它针对 DGX Spark 的 SM121 架构打包了特定的优化 vLLM 构建，并给出实测吞吐结果和完整证据链，适合需要在双 GB10 硬件上可复现地高效部署该模型的用户。
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.
| 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 |
| 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/ |
docker pull ghcr.io/r0b0tlab/deepseek-v4-flash-dspark-v026-sm121:v0.26.0-sm121-optimized
# digest sha256:269a1c6ac581f4fb2fe13110d2468159c3ca65c07ffb09b4080549673f625841
sha256:f904b9445aa407420b26f70ee5c2075c7d04ef0fdd38a728b20985ee331303a6vllm 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
max_num_seqs=16, max_num_batched_tokens=16384, K=6max_num_seqs=1, max_num_batched_tokens=8192, K=6Raw harness outputs live under evidence/benchmarks/ (c1 decode JSON/logs, concurrent ladder, staggered c16, correctness, GSM8K smoke).
DeepSeek-V4-Flash-DSpark optimized vLLM 0.26.0 SM121 dual-GB10 evidence (NVFP4 KV, B12X, DSpark K6)
根据分类、Topic 和编程语言匹配的相似项目。
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.
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.
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.