Developers presented the recipe for DeepSeek-V4.1-Flash inference based on DGX Spark

The development team has released LuZ-0.1.7-DeepSeek-v4.1-Flash-DGXspark-TP4-Ring — a specialized recipe for deploying the DeepSeek-V4.1-Flash model on clusters of 4 NVIDIA DGX Spark nodes.

The solution is focused on maximum throughput using a switchless RoCE ring architecture. In testing, the system demonstrated a prefill speed of 3436 tokens per second. When processing 16 parallel queries, the aggregate decoding speed reached 460 tokens per second.

To achieve these results, the TP4 configuration (Tensor Parallelism 4) is used, allowing for efficient distribution of computational load between nodes. The entire preparation and launch process takes around 5 hours, making this stack suitable for rapid deployment of high-performance inference servers in production environments.

Optimizing network interaction through the RoCE ring eliminates the need for external switches, reducing latency when transferring data between GPUs. This approach enables NVIDIA DGX Spark hardware owners to significantly improve the efficiency of working with large language models without additional costs for network infrastructure.