{ "generated_utc": "2026-11-13T11:33:30Z", "spec_sha256": "fda0f7ad167e19c11d51a5e091af9ad8ebbd59c6bf6bca5d16967602a8af0140", "spec": { "name": "DLGM #458 vs Fusion #879: one combined index", "version": 1, "author": "claude-fable52 (fable52-asi-fusion-20261112), for Scott", "frozen_at_utc": "2026-11-12T01:59:02Z", "frozen_note": "Weights, scoring rule and constants fixed while Fusion seed 1 was in its first minutes on the GPU queue and before the author had read any Fusion result under this objective.", "candidates": { "dlgm": { "label": "DLGM #458", "upstream": "ASI-Arch #458 DualScaleMemoryRouter (delta_net_dlgm)", "faithful": "dlgm458", "lowrank": "dlgm458_lr" }, "fusion": { "label": "Fusion #879", "upstream": "ASI-Arch #879 FusionGatedFIRNet (delta_net_psfr)", "faithful": "fusion879", "lowrank": "fusion879_lr" } }, "baseline_runs": [ "base", "base_r2", "base_r3" ], "quality_mix": { "ar": 0.6, "sat": 0.2, "nat": 0.2 }, "components": [ { "key": "own_steps", "weight": 0.35, "floor": 0.02, "unit": "nats", "title": "Your objective, same number of training steps", "measure": "Held-out loss after 1,200 steps in the AGILLM-objective simulator (0.6 AR + 0.2 SAT + 0.2 NAT), gain over the live block. Faithful form, all complete seeds." }, { "key": "own_long_context", "weight": 0.05, "floor": 0.02, "unit": "nats", "title": "Your objective, long context", "measure": "What 4x more context does to the loss (1,024 tokens against 256) at step 1,200, relative to the live block. Faithful form." }, { "key": "own_time", "weight": 0.1, "floor": 0.02, "unit": "nats", "title": "Your objective, same simulator box time", "measure": "The site's equal-time method (step cost = 0.75 x bytes streamed + 0.25 x measured ms per step, against the base run of the same day), gain over the live block. Low-rank form, the one that fits at full scale." }, { "key": "own_small", "weight": 0.15, "floor": 0.02, "unit": "nats per byte", "title": "Your matched small-model run (v65 R6)", "measure": "Held-out cross-entropy gain over the width-96 Transformer at an equal 120 training-CPU-seconds, about 350k parameters, 3 paired seeds; WikiText and FineWeb scored separately and averaged." }, { "key": "up_tasks", "weight": 0.15, "floor": null, "unit": "accuracy points", "title": "ASI-Arch paper: benchmark tasks", "measure": "7-task average, gain over the Gated DeltaNet baseline; the 20M and 340M runs scored separately and averaged. Single runs." }, { "key": "up_loss", "weight": 0.1, "floor": null, "unit": "nats", "title": "ASI-Arch paper: training loss", "measure": "Final training loss, gain over the Gated DeltaNet baseline; the 20M and 340M runs scored separately and averaged. Single runs." }, { "key": "serving", "weight": 0.1, "floor": null, "unit": "rubric", "title": "Serving readiness", "measure": "Rubric: 100 = an exact incremental cache verified against full recompute on trained checkpoints; 50 = a cache that reproduces the same tokens but failed the exact-match gate; 0 = none." } ], "scoring": "For a gain component each candidate scores 100 x max(gain, 0) / max(gain of DLGM, gain of Fusion, floor). The better one gets 100 when its gain clears the noise floor, the other gets its share of that gain, and a candidate that does not beat the shared baseline gets 0. Where a component has sub-experiments they are scored separately and averaged. Index = weighted sum of the component scores (0 to 100).", "headline_rule": "The name on top is the higher index. A margin under 3 points is called level in words.", "robustness": { "random_weightings": 20000, "seed": 879458, "method": "weights drawn uniformly from the simplex over the available components; also every leave-one-component-out index" }, "not_scored": [ "Footprint and arithmetic at 560 layers (reported: the two are equal to well under 1%)", "Only Fusion has 1.3B-parameter upstream results (reported)" ], "constants": { "v65_r6": { "source": "SG memory slot 2342 (ChatGPT lane, 2026-11-09), natural-log cross-entropy per byte", "wikitext": { "dlgm": 1.762221473, "fusion": 1.771213003, "baseline": 1.829924379 }, "fineweb": { "dlgm": 2.103326358, "fusion": 2.124353456, "baseline": 2.188070645 }, "per_seed_fusion_minus_dlgm": { "wikitext": [ -0.005230576, 0.002618733, 0.029586433 ], "fineweb": [ -0.001649001, -0.001725079, 0.066455373 ] } }, "upstream": { "source": "ASI-Arch paper (arXiv 2507.18074v2) Table 9 and data/106.json at commit 3113c51b", "tasks7": { "20M": { "dlgm": 37.74, "fusion": 39.37, "baseline": 36.6 }, "340M": { "dlgm": 44.13, "fusion": 44.02, "baseline": 42.1 } }, "train_loss": { "20M": { "dlgm": 4.39, "fusion": 4.4233, "baseline": 4.5678 }, "340M": { "dlgm": 3.4047, "fusion": 3.4048, "baseline": 3.4768 } }, "note_1p3b": "At 1.3B parameters only Fusion was trained upstream: 52.01 on 14 accuracy metrics against 51.04 for DeltaNet." }, "serving": { "dlgm": 50, "fusion": 100, "dlgm_reason": "Incremental cache attempts v63 and v64 reproduced the same tokens (RMS error about 1e-6) but failed the strict exact-match gate twice; the Transformer baseline failed the same gate.", "fusion_reason": "Exact incremental cache verified against full recompute on three trained checkpoints (worst error 1.3e-14 in float64, 1.0e-5 in float32); fixed state of 103,680 bytes; public at experiments/asi_arch_psfr_cache_v1." } } }, "index": { "dlgm": 74.26304561033795, "fusion": 85.4893848167856 }, "top": "fusion", "margin": 11.226339206447648, "provisional": false, "have_own": true, "verdict": "Fusion #879 is on top of the index: 85.5 to 74.3, a lead of 11.2 points. It stays on top under 86% of 20,000 random re-weightings of the same evidence, so the result does not hinge on my choice of weights (a sensitivity check, not a statistical confidence).", "paired_seed_intervals": { "faithful": { "seeds": [ 1, 2, 3 ], "diffs": [ 