Skip to main content

compile_to_cuda_gemm

logicaffeine_cli::compile

Function compile_to_cuda_gemm 

Source 
pub fn compile_to_cuda_gemm(
    m: usize,
    k: usize,
    n: usize,
) -> Result<CudaGemmArtifact, ParseError>
Expand description

Emit an int8 weight-stationary GEMM CUDA program for O(m×n) = A(m×k)·W(k×n) — the DP4A kernel path of the CUDA tier. Uses a deterministic int8 sample (crate::codegen_cuda::sample_gemm_data) baked into the .cu; its numerics are anchored to the certified eval_systolic reference by cuda_matrix_lock. Refuses (typed) if the RDU-shaped target lacks int8 weight-stationary GEMM.