| # Gemma 4 tiny fixture configuration decision |
|
|
| ## Status |
|
|
| The text-only reduced configuration is generated under |
| `artifacts/gemma4-v0/gemma4-random-model` as a BF16 Hugging Face package and a |
| mixed K-quant GGUF requested through llama.cpp's `Q4_K` alias. |
|
|
| The current Hugging Face implementation is Transformers 5.15.0 |
| `Gemma4ForCausalLM` with `Gemma4TextConfig`. Its constructed trainable parameter |
| count is 6,036,608. |
|
|
| The local source file identifies itself as `general.architecture = gemma4`. |
| Repository and revision provenance are not inferable from the local directory and |
| remain intentionally unset. |
|
|
| ## Source observations |
|
|
| The 12B GGUF v3 header uses alignment 32 and contains 667 tensors. Its core |
| geometry is: |
|
|
| - 48 layers arranged as eight repetitions of five sliding-window layers followed |
| by one global layer. |
| - Hidden width 3840 and FFN width 15360. |
| - 16 query heads. |
| - Local layers: 8 KV heads and head width 256. |
| - Global layers: 1 KV head and head width 512. |
| - Global layers omit a separate `attn_v.weight`; this is part of the schema and |
| must not be filled in by a generic Llama tensor template. |
| - Sliding window 1024, context 262144. |
| - Global/local RoPE bases 1000000/10000. |
| - Global/local RoPE dimensions 512/256. |
| - Final logit softcap 30 and RMS epsilon 1e-6. |
| - Tied token/output embedding; the official tensor inventory has no independent |
| `output.weight`. |
|
|
| The separate projector file identifies itself as `general.architecture = clip` |
| and `general.type = mmproj`. It is not included in the first text-only fixture. |
|
|
| ## Tiny geometry |
|
|
| The selected geometry is defined in `configs/gemma4-random-q4k-v1.json`: |
|
|
| - 6 layers: one complete five-local/one-global schedule. |
| - Hidden width 256 and FFN width 1024. |
| - 4 query heads. |
| - Local layers: 2 KV heads, head width 64. |
| - Global layer: 1 KV head, head width 128, no independent V projection. |
| - Context 128 and sliding window 64. |
| - Vocabulary 128 with PAD/EOS/BOS/UNK/MASK IDs 0/1/2/3/4. |
|
|
| This is a constrained scale-down rather than independently selected fields. It |
| preserves the official layer schedule, the 2:1 global/local head-width ratio, the |
| 2:1 query/local-KV head ratio, the single global KV head, FFN ratio 4, separate |
| RoPE regimes, softcap, tied embeddings, and global shared-KV tensor inventory. |
|
|
| All matrix quantization axes are multiples of 256. This allows actual Q4_K/Q6_K |
| tensors rather than Q4_0 or K-quant fallback caused by undersized dimensions. |
| |
| ## Expected GGUF tensor dimensions |
| |
| GGUF dimensions are listed in reader order. |
| |
| | Role | Local layers 0-4 | Global layer 5 | |
| |---|---:|---:| |
| | `attn_q.weight` | `[256, 256]` | `[256, 512]` | |
| | `attn_k.weight` | `[256, 128]` | `[256, 128]` | |
| | `attn_v.weight` | `[256, 128]` | absent | |
| | `attn_output.weight` | `[256, 256]` | `[512, 256]` | |
| | `attn_q_norm.weight` | `[64]` | `[128]` | |
| | `attn_k_norm.weight` | `[64]` | `[128]` | |
|
|
| Every layer also has FFN gate/up `[256, 1024]`, FFN down `[1024, 256]`, |
| hidden-width norms, and one scalar layer-output scale. |
|
|
| ## Generation gates |
|
|
| Fixture generation is accepted only when: |
|
|
| - the direct writer emits `general.architecture = gemma4` and the exact tensor |
| role schedule above; |
| - a second generation is byte-identical; |
| - the F32 GGUF loads in the pinned llama.cpp revision; |
| - direct-token prefill and cached decode are finite and reproducible; |
| - the mixed Q4_K/Q6_K conversion retains the same geometry and loads successfully; |
| - F32 outputs are compared with an independently generated matched reference. |
|
|
