def get_all_configs()

in experimental/ragged_inference/triton_v2_qk_dotprod.py [0:0]


def get_all_configs():
    return [
        # basic configs for compute-bound matmuls
        triton.Config(
            {"BLOCK_M": 128, "BLOCK_K": 256, "BLOCK_D": 32},
            num_stages=3,
            num_warps=8,
        ),
        triton.Config(
            {"BLOCK_M": 256, "BLOCK_K": 128, "BLOCK_D": 32},
            num_stages=3,
            num_warps=8,
        ),
        triton.Config(
            {"BLOCK_M": 256, "BLOCK_K": 64, "BLOCK_D": 32},
            num_stages=4,
            num_warps=4,
        ),
        triton.Config(
            {"BLOCK_M": 64, "BLOCK_K": 256, "BLOCK_D": 32},
            num_stages=4,
            num_warps=4,
        ),
        triton.Config(
            {"BLOCK_M": 128, "BLOCK_K": 128, "BLOCK_D": 32},
            num_stages=4,
            num_warps=4,
        ),
        triton.Config(
            {"BLOCK_M": 128, "BLOCK_K": 64, "BLOCK_D": 32},
            num_stages=4,
            num_warps=4,
        ),
        triton.Config(
            {"BLOCK_M": 64, "BLOCK_K": 128, "BLOCK_D": 32},
            num_stages=4,
            num_warps=4,
        ),
        triton.Config(
            {"BLOCK_M": 128, "BLOCK_K": 32, "BLOCK_D": 32},
            num_stages=4,
            num_warps=4,
        ),
        triton.Config(
            {"BLOCK_M": 64, "BLOCK_K": 32, "BLOCK_D": 32},
            num_stages=5,
            num_warps=2,
        ),
    ] + get_configs_io_bound()