| { |
| "schema_version": 1, |
| "prediction_seal_sha256": "5ed714caff922a5109346a4242545d19f2824636d0f52bb0bbd2a09b40ed4666", |
| "original_protocol_sha256": "1694e447e842600612604891e7b3d036ff0791fa84c4bdbeba54ed093feefdeb", |
| "events": [ |
| { |
| "kind": "zero_training_preflight", |
| "job_id": "5054195", |
| "source_commit": "e41df1aa1ab289cb5e2b0b3beeb3a03ff66e66d3", |
| "status": "failed", |
| "elapsed_seconds": 5, |
| "allocated_gpus": 4, |
| "training_launched": false, |
| "cause": "Global Slurm GPU ordinals 4..7 differ from the four-device NVML job-view ordinals 0..3." |
| }, |
| { |
| "kind": "read_only_device_diagnostic", |
| "job_id": "5054207", |
| "status": "completed", |
| "elapsed_seconds": 2, |
| "allocated_gpus": 4, |
| "training_launched": false, |
| "finding": "Exactly four devices visible through NVML, matching CUDA_VISIBLE_DEVICES=0,1,2,3; global Slurm GPU indices retained separately. UUIDs resolve from this complete job view." |
| }, |
| { |
| "kind": "zero_training_preflight", |
| "job_id": "5054224", |
| "source_commit": "0bd659b3f69e2f779dbea84bb2da221bd0ff52db", |
| "status": "completed", |
| "elapsed_seconds": 53, |
| "allocated_gpus": 4, |
| "training_launched": false, |
| "checks": "Nonconsecutive two-device subset, then all four; actual CUDA UUIDs, Ray placement, NVML selection and post-exit all-device idle checks passed." |
| }, |
| { |
| "kind": "study_array", |
| "array_job_id": "5054326", |
| "source_commit": "0bd659b3f69e2f779dbea84bb2da221bd0ff52db", |
| "requested_pair_indices": [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17], |
| "started_pair_indices": [0, 1], |
| "never_started_pair_indices": [2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17], |
| "cancelled_at_utc": "2026-09-15T14:33:09+00:00", |
| "reason": "First pair exposed deterministic vLLM integer parsing of UUID masks during import, before model initialization. Cancellation prevented repeated setup failures. Cleanup also raced an allocation-sampler subprocess.", |
| "started_invocation_evidence": "Three training logs diagnose the vLLM UUID-to-integer import failure. Seed151/twoGPU was interrupted during device preflight and has no matching traceback. All four invocations are infrastructure-invalid, not memory outcomes.", |
| "repair_scope": "Device-identification compatibility and sampler shutdown only. Frozen predictions, models, data, settings, seeds, pair order and subsets are unchanged.", |
| "replacement_policy": "No started pair or invocation will be rerun. Only indices never claimed may be launched once from corrected immutable source. Missing repeated labels remain unresolved." |
| }, |
| { |
| "kind": "expanded_zero_training_preflight", |
| "job_id": "5054489", |
| "source_commit": "26418498a5b5405a31c8ce7e4f2b9bd1cd7f9575", |
| "status": "failed", |
| "started_at_utc": "2026-09-15T15:26:01+00:00", |
| "ended_at_utc": "2026-09-15T15:26:40+00:00", |
| "elapsed_seconds": 39, |
| "allocated_gpus": 4, |
| "training_launched": false, |
| "cause": "A training-library import initialized CUDA. The preflight incorrectly equated zero training with zero CUDA-context initialization. The failing import was not identified by this version.", |
| "checks_completed": "Initial allocation idle check and preflight-start record only. No two/four-GPU runtime report or pass record was produced.", |
| "additional_checks": "Actual VERL and vLLM worker imports, UUID-to-NVML round-trip validation, capability queries in driver and Ray actors, preservation of physical UUID masks, and post-exit idle checks." |
| }, |
| { |
| "kind": "remaining_study_array", |
| "array_job_id": "5054693", |
| "source_commit": "26418498a5b5405a31c8ce7e4f2b9bd1cd7f9575", |
| "status": "cancelled_before_any_start", |
| "reason": "The expanded preflight failed; DependencyNeverSatisfied prevented every array element from starting. Cancelled only this pending array before queuing the never-claimed slots behind the corrected preflight.", |
| "started_pair_indices": [], |
| "dependency": "afterok:5054489", |
| "requested_pair_indices": [2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17], |
| "maximum_concurrency": 3, |
| "retry_or_replacement": false, |
| "source_worktree_clean_before_submission": true, |
| "frozen_protocol_matrix_and_prediction_seal_unchanged": true, |
| "runtime_scope": "Device identity compatibility and sampling shutdown only; no model, data, batching, precision, memory-control, seed, order, subset, or predictor changes." |
| }, |
| { |
| "kind": "corrected_import_context_zero_training_preflight", |
| "job_id": "5054938", |
| "source_commit": "d8e0c3551d8572fb64e3b7126fbd7569a987973f", |
