warp.config.enable_mathdx_gemm

warp.config.enable_mathdx_gemm: bool = True

Use libmathdx (cuBLASDx) for tile_matmul on GPU when available.

When False, tile_matmul falls back to a scalar GEMM implementation, which avoids the slow libmathdx LTO compilation at the cost of runtime performance.

This setting can be overridden at the module level by setting the "enable_mathdx_gemm" module option.