vLLM vs SGLang vs TensorRT-LLM for production LLM inference throughput

SGLang claims 29% higher throughput than vLLM on H100s via RadixAttention (16,200 vs 12,500 tokens/sec). TensorRT-LLM leads at high concurrency (13% faster than vLLM at 50 concurrent requests). But vLLM has 3x more contributors and broader hardware support. Which engine actually wins for production workloads? Source: https://news.ycombinator.com/item?id=46649975