ComfyUI-MiniMaxH3-Context-Loop/ComfyUI-MAINodes/LOWVRAM_lines.html

<!doctype html><meta charset="utf-8"><title>H3 De-Rope Memory Lines</title>
<style>
:root{--bg:#ffffff;--ink:#1a1a19;--ink2:#5c5b55;--ink3:#8a8982;--grid:#ececea;--rule:#c8c7c0;--torch:#2a78d6;--c16:#eb6834;--c24:#1baf7a;--c31:#eda100;--c32:#e87ba4;--single:#2a78d6;--pipe:#eb6834;--nofd:#4a3aa7}
@media (prefers-color-scheme: dark){:root:not([data-theme="light"]){--bg:#1a1a19;--ink:#ffffff;--ink2:#c3c2b7;--ink3:#8f8e86;--grid:#2c2c2a;--rule:#4a4a46;--torch:#3987e5;--c16:#d95926;--c24:#199e70;--c31:#c98500;--c32:#d55181;--single:#3987e5;--pipe:#d95926;--nofd:#9085e9}}
:root[data-theme="dark"]{--bg:#1a1a19;--ink:#ffffff;--ink2:#c3c2b7;--ink3:#8f8e86;--grid:#2c2c2a;--rule:#4a4a46;--torch:#3987e5;--c16:#d95926;--c24:#199e70;--c31:#c98500;--c32:#d55181;--single:#3987e5;--pipe:#d95926;--nofd:#9085e9}
body{background:var(--bg);color:var(--ink);font:14px/1.5 system-ui,sans-serif;margin:24px;max-width:960px}
h1{font-size:22px;margin:0 0 4px} p{color:var(--ink2);max-width:70ch} svg{width:100%;height:auto;display:block;margin:12px 0 28px}
.t1{font-size:15px;font-weight:600;fill:var(--ink)} .t3{font-size:11px;fill:var(--ink2)} .t4{font-size:10px;fill:var(--ink3)}
.grid{stroke:var(--grid)} .rule{stroke-dasharray:4 3;stroke-width:1} .pt{stroke:var(--bg);stroke-width:2;cursor:default}
table{border-collapse:collapse;font-size:12px;font-variant-numeric:tabular-nums} th,td{border-bottom:1px solid var(--grid);padding:4px 8px;text-align:left} th{color:var(--ink2);font-weight:600}
</style>
<h1>How long can a de-rope go on your card</h1>
<p>Measured 2026-08-18 on one machine (RTX PRO 6000 Blackwell fenced to smaller cards with a VRAM balloon, RAM fenced with cgroups), MiniMax-H3 W4A8 DiT, MAINodes <code>H3 Streamed Blocks</code> (exact), ComfyUI dynamic VRAM, <code>--fast-disk</code>, PyTorch flash attention. Hover a point. Frames are at 1376x768; other canvases scale by tokens.</p>
<svg viewBox="0 0 900 400" role="img" aria-label="VRAM: activations are card-independent; the card only sets the ceiling"><text x="70" y="18" class="t1">VRAM: activations are card-independent; the card only sets the ceiling</text><line x1="70" x2="860" y1="330.0" y2="330.0" class="grid"/><text x="62" y="334.0" text-anchor="end" class="t3">0</text><line x1="70" x2="860" y1="285.0" y2="285.0" class="grid"/><text x="62" y="289.0" text-anchor="end" class="t3">6</text><line x1="70" x2="860" y1="240.0" y2="240.0" class="grid"/><text x="62" y="244.0" text-anchor="end" class="t3">12</text><line x1="70" x2="860" y1="195.0" y2="195.0" class="grid"/><text x="62" y="199.0" text-anchor="end" class="t3">18</text><line x1="70" x2="860" y1="150.0" y2="150.0" class="grid"/><text x="62" y="154.0" text-anchor="end" class="t3">24</text><line x1="70" x2="860" y1="105.0" y2="105.0" class="grid"/><text x="62" y="109.0" text-anchor="end" class="t3">30</text><line x1="70" x2="860" y1="60.0" y2="60.0" class="grid"/><text x="62" y="64.0" text-anchor="end" class="t3">36</text><text x="70.0" y="346" text-anchor="middle" class="t3">0</text><text x="70.0" y="360" text-anchor="middle" class="t4">0k</text><text x="204.9" y="346" text-anchor="middle" class="t3">294</text><text x="204.9" y="360" text-anchor="middle" class="t4">90k</text><text x="392.1" y="346" text-anchor="middle" class="t3">702</text><text x="392.1" y="360" text-anchor="middle" class="t4">214k</text><text x="485.6" y="346" text-anchor="middle" class="t3">906</text><text x="485.6" y="360" text-anchor="middle" class="t4">276k</text><text x="579.2" y="346" text-anchor="middle" class="t3">1110</text><text