<?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[3090ti部署qwen3.6-27B-MTP-q4_K-M的疑惑]]></title><description><![CDATA[<p dir="auto">启动参数，200K上下文，就是没办法拉满，有大神指导一下吗<br />
MODEL="/home/stephen/models/qwen3.6-27b-mtp/Qwen3.6-27B-Q4_K_M.gguf"<br />
MMR="/home/stephen/models/qwen3.6-27b-mtp/mmproj-BF16.gguf"<br />
LLAMA_SERVER="${LLAMA_SERVER:-/home/stephen/llama.cpp-turbo/build/bin/llama-server}"</p>
<p dir="auto">if [ ! -f "$LLAMA_SERVER" ]; then<br />
echo "ERROR: llama-server not found at $LLAMA_SERVER"<br />
exit 1<br />
fi</p>
<p dir="auto">CUDA_VISIBLE_DEVICES=0 "$LLAMA_SERVER" <br />
-m "$MODEL" <br />
--mmproj "$MMR" <br />
--spec-type draft-mtp <br />
--spec-draft-n-max 2 <br />
--spec-draft-p-min 0.75 <br />
--parallel 1 <br />
--host 127.0.0.1 <br />
--port 8080 <br />
--ctx-size 200704 <br />
--n-gpu-layers 99 <br />
--cache-type-k turbo4 <br />
--cache-type-v turbo4 <br />
--flash-attn on <br />
--no-mmap <br />
--reasoning off <br />
--jinja <br />
"$@"</p>
]]></description><link>https://lcz.me/topic/385/3090ti部署qwen3.6-27b-mtp-q4_k-m的疑惑</link><generator>RSS for Node</generator><lastBuildDate>Thu, 11 Jun 2026 13:58:19 GMT</lastBuildDate><atom:link href="https://lcz.me/topic/385.rss" rel="self" type="application/rss+xml"/><pubDate>Mon, 01 Jun 2026 16:38:59 GMT</pubDate><ttl>60</ttl><item><title><![CDATA[Reply to 3090ti部署qwen3.6-27B-MTP-q4_K-M的疑惑 on Mon, 08 Jun 2026 08:43:12 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/stxpnet" aria-label="Profile: stxpnet">@<bdi>stxpnet</bdi></a></p>
<p dir="auto">3090理論上支持BF16, 不過好像強行用因為表現會比FP16跟TF32更差, 所以沒有人去優化</p>
<p dir="auto"><a href="https://github.com/huggingface/transformers/issues/14608" rel="nofollow ugc">Github</a></p>
]]></description><link>https://lcz.me/post/5707</link><guid isPermaLink="true">https://lcz.me/post/5707</guid><dc:creator><![CDATA[566656661]]></dc:creator><pubDate>Mon, 08 Jun 2026 08:43:12 GMT</pubDate></item><item><title><![CDATA[Reply to 3090ti部署qwen3.6-27B-MTP-q4_K-M的疑惑 on Mon, 08 Jun 2026 07:29:48 GMT]]></title><description><![CDATA[<p dir="auto">还有个问题,你要200K, 最好不要加载投影文件,加了投影文件似乎会拖慢速度,并且智商似乎也有影响.<br />
删除那行 MMR="/home/stephen/models/qwen3.6-27b-mtp/mmproj-BF16.gguf" ,模型只会识图能力(另外似乎3090系不支持BF16格式? 我也是问过AI,所以我之前测试都是下载的F16格式,有些大神制作的模型没有F16投影文件) .<br />
我主要用来编程,改BUG,所以直接不加载投影. 我有另一台二奶机,16G老显卡,如果有识图需求是在上面加载的9B模型来给HERMES识图用.</p>
]]></description><link>https://lcz.me/post/5694</link><guid isPermaLink="true">https://lcz.me/post/5694</guid><dc:creator><![CDATA[stxpnet]]></dc:creator><pubDate>Mon, 08 Jun 2026 07:29:48 GMT</pubDate></item><item><title><![CDATA[Reply to 3090ti部署qwen3.6-27B-MTP-q4_K-M的疑惑 on Mon, 08 Jun 2026 01:09:28 GMT]]></title><description><![CDATA[<blockquote>
<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/stxpnet" aria-label="Profile: stxpnet">@<bdi>stxpnet</bdi></a> <a href="/post/5485">说</a>:</p>
<p dir="auto">noonghunna/club-3090</p>
</blockquote>
<p dir="auto">看起来不错，我要去试一下</p>
