CUDA 12.8 install of dflash[sglang] fails

[Sylvan820]

I am trying to set up the SGLang backend in a CUDA 12.8 environment, but the standard install path fails:

uv pip install -e ".[sglang]"

Environment:

CUDA: 12.8
Python: 3.11
Install command: uv pip install -e ".[sglang]"
Backend: sglang

From pyproject.toml, the SGLang extra currently depends on:

sglang[all] @ git+https://github.com/sgl-project/sglang.git@refs/pull/23000/head#subdirectory=python

However, when installing in a CUDA 12.8 environment, many dependencies appear to resolve toward CUDA 13.0 / newer SGLang-compatible packages, and the installation fails. My guess is that the latest SGLang PR dependency may currently assume or prefer CUDA 13.0-compatible packages, while CUDA 12.8 is still common in cloud environments.

Would it be possible to provide one of the following?

1. A recommended installation command for CUDA 12.8.
2. A known-good constraints file for uv pip install -e ".[sglang]".
3. A pinned SGLang commit/PR that works with CUDA 12.8.
4. An official Docker image for the SGLang backend.
5. Documentation clarifying which CUDA / PyTorch / SGLang versions are expected to work.

For example, a CUDA 12.8 setup guide could specify the exact compatible versions of:

torch
torchvision
torchaudio
sglang
flashinfer
cuda-python
transformers

This would make it much easier to reproduce the SGLang backend environment reliably.

Thanks!