Struct FloatCastAdapter 

pub struct FloatCastAdapter { /* private fields */ }

Adapter that casts every float tensor to one target dtype.

Unlike HalfPrecisionAdapter, which infers the direction from the source dtype (F32 <-> F16 only) and converts a fixed list of module types, this adapter is target-driven: every float tensor (F64, F32, Flex32, F16, BF16) is converted to the given dtype regardless of which module it belongs to. Non-float tensors (ints, bools, quantized) and tensors already in the target dtype pass through unchanged.

This is the right tool for loading a stock checkpoint into a module whose backend runs at a different float precision, e.g. BF16 safetensors into an F16 model. HalfPrecisionAdapter passes BF16 tensors through unconverted and skips custom module types entirely, so the loaded params silently keep the source dtype instead of the module’s.

Examples

Cast everything to the backend’s float element on load:

let adapter = FloatCastAdapter::to(<B::FloatElem as Element>::dtype());
// store.with_from_adapter(adapter);

Chain after a framework adapter:

let adapter = PyTorchToBurnAdapter.chain(FloatCastAdapter::to(DType::F16));

Implementations

impl FloatCastAdapter

pub fn to(target: DType) -> Self

Create an adapter that casts every float tensor to target.

Panics

Panics if target is not a float dtype.

Trait Implementations

impl Clone for FloatCastAdapter

fn clone(&self) -> Self

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for FloatCastAdapter

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

Formats the value using the given formatter.

impl ModuleAdapter for FloatCastAdapter

fn adapt(&self, tensor: PackTensor, _ctx: ModuleContext<'_>) -> PackTensor

Adapt a tensor given where in the module hierarchy it was found.

fn clone_box(&self) -> Box<dyn ModuleAdapter>

Clone the adapter into a boxed trait object

fn get_alternative_param_name( &self, _param_name: &str, _module_type: &str, ) -> Option<String>

Get alternative parameter name to try during matching

fn chain<A>(self, next: A) -> ChainAdapter

where Self: Sized + 'static, A: ModuleAdapter + 'static,

Chain adapters together, applying self first and then next.