|
|
--- |
|
|
license: mit |
|
|
--- |
|
|
|
|
|
# utils wheels for RTX5090 |
|
|
|
|
|
As title, lot of third party librarys which are important for running NN lack prebuilt wheel for RTX5090, so I decide to compile them bymyself and upload them here. |
|
|
|
|
|
***Use At Your Own Risks, Check Official Release To See if There Are Any Official Supports On Your HW Regularly*** |
|
|
|
|
|
### Inlcuded Library |
|
|
* Flash Attention |
|
|
* Xformers (with cutlass/flash attention built-in) |
|
|
* NATTEN |
|
|
* SageAttention |
|
|
* vLLM |
|
|
|
|
|
### IMPORTANT NOTE |
|
|
1. I only ensure those wheels can works on RTX50 series (sm120) GPUs, if your platform is mixed with different sm/cu arch GPUs, you may still need to compile them by yourself |
|
|
2. Env |
|
|
* Pytorch: 2.7.0 |
|
|
* CUDA: 12.8(12.8.1) |
|
|
* CUDNN: 9.8 |
|
|
* Compiler: GCC13 |
|
|
* Tested platform: Ubuntu 22.04 and 24.04 |
|
|
* CPU arch: amd64 (x86-64) |
|
|
3. Not all the wheels are fully functional (due to deps things or source implementation), for example, cutlass w8a8 scaled mm is not working in vllm, you need to use `VLLM_TEST_FORCE_FP8_MARLIN=1` to make VLLM works normally with fp8 weight quantization. If you are using flash attention, you need `VLLM_FLASH_ATTN_VERSION=2` to make it work on 5090 |
|
|
4. If you meet any problem or need wheels for specific setup you can open discussion, but I can't ensure I will do it or not. |
|
|
|
|
|
### Tips |
|
|
* Install `triton==3.3.1` for better RTX50 series support |
|
|
* Install `nvidia-nccl-cu12==2.26.5` for correct multi-gpu deps for RTX50 series |
|
|
* Torch 2.7.0 use 2.26.2 in their requirements which is not compatibile with RTX50 series, you should install this from pypi directly with `pip isntall nvidia-nccl-cu12>2.26.2` |
|
|
* I build all those wheel with `python -m build -n -w .` which is more suitable in modern python packaging, I recommend all the user who want to compile those wheel by themselves follow this scheme. (No matter the project use pyproject or setup.py, build package will works) |