0.04320305585861117, 0.0077471137046813965, 0.02234553098678571 ], "mean": 0.024431900183359428, "lo": -0.019838536922366836, "hi": 0.06870233728908569, "n": 3, "positive": 3 }, "lowrank": { "seeds": [ 1, 2, 3 ], "diffs": [ -0.004479908943175737, -0.0012068748474121094, -0.03876460790634084 ], "mean": -0.014817130565642897, "lo": -0.06650028911223943, "hi": 0.03686602798095363, "n": 3, "positive": 0 }, "lrk": { "seeds": [ 1, 2, 3 ], "diffs": [ 0.022443425655365168, -0.026388692855834783, -0.02135441303253227 ], "mean": -0.008433226744333963, "lo": -0.07515802526855948, "hi": 0.05829157177989156, "n": 3, "positive": 1 }, "lg": { "seeds": [ 1, 2, 3 ], "diffs": [ 0.0006412029266353869, -0.03822736740112376, 0.020174694061280363 ], "mean": -0.005803823471069336, "lo": -0.07966240746131747, "hi": 0.0680547605191788, "n": 3, "positive": 2 }, "hn": { "seeds": [ 1, 2, 3 ], "diffs": [ 0.009809410572051647, 0.04103357791900564, 0.013680362701415483 ], "mean": 0.021507783730824254, "lo": -0.020776247133680387, "hi": 0.0637918145953289, "n": 3, "positive": 3 } }, "trained_minus_fixed_projection": { "dlgm": { "seeds": [ 1, 2, 3 ], "diffs": [ 0.041070604324341176, 0.011099004745483754, -0.061036348342896396 ], "mean": -0.0029555797576904888, "lo": -0.13334443118739053, "hi": 0.12743327167200957, "n": 3, "positive": 2 }, "fusion": { "seeds": [ 1, 2, 3 ], "diffs": [ 0.019268381595611395, -0.0007396697998052204, -0.019507241249083762 ], "mean": -0.0003261764844258626, "lo": -0.048500272282334625, "hi": 0.047847919313482895, "n": 3, "positive": 1 } }, "superiority_proven": false, "full_scale_measured": false, "components": [ { "key": "own_steps", "title": "Your objective, same number of training steps", "weight": 0.35, "unit": "nats", "measure": "Held-out loss after 1,200 steps in the AGILLM-objective simulator (0.6 AR + 0.2 SAT + 0.2 NAT), gain over the live block. Faithful form, all complete seeds.", "floor": 0.02, "score": { "dlgm": 81.93258003159994, "fusion": 100.0 }, "parts": [ { "name": "simulator, faithful form", "dlgm": 0.11079438130060915, "fusion": 0.13522628148396856, "seeds": { "dlgm": 3, "fusion": 3 }, "se": { "dlgm": 0.006967905627775838, "fusion": 0.003548613814228284 } } ], "pending": null }, { "key": "own_long_context", "title": "Your objective, long context", "weight": 0.05, "unit": "nats", "measure": "What 4x more context does to the loss (1,024 tokens against 256) at step 1,200, relative to the live block. Faithful form.", "floor": 0.02, "score": { "dlgm": 100.0, "fusion": 93.92328593027189 }, "parts": [ { "name": "simulator, faithful form", "dlgm": 0.24948746817452566, "fusion": 0.23432682809375582, "seeds": { "dlgm": 3, "fusion": 3 }, "se": null } ], "pending": null }, { "key": "own_time", "title": "Your objective, same simulator box time", "weight": 0.1, "unit": "nats", "measure": "The site's equal-time method (step cost = 0.75 x bytes streamed + 0.25 x measured ms per step, against the base run of the same day), gain over the live block. Low-rank form, the one that fits at full scale.", "floor": 0.02, "score": { "dlgm": 0.0, "fusion": 0.0 }, "parts": [ { "name": "simulator, low-rank form", "dlgm": -0.04696765100960131, "fusion": -0.15047664248354717, "seeds": { "dlgm": 3, "fusion": 3 }, "se": null } ], "pending": null }, { "key": "own_small", "title": "Your matched small-model run (v65 R6)", "weight": 0.15, "unit": "nats per byte", "measure": "Held-out cross-entropy gain over the width-96 Transformer at an equal 120 training-CPU-seconds, about 350k parameters, 3 paired seeds; WikiText and FineWeb scored separately and averaged.", "floor": 0.02, "score": { "dlgm": 100.0, "fusion": 80.95336478825472 }, "parts": [ { "name": "wikitext", "dlgm": 0.06770290599999984, "fusion": 0.05871137599999998 }, { "name": "fineweb", "dlgm": 0.08474428699999992, "fusion": 0.06371718899999967 } ], "pending": null }, { "key": "up_tasks", "title": "ASI-Arch paper: benchmark tasks", "weight": 0.15, "unit": "accuracy points", "measure": "7-task average, gain over the Gated DeltaNet baseline; the 20M and 340M runs scored separately and averaged. Single runs.", "floor": null, "score": { "dlgm": 70.57761732851989, "fusion": 97.29064039408868 }, "parts": [ { "name": "20M parameters", "dlgm": 1.1400000000000006, "fusion": 2.769999999999996 }, { "name": "340M parameters", "dlgm": 2.030000000000001, "fusion": 1.9200000000000017 } ], "pending": null }, { "key": "up_loss", "title": "ASI-Arch paper: training loss", "weight": 0.1, "unit": "nats", "measure": "Final training loss, gain over the Gated DeltaNet baseline; the 20M and 340M runs scored separately and averaged. Single runs.", "floor": null, "score": { "dlgm": 100.0, "fusion": 90.56619742920486 }, "parts": [ { "name": "20M parameters", "dlgm": 0.1778000000000004, "fusion": 0.14449999999999985 }, { "name": "340M parameters", "dlgm": 0.07209999999999983, "fusion": 0.07200000000000006 } ], "pending": null }, { "key": "serving", "title": "Serving readiness", "weight": 0.1, "unit": "rubric", "measure": "Rubric: 100 = an exact incremental cache verified against full recompute on trained checkpoints; 50 = a cache that reproduces the same tokens but failed the exact-match gate; 0 = none.", "floor": null, "score": { "dlgm": 50.0, "fusion": 100.0 }, "parts": [ { "name": "rubric", "dlgm": 50, "fusion": 100 } ], "pending": null } ], "contributions": [ { "key": "own_steps", "title": "Your objective, same number of training steps", "points": 6.323596988940021 }, { "key": "own_long_context", "title": "Your objective, long context", "points": -0.30383570348640565 }, { "key": "own_time", "title": "Your objective, same simulator box time", "points": 0.0 }, { "key": "own_small", "title": "Your matched small-model run (v65 R6)", "points": -2.8569952817617925 }, { "key": "up_tasks", "title": "ASI-Arch paper: benchmark tasks", "points": 4.0069534598353185 }, { "key": "up_loss", "title": "ASI-Arch paper: training loss", "points": -0.9433802570795138 }, { "key": "serving", "title": "Serving readiness", "points": 5.0 } ], "robustness": { "draws": 20000, "share": { "dlgm": 0.13975, "fusion": 0.86025 }, "leave_one_out": [ { "without": "own_steps", "title": "Your objective, same number of training steps", "dlgm": 70.1332963065815, "fusion": 77.6759766412086, "top": "fusion" }, { "without": "own_long_context", "title": "Your objective, long context", "dlgm": 72.90846906351364, "fusion": 85.04549528449685, "top": "fusion" }, { "without": "own_time", "title": "Your objective, same simulator box time", "dlgm": 82.51449512259774, "fusion": 94.988205351984, "top": "fusion" }, { "without": "own_small", "title": "Your matched small-model run (v65 R6)", "dlgm": 69.72123012980936, "fusion": 86.28985893946752, "top": "fusion" }, { "without": "up_tasks", "title": "ASI-Arch paper: benchmark tasks", "dlgm": 74.91341530712938, "fusion": 83.40681030314387, "top": "fusion" }, { "without": "up_loss", "title": "ASI-Arch paper: training loss", "dlgm": 71.40338401148661, "fusion": 84.9252945265168, "top": "fusion" }, { "without": "serving", "title": "Serving readiness", "dlgm": 76.95893956704217, "fusion": 83.87709424087289, "top": "fusion" } ] }, "alternative_forms": { "lowrank": { "form": "low-rank form", "index": { "dlgm": 80.58664259927798, "fusion": 80.70259049862094 }, "margin": 0.1159478993429559, "top": "fusion", "gain": { "dlgm": 0.11663459936777798, "fusion": 0.10181746880213509 }, "seeds": { "dlgm": 3, "fusion": 3 }, "se": { "dlgm": 0.001899975283621722, "fusion": 0.012862404698486806 }, "complete": true }, "lrk": { "form": "low-rank gate, per-channel kernels", "index": { "dlgm": 80.58664259927798, "fusion": 82.97692416859256 }, "margin": 2.390281569314581, "top": "fusion", "gain": { "dlgm": 0.1203629692395533, "fusion": 0.11192974249521932 }, "seeds": { "dlgm": 3, "fusion": 3 }, "se": { "dlgm": 0.01127159307344835, "fusion": 0.004273733658300246 }, "complete": true }, "lg": { "form": "low-rank gate with a trained projection, per-channel kernels", "index": { "dlgm": 80.58664259927798, "fusion": 83.34841638081593 }, "margin": 2.761773781537954, "top": "fusion", "gain": { "dlgm": 0.1174073894818628, "fusion": 0.11160356601079346 }, "seeds": { "dlgm": 3, "fusion": 3 }, "se": { "dlgm": 0.0248580577227395, "fusion": 0.015109449956207239 }, "complete": true }, "hn": { "form": "per-head norm check", "index": { "dlgm": 74.25139652036944, "fusion": 85.793220520272 }, "margin": 11.541823999902562, "top": "fusion", "gain": { "dlgm": 0.10185121297836375, "fusion": 0.123358996709188 }, "seeds": { "dlgm": 3, "fusion": 3 }, "se": { "dlgm": 0.004449143999394207, "fusion": 0.006023065063530372 }, "complete": true } }, "own": { "base": { "ids": [ "base", "base_r2", "base_r3" ], "seeds": 7, "distinct_seeds": 3, "ar": 5.242889966283526, "sat": 5.464196605341775, "nat": 5.8445658683776855, "q": 5.407486474514008, "q_sd": 0.009209629858174919, "ctx_gain": -0.31628663199288504, "ms_per_step": 202.34329359872, "per_day": { "base": { "seeds": 3, "ms_per_step": 194.1464344660441, "started": 1794199733.7860162 }, "base_r2": { "seeds": 3, "ms_per_step": 213.17382653554282, "started": 1794376000.6129856 }, "base_r3": { "seeds": 1, "ms_per_step": 194.4422721862793, "started": 1794449015.1553707 } } }, "forms": { "dlgm458": { "id": "dlgm458", "cand": "dlgm", "form": "faithful", "seeds": 3, "seeds_list": [ 1, 2, 3 ], "ar": 5.096494356791179, "sat": 5.3806353608767195, "nat": 5.813342034816742, "ctx_gain": -0.06679916381835938, "gain_steps": 0.11079438130060915, "gain_sd": 0.012068766569652865, "gain_se": 0.006967905627775838, "gain_seeds": [ 0.09906218051910542, 0.12317367792129552, 0.11014728546142649 ], "gain_by_seed": { "1": 0.09906218051910542, "2": 0.12317367792129552, "3": 0.11014728546142649 }, "ctx_vs_base": 0.24948746817452566, "ms_per_step": 461.06282869974774, "ms_seeds": [ 500.8518695831299, 538.114070892334, 344.2225456237793 ], "base_same_day": [ "base" ], "base_ms": 194.1464344660441, "ms_ratio": 2.374819964980541, "byte_ratio": 1.0165921861864518, "step_ratio": 1.356149130884974, "tokens_eq": 906095.0392661679, "gain_time": -0.003420383028652908, "d_time": { "ar": 0.008720416797797448, "sat": -0.02492029755121017, "nat": -0.01834286798544671 }, "census": { "latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 39820506467, "kv_bytes_per_token": 24903680 }, "max_gpu_mib": 563, "causal_ok": true, "sim_sha": [ "dc5c3c28ddbf1853" ], "ld_mean": 6.31711, "started": 1794205654.7756302, "skipped": 0 }, "dlgm458_lr": { "id": "dlgm458_lr", "cand": "dlgm", "form": "lowrank", "seeds": 3, "seeds_list": [ 1, 2, 3 ], "ar": 5.09007853269577, "sat": 5.379696806271871, "nat": 5.804326971371968, "ctx_gain": -0.0886243184407552, "gain_steps": 0.11663459936777798, "gain_sd": 0.00329085372435791, "gain_se": 0.001899975283621722, "gain_seeds": [ 0.11409821510314977, 0.12035325765609795, 0.11545232534408623 ], "gain_by_seed": { "1": 0.11409821510314977, "2": 0.12035325765609795, "3": 0.11545232534408623 }, "ctx_vs_base": 0.22766231355212985, "ms_per_step": 542.5885915756226, "ms_seeds": [ 530.7273864746094, 542.2871112823486, 554.7512769699097 ], "base_same_day": [ "base" ], "base_ms": 194.1464344660441, "ms_ratio": 2.7947388942160587, "byte_ratio": 1.0100084067076678, "step_ratio": 1.4561910285847657, "tokens_eq": 843845.3306461027, "gain_time": -0.04696765100960131, "d_time": { "ar": -0.04350904453303084, "sat": -0.06228156242071403, "nat": -0.04202955902820005 }, "census": { "latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 109016867, "kv_bytes_per_token": 24903680 }, "max_gpu_mib": 542, "causal_ok": true, "sim_sha": [ "dc5c3c28ddbf1853" ], "ld_mean": 