<?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R9700跑minimaxH3效果分享]]></title><description><![CDATA[<h1>《重构人生》H3 视频 —— 跑通环境配置 &amp; 时间基准</h1>
<blockquote>
<p dir="auto">日期：2026-08-31 凌晨（最终跑通版）<br />
结论：Ref2VA 测试片（含中文台词 + 字幕）已跑通，此文档记录可复现的环境配置与实测耗时。</p>
</blockquote>
<hr />
<h2>一、硬件</h2>
<ul>
<li>CPU：Intel Ultra 7 270K Plus（24 核 / 62GB 内存）</li>
<li>GPU：2× AMD Radeon AI PRO R9700 32GB（gfx1201 / RDNA4，PCIe 5.0 x8+x8，无 P2P）</li>
<li>系统：Ubuntu 24.04</li>
</ul>
<hr />
<h2>二、软件栈（关键版本）</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>组件</th>
<th>版本</th>
</tr>
</thead>
<tbody>
<tr>
<td>torch</td>
<td>2.11.0+rocm7.2（<strong>勿用 2.10+rocm7.0=挂死黑屏，勿用 2.12+rocm7.14=DynamicVRAM 卡死</strong>）</td>
</tr>
<tr>
<td>torchvision / torchaudio</td>
<td>0.26.0+rocm7.2 / 2.11.0+rocm7.2</td>
</tr>
<tr>
<td>ComfyUI</td>
<td>0.33.0</td>
</tr>
<tr>
<td>comfy-kitchen</td>
<td>0.2.31（HIP 后端）</td>
</tr>
<tr>
<td>comfy-aimdo</td>
<td>0.4.13（DynamicVRAM 未启用，健康态）</td>
</tr>
<tr>
<td>Python</td>
<td>3.11.16（venv: ~/comfy-h3/venv）</td>
</tr>
</tbody>
</table>
<p dir="auto">关键软链（comfy-kitchen HIP 原生内核）：</p>
<pre><code class="language-bash">ln -sf libamdhip64.so libamdhip64.so.7   # 在 venv/.../torch/lib/ 下
</code></pre>
<hr />
<h2>三、启动参数（ComfyUI）</h2>
<pre><code class="language-bash">export HSA_XNACK=0
python main.py --listen 127.0.0.1 --port 8188 --reserve-vram 1.0
</code></pre>
<ul>
<li>不带 <code>--gpu-only</code>（DiT 20GB 会因显存碎片 OOM）</li>
<li><code>HSA_XNACK=0</code> 抑制 SVM 页迁移开销</li>
<li>启动脚本：<code>~/comfy-h3/start-comfy-rocm7.sh</code></li>
</ul>
<hr />
<h2>四、模型文件</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>角色</th>
<th>文件</th>
<th>大小</th>
</tr>
</thead>
<tbody>
<tr>
<td>DiT 主模型</td>
<td>minimax_h3_fl2va_pruned_int8_convrot.safetensors</td>
<td>20G</td>
</tr>
<tr>
<td>文本编码器 TE</td>
<td>qwen3vl_32b_minimax_h3_int8_convrot.safetensors（INT8，<strong>必用，勿用 nvfp4</strong>）</td>
<td>26G</td>
</tr>
<tr>
<td>视频 VAE</td>
<td>minimax_h3_video_vae_fp16.safetensors</td>
<td>4.9G</td>
</tr>
<tr>
<td>音频 VAE</td>
<td>minimax_h3_audio_vae_fp32.safetensors</td>
<td>578M</td>
</tr>
<tr>
<td>8步 turbo LoRA</td>
<td>minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors</td>
<td>1.9G</td>
</tr>
</tbody>
</table>
<hr />
<h2>五、双卡摆放（ComfyUI-MultiGPU）</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>组件</th>
<th>设备</th>
<th>节点</th>
</tr>
</thead>
<tbody>
<tr>
<td>DiT</td>
<td>cuda:0</td>
<td>UNETLoaderMultiGPU(device=cuda:0)</td>
</tr>
<tr>
<td>文本编码器 TE</td>
<td>cuda:1</td>
<td>CLIPLoaderMultiGPU(device=cuda:1)</td>
