NVlabs/CuTe

Reference implementation and examples of the CuTe Layout representation and algebra.

What it solves

PyCuTe provides a pure-Python reference implementation of the hierarchical layout-and-tensor algebra used in NVIDIA's CUTLASS 3.x. It allows developers to learn the complex algebra, prototype new data transformations, and generate test vectors for C++ and DSL implementations without requiring a GPU.

How it works

At its core, PyCuTe defines a Layout as a function that maps coordinates to offsets using a Shape (the domain) and a Stride (the mapping logic). It implements a suite of algebraic operations to manipulate these layouts, including:

  • Coalesce: Simplifies layouts to the minimum number of modes.
  • Composition: Indexes one layout through another.
  • Complement: Identifies missing modes to fill a codomain.
  • Logical Divide/Product: Handles tiling and repetition of patterns.
  • Inversion/Nullspace: Analyzes and inverts maps.

It also supports F2 (XOR-swizzle) strides to help visualize and manage shared-memory bank conflicts. A thin Tensor layer pairs these layouts with an Accessor to provide a data model.

Who it’s for

  • Developers working with CUTLASS 3.x or the CuTe DSL.
  • Engineers prototyping GPU tensor operations and memory layouts.
  • Researchers studying hierarchical layout algebra.

Highlights

  • GPU-Independent: Runs in plain Python (3.10+) with no hardware requirements.
  • Visualization Tools: Includes utilities to render layouts as ASCII tables, colored SVGs, or TikZ/PDFs.
  • Comprehensive Algebra: Implements the full layout algebra described in the CuTe Whitepaper.
  • Swizzle Support: Specifically handles XOR-swizzling to visualize bank conflicts.

Related

  • Project
    scverse/pertpy

    A framework for analyzing large-scale single-cell perturbation experiments, providing tools to harmonize data and analyze how cells respond to genetic, drug, or environmental stimuli.

  • Project
    o19s/quepid

    Quepid is a tool that turns search relevance tuning into a repeatable engineering process, helping teams collaborate and test search results to improve app search quality.

  • Project
    drivelineresearch/autoresearch-claude-code

    autoresearch‑claude‑code is a Python framework that runs budget‑constrained auto‑research loops with Claude Code or Codex agents. You define a goal, a deterministic scorer, and files the agent may edit; the system creates a Git branch, measures each change, and automatically commits only statistically‑significant improvements while respecting run‑time and turn limits.

  • Project
    elbwalker/walkerOS

    A warehouse-native event collection layer that unifies tracking across multiple destinations like GA4 and BigQuery using a config-as-code approach.