Struct SafetensorsSource 

pub struct SafetensorsSource { /* private fields */ }

ModelSource over a HuggingFace-format directory.

Standard layout:

<dir>/config.json                   (required)
<dir>/generation_config.json        (optional)
<dir>/tokenizer.json                (required by the runtime)
<dir>/tokenizer_config.json         (optional; chat special tokens)
<dir>/model.safetensors             (single-file), or
<dir>/model.safetensors.index.json  (sharded)

For diffusion sub-dirs (unet/, vae/, text_encoder/) use SafetensorsSource::load_weights_only, which tolerates diffusion_pytorch_model.safetensors (or model.safetensors) and no tokenizer/config.

Files are memory-mapped; ModelSource::open_tensor returns zero-copy views into the mapping.

Implementations

impl SafetensorsSource

pub fn load(dir: impl AsRef<Path>) -> Result<Self>

Opens a model directory (see type docs for the expected layout).

pub fn load_weights_only( dir: impl AsRef<Path>, architecture: &str, ) -> Result<Self>

Opens a directory purely for weight loading. No config.json or tokenizer.json is required; metadata() returns a placeholder and tokenizer() errors. Accepts diffusion-style diffusion_pytorch_model.safetensors as well as model.safetensors.

Trait Implementations

impl ModelSource for SafetensorsSource

fn metadata(&self) -> &ModelMetadata

Architecture + hyperparameter metadata.

fn tensor_names(&self) -> Vec<String>

Names of all tensors available in this source.

fn open_tensor(&self, name: &str) -> Result<TensorReader<'_>>

Opens a tensor by name, returning a lazy reader over the raw bytes.

fn tokenizer(&self) -> Result<TokenizerSpec>

Tokenizer specification (path to tokenizer.json + added tokens).

fn sampler_defaults(&self) -> Option<SamplerConfig>

Sampler defaults from generation_config.json, if present.

fn open_tensor_quant(&self, _name: &str) -> Result<Option<QuantTensor<'_>>>

Raw packed quantized bytes for a tensor, when this source stores it in a quant format that has a device kernel (GGUF Q4_0/Q4_K/Q6_K). None means “no packed representation” — the caller falls back to ModelSource::open_tensor, which dequantizes to float. Sources without packed formats keep this default.