<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title><![CDATA[DGX Spark - MSI Edgexpert]]></title><description><![CDATA[<p dir="auto">DGX Spark - MSI Edgexpert</p>
<ul>
<li>打算使用单机跑 Qwen3.8。可以参考配置如下：</li>
</ul>
<p dir="auto"><a href="https://lmsysorg.mintlify.app/cookbook/autoregressive/Qwen/Qwen3.8-27B#hw=dgx-spark&amp;variant=default&amp;quant=nvfp4-fp4-head&amp;nodes=single&amp;spec=dflash&amp;tier=low-latency&amp;ssmDtype=bfloat16" rel="nofollow ugc">https://lmsysorg.mintlify.app/cookbook/autoregressive/Qwen/Qwen3.8-27B#hw=dgx-spark&amp;variant=default&amp;quant=nvfp4-fp4-head&amp;nodes=single&amp;spec=dflash&amp;tier=low-latency&amp;ssmDtype=bfloat16</a></p>
<p dir="auto">之前试过了 vllm 配置。不行。太慢了。不现实。</p>
<p dir="auto">主要用意是在Hermes agent 使用。可是我试跑后不是很理想。我相信是我的配置问题？</p>
<p dir="auto">请求<img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/1f64f.png?v=301515bb865" class="not-responsive emoji emoji-android emoji--pray" style="height:23px;width:auto;vertical-align:middle" title="🙏" alt="🙏" />有哪位大神可以帮帮忙。</p>
]]></description><link>https://lcz.me/topic/1392</link><generator>RSS for Node</generator><lastBuildDate>Thu, 10 Sep 2026 04:34:26 GMT</lastBuildDate><atom:link href="https://lcz.me/topic/1392.rss" rel="self" type="application/rss+xml"/><pubDate>Fri, 28 Aug 2026 17:45:25 GMT</pubDate><ttl>60</ttl><item><title><![CDATA[Reply to DGX Spark - MSI Edgexpert on Sat, 29 Aug 2026 06:10:01 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/rex-fan" aria-label="Profile: Rex-Fan">@<bdi>Rex-Fan</bdi></a></p>
<p dir="auto">谢谢你的建议。不过是怎么知道 dgx spark 不适合跑稠密模型？还请大神赐点高建。</p>
]]></description><link>https://lcz.me/post/14729</link><guid isPermaLink="true">https://lcz.me/post/14729</guid><dc:creator><![CDATA[chrestkoo dev]]></dc:creator><pubDate>Sat, 29 Aug 2026 06:10:01 GMT</pubDate></item><item><title><![CDATA[Reply to DGX Spark - MSI Edgexpert on Sat, 29 Aug 2026 06:07:53 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/imbiplaza-asus" aria-label="Profile: imbiplaza-ASUS">@<bdi>imbiplaza-ASUS</bdi></a> 是的。就是他的架构关系。导致现在这样</p>
]]></description><link>https://lcz.me/post/14727</link><guid isPermaLink="true">https://lcz.me/post/14727</guid><dc:creator><![CDATA[chrestkoo dev]]></dc:creator><pubDate>Sat, 29 Aug 2026 06:07:53 GMT</pubDate></item><item><title><![CDATA[Reply to DGX Spark - MSI Edgexpert on Sat, 29 Aug 2026 06:07:22 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/kos-or" aria-label="Profile: kos-or">@<bdi>kos-or</bdi></a> 我还没使用这个 qwen 3.8 之前是使用 Qwen3.6-35B-A3B。用的很好。很顺。<br />
我个人觉得普通用还可以。复杂一点的话就要用到 skill, script 强制做对。</p>
<p dir="auto">这是我的 podman 配置：</p>
<pre><code>name: edgeexpert-reasoning

services:
  reasoning-vlm:
    image: ${REASONING_VLLM_IMAGE:-vllm/vllm-openai:latest}
    container_name: edgeexpert-reasoning-vlm
    restart: unless-stopped
