Installation

About

This page shows how to install d9d from PyPI or from a checkout of the repository. Some features need extras: optional dependencies that you install with d9d[<extra>].

From PyPI

1

pip install d9d

1

poetry add d9d

1

uv add d9d

From a Checkout

The main branch can be ahead of the latest release. To get these changes, install d9d from a checkout:

1
2
3

git clone https://github.com/d9d-project/d9d.git
cd d9d
pip install -e .

To work on d9d itself, follow the development setup in CONTRIBUTING.md.

Extras

Install an extra with its name in brackets. The quotes keep the shell from expanding the brackets:

1
2

pip install "d9d[tracker-aim]"
pip install -e ".[tracker-aim]"  # From a checkout.

The extras are:

• d9d[tracker-wandb], d9d[tracker-tensorboard], d9d[tracker-mlflow], d9d[tracker-aim]: The SDKs of the experiment trackers.
• d9d[visualization]: Plotly, for plotting learning rate schedules.
• d9d[linear-attention]: Flash Linear Attention kernels for linear attention.
• d9d[backend-sdpa-flash-attention-2]: FlashAttention 2 kernels as an SDPA backend. You may need to build FlashAttention 2 from source if no prebuilt wheel matches your setup.
• d9d[backend-sdpa-flash-attention-4]: FlashAttention 4 kernels as an SDPA backend.
• d9d[backend-grouped-mm-quack]: QuACK kernels as a grouped GEMM backend for MoE experts.
• d9d[backend-grouped-mm-grouped-gemm]: The grouped_gemm library as a grouped GEMM backend for MoE experts. You must build and install it manually first.
• d9d[backend-expert-parallel-deepep]: DeepEP for expert parallelism. You must build and install it manually first.
• d9d[cce]: Fused cross-entropy kernels for the causal language modelling head. You must install Cut Cross Entropy from its repository first.