Function blockwise_matmul

tests/python/contrib/test_cutlass_gemm.py:248–272

(
    x_fp8_np: np.ndarray,
    x_scale_np: np.ndarray,
    w_np: np.ndarray,
    w_scale_np: np.ndarray,
    block_size: tuple[int, int],
    dtype: str,
)

def blockwise_matmul(
    x_fp8_np: np.ndarray,
    x_scale_np: np.ndarray,
    w_np: np.ndarray,
    w_scale_np: np.ndarray,
    block_size: tuple[int, int],
    dtype: str,
):
    o_np = np.zeros((x_fp8_np.shape[0], w_np.shape[0]), dtype=dtype)
    for j in range(w_scale_np.shape[0]):
        for k in range(w_scale_np.shape[1]):
            o_np[:, j * block_size[0] : min((j + 1) * block_size[0], w_np.shape[0])] += (
                np.matmul(
                    x_fp8_np[
                        :, k * block_size[1] : min((k + 1) * block_size[1], x_fp8_np.shape[1])
                    ].astype(dtype),
                    w_np[
                        j * block_size[0] : min((j + 1) * block_size[0], w_np.shape[0]),
                        k * block_size[1] : min((k + 1) * block_size[1], w_np.shape[1]),
                    ].T.astype(dtype),
                )
                * x_scale_np[:, k : k + 1]
                * w_scale_np[j, k]
            )
    return o_np


def blockwise_bmm(