MiaAI-Lab/DeepSeek-v4-Flash-DSpark-2x-DGX-Spark
DeepSeek-v4-Flash 0731 recipe for 2x DGX Sparks
Build with Backblaze B2
SDKs, agent skills, IDE extensions, and reference pipelines from Backblaze Labs. All open source.
Loading star history...
Where it stands
Rank, growth pace and release cadence relative to everything GStars tracks.
Ranks #350 of 2,845 repositories GStars tracks this week (top 12%).
The project lists 21 contributors on GitHub.
Use Cases & Benefits
- Implements a distributed recipe for running DeepSeek-v4-Flash Vision-Exp model on two DGX Spark nodes with DSpark speculative decoding and 1M-token context ceiling.
- Enables efficient large-context vision-language model serving with native image support and high concurrency using optimized KV cache and NCCL/RoCE networking.
- Use for deploying large-scale multi-modal AI inference clusters requiring long-context vision and text processing on NVIDIA DGX Spark hardware.
- Use for research and development of high-throughput, low-latency multi-node model serving with advanced token batching and speculative decoding.
- Use for integrating vision-expansion AI models with OpenAI-compatible APIs supporting image inputs and fine-grained reasoning token budgets.
Should you depend on it?
License terms, project hygiene and risks, from GitHub metadata and GStars' snapshots.
License
MIT
Permissive: commercial use, modification and redistribution are allowed with attribution.
Community files
- README: present
Watch-outs
- No GitHub releases; pin a commit or tag instead of a release.
General guidance from repository metadata, not legal advice. How we score
From the README
Two-node DGX Spark recipe for
deepseek-ai/DeepSeek-V4-Flash-Vision-Exp: vLLM TP=2, DSpark speculative decoding, 1M-token ceiling,nvfp4_ds_mlaKV. Native image support is a startup hotfix on the Anemll 0.1.1 runtime (ViT + Aligner from the Vision-Exp checkpoint, OpenAIimage_url/<image>path</image>). There is no video encoder in the official weights; GIF is a still frame. The old Qwen3-VL sidecar / MCP path is removed.Numbers: results/RESULTS-2026-08-14.md (dated tables, method, historical lanes). Checkpoint / encoder: docs/DEEPSEEK_V4_FLASH_0731.md.
Run everything from the head node. You need two DGX Sparks, RoCE/NCCL working, and the same image on both. By default each node has its own HuggingFace checkpoint (
preparecopies onto the worker). SetDSPARK_WORKER_HF_NFS=1to keep weights only on the head: the worker mounts that cache over NFSv4 on the ConnectX link (same pattern as Qwen3.8-Flash-vLLM).Set at least:
WORKER_HOST,MASTER_ADDR,NCCL_IB_HCA,NCCL_SOCKET_IFNAME(and matchingTP_/GLOO_IF names),VLLM_HOST_IP,WORKER_VLLM_HOST_IP,HF_CACHE.WORKER_HF_CACHEis the worker checkpoint path. WhenDSPARK_WORKER_HF_NFS=1, it is the worker's local JIT overlay on the NFS mount. If the worker checkout is not the same path, setWORKER_DIR/WORKER_SCRIPT_DIR.
Daily snapshots
Stars gained per day, and how forks and open issues moved, from GStars' daily snapshots
+186 stars over 37 days across 7 snapshot days; best day +36 on Sep 9.
Forks
197+19
Open issues
41+9
Compared with peers in Backend
Fastest-growing repositories GStars tracks in this cohort for the week of Oct 11, by 7-day star growth
| Repository | Stars | +7d | Health | Last release |
|---|---|---|---|---|
| NationalSecurityAgency/ghidra | 92.6K | +7.5K | 75 | 30 days ago |
| DuarteSantos8/openGym | 14.5K | +4.4K | 59 | 11 days ago |
| K-Dense-AI/scientific-agent-skills | 53.6K | +3.2K | 75 | 20 days ago |
| 666ghj/MiroFish | 80.7K | +2.4K | 23 | 226 days ago |
| anthropics/claude-plugins-official | 43.6K | +2.3K | 35 | — |
| MiaAI-Lab/DeepSeek-v4-Flash-DSpark-2x-DGX-Spark | 1.5K | +0 | 23 | — |
| Category median | 53.6K | +3.2K | 59 | — |
Also by MiaAI-Lab
Other repositories from this maintainer that GStars tracks
Discover Repositories
Search across tracked repositories by name or description