]]></description><link>https://lcz.me/post/5627</link><guid isPermaLink="true">https://lcz.me/post/5627</guid><dc:creator><![CDATA[Don Zhu 0]]></dc:creator><pubDate>Mon, 08 Jun 2026 01:09:28 GMT</pubDate></item><item><title><![CDATA[Reply to 3090ti部署qwen3.6-27B-MTP-q4_K-M的疑惑 on Sun, 07 Jun 2026 11:16:50 GMT]]></title><description><![CDATA[<p dir="auto">自己起的命令行不知道为什么不行， 可以克隆noonghunna/club-3090 这个大神的REPO，它会在系统盘拉取容器，轻松突破200K，后期你跑顺了也可以自己去找找yml文件，修改其中的配置。但是智商会有一定损伤的。 目前我日常就使用这个llamacpp/mtpiq4nl     (IQ4_NL MTP)<br />
。</p>
]]></description><link>https://lcz.me/post/5485</link><guid isPermaLink="true">https://lcz.me/post/5485</guid><dc:creator><![CDATA[stxpnet]]></dc:creator><pubDate>Sun, 07 Jun 2026 11:16:50 GMT</pubDate></item><item><title><![CDATA[Reply to 3090ti部署qwen3.6-27B-MTP-q4_K-M的疑惑 on Thu, 04 Jun 2026 15:14:43 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/comen" aria-label="Profile: comeN">@<bdi>comeN</bdi></a> 现在就是保持200K上下文<br />
comfyui还要花时间折腾，3090ti跑模型，3060跑comfyui，跑通之后再考虑换显卡</p>
]]></description><link>https://lcz.me/post/5026</link><guid isPermaLink="true">https://lcz.me/post/5026</guid><dc:creator><![CDATA[暧昧光影]]></dc:creator><pubDate>Thu, 04 Jun 2026 15:14:43 GMT</pubDate></item><item><title><![CDATA[Reply to 3090ti部署qwen3.6-27B-MTP-q4_K-M的疑惑 on Tue, 02 Jun 2026 14:01:16 GMT]]></title><description><![CDATA[<p dir="auto">一般200K就够用，超过200K的大任务你最好交给DeepSeek-v4-flash的API解决问题。</p>
]]></description><link>https://lcz.me/post/4680</link><guid isPermaLink="true">https://lcz.me/post/4680</guid><dc:creator><![CDATA[comeN]]></dc:creator><pubDate>Tue, 02 Jun 2026 14:01:16 GMT</pubDate></item><item><title><![CDATA[Reply to 3090ti部署qwen3.6-27B-MTP-q4_K-M的疑惑 on Tue, 02 Jun 2026 06:16:36 GMT]]></title><description><![CDATA[<p dir="auto">换成 IQ4-XS就能开到256K上下文了，智商影响不大，个别版本跟K M 参数非常接近</p>
]]></description><link>https://lcz.me/post/4639</link><guid isPermaLink="true">https://lcz.me/post/4639</guid><dc:creator><![CDATA[asd2667]]></dc:creator><pubDate>Tue, 02 Jun 2026 06:16:36 GMT</pubDate></item><item><title><![CDATA[Reply to 3090ti部署qwen3.6-27B-MTP-q4_K-M的疑惑 on Tue, 02 Jun 2026 06:00:48 GMT]]></title><description><![CDATA[<blockquote>
<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/%E6%9A%A7%E6%98%A7%E5%85%89%E5%BD%B1" aria-label="Profile: 暧昧光影">@<bdi>暧昧光影</bdi></a> <a href="/post/4630">说</a>:</p>
<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/applejuice" aria-label="Profile: applejuice">@<bdi>applejuice</bdi></a> 我是双卡配置，只用单独3090ti跑llama.cpp，turbo4之后上下文最多到200k，256不行。vllm问了ai肯定不行，128k都够呛。<br />
如果不加多模态投影，省点显存的话，回头试试是否可以256k拉满</p>
</blockquote>
<p dir="auto">单卡 我就没试过了<br />
单卡我才开65k 上下文 ComfyUI 上线的时候 过渡用罢了<br />
其他时候都是2张卡跑LLM</p>
]]></description><link>https://lcz.me/post/4637</link><guid isPermaLink="true">https://lcz.me/post/4637</guid><dc:creator><![CDATA[applejuice]]></dc:creator><pubDate>Tue, 02 Jun 2026 06:00:48 GMT</pubDate></item><item><title><![CDATA[Reply to 3090ti部署qwen3.6-27B-MTP-q4_K-M的疑惑 on Tue, 02 Jun 2026 04:55:16 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/applejuice" aria-label="Profile: applejuice">@<bdi>applejuice</bdi></a> 我是双卡配置，只用单独3090ti跑llama.cpp，turbo4之后上下文最多到200k，256不行。vllm问了ai肯定不行，128k都够呛。<br />
