Model Families

Model Families and Synapse Sources

src/types.rs + src/moe/adapter.rs.

ModelFamily

Variant	Arch slug / GGUF arch match
Olmoe	olmoe
Qwen3Moe	qwen3moe
Gemma4	gemma4
DeepSeek2	deepseek2
LlamaMoe	llama

slug() returns snake-case labels (qwen3_moe, llama_moe, …).

Family override must match inferred architecture or load fails with InvalidConfig.

Adapter resolution (resolve_adapter)

Reads GGUF metadata keys:

• {arch}.embedding_length → hidden_size
• {arch}.block_count → num_layers
• {arch}.expert_count → num_experts
• {arch}.expert_used_count → optional (default 1)

Token embedding tensor: token_embd.weight or tok_embeddings.weight.

Routing tensor: first *ffn_gate_inp.weight or *ffn_gate.weight — must be rank-2 F32.

Synapse source labels

Preferred tensor: blk.0.attn_q.weight when present.

Condition	synapse_source	real_gpu_synapse_tensor
attn_q is F16 rank-2 containing hidden_size (GQA-relaxed)	real	attn_q name
attn_q present but not real F16	routing-f32	routing tensor name
no attn_q	synthetic-fallback	none

Public metadata for SAAQ / Surrogate_Viz consumers:

• preferred_gpu_synapse_tensor_name
• real_gpu_synapse_tensor_name
• synapse_source

---

Last updated: July 29, 2026 Updated by: Grok Build: Grok 4.5 Package tip reference: 37c5a21 (main)