<?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[双卡 3090 qwen 3.8-27b dflash2 131k上下文 ，100+ t/s 有大神已经搞出来了？]]></title><description><![CDATA[<p dir="auto"><img src="https://upload.lcz.me/uploads/c01a2e96-c140-4e7a-aaf2-d64e72fef894.jpeg" alt="9013f100-a65b-43f7-81d8-4c8b8fc3f8ee-image.jpeg" class=" img-fluid img-markdown" /></p>
<p dir="auto"><a href="https://www.reddit.com/r/LocalLLaMA/comments/1vsccit/qwen3827b_on_2x_3090_vllm_dflash2_218_toks_single/" rel="nofollow ugc">https://www.reddit.com/r/LocalLLaMA/comments/1vsccit/qwen3827b_on_2x_3090_vllm_dflash2_218_toks_single/</a></p>
<p dir="auto"><img src="https://upload.lcz.me/uploads/71bcc9cb-d6b5-4c1f-8e7f-ca64b1866ade.jpeg" alt="d275f82b-5ab0-470a-8e96-ab666e752ffa-image.jpeg" class=" img-fluid img-markdown" /></p>
<p dir="auto"><a href="https://github.com/oceanplexian/vllm/pull/1" rel="nofollow ugc">https://github.com/oceanplexian/vllm/pull/1</a></p>
<p dir="auto">感觉变革太快了。比起3.6时代还要疯狂啊！<br />
如果有256K或200K上下文，我都愿意换用它这个方案！</p>
<p dir="auto">我自己目前是用双卡 qwen 3.8 w4a16的 awq模型，只有66t/s +256K上下文 ，还没机会测试超过131K的上下文。<br />
其实我目前的感觉，没有nvlink的话，100K前还好，过了100K，nccl提供那点带宽就供给不上了，直接导致prefill 速度和生成速度都断崖式下降。</p>
]]></description><link>https://lcz.me/topic/1201/双卡-3090-qwen-3.8-27b-dflash2-131k上下文-100-t-s-有大神已经搞出来了</link><generator>RSS for Node</generator><lastBuildDate>Fri, 21 Aug 2026 23:15:23 GMT</lastBuildDate><atom:link href="https://lcz.me/topic/1201.rss" rel="self" type="application/rss+xml"/><pubDate>Wed, 19 Aug 2026 15:13:59 GMT</pubDate><ttl>60</ttl><item><title><![CDATA[Reply to 双卡 3090 qwen 3.8-27b dflash2 131k上下文 ，100+ t/s 有大神已经搞出来了？ on Thu, 20 Aug 2026 04:11:58 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">我在国外，邮寄需要一个星期多<br />
前两天搞到卖家手上，不知道他们几时才能给回复</p>
<p dir="auto">现在用着</p>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>Item</th>
<th>Value</th>
</tr>
</thead>
<tbody>
<tr>
<td>Model</td>
<td>Qwen3.8-27B UD-Q4_K_M (16.46 GB, sha256-verified, nextn head present)</td>
</tr>
<tr>
<td>Engine</td>
<td>llama.cpp b10498-60addddf3, static CUDA sm_86</td>
</tr>
<tr>
<td>Context</td>
<td>135,168</td>
</tr>
<tr>
<td>KV</td>
<td>q8_0 / q8_0</td>
</tr>
<tr>
<td>Vision</td>
<td>on (mmproj-F16)</td>
</tr>
<tr>
<td>MTP</td>
<td>draft-mtp, n-max 2</td>
</tr>
</tbody>
</table>
]]></description><link>https://lcz.me/post/13026</link><guid isPermaLink="true">https://lcz.me/post/13026</guid><dc:creator><![CDATA[applejuice]]></dc:creator><pubDate>Thu, 20 Aug 2026 04:11:58 GMT</pubDate></item><item><title><![CDATA[Reply to 双卡 3090 qwen 3.8-27b dflash2 131k上下文 ，100+ t/s 有大神已经搞出来了？ on Thu, 20 Aug 2026 00:25:16 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> 你现在那个卡还没修好吗？ 单卡推理跑什么框架和模型？</p>
]]></description><link>https://lcz.me/post/12983</link><guid isPermaLink="true">https://lcz.me/post/12983</guid><dc:creator><![CDATA[stxpnet]]></dc:creator><pubDate>Thu, 20 Aug 2026 00:25:16 GMT</pubDate></item><item><title><![CDATA[Reply to 双卡 3090 qwen 3.8-27b dflash2 131k上下文 ，100+ t/s 有大神已经搞出来了？ on Wed, 19 Aug 2026 16:28:25 GMT]]></title><description><![CDATA[<p dir="auto">我觉得2000prefill 60decode 才是最适合的agent 用. 过了1500prefill 50decode 这条线 我宁愿长上下文 跟 更高准确率</p>
<p dir="auto">一张卡 我昨天读我的本地文件 结果131k context 压缩了好多次<br />
一张卡prefill 好像最多1000 总共花了50分钟...</p>
]]></description><link>https://lcz.me/post/12949</link><guid isPermaLink="true">https://lcz.me/post/12949</guid><dc:creator><![CDATA[applejuice]]></dc:creator><pubDate>Wed, 19 Aug 2026 16:28:25 GMT</pubDate></item><item><title><![CDATA[Reply to 双卡 3090 qwen 3.8-27b dflash2 131k上下文 ，100+ t/s 有大神已经搞出来了？ on Wed, 19 Aug 2026 15:46:26 GMT]]></title><description><![CDATA[<p dir="auto">这些都是意义不大的特殊场景，实际用到的情况不多。来，回复下我，我来给你第100分，谢谢主人<img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/1f602.png?v=60716d54ab2" class="not-responsive emoji emoji-android emoji--joy" style="height:23px;width:auto;vertical-align:middle" title="😂" alt="😂" />，地上有个🧼</p>
]]></description><link>https://lcz.me/post/12936</link><guid isPermaLink="true">https://lcz.me/post/12936</guid><dc:creator><![CDATA[terry]]></dc:creator><pubDate>Wed, 19 Aug 2026 15:46:26 GMT</pubDate></item></channel></rss>