{ "title": "Qwen3-Omni Full-Parameter Feasibility Gates", "generated_at_utc": "2026-06-13T18:14:32+00:00", "status": "pass", "decision": "full_parameter_feasible_for_guarded_short_runs_not_promoted", "interpretation": "The full-parameter gates prove that Qwen3-Omni full-parameter FSDP can load, prepare, run backward/optimizer steps, and complete guarded pilots up to 256 optimizer steps on an 8-GPU remote worker. They do not prove a production full-parameter fine-tune, and they intentionally save no full checkpoints or public weights.", "aggregate": { "run_count": 7, "passed_run_count": 6, "preempted_run_count": 1, "review_or_missing_run_count": 0, "completed_full_parameter_train_steps": 489, "longest_passed_run_id": "xperience10m_qwen3_omni_128ep_fullparam_pilot256_after_qwen_v6_preemptible_8gpu_20260611", "longest_passed_steps": 256, "num_processes": [ 8 ], "checkpoint_saved": false }, "runs": [ { "id": "fullparam_smoke_1step", "title": "Full-Parameter 1-Step Feasibility Smoke", "status": "passed", "scope": "1 optimizer step over 8 train samples", "summary_path": "results/omni_finetune/xperience10m_qwen3_omni_128ep_fullparam_smoke_preemptible_8gpu_20260609/fullparam_feasibility_summary.json", "run_id": "xperience10m_qwen3_omni_128ep_fullparam_smoke_preemptible_8gpu_20260609", "purpose": "Full-parameter Qwen3-Omni feasibility gate: load, FSDP prepare, backward/optimizer step, and no checkpoint save.", "tuning_mode": "full", "training_objective": null, "num_processes": 8, "num_train_samples": 8, "configured_max_train_steps": 1, "observed_train_steps": 1, "first_step_loss": null, "final_step_loss": 1.2726006507873535, "epoch_train_loss": 1.2726006507873535, "min_step_loss": null, "max_step_loss": null, "model_load_seconds": null, "accelerator_prepare_seconds": null, "train_loop_seconds": null, "save_mode": "none", "checkpoint_saved": false, "checkpoint_policy": "no full-parameter checkpoint or public weights; save_mode=none", "preempt_event": null, "parent_resume_event": null }, { "id": "fullparam_shorttrain8", "title": "Full-Parameter 8-Step Short Train", "status": "passed", "scope": "8 optimizer steps over 64 train samples", "summary_path": "results/omni_finetune/xperience10m_qwen3_omni_128ep_fullparam_shorttrain8_preemptible_8gpu_20260609/fullparam_shorttrain8_summary.json", "run_id": "xperience10m_qwen3_omni_128ep_fullparam_shorttrain8_preemptible_8gpu_20260609", "purpose": "guarded full-parameter Qwen3-Omni short train using all 8 remote GPUs while Qwen v5 export was CPU-bound", "tuning_mode": "full", "training_objective": "structured_episode_understanding_json_qa", "num_processes": 8, "num_train_samples": 64, "configured_max_train_steps": 8, "observed_train_steps": 8, "first_step_loss": 1.2701153755187988, "final_step_loss": 1.180522084236145, "epoch_train_loss": 1.2190196067094803, "min_step_loss": 1.1532728672027588, "max_step_loss": 1.2701153755187988, "model_load_seconds": 111.21318578720093, "accelerator_prepare_seconds": null, "train_loop_seconds": null, "save_mode": "none", "checkpoint_saved": false, "checkpoint_policy": "no full-parameter checkpoint or public weights; save_mode=none", "preempt_event": null, "parent_resume_event": null }, { "id": "fullparam_pilot32", "title": "Full-Parameter 32-Step Pilot", "status": "passed", "scope": "32 optimizer steps over 256 train samples", "summary_path": "results/omni_finetune/xperience10m_qwen3_omni_128ep_fullparam_pilot32_preemptible_8gpu_20260609/fullparam_pilot32_summary.json", "run_id": "xperience10m_qwen3_omni_128ep_fullparam_pilot32_preemptible_8gpu_20260609", "purpose": "guarded