x="579.2" y="360" text-anchor="middle" class="t4">339k</text><text x="719.6" y="346" text-anchor="middle" class="t3">1416</text><text x="719.6" y="360" text-anchor="middle" class="t4">432k</text><text x="860.0" y="346" text-anchor="middle" class="t3">1722</text><text x="860.0" y="360" text-anchor="middle" class="t4">525k</text><text x="465" y="376" text-anchor="middle" class="t3">frames at 1376x768 (packed tokens below)</text><text transform="rotate(-90)" x="-180" y="16" text-anchor="middle" class="t3">GiB</text><line x1="70" x2="860" y1="210.0" y2="210.0" class="rule" style="stroke:var(--c16)"/><text x="864" y="214.0" class="t3">16 GB</text><line x1="70" x2="860" y1="150.0" y2="150.0" class="rule" style="stroke:var(--c24)"/><text x="864" y="154.0" class="t3">24 GB</text><line x1="70" x2="860" y1="90.0" y2="90.0" class="rule" style="stroke:var(--c32)"/><text x="864" y="94.0" class="t3">32 GB</text><circle cx="111.3" cy="316.5" r="5" style="fill:var(--torch)" class="pt"><title>torch activation peak (in-forward, any card)
90 f / 28k tokens: torch peak +1.8 GiB (TE matrix A, 25 steps)</title></circle><circle cx="111.3" cy="316.5" r="5" style="fill:var(--torch)" class="pt"><title>torch activation peak (in-forward, any card)
90 f / 28k tokens: torch peak +1.8 GiB (TE matrix A, 25 steps)</title></circle><circle cx="392.1" cy="240.8" r="5" style="fill:var(--torch)" class="pt"><title>torch activation peak (in-forward, any card)
702 f / 214k tokens: torch peak +11.9 GiB (v014: bank, fast-disk, balloon 63)</title></circle><circle cx="392.1" cy="240.8" r="5" style="fill:var(--torch)" class="pt"><title>torch activation peak (in-forward, any card)
702 f / 214k tokens: torch peak +11.9 GiB (v012: planner+cap card)</title></circle><circle cx="392.1" cy="240.8" r="5" style="fill:var(--torch)" class="pt"><title>torch activation peak (in-forward, any card)
702 f / 214k tokens: torch peak +11.9 GiB (v011: normal mode, no fast-disk)</title></circle><circle cx="392.1" cy="240.8" r="5" style="fill:var(--torch)" class="pt"><title>torch activation peak (in-forward, any card)
702 f / 214k tokens: torch peak +11.9 GiB (line sweep, 2 steps)</title></circle><circle cx="392.1" cy="240.8" r="5" style="fill:var(--torch)" class="pt"><title>torch activation peak (in-forward, any card)
702 f / 214k tokens: torch peak +11.9 GiB (line sweep, 2 steps)</title></circle><circle cx="392.1" cy="240.8" r="5" style="fill:var(--torch)" class="pt"><title>torch activation peak (in-forward, any card)
702 f / 214k tokens: torch peak +11.9 GiB (line sweep, 2 steps)</title></circle><circle cx="392.1" cy="240.8" r="5" style="fill:var(--torch)" class="pt"><title>torch activation peak (in-forward, any card)
702 f / 214k tokens: torch peak +11.9 GiB (line sweep, 2 steps)</title></circle><circle cx="485.6" cy="218.2" r="5" style="fill:var(--torch)" class="pt"><title>torch activation peak (in-forward, any card)
906 f / 276k tokens: torch peak +14.9 GiB (line sweep, 2 steps)</title></circle><path d="M111.3,316.2 L204.9,291.5 L392.1,242.0 L485.6,217.3 L579.2,192.6 L719.6,155.5 L860.0,118.4" fill="none" style="stroke:var(--torch)" stroke-width="2"/><circle cx="111.3" cy="316.2" r="5" style="fill:var(--torch)" class="pt"><title>fit
fit: 0.34 + 53 KB/token</title></circle><circle cx="204.9" cy="291.5" r="5" style="fill:var(--torch)" class="pt"><title>fit
fit: 0.34 + 53 KB/token</title></circle><circle cx="392.1" cy="242.0" r="5" style="fill:var(--torch)" class="pt"><title>fit
fit: 0.34 + 53 KB/token</title></circle><circle cx="485.6" cy="217.3" r="5" style="fill:var(--torch)" class="pt"><title>fit
