SGLang 公布 V4.1 Flash 推理跑分：prefill 最高 1.04×，decode 步长缩短 7.4%
SGLang 在 4 张 B300/GB300 上 prefill 吞吐达 DeepSeek 自研栈的 1.04×，64 张 GB300 上 decode 步长缩短最多 7.4%，vLLM 也追平其 decode 曲线。
🔔 The Daily Semi-Cap — October 24, 2026 📊 SMH: $527.20 (last market close: October 23, 2026; +0.74% vs October 22 close) • Open inference engines are catching DeepSeek’s in-house stack. SGLang’s new DeepSeek-V4.1-Flash benchmarks show up to 1.04× prefill throughput on four B300/GB300 GPUs and up to 7.4% shorter decode steps on 64 GB300s. vLLM also matched DeepInfer’s decode curves. These are specific benchmark configurations, not universal production gains, but faster serving can improve the economics of installed accelerators without waiting for new silicon. • Datacenter opposition cuts across party lines: 63% of voters oppose an AI facility near home, versus 13% supporting one. Even among MAGA supporters, 40% oppose a facility within 10 miles, against 29% supporting it. The proximity question covered roughly 500 voters. For infrastructure suppliers, this highlights a project-timing risk: enthusiasm for AI does not automatically translate into local permission to build its power-hungry physical footprint.