Astral GPU Indexes

oritsnile1 pts0 comments

Astral GPU indexes

Astral GPU indexes

Pre-built, GPU-enabled wheels for the Python ecosystem.

Pre-built wheels for popular GPU-enabled packages in the PyTorch ecosystem, like FlashAttention, across a range of Python versions, CPU architectures, CUDA versions, and PyTorch versions, based on open source build pipelines.

Following the PyTorch convention, the Astral GPU indexes use a separate index URL for each supported CUDA version (e.g., /simple/cu128), and publish a separate wheel version for each supported CUDA and PyTorch combination (e.g., +cu.12.8.torch.2.11).

At present, the Astral GPU indexes include builds for the following packages:

Adan

causal-conv1d

DeepEP

DeepGEMM

DeepSpeed

Detectron2

Flash Attention

Flash Attention 3

Grouped GEMM

Mamba SSM

MegaBlocks

MMCV

PyCUDA

pyg-lib

PyTorch Scatter

PyTorch3D

SageAttention2++

SageAttention3

Transformer Engine

Transformer Engine CUDA 12 Core

Transformer Engine CUDA 13 Core

Transformer Engine PyTorch

vLLM

Available indexes

CPU<br>cpu / 3 packages

CUDA 11.8<br>cu118 / 1 package

CUDA 12.1<br>cu121 / 12 packages

CUDA 12.4<br>cu124 / 14 packages

CUDA 12.6<br>cu126 / 18 packages

CUDA 12.8<br>cu128 / 22 packages

CUDA 12.9<br>cu129 / 18 packages

CUDA 13.0<br>cu130 / 17 packages

CUDA 13.2<br>cu132 / 8 packages

Using an index with uv

Use uv add to pin a package to the Astral GPU index.

Copy

uv add flash-attn==2.8.3.post1+cu.12.6.torch.2.12 --index astral-cu126=https://wheels.astral.sh/simple/cu126/

The uv pip interface

Add the Astral GPU index as an extra --index at install time.

Copy

uv pip install \<br>flash-attn==2.8.3.post1+cu.12.6.torch.2.12 \<br>--index astral-cu126=https://wheels.astral.sh/simple/cu126/

pip

Add the Astral GPU index as an extra --extra-index-url at install time.

Copy

python -m pip install \<br>flash-attn==2.8.3.post1+cu.12.6.torch.2.12 \<br>--extra-index-url https://wheels.astral.sh/simple/cu126/

Manifest

Package versions available in the CUDA 12.6 index.

CUDA 12.6<br>18 packages

adan0.0.2+cu.12.6.torch.2.11, 0.0.2+cu.12.6.torch.2.10, 0.0.2+cu.12.6.torch.2.9, 0.0.2+cu.12.6.torch.2.8, 0.0.2+cu.12.6.torch.2.7, 0.0.2+cu.12.6.torch.2.6

causal-conv1d1.6.2.post1+cu.12.6.torch.2.11, 1.6.2.post1+cu.12.6.torch.2.10, 1.6.2.post1+cu.12.6.torch.2.9, 1.6.2.post1+cu.12.6.torch.2.8, 1.6.2.post1+cu.12.6.torch.2.7, 1.6.2.post1+cu.12.6.torch.2.6, 1.6.1+cu.12.6.torch.2.10, 1.6.1+cu.12.6.torch.2.9, 1.6.1+cu.12.6.torch.2.8, 1.6.1+cu.12.6.torch.2.7, 1.6.1+cu.12.6.torch.2.6, 1.6.0+cu12.6torch2.10.0cxx11abitrue, 1.6.0+cu12.6torch2.9.0cxx11abitrue, 1.6.0+cu12.6torch2.8.0cxx11abitrue, 1.6.0+cu12.6torch2.7.1cxx11abitrue, 1.6.0+cu12.6torch2.6.0cxx11abitrue, 1.5.4+cu12.6torch2.9.0cxx11abitrue, 1.5.4+cu12.6torch2.8.0cxx11abitrue, 1.5.4+cu12.6torch2.7.1cxx11abitrue, 1.5.4+cu12.6torch2.6.0cxx11abitrue

