=== CONFIG A ===
{
  "_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: []

=== TENSOR LAYOUT: FIRST LAYER AND NON-LAYER TENSORS ===
lm_head.weight (262144, 640) BF16
model.embed_tokens.weight (262144, 640) BF16
model.layers.0.input_layernorm.weight (640,) BF16
model.layers.0.mlp.down_proj.weight (640, 2048) BF16
model.layers.0.mlp.gate_proj.weight (2048, 640) BF16
model.layers.0.mlp.up_proj.weight (2048, 640) BF16
model.layers.0.post_attention_layernorm.weight (640,) BF16
model.layers.0.post_feedforward_layernorm.weight (640,) BF16
model.layers.0.pre_feedforward_layernorm.weight (640,) BF16
model.layers.0.self_attn.k_norm.weight (256,) BF16
model.layers.0.self_attn.k_proj.weight (256, 640) BF16
model.layers.0.self_attn.o_proj.weight (640, 1024) BF16
model.layers.0.self_attn.q_norm.weight (256,) BF16
model.layers.0.self_attn.q_proj.weight (1024, 640) BF16
model.layers.0.self_attn.v_proj.weight (256, 640) BF16
model.norm.weight (640,) BF16

=== NON-LAYER TENSORS, INCLUDING EMBEDDING IDENTITY ===
lm_head.weight shape (262144, 640) changed 0 / 167772160 ||B-A|| 0.0 relative_difference 0.0
model.embed_tokens.weight shape (262144, 640) changed 0 / 167772160 ||B-A|| 0.0 relative_difference 0.0
model.norm.weight shape (640,) changed 0 / 640 ||B-A|| 0.0 relative_difference 0.0

=== PER-LAYER STATIC DIFFERENCES ===
layer part       parameters changed_frac  ||B-A||  relative_diff  ||A||  ||B||
 0 attention      1638912     0.75389   0.29571      0.005216   56.6955   56.6950
 0 mlp            3932160     0.76593   0.46554      0.006859   67.8686   67.8684
 0 norm/other        2560     0.04141   0.00477      0.000012  410.0546  410.0546
 1 attention      1638912     0.78508   0.29445      0.004621   63.7205   63.7198
 1 mlp            3932160     0.78431   0.46855      0.007165   65.3925   65.3925
 1 norm/other        2560     0.02656   0.00427      0.000004 1058.3784 1058.3784
 2 attention      1638912     0.76083   0.29718      0.005071   58.6053   58.6052
 2 mlp            3932160     0.77652   0.46182      0.006999   65.9808   65.9808
 2 norm/other        2560     0.02109   0.00318      0.000003 1079.7547 1079.7547
 3 attention      1638912     0.79866   0.29629      0.006203   47.7629   47.7629
 3 mlp            3932160     0.79081   0.46207      0.007081   65.2501   65.2500
 3 norm/other        2560     0.01211   0.00278      0.000002 1437.8398 1437.8398
 4 attention      1638912     0.79663   0.29869      0.004986   59.9031   59.9030
 4 mlp            3932160     0.80191   0.46884      0.007236   64.7887   64.7887
 4 norm/other        2560     0.00898   0.00174      0.000001 1530.8081 1530.8081
 5 attention      1638912     0.80262   0.31690      0.004250   74.5611   74.5615
 5 mlp            3932160     0.77835   0.47275      0.006954   67.9788   67.9788
 5 norm/other        2560     0.00742   0.00159      0.000001 1252.9466 1252.9466
 6 attention      1638912     0.78989   0.31917      0.005941   53.7198   53.7196
 6 mlp            3932160     0.79351   0.47239      0.007217   65.4549   65.4548
 6 norm/other        2560     0.00820   0.00209      0.000002 1238.5679 1238.5679
 7 attention      1638912     0.77745   0.30760      0.005382   57.1527   57.1528
 7 mlp            3932160     0.78568   0.46554      0.007048   66.0538   66.0538
 7 norm/other        2560     0.00742   0.00249      0.000002 1197.0734 1197.0734
