Deepseek-V4-Flash-Vision-Exp FP4 SGLANG DSPARK 1Mctx

james.meadlock · September 3, 2026, 3:36am

Hopefully it can get a lot faster!

el8 · September 3, 2026, 12:51pm

You should join us on the Local Inference Lab Discord, lots of work happening there on the Station:

james.meadlock · September 3, 2026, 2:52pm

thx will check it out!

james.meadlock · September 4, 2026, 9:59pm

slight improvements… Recipe v2 from the DSFVE inner loop — J&M Labs

github.com/jmeadlock/dsf-vision-exp-gb300

RECIPE_V2.md

main

# DSFVE single-GB300 Recipe v2

Recipe v2 is the evidence-hardened default for serving `deepseek-ai/DeepSeek-V4-Flash-Vision-Exp` on one NVIDIA GB300 with SGLang and DSpark.

It is **not** a measured speed promotion. Iteration 13 measured equivalent C8-C64 throughput inside the frozen ±3% band. The improvement is operational: fewer ineffective runtime dependencies, explicit static verification, checkpoint-default NextN retained, immutable image/model pins, and the DSML encoding patch kept in the launch path.

## Default launch

```bash
docker pull lmsysorg/sglang@sha256:7ac467a50508b7029a23e846c150998fdd26d95c1cfd377ea7e74e28374486a6
python3 -m venv ~/hfenv && ~/hfenv/bin/pip install 'huggingface_hub[cli]'
~/hfenv/bin/hf download deepseek-ai/DeepSeek-V4-Flash-Vision-Exp \
  --revision 6821d6ad3681a4b137b066b76094fa82ebd0a380 \
  --local-dir ~/models/DeepSeek-V4-Flash-Vision-Exp/6821d6ad3681a4b137b066b76094fa82ebd0a380/original
mkdir -p ~/ds4f-vision-exp && cp -r patches ~/ds4f-vision-exp/
echo "your-key" > ~/.glm_api_key

./launch-dsfv.sh dspark
```

This file has been truncated. show original

Topic		Replies	Views
Deepseek 4.1 Flash DGX Station / GB300 / GB200 deepseek	1	742	September 18, 2026
DS V4 Flash 0731 Vision Exp, MiaAI Lab vs Eugr's receipe DGX Spark / GB10 deepseek	1	591	September 11, 2026
DeepSeek V4 Flash with Vision DGX Spark / GB10 deepseek	7	3047	August 10, 2026
DeepSeek-V4.1-Flash 552B MoE on 4x DGX Spark TP4 - 77.2 tok/s c1 on PEAK 52 code · 72 tok/s on a warm code run · 47 math · 39 reasoning · 23 prose DGX Spark / GB10 Projects deepseek	5	2136	September 12, 2026
Let’s optimize DeepSeek-V4.1-Flash-1017 for 2× DGX Spark DGX Spark / GB10 cuda , llama , deepseek	7	1292	November 4, 2026
New DeepSeek-V4-Flash-DSpark DGX Spark / GB10 deepseek	5	5329	June 29, 2026
DeepSeek v4 Flash Vision Exp is Released as Open Weights DGX Spark / GB10 deepseek	452	21235	October 11, 2026
Official NVidia optimized DeepSeek-V4-Flash models? DGX Spark / GB10 deepseek	28	4030	July 11, 2026
DeepSeek-V4-Flash-Vision-Exp on one DGX Spark: text + images, 245k context, CUDA graphs — plus a pack that keeps 90 % of the original model DGX Spark / GB10 deepseek	20	2271	November 3, 2026
C1 1058pp/s, 52 tg/s on 1x dgx spark on Deepseek V4 flash 0731 full 256 experts DGX Spark / GB10 Projects deepseek	94	7745	September 5, 2026

Deepseek-V4-Flash-Vision-Exp FP4 SGLANG DSPARK 1Mctx

Related topics