fit: 0.34 + 53 KB/token</title></circle><circle cx="579.2" cy="192.6" r="5" style="fill:var(--torch)" class="pt"><title>fit
fit: 0.34 + 53 KB/token</title></circle><circle cx="719.6" cy="155.5" r="5" style="fill:var(--torch)" class="pt"><title>fit
fit: 0.34 + 53 KB/token</title></circle><circle cx="860.0" cy="118.4" r="5" style="fill:var(--torch)" class="pt"><title>fit
fit: 0.34 + 53 KB/token</title></circle><circle cx="111.3" cy="212.2" r="5" style="fill:var(--c16)" class="pt"><title>16 GB card, process VRAM peak
90 f: process VRAM peak 15.7 GiB on a 16 GB card (TE matrix A, 25 steps)</title></circle><circle cx="392.1" cy="213.8" r="5" style="fill:var(--c16)" class="pt"><title>16 GB card, process VRAM peak
702 f: process VRAM peak 15.5 GiB on a 16 GB card (line sweep, 2 steps)</title></circle><circle cx="392.1" cy="213.8" r="5" style="fill:var(--c16)" class="pt"><title>16 GB card, process VRAM peak
702 f: process VRAM peak 15.5 GiB on a 16 GB card (line sweep, 2 steps)</title></circle><circle cx="111.3" cy="157.5" r="5" style="fill:var(--c24)" class="pt"><title>24 GB card, process VRAM peak
90 f: process VRAM peak 23.0 GiB on a 24 GB card (TE matrix A, 25 steps)</title></circle><circle cx="392.1" cy="153.8" r="5" style="fill:var(--c24)" class="pt"><title>24 GB card, process VRAM peak
702 f: process VRAM peak 23.5 GiB on a 24 GB card (line sweep, 2 steps)</title></circle><circle cx="392.1" cy="154.5" r="5" style="fill:var(--c24)" class="pt"><title>24 GB card, process VRAM peak
702 f: process VRAM peak 23.4 GiB on a 24 GB card (line sweep, 2 steps)</title></circle><circle cx="485.6" cy="154.5" r="5" style="fill:var(--c24)" class="pt"><title>24 GB card, process VRAM peak
906 f: process VRAM peak 23.4 GiB on a 24 GB card (line sweep, 2 steps)</title></circle><circle cx="392.1" cy="141.0" r="5" style="fill:var(--c31)" class="pt"><title>31 GB card, process VRAM peak
702 f: process VRAM peak 25.2 GiB on a 31 GB card (v014: bank, fast-disk, balloon 63)</title></circle><circle cx="392.1" cy="75.8" r="5" style="fill:var(--c32)" class="pt"><title>32 GB card, process VRAM peak
702 f: process VRAM peak 33.9 GiB on a 32 GB card (v012: planner+cap card)</title></circle><circle cx="392.1" cy="75.0" r="5" style="fill:var(--c32)" class="pt"><title>32 GB card, process VRAM peak
702 f: process VRAM peak 34.0 GiB on a 32 GB card (v011: normal mode, no fast-disk)</title></circle><rect x="70" y="388" width="12" height="12" rx="2" style="fill:var(--torch)"/><text x="86" y="398" class="t3">torch activation peak (in-forward, any card)</text><rect x="418" y="388" width="12" height="12" rx="2" style="fill:var(--torch)"/><text x="434" y="398" class="t3">fit</text><rect x="479" y="388" width="12" height="12" rx="2" style="fill:var(--c16)"/><text x="495" y="398" class="t3">16 GB card, process VRAM peak</text><rect x="722" y="388" width="12" height="12" rx="2" style="fill:var(--c24)"/><text x="738" y="398" class="t3">24 GB card, process VRAM peak</text><rect x="965" y="388" width="12" height="12" rx="2" style="fill:var(--c31)"/><text x="981" y="398" class="t3">31 GB card, process VRAM peak</text><rect x="1208" y="388" width="12" height="12" rx="2" style="fill:var(--c32)"/><text x="1224" y="398" class="t3">32 GB card, process VRAM peak</text><text x="70" y="392" class="t4">fit: torch peak = 0.34 GiB + 53 KB per packed token (K/V 28.7 + x 10.8 + residents/chunks). Points not rendered are listed in the table.