| # SmolLM3 random-model configuration decision |
|
|
| The source package at `/home/codex/llm_models/hfmodels/SmolLM3-3B` identifies |
| itself as `SmolLM3ForCausalLM` with 3,075,098,624 parameters. The generated |
| fixture is published as `smollm3-random-model`. |
|
|
| The source has hidden width 2048, FFN width 11008, 36 layers, 16 query heads, |
| four KV heads, context 65536, RoPE base 5000000, and tied embeddings. Its |
| attention schedule repeats three RoPE layers followed by one no-RoPE layer. |
|
|
| The reduced geometry uses hidden width 256, FFN width 1376, four layers, four |
| query heads, one KV head, head dimension 64, context 128, and vocabulary 128. |
| This preserves: |
|
|
| - the query/KV head ratio of four; |
| - the exact FFN ratio of 43/8; |
| - one complete three-RoPE/one-no-RoPE schedule; |
| - RoPE base, RMS epsilon, bias settings, and tied embeddings; |
| - matrix quantization axes divisible by 32 for Q4_0. |
| |
| The official Transformers 5.14.1 `SmolLM3ForCausalLM` construction contains |
| 4,917,504 trainable parameters. The retained Hugging Face package is BF16. The |
| GGUF path uses the pinned official llama.cpp SmolLM3 converter, a temporary F16 |
| GGUF, and Q4_0 quantization. |
|
|
| The Hugging Face package contains the deterministic 128-token `tokenizer.json`. |
| For official converter compatibility it also carries the existing 128-token |
| SentencePiece auxiliary artifact whose EOG set is `{1}`. Both are synthetic; |
| direct token IDs are the numerical reference interface. |
|
|
