Shared "Llama-shaped" architecture config for the M4 family group.
All the families listed in PLAN.md M4 share the same skeleton —
GQA + RoPE + RMSNorm + SwiGLU FFN, differing mainly in RoPE base/
scaling, sliding-window setting, and exact tensor-name conventions:
| family | llama.cpp arch tag | RoPE scaling | notes |
|---|---|---|---|
| Mistral 3+ | mistral3 / mistral4 |
Default | Sliding-window (5K for 3.5) |
| Phi 3 / 4 | phi3 |
Default + YaRN | (phi4 reuses phi3 upstream) |
| Bonsai | (llama) |
Default | Ships as llama-tagged GGUF |
| OmniCoder | (qwen3) |
Default | Ships as qwen3-tagged GGUF |
| Granite | granite |
Default | IBM Llama-shaped |
| Command-R | command-r / cohere2 |
Default | Cohere Llama-shaped |
[LlamaBaseConfig] captures the union of fields the rlx-models M4
stub crates need; [from_gguf_path] reads them straight from a GGUF
header (no full weight load).
Existing rlx-llama32 / rlx-qwen3 / rlx-gemma keep their own
per-family configs for now — this crate is the new shared base
the M4 stubs grow against. A future cleanup may migrate the existing
crates to use it, but that's an invasive refactor explicitly out of
scope for the M4 stub work.