File size: 12,181 Bytes
98cf463 69865f3 98cf463 63c5594 98cf463 728eb2a 98cf463 728eb2a 98cf463 728eb2a 63c5594 728eb2a 98cf463 63c5594 98cf463 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 | {
"title": "Qwen3-Omni Full-Parameter Feasibility Gates",
"generated_at_utc": "2026-06-18T12:53:13+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."
]
}
|