[Blog] Optimizing DeepSeek V4.1 Flash on 8× H20: EP8 MoE Dispatch and vLLM Serving Performance

[luoyuctl]

Summary

• Add an 8×H20 deployment case study tracing the MXFP4/Humming W4A8 path through FP8 conversion, Engram offload, and routed-row-aware MoE dispatch.
• Report separate isolated operator results (up to 3.24×) and three-run whole-recipe serving comparisons against Humming: Decode c256 4,725 → 6,459 output tok/s; fixed-1K Prefill c64 9,875 → 15,048 input tok/s. These gains have different scopes and must not be multiplied.
• Include figures, per-run data, paired 500-item GSM8K sanity-check results, methods, and environment/configuration details.

Related work and duplication check

No open PR in this blog repository was found for this DeepSeek V4.1 Flash on H20 case study. This is a blog post, not a duplicate of FlashInfer PR #5560 or vLLM issue #58799.

Validation

• Production Jekyll build: passed.
• node --check assets/repro/deepseek-v41-h20/quality_eval.mjs: passed.
• Referenced article assets and JSON/JSONL parsing: passed.
• git diff --check: passed.
• Reported serving results are means of three runs per concurrency tier; paired GSM8K numeric accuracy was 476/494 (Humming) versus 475/494 (tuned).

AI assistance was used to draft and organize the article and reproduction materials; the human submitter reviewed the changes and reported measurements.

[luoyuctl]

@ywang96 @mgoin @esmeetu Could you review this 8×H20 DeepSeek V4.1 Flash serving case study when you have time? It includes a three-run Humming W4A8 vs FP8 recipe comparison, an isolated EP8 MoE operator ablation, and the reproduction data. I would especially appreciate feedback on the benchmark methodology and how the operator and whole-recipe results are attributed. Thank you!

[luoyuctl]

@NickLucche Would you have time to review this post? It is a serving case study for DeepSeek-V4.1-Flash on 8×H20 (three-run recipe comparison plus an isolated EP8 MoE dispatch ablation). The post is self-contained; the main thing I'd like a second opinion on is whether the whole-recipe and operator-only results are attributed clearly enough. Happy to shorten it if that helps.