    gpus: all
    ipc: host
    ports:
      - "${BIND_ADDRESS:-0.0.0.0}:8000:8000"
    environment:
      HF_TOKEN: ${HF_TOKEN:?Set HF_TOKEN in the .env file}
      VLLM_API_KEY: ${VLLM_API_KEY:?Set VLLM_API_KEY in the .env file}
    volumes:
      - ${HF_CACHE_DIR:-/home/koo/.cache/huggingface}:/root/.cache/huggingface
    command:
      - ${REASONING_MODEL:-nvidia/Qwen3.6-35B-A3B-NVFP4}
      - --served-model-name
      - ${REASONING_SERVED_MODEL_NAME:-nvidia/Qwen3.6-35B-A3B-NVFP4}
      - --host
      - "0.0.0.0"
      - --port
      - "8000"
      - --tensor-parallel-size
      - "1"
      - --trust-remote-code
      - --kv-cache-dtype
      - fp8
      - --attention-backend
      - flashinfer
      - --moe-backend
      - marlin
      - --gpu-memory-utilization
      - ${REASONING_GPU_MEMORY_UTILIZATION:-0.4}
      - --max-model-len
      - "262144"
      - --max-num-seqs
      - ${REASONING_MAX_NUM_SEQS:-8}
      - --max-num-batched-tokens
      - "8192"
      - --enable-chunked-prefill
      - --async-scheduling
      - --enable-prefix-caching
      - --speculative-config
      - '{"method":"mtp","num_speculative_tokens":3,"moe_backend":"triton"}'
      - --load-format
      - fastsafetensors
      - --reasoning-parser
      - qwen3
      - --tool-call-parser
      - qwen3_xml
      - --enable-auto-tool-choice

    healthcheck:
      test:
        - CMD-SHELL
        - &gt;-
          python3 -c "import urllib.request;
          urllib.request.urlopen('http://127.0.0.1:8000/health', timeout=5)"

      interval: 30m
      timeout: 10s
      retries: 3
      start_period: 30m

</code></pre>
<p dir="auto">还没用过 ornith-ai/Ornith-1.5-35B-A3B。请问有谁用过这个吗？</p>
]]></description><link>https://lcz.me/post/14725</link><guid isPermaLink="true">https://lcz.me/post/14725</guid><dc:creator><![CDATA[chrestkoo dev]]></dc:creator><pubDate>Sat, 29 Aug 2026 06:07:22 GMT</pubDate></item><item><title><![CDATA[Reply to DGX Spark - MSI Edgexpert on Sat, 29 Aug 2026 02:02:24 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/chrestkoo-dev" aria-label="Profile: chrestkoo-dev">@<bdi>chrestkoo-dev</bdi></a> dgx spark 不适合跑稠密模型，看看moe模型吧。单机器可以等等看社区会不会给出一个 qwen 3.8 next 的配方，目前pro 6000 单卡已经可以跑了</p>
]]></description><link>https://lcz.me/post/14681</link><guid isPermaLink="true">https://lcz.me/post/14681</guid><dc:creator><![CDATA[Rex Fan]]></dc:creator><pubDate>Sat, 29 Aug 2026 02:02:24 GMT</pubDate></item><item><title><![CDATA[Reply to DGX Spark - MSI Edgexpert on Sat, 29 Aug 2026 00:47:58 GMT]]></title><description><![CDATA[<p dir="auto">如果没有nvidia 架构的需求, 我觉得Mac Studio max or ultra 才是最好的</p>
]]></description><link>https://lcz.me/post/14675</link><guid isPermaLink="true">https://lcz.me/post/14675</guid><dc:creator><![CDATA[imbiplaza ASUS]]></dc:creator><pubDate>Sat, 29 Aug 2026 00:47:58 GMT</pubDate></item><item><title><![CDATA[Reply to DGX Spark - MSI Edgexpert on Sat, 29 Aug 2026 00:06:34 GMT]]></title><description><![CDATA[<blockquote>
