision": 1,
  "released_versions": {
    "loggetta": "0.3.1",
    "experts4bit-qlora": "0.50.0",
    "grouped-nf4-gemm": "0.43.0"
  },
  "claims": [
    {
      "id": "e4b.train.h2h.unsloth.qwen3.5090.2026-10-08.field-device-ratio",
      "package": "experts4bit-qlora",
      "area": "train",
      "model": "Qwen/Qwen3-30B-A3B @ ad44e777bcd1",
      "claim": "TC1 amendment 69 (#835), one rented RTX 5090 (tc1-5090-137, AMD EPYC 7713, Vast machine 55913): TC1's field recipe (alpaca, seq 2048, micro-batch 2 x accum 4) over 60 load-gated steps with both frameworks on one stack (venv-unsloth, torch 2.12.1), e4b's matched arm at every current default against Unsloth 2026.9.14's matched arm (grouped_mm), two draws each in ABBA order, every arm profiled. Unsloth's device time per profiled step is 1.924x e4b's (3259.1 / 1693.9 ms; P205 HELD, [1.7, 2.7]); s/step 9.780 against 3.489, a wall ratio of 2.803 on this host (P206 HELD, [2.3, 3.3]). Unsloth runs 14.09x e4b's CPU ops per step (P207 HELD, >= 8) and its GPU is busy for 0.333 of its timed step against e4b's 0.486, so the wall ratio is 1.457x the device ratio (P208 HELD, >= 1.05). Held-out at N 0.75643 against 0.75761 (P209 HELD, within 0.01). e4b peaks 26.16 GB against Unsloth's 24.27.",
      "value": 1.924,
      "unit": "ratio, device ms per profiled step, Unsloth matched arm / e4b matched arm at its defaults, TC1 field recipe (medians of two draws); wall s/step ratio 2.803 on this host",
      "hardware": "RTX 5090 (sm_120), one rented Vast.ai verified/secure host (tc1-5090-137: instance 54809007, AMD EPYC 7713, Vast machine 55913); e4b @ 737f02f + grouped-nf4-gemm @ 8371c38, torch 2.12.1+cu130; Unsloth 2026.9.14",
      "conditions": "bench/tc1 token qwen3pos69 (TC1 amendment 69): TC1's tokens and field recipe, TC1_STEPS=60, TC1_LOAD_GATE=6.0 / TC1_LOAD_RETRIES=2, e4b fused_attn4_m_fd and Unsloth ckpt_unsloth_m_fv, matched set (fp32 adapters, matched init), two draws each, the profile instrument (3 warm + 3 profiled steps, before the timed window) on every arm; validity (pos69_why: every e4b default present and unset) and scoring by bench/tc1/tc1_reduce.py",
      "measured_on": "2026-10-08",
      "status": "measured",
      "tier": "measured",
      "evidence": [
        "bench/h2h-2026-10-02/tc1/README.md",
        "bench/h2h-2026-10-02/tc1/RESULTS-tc1-pos69.md",
        "bench/h2h-2026-10-02/tc1/receipts/tc1-5090-137/RESULTS-tc1.md",
        "bench/tc1/TC1-PREREG.md",
        "bench/tc1/tc1_reduce.py"
      ],
      "notes": "Vast invoiced $2.12.",
      "evidence_urls": [
        "https://github.com/pjordanandrsn/experts4bit-qlora/blob/72e6dfeb36dfefbfdb84d71b77b1399c64709581/bench/h2h-2026-10-02/tc1/README.md",
        "https://github.com/pjordanandrsn/experts4bit-qlora/blob/72e6dfeb36dfefbfdb84d71b77b1399c64709581/bench/h2h-2026-10-02/tc1/RESULTS-tc1-pos69.md",
        "https://github.com/pjordanandrsn/experts4bit-qlora/blob/72e6dfeb36dfefbfdb84d71b77b1399c64709581/bench/h2h-2026-10-02/tc1/receipts/tc1-5090-137/RESULTS-tc1.md",
        "https://github.com/pjordanandrsn/experts4bit-qlora/blob/72e6dfeb36dfefbfdb84d71b77b1399c64709581/bench/tc1/TC1-PREREG.md",
        "https://github.com/pjordanandrsn/experts4bit-qlora/blob/72e6dfeb36dfefbfdb84d71b77b1399c64709581/bench/tc1/tc1_reduce.py"
      ]
    },
    {
      "id": "e4b.train.h2h.unsloth.qwen3.5090.2026-10-08.packed-4k-device-ratio",
      "package": "experts4bit-qlora",
      "area": "train",
      "model": "Qwen/Qwen3-30B-A3B @ ad44e777bcd1",
