Kompas

SGLang + DFlash on DGX Spark (Qwen3-Coder-Next NVFP4)

Running Qwen3-Coder-Next-NVFP4-GB10 with DFlash speculative decoding on SGLang. Tested on a Lenovo ThinkStation PGX (NVIDIA GB10 Grace Blackwell, 128 GB unified memory).

What you get

Test	SGLang DFlash	vLLM DFlash+Marlin	Delta
Short code (307 tok)	150 tok/s	108 tok/s	+38%

Kompas

Qwen3 Coder Next + DFlash on DGX Spark: 108 tok/s with a 2 line vLLM patch

I got DFlash speculative decoding working with Qwen3 Coder Next on my DGX Spark (GB10, 128 GB unified memory). Result: 88-108 tok/s depending on task complexity, up from 62 tok/s without DFlash.

Tool calling works too (--enable-auto-tool-choice --tool-call-parser qwen3_coder), tested at 89 tok/s with DFlash active. Useful if you're running coding agents that need function calls.

The fix turned out to be surprisingly simple: two lines of Python.

The Problem

Kompas

(() => {
  function formatDate(date = new Date()) {
    return date.toISOString().split("T")[0];
  }

  function escapeMarkdown(text) {
    return text
      .replace(/\\/g, "\\\\")
      .replace(/\*/g, "\\*")
      .replace(/_/g, "\\_")