Gemma 3 random-model configuration decision
The source is gemma-3-4b-it-Q4_K_M.gguf. It identifies itself as
general.architecture = gemma3 and uses 34 layers, hidden width 2560, FFN width
10240, eight query heads, four KV heads, head dimension 256, sliding window
1024, context 131072, linear RoPE scaling factor 8, and tied embeddings.
The selected reduced geometry is six layers, hidden width 256, FFN width 1024, four query heads, two KV heads, head dimension 64, sliding window 64, context 128, and vocabulary 128. It preserves the FFN ratio of four, GQA ratio of two, one complete five-sliding/one-full attention period, linear RoPE scaling, RMS epsilon, four norm roles per layer, Q/K head norms, and tied embeddings.
The source EOS token ID is 106 and remains 106. The synthetic HF tokenizer and
converter SentencePiece tokenizer both cover IDs 0..127 and represent
PAD/BOS/UNK/EOS as 0/2/3/106. They are test tokenizers, not copies of the
source linguistic vocabulary.
All matrix quantization axes used by the reduced model are divisible by 256.
The retained GGUF uses llama.cpp Q4_K, an alias for the mixed Q4_K_M profile,
without --pure. Actual tensor types are recorded after generation.
The official Transformers 5.15.0 Gemma3ForCausalLM construction contains
5,938,176 trainable parameters. Norm parameters are initialized to zero because
Gemma3 RMSNorm applies 1 + weight; other one-dimensional parameters use their
model defaults and matrices use the deterministic tlfloat LCG64 equation.