cutile-python-dsl.md

id	doc-cutile-python-dsl
title	cuTile Python Documentation
url	https://docs.nvidia.com/cuda/cutile-python/
source_category	official-doc
architectures	sm80 sm90 sm100 sm110 sm120
tags	cutile python gemm tma
retrieved_at	2026-08-18

cuTile Python documentation

cuTile is NVIDIA's tile-based GPU programming model and Python DSL. A function decorated with @ct.kernel executes across a logical grid of blocks and is queued from the host with ct.launch. Global arrays are mutable, strided storage; tiles are immutable kernel values without a programmer-defined physical layout. Tile dimensions must be compile-time powers of two.

The current documentation says cuTile automatically uses hardware capabilities such as Tensor Cores and tensor memory accelerators and is portable across supported NVIDIA architectures. That is a compiler/runtime abstraction: it does not guarantee that every load becomes TMA or that every matrix operation lowers to a particular PTX instruction.

This contiguous excerpt is from the official vector-add example:

@ct.kernel
def vector_add(a, b, c, tile_size: ct.Constant[int]):
    pid = ct.bid(0)
    a_tile = ct.load(a, index=(pid,), shape=(tile_size,))
    b_tile = ct.load(b, index=(pid,), shape=(tile_size,))
    result = a_tile + b_tile
    ct.store(c, index=(pid, ), tile=result)

As of the access date, the quickstart requires compute capability 8.x, 9.x, 10.x, 11.x, or 12.x, driver R580 or later, and CUDA Toolkit 13.1 or later (or matching Python packages for the compiler dependencies). The latest listed cuTile Python release was 1.5.0, dated 2026-07-08. CUDA Toolkit 13.3 added Hopper support and block-scaled MMA to the documented Python stack.

Because the API is evolving, examples should be pinned to a cuTile release; older local claims about compiler-selected worker counts, automatic swizzles, or guaranteed tcgen05/TMEM lowering are not treated as public contracts.