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 B DIFFERENCES

TOKENIZER FILE COMPARISON
tokenizer.json identical: True
tokenizer_config.json identical: True

TENSOR KEY DIFFERENCES [] []
Number of tensors: 237 237

EMBEDDING/OUTPUT lm_head.weight (262144, 640) bit-identical: True

EMBEDDING/OUTPUT model.embed_tokens.weight (262144, 640) bit-identical: True

TENSOR FAMILIES
lm_head.weight [(262144, 640)]
model.embed_tokens.weight [(262144, 640)]
model.layers.*.input_layernorm.weight [(640,)]
model.layers.*.mlp.down_proj.weight [(640, 2048)]
model.layers.*.mlp.gate_proj.weight [(2048, 640)]
model.layers.*.mlp.up_proj.weight [(2048, 640)]
model.layers.*.post_attention_layernorm.weight [(640,)]
model.layers.*.post_feedforward_layernorm.weight [(640,)]
model.layers.*.pre_feedforward_layernorm.weight [(640,)]
model.layers.*.self_attn.k_norm.weight [(256,)]
model.layers.*.self_attn.k_proj.weight [(256, 640)]
model.layers.*.self_attn.o_proj.weight [(640, 1024)]
model.layers.*.self_attn.q_norm.weight [(256,)]
model.layers.*.self_attn.q_proj.weight [(1024, 640)]
model.layers.*.self_attn.v_proj.weight [(256, 640)]
model.norm.weight [(640,)]

LAYER DIFFERENCES: ||B-A||, relative norm, changed fraction
 0     0.551539    0.001315  0.76205
 1     0.553410    0.000521  0.78419
 2     0.549186    0.000507  0.77156
 3     0.548911    0.000381  0.79276
 4     0.555902    0.000363  0.79999
 5     0.569138    0.000453  0.78514
 6     0.570107    0.000459  0.79208
 7     0.557993    0.000465  0.78290
 8     0.552039    0.000399  0.79206
 9     0.571160    0.000391  0.78384
10     0.578741    0.000236  0.79197
11     0.566308    0.000095  0.78514
12     0.585335    0.000080  0.79162
13     0.588595    0.000086  0.79825
14     0.594680    0.000076  0.79819
15     0.604158    0.000062  0.79195
16     0.590751    0.000062  0.79505
17     0.575472    0.000056  0.79402

