Lab

Open experiments, micro-drops, and engineering notes. Our lab is where we explore, prototype, and share what we're learning.

Filter#llm inference #pytorch #representation engineering #vllm

experimentMay 18, 2026

Per-request activation steering for Gemma 4 31B with vLLM

Training contrastive residual-stream vectors for Gemma 4 31B and applying them inside compiled vLLM inference without recompilation.

#representation engineering#llm inference#pytorch+1