https://oshea00.github.io/tags/spark/ Recent content in Spark on MikesBlog Hugo -- 0.165.0 en-us Thu, 13 Aug 2026 00:00:00 +0000 https://oshea00.github.io/posts/running-cool-dgx-cluster/ Thu, 13 Aug 2026 00:00:00 +0000 https://oshea00.github.io/posts/running-cool-dgx-cluster/ <figure class="align-center "> <img loading="lazy" src="images/cutebutoverheated.png#center"/> </figure> <p>I recently ran into an interesting thermal issue while running two NVIDIA DGX Sparks as a small distributed inference cluster using llama.cpp.</p> <p>The two Sparks are directly connected over QSFP and use RDMA as the transport. With <code>llama-server</code> running on Spark1, and <code>ggml-rpc-server</code> running on Spark2, I&rsquo;m using llama.cpp&rsquo;s layer-split configuration with the <code>DeepSeek-V4-Flash-0731-UD-Q4_K_XL</code> model. The model is too large to run on a single Spark, thus the need to layer it over two Sparks.</p>