pipenaut 0.1.0

Experience pipeline parallelism on your laptop. Naive, GPipe, 1F1B — one command.

pipenaut

Experience pipeline parallelism on your laptop. Naive, GPipe, 1F1B — one command.

No GPUs needed. No cluster setup. Just pip install and run.

Quick Start

pip install pipenaut

# Compare all 3 pipeline schedules
pipenaut compare --workers 4

# Run a single schedule
pipenaut run --schedule 1f1b --workers 4 --steps 20

# Learn about a schedule
pipenaut explain 1f1b

What is this?

Pipeline parallelism splits a neural network across multiple devices (or processes), with each device running one "stage" of the model. Data flows through the pipeline like an assembly line.

pipenaut lets you experience 3 classic pipeline schedules on your local CPU:

Schedule	Strategy	Bubble Overhead
Naive	Forward all → Backward all (sequential)	~75%
GPipe	Forward all chunks → Backward all chunks	~43%
1F1B	Interleave forward and backward	~19%

Commands

pipenaut compare

Runs all 3 schedules and prints a side-by-side comparison with pipeline timelines and a results table.

pipenaut run

Run a single schedule with detailed output.

Flag	Default	Description
--schedule, -s	1f1b	Schedule: naive, gpipe, 1f1b
--workers, -w	4	Number of pipeline stages
--steps	30	Training steps
--chunks, -c	8	Micro-batches
--batch-size	32	Batch size
--dim	128	Hidden dimension
--layers	16	Total model layers

pipenaut explain

Learn how a schedule works with ASCII diagrams and explanations.

pipenaut explain naive
pipenaut explain gpipe
pipenaut explain 1f1b

Requirements

• Python >= 3.10
• PyTorch >= 2.0
• Works on macOS, Linux, and Windows
• CPU only (CUDA used automatically if available)

License

MIT