Crate ferrotensor_loader 

Model loading for Ferrotensor: Hugging Face config.json parsing (Lfm2Config), memory-mapped safetensors weights (single file or sharded, F32/F16/BF16 read as F32, int-quantized weights dequantized on the GPU) through WeightsContext, and a Tokenizer / Sampler pair for text in and out.

Re-exports

pub use tokenizer::argmax;

pub use tokenizer::Sampler;

pub use tokenizer::Tokenizer;

Modules

tokenizer

Tokenizer for LFM2.5 models

Structs

Lfm2Config

Model hyperparameters read from config.json (Llama and LFM2 layouts).

ModelLoader

Reads model config files and raw safetensors files.

QuantizationConfig

Quantization settings from the model config.

TensorMetadata

Shape and dtype of a stored tensor.

WeightsContext

Memory-mapped safetensors files and an index from tensor name to file.

Enums

LoaderError

Errors returned while loading model configs, weights and tensors.

Type Aliases

LlamaConfig

Alias kept for older code that used the Llama-specific config name.