docs/development-artifacts/loop-learning-development-v1/preflight/diagnostic-gpu.json

{
  "kind": "current_weights_occurrence_gradient_geometry",
  "arm": "shared",
  "base_unique_blocks": 2,
  "executed_blocks": 4,
  "record_count": 2,
  "record_ids": [
    "0002f4903b87da3069087041",
    "000e12818b092204b897db23"
  ],
  "encoded_records_sha256": "89be38ecbd94a64ba7db12c4d43a5850594589f7fee191f149e255c7f6c344cd",
  "training_batch_selection": "fixed training-only batch supplied by the caller",
  "loss_definition": "mean of per-example mean answer-and-EOS cross entropy",
  "dropout_enabled": false,
  "parameter_dtype": "torch.float32",
  "autocast_enabled": false,
  "reference_loss": 10.937704086303711,
  "expanded_loss": 10.937704086303711,
  "forward_maximum_absolute_error": 0.0,
  "loss_absolute_error": 0.0,
  "gradient_reconstruction": {
    "maximum_absolute_error": 2.9802322387695312e-08,
    "maximum_relative_l2_error": 3.227223353537611e-08,
    "unique_parameter_tensors": 28,
    "actually_tied_parameter_tensors": 24,
    "trainable_parameter_coordinates": 14511360,
    "interpretation": "Each reference gradient equals the sum over its actual execution aliases. For an untied parameter this is a one-occurrence equality."
  },
  "groups": [
    {
      "base_block": 0,
      "component": "attention",
      "executed_block_indices": [
        0,
        2
      ],
      "loop_iterations": [
        0,
        1
      ],
      "actual_unique_modules": 1,
      "actually_shared_across_occurrences": true,
      "shared_gradient_maximum_absolute_error": 2.9802322387695312e-08,
      "occurrence_gradient_norms": [
        5.884517848774616,
        3.3358956273221856
      ],
      "gradient_dot_products": [
        [
          34.62755031254703,
          4.81165064074796
        ],
        [
          4.81165064074796,
          11.128199636387277
        ]
      ],
      "gradient_cosines": [
        [
          1.0,
          0.24511548504999212
        ],
        [
          0.24511548504999212,
          1.0
        ]
      ],
      "summed_gradient_norm": 7.441710235586322,
      "summed_gradient_squared_norm": 55.37905123043022,
      "sum_individual_squared_norms": 45.75574994893431,
      "twice_pairwise_inner_product_sum": 9.623301281495912,
      "squared_norm_ratio": 1.210318949907629,
      "gradient_coordinates": 263168
    },
    {
      "base_block": 0,
      "component": "mlp",
      "executed_block_indices": [
        0,
        2
      ],
      "loop_iterations": [
        0,
        1
      ],
      "actual_unique_modules": 1,
      "actually_shared_across_occurrences": true,
      "shared_gradient_maximum_absolute_error": 2.9802322387695312e-08,
      "occurrence_gradient_norms": [
        11.135506360484815,
        6.773574624507856
      ],
      "gradient_dot_products": [
        [
          123.99950190439775,
          33.52081695925861
        ],
        [
          33.52081695925861,
          45.88131319377674
        ]
      ],
      "gradient_cosines": [
        [
          0.9999999999999999,
          0.44441294883734056
        ],
        [
          0.44441294883734056,
          1.0
        ]
      ],
      "summed_gradient_norm": 15.392285373416502,
      "summed_gradient_squared_norm": 236.9224490166916,
      "sum_individual_squared_norms": 169.88081509817448,
      "twice_pairwise_inner_product_sum": 67.04163391851725,
      "squared_norm_ratio": 1.3946392291547083,
      "gradient_coordinates": 525568
    },
    {
      "base_block": 1,
      "component": "attention",
      "executed_block_indices": [
        1,
        3
      ],
      "loop_iterations": [
        0,
        1
      ],
      "actual_unique_modules": 1,
      "actually_shared_across_occurrences": true,
      "shared_gradient_maximum_absolute_error": 2.9802322387695312e-08,
      "occurrence_gradient_norms": [
        4.657955892311537,
        3.407568762723621
      ],
      "gradient_dot_products": [
        [
          21.696553094719764,
          7.090349297373162
        ],
        [
          7.090349297373162,
          11.61152487268979
        ]
      ],
      "gradient_cosines": [
        [
          0.9999999999999999,
          0.44671201182243164
        ],
        [
          0.44671201182243164,
          1.0
        ]
      ],
      "summed_gradient_norm": 6.891210094181998,
      "summed_gradient_squared_norm": 47.48877656215586,
      "sum_individual_squared_norms": 33.30807796740955,
      "twice_pairwise_inner_product_sum": 14.180698594746325,
      "squared_norm_ratio": 1.4257435270993866,
      "gradient_coordinates": 263168
    },
    {
      "base_block": 1,
      "component": "mlp",
      "executed_block_indices": [
        1,
        3
      ],
      "loop_iterations": [
        0,
        1
      ],
      "actual_unique_modules": 1,
      "actually_shared_across_occurrences": true,
      "shared_gradient_maximum_absolute_error": 1.4901161193847656e-08,
      "occurrence_gradient_norms": [
        7.9838843686438175,
        6.226469746422302
      ],
      "gradient_dot_products": [
        [
          63.74240961187508,
          28.91817685628013
        ],
        [
          28.91817685628013,
          38.768925503112214
        ]
      ],
      "gradient_cosines": [
        [
          0.9999999999999999,
          0.5817210626292507
        ],
        [
          0.5817210626292507,
          1.0
        ]
      ],
      "summed_gradient_norm": 12.66284679002109,
      "summed_gradient_squared_norm": 160.34768882754742,
      "sum_individual_squared_norms": 102.5113351149873,
      "twice_pairwise_inner_product_sum": 57.83635371256025,
      "squared_norm_ratio": 1.5641947170787005,
      "gradient_coordinates": 525568
    }
  ],
  "interpretation": "Local derivatives at the same current weights. Attention and MLP reports exclude their LayerNorm parameters; the complete gradient reconstruction includes LayerNorm, embeddings and output readout. Sums for independent modules describe a coordinated perturbation, not an actual shared parameter. No optimizer update, held-out prediction, or causal claim about learning efficiency is made by this diagnostic.",
  "diagnostic_wall_seconds": 0.7275393940508366
}