cuBLAS & Quantized FP8 Matrix Multiplication in Rust

Zero-Overhead cuBLAS Bindings

While custom PTX is vital for unique activations, dense GEMM (General Matrix Multiply) operations are best delegated to cuBLAS, NVIDIA's hand-tuned assembly library. cudarc exposes native cuBLAS bindings in safe Rust:

use cudarc::cublas::{CudaBlas, GemmConfig, Gemm};

pub fn run_cublas_gemm(
    blas: &CudaBlas,
    a: &CudaSlice,
    b: &CudaSlice,
    c: &mut CudaSlice,
    m: usize, n: usize, k: usize,
) -> Result<(), Box> {
    let cfg = GemmConfig {
        alpha: 1.0,
        beta: 0.0,
        m: m as i32,
        n: n as i32,
        k: k as i32,
        lda: k as i32,
        ldb: n as i32,
        ldc: n as i32,
    };

    // Invokes cuBLAS SGEMM with zero pointer allocation overhead
    unsafe { blas.gemm(cfg, a, b, c) }?;
    Ok(())
}