6.320846666666667, "started": 1794204421.818115, "skipped": 0 }, "dlgm458_lrk": { "id": "dlgm458_lrk", "cand": "dlgm", "form": "lrk", "seeds": 3, "seeds_list": [ 1, 2, 3 ], "ar": 5.085072795550029, "sat": 5.372280518213908, "nat": 5.808118621508281, "ctx_gain": -0.0705954631169637, "gain_steps": 0.1203629692395533, "gain_sd": 0.019522971885453978, "gain_se": 0.01127159307344835, "gain_seeds": [ 0.09803191423416191, 0.13420046567916888, 0.12885652780532908 ], "gain_by_seed": { "1": 0.09803191423416191, "2": 0.13420046567916888, "3": 0.12885652780532908 }, "ctx_vs_base": 0.24569116887592135, "ms_per_step": 411.5292231241862, "ms_seeds": [ 418.8896417617798, 408.05137157440186, 407.64665603637695 ], "base_same_day": [ "base_r3" ], "base_ms": 194.4422721862793, "ms_ratio": 2.11645964890769, "byte_ratio": 1.0102896442269542, "step_ratio": 1.2868321453971383, "tokens_eq": 954903.096254851, "gain_time": 0.012888796601531104, "d_time": { "ar": 0.02774354990215233, "sat": -0.009556191631975963, "nat": -0.009230475066825505 }, "census": { "latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 224840867, "kv_bytes_per_token": 24903680 }, "max_gpu_mib": 544, "causal_ok": true, "sim_sha": [ "dc5c3c28ddbf1853" ], "ld_mean": 6.349620000000001, "started": 1794454860.3268344, "skipped": 0 }, "dlgm458_lg": { "id": "dlgm458_lg", "cand": "dlgm", "form": "lg", "seeds": 3, "seeds_list": [ 1, 2, 3 ], "ar": 5.090413510799408, "sat": 5.372284630934398, "nat": 5.806870261828105, "ctx_gain": -0.05846289793650309, "gain_steps": 0.1174073894818628, "gain_sd": 0.043055418953264715, "gain_se": 0.0248580577227395, "gain_seeds": [ 0.13910251855850309, 0.14529947042465263, 0.06782017946243268 ], "gain_by_seed": { "1": 0.13910251855850309, "2": 0.14529947042465263, "3": 0.06782017946243268 }, "ctx_vs_base": 0.2578237340563819, "ms_per_step": 321.2440808614095, "ms_seeds": [ 317.42703914642334, 323.14610481262207, 323.1590986251831 ], "base_same_day": [ "base_r3" ], "base_ms": 194.4422721862793, "ms_ratio": 1.652130872826109, "byte_ratio": 1.0088524445620068, "step_ratio": 1.1696720516280323, "tokens_eq": 1050550.8773075917, "gain_time": 0.06742646113751202, "d_time": { "ar": 0.09417499145206776, "sat": 0.0392256812436953, "nat": 0.01538165008766157 }, "census": { "latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 299551907, "kv_bytes_per_token": 24903680 }, "max_gpu_mib": 545, "causal_ok": true, "sim_sha": [ "dc5c3c28ddbf1853" ], "ld_mean": 6.34036, "started": 1794488100.6403992, "skipped": 0 }, "dlgm458_hn": { "id": "dlgm458_hn", "cand": "dlgm", "form": "hn", "seeds": 3, "seeds_list": [ 1, 2, 3 ], "ar": 5.109130680561066, "sat": 5.389476478099823, "nat": 5.811307787895203, "ctx_gain": -0.07320030530293782, "gain_steps": 0.10185121297836375, "gain_sd": 0.00770614345714096, "gain_se": 0.004449143999394207, "gain_seeds": [ 0.10957328081131035, 0.09416109323501676, 0.10181926488876414 ], "gain_by_seed": { "1": 0.10957328081131035, "2": 0.09416109323501676, "3": 0.10181926488876414 }, "ctx_vs_base": 0.24308632668994723, "ms_per_step": 361.5771532058716, "ms_seeds": [ 421.20420932769775, 329.83553409576416, 333.69171619415283 ], "base_same_day": [ "base_r3" ], "base_ms": 194.4422721862793, "ms_ratio": 1.8595604193489057, "byte_ratio": 1.0115857690781647, "step_ratio": 1.22357943164585, "tokens_eq": 1004266.6362470051, "gain_time": 0.017726394274140134, "d_time": { "ar": 0.03333417234951508, "sat": -0.0026497812322592296, "nat": -0.008720764445585338 }, "census": { "latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 39820506467, "kv_bytes_per_token": 24903680 }, "max_gpu_mib": 563, "causal_ok": true, "sim_sha": [ "dc5c3c28ddbf1853" ], "ld_mean": 6.303951666666666, "started": 1794457697.1703331, "skipped": 0 }, "fusion879": { "id": "fusion879", "cand": "fusion", "form": "faithful", "seeds": 3, "seeds_list": [ 1, 2, 3 ], "ar": 5.071844935417175, "sat": 5.359661897023519, "nat": 5.786104261875153, "ctx_gain": -0.08195980389912923, "gain_steps": 0.13522628148396856, "gain_sd": 0.006146379422684173, "gain_se": 0.003548613814228284, "gain_seeds": [ 0.1422652363777166, 0.13092079162597692, 0.1324928164482122 ], "gain_by_seed": { "1": 0.1422652363777166, "2": 0.13092079162597692, "3": 0.1324928164482122 }, "ctx_vs_base": 0.23432682809375582, "ms_per_step": 370.39339542388916, "ms_seeds": [ 290.7313108444214, 291.6107177734375, 528.8381576538086 ], "base_same_day": [ "base_r3" ], "base_ms": 194.4422721862793, "ms_ratio": 1.90490160014714, "byte_ratio": 1.0084644888242298, "step_ratio": 1.2325737666549572, "tokens_eq": 996938.3036074192, "gain_time": 0.03187247215024751, "d_time": { "ar": 0.05450640171746013, "sat": 0.00495994869468408, "nat": -0.009116793095826914 }, "census": { "latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 41081291747, "kv_bytes_per_token": 24903680 }, "max_gpu_mib": 565, "causal_ok": true, "sim_sha": [ "dc5c3c28ddbf1853" ], "ld_mean": 6.3521350000000005, "started": 1794449191.0464382, "skipped": 0 }, "fusion879_lr": { "id": "fusion879_lr", "cand": "fusion", "form": "lowrank", "seeds": 3, "seeds_list": [ 1, 2, 3 ], "ar": 5.107352991898854, "sat": 5.390841384728749, "nat": 5.815444668134053, "ctx_gain": -0.117959459622701, "gain_steps": 0.10181746880213509, "gain_sd": 0.02227833844529179, "gain_se": 0.012862404698486806, "gain_seeds": [ 0.10961830615997403, 0.11914638280868584, 0.07668771743774538 ], "gain_by_seed": { "1": 0.10961830615997403, "2": 0.11914638280868584, "3": 0.07668771743774538 }, "ctx_vs_base": 0.19832717237018405, "ms_per_step": 743.5529629389445, "ms_seeds": [ 671.6419458389282, 798.8896369934082, 760.1273059844971 ], "base_same_day": [ "base_r3" ], "base_ms": 194.4422721862793, "ms_ratio": 3.8240293871211657, "byte_ratio": 1.0090552396067543, "step_ratio": 1.7127987764853572, "tokens_eq": 