      "claim": "TC1 amendment 68 (#835), one rented RTX 5090 (tc1-5090-136, an AMD engineering-sample CPU, Vast machine 13828): e4b's matched arm at its defaults against Unsloth 2026.9.14's matched arm on packed rows of 4,096 real tokens, both in venv-unsloth (torch 2.12.1), two draws each in ABBA order, every arm profiled, 40 load-gated steps. Unsloth's device time per profiled step is 1.160x e4b's (11649.4 / 10042.8 ms; P200 HELD, [1.06, 1.24]; amendment 66's host read 1.141). Unsloth runs 7.09x e4b's CPU ops per step (P202 HELD, >= 5) and its GPU is busy for 0.783 of its timed step against e4b's 0.993 (P201 HELD, gap >= 0.08), so its wall time depends on the host far more than e4b's: s/step 14.886 against 10.111 here, a wall ratio of 1.472. Held-out at N within 0.0001 (P204 HELD). P203 FALSIFIED as registered: Unsloth's device time was 14.7 % above amendment 66's absolute reference, but e4b's rose 12.9 % on the same card, so the bound tested the card.",
      "value": 1.16,
      "unit": "ratio, device ms per profiled step, Unsloth matched arm / e4b matched arm at its defaults, packed 4,096-token rows (medians of two draws); wall s/step ratio 1.472 on this host",
      "hardware": "RTX 5090 (sm_120), one rented Vast.ai verified/secure host (tc1-5090-136: instance 54768641, an AMD engineering-sample CPU, a 30.7-CPU quota, Vast machine 13828); e4b @ 0b29952 + grouped-nf4-gemm @ 8944aa8, torch 2.12.1+cu130; Unsloth 2026.9.14",
      "conditions": "bench/tc1 token qwen3pos68 (TC1 amendment 68): packed rows (TC1_PACK=1, seq 4096, micro-batch 1 x accum 4, TC1_FREE_OUTPUTS=1), held-out at 0 and 40, TC1_LOAD_GATE=6.0 / TC1_LOAD_RETRIES=2, e4b fused_attn4_m_pd and Unsloth ckpt_unsloth_m_pv (grouped_mm), two draws each, --phase-peaks 1 and the profile instrument (3 warm + 3 profiled steps, before the timed window) on every arm; validity (pos68_why) and scoring by bench/tc1/tc1_reduce.py",
      "measured_on": "2026-10-08",
      "status": "measured",
      "tier": "measured",
      "evidence": [
        "bench/h2h-2026-10-02/tc1/README.md",
        "bench/h2h-2026-10-02/tc1/RESULTS-tc1-pos68.md",
        "bench/h2h-2026-10-02/tc1/receipts/tc1-5090-136/RESULTS-tc1.md",
        "bench/tc1/TC1-PREREG.md",
        "bench/tc1/tc1_reduce.py"
      ],
      "notes": "Vast invoiced $1.96.",
      "evidence_urls": [
        "https://github.com/pjordanandrsn/experts4bit-qlora/blob/72e6dfeb36dfefbfdb84d71b77b1399c64709581/bench/h2h-2026-10-02/tc1/README.md",
        "https://github.com/pjordanandrsn/experts4bit-qlora/blob/72e6dfeb36dfefbfdb84d71b77b1399c64709581/bench/h2h-2026-10-02/tc1/RESULTS-tc1-pos68.md",
        "https://github.com/pjordanandrsn/experts4bit-qlora/blob/72e6dfeb36dfefbfdb84d71b77b1399c64709581/bench/h2h-2026-10-02/tc1/receipts/tc1-5090-136/RESULTS-tc1.md",
        "https://github.com/pjordanandrsn/experts4bit-qlora/blob/72e6dfeb36dfefbfdb84d71b77b1399c64709581/bench/tc1/TC1-PREREG.md",
        "https://github.com/pjordanandrsn/experts4bit-qlora/blob/72e6dfeb36dfefbfdb84d71b77b1399c64709581/bench/tc1/tc1_reduce.py"
      ]
    },
    {
      "id": "e4b.serve.p117.wide-bucket-quality.qwen3.5090.2026-10-08",
      "package": "experts4bit-qlora",
      "area": "serve",
      "model": "Qwen/Qwen3-30B-A3B",
      "claim": "Lane P117, one RTX 5090: on SC2e's served stack (Qwen3-30B-A3B int4 experts and int4 RTN attention, the T=1 folds, fused q/k/v), decoding 32 or 64 rows in one graph (E4B_PAGED_BUCKETS=auto) costs no measurable quality against decode in pieces of at most 16 rows. Teacher-forced over 64 wikitext windows (512 + 128, decoded together) W32 reads +0.0006 nats, W64 -0.0032 and W64pad (48 windows padded to 64) -0.0033, inside the 16-row arithmetic's own neutral perturbations (half-batch -0.0025, prefill split -0.0007; B_floor 0.0025, S_floor 0.0148); the captured 64-row replay emits the eager W64 tokens at all 8,128 positions; a halved decode scale reads +1.147. AT_PARITY.",
