ds4-on-spark
Entrpi/ds4, a Blackwell CUDA perf fork of antirez/ds4 on NVIDIA DGX Spark: one-command install, ~3x upstream prefill, ~1.5x decode, DSpark, and full continuous
What is ds4-on-spark?
Entrpi/ds4, a Blackwell CUDA perf fork of antirez/ds4 on NVIDIA DGX Spark: one-command install, ~3x upstream prefill, ~1.5x decode, DSpark, and full continuous batch support
Overview of ds4-on-spark
What it is
ds4-on-spark is a packaging repository for Entrpi/ds4, a Blackwell CUDA performance fork of antirez/ds4 (DwarfStar 4), pinned at release v0.6.5. It builds the upstream engine into a batched multi-request server and runs DeepSeek-V4-Flash entirely on-device on an NVIDIA DGX Spark (GB10, SM121, 128 GB unified memory, ~119 GiB usable). The sm_120 target (RTX PRO 6000 / 5090-class) also builds. Upstream remains the architectural foundation and the model recipe.
Key features
- Reported 2.4-3.3x upstream prefill throughput and 1.33-1.47x decode speed across the 2k-128k context range.
- Continuous batching, prefix caching with disk-persisted KV banks that survive restarts, and DSpark lossless speculative decode.
- A memory governor measured up to 3 million tokens of active context resident at once (2.26 million with shipped defaults); unused context is demand-mapped.
- Four APIs from one continuous batch: OpenAI chat completions and completions, OpenAI Responses, and Anthropic Messages, with streaming, tool calls, and thinking on the batched path.
- A one-command installer that verifies the host, clones the fork, builds ds4, ds4-server and ds4-bench for sm_121, downloads the ~81 GiB DeepSeek-V4-Flash-0731 Q2 GGUF and the ~6.5 GiB DSpark drafter, runs a smoke test, and installs the ds4-serve launcher.
- A default launch context of 512k, with the model's full 1M window available via -c 1048576.
Use cases
Serving DeepSeek-V4-Flash on a single DGX Spark for agent harnesses such as OpenAI Codex and Claude Code, deep document ingestion, and multi-request batched serving.
Getting started
bashcurl -sSL https://raw.githubusercontent.com/entrpi/ds4-on-spark/main/install.sh | bash -s -- --startds4-serveds4-serve -c 1048576
License and pricing
MIT. Open source; no pricing is stated.
Based on the project's public README and repository metadata.