vLLM Pascal fork revives Tesla P40

This is exactly the kind of open-source infrastructure work that keeps local AI practical outside the latest GPU generation.

• –The repo includes a concrete build path for Python 3.12, CUDA 12.1, and PyTorch 2.5.1, which makes it more than a brag post
• –Pascal support has been a long-running gap around vLLM, so a working fork for `sm_61` hardware solves a real pain point for budget homelab users
• –Real-time Qwen3-ASR on a Tesla P40 is more interesting than the card itself because it lowers the cost floor for local transcription systems
• –The bigger caveat is maintainability: this is an unofficial fork, and the author already flags vision and newer Qwen3.5-style workloads as much harder