Crate sapient_models 

sapient-models — pre-built LLM architecture graph builders.

Each architecture module builds a SAPIENT Graph from a ModelInfo, matching the exact HuggingFace architecture for that model family.

Supported architectures

HuggingFace class	Module	Models
LlamaForCausalLM	llama	Llama 1/2/3, Mistral, CodeLlama, Vicuna, WizardLM
PhiForCausalLM	phi	Phi-1/2/3/3.5
GemmaForCausalLM	gemma	Gemma, Gemma 2
GPT2LMHeadModel	gpt2	GPT-2, CodeGen, GPT-J
BertForMaskedLM	bert	BERT, RoBERTa, DistilBERT
Qwen2ForCausalLM	qwen	Qwen, Qwen2, Qwen2.5
MixtralForCausalLM	mixtral	Mixtral-8x7B, Mixtral-8x22B

Re-exports

pub use forward::mac_gpu_support;

pub use forward::total_system_ram_bytes;

pub use forward::ForwardEngine;

pub use forward::LlmBackendKind;

pub use forward::MacGpuSupport;

pub use registry::build_graph;

pub use registry::ModelGraph;

Modules

architectures

forward

Real transformer forward passes for text generation.

gguf_weights

Map llama.cpp GGUF tensor names to HuggingFace layout for native forward passes.

registry

Architecture registry — dispatch from ArchType → graph builder.

weights

HuggingFace safetensors weight loading and key resolution.