| "status": "completed", |
| "started_at_utc": "2026-09-15T15:46:24+00:00", |
| "ended_at_utc": "2026-09-15T15:47:47+00:00", |
| "elapsed_seconds": 83, |
| "allocated_gpus": 4, |
| "training_launched": false, |
| "correction": "Explicitly disposable preflight children may initialize CUDA through imports. Per-import transitions are recorded; actual task UUIDs, NVML capabilities and Ray placement remain checked. All-four-device idle checks now follow both successful and failed child exits.", |
| "checks_completed": "Both task sizes passed CUDA/Ray/NVML identity and capability checks, with two all-four-device idle samples after each disposable child. Recorded import-created CUDA contexts arose while importing verl.workers.rollout.vllm_rollout.vllm_async_server.", |
| "two_gpu_runtime_report_sha256": "77f617e346869b6a617eb316468a7ac74b9156ae7e168df3ea089aba5153c81d", |
| "two_gpu_post_child_idle_sha256": "6c4adc25407078a374b10cebf6d810db8ee4cf724215679fa570128e5f16cd32", |
| "four_gpu_runtime_report_sha256": "64e83062d19c77987a2215851b2335a8d25ca6ac79020f72073d6cc1109ac97f", |
| "four_gpu_post_child_idle_sha256": "df025b3e93ead549afe56be266c4b0407a96d4cea243e1eab50dc29d8af91cec", |
| "actual_training_bridge_and_trainer_probe_exec_unchanged": true, |
| "frozen_protocol_matrix_and_prediction_seal_unchanged": true |
| }, |
| { |
| "kind": "remaining_study_array", |
| "array_job_id": "5054951", |
| "source_commit": "d8e0c3551d8572fb64e3b7126fbd7569a987973f", |
| "status": "completed_all_claims_terminal", |
| "first_started_at_utc": "2026-09-15T15:48:51+00:00", |
| "last_ended_at_utc": "2026-09-15T16:22:37+00:00", |
| "dependency": "afterok:5054938", |
| "requested_pair_indices": [2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17], |
| "maximum_concurrency": 3, |
| "retry_or_replacement": false, |
| "source_worktree_clean_before_submission": true, |
| "frozen_protocol_matrix_and_prediction_seal_unchanged": true |
| }, |
| { |
| "kind": "host_memory_censoring_diagnostic", |
| "observed_at_utc": "2026-09-15T16:08:43+00:00", |
| "array_job_id": "5054951", |
| "source_commit": "d8e0c3551d8572fb64e3b7126fbd7569a987973f", |
| "host_memory_limit_mib": 196608, |
| "confirmed_jobs": [ |
| { |
| "job_id": "5054954", |
| "pair_index": 4, |
| "pair_id": "est-qwen25-7b-gsm8k-c2-s152", |
| "node": "acn18", |
| "scheduler_oom_kill_events": 4, |
| "scheduler_max_rss_mib": 196607.84, |
| "ended_at_utc": "2026-09-15T15:57:33+00:00" |
| }, |
| { |
| "job_id": "5054981", |
| "pair_index": 6, |
| "pair_id": "est-qwen25-7b-gsm8k-c2-s151", |
| "node": "acn11", |
| "scheduler_oom_kill_events": 3, |
| "scheduler_max_rss_mib": 196607.59, |
| "ended_at_utc": "2026-09-15T16:02:35+00:00" |
| } |
| ], |
| "finding": "Both four-task-GPU c2 periods reached generation-engine graph capture and then lost workers at the host-memory cgroup limit. Neither completed a training step or reported CUDA out-of-memory. Partial device peaks do not establish a completed memory label.", |
| "eligibility": "Keep four-GPU outcomes unresolved. Earlier two-GPU admission failures remain separately subject to independent evidence validation. Retain costs and all raw evidence.", |
| "additional_scheduler_evidence": "Array index 7, job 5054998, also ended OUT_OF_MEMORY at 16:05:07 UTC with four oom_kill events; its detailed period diagnosis is not included in the two-job finding above.", |
| "original_study_resources_and_prediction_seal_unchanged": true, |
| "new_followup_launched": false |
| } |
| ], |
| "status": "complete_independently_collected", |
| "final_outcomes": { |
| "planned_slots": 36, |
| "launched_invocations": 34, |
| "eligible_memory_admission_failures": 23, |
| "completed_invocations": 0, |
| "unresolved_slots": 13, |
| "resolved_configurations": 6, |
| "resolved_configuration_states": "all memory_failure", |
| "unresolved_configurations": 6, |
| "unresolved_slot_causes": { |
| "initial_setup_or_interrupted_invocation": 4, |
| "host_memory_exhaustion": 7, |
| "scheduler_identity_nonlaunch": 2 |
| }, |
| "same_setting_retries_or_replacements": 0, |
| "accounted_outer_allocations": 18, |
| "outer_allocation_gpu_seconds": 23296, |
| "batch_step_gpu_seconds_separate_not_additive": 23308, |
| "preflight_and_diagnostic_gpu_seconds_separate": 728, |
| "results_directory": "benchmark/estimation/results", |
| "accounting_input": "benchmark/estimation/allocation_accounting.json", |
| "accounting_raw_source": "benchmark/estimation/allocation_accounting_sacct.psv" |
| }, |
| "study_resource_cap_unchanged_reserved_gpu_hours": 66, |
| "preflight_cost_reported_separately": true |
| } |
|
|
