Skip to content

Add blog post: Running a big LLM across multiple GPUs with vLLM #265

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
saiyam1814 merged 20 commits into from
Sep 1, 2026
Merged

Add blog post: Running a big LLM across multiple GPUs with vLLM #265

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
20 commits
Select commit Hold shift + click to select a range
58de3d4
Add blog post: Running a single LLM across two GPUs with vLLM
saiyam1814 Aug 18, 2026
8b1d2a5
Drop draft flag so the post renders on the Cloudflare preview
saiyam1814 Aug 18, 2026
649688c
Rework: big model across multiple GPUs, measured on 4 cards
saiyam1814 Aug 18, 2026
a6ed090
Answer the shard question: fixed at upload, and why 5-10GB is the norm
saiyam1814 Aug 18, 2026
81bd69c
Regenerate feeds for the shard section
saiyam1814 Aug 18, 2026
2eb05a8
Restructure the post into a runbook track and a deep-dive track so ac…
shkatara Aug 23, 2026
98e2502
updated benchmark and info in right places
shkatara Aug 23, 2026
204e9fa
updated benchmark and info in right places
shkatara Aug 23, 2026
08b151e
Cut the post down to the runbook so readers with a deadline are not p…
shkatara Aug 28, 2026
c793d7d
Restore the evidence and the why, and drop the part-two promises
saiyam1814 Aug 31, 2026
abb7e00
Regenerate the cake diagram as SVG in the house Excalidraw style
saiyam1814 Aug 31, 2026
964f340
Fix the review findings: a broken table, a duplicated line, and HF_HOME
saiyam1814 Aug 31, 2026
5f2a7ec
Trim the deep mechanics back out, and say what the 2 in the KV formul…
saiyam1814 Aug 31, 2026
c87a2ee
Put the cake illustration back, with the duplicated axis label removed
saiyam1814 Aug 31, 2026
00b7452
Cut duplicated log lines and repeated benchmark boilerplate
saiyam1814 Aug 31, 2026
4c68fc0
Credit Utho, drop the eviction story, merge the flags step into the c…
saiyam1814 Aug 31, 2026
7299f07
Remove docker explaination and reduce paragraph lengths
shkatara Sep 1, 2026
fe68ffe
Drop --no-enable-prefix-caching from the benchmark command; it is not…
saiyam1814 Sep 1, 2026
4a5e09d
Merge remote-tracking branch 'origin/main' into blog/two-gpu-vllm-pub
saiyam1814 Sep 1, 2026
e65253f
Re-measure all three configurations in one sitting at 640 prompts
saiyam1814 Sep 1, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
3 changes: 3 additions & 0 deletions .gitignore
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,3 +30,6 @@ yarn-error.log*
# wrangler local dev
.dev.vars
.wrangler/

# Scratch notes for unpublished posts, never rendered by lib/blog.js
/content/blog/_drafts/
290 changes: 290 additions & 0 deletions components/MoeExpertRoutingAnimation.jsx
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,290 @@
const gpus = [0, 1, 2, 3];
const PER_GPU = 32;

export default function MoeExpertRoutingAnimation() {
return (
<figure className="moeroute" aria-labelledby="moeroute-title">
<style>{`
.moeroute {
margin: 2rem 0;
border: 1px solid var(--border-subtle);
border-radius: 8px;
background:
linear-gradient(135deg, rgba(168, 85, 247, 0.1), rgba(5, 202, 255, 0.08)),
var(--bg-card);
box-shadow: var(--card-shadow);
overflow: hidden;
}

.moeroute * { box-sizing: border-box; }
.moeroute__inner { padding: clamp(1rem, 3vw, 1.5rem); }

.moeroute__eyebrow {
margin: 0 0 0.35rem;
color: var(--accent);
font-size: 0.72rem;
font-weight: 800;
text-transform: uppercase;
}

.moeroute__title {
margin: 0;
color: var(--text-primary);
font-size: clamp(1.15rem, 2.5vw, 1.55rem);
line-height: 1.2;
}

.moeroute__subtitle {
max-width: 46rem;
margin: 0.55rem 0 1.35rem;
color: var(--text-secondary);
font-size: 0.95rem;
line-height: 1.6;
}

.moeroute__stage {
display: grid;
gap: 0.8rem;
padding: 1rem;
border: 1px solid var(--border-subtle);
border-radius: 8px;
background:
linear-gradient(var(--border-subtle) 1px, transparent 1px),
linear-gradient(90deg, var(--border-subtle) 1px, transparent 1px),
color-mix(in srgb, var(--bg-elevated) 86%, transparent);
background-size: 42px 42px;
}

.moeroute__token {
padding: 0.55rem 0.7rem;
border: 1px dashed color-mix(in srgb, var(--accent) 50%, var(--border-medium));
border-radius: 8px;
background: color-mix(in srgb, var(--accent) 8%, var(--bg-card));
color: var(--text-primary);
font-size: 0.78rem;
font-weight: 900;
text-align: center;
}