如果不加多模态投影，省点显存的话，回头试试是否可以256k拉满</p>
]]></description><link>https://lcz.me/post/4630</link><guid isPermaLink="true">https://lcz.me/post/4630</guid><dc:creator><![CDATA[暧昧光影]]></dc:creator><pubDate>Tue, 02 Jun 2026 04:55:16 GMT</pubDate></item><item><title><![CDATA[Reply to 3090ti部署qwen3.6-27B-MTP-q4_K-M的疑惑 on Tue, 02 Jun 2026 04:51:53 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/xiaote" aria-label="Profile: Xiaote">@<bdi>Xiaote</bdi></a> turbo4上下文能拉到200k，256k会oom</p>
]]></description><link>https://lcz.me/post/4629</link><guid isPermaLink="true">https://lcz.me/post/4629</guid><dc:creator><![CDATA[暧昧光影]]></dc:creator><pubDate>Tue, 02 Jun 2026 04:51:53 GMT</pubDate></item><item><title><![CDATA[Reply to 3090ti部署qwen3.6-27B-MTP-q4_K-M的疑惑 on Mon, 01 Jun 2026 19:05:39 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/%E6%9A%A7%E6%98%A7%E5%85%89%E5%BD%B1" aria-label="Profile: 暧昧光影">@<bdi>暧昧光影</bdi></a> 关于你的两个问题：</p>
<p dir="auto"><strong>1. 200K上下文拉不满的原因：</strong><br />
3090ti只有24G显存，Qwen3.6-27B-Q4_K_M模型本身占约16-17G，剩下7-8G给KV Cache。<br />
你开了MTP（draft-mtp）还会额外占用draft model的显存，实际留给KV Cache的空间更少。</p>
<p dir="auto">即使用了 <code>--cache-type-k turbo4 --cache-type-v turbo4</code>（Q4量化KV Cache），200K上下文在27B模型上也远远超过24G的承载能力。粗略估算：27B模型80层左右，Q4 KV Cache每个token约0.5字节×4096隐藏层×2(K+V)×80层 ≈ 0.3MB/token。200K tokens ≈ 61GB显存，这还没算模型本身。</p>
<p dir="auto"><strong>建议解决方案：</strong></p>
<ul>
<li>先用 <code>--ctx-size 32768</code> 跑（32K上下文），绝大部分场景完全够用</li>
<li>如果确实需要长上下文，加 <code>--no-kv-offload</code> 把KV Cache放到系统内存，但推理速度会下降</li>
<li>或者先用普通版（非MTP）的Qwen3.6-27B跑通，确认基础设置没问题再开MTP——MTP本身也会多吃显存</li>
<li>另外可以加上 <code>--cont-batching</code> 配合context shifting，长对话时自动丢弃早期token</li>
</ul>
<p dir="auto"><strong>2. CUDA设备编号问题：</strong><br />
这是正常的。nvidia-smi的GPU编号和CUDA的设备编号遵循不同的枚举顺序（PCIe总线顺序vs驱动加载顺序）。你可以用 <code>nvidia-smi -q -d INDEX</code> 查看每个GPU的Bus-ID，然后在代码里根据Bus-ID确认对应关系。或者直接试试 CUDA_VISIBLE_DEVICES=1 指定另一张卡，看能效有没有区别。</p>
]]></description><link>https://lcz.me/post/4591</link><guid isPermaLink="true">https://lcz.me/post/4591</guid><dc:creator><![CDATA[Xiaote]]></dc:creator><pubDate>Mon, 01 Jun 2026 19:05:39 GMT</pubDate></item><item><title><![CDATA[Reply to 3090ti部署qwen3.6-27B-MTP-q4_K-M的疑惑 on Mon, 01 Jun 2026 17:20:43 GMT]]></title><description><![CDATA[<p dir="auto">不明白什么是 上下文不能拉满？</p>
<p dir="auto">下面是我用的CUDA 跟驱动给你参考<br />
我是有nvlink但是据了解设置没什么分别</p>
<pre><code>+-----------------------------------------------------------------------------+
| NVIDIA-SMI 595.71.05              Driver Version: 595.71.05    CUDA: 13.2   |
+-----------------------------------------------------------------------------+
</code></pre>
<p dir="auto">之前跑的llama.cpp</p>
<h2>Flags (in docker-compose.yml)</h2>
<pre><code>--model /models/heretic-gptq-int4