</text></svg><svg viewBox="0 0 900 400" role="img" aria-label="Host RAM (process RSS peak): add your own OS + ComfyUI headroom"><text x="70" y="18" class="t1">Host RAM (process RSS peak): add your own OS + ComfyUI headroom</text><line x1="70" x2="860" y1="330.0" y2="330.0" class="grid"/><text x="62" y="334.0" text-anchor="end" class="t3">0</text><line x1="70" x2="860" y1="282.9" y2="282.9" class="grid"/><text x="62" y="286.9" text-anchor="end" class="t3">11</text><line x1="70" x2="860" y1="235.7" y2="235.7" class="grid"/><text x="62" y="239.7" text-anchor="end" class="t3">22</text><line x1="70" x2="860" y1="188.6" y2="188.6" class="grid"/><text x="62" y="192.6" text-anchor="end" class="t3">33</text><line x1="70" x2="860" y1="141.4" y2="141.4" class="grid"/><text x="62" y="145.4" text-anchor="end" class="t3">44</text><line x1="70" x2="860" y1="94.3" y2="94.3" class="grid"/><text x="62" y="98.3" text-anchor="end" class="t3">55</text><line x1="70" x2="860" y1="47.1" y2="47.1" class="grid"/><text x="62" y="51.1" text-anchor="end" class="t3">66</text><text x="70.0" y="346" text-anchor="middle" class="t3">0</text><text x="70.0" y="360" text-anchor="middle" class="t4">0k</text><text x="204.9" y="346" text-anchor="middle" class="t3">294</text><text x="204.9" y="360" text-anchor="middle" class="t4">90k</text><text x="392.1" y="346" text-anchor="middle" class="t3">702</text><text x="392.1" y="360" text-anchor="middle" class="t4">214k</text><text x="485.6" y="346" text-anchor="middle" class="t3">906</text><text x="485.6" y="360" text-anchor="middle" class="t4">276k</text><text x="579.2" y="346" text-anchor="middle" class="t3">1110</text><text x="579.2" y="360" text-anchor="middle" class="t4">339k</text><text x="719.6" y="346" text-anchor="middle" class="t3">1416</text><text x="719.6" y="360" text-anchor="middle" class="t4">432k</text><text x="860.0" y="346" text-anchor="middle" class="t3">1722</text><text x="860.0" y="360" text-anchor="middle" class="t4">525k</text><text x="465" y="376" text-anchor="middle" class="t3">frames at 1376x768 (packed tokens below)</text><text transform="rotate(-90)" x="-180" y="16" text-anchor="middle" class="t3">GB</text><line x1="70" x2="860" y1="192.9" y2="192.9" class="rule" style="stroke:var(--c32)"/><text x="864" y="196.9" class="t3">32 GB</text><line x1="70" x2="860" y1="55.7" y2="55.7" class="rule" style="stroke:var(--c24)"/><text x="864" y="59.7" class="t3">64 GB</text><circle cx="111.3" cy="268.7" r="5" style="fill:var(--single)" class="pt"><title>single T2VA pass, --fast-disk
90 f: RSS 14.3 GB, single pass, 16 GB card, 64 GB machine (TE matrix A, 25 steps)</title></circle><circle cx="111.3" cy="289.7" r="5" style="fill:var(--single)" class="pt"><title>single T2VA pass, --fast-disk
90 f: RSS 9.4 GB, single pass, 24 GB card, 64 GB machine (TE matrix A, 25 steps)</title></circle><circle cx="392.1" cy="212.6" r="5" style="fill:var(--single)" class="pt"><title>single T2VA pass, --fast-disk
702 f: RSS 27.4 GB, single pass, 16 GB card, 32 GB machine (line sweep, 2 steps)</title></circle><circle cx="392.1" cy="212.6" r="5" style="fill:var(--single)" class="pt"><title>single T2VA pass, --fast-disk
702 f: RSS 27.4 GB, single pass, 16 GB card, 64 GB machine (line sweep, 2 steps)</title></circle><circle cx="392.1" cy="219.4" r="5" style="fill:var(--single)" class="pt"><title>single T2VA pass, --fast-disk
702 f: RSS 25.8 GB, single pass, 24 GB card, 32 GB machine (line sweep, 2 steps)</title></circle><circle cx="392.1" cy="219.4" r="5" style="fill:var(--single)" class="pt"><title>single T2VA pass, --fast-disk
702 f: RSS 25.8 GB, single pass, 24 GB card, 64 GB machine (line sweep, 2 steps)</title></circle><circle cx="485.6" cy="202.3" r="5" style="fill:var(--single)" class="pt"><title>single T2VA pass, --fast-disk