TOP 45 TENSORS BY DIFFERENCE NORM (NOT BEHAVIORAL IMPORTANCE)
model.layers.14.mlp.gate_proj.weight                           (2048, 640)      delta=0.298669 rel=0.006685 changed=0.7647 max=0.001465
model.layers.15.mlp.up_proj.weight                             (2048, 640)      delta=0.298319 rel=0.006677 changed=0.7460 max=0.001465
model.layers.15.mlp.gate_proj.weight                           (2048, 640)      delta=0.298282 rel=0.006674 changed=0.7641 max=0.001953
model.layers.14.mlp.up_proj.weight                             (2048, 640)      delta=0.296389 rel=0.006643 changed=0.7445 max=0.001465
model.layers.16.mlp.up_proj.weight                             (2048, 640)      delta=0.295437 rel=0.006600 changed=0.7442 max=0.001465
model.layers.16.mlp.gate_proj.weight                           (2048, 640)      delta=0.294908 rel=0.006614 changed=0.7606 max=0.001228
model.layers.13.mlp.gate_proj.weight                           (2048, 640)      delta=0.292737 rel=0.006554 changed=0.7593 max=0.001221
model.layers.13.mlp.up_proj.weight                             (2048, 640)      delta=0.292338 rel=0.006557 changed=0.7427 max=0.001465
model.layers.12.mlp.gate_proj.weight                           (2048, 640)      delta=0.291493 rel=0.006510 changed=0.7676 max=0.001465
model.layers.17.mlp.up_proj.weight                             (2048, 640)      delta=0.290303 rel=0.006465 changed=0.7434 max=0.001221
model.layers.17.mlp.gate_proj.weight                           (2048, 640)      delta=0.290025 rel=0.006571 changed=0.7654 max=0.001465
model.layers.6.mlp.gate_proj.weight                            (2048, 640)      delta=0.286932 rel=0.006417 changed=0.7615 max=0.001221
model.layers.12.mlp.up_proj.weight                             (2048, 640)      delta=0.286725 rel=0.006414 changed=0.7468 max=0.001465
model.layers.15.mlp.down_proj.weight                           (640, 2048)      delta=0.284161 rel=0.015485 changed=0.8867 max=0.001282
model.layers.11.mlp.gate_proj.weight                           (2048, 640)      delta=0.283441 rel=0.006330 changed=0.7608 max=0.001465
model.layers.10.mlp.gate_proj.weight                           (2048, 640)      delta=0.283137 rel=0.006314 changed=0.7636 max=0.001465
model.layers.5.mlp.gate_proj.weight                            (2048, 640)      delta=0.283041 rel=0.006315 changed=0.7590 max=0.001465
model.layers.9.mlp.gate_proj.weight                            (2048, 640)      delta=0.281397 rel=0.006269 changed=0.7623 max=0.001343
model.layers.7.mlp.gate_proj.weight                            (2048, 640)      delta=0.281146 rel=0.006277 changed=0.7582 max=0.001465
model.layers.9.mlp.up_proj.weight                              (2048, 640)      delta=0.280847 rel=0.006290 changed=0.7350 max=0.001465
model.layers.4.mlp.gate_proj.weight                            (2048, 640)      delta=0.280535 rel=0.006262 changed=0.7538 max=0.001251
model.layers.3.mlp.gate_proj.weight                            (2048, 640)      delta=0.279027 rel=0.006235 changed=0.7472 max=0.001221