      "value": -0.003174810658067649,
      "unit": "mean NLL bias of one 64-row decode piece (W64) against four 16-row pieces (R), nats, over 64 windows x 128 positions",
      "hardware": "RTX 5090 (sm_120, driver 595.91.07, 32607 MiB), one rented Vast.ai host (instance 54778599, AMD Ryzen 9 9950X, 32 threads, 123 GiB RAM), image pytorch/pytorch:2.8.0-cuda12.8-cudnn9-devel",
      "conditions": "e4b @9ea9528 (0.48.0 + P117's registration #1335) + grouped-nf4-gemm @b4f93f1 (v0.42.0); torch 2.8.0+cu128, triton 3.4.0, transformers 5.17.0; Qwen/Qwen3-30B-A3B @ad44e77, NF4 arena baked on the box, int4 stores repacked at load; premise tests/test_decode_graph_buckets.py 7 passed on the card; arms R (buckets 1-16), rep, floor (half buckets 1-8 / chunk 256 / rev), W32 (1-32), W64 (1-64), W64pad, mutant_scale, G64 (captured); every pass device-grouped, capture=False except G64; rule AT_PARITY iff W32, W64 and W64pad each have bias <= B_floor + 0.01 and spread <= 2 x max(S_floor, 0.005); verdict by bench/p117/p117_reduce.py",
      "measured_on": "2026-10-08",
      "status": "measured",
      "tier": "measured",
      "evidence": [
        "bench/p117/RESULTS-p117.md",
        "bench/p117/PREREG-p117.md",
        "bench/p117/p117_reduce.py",
        "bench/p117/p117_box.py",
        "bench/p117/receipts/p117-5090-1/verdict.json",
        "bench/p117/receipts/p117-5090-1/box.json",
        "bench/p117/receipts/p117-5090-1/summary.txt",
        "bench/p117/receipts/p117-5090-1/SHA256SUMS"
      ],
      "notes": "Pre-registered in #1335 (e4b#846), the read the ruling on #1320 required before E4B_PAGED_BUCKETS=auto may be a default. With SC2e (bench/h2h-2026-10-02/sc2e: 64 slots with buckets up to 64 serve 12 req/s, against 8 on the default list) this licenses E4B_PAGED_BUCKETS=auto as serve_paged's default. Read on one model and card; other models ride this read. Predictions held for R's repeat, the floor, FUNCTION, the mutant, engagement, the box's time and the verdict; missed for rev (bit-identical: 64 windows reverse into the same 16-row sets) and narrowly for W64 and W64pad's bias (-0.0032 / -0.0033 against [-0.003, +0.003], the better side). Lane cost $1.068.",
      "evidence_urls": [
        "https://github.com/pjordanandrsn/experts4bit-qlora/blob/72e6dfeb36dfefbfdb84d71b77b1399c64709581/bench/p117/RESULTS-p117.md",
        "https://github.com/pjordanandrsn/experts4bit-qlora/blob/72e6dfeb36dfefbfdb84d71b77b1399c64709581/bench/p117/PREREG-p117.md",
        "https://github.com/pjordanandrsn/experts4bit-qlora/blob/72e6dfeb36dfefbfdb84d71b77b1399c64709581/bench/p117/p117_reduce.py",
        "https://github.com/pjordanandrsn/experts4bit-qlora/blob/72e6dfeb36dfefbfdb84d71b77b1399c64709581/bench/p117/p117_box.py",
        "https://github.com/pjordanandrsn/experts4bit-qlora/blob/72e6dfeb36dfefbfdb84d71b77b1399c64709581/bench/p117/receipts/p117-5090-1/verdict.json",
        "https://github.com/pjordanandrsn/experts4bit-qlora/blob/72e6dfeb36dfefbfdb84d71b77b1399c64709581/bench/p117/receipts/p117-5090-1/box.json",
        "https://github.com/pjordanandrsn/experts4bit-qlora/blob/72e6dfeb36dfefbfdb84d71b77b1399c64709581/bench/p117/receipts/p117-5090-1/summary.txt",
        "https://github.com/pjordanandrsn/experts4bit-qlora/blob/72e6dfeb36dfefbfdb84d71b77b1399c64709581/bench/p117/receipts/p117-5090-1/SHA256SUMS"
      ]
    }
  ]
}