717422.2780106622, "gain_time": -0.15047664248354717, "d_time": { "ar": -0.1691068337559516, "sat": -0.1609021986964967, "nat": -0.0841605124533844 }, "census": { "latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 114233507, "kv_bytes_per_token": 24903680 }, "max_gpu_mib": 543, "causal_ok": true, "sim_sha": [ "dc5c3c28ddbf1853" ], "ld_mean": 6.294751666666667, "started": 1794451417.762637, "skipped": 0 }, "fusion879_lrk": { "id": "fusion879_lrk", "cand": "fusion", "form": "lrk", "seeds": 3, "seeds_list": [ 1, 2, 3 ], "ar": 5.098268588383992, "sat": 5.383000214894612, "nat": 5.799977680047353, "ctx_gain": -0.08848289648691814, "gain_steps": 0.11192974249521932, "gain_sd": 0.0074023238341932324, "gain_se": 0.004273733658300246, "gain_seeds": [ 0.12047533988952708, 0.1078117728233341, 0.10750211477279681 ], "gain_by_seed": { "1": 0.12047533988952708, "2": 0.1078117728233341, "3": 0.10750211477279681 }, "ctx_vs_base": 0.2278037355059669, "ms_per_step": 393.4518098831177, "ms_seeds": [ 521.0442543029785, 327.5418281555176, 331.76934719085693 ], "base_same_day": [ "base_r3" ], "base_ms": 194.4422721862793, "ms_ratio": 2.023489056464962, "byte_ratio": 1.005387294449588, "step_ratio": 1.2599127349534316, "tokens_eq": 975305.6429304355, "gain_time": 0.007785049140592547, "d_time": { "ar": 0.024047007585075164, "sat": -0.018011348848591346, "nat": -0.015204428203671405 }, "census": { "latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 294918947, "kv_bytes_per_token": 24903680 }, "max_gpu_mib": 545, "causal_ok": true, "sim_sha": [ "dc5c3c28ddbf1853" ], "ld_mean": 6.303936666666668, "started": 1794454323.621781, "skipped": 0 }, "fusion879_lg": { "id": "fusion879_lg", "cand": "fusion", "form": "lg", "seeds": 3, "seeds_list": [ 1, 2, 3 ], "ar": 5.097694794336955, "sat": 5.382589836915334, "nat": 5.803740322589874, "ctx_gain": -0.09531323115030925, "gain_steps": 0.11160356601079346, "gain_sd": 0.026170334998570285, "gain_se": 0.015109449956207239, "gain_seeds": [ 0.13974372148513847, 0.10707210302352888, 0.08799487352371305 ], "gain_by_seed": { "1": 0.13974372148513847, "2": 0.10707210302352888, "3": 0.08799487352371305 }, "ctx_vs_base": 0.22097340084257577, "ms_per_step": 280.0167401631673, "ms_seeds": [ 279.8255681991577, 279.296875, 280.92777729034424 ], "base_same_day": [ "base_r3" ], "base_ms": 194.4422721862793, "ms_ratio": 1.440102180532565, "byte_ratio": 1.0070713750385758, "step_ratio": 1.1153290764120731, "tokens_eq": 1101737.61806063, "gain_time": 0.06582635051362437, "d_time": { "ar": 0.09544949683010362, "sat": 0.039539397833976864, "nat": 0.0032438642438341603 }, "census": { "latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 369629987, "kv_bytes_per_token": 24903680 }, "max_gpu_mib": 546, "causal_ok": true, "sim_sha": [ "dc5c3c28ddbf1853" ], "ld_mean": 6.313475, "started": 1794481858.6687171, "skipped": 0 }, "fusion879_hn": { "id": "fusion879_hn", "cand": "fusion", "form": "hn", "seeds": 3, "seeds_list": [ 1, 2, 3 ], "ar": 5.083030541737874, "sat": 5.370459159215291, "nat": 5.801086604595184, "ctx_gain": -0.06132117907206217, "gain_steps": 0.123358996709188, "gain_sd": 0.010432254707327672, "gain_se": 0.006023065063530372, "gain_seeds": [ 0.119382691383362, 0.1351946711540224, 0.11549962759017962 ], "gain_by_seed": { "1": 0.119382691383362, "2": 0.1351946711540224, "3": 0.11549962759017962 }, "ctx_vs_base": 0.25496545292082284, "ms_per_step": 317.21266110738117, "ms_seeds": [ 366.3138151168823, 292.2466993331909, 293.0774688720703 ], "base_same_day": [ "base_r3" ], "base_ms": 194.4422721862793, "ms_ratio": 1.6313976253244231, "byte_ratio": 1.0086055636379672, "step_ratio": 1.164303579059581, "tokens_eq": 1055394.8489899114, "gain_time": 0.05620047119029117, "d_time": { "ar": 0.07986434251431085, "sat": 0.028798838978344676, "nat": 0.012610489430178617 }, "census": { "latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 41081291747, "kv_bytes_per_token": 24903680 }, "max_gpu_mib": 565, "causal_ok": true, "sim_sha": [ "dc5c3c28ddbf1853" ], "ld_mean": 6.327065, "started": 1794457306.9480817, "skipped": 0 } } }, "queue": { "state": "queue empty", "current": null, "last_update": 1794569598.2358065, "progress": { "fusion879": { "done": [ 1, 2, 3 ], "running": {}, "failed": [] }, "fusion879_lr": { "done": [ 1, 2, 3 ], "running": {}, "failed": [] }, "base_r3": { "done": [ 1 ], "running": {}, "failed": [] }, "fusion879_lrk": { "done": [ 1, 2, 3 ], "running": {}, "failed": [] }, "dlgm458_lrk": { "done": [ 1, 2, 3 ], "running": {}, "failed": [] }, "fusion879_lg": { "done": [ 1, 2, 3 ], "running": {}, "failed": [] }, "dlgm458_lg": { "done": [ 1, 2, 3 ], "running": {}, "failed": [] }, "fusion879_hn": { "done": [ 1, 2, 3 ], "running": {}, "failed": [] }, "dlgm458_hn": { "done": [ 1, 2, 3 ], "running": {}, "failed": [] } } }, "fusion_state": [ { "fusion_state_step": 150, "lambda_min": 0.99996, "lambda_mean": 0.99996, "temp_min": 0.9769, "temp_max": 1.0173, "mean_gate_short_long_delta_value": [ 0.2567, 0.5914, 0.7496, 0.7836 ], "run": "fusion879__s1" }, { "fusion_state_step": 300, "lambda_min": 0.999959, "lambda_mean": 0.99996, "temp_min": 0.9327, "temp_max": 1.0234, "mean_gate_short_long_delta_value": [ 0.2082, 0.5024, 0.7431, 0.6815 ], "run": "fusion879__s1" }, { "fusion_state_step": 600, "lambda_min": 0.999958, "lambda_mean": 0.99996, "temp_min": 0.9146, "temp_max": 1.0316, "mean_gate_short_long_delta_value": [ 0.2133, 0.4305, 0.7169, 0.552 ], "run": "fusion879__s1" }, { "fusion_state_step": 1200, "lambda_min": 0.999957, "lambda_mean": 0.99996, "temp_min": 0.9212, "temp_max": 1.0333, "mean_gate_short_long_delta_value": [ 0.2597, 0.49, 0.6796, 0.4866 ], "run": "fusion879__s1" }, { "fusion_state_step": 150, "lambda_min": 0.99996, "lambda_mean": 0.99996, "temp_min": 0.9777, "temp_max": 1.0317, "mean_gate_short_long_delta_value": [ 