Why it is worth attention

It documents and packages a specific optimized vLLM build for DGX Spark's SM121, with measured throughput results and a full evidence trail, making it useful for reproducible high-performance local serving of this model on dual-GB10 hardware.

Who it is for

Developers deploying DeepSeek models on NVIDIA DGX Spark/GB10
vLLM users needing SM121-specific container images
ML infrastructure engineers evaluating optimized MoE/MLA serving
Reproducibility-focused researchers comparing vLLM builds

Use cases

Reproduce the reported c1/c16 throughput benchmarks on dual-GB10 hardware
Launch DeepSeek-V4-Flash-DSpark serving via the provided container and profiles
Use long-context or production decode profiles for deployment
Inspect raw benchmark outputs and sealed evidence for validation

Strengths

Clear performance figures: c1 decode median 80.58 tok/s and c16 aggregate 402.0 tok/s
Includes raw benchmark harness outputs, scripts, profiles, and sealed metrics JSON
Provides a container image with pinned vLLM version and runtime identity
Covers long-request correctness and staggered concurrency results

Considerations

Not an official DeepSeek/vLLM release and model weights are not redistributed
Results are specific to a dual-GB10 software pin set; no guarantees on other hardware or revisions
GSM8K check is only a 50-sample smoke test, not a full evaluation

README quick start

DeepSeek-V4-Flash-DSpark · vLLM 0.26 SM121 (r0b0tlab)

Full reproducibility package + optimized container for dual NVIDIA GB10 (DGX Spark / SM121) serving of DeepSeek-V4-Flash-DSpark on vLLM 0.26.0+dspark.sm121.2.

Maintained by r0b0tlab.
Not an official DeepSeek or vLLM release. Model weights are not redistributed.

Headline results

Metric	Result	Prior floor
Dedicated c1 decode median (2048 tok × 5)	80.58 tok/s	77.76 (v0.25 K6)
Static c16 aggregate best-of-2	402.0 tok/s	342.70 (v0.25)
Exact ~300K long request	QUALIFIED	—
Staggered c16	16/16	—
GSM8K smoke N=50	50/50	smoke only

Quick links

Artifact	Path
HTML report	docs/report.html
Claims	docs/CLAIMS.md
Reproducibility	docs/REPRODUCIBILITY.md
Container pull	CONTAINER.md
Metrics JSON	evidence/publication-summary.json
Thin-image recipe	recipe/
Launch scripts	scripts/
Profiles	profiles/
Benchmarks	benchmarks/

Container

docker pull ghcr.io/r0b0tlab/deepseek-v4-flash-dspark-v026-sm121:v0.26.0-sm121-optimized
# digest sha256:269a1c6ac581f4fb2fe13110d2468159c3ca65c07ffb09b4080549673f625841

Local build ID: sha256:f904b9445aa407420b26f70ee5c2075c7d04ef0fdd38a728b20985ee331303a6
Details: CONTAINER.md

Runtime identity

vllm              0.26.0+dspark.sm121.2
kv_cache_dtype    nvfp4_ds_mla
moe_backend       flashinfer_b12x
mtp               dspark K=6
max_model_len     327680
model_revision    913f0657a874f76844e2e91cbe706dbcaceeb6d7
upstream_vllm     568afb3a13806beb53bb2e6bd518269357b237c0

Profiles (do not mix rows)

Production decode — max_num_seqs=16, max_num_batched_tokens=16384, K=6
C1 long-context — max_num_seqs=1, max_num_batched_tokens=8192, K=6

Benchmark results

Raw harness outputs live under evidence/benchmarks/ (c1 decode JSON/logs, concurrent ladder, staggered c16, correctness, GSM8K smoke).

What's included for reproducibility

Knownfix thin-image Dock

Description

DeepSeek-V4-Flash-DSpark optimized vLLM 0.26.0 SM121 dual-GB10 evidence (NVFP4 KV, B12X, DSpark K6)

Related repositories

Similar projects matched by category, topics, and programming language.

lopopolo

Featured

harness-engineering

Harness Engineering is a methodology for improving coding agent outputs by carefully crafting the environment around them—providing curated context, tools, and executable constraints that encode an organization’s nonfunctional requirements and cumulative lessons.

AI & Machine LearningAI Agents

2,390

slvDev

Featured

esp32-ai

A 28.9 million parameter language model runs on an $8 ESP32-S3 microcontroller entirely on-device, generating simple stories at about 9.5 tokens per second.

AI & Machine LearningLarge Language Models

1,960

littledivy

Featured

mimic

mimic captures traffic from any iOS or web app and automatically generates a Python client library that lets you call the app's API like a regular library.

AI & Machine Learning

1,482