---
type: "post"
id: "full-engineering-writeup-covers-dep2-prefill-wide-ep-b2079d"
slug: "full-engineering-writeup-covers-dep2-prefill-wide-ep-b2079d"
title: "Full engineering writeup covers DEP2 prefill, Wide-EP decode, and CPU overhead fixes."
description: "vLLM published an engineering write-up on serving DeepSeek-V4.1-Flash with prefill/decode disaggregation and Wide-EP decode across an NVIDIA GB300 NVL72 rack."
retrieval_nugget: "vLLM published an engineering write-up on serving DeepSeek-V4.1-Flash with prefill/decode disaggregation and Wide-EP decode across an NVIDIA GB300 NVL72 rack."
published_at: "2026-10-24"
updated_at: "2026-10-24"
record_date: "2026-10-24"
date_kind: "published_at"
topics: ["harness-engineering","agent-runtime"]
entities: ["vLLM","DeepSeek-V4.1-Flash","DeepInfer","NVIDIA GB300 NVL72","SemiAnalysis AgentX"]
related_patterns: ["harness-architecture-outlives-model-choice"]
novelty_label: "new_fact"
what_changed: "The writeup adds rack-scale GB300 results, DEP2 prefill and Wide-EP/EPLB scaling details, CPU-overhead fixes, and a 826 tokens/s single-user decode path."
prior_refs: ["memory:56f8e800-0832-87eb-a8d5-c2519d7104fa","memory:e5f1ce5e-39ce-8087-ab11-b825e426defd","memory:eb27678d-1635-8fbf-a995-67cecf7d972b"]
source_url: "https://vllm.ai/blog/2026-10-23-deepseek-v41-flash-pd-wide-ep"
source_urls: ["https://vllm.ai/blog/2026-10-23-deepseek-v41-flash-pd-wide-ep"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:full-engineering-writeup-covers-dep2-prefill-wide-ep-b2079d"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/full-engineering-writeup-covers-dep2-prefill-wide-ep-b2079d/","markdown":"https://newruntime.com/posts/full-engineering-writeup-covers-dep2-prefill-wide-ep-b2079d.md","json":"https://newruntime.com/posts/full-engineering-writeup-covers-dep2-prefill-wide-ep-b2079d.json"}
---

# Full engineering writeup covers DEP2 prefill, Wide-EP decode, and CPU overhead fixes.

## Retrieval answer

vLLM published an engineering write-up on serving DeepSeek-V4.1-Flash with prefill/decode disaggregation and Wide-EP decode across an NVIDIA GB300 NVL72 rack.

vLLM published an engineering write-up on serving DeepSeek-V4.1-Flash with prefill/decode disaggregation and Wide-EP decode across an NVIDIA GB300 NVL72 rack.

The team reports 954,000 tokens per second per GPU on the SemiAnalysis AgentX benchmark, with P90 per-user throughput of 196 tokens per second. It describes this as a 7.9x increase over its earlier aggregated single-node result, and cites DeepSeek off-peak pricing of $38.96 per GPU-hour. A specialized small-batch decode path reached 826 tokens per second for one user.

For comparison, vLLM says its GB200 decode results were within 6% of DeepInfer's published throughput across batches of 8 to 256 at 131K context. Its prefill comparison uses different hardware and configurations, so the post says it does not treat the difference as an engine-to-engine speedup.

The post attributes the work to DEP2 prefill, Wide-EP decode, expert parallel load balancing, and reductions in CPU overhead. It does not establish equivalent performance for other models, workloads, hardware, pricing tiers, or production deployments.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Harness Engineering](https://newruntime.com/topics/harness-engineering.md): Explore the harness-engineering topic hub.
- topic hub: [Agent Runtime](https://newruntime.com/topics/agent-runtime.md): Explore the agent-runtime topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)