DeepSpeed 0.17.4 — Windows x86_64 (Python 3.12, CUDA 12.9)

This repository provides a prebuilt DeepSpeed 0.17.4 wheel for Windows 64-bit, built against:

• Python: 3.12
• CUDA Toolkit: 12.9 (NVCC 12.9.41)
• PyTorch: 2.8.0+cu129

The wheel includes compiled CUDA kernels, with Sparse Attention and Evoformer Attention disabled for Windows compatibility.

---

📦 Wheel

• Filename: deepspeed-0.17.4+c4b1a8cb-cp312-cp312-win_amd64.whl
• SHA256 checksum: 6d0964a6b204bfd9eff2bf972d0c6587fe6c66c0f29c0972b0f6080a670b1ec3

---

🛠 Build Environment

Component	Version / Details
OS	Windows 10/11 x64
Python	3.12.10 (venv)
CUDA Toolkit	12.9 (NVCC 12.9.41)
NVIDIA Driver	Compatible with CUDA 12.9
Visual Studio	2022 Build Tools (MSVC v143), x64 Native Tools prompt
PyTorch	2.8.0+cu129
DeepSpeed commit	c4b1a8cb (tag v0.17.4)
Build frontend	pypa/build

---

⚙ Build Flags Used

set DS_BUILD_OPS=1
set DS_BUILD_SPARSE_ATTN=0
set DS_BUILD_EVOFORMER_ATTN=0
set TORCH_CUDA_ARCH_LIST=8.9+PTX

These flags:

• Enable main DeepSpeed CUDA ops
• Disable Windows-problematic ops (SPARSE_ATTN, EVOFORMER_ATTN)
• Target NVIDIA Ada GPUs (8.9) with PTX fallback

---

📥 Installation

1. Install a PyTorch build for CUDA 12.9:

   pip install torch torchvision --index-url https://download.pytorch.org/whl/cu129

2. Install the wheel:

   pip install deepspeed-0.17.4+c4b1a8cb-cp312-cp312-win_amd64.whl

3. Verify:

   python -c "import torch, deepspeed; print(torch.__version__, torch.version.cuda); print('DeepSpeed', deepspeed.__version__)"