full-parameter Qwen3-Omni 32-step pilot using all 8 remote GPUs while Qwen v5 export was CPU-bound", "tuning_mode": "full", "training_objective": "structured_episode_understanding_json_qa", "num_processes": 8, "num_train_samples": 256, "configured_max_train_steps": 32, "observed_train_steps": 32, "first_step_loss": 1.2583622932434082, "final_step_loss": 0.2206273376941681, "epoch_train_loss": 0.8451133379712701, "min_step_loss": 0.2206273376941681, "max_step_loss": 1.26246976852417, "model_load_seconds": 116.43117642402649, "accelerator_prepare_seconds": 1122.333820104599, "train_loop_seconds": 440.2827479839325, "save_mode": "none", "checkpoint_saved": false, "checkpoint_policy": "no full-parameter checkpoint or public weights; save_mode=none", "preempt_event": null, "parent_resume_event": null }, { "id": "fullparam_pilot64", "title": "Full-Parameter 64-Step Pilot", "status": "passed", "scope": "64 optimizer steps over 512 train samples", "summary_path": "results/omni_finetune/xperience10m_qwen3_omni_128ep_fullparam_pilot64_preemptible_8gpu_20260609/fullparam_pilot64_summary.json", "run_id": "xperience10m_qwen3_omni_128ep_fullparam_pilot64_preemptible_8gpu_20260609", "purpose": "guarded full-parameter Qwen3-Omni 64-step pilot using all 8 remote GPUs while Qwen v5 long multiscale export was CPU-bound", "tuning_mode": "full", "training_objective": "structured_episode_understanding_json_qa", "num_processes": 8, "num_train_samples": 512, "configured_max_train_steps": 64, "observed_train_steps": 64, "first_step_loss": 1.2597532272338867, "final_step_loss": 0.011219973675906658, "epoch_train_loss": 0.4434075650788145, "min_step_loss": 0.011219973675906658, "max_step_loss": 1.2597532272338867, "model_load_seconds": 80.75278043746948, "accelerator_prepare_seconds": 811.6013672351837, "train_loop_seconds": 903.8602497577667, "save_mode": "none", "checkpoint_saved": false, "checkpoint_policy": "no full-parameter checkpoint or public weights; save_mode=none", "preempt_event": null, "parent_resume_event": null }, { "id": "fullparam_pilot128_preempted", "title": "Full-Parameter 128-Step Opportunistic Pilot", "status": "preempted_for_qwen_v5_handoff", "scope": "planned 128 optimizer steps over 1024 train samples; preempted for Qwen v5 handoff", "summary_path": "results/omni_finetune/xperience10m_qwen3_omni_128ep_fullparam_pilot128_preemptible_8gpu_20260609/fullparam_pilot128_summary.json", "run_id": "xperience10m_qwen3_omni_128ep_fullparam_pilot128_preemptible_8gpu_20260609", "purpose": "opportunistic guarded full-parameter Qwen3-Omni 128-step pilot launched only while Qwen v5 export was CPU-bound; preempted for main Qwen v5 handoff", "tuning_mode": "full", "training_objective": "structured_episode_understanding_json_qa", "num_processes": 8, "num_train_samples": 1024, "configured_max_train_steps": 128, "observed_train_steps": 0, "first_step_loss": null, "final_step_loss": null, "epoch_train_loss": null, "min_step_loss": null, "max_step_loss": null, "model_load_seconds": 65.33229899406433, "accelerator_prepare_seconds": null, "train_loop_seconds": null, "save_mode": "none", "checkpoint_saved": false, "checkpoint_policy": "no full-parameter checkpoint or public weights; save_mode=none", "preempt_event": { "event": "fullparam_pilot128_preempt_for_qwen_v5_handoff", "guard_pid": 3757690, "kind": "fullparam_pilot128_handoff_guard", "time": 1780954910, "watched_pid": 3753782 }, "parent_resume_event": { "event": "fullparam_pilot128_resumed_parent_after_preempt", "guard_pid": 3757690, "kind": "fullparam_pilot128_handoff_guard", "time": 