Struct QuantPreview 

pub struct QuantPreview<'model> { /* private fields */ }

Ask llama.cpp what it would do, without writing a file.

crate::model_quantize is all-or-nothing: it reads a model, quantizes every tensor, and writes the result. This exposes the decision layer underneath — which tensors get quantized at all, and to which ggml type — so a tool can show the plan, or estimate output size, before committing to a run that may take minutes and tens of gigabytes.

The k-quant mixes are why this is not simply LlamaFtype::default_ggml_type: they deliberately keep attention and output tensors at higher precision, so the per-tensor answer differs from the ftype’s nominal type.

Wraps llama_quant_init / llama_quant_free.

Implementations

impl<'model> QuantPreview<'model>

pub fn new( model: &'model LlamaModel, params: &QuantizeParams, ) -> Result<Self, QuantPreviewError>

Build a preview for model under params.

Errors

Returns QuantPreviewError::Init if llama.cpp could not build the quantization state, which happens for a model it cannot quantize.

pub fn allows_quantization(&self, tensor: &GgmlTensor) -> bool

Whether this tensor would be quantized at all.

llama.cpp skips 1-D tensors, tensors below a size threshold, and ones whose name marks them as needing full precision — so a false here is the usual reason a tensor keeps its original type.

Requires the ggml feature, which is what exposes crate::ggml::GgmlTensor.

pub fn compute_types( &self, tensors: &[&GgmlTensor], ftype: LlamaFtype, ) -> Result<Vec<Option<GgmlType>>, QuantPreviewError>

Compute the storage type each tensor would be assigned under ftype.

Every tensor passed must already satisfy Self::allows_quantization — upstream states the caller filters first, and does not re-check.

An entry is None when llama.cpp picks a ggml type this crate’s GgmlType does not know.

Errors

Returns QuantPreviewError::NotQuantizable naming the first tensor that fails the filter, rather than letting llama.cpp decide what to do with it.

Requires the ggml feature.

Trait Implementations

impl<'model> Debug for QuantPreview<'model>

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Drop for QuantPreview<'_>

fn drop(&mut self)

Executes the destructor for this type.

fn pin_drop(self: Pin<&mut Self>)

🔬This is a nightly-only experimental API. (pin_ergonomics)

Execute the destructor for this type, but different to Drop::drop, it requires self to be pinned.