README.md

cuBLAS LT — Lightweight BLAS Examples

1 example demonstrating lightweight GEMM with algorithm heuristics. Enable with -Dcublaslt=true.

Build & Run

zig build run-cublaslt-lt_sgemm -Dcublaslt=true

---

Examples

Example	File	Description
lt_sgemm	lt_sgemm.zig	SGEMM via cuBLASLt with automatic algorithm heuristic selection

---

When to Use cuBLAS LT

cuBLAS LT extends cuBLAS with:

• Algorithm search: getHeuristics selects the fastest GEMM algorithm for your matrix shapes
• Mixed precision: FP16/BF16 input with FP32 accumulation
• Custom epilogues: ReLU, bias addition, GELU fused into GEMM
• Layout control: Row/column major, transposed, strided

Key API

const lt = @import("zcuda").cublaslt;

const blas_lt = try lt.CublasLtContext.init(ctx);
defer blas_lt.deinit();

// Create matmul descriptor
const matmul_desc = try blas_lt.createMatmulDesc(.{
    .compute_type = .compute_32f,
    .scale_type   = .real_32f,
    .transa       = .no_transpose,
    .transb       = .no_transpose,
});

// Query heuristic
const algos = try blas_lt.getHeuristics(matmul_desc, a_layout, b_layout, c_layout, d_layout, .{});
const algo = algos[0].algo;

// Execute
try blas_lt.matmulWithAlgo(matmul_desc, alpha, d_a, a_layout, d_b, b_layout,
    beta, d_c, c_layout, d_d, d_layout, algo, workspace, stream);

→ Full API reference: docs/cublaslt/README.md