{
  "tile_config": {
    "tile_m": [128, 256],
    "tile_n": [128, 256],
    "tile_k": [32, 64],
    "warp_m": [2, 4],
    "warp_n": [2, 4],
    "warp_k": [1],
    "warp_tile_m": [16, 32],
    "warp_tile_n": [16, 32],
    "warp_tile_k": [16]
  },
  "trait_config": {
    "pipeline": ["compv4"],
    "epilogue": ["cshuffle"],
    "scheduler": ["intrawave"],
    "pad_m": [false],
    "pad_n": [false],
    "pad_k": [false],
    "persistent": [false, true]
  },
  "multi_d_config": {
    "elementwise_ops": ["MultiDAdd", "Relu", "Gelu"],
    "num_d_tensors": [1, 2]
  }
}