deepspeed0.19.2+cu.12.6.torch.2.12, 0.19.2+cu.12.6.torch.2.11, 0.19.2+cu.12.6.torch.2.10, 0.19.2+cu.12.6.torch.2.9, 0.19.2+cu.12.6.torch.2.8, 0.19.2+cu.12.6.torch.2.7, 0.19.1+cu.12.6.torch.2.12, 0.19.1+cu.12.6.torch.2.11, 0.19.1+cu.12.6.torch.2.10, 0.19.1+cu.12.6.torch.2.9, 0.19.1+cu.12.6.torch.2.8, 0.19.1+cu.12.6.torch.2.7, 0.19.0+cu.12.6.torch.2.12, 0.19.0+cu.12.6.torch.2.11, 0.19.0+cu.12.6.torch.2.10, 0.19.0+cu.12.6.torch.2.9, 0.19.0+cu.12.6.torch.2.8, 0.19.0+cu.12.6.torch.2.7, 0.18.9+cu.12.6.torch.2.11, 0.18.9+cu.12.6.torch.2.10, 0.18.9+cu.12.6.torch.2.9, 0.18.9+cu.12.6.torch.2.8, 0.18.9+cu.12.6.torch.2.7, 0.18.8+cu.12.6.torch.2.10, 0.18.8+cu.12.6.torch.2.9, 0.18.8+cu.12.6.torch.2.8, 0.18.8+cu.12.6.torch.2.7, 0.18.7+cu.12.6.torch.2.10, 0.18.7+cu.12.6.torch.2.9, 0.18.7+cu.12.6.torch.2.8, 0.18.7+cu.12.6.torch.2.7, 0.18.6+cu.12.6.torch.2.10, 0.18.6+cu.12.6.torch.2.9, 0.18.6+cu.12.6.torch.2.8, 0.18.6+cu.12.6.torch.2.7, 0.18.5+cu.12.6.torch.2.10, 0.18.5+cu.12.6.torch.2.9, 0.18.5+cu.12.6.torch.2.8, 0.18.5+cu.12.6.torch.2.7, 0.18.4+cu12.6torch2.10.0cxx11abitrue, 0.18.4+cu12.6torch2.9.0cxx11abitrue, 0.18.4+cu12.6torch2.8.0cxx11abitrue, 0.18.4+cu12.6torch2.7.1cxx11abitrue, 0.18.3+cu12.6torch2.9.0cxx11abitrue, 0.18.3+cu12.6torch2.8.0cxx11abitrue, 0.18.3+cu12.6torch2.7.1cxx11abitrue, 0.18.2+cu12.6torch2.9.0cxx11abitrue, 0.18.2+cu12.6torch2.8.0cxx11abitrue, 0.18.2+cu12.6torch2.7.1cxx11abitrue, 0.18.1+cu12.6torch2.9.0cxx11abitrue, 0.18.1+cu12.6torch2.8.0cxx11abitrue, 0.18.1+cu12.6torch2.7.1cxx11abitrue

detectron20.6+cu.12.6.torch.2.12, 0.6+cu.12.6.torch.2.11, 0.6+cu.12.6.torch.2.10, 0.6+cu.12.6.torch.2.9, 0.6+cu.12.6.torch.2.8, 0.6+cu.12.6.torch.2.7, 0.6+cu.12.6.torch.2.6

flash-attn2.8.3.post1+cu.12.6.torch.2.12, 2.8.3.post1+cu.12.6.torch.2.11, 2.8.3.post1+cu.12.6.torch.2.10, 2.8.3.post1+cu.12.6.torch.2.9, 2.8.3.post1+cu.12.6.torch.2.8, 2.8.3.post1+cu.12.6.torch.2.7, 2.8.3.post1+cu.12.6.torch.2.6, 2.8.3+cu.12.6.torch.2.11, 2.8.3+cu.12.6.torch.2.10, 2.8.3+cu.12.6.torch.2.9, 2.8.3+cu.12.6.torch.2.8, 2.8.3+cu.12.6.torch.2.7, 2.8.3+cu.12.6.torch.2.6

grouped-gemm0.3.0+cu.12.6.torch.2.7

mamba-ssm2.3.2.post1+cu.12.6.torch.2.11, 2.3.2.post1+cu.12.6.torch.2.10,...

torch cu12 6torch2 post1 0cxx11abitrue cuda

Related Articles