OUTPUT of run 1 of at most 20:
```
=== 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
}
Configs identical: True

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

=== TENSOR SCHEMA ===
lm_head.weight count= 1 shapes/dtypes= [((262144, 640), 'BF16')]
model.embed_tokens.weight count= 1 shapes/dtypes= [((262144, 640), 'BF16')]
model.layers.<L>.input_layernorm.weight count= 18 shapes/dtypes= [((640,), 'BF16')]
model.layers.<L>.mlp.down_proj.weight count= 18 shapes/dtypes= [((640, 2048), 'BF16')]
model.layers.<L>.mlp.gate_proj.weight count= 18 shapes/dtypes= [((2048, 640), 'BF16')]
model.layers.<L>.mlp.up_proj.weight count= 18 shapes/dtypes= [((2048, 640), 'BF16')]
model.layers.<L>.post_attention_layernorm.weight count= 18 shapes/dtypes= [((640,), 'BF16')]
model.layers.<L>.post_feedforward_layernorm.weight count= 18 shapes/dtypes= [((640,), 'BF16')]
model.layers.<L>.pre_feedforward_layernorm.weight count= 18 shapes/dtypes= [((640,), 'BF16')]
model.layers.<L>.self_attn.k_norm.weight count= 18 shapes/dtypes= [((256,), 'BF16')]
model.layers.<L>.self_attn.k_proj.weight count= 18 shapes/dtypes= [((256, 640), 'BF16')]
model.layers.<L>.self_attn.o_proj.weight count= 18 shapes/dtypes= [((640, 1024), 'BF16')]
model.layers.<L>.self_attn.q_norm.weight count= 18 shapes/dtypes= [((256,), 'BF16')]
model.layers.<L>.self_attn.q_proj.weight count= 18 shapes/dtypes= [((1024, 640), 'BF16')]
model.layers.<L>.self_attn.v_proj.weight count= 18 shapes/dtypes= [((256, 640), 'BF16')]
model.norm.weight count= 1 shapes/dtypes= [((640,), 'BF16')]

=== TOKENIZER CONSISTENCY ===
tokenizer.json byte-identical: True
tokenizer_config.json byte-identical: True
bos_token '<bos>'
eos_token '<eos>'
pad_token '<pad>'

=== WEIGHT AUDIT ===

[stderr]
Traceback (most recent call last):
  File "<stdin>", line 171, in <module>
  File "<stdin>", line 154, in square_sum
  File "/env/lib/python3.12/site-packages/numpy/_core/einsumfunc.py", line 1423, in einsum
    return c_einsum(*operands, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ValueError: output has more dimensions than subscripts given in einstein sum, but no '...' ellipsis provided to broadcast the extra dimensions.

```