Establish cicd workflow to build and publish FastVideo and STA Kernel

[PorridgeSwim]

Automated PyPI Publishing Workflows

This PR introduces automatic PyPI publishing for FastVideo and Sliding Tile Attention through GitHub Actions workflows. The workflows are triggered when version numbers are changed in their respective configuration files.

Key Features

• Version-Based Triggers: Packages are only published when their version numbers are intentionally updated
• Automated Build Process: Handles all build steps automatically without manual intervention
• Secure Deployment: Uses PyPI's OIDC Trusted Publishing method for secure authentication
• Compatibility Support:
  • Supports multiple Python versions (3.10-3.13)
  • Compatible with various PyTorch and CUDA versions
  • Generates appropriately named packages with compatibility indicators

Benefits

• Streamlines release process by eliminating manual publishing steps
• Reduces possibility of human error during package deployment
• Ensures consistent versioning and release practices
• Improves security by eliminating the need for stored API tokens

Successful Test Runs

The workflows have been tested successfully:

• FastVideo workflow: https://github.com/hao-ai-lab/FastVideo/actions/runs/13622134203
• Sliding Tile Attention workflow: https://github.com/hao-ai-lab/FastVideo/actions/runs/13622134185

These workflows will save development time and standardize our package release process across projects.

[PorridgeSwim]

Addressing Test Failures

Code Style Issues

Will update code to pass YAPF tests.

Package Dependency Error

ERROR: Could not find a version that satisfies the requirement st_attn>=0.0.1 (from fastvideo) (from versions: 0.0.1.dev1, 0.0.1.dev2)

Only development versions uploaded to PyPI; stable version required.

CUDA Requirement Issue

st_attn kernel needs CUDA for compilation. GitHub workflow (.github/workflows/test.yml) lacks CUDA configuration.

Action Plan

1. Fix YAPF style issues
2. Add CUDA configuration to test.yml
3. Upload stable package version to PyPI