model.layers.8.mlp.gate_proj.weight                            (2048, 640)      delta=0.278643 rel=0.006203 changed=0.7623 max=0.001221
model.layers.14.mlp.down_proj.weight                           (640, 2048)      delta=0.278398 rel=0.018078 changed=0.9032 max=0.001343
model.layers.16.mlp.down_proj.weight                           (640, 2048)      delta=0.278261 rel=0.014094 changed=0.8763 max=0.001251
model.layers.6.mlp.up_proj.weight                              (2048, 640)      delta=0.278210 rel=0.006233 changed=0.7361 max=0.001465
model.layers.5.mlp.up_proj.weight                              (2048, 640)      delta=0.277837 rel=0.006216 changed=0.7381 max=0.001953
model.layers.1.mlp.gate_proj.weight                            (2048, 640)      delta=0.277353 rel=0.006171 changed=0.7289 max=0.001465
model.layers.11.mlp.up_proj.weight                             (2048, 640)      delta=0.276536 rel=0.006190 changed=0.7360 max=0.001465
model.layers.4.mlp.up_proj.weight                              (2048, 640)      delta=0.276299 rel=0.006184 changed=0.7363 max=0.001465
model.layers.10.mlp.up_proj.weight                             (2048, 640)      delta=0.276123 rel=0.006174 changed=0.7345 max=0.001953
model.layers.1.mlp.up_proj.weight                              (2048, 640)      delta=0.275728 rel=0.006145 changed=0.7285 max=0.001221
model.layers.2.mlp.gate_proj.weight                            (2048, 640)      delta=0.275602 rel=0.006151 changed=0.7288 max=0.001221
model.layers.7.mlp.up_proj.weight                              (2048, 640)      delta=0.274724 rel=0.006158 changed=0.7309 max=0.001465
model.layers.8.mlp.up_proj.weight                              (2048, 640)      delta=0.273235 rel=0.006127 changed=0.7301 max=0.001465
model.layers.0.mlp.gate_proj.weight                            (2048, 640)      delta=0.273165 rel=0.006113 changed=0.7296 max=0.001221
model.layers.13.mlp.down_proj.weight                           (640, 2048)      delta=0.272982 rel=0.016468 changed=0.8942 max=0.001190
model.layers.0.mlp.up_proj.weight                              (2048, 640)      delta=0.272349 rel=0.006085 changed=0.7275 max=0.001099
model.layers.2.mlp.up_proj.weight                              (2048, 640)      delta=0.271393 rel=0.006066 changed=0.7248 max=0.001221
model.layers.3.mlp.up_proj.weight                              (2048, 640)      delta=0.268005 rel=0.005989 changed=0.7292 max=0.001221
model.layers.12.mlp.down_proj.weight                           (640, 2048)      delta=0.265868 rel=0.011741 changed=0.8554 max=0.001160
model.layers.17.mlp.down_proj.weight                           (640, 2048)      delta=0.263536 rel=0.014429 changed=0.8804 max=0.001282
model.layers.0.mlp.down_proj.weight                            (640, 2048)      delta=0.260645 rel=0.010586 changed=0.8407 max=0.001099
model.layers.1.mlp.down_proj.weight                            (640, 2048)      delta=0.258052 rel=0.016577 changed=0.8955 max=0.001221
model.layers.5.mlp.down_proj.weight                            (640, 2048)      delta=0.257267 rel=0.010374 changed=0.8380 max=0.001099

