Bump torch to 2.9.1 with auto-patched cuDNN 9.17

[yondonfu]

Summary

• Upgrade torch 2.8.0 → 2.9.1, torchvision 0.23.0 → 0.24.1
• Upgrade torchao 0.13.0 → 0.15.0, triton 3.4 → 3.5.1
• Bump Python 3.10 → 3.12 (flash-attn 2.8.3 wheels for torch 2.9 only support cp312)
• Add nvidia-cudnn-cu12 override (9.15+) to fix Conv3D bf16 performance regression
• Add automatic cuDNN patching for Windows via .pth file

Why

PyTorch 2.9.1 has a Conv3D bf16 performance regression with cuDNN < 9.15. On Windows, PyTorch bundles cuDNN in torch/lib and loads it by full path, ignoring pip packages. The .pth file automatically copies the newer cuDNN DLLs at Python startup.

The flash-attn 2.8.3 prebuilt wheels for torch 2.9 are only available for Python 3.12 (cp312), requiring the Python version bump.

Changes

• .python-version: 3.10.12 → 3.12.8
• pyproject.toml:
  • requires-python → >=3.12
  • Bump deps + add cuDNN override + force-include for .pth
  • Update wheel URLs for cp312 (flash-attn, sageattention)
  • ruff target-version → py312
• .github/workflows/lint.yml: Python 3.10 → 3.12
• src/scope/core/patches/cudnn.py: Use importlib.util.find_spec() to find package paths WITHOUT importing torch (prevents DLL locking)
• patches.pth: Installed to site-packages, runs at Python startup

Test plan

• Fresh uv sync installs all deps correctly
• cuDNN version reports 91701 (9.17.1) automatically on fresh install
• Pipeline FPS restored to ~10 (from ~7)
• pytest passes
• ruff check passes
• Linux CI passes
• Windows CI passes
• Linux benchmarking
• Windows benchmarking

🤖 Generated with Claude Code

[yondonfu]

Linux benchmarking

LongLive test script comparable throughput + latency vs. main

Windows benchmarking

LongLive, StreamDiffusionV2, MemFlow, Krea comparable throughput + latency vs main