Crate ferrum_models 

Ferrum 模型层

该 crate 负责围绕 ferrum-interfaces/ferrum-types 定义的核心抽象 提供模型定义解析、构建器与权重加载占位实现，确保上层可以在 重构阶段编译。

Re-exports

pub use common::DecoderOnlyLLM;

pub use common::LlmRuntimeConfig;

pub use definition::ConfigManager;

pub use definition::ModelDefinition;

pub use executor::BertModelExecutor;

pub use executor::ClipModelExecutor;

pub use executor::LlmExecutor;

pub use executor::StubModelExecutor;

pub use executor::TtsModelExecutor;

pub use executor::WhisperModelExecutor;

pub use hf_download::HfDownloader;

pub use image_processor::ClipImageProcessor;

pub use loader::SafeTensorsLoader;

pub use lora::default_lora_model_id;

pub use lora::load_runtime_lora_adapter;

pub use lora::load_startup_lora_adapter;

pub use lora::load_startup_lora_adapters;

pub use lora::render_lora_model_id;

pub use lora::ActiveLoraAdapter;

pub use lora::LoraAdapterConfig;

pub use lora::RuntimeLoraAdapter;

pub use lora::StartupLoraAdapter;

pub use lora::StartupLoraSpec;

pub use multimodal::BertModelWrapper;

pub use multimodal::ClipModelWrapper;

pub use multimodal::WhisperModelWrapper;

pub use registry::Architecture;

pub use registry::DefaultModelRegistry;

pub use registry::ModelAlias;

pub use registry::ModelDiscoveryEntry;

pub use registry::ModelFormatType;

pub use source::DefaultModelSourceResolver;

pub use source::ModelFormat;

pub use source::ModelSourceConfig;

pub use source::ModelSourceResolver;

pub use source::ResolvedModelSource;

pub use tensor_wrapper::CandleTensorWrapper;

pub use tokenizer::TokenizerFactory;

pub use tokenizer::TokenizerHandle;

Modules

audio_processor

Audio preprocessing for Whisper ASR.

common

Cross-model traits and helpers (Model-as-Code shared infrastructure).

definition

Model definition and configuration parsing

executor

Model executor implementations.

gguf_config

Build a LlamaFamilyConfig from a GGUF file’s metadata.

gguf_engine_loader

GGUF → engine glue. Lets ferrum serve / ferrum bench accept a .gguf path (or an alias resolving to one) and produce a Box<dyn DecoderOnlyLLM> that the existing LlmExecutor + ContinuousBatchEngine can drive.

gguf_runtime

Thin dispatch wrapper around candle-transformers’ quantized GGUF model loaders. Exists so the rest of ferrum can hand a .gguf path to something that produces a Tensor from token ids without caring whether the underlying arch is Qwen3 / Qwen3-MoE / Llama-3.x / Mistral.

hf_download

HuggingFace model downloader with proxy and resume support

image_processor

Image preprocessing for CLIP models.

loader

Weight loading from SafeTensors files

lora

Startup LoRA adapter loader and validator.

mel

Mel spectrogram computation matching Python whisper exactly.

models

Model-as-Code implementations.

moe

Mixture-of-Experts runtime primitives.

moe_config

Mixture-of-Experts (MoE) configuration types.

multimodal

Legacy architecture wrappers (candle-based, non-decoder model types).

registry

Model registry and alias management

source

Model source resolution and downloading with progress tracking

tensor_wrapper

Candle Tensor wrapper implementing TensorLike

tokenizer

Tokenizer 占位实现

weight_format

Dim 3 polymorphism point — weight-format detection for the executor factory.

Structs

AttentionConfig

Attention configuration for model architecture

ModelConfig

Model configuration for runtime

ModelInfo

Model information and metadata

RopeScaling

RoPE (Rotary Position Embedding) scaling configuration

Enums

Activation

Activation function type

ModelType

Model type enumeration

NormType

Normalization type used in the model

Traits

ModelExecutor

Core model executor trait focusing on tensor operations

Type Aliases

Result

Result type used throughout Ferrum