Module models 

Model-as-Code implementations.

Each module defines one model family as explicit Rust code: structs for weights + forward methods using the Backend trait and Linear trait directly. This replaces the earlier “generic ModelRunner + TransformerConfig” approach, which could not express MoE / MLA / multimodal / quantization cleanly.

Current coverage:

• llama_family — Llama / Llama-2 / Llama-3 / Qwen2 / Qwen2.5 / Qwen3 (standard GQA + SwiGLU + RoPE, optional QK-norm).

Planned (Phase D):

• mistral — sliding-window attention variant.
• deepseek_v3 — MLA compressed KV + MoE expert routing.
• qwen_vl — ViT backbone + LLM (multimodal).

Re-exports

pub use llama_family::LlamaFamilyConfig;

pub use llama_family::LlamaFamilyModel;

pub use llama_family_pipeline::LlamaFamilyPipelineModel;

pub use llama_family_pipeline::LlamaPipelineMode;

pub use llama_family_pipeline::LlamaPipelinePlacement;

pub use llama_family_pipeline::LlamaPipelineStageBridge;

pub use llama_family_pipeline::LlamaPipelineStagePlacement;

pub use llama_family_pipeline::LlamaPipelineTransport;

pub use qwen3_moe::Qwen3MoeModel;

Modules

llama_family

Llama-family decoder model as explicit code.

llama_family_forward_batched

Batched-decode forward methods for LlamaFamilyModel.

llama_family_pipeline

qwen3_moe

Qwen3MoeModel<B> — Qwen3-MoE family decoder (Qwen3-30B-A3B and friends).

qwen3_moe_forward_unified

Qwen3-MoE unified mixed-batch forward — vLLM-style.

qwen3_moe_forward_unified_layer

One Qwen3-MoE unified forward transformer layer.

qwen3_moe_profile

Profiling counters shared by the Qwen3-MoE forward paths.

qwen3_moe_runtime

Runtime environment snapshot for Qwen3-MoE model paths.