[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-94434":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":10,"language":11,"languages":10,"totalLinesOfCode":10,"stars":12,"forks":13,"watchers":14,"openIssues":15,"contributorsCount":15,"subscribersCount":15,"size":15,"stars1d":16,"stars7d":17,"stars30d":18,"stars90d":15,"forks30d":15,"starsTrendScore":19,"compositeScore":20,"rankGlobal":10,"rankLanguage":10,"license":10,"archived":21,"fork":21,"defaultBranch":22,"hasWiki":23,"hasPages":21,"topics":24,"createdAt":10,"pushedAt":10,"updatedAt":25,"readmeContent":26,"aiSummary":27,"trendingCount":15,"starSnapshotCount":15,"syncStatus":28,"lastSyncTime":29,"discoverSource":30},94434,"DeepSeek-v4-Flash-One-DGX-Spark","MiaAI-Lab\u002FDeepSeek-v4-Flash-One-DGX-Spark","MiaAI-Lab","DeepSeek-v4-Flash-One-DGX-Spark - DwarfStar 4 Engine","https:\u002F\u002Fx.com\u002FMiaAI_lab",null,"Shell",175,19,1,0,39,61,63,139,87.14,false,"main",true,[],"2026-08-24 04:01:22","# DeepSeek-v4-Flash-One-DGX-Spark - DwarfStar 4 Engine\n\n\u003Cp align=\"center\">\n  \u003Ca href=\"https:\u002F\u002Fx.com\u002FMiaAI_lab\" target=\"_blank\" rel=\"noopener noreferrer\" style=\"display:inline-block;margin:0 8px;vertical-align:middle;\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FFollow%20me%20on%20X-000000?style=for-the-badge&logo=x&logoColor=white\" alt=\"Follow Mia on X\" height=\"28\" style=\"height:28px;width:auto;vertical-align:middle;border:0;\" \u002F>\u003C\u002Fa>\n\u003C\u002Fp>\n\nThin, idempotent launcher scripts for running the **DeepSeek-V4-Flash** server built for the NVIDIA DGX Spark (GB10 \u002F SM121) — the DwarfStar 4 (C\u002FCUDA) engine that serves an OpenAI-compatible `\u002Fv1` API on `:8888`.\n\nThis is based on **antirez\u002Fds4** (DwarfStar 4) and its DGX Spark fork:\n\n- [antirez\u002Fds4](https:\u002F\u002Fgithub.com\u002Fantirez\u002Fds4) — upstream DwarfStar 4 engine (MIT-licensed, C\u002FCUDA)\n- [Entrpi\u002Fds4-on-spark](https:\u002F\u002Fgithub.com\u002FEntrpi\u002Fds4-on-spark) — DGX Spark one-command install, benchmarks, and roofline analysis (what `start.sh` pulls)\n- [Entrpi\u002Fds4 (batched-serving)](https:\u002F\u002Fgithub.com\u002FEntrpi\u002Fds4\u002Ftree\u002Fbatched-serving) — the DGX-Spark-optimized CUDA perf fork used here\n\n> **Note:** this repo does **not** use vLLM. `ds4-server` exposes the same `\u002Fv1` API that `vllm serve` does, but vLLM cannot read this repo's asymmetric GGUF, so the repo ships its own server. These scripts are just thin wrappers over that server's official installer.\n\n## Requirements\n\n- A NVIDIA DGX Spark (GB10 \u002F SM121)\n- Bash\n- `curl`\n- Disk space for the ~110 GiB GGUF weight set\n- `ss` (for `stop.sh`)\n\n## Background\n\n**DwarfStar 4** is a small, self-contained native inference engine optimized for DeepSeek V4 Flash, written by Salvatore Sanfilippo ([antirez](https:\u002F\u002Fx.com\u002Fantirez)) — deliberately narrow, not a generic GGUF runner. [Entrpi](https:\u002F\u002Fgithub.com\u002FEntrpi) maintains a DGX-Spark-optimized CUDA perf fork of it, plus the [ds4-on-spark](https:\u002F\u002Fgithub.com\u002FEntrpi\u002Fds4-on-spark) installer this repo wraps, which serves DeepSeek-V4-Flash entirely on-device on a GB10 \u002F SM121 DGX Spark (RTX PRO 6000 \u002F 5090-class `sm_120` also builds).