vLLM compile deep dive (it's all torch.compile)

​Join us for a focused deep dive into compiling vLLM models with torch.compile. We'll walk through why torch.compile matters for large-model performance and how vLLM leverages it, demonstrate concrete compilation strategies and flags, and profile real-world gains on inference workloads.

​This session is best for people comfortable with vLLM or PyTorch and basic LLM concepts who want actionable techniques to speed up inference. This is mostly a theoretical session and we are open for discussion afterwards; questions and specific pain points are welcome. Meet at NS library on March 9.