deepseek-v4.1IMAGE

4×DGX Spark 跑 V4.1 Flash：吞吐提升但 16 用户降 15%

在 4 台 DGX Spark 上用 TensorFold v0.3.5 与 4.0bpw 权重部署 V4.1 Flash，加两处补丁后整体变快，代价是 16 并发吞吐下降约 15%。

Tech2Wild

@Tech2Wild1 天前收录原推发布于 2026.10.27

4×DGX Spark 跑 V4.1 Flash：吞吐提升但 16 用户降 15%

Tech2Wild @Tech2Wild原推发布于 2026.10.27

Before vs now on our 4x DGX Spark DeepSeek V4.1 Flash fleet. Full credit to @bertholomusai for the TensorFold v0.3.5 recipe, engine and 4.0bpw weights. We added 2 small patches, both RoCE lanes and a 29.4M KV pool. One tradeoff: 16-user throughput is down ~15%.

x.com/tech2wild/status/2114866354612228293采集于 2026.10.27 12:08