Qwen3.8-Flash-Next 单台 DGX Spark 部署配方更新
该配方更新了单台 DGX Spark 上 262K 上下文的吞吐数据、NVFP4 官方权重支持、24/7 守护重启、缓存 token 上报与密钥安全等改进。
Big update to the @Alibaba_Qwen Qwen3.8-Flash-Next single DGX Spark recipe! 𝗪𝗵𝗮𝘁'𝘀 𝗻𝗲𝘄 🎁 • Measured on one DGX Spark, 262K context, MTP k=3, aggregate tok/s at 1 / 2 / 4 / 8 streams: Prose: 38.0 / 61.1 / 89.2 / 117.4 Code: 53.8 / 87.5 / 131.8 / 180.2 • Long context holds: MTP keeps working at a 185K-token prompt (35.4 tok/s decode), prefill ~2,000 tok/s from 4K to 185K • ~1M-token KV pool at the full 262K context (FP8 KV) • NVIDIA's official NVFP4 checkpoint now runs on one Spark, with chat, tool calls and vision working • 24/7 mode: a supervisor restarts the server on a crash, detects a hung engine with a real generation probe, and stops retrying after repeated failures • Cached-token reporting in every response • Optional API key, bind address and a fixed chat template for agent clients 𝗠𝗼𝗿𝗲 𝘀𝘁𝗮𝗯𝗹𝗲 ⚖️ • Measured memory reserve table per checkpoint and concurrency • NVIDIA checkpoint peak memory down from 101 to 91 GiB • The first-launch kernel compile can no longer run the box out of memory 𝗦𝗮𝗳𝗲𝗿 🛟 • Your HF token is no longer written to disk • Downloads are sha256-checked file by file, even on re-runs, and the script tells you exactly which file to fetch again 12+ bugs found and fixed. Big thanks to malvavisc0 for going back and forth with me on PR 41 addressing the issues I flagged 🙌