SGLang handles higher concurrency better than vLLM on 4x RTX 5060 Ti

A user benchmarking Qwen3.6 27B (INT8) on four RTX 5060 Ti GPUs found that SGLang manages higher concurrency more effectively than vLLM, which previously suffered from time-to-first-token (TTFT) issues.

• The setup achieved a mean end-to-end latency of 13679.23 ms and a mean TTFT of 7504.07 ms at 8 concurrent requests.
• Output token throughput reached 127.63 tok/s, with a peak output throughput of 76.00 tok/s.
• Configuration required disabling multimodal features, using speculative decoding (NEXTN), and applying a PCIe P2P driver patch via Proxmox VM passthrough.

This configuration provides a working reference for users planning to deploy large models on consumer-grade hardware with limited NVLink bandwidth.