OUTPUT of run 1 of at most 20:
```
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
}
CONFIGS IDENTICAL: True
TENSOR KEYS IDENTICAL: True

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

EMBEDDING/OUTPUT CHECK: lm_head.weight exactly_identical= True norm_A= 509.9050843292101 norm_B= 509.9050843292101

EMBEDDING/OUTPUT CHECK: model.embed_tokens.weight exactly_identical= True norm_A= 509.9050843292101 norm_B= 509.9050843292101

LAYER DIFFERENCES (descriptive only, not importance)
layer  ||A-B||  relative_difference  fraction_changed
 0     0.55054    0.001312    0.7622
 1     0.55837    0.000526    0.7859
 2     0.56934    0.000526    0.7792
 3     0.55865    0.000388    0.7957
 4     0.57096    0.000372    0.8049
 5     0.57980    0.000461    0.7884
 6     0.58113    0.000468    0.7950
 7     0.56934    0.000474    0.7871
 8     0.57297    0.000414    0.7988
 9     0.57884    0.000396    0.7872
10     0.58862    0.000240    0.7956
11     0.58074    0.000097    0.7906
12     0.59754    0.000082    0.7955
13     0.59400    0.000086    0.8002
14     0.59747    0.000076    0.7990
15     0.60814    0.000063    0.7934
16     0.58132    0.000061    0.7922
17     0.55277    0.000054    0.7871

24 LARGEST RELATIVE TENSOR DIFFERENCES (not a causal ranking)
model.layers.4.mlp.down_proj.weight: delta=0.26156, relative=0.018788, changed=0.9183, norm_A=13.921, norm_B=13.922
model.layers.14.mlp.down_proj.weight: delta=0.28222, relative=0.018327, changed=0.9041, norm_A=15.399, norm_B=15.399
model.layers.8.mlp.down_proj.weight: delta=0.25419, relative=0.017131, changed=0.8990, norm_A=14.838, norm_B=14.838
model.layers.1.mlp.down_proj.weight: delta=0.26384, relative=0.016949, changed=0.8977, norm_A=15.567, norm_B=15.567
model.layers.10.mlp.down_proj.weight: delta=0.26087, relative=0.016746, changed=0.8996, norm_A=15.578, norm_B=15.578
model.layers.13.mlp.down_proj.weight: delta=0.27357, relative=0.016504, changed=0.8936, norm_A=16.576, norm_B=16.576
model.layers.3.mlp.down_proj.weight: delta=0.25818, relative=0.016254, changed=0.8985, norm_A=15.884, norm_B=15.884
model.layers.15.mlp.down_proj.weight: delta=0.28232, relative=0.015384, changed=0.8859, norm_A=18.351, norm_B=18.351
model.layers.6.mlp.down_proj.weight: delta=0.25525, relative=0.01492, changed=0.8853, norm_A=17.108, norm_B=17.108
model.layers.2.mlp.down_proj.weight: delta=0.26083, relative=0.014062, changed=0.8802, norm_A=18.548, norm_B=18.548
model.layers.16.mlp.down_proj.weight: delta=0.27017, relative=0.013684, changed=0.8719, norm_A=19.743, norm_B=19.744
model.layers.17.mlp.down_proj.weight: delta=0.2489, relative=0.013628, changed=0.8729, norm_A=18.264, norm_B=18.264
model.layers.7.mlp.down_proj.weight: delta=0.2565, relative=0.013404, changed=0.8714, norm_A=19.136, norm_B=19.136
model.layers.11.mlp.down_proj.weight: delta=0.25417, relative=0.012376, changed=0.8608, norm_A=20.538, norm_B=20.537
model.layers.9.mlp.down_proj.weight: delta=0.25952, relative=0.012282, changed=0.8602, norm_A=21.13, norm_B=21.129
model.layers.12.mlp.down_proj.weight: delta=0.26852, relative=0.011859, changed=0.8569, norm_A=22.644, norm_B=22.644
model.layers.3.self_attn.o_proj.weight: delta=0.17832, relative=0.011091, changed=0.8513, norm_A=16.078, norm_B=16.078
model.layers.4.self_attn.o_proj.weight: delta=0.1894, relative=0.011019, changed=0.8643, norm_A=17.187, norm_B=17.189
model.layers.0.mlp.down_proj.weight: delta=0.26313, relative=0.010687, changed=0.8427, norm_A=24.623, norm_B=24.623
model.layers.5.mlp.down_proj.weight: delta=0.26109, relative=0.010528, changed=0.8403, norm_A=24.799, norm_B=24.799
