PyPTO

A high-performance programming framework for tile-centric computing.

PyPTO lets you write NPU kernels and their orchestration in Python, compiles them through a multi-level IR, and hands the resulting task graph to a runtime that executes it on device.

Where to start

You are...	Start here
New to PyPTO	Getting Started — install, first program, compile and run
Writing kernels	Language Guide — types, functions, control flow, memory, scopes
Looking for an operator	Operation Reference — the pl.* / pl.tensor.* / pl.tile.* surface
Chasing a wrong result	Torch Codegen Debug Guide — run the IR through PyTorch and compare
Working on the compiler	Developer documentation — IR, passes, code generation
Reading generated code	PTO ISA reference — cluster architecture and instruction semantics

PyPTO and the runtime

PyPTO is the compiler and programming language. The runtime that schedules and executes what it produces is a separate project, hw-native-sys/simpler, included here as the runtime/ submodule.

The split decides what you will find where:

Concern	Owner	Documentation
The pl.* language, ir.compile(), IR and passes	PyPTO	This site
pypto.runtime — ChipWorker, DeviceTensor, RunConfig, benchmark	PyPTO (its own Python layer)	This site
Scheduler internals, graph building, message queue, tensormap and ring buffers	simpler	https://hw-native-sys.github.io/simpler/

pypto.runtime is PyPTO's own Python wrapper, not a simpler API — it is documented here. What is linked out is simpler's internals.

About this documentation

• English is authoritative. docs/zh/ mirrors it page for page; use the language selector in the header to switch.
• The markdown under docs/ is the single source of truth and is readable directly on GitHub. This site is a build artifact and is never committed.
• The user manual is being expanded from the four guides above into a full chaptered manual (tutorials, distributed programming, performance, accuracy debugging). See issue #2120 for the plan.