 8 attention      1638912     0.78384   0.30316      0.005371   56.4418   56.4420
 8 mlp            3932160     0.79600   0.46134      0.007096   65.0143   65.0144
 8 norm/other        2560     0.00898   0.00212      0.000002 1382.3530 1382.3530
 9 attention      1638912     0.78261   0.32176      0.005523   58.2544   58.2540
 9 mlp            3932160     0.78486   0.47190      0.007070   66.7427   66.7419
 9 norm/other        2560     0.01016   0.00199      0.000001 1457.9683 1457.9683
10 attention      1638912     0.77694   0.33535      0.004679   71.6774   71.6771
10 mlp            3932160     0.79874   0.47168      0.007232   65.2214   65.2216
10 norm/other        2560     0.00859   0.00213      0.000001 2449.4564 2449.4564
11 attention      1638912     0.78826   0.32164      0.003811   84.3949   84.3952
11 mlp            3932160     0.78435   0.46610      0.007009   66.5039   66.5038
11 norm/other        2560     0.00703   0.00211      0.000000 5974.8339 5974.8339
12 attention      1638912     0.79688   0.32365      0.005359   60.3915   60.3919
12 mlp            3932160     0.78994   0.48771      0.007258   67.2000   67.2001
12 norm/other        2560     0.00703   0.00179      0.000000 7330.5102 7330.5102
13 attention      1638912     0.79827   0.31742      0.005829   54.4589   54.4588
13 mlp            3932160     0.79874   0.49566      0.007597   65.2469   65.2463
13 norm/other        2560     0.02773   0.00351      0.000001 6870.6222 6870.6222
14 attention      1638912     0.78523   0.31476      0.004778   65.8748   65.8751
14 mlp            3932160     0.80409   0.50453      0.007764   64.9878   64.9876
14 norm/other        2560     0.03633   0.00382      0.000000 7862.6967 7862.6967
15 attention      1638912     0.77643   0.32602      0.005164   63.1300   63.1301
15 mlp            3932160     0.79893   0.50864      0.007729   65.8064   65.8061
15 norm/other        2560     0.01172   0.00253      0.000000 9693.7315 9693.7315
16 attention      1638912     0.79948   0.31193      0.005640   55.3094   55.3094
16 mlp            3932160     0.79371   0.50168      0.007579   66.1931   66.1925
16 norm/other        2560     0.00937   0.00254      0.000000 9574.1642 9574.1642
17 attention      1638912     0.78958   0.30548      0.004623   66.0831   66.0830
17 mlp            3932160     0.79638   0.48769      0.007439   65.5570   65.5566
17 norm/other        2560     0.01094   0.00235      0.000000 10309.0641 10309.0641

=== TOP 24 TENSORS BY RELATIVE DIFFERENCE (NOT CAUSAL IMPORTANCE) ===
model.layers.4.mlp.down_proj.weight                               rel=0.018281 rms=0.0002223 max=0.0012207 changed=0.9156
model.layers.14.mlp.down_proj.weight                              rel=0.018078 rms=0.0002432 max=0.0013428 changed=0.9032
model.layers.8.mlp.down_proj.weight                               rel=0.016582 rms=0.0002149 max=0.0012207 changed=0.8956
model.layers.1.mlp.down_proj.weight                               rel=0.016577 rms=0.0002254 max=0.0012207 changed=0.8955
model.layers.10.mlp.down_proj.weight                              rel=0.016500 rms=0.0002245 max=0.0019531 changed=0.8981
model.layers.13.mlp.down_proj.weight                              rel=0.016468 rms=0.0002384 max=0.0011902 changed=0.8942
model.layers.3.mlp.down_proj.weight                               rel=0.015904 rms=0.0002207 max=0.0019531 changed=0.8960