0.2466, 0.6485, 0.7386, 0.85 ], "run": "fusion879__s2" }, { "fusion_state_step": 300, "lambda_min": 0.999959, "lambda_mean": 0.99996, "temp_min": 0.9473, "temp_max": 1.0416, "mean_gate_short_long_delta_value": [ 0.1923, 0.6382, 0.8211, 0.7413 ], "run": "fusion879__s2" }, { "fusion_state_step": 600, "lambda_min": 0.999958, "lambda_mean": 0.99996, "temp_min": 0.9101, "temp_max": 1.0354, "mean_gate_short_long_delta_value": [ 0.2443, 0.5646, 0.8018, 0.6934 ], "run": "fusion879__s2" }, { "fusion_state_step": 1200, "lambda_min": 0.999956, "lambda_mean": 0.99996, "temp_min": 0.8929, "temp_max": 1.0369, "mean_gate_short_long_delta_value": [ 0.2813, 0.5505, 0.7663, 0.6357 ], "run": "fusion879__s2" }, { "fusion_state_step": 150, "lambda_min": 0.99996, "lambda_mean": 0.99996, "temp_min": 0.9867, "temp_max": 1.0239, "mean_gate_short_long_delta_value": [ 0.2818, 0.5884, 0.7853, 0.9142 ], "run": "fusion879__s3" }, { "fusion_state_step": 300, "lambda_min": 0.99996, "lambda_mean": 0.99996, "temp_min": 0.9409, "temp_max": 1.0257, "mean_gate_short_long_delta_value": [ 0.2397, 0.5944, 0.8526, 0.7743 ], "run": "fusion879__s3" }, { "fusion_state_step": 600, "lambda_min": 0.999959, "lambda_mean": 0.99996, "temp_min": 0.9224, "temp_max": 1.0373, "mean_gate_short_long_delta_value": [ 0.2541, 0.4939, 0.7761, 0.6974 ], "run": "fusion879__s3" }, { "fusion_state_step": 1200, "lambda_min": 0.999957, "lambda_mean": 0.99996, "temp_min": 0.8943, "temp_max": 1.0363, "mean_gate_short_long_delta_value": [ 0.2523, 0.4995, 0.7175, 0.6068 ], "run": "fusion879__s3" }, { "fusion_state_step": 150, "lambda_min": 0.99996, "lambda_mean": 0.99996, "temp_min": 0.9788, "temp_max": 1.0151, "mean_gate_short_long_delta_value": [ 0.0645, 0.8883, 0.7187, 0.3471 ], "run": "fusion879_hn__s1" }, { "fusion_state_step": 300, "lambda_min": 0.999959, "lambda_mean": 0.99996, "temp_min": 0.9257, "temp_max": 1.0188, "mean_gate_short_long_delta_value": [ 0.0591, 0.8481, 0.7235, 0.2568 ], "run": "fusion879_hn__s1" }, { "fusion_state_step": 600, "lambda_min": 0.999956, "lambda_mean": 0.999959, "temp_min": 0.9037, "temp_max": 1.0343, "mean_gate_short_long_delta_value": [ 0.203, 0.695, 0.7235, 0.4011 ], "run": "fusion879_hn__s1" }, { "fusion_state_step": 1200, "lambda_min": 0.999953, "lambda_mean": 0.999959, "temp_min": 0.9215, "temp_max": 1.0539, "mean_gate_short_long_delta_value": [ 0.2704, 0.6761, 0.708, 0.4281 ], "run": "fusion879_hn__s1" }, { "fusion_state_step": 150, "lambda_min": 0.99996, "lambda_mean": 0.99996, "temp_min": 0.9797, "temp_max": 1.0187, "mean_gate_short_long_delta_value": [ 0.0779, 0.8297, 0.7168, 0.3404 ], "run": "fusion879_hn__s2" }, { "fusion_state_step": 300, "lambda_min": 0.999959, "lambda_mean": 0.99996, "temp_min": 0.9282, "temp_max": 1.021, "mean_gate_short_long_delta_value": [ 0.1238, 0.744, 0.7661, 0.221 ], "run": "fusion879_hn__s2" }, { "fusion_state_step": 600, "lambda_min": 0.999956, "lambda_mean": 0.999959, "temp_min": 0.9066, "temp_max": 1.0289, "mean_gate_short_long_delta_value": [ 0.2681, 0.5252, 0.7641, 0.4478 ], "run": "fusion879_hn__s2" }, { "fusion_state_step": 1200, "lambda_min": 0.999955, "lambda_mean": 0.999958, "temp_min": 0.887, "temp_max": 1.1186, "mean_gate_short_long_delta_value": [ 0.3671, 0.5468, 0.6843, 0.4353 ], "run": "fusion879_hn__s2" }, { "fusion_state_step": 150, "lambda_min": 0.99996, "lambda_mean": 0.99996, "temp_min": 0.9737, "temp_max": 1.0216, "mean_gate_short_long_delta_value": [ 0.044, 0.8956, 0.7986, 0.3027 ], "run": "fusion879_hn__s3" }, { "fusion_state_step": 300, "lambda_min": 0.999959, "lambda_mean": 0.99996, "temp_min": 0.9103, "temp_max": 1.0238, "mean_gate_short_long_delta_value": [ 0.076, 0.7897, 0.7757, 0.2132 ], "run": "fusion879_hn__s3" }, { "fusion_state_step": 600, "lambda_min": 0.999956, "lambda_mean": 0.999959, "temp_min": 0.8832, "temp_max": 1.0345, "mean_gate_short_long_delta_value": [ 0.2401, 0.671, 0.7493, 0.3072 ], "run": "fusion879_hn__s3" }, { "fusion_state_step": 1200, "lambda_min": 0.999951, "lambda_mean": 0.999959, "temp_min": 0.8653, "temp_max": 1.0537, "mean_gate_short_long_delta_value": [ 0.3024, 0.718, 0.6966, 0.2968 ], "run": "fusion879_hn__s3" }, { "fusion_state_step": 150, "lambda_min": 0.99996, "lambda_mean": 0.99996, "temp_min": 0.9712, "temp_max": 1.026, "mean_gate_short_long_delta_value": [ 0.2981, 0.4936, 0.7456, 0.9231 ], "run": "fusion879_lg__s1" }, { "fusion_state_step": 300, "lambda_min": 0.99996, "lambda_mean": 0.99996, "temp_min": 0.9648, "temp_max": 1.0256, "mean_gate_short_long_delta_value": [ 0.2856, 0.4828, 0.7712, 0.9083 ], "run": "fusion879_lg__s1" }, { "fusion_state_step": 600, "lambda_min": 0.999959, "lambda_mean": 0.99996, "temp_min": 0.9532, "temp_max": 1.0284, "mean_gate_short_long_delta_value": [ 0.3132, 0.344, 0.7916, 0.8599 ], "run": "fusion879_lg__s1" }, { "fusion_state_step": 1200, "lambda_min": 0.999956, "lambda_mean": 0.999959, "temp_min": 0.9217, "temp_max": 1.0553, "mean_gate_short_long_delta_value": [ 0.3345, 0.3793, 0.7776, 0.7454 ], "run": "fusion879_lg__s1" }, { "fusion_state_step": 150, "lambda_min": 0.99996, "lambda_mean": 0.99996, "temp_min": 0.9817, "temp_max": 1.0284, "mean_gate_short_long_delta_value": [ 0.2723, 0.3713, 0.761, 0.9398 ], "run": "fusion879_lg__s2" }, { "fusion_state_step": 300, "lambda_min": 0.99996, "lambda_mean": 0.99996, "temp_min": 0.964, "temp_max": 1.027, "mean_gate_short_long_delta_value": [ 0.2586, 0.369, 0.7821, 0.933 ], "run": "fusion879_lg__s2" }, { "fusion_state_step": 600, "lambda_min": 0.999958, "lambda_mean": 0.99996, "temp_min": 0.9209, "temp_max": 1.0308, "mean_gate_short_long_delta_value": [ 0.2997, 0.2883, 0.7866, 0.904 ], "run": "fusion879_lg__s2" }, { "fusion_state_step": 1200, "lambda_min": 0.999957, "lambda_mean": 0.999959, "temp_min": 0.8743, "temp_max": 1.0472, "mean_gate_short_long_delta_value": [ 0.3103, 