DGX Spark + Qwen27B 60tps! sglang+Dflash2 is over now.

61.252.***.***
29

Reddit 60tps

https://www.reddit.com/r/LocalLLM/comments/1vvtquz/dgx_spark_qwen_38_27b_nvfp4_at_60toks_generation/

The version I'm using is this

https://forums.developer.nvidia.com/t/qwen3-8-27b-nvfp4-on-single-dual-dgx-spark-sglang-dflash2-fully-openai-compatible/380732

Workload (single stream)

1x Spark

2x Spark (TP=2)

Code Generation

52~61 tokens/sec

87 tokens/sec

Prose/Essay

26 tokens/sec

41 tokens/sec

Thought Conversation

34~49 tokens/sec

49 tokens/sec

Thinking about code, ("none")

52 tokens/sec

approximately 80 tokens/sec

TTFT, short prompt

approximately 0.16s

Similar

TTFT, 16K prefix repeat

0.44s

0.74s

Context Window

262,144

262,144

I'm using more conservative settings (for quant, embedding, using 8B models together, etc.) and actual measurements show 36~46 tps

Flash 2+ SLang is really .. incredibly fast. What a technological advancement.

Today's tip: Qwen38 has a compact bug

  • When compressing a 150K token session with Qwen3.8 27B in DeepSeek Harness, a problem was reported where auto compact did not finish, almost the same as before.

    • Cause: The compressor inherits the reasoning setting of the original conversation and consumes the output budget during the process.

    • Solution: Request a summary only with reasoningEffort: off

    • Remove old reasoning, remove images, summarize large tool results.

    • This method was reported to have stabilized the compression of 128K·150K·256K sessions.
      DeepSeek Harness Qwen3.8 compaction case

  • In Qwen Code, there was also a problem reported where auxiliary requests such as title, summary, and recap inherited the main model settings and unnecessarily engaged in thinking. The solution is "each auxiliary request should have its own model, listening, and extra body settings."
    Qwen Code side-query setting inheritance problem

  • Another Qwen Code issue found that the general OpenAI request structure did not include Qwen-specific enable_thinking, so the OFF setting was not actually being sent to the server. In other words, simply displaying OFF in the UI is not enough; the provider adapter must explicitly insert it into the request body.
    Qwen Code enable_thinking delivery defect

The following was added to Vector Build to resolve the issue:

{

"chat_template_kwargs": {

"enable_thinking": false,

"preserve_thinking": false

}

}

로그인한 회원만 댓글 등록이 가능합니다.

개발한당

KR | ID | EN
  • IDR
  • KOR
7.86 =0.00

2026.08.23 KEB 하나은행 고시회차 1420회

다가오는 한인 행사일정

  • 등록 된 일정이 없어요!