File size: 11,490 Bytes
d8321f8 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 | {
"allocation_gpus": 4,
"cache_policy": "Fresh process and Ray/runtime identifiers per period. Shared immutable model snapshot and identical preparation; writable compiler/cache policy identical between periods and recorded by runner. Order randomization does not eliminate all thermal/cache carryover. No hardware-confinement claim from a mask alone.",
"classifier": {
"C": 1.0,
"class_weight": "balanced",
"convergence": "ConvergenceWarning or nonfinite parameters is a hard failure",
"decision": "argmax; no probability calibration or safety interpretation",
"implementation": "sklearn.linear_model.LogisticRegression",
"max_iter": 10000,
"multi_class": "multinomial",
"penalty": "l2",
"random_state": 20260915,
"scaling": "Training-fold StandardScaler",
"solver": "lbfgs"
},
"collection_order_seed": 2026091503,
"comparison": "Two versus four task GPUs on the same four-GPU allocation; matched global prompts, compound level, seed, source and horizon. Generated responses may differ; this is the total execution consequence, not identical tensor work or pure memory sharding.",
"component_regression": "Ordinary nonnegative least squares with intercept, maxiter10000, float64. Only configurations with three completed repeats supply numerical targets, their maximum external completed peak. Feature divisors=max(training-completion feature max,1). This is empirical whole-run component regression, not exact tensor liveness or a sum/max of measured stage peaks. Intercept and coefficients jointly fit runtime/activation effects; coefficients are not causal components.",
"cost": "Report full permitted fitting pool's actual attempts separately from donor-copy selected subset. Completed-only regression still reports full pool acquisition. Existing source reuse has zero incremental GPU collection; a cold start does not. CPU fitting/prediction time separately.",
"cost_scope": "Both periods reserve four GPUs. Report task-provisioned GPU-time (2*t2+4*t4), reservation during invocations 4*(t2+t4), and whole allocation4*T separately. Include failed/unresolved work; queue wait is separate. None is utilization-weighted time.",
"data_sha256": {
"data/codecontests/heavy_tail/test.parquet": "e9d077a023de7f80aa013c5541c33b5978a43e203ce916045bc72c4fe5c62a82",
"data/codecontests/heavy_tail/train.parquet": "0bc6ec8fd7ef1618a604417faa4bb13cb9a0117dcffb75685b2c80c1da95f697",
"data/gsm8k/test.parquet": "0dcd50ed32caa9d8f434d4e3ef8031bc3e49b0f823655a53be5c002e727f17dd",
"data/gsm8k/train.parquet": "8fb13a0cd8621e5fcacdee3ffe28b3a369975478ea7ab514781f3f190cba2488",
"data/math/test.parquet": "e0cb48b1e49874c71f4a02dff944eba08844067d98dc5bdc89c004644e6b597d",
"data/math/train.parquet": "4aa303577e968be6f810600730c970b14be659a9ae9072cc2afefd04d67e7cb8"
},
"device_capacity_mib": 40960,
"donor_rule": "Same workload and compound level required. Prefer same GPU count, then same architecture-family prior, then smallest absolute log parameter-count ratio; tie-break by configuration_id. A missing exact GPU match uses a disclosed different-GPU donor. No outcomes affect donor selection.",
"evaluation_seeds": [
151,
152,
153
],
"failure_policy": "No reruns/replacements/requeues. Each period has a persistent launch ledger. Continue the second period after any first-period outcome only after all four devices pass identity/cleanup/idle checks. Otherwise retain not-started. Missing GPU identity/evidence is unresolved, not OOM. Early validated OOM need not reach full worker initialization. Do not drop incomplete pairs.",
