Struct MatmulRequest 

pub struct MatmulRequest<T: GemmSupported> {

    pub a: GpuRef<T>,
    pub b: GpuRef<T>,
    pub c: GpuRef<T>,
    pub d: Option<GpuRef<T>>,
    pub m: i32,
    pub n: i32,
    pub k: i32,
    pub alpha: T::Scalar,
    pub beta: T::Scalar,
    pub transa: bool,
    pub transb: bool,
    pub lda: i64,
    pub ldb: i64,
    pub ldc: i64,
    pub ldd: i64,
    pub epilogue: Epilogue,
    pub bias: Option<GpuRef<T>>,
    pub gelu_aux: Option<GpuRef<T>>,
    pub scales: ScaleSet,
    pub workspace_size: usize,
    pub reply: Sender<Result<(), GpuError>>,
}

Typed matmul request. Public surface; instantiated by callers.

Fields

a: GpuRef<T>

b: GpuRef<T>

c: GpuRef<T>

d: Option<GpuRef<T>>

Optional explicit D output buffer. cuBLASLt allows out-of-place matmul where the result lands in D rather than in-place into C. Required for fp8 (the scale-back step produces a different dtype than the accumulator).

m: i32

n: i32

k: i32

alpha: T::Scalar

beta: T::Scalar

transa: bool

transb: bool

lda: i64

ldb: i64

ldc: i64

ldd: i64

epilogue: Epilogue

bias: Option<GpuRef<T>>

gelu_aux: Option<GpuRef<T>>

scales: ScaleSet

workspace_size: usize

Hint for the heuristic: maximum workspace bytes the algorithm search may use. A reasonable default is 4 * 1024 * 1024 (cuBLASLt’s standard 4 MiB minimum).

reply: Sender<Result<(), GpuError>>

Trait Implementations

impl BlasLtDispatch for MatmulRequest<f32>

fn dtype_kind(&self) -> DTypeKind

fn dispatch(self: Box<Self>, ctx: &BlasLtDispatchCtx<'_>)

impl BlasLtDispatch for MatmulRequest<f64>

f64 (cudarc 0.19.4 has no Matmul<f64> impl).

fn dtype_kind(&self) -> DTypeKind

fn dispatch(self: Box<Self>, _ctx: &BlasLtDispatchCtx<'_>)

impl<T: GemmSupported> Debug for MatmulRequest<T>

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

Formats the value using the given formatter.