Entrpi / ds4-on-spark

What is ds4-on-spark?

This is a deployment toolkit for running DeepSeek-V4-Flash on NVIDIA DGX Spark hardware. It wraps antirez/ds4 with shell scripts that handle installation, benchmarking, and performance analysis on GB10 GPUs. A single curl command clones the engine, builds the CUDA binaries, downloads the quantized model weights, and runs a smoke test. The project includes detailed throughput measurements across different context lengths and a roofline analysis showing how close the inference engine runs to the hardware bandwidth ceiling.

Why is it gaining traction?

The hook is the roofline analysis. Most inference benchmarks report "tokens per second" without explaining why the numbers are what they are. This project measures actual memory bandwidth on the Spark, breaks down bytes-per-token by quantization bucket, and shows that steady-state decode hits roughly 95% of the bandwidth roofline. That kind of analysis is rare and useful for anyone trying to understand whether further optimization is even possible. The one-command install also lowers the barrier for reproducing the numbers.

Who should use this?

ML engineers evaluating DeepSeek-V4-Flash on Blackwell hardware will find the benchmarks directly applicable. Researchers studying MoE inference efficiency will value the roofline decomposition. Anyone running ds4 on a DGX Spark and wanting reproducible performance numbers should start here. If you need concurrent request handling, look elsewhere—the server serializes clients by design.

Verdict

The technical content is solid and the analysis is genuinely useful, but the repository has 14 stars and a credibility score of 0.699999988079071%, signaling early-stage work. Use it as a reference for benchmarking methodology and hardware-specific performance expectations, but treat the scripts as a starting point rather than production-ready infrastructure.