model.layers.15.mlp.down_proj.weight                              rel=0.015485 rms=0.0002482 max=0.0012817 changed=0.8867
model.layers.6.mlp.down_proj.weight                               rel=0.014720 rms=0.0002200 max=0.0010986 changed=0.8829
model.layers.17.mlp.down_proj.weight                              rel=0.014429 rms=0.0002302 max=0.0012817 changed=0.8804
model.layers.16.mlp.down_proj.weight                              rel=0.014094 rms=0.0002431 max=0.0012512 changed=0.8763
model.layers.2.mlp.down_proj.weight                               rel=0.013604 rms=0.0002204 max=0.0019531 changed=0.8759
model.layers.7.mlp.down_proj.weight                               rel=0.013034 rms=0.0002179 max=0.0012207 changed=0.8679
model.layers.9.mlp.down_proj.weight                               rel=0.012032 rms=0.0002221 max=0.0012207 changed=0.8573
model.layers.11.mlp.down_proj.weight                              rel=0.011970 rms=0.0002147 max=0.0010986 changed=0.8562
model.layers.12.mlp.down_proj.weight                              rel=0.011741 rms=0.0002322 max=0.0011597 changed=0.8554
model.layers.3.self_attn.o_proj.weight                            rel=0.010839 rms=0.0002153 max=0.0010986 changed=0.8487
model.layers.4.self_attn.o_proj.weight                            rel=0.010615 rms=0.0002254 max=0.0011082 changed=0.8586
model.layers.0.mlp.down_proj.weight                               rel=0.010586 rms=0.0002277 max=0.0010986 changed=0.8407
model.layers.16.self_attn.o_proj.weight                           rel=0.010531 rms=0.0002271 max=0.0011597 changed=0.8394
model.layers.5.mlp.down_proj.weight                               rel=0.010374 rms=0.0002247 max=0.0010986 changed=0.8380
model.layers.1.self_attn.o_proj.weight                            rel=0.009584 rms=0.0002189 max=0.0010452 changed=0.8313
model.layers.13.self_attn.o_proj.weight                           rel=0.008842 rms=0.0002382 max=0.0012283 changed=0.8228
model.layers.8.self_attn.o_proj.weight                            rel=0.008573 rms=0.0002225 max=0.0010986 changed=0.8378

=== NORMALISATION PARAMETER DIFFERENCES ===
model.layers.0.input_layernorm.weight A[min,mean,max]= [6.375, 12.474902153015137, 61.75] B[min,mean,max]= [6.375, 12.474902153015137, 61.75] largest (index,A,B)= [(639, 9.5, 9.5), (638, 9.4375, 9.4375), (217, 15.5625, 15.5625), (216, 8.875, 8.875)]
model.layers.0.post_attention_layernorm.weight A[min,mean,max]= [-1.0, -0.36056631803512573, 51.75] B[min,mean,max]= [-1.0, -0.36056479811668396, 51.75] largest (index,A,B)= [(103, -0.099609375, -0.09912109375), (195, -0.1220703125, -0.12158203125), (639, 7.09375, 7.09375), (211, -0.890625, -0.890625)]
model.layers.0.post_feedforward_layernorm.weight A[min,mean,max]= [-0.98828125, 0.8047735095024109, 152.0] B[min,mean,max]= [-0.98828125, 0.8047771453857422, 152.0] largest (index,A,B)= [(244, -0.1298828125, -0.130859375), (304, -0.16015625, -0.1611328125), (435, 0.1806640625, 0.181640625), (462, 0.197265625, 0.1962890625)]
model.layers.0.pre_feedforward_layernorm.weight A[min,mean,max]= [-1.0, 4.830193996429443, 19.625] B[min,mean,max]= [-1.0, 4.830193996429443, 19.625] largest (index,A,B)= [(639, 0.76953125, 0.76953125), (638, 4.5, 4.5), (217, 5.78125, 5.78125), (216, 4.84375, 4.84375)]
