OUTPUT of run 1 of at most 60:
```
=== ARCHITECTURE ===
A config:
{
  "_sliding_window_pattern": 6,
  "architectures": [
    "Gemma3ForCausalLM"
  ],
  "attention_bias": false,
  "attention_dropout": 0.0,
  "attn_logit_softcapping": null,
  "bos_token_id": 2,
  "dtype": "float32",
  "eos_token_id": 1,
  "final_logit_softcapping": null,
  "head_dim": 256,
  "hidden_activation": "gelu_pytorch_tanh",
  "hidden_size": 640,
  "initializer_range": 0.02,
  "intermediate_size": 2048,
  "layer_types": [
    "sliding_attention",
    "sliding_attention",
    "sliding_attention",
    "sliding_attention",
    "sliding_attention",
    "full_attention",
    "sliding_attention",
    "sliding_attention",
    "sliding_attention",
    "sliding_attention",
    "sliding_attention",
    "full_attention",
    "sliding_attention",
    "sliding_attention",
    "sliding_attention",
    "sliding_attention",
    "sliding_attention",
    "full_attention"
  ],
  "max_position_embeddings": 32768,
  "model_type": "gemma3_text",
  "num_attention_heads": 4,
  "num_hidden_layers": 18,
  "num_key_value_heads": 1,
  "pad_token_id": 0,
  "query_pre_attn_scalar": 256,
  "rms_norm_eps": 1e-06,
  "rope_parameters": {
    "full_attention": {
      "rope_theta": 1000000.0,
      "rope_type": "default"
    },
    "sliding_attention": {
      "rope_theta": 10000.0,
      "rope_type": "default"
    }
  },
  "sliding_window": 512,
  "tie_word_embeddings": true,
  "transformers_version": "5.20.2",
  "use_bidirectional_attention": false,
  "use_cache": true,
  "vocab_size": 262144
}
Config differences: {}

Keys only in A: []
Keys only in B: []
Shape mismatches: []

=== TENSOR INVENTORY: suffix, shapes and counts ===
input_layernorm.weight {'(640,)': 18}
mlp.down_proj.weight {'(640, 2048)': 18}
mlp.gate_proj.weight {'(2048, 640)': 18}
mlp.up_proj.weight {'(2048, 640)': 18}
model.norm.weight {'(640,)': 1}
post_attention_layernorm.weight {'(640,)': 18}
post_feedforward_layernorm.weight {'(640,)': 18}
pre_feedforward_layernorm.weight {'(640,)': 18}
self_attn.k_norm.weight {'(256,)': 18}
self_attn.k_proj.weight {'(256, 640)': 18}
self_attn.o_proj.weight {'(640, 1024)': 18}
self_attn.q_norm.weight {'(256,)': 18}
self_attn.q_proj.weight {'(1024, 640)': 18}
self_attn.v_proj.weight {'(256, 640)': 18}

=== OVERALL ===
Tensors: 235
Present scalar parameters: 100326016
Exactly differing scalar parameters: 79105310
Differing fraction: 0.7884825208249074
Exactly identical tensors: 31
Global ||B-A||_F: 2.42016001485264

=== LAYER SUMMARIES ===
layer    parameters   changed_frac   ||A||       ||B||       ||B-A||    symmetric_rel_delta   delta_energy_share
   -1           640      0.000000   1355.7502   1355.7502     0.00000             0.0000000             0.000000
    0       5573632      0.762054    419.4822    419.4821     0.55154             0.0013148             0.051935
    1       5573632      0.784188   1062.3094   1062.3094     0.55341             0.0005210             0.052288
    2       5573632      0.771561   1083.3551   1083.3551     0.54919             0.0005069             0.051493
    3       5573632      0.792763   1440.1118   1440.1118     0.54891             0.0003812             0.051442
    4       5573632      0.799995   1533.3491   1533.3491     0.55590             0.0003625             0.052760
    5       5573632      0.785135   1257.0027   1257.0027     0.56914             0.0004528             0.055303
    6       5573632      0.792085   1241.4591   1241.4591     0.57011             0.0004592             0.055491
    7       5573632      0.782902   1200.2559   1200.2559     0.55799             0.0004649             0.053158
    8       5573632      0.792063   1385.0315   1385.0315     0.55204             0.0003986             0.052030
    9       5573632      0.783840   1460.6573   1460.6572     0.57116             0.0003910             0.055696
   10       5573632      0.791970   2451.3727   2451.3727     0.57874             0.0002361             0.057185
   11       5573632      0.785140   5975.7999   5975.7999     0.56631             0.0000948             0.054754
   12       5573632      0.791623   7331.0670   7331.0670     0.58533             0.0000798             0.058495
   13       5573632      0.798250   6871.1478   6871.1478     0.58860             0.0000857             0.059149
   14       5573632      0.798190   7863.2412   7863.2412     0.59468             0.0000756             0.060378
   15       5573632      0.791949   9694.1604   9694.1604     0.60416             0.0000623             0.062318
