File size: 3,486 Bytes
cf9609e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 | # Gemma 4 tiny fixture configuration decision
## Status
The text-only tiny configuration is fixed and generated under
`artifacts/gemma4-v0/gemma4-random-model` as a BF16 Hugging Face package and a
Q4_0 GGUF fixture.
The fixed Hugging Face implementation is Transformers 5.14.1
`Gemma4ForCausalLM` with `Gemma4TextConfig`. Its constructed trainable parameter
count is 1,519,168.
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-tiny-v0.json`:
- 6 layers: one complete five-local/one-global schedule.
- Hidden width 128 and FFN width 512.
- 4 query heads.
- Local layers: 2 KV heads, head width 32.
- Global layer: 1 KV head, head width 64, 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 dimensions are multiples of 32. This is required so that a direct F32
GGUF can subsequently be quantized through the pinned Q4_0 path without changing
model geometry merely to satisfy quantization blocks.
## Expected GGUF tensor dimensions
GGUF dimensions are listed in reader order.
| Role | Local layers 0-4 | Global layer 5 |
|---|---:|---:|
| `attn_q.weight` | `[128, 128]` | `[128, 256]` |
| `attn_k.weight` | `[128, 64]` | `[128, 64]` |
| `attn_v.weight` | `[128, 64]` | absent |
| `attn_output.weight` | `[128, 128]` | `[256, 128]` |
| `attn_q_norm.weight` | `[32]` | `[64]` |
| `attn_k_norm.weight` | `[32]` | `[64]` |
Every layer also has FFN gate/up `[128, 512]`, FFN down `[512, 128]`,
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 Q4_0 conversion retains the same geometry and loads successfully;
- F32 outputs are compared with an independently generated matched reference.
|