1780954925, "watched_pid": 3753782 } }, { "id": "fullparam_pilot128_after_qwen_v5", "title": "Full-Parameter 128-Step Post-Qwen-v5 Pilot", "status": "passed", "scope": "128 optimizer steps over 1024 train samples after verified Qwen v5 handoff", "summary_path": "results/omni_finetune/xperience10m_qwen3_omni_128ep_fullparam_pilot128_after_qwen_v5_preemptible_8gpu_20260609/training_metadata.json", "progress_path": "results/omni_finetune/xperience10m_qwen3_omni_128ep_fullparam_pilot128_after_qwen_v5_preemptible_8gpu_20260609/progress.jsonl", "run_id": "xperience10m_qwen3_omni_128ep_fullparam_pilot128_after_qwen_v5_preemptible_8gpu_20260609", "purpose": "post_verified_qwen_v5_full_parameter_feasibility_pilot", "tuning_mode": "full", "training_objective": "structured_episode_understanding_json_qa", "num_processes": 8, "num_train_samples": 1024, "configured_max_train_steps": 128, "observed_train_steps": 128, "first_step_loss": 1.2273844480514526, "final_step_loss": 0.0136940386146307, "epoch_train_loss": 0.21579630990163423, "min_step_loss": 0.004702376667410135, "max_step_loss": 1.2273844480514526, "model_load_seconds": null, "accelerator_prepare_seconds": null, "train_loop_seconds": null, "save_mode": "none", "checkpoint_saved": false, "checkpoint_policy": "no full-parameter checkpoint or public weights; save_mode=none", "preempt_event": null, "parent_resume_event": null, "progress_events": { "max_steps_reached": true, "save_skipped": true, "complete": true } }, { "id": "fullparam_pilot256_after_qwen_v6", "title": "Full-Parameter 256-Step Post-Qwen-v6 Pilot", "status": "passed", "scope": "256 optimizer steps over 2048 train samples after verified Qwen v6 handoff", "summary_path": "results/omni_finetune/xperience10m_qwen3_omni_128ep_fullparam_pilot256_after_qwen_v6_preemptible_8gpu_20260611/training_metadata.json", "progress_path": "results/omni_finetune/xperience10m_qwen3_omni_128ep_fullparam_pilot256_after_qwen_v6_preemptible_8gpu_20260611/progress.jsonl", "run_id": "xperience10m_qwen3_omni_128ep_fullparam_pilot256_after_qwen_v6_preemptible_8gpu_20260611", "purpose": "post_verified_qwen_v6_full_parameter_feasibility_pilot", "tuning_mode": "full", "training_objective": "structured_episode_understanding_json_qa", "num_processes": 8, "num_train_samples": 2048, "configured_max_train_steps": 256, "observed_train_steps": 256, "first_step_loss": 1.2249484062194824, "final_step_loss": 0.009560374543070793, "epoch_train_loss": 0.1157912792496063, "min_step_loss": 0.002165592508390546, "max_step_loss": 1.2249484062194824, "model_load_seconds": null, "accelerator_prepare_seconds": null, "train_loop_seconds": null, "save_mode": "none", "checkpoint_saved": false, "checkpoint_policy": "no full-parameter checkpoint or public weights; save_mode=none", "preempt_event": null, "parent_resume_event": null, "progress_events": { "max_steps_reached": true, "save_skipped": true, "complete": true } } ], "publication_policy": { "public_summary_allowed": true, "publish_full_parameter_weights": false, "publish_full_checkpoints": false, "reason": "All completed full-parameter gate runs used save_mode=none; the preempted pilot saved nothing. These are feasibility evidence only." }, "next_steps": [ "Keep the verified Qwen3-Omni LoRA adapter as the published production result for the 128-episode suite.", "For a production full-parameter run, add a sharded checkpoint/resume plan before any long training launch.", "Run a separate checkpointed full-parameter pilot only when GPUs are not needed by verified LoRA evaluation/publication work." ] }