Struct CudaModel 

pub struct CudaModel<'a> {
    pub dag: Dag<CompiledNode>,
    pub names: HashMap<usize, String>,
    /* private fields */
}

A compiled model: a DAG of CompiledNodes, not yet loaded into GPU memory. Call CudaModel::load to load it and get a runnable LoadedModel.

Fields

dag: Dag<CompiledNode>

The compiled node DAG.

names: HashMap<usize, String>

DAG node index → dotted name (e.g. "model.0.conv"), populated from the source Graph::names field during compilation.

Implementations

impl<'a> CudaModel<'a>

pub fn new(dag: Dag<CompiledNode>) -> Result<Self>

Wraps a compiled node DAG as a CudaModel, with no node names.

pub fn with_names( dag: Dag<CompiledNode>, names: HashMap<usize, String>, ) -> Result<Self>

Wraps a compiled node DAG as a CudaModel, with the given node-index → name mapping.

pub fn load( self, _device: &CudaDevice<'_>, batch_size: usize, ) -> Result<LoadedModel>

Load all compiled PTX kernels into GPU memory and pre-allocate zero-initialised parameter buffers, producing a LoadedModel ready for inference.

batch_size resolves dynamic (None) shape dimensions when computing parameter buffer sizes.

Trait Implementations

impl<'a> Model<'a> for CudaModel<'a>

type Input = TensorRef

This model’s input type.

type Output = TensorRef

This model’s output type.

fn forward(&self, _input: Self::Input) -> Result<Self::Output>

Runs inference, producing Output from input.