--served-model-name qwen3.6-27b-heretic
--quantization gptq_marlin
--dtype float16
--tensor-parallel-size 2              # both 3090s, real NVLink usage
--max-model-len 262144                # 262K context
--gpu-memory-utilization 0.92
--max-num-seqs 2                      # 2 concurrent streams
--max-num-batched-tokens 8192
--kv-cache-dtype fp8_e5m2             # 1 byte/token KV
--trust-remote-code
--reasoning-parser qwen3              # routes &lt;think&gt; → reasoning_content
--enable-auto-tool-choice
--tool-call-parser qwen3_coder        # native Qwen3 tool format
--enable-prefix-caching               # repeated prompts share KV
--enable-chunked-prefill              # long prefill doesn't block decode
--disable-custom-all-reduce           # MANDATORY for cross-NUMA setup
--host 0.0.0.0
--port 8000                           # container-side; mapped to host :8011
</code></pre>
<p dir="auto">我的vllm设置</p>
<h2>Flags (in docker-compose.yml)</h2>
<pre><code>--model /models/heretic-gptq-int4
--served-model-name qwen3.6-27b-heretic
--quantization gptq_marlin
--dtype float16
--tensor-parallel-size 2              # both 3090s, real NVLink usage
--max-model-len 262144                # 262K context
--gpu-memory-utilization 0.92
--max-num-seqs 2                      # 2 concurrent streams
--max-num-batched-tokens 8192
--kv-cache-dtype fp8_e5m2             # 1 byte/token KV
--trust-remote-code
--reasoning-parser qwen3              # routes &lt;think&gt; → reasoning_content
--enable-auto-tool-choice
--tool-call-parser qwen3_coder        # native Qwen3 tool format
--enable-prefix-caching               # repeated prompts share KV
--enable-chunked-prefill              # long prefill doesn't block decode
--disable-custom-all-reduce           # MANDATORY for cross-NUMA setup
--host 0.0.0.0
--port 8000                           # container-side; mapped to host :8011
</code></pre>
<h3>Environment</h3>
<pre><code>PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True,max_split_size_mb:512
VLLM_USE_FLASHINFER_SAMPLER=1
VLLM_NO_USAGE_STATS=1
OMP_NUM_THREADS=1
</code></pre>
<hr />
]]></description><link>https://lcz.me/post/4587</link><guid isPermaLink="true">https://lcz.me/post/4587</guid><dc:creator><![CDATA[applejuice]]></dc:creator><pubDate>Mon, 01 Jun 2026 17:20:43 GMT</pubDate></item><item><title><![CDATA[Reply to 3090ti部署qwen3.6-27B-MTP-q4_K-M的疑惑 on Mon, 01 Jun 2026 16:41:15 GMT]]></title><description><![CDATA[<p dir="auto">我是双显卡CUDA_VISIBLE_DEVICES=0指定运行在3090ti，另外有个问题ubuntu2404，cuda12.8，驱动是580版本。cuda0对应nvidia-smi里面的GPU1，cuda1反而是gpu0。不知道是个什么原因，问ai也解决不了。。。。<a class="plugin-mentions-user plugin-mentions-a" href="/user/terry" aria-label="Profile: terry">@<bdi>terry</bdi></a></p>
]]></description><link>https://lcz.me/post/4585</link><guid isPermaLink="true">https://lcz.me/post/4585</guid><dc:creator><![CDATA[暧昧光影]]></dc:creator><pubDate>Mon, 01 Jun 2026 16:41:15 GMT</pubDate></item></channel></rss>