[Discussion] As vLLM-Omni is also complex, I tried building a minimal version (~1k LOC) inspired by nano-vllm

Inspired by nano-vLLM, I built nano-vLLM-omni — a minimal vLLM-Omni-style diffusion engine in ~1,079 lines of Python.

It keeps the core ideas (request / scheduler / runner / pipeline + step-wise diffusion execution), but strips things down to make the execution flow easier to follow and modify.

Currently supports Wan2.2-TI2V-5B on a single RTX 3090. In a simple benchmark, it runs ~9.1% faster than the official vllm-omni path (likely due to reduced abstraction overhead).

Repo: https://github.com/Rising0321/nano-vllm-omni

Curious how you think about the trade-off between architectural simplicity vs scalability in these systems — especially for single-GPU / local-first use cases.