"feature_scope": "Architecture/checkpoint-header and configured shape volumes only. Largest-unit proxy=max(largest decoder block, root non-block elements, largest weight tensor); not an observed wrapping map. Checkpoint-boundary and vocabulary-output volumes are predictive covariates, not exact live activation totals. No source identity, dataset name, compound level, actual generated length or outcome appears as a numeric feature.",
"features": [
"actor_shards_gib",
"adapter_optimizer_gib",
"generation_budget_gib",
"generation_replica_gib",
"largest_unit_proxy_gib",
"checkpoint_boundary_volume_gib",
"vocabulary_output_volume_gib",
"actor_parameter_offload"
],
"frozen_at_utc": "2026-09-15T13:33:20.483317+00:00",
"gpus": "NVIDIA A100-SXM4-40GB",
"input_sha256": {
"benchmark/attempts.csv": "397c8caaf0d3abe06587fd649c181fe71d5b8b5f5e24e0591cc7e56d38e275c1",
"benchmark/configurations.csv": "3a77c96a67b82e01d23fb6f7a440cba23041bed1dae2749072cb9abb5b5d239d",
"benchmark/estimation/matrix.csv": "1ff175e22c76bf1af3857b4ae57cd4f78bafe0d280a295d696e96815ef85e9c8",
"benchmark/estimation/model_metadata.json": "2f0ce16c13e72900d2a60f9b7ad28d4dc5ed43d7915d8a88925ba9c5ffd684ad",
"benchmark/estimation/targets.csv": "f2481278127fa726f24297d81fc9df46e999315fe543167ed398e671fffa522f",
"benchmark/outcomes.csv": "af05cd91ed1de843735cc188e5e54dbd1645b3ed91c97f503ac3de67bc5f61c9",
"memory_tuner/estimation_baselines.py": "0658f92a18147fd67909ca49fa3dc32020217cfd61d34651161c95baa80dd55a",
"memory_tuner/model_memory_metadata.py": "8845a1197b49d9463ebe5dc8c953463dd2106361c54e7b12244057c48aaa2712",
"memory_tuner/plan_estimation_study.py": "16855f584441528870fcca38c3a19eafd1234c3eacb9b43cc9eebdd4bd6ad862"
},
"label_mapping": "Component regression first rejects a positive resident-actor startup-budget check as memory_failure. Otherwise regression peak > capacity predicts memory_failure; otherwise > margin predicts above_margin; otherwise within_margin. This is an evaluated heuristic, not a claim that every allocation failure requires an observed capacity crossing.",
"memory_margin_limit_mib": 38912,
"model": "Qwen/Qwen2.5-7B-Instruct",
"model_revision": "a09a35458c702b33eeacc393d103063234e8bc28",
"outcome_rule": "A configuration has a repeated memory label only if all three seed slots have validated memory outcomes. Any diagnosed memory failure then implies memory_failure; otherwise any completed peak above38912MiB implies above_margin; otherwise within_margin. Unresolved/not-started slots remain explicit with known failure/margin flags, not replaced or discarded.",
"pair_order_seed": 2026091501,
"planned_pairs": 18,
"planned_processes": 36,
"prior_visibility": "Old labels and prior manuscript findings were public during design. Retrospective holdout is not blind method development. Historical7B exploratory work used other workflows and is excluded from fitting/selection. New standardGRPO predictions are frozen before their outcomes, not an unseen-family/hardware evaluation.",
"prospective_design": "All 90 old configurations fit predictions for 12 new 7B configurations before execution.",
"protocol_version": "estimation-1.0",
"randomization": "9 two-GPU-first and 9 four-GPU-first pairs, balanced as closely as possible within each three-seed workload/level cell. Each of six two-of-four allocation-ordinal subsets occurs three times, assigned independently of period order. Physical UUIDs resolve before either period.",
"resource_cap": {
"gpus_per_allocation": 4,
"idle_timeout_seconds": 180,
"initial_allocations": 18,
"max_concurrent_allocations": 3,
"maximum_reserved_gpu_hours": 66,
"payload_timeout_seconds": 1500,
