Struct CutlassInner 

pub struct CutlassInner {
    pub plan_cache: Arc<PlanCache>,
    pub compile_sink: Option<CompileSink>,
    pub dispatched: Mutex<u64>,
}

Inner state of a CutlassActor.

Fields

plan_cache: Arc<PlanCache>

compile_sink: Option<CompileSink>

Optional NVRTC-shaped sink: when present, the actor forwards rendered .cu source to it for compilation. Left as a generic Box<dyn Fn ...> so the cutlass crate doesn’t pull atomr-accel-cuda::nvrtc into its compile graph when the feature is off.

dispatched: Mutex<u64>

Counter of dispatched messages — exposed for the actor::tests::cutlass_msg_constructs test and for telemetry.

Implementations

impl CutlassInner

pub fn new(plan_cache_capacity: usize) -> Self

pub fn dispatched(&self) -> u64

Examples found in repository

examples/cutlass_gemm_fp8.rs (line 27)

fn main() {
    let actor = CutlassActor::new(16);
    let req = GemmRequest::<F8E4m3>::new(GemmShape::new(4096, 4096, 4096), SmArch::Sm90a)
        .with_epilogue(GemmEpilogue::LinearReLU {
            alpha: 1.0,
            beta: 0.0,
        });

    println!("plan key: {:?}", req.plan_key());
    let (src, name) = req.render_cu();
    println!("kernel:   {name}");
    println!("--- generated .cu ---");
    println!("{src}");

    actor.handle(CutlassMsg::Gemm(Box::new(req.clone())));
    actor.handle(CutlassMsg::Gemm(Box::new(req)));

    println!("dispatched: {}", actor.inner().dispatched());
    println!("plan cache len: {}", actor.inner().plan_cache.len());
}