model.layers.0.self_attn.k_norm.weight A[min,mean,max]= [-1.828125, 0.9714725017547607, 5.84375] B[min,mean,max]= [-1.828125, 0.9714727401733398, 5.84375] largest (index,A,B)= [(9, 0.0615234375, 0.062255859375), (147, 0.125, 0.12451171875), (200, -0.05908203125, -0.059326171875), (7, 0.0576171875, 0.057373046875)]
model.layers.0.self_attn.q_norm.weight A[min,mean,max]= [-3.78125, 0.9641852378845215, 3.5625] B[min,mean,max]= [-3.78125, 0.9641826152801514, 3.5625] largest (index,A,B)= [(56, -0.08251953125, -0.0830078125), (199, -0.0380859375, -0.038330078125), (104, 0.05859375, 0.058837890625), (46, 0.046630859375, 0.04638671875)]
model.layers.1.input_layernorm.weight A[min,mean,max]= [-0.171875, 24.65911293029785, 47.5] B[min,mean,max]= [-0.171875, 24.65911293029785, 47.5] largest (index,A,B)= [(639, 1.296875, 1.296875), (638, 20.75, 20.75), (217, 33.25, 33.25), (216, 22.5, 22.5)]
model.layers.1.post_attention_layernorm.weight A[min,mean,max]= [-1.0078125, 2.750606060028076, 103.0] B[min,mean,max]= [-1.0078125, 2.7506070137023926, 103.0] largest (index,A,B)= [(389, 0.1884765625, 0.189453125), (486, -0.04052734375, -0.041015625), (95, 0.1201171875, 0.11962890625), (557, -0.0703125, -0.07080078125)]
model.layers.1.post_feedforward_layernorm.weight A[min,mean,max]= [-1.0, 1.8786900043487549, 195.0] B[min,mean,max]= [-1.0, 1.8786838054656982, 195.0] largest (index,A,B)= [(505, 0.146484375, 0.1455078125), (363, 0.197265625, 0.1982421875), (308, -0.166015625, -0.1650390625), (553, 0.2177734375, 0.21875)]
model.layers.1.pre_feedforward_layernorm.weight A[min,mean,max]= [-1.0, 28.950634002685547, 55.75] B[min,mean,max]= [-1.0, 28.950634002685547, 55.75] largest (index,A,B)= [(639, 1.734375, 1.734375), (638, 31.0, 31.0), (217, 36.5, 36.5), (216, 31.125, 31.125)]
model.layers.1.self_attn.k_norm.weight A[min,mean,max]= [-1.0234375, 1.6646957397460938, 7.875] B[min,mean,max]= [-1.0234375, 1.6646957397460938, 7.875] largest (index,A,B)= [(255, 1.828125, 1.828125), (254, 2.203125, 2.203125), (93, 0.875, 0.875), (92, 3.1875, 3.1875)]
model.layers.1.self_attn.q_norm.weight A[min,mean,max]= [-1.1171875, 1.6944160461425781, 14.4375] B[min,mean,max]= [-1.1171875, 1.6944160461425781, 14.4375] largest (index,A,B)= [(255, 1.4765625, 1.4765625), (254, 1.3671875, 1.3671875), (93, 1.7890625, 1.7890625), (92, 0.78125, 0.78125)]
model.layers.10.input_layernorm.weight A[min,mean,max]= [-0.263671875, 16.899927139282227, 100.5] B[min,mean,max]= [-0.263671875, 16.899927139282227, 100.5] largest (index,A,B)= [(639, 15.0625, 15.0625), (638, 9.375, 9.375), (217, 9.3125, 9.3125), (216, 12.8125, 12.8125)]
model.layers.10.post_attention_layernorm.weight A[min,mean,max]= [-71.5, 42.12190246582031, 482.0] B[min,mean,max]= [-71.5, 42.12190246582031, 482.0] largest (index,A,B)= [(54, 0.08935546875, 0.08984375), (356, -0.0133056640625, -0.01312255859375), (195, -0.00848388671875, -0.0084228515625), (639, 16.75, 16.75)]
model.layers.10.post_feedforward_layernorm.weight A[min,mean,max]= [-1.2109375, 55.94256591796875, 592.0] B[min,mean,max]= [-1.2109375, 55.94256591796875, 592.0] largest (index,A,B)= [(639, 39.25, 39.25), (638, 72.5, 72.5), (217, 51.5, 51.5), (216, 53.0, 53.0)]
model.layers.10.pre_feedforward_layernorm.weight A[min,mean,max]= [-1.03125, 3.502417802810669, 38.25] B[min,mean,max]= [-1.03125, 3.5024161338806152, 38.25] largest (index,A,B)= [(45, 0.13671875, 0.1376953125), (270, -0.166015625, -0.1669921875), (338, -0.10595703125, -0.
[output cut at 15,000 characters]