{
  "reason": "MPS dynamic-length memory growth to80GB physical footprint and increasing step times; stopped before any heldout evaluation",
  "completed_checkpoint": false,
  "action": "Restart both sizes from base with128-token masked padding and explicit unusedMPS cache release each optimizerstep, after development/training-only diagnostic. Frozeninputs,optimizer,epochs,seed unchanged.",
  "partial_attempt_status": {
    "state": "failed",
    "started": "2026-09-21T00:39:03-0700",
    "stages": [
      {
        "name": "train1000",
        "command": [
          "train_public_laya.py",
          "--data",
          "/Users/alexanderollman/jevtest/train/moderation-study/aegis_1000",
          "--out",
          "/Users/alexanderollman/jevtest/models/laya-moderation-aegis-1000-s7"
        ],
        "started": 1789976343.805696,
        "pid": 23292,
        "exit_code": -2,
        "seconds": 1120.7
      }
    ]
  },
  "second_attempt": {
    "status": {
      "state": "failed",
      "started": "2026-09-21T01:02:58-0700",
      "stages": [
        {
          "name": "train1000",
          "command": [
            "train_public_laya.py",
            "--data",
            "/Users/alexanderollman/jevtest/train/moderation-study/aegis_1000",
            "--out",
            "/Users/alexanderollman/jevtest/models/laya-moderation-aegis-1000-s7",
            "--pad-to-multiple",
            "128"
          ],
          "started": 1789977778.668488,
          "pid": 6144,
          "exit_code": -2,
          "seconds": 792.6
        }
      ]
    },
    "reason": "Masked padding alone did not bound retained allocator cache; driver memory reached106GB with6.8GB live tensors and step36 took119seconds. Stopped before checkpoint or heldout scoring."
  }
}