</tr>
<tr>
<td>视频 VAE</td>
<td>cuda:1</td>
<td>VAELoaderMultiGPU(device=cuda:1)</td>
</tr>
<tr>
<td>音频 VAE</td>
<td>cuda:1</td>
<td>VAELoaderMultiGPU(device=cuda:1)</td>
</tr>
</tbody>
</table>
<hr />
<h2>六、工作流参数（8步 turbo）</h2>
<ul>
<li>采样器 res_multistep + scheduler simple + steps=8 + denoise=1.0 + LoRA strength 1.0</li>
<li>提交：<code>~/comfy-h3/gen_h3.sh</code>（T2V）/ <code>~/comfy-h3/gen_h3_ref2va.py</code>（Ref2VA）</li>
<li>帧网格 17k+5：5s=124帧、10s=243帧</li>
</ul>
<p dir="auto"><strong>Ref2VA 测试片参数</strong>：768×1344 竖屏 / 243帧(10s) / 8步 / ref-size=match / 3 参考图（林远/赵天成/会议室）</p>
<hr />
<h2>七、实测时间基准</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>任务</th>
<th>分辨率/时长</th>
<th>采样速度</th>
<th>全程耗时</th>
</tr>
</thead>
<tbody>
<tr>
<td>H3 T2V 冒烟</td>
<td>864×480 / 5s(124帧)</td>
<td>11.6s/it</td>
<td><strong>176s</strong>（约 3 分钟）</td>
</tr>
<tr>
<td>Krea-2 生图</td>
<td>768×1344</td>
<td>1.35s/it</td>
<td>30s</td>
</tr>
<tr>
<td><strong>Ref2VA 测试片</strong></td>
<td>768×1344 / 10s(243帧)</td>
<td>190s/it</td>
<td><strong>27分03秒</strong></td>
</tr>
</tbody>
</table>
<ul>
<li>竖屏 768×1344 像素 ≈ 864×480 的 2.5 倍，加上 Ref2VA 参考图参与每个采样步，故 190s/it（T2V 横屏的 16 倍/步）。</li>
<li>加载阶段（TE 26GB + DiT 20GB + 参考图）约 1.5 分钟；采样 8 步约 25 分钟；VAE 解码约 2 分钟。</li>
<li>dmesg SVM 计数全程 0（无卡死、无页迁移死循环）。</li>
</ul>
<hr />
<h2>八、跑通验证</h2>
<ol>
<li>TE 加载：<code>loaded completely; 30714.80 MB usable, 25883.83 MB loaded, full load: True</code>（健康态，非 DynamicVRAM 分阶段）</li>
<li>身份三锚定：林远（侧分短发/细框眼镜/深灰T恤）、赵天成（中分油头/金丝眼镜/圆脸/藏青西装）、会议室（玻璃桌/红色 SYSTEM FAILURE 屏/冷白光）全部照图生成，零漂移、无第三人。</li>
<li>中文台词：<code>&lt;d&gt;[Chinese] 林远，数据库是你动的吧？&lt;/d&gt;</code> + <code>&lt;d&gt;[Chinese] 我没碰过生产数据。&lt;/d&gt;</code>，人声时段音量 +3~4dB 确认念出。</li>
<li>字幕：ASS（Noto Sans CJK SC，白字黑边）时间轴对齐台词。</li>
</ol>
<hr />
<h2>九、一句话可复现</h2>
<p dir="auto"><strong>torch 2.11.0+rocm7.2 + ComfyUI 0.33 + INT8 TE + HSA_XNACK=0 + 8步 turbo + 双卡 MultiGPU，Ref2VA 竖屏 10s 约 27 分钟出片。</strong></p>
]]></description><link>https://lcz.me/topic/1427</link><generator>RSS for Node</generator><lastBuildDate>Mon, 07 Sep 2026 19:02:19 GMT</lastBuildDate><atom:link href="https://lcz.me/topic/1427.rss" rel="self" type="application/rss+xml"/><pubDate>Mon, 31 Aug 2026 00:11:47 GMT</pubDate><ttl>60</ttl><item><title><![CDATA[Reply to 双R9700跑minimaxH3效果分享 on Mon, 31 Aug 2026 02:58:58 GMT]]></title><description><![CDATA[<p dir="auto">直出的视频截图，看着还行，都不用超分了</p>
