Server 1

Last night's deploy raised max_model_len (8K→16K) and max_num_seqs (12→24) — that roughly 4×'d the KV cache. GPU memory peaked at 94% and first-token latency is up 2.3×.