Recipe by Tech2Wild tonyd2wild · based on DeepSeek, vLLM, FlashInfer, local-inference-lab (Jason Cook, Luke Alonso), bot-lab-21, ecohash-co, NVIDIA

License: MIT · DeepSeek: MIT · vLLM: Apache-2.0 · FlashInfer: Apache-2.0 · local-inference-lab (Jason Cook, Luke Alonso): Apache-2.0.

TESTED BY Tech2Wild on 4x DGX Spark (GB10), vLLM dsv41-feat TP4, native FP8/FP4 1017 weights, DSpark · their results (log or results file)

DeepSeek-V4.1-Flash-1017 on four DGX Sparks (vLLM)

ADVANCED: four DGX Sparks, Linux terminal, the author’s own launch scripts.

FORFour DGX Sparks on RoCE, with the stock 1017 weights.

NEEDThe author’s scripts name their own hosts (“Reddie”, “Asusi”). Adapt them to yours.

WHYA second, independent test of the same model on four Sparks: 36.1 tok/s prose, 76.3 code, one request.

NOTEThe repo’s default lane now uses a 3.5-bit file. We list only this lane, which runs the weights as shipped.

GET

git clone https://github.com/tonyd2wild/DeepSeek-V4.1-Flash-vLLM-DGX-Spark.git
cd DeepSeek-V4.1-Flash-vLLM-DGX-Spark/runs/2026-10-17-1017

THENFollow “How to run it” in that folder’s README: restore_1017.sh 1017g (the serving config), then postcheck1017.sh.

READYThe head logs “Application startup complete”. About 11 minutes.

CHECKpostcheck1017.sh: count, code, a tool call and an image.

SAFETYThe scripts run as root and stop other serving stacks on all four nodes. Read them first.

OR · LET A CODING AGENT DO IT

PASTEInto Claude Code, Codex or Cursor:

Help me run Tech2Wild’s vLLM lane for DeepSeek-V4.1-Flash-1017 on my four DGX Sparks: https://github.com/tonyd2wild/DeepSeek-V4.1-Flash-vLLM-DGX-Spark/tree/main/runs/2026-10-17-1017
Use only this native-weights lane (boot1017g), not the EXL3 or uncensored lanes.
1. Read the README and the scripts in launch/ and tools/. List every host name, path and root command they assume, and map them to my machines. Stop and show me the plan.
2. After I approve, run the gate (prelaunch-1017.sh), the launch, and postcheck1017.sh, and tell me the results.
Ask me before running anything as root or stopping other programs.

If it fails

The gate refuses to launch.
It lists why: another stack holds the GPUs, a node lacks the model or Engram copy, or a GPU tests slow. Fix that item first.

From the original README, shortened. Read it for full details.

What they tested

One request: prose 36.1 tok/s, code 76.3; 48.5 tok/s average over 8 categories; 165.7 total at 6 requests (bench JSON and markdown) · evidence
Head log, postcheck, args and env of the serving boot · evidence
Vision and tool checks 7/7; 69-scenario tool eval 98.6 · evidence

Not shown in their results: The launch scripts are written for the author’s own hosts. Longest tested prompt is 93,335 tokens (cold prefill), under our 128K bar for long documents.

RunCards does not test recipes. We only list recipes whose authors publish their own test results. Checked 2026-10-19.

Who made what

DeepSeek: the DeepSeek-V4.1-Flash-1017 model
vLLM: the engine and day-0 DeepSeek-V4.1 support (vllm-project/vllm#56214, Yongye Zhu and others)
FlashInfer: the SM120 sparse-MLA decode kernel
local-inference-lab (Jason Cook, Luke Alonso): the b12x RoCEnante one-shot RoCE collectives
bot-lab-21: the NCCL_MAX_NCHANNELS=8 measurement
ecohash-co: the expandable_segments finding
NVIDIA: the Connect Multiple Sparks playbook for cables and NVIDIA Sync

From the repo’s CREDITS.md and the run folder’s Credits section.