cutile

NVIDIA CuTile Python

Tile-based GPU programming DSL in Python. Write kernels that operate on tiles (multidimensional array chunks) instead of individual threads. The compiler automatically leverages tensor cores, TMA, and hardware features. Portable across NVIDIA Blackwell architectures.

Official Sources:

• Documentation
• GitHub
• TileGym (tutorials/benchmarks)
• Blog: Simplify GPU Programming
• Blog: Tuning Flash Attention

Requirements

• GPU: Compute capability 10.x or 12.x (Blackwell)
• NVIDIA Driver: r580+
• CUDA Toolkit: 13.1+
• Python: 3.10-3.13

Quick Start

pip install cuda-tile
pip install cupy-cuda13x  # or torch for PyTorch integration

import cuda.tile as ct
import cupy

TILE_SIZE = 16

@ct.kernel
def vector_add(a, b, result):
    pid = ct.bid(0)
    a_tile = ct.load(a, index=(pid,), shape=(TILE_SIZE,))
    b_tile = ct.load(b, index=(pid,), shape=(TILE_SIZE,))
    ct.store(result, index=(pid,), tile=a_tile + b_tile)

# Host code
n = 1024
a = cupy.random.rand(n).astype(cupy.float32)
b = cupy.random.rand(n).astype(cupy.float32)
c = cupy.empty(n, dtype=cupy.float32)
grid = (ct.cdiv(n, TILE_SIZE), 1, 1)
ct.launch(cupy.cuda.get_current_stream(), grid, vector_add, (a, b, c))

Core Concepts

Kernels and Launch

@ct.kernel
def my_kernel(a, b, c, TILE: ct.Constant[int]):
    pid = ct.bid(0)  # block index along axis 0
    # ... tile operations ...

# Launch: (stream, grid, kernel, args)
ct.launch(stream, (num_blocks_x, num_blocks_y, num_blocks_z), my_kernel, (a, b, c, 128))

• @ct.kernel marks GPU entry points (cannot call directly)
• ct.bid(axis) returns block index (0, 1, or 2) as int32
• ct.num_blocks(axis) returns grid size along axis
• ct.Constant[int] marks compile-time constant parameters (generates distinct kernel per value)
• Grid is a 3-tuple (x, y, z)