model.layers.16.self_attn.o_proj.weight: delta=0.17979, relative=0.0103, changed=0.8358, norm_A=17.455, norm_B=17.455
model.layers.1.self_attn.o_proj.weight: delta=0.18141, relative=0.0098111, changed=0.8358, norm_A=18.492, norm_B=18.489
model.layers.8.self_attn.o_proj.weight: delta=0.19099, relative=0.0090905, changed=0.8478, norm_A=21.01, norm_B=21.01
model.layers.13.self_attn.o_proj.weight: delta=0.19266, relative=0.0088338, changed=0.8224, norm_A=21.809, norm_B=21.81

PER-LAYER PROJECTION DIFFERENCES
entries are ||A-B|| / mean(||A||,||B||)
layer       q         k         v         o      gate        up      down
 0   0.005915  0.005655  0.005363  0.006141  0.006179  0.006074  0.010687
 1   0.005859  0.005537  0.006016  0.009811  0.006275  0.006242  0.016949
 2   0.006604  0.006293  0.006085  0.007668  0.006358  0.006270  0.014062
 3   0.006467  0.006326  0.005996  0.011091  0.006280  0.006075  0.016254
 4   0.006370  0.005960  0.006145  0.011019  0.006423  0.006343  0.018788
 5   0.006593  0.006402  0.006425  0.007864  0.006500  0.006353  0.010528
 6   0.007141  0.006933  0.006198  0.007729  0.006535  0.006297  0.014920
 7   0.006672  0.006438  0.006157  0.007361  0.006401  0.006306  0.013404
 8   0.006543  0.006344  0.006418  0.009091  0.006440  0.006370  0.017131
 9   0.006859  0.006759  0.006217  0.007121  0.006401  0.006358  0.012282
10   0.007106  0.007055  0.006152  0.006645  0.006431  0.006330  0.016746
11   0.006604  0.006644  0.006250  0.006588  0.006547  0.006371  0.012376
12   0.007359  0.007490  0.005640  0.007272  0.006594  0.006501  0.011859
13   0.006716  0.006525  0.006332  0.008834  0.006655  0.006614  0.016504
14   0.006652  0.006518  0.006363  0.006632  0.006763  0.006675  0.018327
15   0.007184  0.007332  0.006486  0.007355  0.006709  0.006708  0.015384
16   0.006704  0.006786  0.006379  0.010300  0.006507  0.006482  0.013684
17   0.006295  0.006198  0.005576  0.007182  0.006356  0.006235  0.013628

TOKENIZER STRUCTURE
vocabulary size: 262145
special tokens: {0: '<pad>', 1: '<eos>', 2: '<bos>', 3: '<unk>', 4: '<mask>', 105: '<start_of_turn>', 106: '<end_of_turn>', 255999: '<start_of_image>', 256000: '<end_of_image>', 262144: '<image_soft_token>'}
normalizer {"type": "Replace", "pattern": {"String": " "}, "content": "▁"}
pre_tokenizer {"type": "Split", "pattern": {"String": " "}, "behavior": "MergedWithPrevious", "invert": false}
post_processor {"type": "TemplateProcessing", "single": [{"SpecialToken": {"id": "<bos>", "type_id": 0}}, {"Sequence": {"id": "A", "type_id": 0}}], "pair": [{"SpecialToken": {"id": "<bos>", "type_id": 0}}, {"Sequence": {"id": "A", "type_id": 0}}, {"SpecialToken": {"id": "<bos>", "type_id": 1}}, {"Sequence": {"id": "B", "type_id": 1}}], "special_tokens": {"<bos>": {"id": "<bos>", "ids": [2], "tokens": ["<bos>"]}}}
decoder {"type": "Sequence", "decoders": [{"type": "Replace", "pattern": {"String": "▁"}, "content": " "}, {"type": "ByteFallback"}, {"type": "Fuse"}]}

CANDIDATE ANSWER/AGREEMENT TOKENIZATIONS (no model execution)
'A' [236776] ['A']
'B' [236799] ['B']
'C' [236780] ['C']
'D' [236796] ['D']
' A' [562] ['▁A']
' B' [603] ['▁B']
' C' [565] ['▁C']
' D' [622] ['▁D']
'yes' [4443] ['yes']
'no' [1904] ['no']
' yes' [11262] ['▁yes']
' no' [951] ['▁no']
'true' [3397] ['true']
'false' [4530] ['false']
'correct' [19448] ['correct']
'incorrect' [111863] ['incorrect']
'agree' [88047] ['agree']
'disagree' [2763, 88047] ['dis', 'agree']
'Paris' [50429] ['Paris']
'London' [27822] ['London']
'2' [236778] ['2']
'3' [236800] ['3']
'4' [236812] ['4']
'5' [236810] ['5']

CHAT TEMPLATE
None

```