4 coding agents on one RTX 3090: ~290 tok/s total from a custom CUDA engine for Qwen3.8-27B (3-4x llama-server --parallel) · AI Intel