<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/chrestkoo-dev" aria-label="Profile: chrestkoo-dev">@<bdi>chrestkoo-dev</bdi></a> <a href="/post/14654">said</a>:</p>
<p dir="auto">DGX Spark - MSI Edgexpert</p>
<p dir="auto">打算使用单机跑 Qwen3.8。可以参考配置如下：</p>
</blockquote>
<p dir="auto">DGX Spark 128GB Bandwidth 273.2 GB/s<br />
RTX 5050 8GB Bandwidth 320.0 GB/s<br />
RTX 5060 Ti 16GB Bandwidth 448.0 GB/s<br />
AMD RX7900 XTX 24GB Bandwidth 960.0 GB/s<br />
RTX 5070 Ti 16GB Bandwidth 896.0 GB/s<br />
RTX Pro 5000 48GB Bandwidth 1340 GB/s<br />
RTX 5090 Bandwidth 1790 GB/s</p>
<p dir="auto">DGX Spark 大概要再買第二台 速度表現才會到達能用的程度 (你可以參考雙機器的速度)<br />
這機子剛出來的時候研究過 基本上是為了微調LLM使用的</p>
<p dir="auto">有一種解套的方式就是跑Qwen3.6-35B-A3B 試看看 但不知道夠聰明不 能否應付Agent的場景 ？</p>
<p dir="auto">或用這一個 ornith-ai/Ornith-1.5-35B-A3B 高精度 這個效果可能更好<br />
<a href="https://huggingface.co/ornith-ai/Ornith-1.5-35B-A3B" rel="nofollow ugc">https://huggingface.co/ornith-ai/Ornith-1.5-35B-A3B</a></p>
]]></description><link>https://lcz.me/post/14671</link><guid isPermaLink="true">https://lcz.me/post/14671</guid><dc:creator><![CDATA[kos or]]></dc:creator><pubDate>Sat, 29 Aug 2026 00:06:34 GMT</pubDate></item><item><title><![CDATA[Reply to DGX Spark - MSI Edgexpert on Fri, 28 Aug 2026 19:11:05 GMT]]></title><description><![CDATA[<p dir="auto">方向没错，你贴的这条 LMSYS cookbook 配置就是 DGX Spark 跑 Qwen3.8-27B 的官方推荐档。问题大概率出在"用了 vLLM"上，不是你的配置：</p>
<p dir="auto"><strong>1. vLLM 为什么慢</strong><br />
Qwen3.8-27B 是 Mamba 混合架构（SSM + attention 并存），vLLM 对混合架构的 fused kernel 支持不成熟，跑起来会落慢路径，速度完全没法看。这是框架问题，换谁都一样。</p>
<p dir="auto"><strong>2. 正确姿势：SGLang 按 cookbook 参数跑</strong></p>
<ul>
<li>量化用 <strong>NVFP4-FP4-head</strong>：27B 权重约 14GB，128G 统一内存随便塞，还能留足长上下文；</li>
<li><code>--spec-algorithm dflash</code>（就是链接里 spec=dflash，混合架构的融合 attention 路径）；</li>
<li><code>--ssm-dtype bfloat16</code>（Mamba 部分精度，跟 cookbook 保持一致）；</li>
<li>单机 single node + low-latency tier 的其它参数照抄。</li>
</ul>
<p dir="auto"><strong>3. Hermes 接入</strong><br />
SGLang 起的就是 OpenAI 兼容端点。Hermes 里配一个 OpenAI-compatible provider，base_url 指到 SGLang 端口（默认 8000），模型名填 qwen3.8-27B。max_tokens 设大（16K+），本地模型别用小窗口。</p>
<p dir="auto"><strong>4. 备选</strong><br />
SGLang 装起来费劲的话，llama.cpp 跑 UD-Q5_K_XL 的 GGUF（约 16GB）在 DGX Spark 上也能跑，速度比 DFLASH 档慢一些但胜在稳，Hermes 同样走 OpenAI 兼容端点。</p>
<p dir="auto"><strong>5. 小提示</strong><br />
cookbook 的 variant=default 参数要和你的 SGLang 版本对得上，先原样跑通确认能出数，再动 variant 调优。另外 MSI Edgexpert 是管理/监控那层，不参与模型推理，不用管它。</p>
]]></description><link>https://lcz.me/post/14659</link><guid isPermaLink="true">https://lcz.me/post/14659</guid><dc:creator><![CDATA[Xiaote]]></dc:creator><pubDate>Fri, 28 Aug 2026 19:11:05 GMT</pubDate></item></channel></rss>