{
  "repo_id": "FlyinGodzilla/AMOR-GatedDeltaNet-1.5B",
  "source_filename": "amor_gdn_1p5b_step_0062558.pt",
  "source_sha256": "a635e9fbf742f17f71b607565440ed8c338a90a00173b1f0813e79806e45f0f7",
  "step": 62558,
  "tokens_seen": 30748520448,
  "n_params": 1524025472,
  "tensor_dtypes": {
    "torch.float32": 458
  },
  "gate_buffers": {
    "0": {
      "running_threshold": 0.9387496709823608,
      "running_std": 0.028903765603899956
    },
    "1": {
      "running_threshold": 0.9252517819404602,
      "running_std": 0.0452539287507534
    },
    "2": {
      "running_threshold": 0.9380253553390503,
      "running_std": 0.027104970067739487
    }
  },
  "optimizer_included": false,
  "precision_changed": false,
  "validation": {
    "exported_tensors_equal": true,
    "strict_model_load": true,
    "generation": true,
    "cuda": {
      "versions": {
        "torch": "2.6.0+cu124",
        "transformers": "5.0.0",
        "mamba-ssm": "2.3.0",
        "causal-conv1d": "1.6.0",
        "flash-linear-attention": "0.4.2",
        "triton": "3.2.0",
        "safetensors": "0.7.0"
      },
      "tokenizer_revision": "5590cc63198b0f111f4289daf006da0a12044173",
      "strict_model_load": true,
      "generation": true,
      "device": "NVIDIA H100 NVL",
      "torch": "2.6.0+cu124",
      "autocast": "bfloat16",
      "prefill_max_abs_diff": 0.0,
      "decode_max_abs_diff": 0.0625,
      "generated_token_ids": [
        264,
        502,
        4668,
        304,
        279,
        469,
        1950,
        430
      ],
      "published_cli_passed": true,
      "test_scope": "Short CUDA smoke test; not a benchmark rerun",
      "comparison_tolerance": {
        "rtol": 0.02,
        "atol": 0.1
      }
    },
    "router_release": {
      "exact_router_export": true,
      "base_model_association": true,
      "default_forward_max_abs_diff": 0.0,
      "modes": [
        {
          "router": false,
          "cases": [
            {
              "batch": 1,
              "length": 16,
              "prefill_max_abs_diff": 0.0,
              "decode_max_abs_diff": 0.0625
            },
            {
              "batch": 1,
              "length": 128,
              "prefill_max_abs_diff": 0.03125,
              "decode_max_abs_diff": 0.125
            },
            {
              "batch": 2,
              "length": 64,
              "prefill_max_abs_diff": 0.0
            }
          ],
          "canonical_router_exact": null
        },
        {
          "router": true,
          "cases": [
            {
              "batch": 1,
              "length": 16,
              "prefill_max_abs_diff": 0.0,
              "decode_max_abs_diff": 0.0625
            },
            {
              "batch": 1,
              "length": 128,
              "prefill_max_abs_diff": 0.0,
              "decode_max_abs_diff": 0.125
            },
            {
              "batch": 2,
              "length": 64,
              "prefill_max_abs_diff": 0.0
            }
          ],
          "canonical_router_exact": true
        }
      ],
      "frozen_thresholds_preserved": true,
      "training_router_rejected": true,
      "gpu": "NVIDIA H100 NVL",
      "torch": "2.6.0+cu124",
      "generation_cli_modes_passed": [
        "entropy",
        "router"
      ]
    },
    "current_release": {
      "strict_model_load": true,
      "entropy_and_router_generation_passed": true,
      "unit_tests_passed": 17,
      "device": "NVIDIA H100 NVL"
    }
  },
  "implementation": "AMOR shared release: native entropy gate and optional fitted inference router",
  "router": {
    "variant": "distill",
    "hidden_dim": 512,
    "activation": "silu",
    "weights": "router.safetensors",
    "config": "router_config.json",
    "sha256": "f67c3a65f7b92cb884d55cedaf49d39dca9a85b7a205c2492c9f8d2fc5b0fb9b",
    "base_weight_sha256": "eec1f7d2e406b4df732714cb66ec968cd2bfa5c3f24888f3141e51ae1f001bdd",
    "precision": "float32",
    "inference_only": true
  },
  "weight_layout": {
    "format": "safetensors-sharded-v1",
    "index": "model.safetensors.index.json",
    "shards": 8,
    "weight_identity_sha256": "eec1f7d2e406b4df732714cb66ec968cd2bfa5c3f24888f3141e51ae1f001bdd",
    "original_single_file_sha256": "87f2506b6d1f4ac36936b715920e80e348091fa9620391267a67ef27d17c7be5",
    "all_tensor_bytes_identical": true
  }
}