ALL ONE-DIMENSIONAL TENSOR DIFFERENCES
model.layers.0.input_layernorm.weight                          delta=0.000000 mean(B-A)=+0.0000000 max=0.000000
model.layers.0.post_attention_layernorm.weight                 delta=0.000691 mean(B-A)=+0.0000015 max=0.000488
model.layers.0.post_feedforward_layernorm.weight               delta=0.004725 mean(B-A)=+0.0000036 max=0.000977
model.layers.0.pre_feedforward_layernorm.weight                delta=0.000000 mean(B-A)=+0.0000000 max=0.000000
model.layers.0.self_attn.k_norm.weight                         delta=0.000971 mean(B-A)=+0.0000002 max=0.000732
model.layers.0.self_attn.q_norm.weight                         delta=0.000649 mean(B-A)=-0.0000026 max=0.000488
model.layers.1.input_layernorm.weight                          delta=0.000000 mean(B-A)=+0.0000000 max=0.000000
model.layers.1.post_attention_layernorm.weight                 delta=0.002124 mean(B-A)=+0.0000007 max=0.000977
model.layers.1.post_feedforward_layernorm.weight               delta=0.003702 mean(B-A)=-0.0000059 max=0.000977
model.layers.1.pre_feedforward_layernorm.weight                delta=0.000000 mean(B-A)=+0.0000000 max=0.000000
model.layers.1.self_attn.k_norm.weight                         delta=0.000000 mean(B-A)=+0.0000000 max=0.000000
model.layers.1.self_attn.q_norm.weight                         delta=0.000000 mean(B-A)=+0.0000000 max=0.000000
model.layers.10.input_layernorm.weight                         delta=0.000000 mean(B-A)=+0.0000000 max=0.000000
model.layers.10.post_attention_layernorm.weight                delta=0.000525 mean(B-A)=+0.0000011 max=0.000488
model.layers.10.post_feedforward_layernorm.weight              delta=0.000000 mean(B-A)=+0.0000000 max=0.000000
model.layers.10.pre_feedforward_layernorm.weight               delta=0.002068 mean(B-A)=-0.0000018 max=0.000977
model.layers.10.self_attn.k_norm.weight                        delta=0.001155 mean(B-A)=-0.0000005 max=0.000488
model.layers.10.self_attn.q_norm.weight                        delta=0.001676 mean(B-A)=+0.0000072 max=0.000977
model.layers.11.input_layernorm.weight                         delta=0.001663 mean(B-A)=+0.0000026 max=0.000977
model.layers.11.post_attention_layernorm.weight                delta=0.000000 mean(B-A)=+0.0000000 max=0.000000
model.layers.11.post_feedforward_layernorm.weight              delta=0.000000 mean(B-A)=+0.0000000 max=0.000000
model.layers.11.pre_feedforward_layernorm.weight               delta=0.001293 mean(B-A)=-0.0000022 max=0.000977
model.layers.11.self_attn.k_norm.weight                        delta=0.000793 mean(B-A)=-0.0000026 max=0.000488
model.layers.11.self_attn.q_norm.weight                        delta=0.000948 mean(B-A)=+0.0000026 max=0.000488
model.layers.12.input_layernorm.weight                         delta=0.000901 mean(B-A)=+0.0000011 max=0.000488
model.layers.12.post_attention_layernorm.weight                delta=0.000000 mean(B-A)=+0.0000000 max=0.000000
model.layers.12.post_feedforward_layernorm.weight              delta=0.000000 mean(B-A)=+0.0000000 max=0.000000
model.layers.12.pre_feedforward_layernorm.weight               delta=0.001550 mean(B-A)=+0.0000048 max=0.000977
model.layers.12.self_attn.k_norm.weight                        delta=0.000986 mean(B-A)=-0.0000044 max=0.000977
model.layers.12.self_attn.q_norm.weight                        delta=0.001029 mean(B-A)=+0.0000033 max=0.000732
model.layers.13.input_layernorm.weight                         delta=0.002412 mean(B-A)=-0.0000016 max=0.000732
model.layers.13.post_attention_layernorm.weight                delta=0.000000 mean(B-A)=+0.0000000 max=0.000000
model.layers.13.post_feedforward_layernorm.weight              delta=0.000000 mean(B-A)=+0.0000000 max=0.000000
model.layers.13.pre_feedforward_layernorm.weight               delta=0.002547 mean(B-A)=-0.0000044 max=0.000977
model.layers.13.self_attn.k_norm.weight                        delta=0.001134 mean(B-A)=-0.0000002 max=0.000977
model.layers.13.self_attn.q_norm.weight                        delta=0.001942 mean(B-A)=-0.0000043 max=0.000977
model.layers.14.input_layernorm.weight                         delta=0.002092 mean(B-A)=+0.0000009 max=0.000977
model.layers.14.post_attention_layernorm.weight                delta=0.000000 mean(B-A)=+0.0000000 max=0.000000
model.layers.14.post_feedforward_layernorm.weight              delta=0.000000 mean(B-A)=+0.0000000 max=0.000000
model.layers.14.pre_feedforward_layernorm.weight               delta=0.003197 mean(B-A)=-0.0000066 max=0.000977
model.layers.14.self_attn.k_norm.weight                        delta=0.001896 mean(B-A)=+0.0000031 max=0.000977
model.layers.14.self_attn.q_norm.weight                        delta=0.001234 mean(B-A)=+0.0000041 max=0.000977
model.layers.15.input_layernorm.weight                         delta=0.000977 mean(B-A)=+0.0000015 max=0.000977
model.layers.15.post_attention_layernorm.weight                delta=0.000000 mean(B-A)=+0.0000000 max=0.000000
model.layers.15.post_feedforward_layernorm.weight              delta=0.000000 mean(B-A)=+0.0000000 max=0.000000
model.layers.15.pre_feedforward_layernorm.weight               delta=0.002331 mean(B-A)=+0.0000032 max=0.000977
model.layers.15.self_attn.k_norm.weight                        delta=0.001865 mean(B-A)=-0.0000014 max=0.000977
model.layers.15.self_attn.q_norm.weight                        delta=0.001384 mean(B-A)=-0.0000007 max=0.000977
model.layers.16.input_layernorm.weight                        
[output cut at 15,000 characters]