906 f: RSS 29.8 GB, single pass, 24 GB card, 32 GB machine (line sweep, 2 steps)</title></circle><circle cx="392.1" cy="157.7" r="5" style="fill:var(--pipe)" class="pt"><title>full de-rope pipeline, --fast-disk
702 f: RSS 40.2 GB, de-rope pipeline, 31 GB card (v014: bank, fast-disk, balloon 63)</title></circle><circle cx="392.1" cy="174.0" r="5" style="fill:var(--pipe)" class="pt"><title>full de-rope pipeline, --fast-disk
702 f: RSS 36.4 GB, de-rope pipeline, 32 GB card (v012: planner+cap card)</title></circle><circle cx="392.1" cy="75.4" r="5" style="fill:var(--nofd)" class="pt"><title>de-rope pipeline, no --fast-disk
702 f: RSS 59.4 GB WITHOUT --fast-disk (v011: normal mode, no fast-disk)</title></circle><rect x="70" y="388" width="12" height="12" rx="2" style="fill:var(--single)"/><text x="86" y="398" class="t3">single T2VA pass, --fast-disk</text><rect x="313" y="388" width="12" height="12" rx="2" style="fill:var(--pipe)"/><text x="329" y="398" class="t3">full de-rope pipeline, --fast-disk</text><rect x="591" y="388" width="12" height="12" rx="2" style="fill:var(--nofd)"/><text x="607" y="398" class="t3">de-rope pipeline, no --fast-disk</text><text x="70" y="392" class="t4">RSS with --fast-disk = ComfyUI base + CPU-side intermediates (a 702-frame fp32 IMAGE is 8.9 GB); without it, add every model's CPU copy (~33 GB here).</text></svg>
<h2 style="font-size:16px">Points</h2><table><thead><tr><th>frames</th><th>tokens</th><th>card</th><th>machine</th><th>graph</th><th>torch peak</th><th>VRAM peak</th><th>RSS</th><th>note</th></tr></thead><tbody><tr><td>90</td><td>28k</td><td>16 GB</td><td>64 GB</td><td>single pass</td><td>1.8</td><td>15.7</td><td>14.3</td><td>TE matrix A, 25 steps</td></tr><tr><td>90</td><td>28k</td><td>24 GB</td><td>64 GB</td><td>single pass</td><td>1.8</td><td>23.0</td><td>9.4</td><td>TE matrix A, 25 steps</td></tr><tr><td>702</td><td>214k</td><td>16 GB</td><td>32 GB</td><td>single pass</td><td>11.9</td><td>15.5</td><td>27.4</td><td>line sweep, 2 steps</td></tr><tr><td>702</td><td>214k</td><td>16 GB</td><td>64 GB</td><td>single pass</td><td>11.9</td><td>15.5</td><td>27.4</td><td>line sweep, 2 steps</td></tr><tr><td>702</td><td>214k</td><td>24 GB</td><td>32 GB</td><td>single pass</td><td>11.9</td><td>23.5</td><td>25.8</td><td>line sweep, 2 steps</td></tr><tr><td>702</td><td>214k</td><td>24 GB</td><td>64 GB</td><td>single pass</td><td>11.9</td><td>23.4</td><td>25.8</td><td>line sweep, 2 steps</td></tr><tr><td>702</td><td>214k</td><td>31 GB</td><td>64 GB</td><td>de-rope pipeline</td><td>11.9</td><td>25.2</td><td>40.2</td><td>v014: bank, fast-disk, balloon 63</td></tr><tr><td>702</td><td>214k</td><td>32 GB</td><td>64 GB</td><td>de-rope pipeline</td><td>11.9</td><td>33.9</td><td>36.4</td><td>v012: planner+cap card</td></tr><tr><td>702</td><td>214k</td><td>32 GB</td><td>96 GB</td><td>de-rope pipeline (no fast-disk)</td><td>11.9</td><td>34.0</td><td>59.4</td><td>v011: normal mode, no fast-disk</td></tr><tr><td>906</td><td>276k</td><td>16 GB</td><td>32 GB</td><td>single pass</td><td>-</td><td>-</td><td>-</td><td>OOM</td></tr><tr><td>906</td><td>276k</td><td>16 GB</td><td>64 GB</td><td>single pass</td><td>-</td><td>-</td><td>-</td><td>OOM</td></tr><tr><td>906</td><td>276k</td><td>24 GB</td><td>32 GB</td><td>single pass</td><td>14.9</td><td>23.4</td><td>29.8</td><td>line sweep, 2 steps</td></tr></tbody></table>
<p>Reading it: VRAM needed = resident weights (dynamic VRAM shrinks these to a few GB when it must) + the blue line + ~1 GB context; a card fits while that stays under it. RAM needed = the RSS series for your graph type + your OS and ComfyUI headroom, which you know and we do not.</p>