oxfuse_runtime/
backend.rs

1//! The portable backend interface. Kept minimal on purpose: it assumes host
2//! tensors, so a device-memory backend would need to revise it.
3
4use crate::Tensor;
5use oxfuse_core::{Error, Graph};
6
7pub trait Backend {
8    fn compile(&self, graph: &Graph) -> Result<Box<dyn Executable>, Error>;
9}
10
11pub trait Executable {
12    /// `inputs` follow the graph's declared input order; the result follows
13    /// the graph's output order.
14    fn run(&self, inputs: &[Tensor]) -> Result<Vec<Tensor>, Error>;
15
16    /// Runs the graph once per element of `batches` and returns the results in
17    /// the same order; every result equals what `run` returns for that element.
18    /// The default runs them one after another. The GPU backend overrides it to
19    /// submit the whole batch at once and wait once, which hides the per-run
20    /// wait; its memory use grows with the batch.
21    fn run_many(&self, batches: &[Vec<Tensor>]) -> Result<Vec<Vec<Tensor>>, Error> {
22        batches.iter().map(|inputs| self.run(inputs)).collect()
23    }
24}
oxfuse_runtime/backend.rs

oxfuse_runtime/
backend.rs
