accelrod 0.1.0

A library for benchmarking the tflops of accelerators

Accelrod

Installation

pip install -U accelrod

[all] includes nbformat to visualize the result in notebook

pip install -U accelrod[all]

Example

import torch

from accelrod.benchmark import benchrun, plot_result

# without max_dimension, the matrix size will be determined based on available VRAM
result = benchrun(
    algorithm="GEMM",
    device="auto",
    as_dataframe="pandas",
    params={
        # "max_dimension": 2048,
        "dtype": [torch.float32, torch.float16, torch.bfloat16],
    },
)

# plot in notebook
import plotly.offline as pyo

pyo.init_notebook_mode()
plot_result(result)

Developer guide

Setup working env / dependencies

Sync the project's dependencies with the environment.

uv sync

lock dependencies declared in a pyproject.toml

uv pip compile pyproject.toml -o requirements.txt

Add dev dependencies

uv add --dev ruff pytest

Editable Installs

uv pip install -e .[all,dev]

Test package with local build

uv pip install dist/accelrod-0.1.0-py3-none-any.whl uv run --with accelrod --no-project --refresh-package accelrod -- python -c "import accelrod"