]]></description><link>https://lcz.me/post/15052</link><guid isPermaLink="true">https://lcz.me/post/15052</guid><dc:creator><![CDATA[罗冰寒]]></dc:creator><pubDate>Mon, 31 Aug 2026 02:58:58 GMT</pubDate></item><item><title><![CDATA[Reply to 双R9700跑minimaxH3效果分享 on Mon, 31 Aug 2026 02:57:35 GMT]]></title><description><![CDATA[<p dir="auto"><img src="https://upload.lcz.me/uploads/f4521248-e1f6-4913-a94e-2a188056a7d2.jpeg" alt="a7045546-9ce8-43e3-833e-b901f432ce19-image.jpeg" class=" img-fluid img-markdown" /></p>
<p dir="auto"><img src="https://upload.lcz.me/uploads/3653ecc7-234c-4c21-b8d7-4e7dd4d158fe.jpeg" alt="88eef9f7-72d0-4dc1-886c-cc398919ebf1-image.jpeg" class=" img-fluid img-markdown" /></p>
]]></description><link>https://lcz.me/post/15051</link><guid isPermaLink="true">https://lcz.me/post/15051</guid><dc:creator><![CDATA[罗冰寒]]></dc:creator><pubDate>Mon, 31 Aug 2026 02:57:35 GMT</pubDate></item><item><title><![CDATA[Reply to 双R9700跑minimaxH3效果分享 on Mon, 31 Aug 2026 01:51:45 GMT]]></title><description><![CDATA[<p dir="auto">看到r9700pro 原来比我慢一倍，我的是 12s, 1.0m 7m - 8m 出视频，平均是70s/it - 80s/it</p>
<p dir="auto">看来nvidia 真的是用钱换时间。。。</p>
<p dir="auto">毕竟r9700pro也是便宜一倍</p>
<p dir="auto">现在我用我的studio 做视频，他会自己分割用多少秒数来生成，现在我都是做45s - 60s 为主</p>
<p dir="auto">看背后走46s/it 。。。别羡慕哥。。。<br />
。。</p>
<p dir="auto"><img src="https://upload.lcz.me/uploads/43d9a9bf-5bfd-43a7-b482-2384a46ec2c1.jpeg" alt="0f451de7-513f-41f2-87e4-8136a5d3282e-image.jpeg" class=" img-fluid img-markdown" /></p>
<p dir="auto">这是关于花4000万买一条香蕉的故事<br />
我也来乘热度</p>
<p dir="auto"><img src="https://upload.lcz.me/uploads/5c092090-bb7e-4294-8a2e-93daaf4954ed.jpeg" alt="37b66194-9d2b-4547-a837-6560446ec901-image.jpeg" class=" img-fluid img-markdown" /></p>
]]></description><link>https://lcz.me/post/15038</link><guid isPermaLink="true">https://lcz.me/post/15038</guid><dc:creator><![CDATA[imbiplaza ASUS]]></dc:creator><pubDate>Mon, 31 Aug 2026 01:51:45 GMT</pubDate></item><item><title><![CDATA[Reply to 双R9700跑minimaxH3效果分享 on Mon, 31 Aug 2026 01:19:52 GMT]]></title><description><![CDATA[<p dir="auto">截图也行啊</p>
]]></description><link>https://lcz.me/post/15034</link><guid isPermaLink="true">https://lcz.me/post/15034</guid><dc:creator><![CDATA[terry]]></dc:creator><pubDate>Mon, 31 Aug 2026 01:19:52 GMT</pubDate></item><item><title><![CDATA[Reply to 双R9700跑minimaxH3效果分享 on Mon, 31 Aug 2026 00:13:52 GMT]]></title><description><![CDATA[<p dir="auto">视频超过2M上传不了，一会压缩下</p>
]]></description><link>https://lcz.me/post/15025</link><guid isPermaLink="true">https://lcz.me/post/15025</guid><dc:creator><![CDATA[罗冰寒]]></dc:creator><pubDate>Mon, 31 Aug 2026 00:13:52 GMT</pubDate></item></channel></rss>