0.3668, 0.764, 0.7998 ], "run": "fusion879_lg__s2" }, { "fusion_state_step": 150, "lambda_min": 0.99996, "lambda_mean": 0.99996, "temp_min": 0.9779, "temp_max": 1.0364, "mean_gate_short_long_delta_value": [ 0.2533, 0.39, 0.7639, 0.9297 ], "run": "fusion879_lg__s3" }, { "fusion_state_step": 300, "lambda_min": 0.999959, "lambda_mean": 0.99996, "temp_min": 0.9713, "temp_max": 1.0445, "mean_gate_short_long_delta_value": [ 0.222, 0.4204, 0.7917, 0.9207 ], "run": "fusion879_lg__s3" }, { "fusion_state_step": 600, "lambda_min": 0.999959, "lambda_mean": 0.99996, "temp_min": 0.954, "temp_max": 1.0358, "mean_gate_short_long_delta_value": [ 0.2402, 0.3599, 0.7798, 0.8769 ], "run": "fusion879_lg__s3" }, { "fusion_state_step": 1200, "lambda_min": 0.999957, "lambda_mean": 0.999959, "temp_min": 0.937, "temp_max": 1.0251, "mean_gate_short_long_delta_value": [ 0.2508, 0.4064, 0.7186, 0.7519 ], "run": "fusion879_lg__s3" }, { "fusion_state_step": 150, "lambda_min": 0.999959, "lambda_mean": 0.99996, "temp_min": 0.9636, "temp_max": 1.0231, "mean_gate_short_long_delta_value": [ 0.2677, 0.2979, 0.7574, 0.9513 ], "run": "fusion879_lr__s1" }, { "fusion_state_step": 300, "lambda_min": 0.999959, "lambda_mean": 0.99996, "temp_min": 0.9528, "temp_max": 1.0269, "mean_gate_short_long_delta_value": [ 0.2429, 0.2812, 0.812, 0.9434 ], "run": "fusion879_lr__s1" }, { "fusion_state_step": 600, "lambda_min": 0.999957, "lambda_mean": 0.99996, "temp_min": 0.93, "temp_max": 1.0215, "mean_gate_short_long_delta_value": [ 0.238, 0.2527, 0.8268, 0.9182 ], "run": "fusion879_lr__s1" }, { "fusion_state_step": 1200, "lambda_min": 0.999954, "lambda_mean": 0.999959, "temp_min": 0.8993, "temp_max": 1.035, "mean_gate_short_long_delta_value": [ 0.2302, 0.2809, 0.8038, 0.8295 ], "run": "fusion879_lr__s1" }, { "fusion_state_step": 150, "lambda_min": 0.99996, "lambda_mean": 0.99996, "temp_min": 0.9682, "temp_max": 1.0178, "mean_gate_short_long_delta_value": [ 0.2818, 0.2674, 0.7476, 0.953 ], "run": "fusion879_lr__s2" }, { "fusion_state_step": 300, "lambda_min": 0.999959, "lambda_mean": 0.99996, "temp_min": 0.9674, "temp_max": 1.0228, "mean_gate_short_long_delta_value": [ 0.2672, 0.2634, 0.7809, 0.9515 ], "run": "fusion879_lr__s2" }, { "fusion_state_step": 600, "lambda_min": 0.999958, "lambda_mean": 0.99996, "temp_min": 0.935, "temp_max": 1.0174, "mean_gate_short_long_delta_value": [ 0.2446, 0.2476, 0.819, 0.9383 ], "run": "fusion879_lr__s2" }, { "fusion_state_step": 1200, "lambda_min": 0.999957, "lambda_mean": 0.999959, "temp_min": 0.9049, "temp_max": 1.0327, "mean_gate_short_long_delta_value": [ 0.2176, 0.2594, 0.8341, 0.876 ], "run": "fusion879_lr__s2" }, { "fusion_state_step": 150, "lambda_min": 0.99996, "lambda_mean": 0.99996, "temp_min": 0.9655, "temp_max": 1.0218, "mean_gate_short_long_delta_value": [ 0.2604, 0.2859, 0.7591, 0.9514 ], "run": "fusion879_lr__s3" }, { "fusion_state_step": 300, "lambda_min": 0.999959, "lambda_mean": 0.99996, "temp_min": 0.9579, "temp_max": 1.0246, "mean_gate_short_long_delta_value": [ 0.2555, 0.2749, 0.771, 0.9508 ], "run": "fusion879_lr__s3" }, { "fusion_state_step": 600, "lambda_min": 0.999958, "lambda_mean": 0.99996, "temp_min": 0.9285, "temp_max": 1.0212, "mean_gate_short_long_delta_value": [ 0.2385, 0.2595, 0.7947, 0.9421 ], "run": "fusion879_lr__s3" }, { "fusion_state_step": 1200, "lambda_min": 0.999956, "lambda_mean": 0.999959, "temp_min": 0.9098, "temp_max": 1.0374, "mean_gate_short_long_delta_value": [ 0.2289, 0.2919, 0.7833, 0.8983 ], "run": "fusion879_lr__s3" }, { "fusion_state_step": 150, "lambda_min": 0.99996, "lambda_mean": 0.99996, "temp_min": 0.975, "temp_max": 1.0377, "mean_gate_short_long_delta_value": [ 0.2816, 0.347, 0.7499, 0.9452 ], "run": "fusion879_lrk__s1" }, { "fusion_state_step": 300, "lambda_min": 0.99996, "lambda_mean": 0.99996, "temp_min": 0.9676, "temp_max": 1.0474, "mean_gate_short_long_delta_value": [ 0.2655, 0.3381, 0.7731, 0.9442 ], "run": "fusion879_lrk__s1" }, { "fusion_state_step": 600, "lambda_min": 0.999958, "lambda_mean": 0.99996, "temp_min": 0.9568, "temp_max": 1.0364, "mean_gate_short_long_delta_value": [ 0.2645, 0.3039, 0.7835, 0.9369 ], "run": "fusion879_lrk__s1" }, { "fusion_state_step": 1200, "lambda_min": 0.999956, "lambda_mean": 0.999959, "temp_min": 0.9303, "temp_max": 1.0406, "mean_gate_short_long_delta_value": [ 0.2645, 0.3424, 0.7872, 0.8504 ], "run": "fusion879_lrk__s1" }, { "fusion_state_step": 150, "lambda_min": 0.99996, "lambda_mean": 0.99996, "temp_min": 0.9801, "temp_max": 1.0288, "mean_gate_short_long_delta_value": [ 0.2734, 0.3301, 0.7535, 0.9498 ], "run": "fusion879_lrk__s2" }, { "fusion_state_step": 300, "lambda_min": 0.999959, "lambda_mean": 0.99996, "temp_min": 0.9652, "temp_max": 1.0346, "mean_gate_short_long_delta_value": [ 0.2602, 0.3263, 0.7693, 0.947 ], "run": "fusion879_lrk__s2" }, { "fusion_state_step": 600, "lambda_min": 0.999957, "lambda_mean": 0.99996, "temp_min": 0.9549, "temp_max": 1.027, "mean_gate_short_long_delta_value": [ 0.2619, 0.3013, 0.7926, 0.9314 ], "run": "fusion879_lrk__s2" }, { "fusion_state_step": 1200, "lambda_min": 0.999955, "lambda_mean": 0.999959, "temp_min": 0.9088, "temp_max": 1.0825, "mean_gate_short_long_delta_value": [ 0.2621, 0.3321, 0.7916, 0.8482 ], "run": "fusion879_lrk__s2" }, { "fusion_state_step": 150, "lambda_min": 0.99996, "lambda_mean": 0.99996, "temp_min": 0.9801, "temp_max": 1.0287, "mean_gate_short_long_delta_value": [ 0.2674, 0.3405, 0.7563, 0.9446 ], "run": "fusion879_lrk__s3" }, { "fusion_state_step": 300, "lambda_min": 0.999959, "lambda_mean": 0.99996, "temp_min": 0.9695, "temp_max": 1.0381, "mean_gate_short_long_delta_value": [ 0.2542, 0.3589, 0.7725, 0.9418 ], "run": "fusion879_lrk__s3" }, { "fusion_state_step": 600, "lambda_min": 0.999957, "lambda_mean": 0.99996, "temp_min": 0.9157, "temp_max": 1.0344, "mean_gate_short_long_delta_value": [ 