\n\nThanks to Bleys Goodson ([@bleysg on X](https:\u002F\u002Fx.com\u002Fbleysg)).\n\n## Quick start\n\n```bash\n.\u002Fstart.sh    # full DSpark stack on :8888\n```\n\nFirst run does the heavy lifting: clones and builds the pinned fork, downloads the ~110 GiB GGUF set, smoke-tests, installs `ds4-serve`, and starts the server on `:8888`. Later runs fast-forward the clone to the pinned tag, skip GGUFs already on disk, and just start the server.\n\n## Usage\n\n```bash\nPORT=8889 .\u002Fstart.sh            # different port\nCTX=262144 .\u002Fstart.sh           # smaller context budget  (KV ≈ 9.5 KiB\u002Ftoken)\n.\u002Fstart.sh --no-dspark          # plain continuous decode (passes through)\n```\n\nEnvironment variables (all optional):\n\n| Variable       | Default    | Meaning                                   |\n|----------------|------------|-------------------------------------------|\n| `PORT`         | `8888`     | Server port                               |\n| `CTX`          | `262144`   | Context budget (KV ≈ 9.5 KiB\u002Ftoken)       |\n| `DS4_SRC_DIR`  | `~\u002Fcode\u002Fds4` | Source directory for the pinned clone   |\n| `DS4_GGUF_DIR` | `~\u002Fgguf`     | Weights directory                       |\n\n## Stopping and restarting\n\n```bash\n.\u002Fstop.sh                     # stop server on :8888 (wait until port is freed)\nPORT=8889 .\u002Fstop.sh           # stop server on a different port\n\npkill -x ds4-server; .\u002Fstart.sh   # force a clean restart\n```\n\n`start.sh` is idempotent: if the server is already answering on the port, it reports the running model and exits without touching anything.\n\n## Checking status\n\n```bash\ncurl http:\u002F\u002F127.0.0.1:8888\u002Fv1\u002Fmodels\n```\n\n## Performance\n\n![Performance on a single NVIDIA DGX Spark](bench.jpg)\n\nMeasured on a single NVIDIA DGX Spark (GB10 \u002F SM121).\n\n## Logs\n\nServer logs go to `~\u002Fds4-server.log`. Check there if the server doesn't come up:\n\n```\n!!! Not reachable yet — check $HOME\u002Fds4-server.log\n```\n\n## Files\n\n| File         | Purpose                                     |\n|--------------|---------------------------------------------|\n| `start.sh`   | Fetch installer, build, download weights, serve on `:8888` |\n| `stop.sh`    | Stop the `ds4-server` process              |\n| `bench.jpg`  | Decode throughput benchmark (tok\u002Fs vs context) on a single DGX Spark |\n","这是一个为NVIDIA DGX Spark（GB10\u002FSM121）硬件定制的轻量级启动脚本集合，用于部署DeepSeek-V4-Flash模型的原生推理服务。项目基于DwarfStar 4（C\u002FCUDA实现）引擎，提供OpenAI兼容的\u002Fv1 API接口，支持端到端本地部署、端口与上下文长度灵活配置，并内置GGUF权重下载、构建缓存与服务启停管理。不依赖vLLM等通用推理框架，专为DGX Spark平台优化CUDA性能与内存带宽利用。适用于需要在单台DGX Spark设备上快速部署低延迟、高吞吐DeepSeek-V4-Flash服务的AI基础设施场景。",2,"2026-08-08 02:30:12","CREATED_QUERY"]