wan2gp-blackwell-attn-wheels

Wan2GP accelerated-attention wheels (RTX 50-series / Blackwell)

Prebuilt CUDA wheels for Wan2GP / WanGP so you don't have to compile them on every instance:

⚠️ Compatibility — these are pinned to one exact stack

• Python 3.12
• PyTorch 2.7.1+cu128
• GPU arch sm120 (compute capability 12.0) — RTX 50-series / Blackwell only

They will not load on a different Python/torch version or a non-Blackwell GPU (e.g. a 4090 is sm89). Rebuild from source for other stacks.

Install

pip install --no-deps \
  https://huggingface.co/deg1988/wan2gp-blackwell-attn-wheels/resolve/main/spas_sage_attn-0.1.0-cp312-cp312-linux_x86_64.whl \
  https://huggingface.co/deg1988/wan2gp-blackwell-attn-wheels/resolve/main/sageattention-2.2.0-cp312-cp312-linux_x86_64.whl

Then restart WanGP and set Attention mode → sage2.

Vast.ai provisioning (auto-install at boot)

Launch the instance with PROVISIONING_MANIFEST pointing at provision-attn.yaml (in this repo) and set:

ATTN_WHEEL_BASE=https://huggingface.co/deg1988/wan2gp-blackwell-attn-wheels/resolve/main

How these were built

Compiled on a Vast.ai Wan2GP instance (RTX 5090) with:

export TORCH_CUDA_ARCH_LIST="12.0"          # Blackwell sm120
export NVCC_APPEND_FLAGS="-DNDEBUG"          # work around cuda_fp8.hpp __assert_fail (CUDA 12.9 + glibc 2.39)
export CPATH="<venv>/nvidia/*/include:$CPATH"  # this image omits system cusparse/cublas/cusolver headers
pip wheel --no-build-isolation --no-deps -w ./wheels \
  git+https://github.com/woct0rdho/SpargeAttn.git \
  git+https://github.com/thu-ml/SageAttention.git