   16       5573632      0.795049   9574.5528   9574.5528     0.59075             0.0000617             0.059583
   17       5573632      0.794017  10309.4843  10309.4843     0.57547             0.0000558             0.056541

=== TOP 20 BY ABSOLUTE DIFFERENCE (NOT CAUSAL IMPORTANCE) ===
tensor | shape | changed_frac | ||B-A|| | symmetric_rel_delta | cosine
model.layers.14.mlp.gate_proj.weight | (2048, 640) | 0.764654 | 0.298669 | 0.00668544 | 0.999977652
model.layers.15.mlp.up_proj.weight | (2048, 640) | 0.746013 | 0.298319 | 0.006677103 | 0.999977708
model.layers.15.mlp.gate_proj.weight | (2048, 640) | 0.764095 | 0.298282 | 0.006673806 | 0.999977730
model.layers.14.mlp.up_proj.weight | (2048, 640) | 0.744450 | 0.296389 | 0.00664335 | 0.999977933
model.layers.16.mlp.up_proj.weight | (2048, 640) | 0.744233 | 0.295437 | 0.006600201 | 0.999978219
model.layers.16.mlp.gate_proj.weight | (2048, 640) | 0.760617 | 0.294908 | 0.006614133 | 0.999978127
model.layers.13.mlp.gate_proj.weight | (2048, 640) | 0.759324 | 0.292737 | 0.006554407 | 0.999978520
model.layers.13.mlp.up_proj.weight | (2048, 640) | 0.742742 | 0.292338 | 0.006557245 | 0.999978501
model.layers.12.mlp.gate_proj.weight | (2048, 640) | 0.767620 | 0.291493 | 0.006510115 | 0.999978809
model.layers.17.mlp.up_proj.weight | (2048, 640) | 0.743374 | 0.290303 | 0.006465188 | 0.999979101
model.layers.17.mlp.gate_proj.weight | (2048, 640) | 0.765401 | 0.290025 | 0.00657134 | 0.999978409
model.layers.6.mlp.gate_proj.weight | (2048, 640) | 0.761549 | 0.286932 | 0.006417006 | 0.999979411
model.layers.12.mlp.up_proj.weight | (2048, 640) | 0.746779 | 0.286725 | 0.006414129 | 0.999979429
model.layers.15.mlp.down_proj.weight | (640, 2048) | 0.886674 | 0.284161 | 0.01548496 | 0.999880108
model.layers.11.mlp.gate_proj.weight | (2048, 640) | 0.760811 | 0.283441 | 0.006330282 | 0.999979964
model.layers.10.mlp.gate_proj.weight | (2048, 640) | 0.763615 | 0.283137 | 0.006313775 | 0.999980068
model.layers.5.mlp.gate_proj.weight | (2048, 640) | 0.758973 | 0.283041 | 0.006315441 | 0.999980058
model.layers.9.mlp.gate_proj.weight | (2048, 640) | 0.762288 | 0.281397 | 0.006269262 | 0.999980348
model.layers.7.mlp.gate_proj.weight | (2048, 640) | 0.758185 | 0.281146 | 0.006276537 | 0.999980303
model.layers.9.mlp.up_proj.weight | (2048, 640) | 0.734996 | 0.280847 | 0.006290321 | 0.999980216

=== TOP 20 BY RELATIVE DIFFERENCE (NOT CAUSAL IMPORTANCE) ===
tensor | shape | changed_frac | ||B-A|| | symmetric_rel_delta | cosine
model.layers.4.mlp.down_proj.weight | (640, 2048) | 0.915636 | 0.254494 | 0.0182807 | 0.999832908
model.layers.14.mlp.down_proj.weight | (640, 2048) | 0.903156 | 0.278398 | 0.01807846 | 0.999836585
model.layers.8.mlp.down_proj.weight | (640, 2048) | 0.895577 | 0.246041 | 0.01658212 | 0.999862517
model.layers.1.mlp.down_proj.weight | (640, 2048) | 0.895543 | 0.258052 | 0.01657688 | 0.999862604
model.layers.10.mlp.down_proj.weight | (640, 2048) | 0.898067 | 0.257038 | 0.01649976 | 0.999863879
model.layers.13.mlp.down_proj.weight | (640, 2048) | 0.894163 | 0.272982 | 0.01646839 | 0.999864396
model.layers.3.mlp.down_proj.weight | (640, 2048) | 0.896035 | 0.252629 | 0.01590446 | 0.999873525
model.layers.15.mlp.down_proj.weight | (640, 2048) | 0.886674 | 0.284161 | 0.01548496 | 0.999880108
model.layers.6.mlp.down_proj.weight | (640, 2048) | 0.882903 | 0.251827 | 0.01471984 | 0.999891663
model.layers.17.mlp.down_proj.weight | (640, 2048) | 0.880357 | 0.263536 | 0.01442912 | 0.999895900
model.layers.16.mlp.down_proj.weight | (640, 2048) | 0.876293 | 0.278261 | 0.01409358 | 0.999900686
model.layers.2.mlp.down_proj.weight | (640, 2048) | 0.875922 | 0.252326 | 0.01360399 | 0.999907466
model.layers.7.mlp.down_proj.weight | (640, 2048) | 0.867943 | 0.249428 | 0.01303433 | 0.999915053
model.layers.9.mlp.down_proj.weight | (640, 2048) | 0.857288 | 0.254224 | 0.01203169 | 0.999927619
model.layers.11.mlp.down_proj.weight | (640, 2048) | 0.856242 | 0.245844 | 0.01197049 | 0.999928354
model.layers.12.mlp.down_proj.weight | (640, 2048) | 0.855425 | 0.265868 | 0.01174139 | 0.999931070
model.layers.3.self_attn.o_proj.weight | (640, 1024) | 0.848694 | 0.174264 | 0.0108386 | 0.999941263
model.layers.4.self_attn.o_proj.weight | (640, 1024) | 0.858575 | 0.182444 | 0.01061527 | 0.999943658
model.layers.0.mlp.down_proj.weight | (640, 2048) | 0.840711 | 0.260645 | 0.01058558 | 0.999943973
model.layers.16.self_attn.o_proj.weight | (640, 1024) | 0.839386 | 0.183825 | 0.01053135 | 0.999944545

=== NON-LAYER TENSORS ===
model.norm.weight (640,) relative_delta= 0.0 changed= 0

No inputs, hidden-state evaluations, nonlinearities, or model execution were used.

```