"requeue": false,
"retries": 0,
"wall_minutes_per_allocation": 55
},
"retrospective_design": "Three leave-model-family-out folds, 60 fitting / 30 test configurations each.",
"source_configuration_ids": [
"cfg-072bd7669d20f8fa",
"cfg-0d9ba9215ab5683e",
"cfg-0f540d3f6fe4661b",
"cfg-0f8a31ef7ac4c551",
"cfg-10644b5d0a82d5c3",
"cfg-1432bb8a4d9c3b8c",
"cfg-17e69bcf1b041bd9",
"cfg-18285c3cb50308d4",
"cfg-1c5278f8be90a50a",
"cfg-2afbe250af9cd3e1",
"cfg-301e4dae7048c6ec",
"cfg-30f98456fdb322c1",
"cfg-32977d5aa42a41cb",
"cfg-338d9224eb9d2f3a",
"cfg-352f4dbc55fbcd3c",
"cfg-371ae37408089efe",
"cfg-399f6b038d6f65f9",
"cfg-3c86a11fdb8455cb",
"cfg-3dc042a103546eac",
"cfg-3f0f48a3c570a0c8",
"cfg-49d9bccfe873f5c6",
"cfg-5244d13eac3116ad",
"cfg-55811d23c5144bb7",
"cfg-5839ccf95c2c1722",
"cfg-5ff9621270e710ed",
"cfg-61c14de21a34d0e6",
"cfg-62f371a80968a0ba",
"cfg-6a4562db98b38d6f",
"cfg-6d54ac65d934f404",
"cfg-714c540f76d9525d",
"cfg-73c06764ba28a56b",
"cfg-78c440d3891ddbc2",
"cfg-7a0e81934d3398ee",
"cfg-7a9a974958c57abc",
"cfg-7df2428553b76877",
"cfg-7f23cab692546e3f",
"cfg-835067f7405e89a2",
"cfg-896685010703c0fb",
"cfg-89ca48ce88ea694e",
"cfg-8bd7050cd7f49e78",
"cfg-9160dc45a0171a10",
"cfg-93d4f41e600c9079",
"cfg-966f5804dc918967",
"cfg-96ccf395b84393f7",
"cfg-97107ceba8bf0cbe",
"cfg-9758870c0a0b4d69",
"cfg-9c461415e6bb9702",
"cfg-9f624ec22aa3dde0",
"cfg-a026693bae37d30b",
"cfg-a23a9ffc1f670139",
"cfg-a24e9cdd2f6143f2",
"cfg-a3987314cb2f9f12",
"cfg-a94bf7877633dfda",
"cfg-aa33c72e0a449b9c",
"cfg-acfa4acda15c3599",
"cfg-b5007322fb2e1772",
"cfg-b5352215babfeba6",
"cfg-b56fb353c2f2fa3e",
"cfg-b8dd70ab682c13b2",
"cfg-b8e90bd69c9744ea",
"cfg-bae2977da7957f72",
"cfg-bfb83ff32b7fd921",
"cfg-c0605c1607caa079",
"cfg-c2a8e6e8e140e109",
"cfg-c3f0ad157d299d39",
"cfg-ca71b56dc3189305",
"cfg-cad8341dbbff10ba",
"cfg-cc2c35944156e5d4",
"cfg-cc5f4f0c2e1256ba",
"cfg-cc900fb7731c3f49",
"cfg-cd3974595f9db882",
"cfg-d02503f8c9804a2d",
"cfg-d138048feeb8fe5d",
"cfg-d4a213515805a9a7",
"cfg-d9397d15f234d541",
"cfg-db87c67536995ab0",
"cfg-e117f15dd5ccfb30",
"cfg-e2a205be8efba055",
"cfg-e8893b12558f51dd",
"cfg-e91942abb5b4f847",
"cfg-ea5a23a40ad5ca50",
"cfg-ea92cccc91029465",
"cfg-f009e9e3891c06f4",
"cfg-f03e4e98c682240e",
"cfg-f439d2dd90137a80",
"cfg-f75b68f479767290",
"cfg-fad1b555f8b37235",
"cfg-fbe1c0422f98cb38",
"cfg-fd4d4789145230df",
"cfg-fd5bb0db68b0dae3"
],
"source_configurations": 90,
"source_selection": "Only 72 boundary and 18 four_gpu source configurations; all prior7B exploration, other controls and the 48-invocation prospective admission panel are excluded. No new two-GPU outcome updates predictions for four GPUs or vice versa.",
"source_studies": [
"boundary",
"four_gpu"
],
"startup_check": "Separate one-sided diagnostic only for resident actor: 4(P+R)/G+u*C>C. Shared-reference counted once; generation budget includes its replica/cache. Assumes actor shards resident before vLLM free-memory check. Passing does not predict completion or margin compliance; no target observation used.",
"status": "specified_before_fitting_and_new_standard_grpo_outcomes",
"subset_seed": 2026091502,
"target_configuration_ids": [
"est-qwen25-7b-gsm8k-c2-2gpu",
"est-qwen25-7b-gsm8k-c2-4gpu",
"est-qwen25-7b-gsm8k-c3-2gpu",
"est-qwen25-7b-gsm8k-c3-4gpu",
"est-qwen25-7b-math-c2-2gpu",
"est-qwen25-7b-math-c2-4gpu",
"est-qwen25-7b-math-c3-2gpu",
"est-qwen25-7b-math-c3-4gpu",
"est-qwen25-7b-code_heavy_tail-c2-2gpu",
"est-qwen25-7b-code_heavy_tail-c2-4gpu",
"est-qwen25-7b-code_heavy_tail-c3-2gpu",
"est-qwen25-7b-code_heavy_tail-c3-4gpu"
],
"target_configurations": 12,
"task_gpu_counts": [
2,
4
]
}
|