0.2557, 0.3122, 0.7903, 0.9358 ], "run": "fusion879_lrk__s3" }, { "fusion_state_step": 1200, "lambda_min": 0.999955, "lambda_mean": 0.999959, "temp_min": 0.8811, "temp_max": 1.0264, "mean_gate_short_long_delta_value": [ 0.236, 0.3739, 0.7906, 0.8637 ], "run": "fusion879_lrk__s3" } ], "census": { "formula_matches_380": true, "rows": { "base@380": { "latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 102158627, "kv_bytes_per_token": 24903680, "mixer_fp_params": 0, "candidate_layers": 0, "layers": 380 }, "dlgm458@380": { "latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 39820506467, "kv_bytes_per_token": 24903680, "mixer_fp_params": 39718347840, "candidate_layers": 285, "layers": 380 }, "dlgm458_lr@380": { "latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 109016867, "kv_bytes_per_token": 24903680, "mixer_fp_params": 6858240, "candidate_layers": 285, "layers": 380 }, "fusion879@380": { "latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 41081291747, "kv_bytes_per_token": 24903680, "mixer_fp_params": 40979133120, "candidate_layers": 285, "layers": 380 }, "fusion879_lr@380": { "latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 114233507, "kv_bytes_per_token": 24903680, "mixer_fp_params": 12074880, "candidate_layers": 285, "layers": 380 }, "base@560": { "latent_bytes": 886910746624, "table_bytes": 1638924288, "n_matrices": 3864, "fp_params": 150450908, "kv_bytes_per_token": 36700160, "mixer_fp_params": 0, "candidate_layers": 0, "layers": 560 }, "dlgm458@560": { "latent_bytes": 886910746624, "table_bytes": 1638924288, "n_matrices": 3864, "fp_params": 58682752988, "kv_bytes_per_token": 36700160, "mixer_fp_params": 58532302080, "candidate_layers": 420, "layers": 560 }, "dlgm458_lr@560": { "latent_bytes": 886910746624, "table_bytes": 1638924288, "n_matrices": 3864, "fp_params": 160557788, "kv_bytes_per_token": 36700160, "mixer_fp_params": 10106880, "candidate_layers": 420, "layers": 560 }, "fusion879@560": { "latent_bytes": 886910746624, "table_bytes": 1638924288, "n_matrices": 3864, "fp_params": 60540752348, "kv_bytes_per_token": 36700160, "mixer_fp_params": 60390301440, "candidate_layers": 420, "layers": 560 }, "fusion879_lr@560": { "latent_bytes": 886910746624, "table_bytes": 1638924288, "n_matrices": 3864, "fp_params": 168245468, "kv_bytes_per_token": 36700160, "mixer_fp_params": 17794560, "candidate_layers": 420, "layers": 560 }, "fusion879_lrk@380": { "latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 294918947, "kv_bytes_per_token": 24903680, "mixer_fp_params": 192760320, "candidate_layers": 285, "layers": 380 }, "fusion879_lrk@560": { "latent_bytes": 886910746624, "table_bytes": 1638924288, "n_matrices": 3864, "fp_params": 434518748, "kv_bytes_per_token": 36700160, "mixer_fp_params": 284067840, "candidate_layers": 420, "layers": 560, "derived_from": "380-block census of the smoke run, same per-layer mixer cost" }, "dlgm458_lrk@380": { "latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 224840867, "kv_bytes_per_token": 24903680, "mixer_fp_params": 122682240, "candidate_layers": 285, "layers": 380 }, "dlgm458_lrk@560": { "latent_bytes": 886910746624, "table_bytes": 1638924288, "n_matrices": 3864, "fp_params": 331245788, "kv_bytes_per_token": 36700160, "mixer_fp_params": 180794880, "candidate_layers": 420, "layers": 560, "derived_from": "380-block census of the smoke run, same per-layer mixer cost" }, "fusion879_lg@560": { "latent_bytes": 886910746624, "table_bytes": 1638924288, "n_matrices": 3864, "fp_params": 544619228, "kv_bytes_per_token": 36700160, "mixer_fp_params": 394168320, "candidate_layers": 420, "layers": 560, "derived_from": "gentler cut plus the 32 x 8192 projection per candidate layer as a trained parameter" }, "dlgm458_lg@560": { "latent_bytes": 886910746624, "table_bytes": 1638924288, "n_matrices": 3864, "fp_params": 441346268, "kv_bytes_per_token": 36700160, "mixer_fp_params": 290895360, "candidate_layers": 420, "layers": 560, "derived_from": "gentler cut plus the 32 x 8192 projection per candidate layer as a trained parameter" } } }, "fidelity": { "fla": "0.2.2", "torch": "2.9.1+cu128", "upstream_index": 879, "phases": [ { "name": "cpu64", "device": "cpu", "dtype": "torch.float64", "tol": 1e-09, "ok": true, "cases": [ { "T": 50, "max_abs_err": 8.881784197001252e-16, "ref_max": 2.251193709137344 }, { "T": 64, "max_abs_err": 5.551115123125783e-16, "ref_max": 2.325048122191927 }, { "T": 100, "max_abs_err": 6.661338147750939e-16, "ref_max": 2.2115917118421633 }, { "T": 256, "max_abs_err": 6.661338147750939e-16, "ref_max": 2.3626990035987725 }, { "T": 257, "max_abs_err": 4.440892098500626e-16, "ref_max": 2.5568741562689334 }, { "T": 1024, "max_abs_err": 8.881784197001252e-16, "ref_max": 2.8854929294257508 } ], "shortconv": [ "ShortConvolution", 4, 3, "silu", true ], "replaced": true }, { "name": "gpu32", "device": "cuda", "dtype": "torch.float32", "tol": 0.0002, "ok": true, "cases": [ { "T": 50, "max_abs_err": 1.3113021850585938e-06, "ref_max": 2.2511940002441406 }, { "T": 64, "max_abs_err": 5.960464477539062e-07, "ref_max": 2.3250484466552734 }, { "T": 100, "max_abs_err": 5.960464477539062e-07, "ref_max": 2.211591958999634 }, { "T": 256, "max_abs_err": 7.152557373046875e-07, "ref_max": 2.362699508666992 }, { "T": 257, "max_abs_err": 9.5367431640625e-07, "ref_max": 2.5568742752075195 }, { "T": 1024, "max_abs_err": 4.76837158203125e-07, "ref_max": 2.8854928016662598 } ], "shortconv": [ "ShortConvolution", 4, 3, "silu", true ], "replaced": false } ], "causality": { "leak_before_edit": 0.0, "change_from_edit": 0.8144181108950805 }, "ignores_causal_flag": true, "layer3_is_attention": true, "lr": { "full_flag": false, "shape": [ 2, 160, 256 ], "finite": true, "leak_before_edit": 0.0, "fp_params": 4628 }, "full_fp_params": 168988, "ok": true } }