.moeroute__token em {
display: block;
margin-top: 0.15rem;
color: var(--text-muted);
font-size: 0.65rem;
font-style: normal;
font-weight: 800;
}

.moeroute__router {
padding: 0.45rem;
border: 1px solid color-mix(in srgb, #a855f7 45%, var(--border-subtle));
border-radius: 6px;
background: color-mix(in srgb, #a855f7 9%, var(--bg-card));
color: color-mix(in srgb, #a855f7 85%, var(--text-primary));
font-size: 0.72rem;
font-weight: 950;
text-align: center;
}

.moeroute__gpus {
display: grid;
grid-template-columns: repeat(4, minmax(0, 1fr));
gap: 0.55rem;
}

.moeroute__gpu {
min-width: 0;
padding: 0.55rem;
border: 1px solid color-mix(in srgb, var(--c) 42%, var(--border-subtle));
border-radius: 7px;
background: color-mix(in srgb, var(--c) 6%, var(--bg-card));
}

.moeroute__gpu-head {
margin-bottom: 0.4rem;
color: var(--text-primary);
font-size: 0.72rem;
font-weight: 950;
}

.moeroute__gpu-head em {
display: block;
color: var(--text-muted);
font-size: 0.58rem;
font-style: normal;
font-weight: 800;
}

.moeroute__experts {
display: grid;
grid-template-columns: repeat(8, 1fr);
gap: 2px;
}

.moeroute__exp {
aspect-ratio: 1;
border-radius: 2px;
background: color-mix(in srgb, var(--c) 16%, var(--bg-elevated));
}

.moeroute__exp[data-hot='yes'] {
background: color-mix(in srgb, var(--c) 78%, transparent);
animation: moeFire 4s ease-in-out infinite;
animation-delay: var(--d);
}

.moeroute__count {
margin-top: 0.4rem;
color: var(--text-secondary);
font-size: 0.62rem;
font-weight: 850;
text-align: center;
}

.moeroute__note {
display: grid;
grid-template-columns: repeat(3, minmax(0, 1fr));
gap: 0.55rem;
}

.moeroute__cell {
min-width: 0;
padding: 0.65rem;
border: 1px solid var(--border-subtle);
border-radius: 7px;
background: var(--bg-card);
}

.moeroute__k {
display: block;
color: var(--text-muted);
font-size: 0.58rem;
font-weight: 900;
text-transform: uppercase;
}

.moeroute__v {
display: block;
margin-top: 0.18rem;
color: var(--accent);
font-size: 0.9rem;
font-weight: 950;
}

.moeroute__d {
display: block;
margin-top: 0.22rem;
color: var(--text-muted);
font-size: 0.63rem;
line-height: 1.35;
}

.moeroute figcaption {
padding: 0 1.5rem 1.15rem;
color: var(--text-muted);
font-size: 0.78rem;
line-height: 1.45;
}

@keyframes moeFire {
0%, 100% { opacity: 0.25; transform: scale(0.9); }
35%, 55% { opacity: 1; transform: scale(1); }
}

@media (max-width: 700px) {
.moeroute__gpus { grid-template-columns: 1fr 1fr; }
.moeroute__note { grid-template-columns: 1fr; }
}

@media (prefers-reduced-motion: reduce) {
.moeroute__exp[data-hot='yes'] {
animation-duration: 0.001ms !important;
animation-iteration-count: 1 !important;
opacity: 1;
transform: scale(1);
}
}
`}</style>
<div className="moeroute__inner">
<p className="moeroute__eyebrow">Expert routing animation</p>
<h3 className="moeroute__title" id="moeroute-title">
Why a 235B model only does 22B of work per token
</h3>
<p className="moeroute__subtitle">
Every layer of this model has 128 small expert networks, and a tiny router picks just 8 of
them for each token. The other 120 sit still. That is the whole trick of a mixture of
experts: you pay for 235B parameters in memory, but only about 22B of arithmetic per token.
</p>

<div className="moeroute__stage" role="img" aria-label="A token being routed to 8 of 128 experts spread over 4 GPUs">
<div className="moeroute__token">
one token arrives
<em>it has already been through attention for this layer</em>
</div>

<div className="moeroute__router">router scores all 128 experts, keeps the top 8</div>

<div className="moeroute__gpus">
{gpus.map((gpu) => {
// 2 of this GPU's 32 experts are picked, so 8 across 4 GPUs
const hot = [3 + gpu, 18 + ((gpu * 5) % 10)];
return (
<div
className="moeroute__gpu"
key={gpu}
style={{ '--c': gpu % 2 === 0 ? '#a855f7' : '#0098cc' }}
>
<p className="moeroute__gpu-head">
GPU {gpu}
<em>experts {gpu * PER_GPU}-{gpu * PER_GPU + PER_GPU - 1}</em>
</p>
<div className="moeroute__experts" aria-hidden="true">
{Array.from({ length: PER_GPU }, (_, i) => (
<div
className="moeroute__exp"
key={i}
data-hot={hot.includes(i) ? 'yes' : 'no'}
style={{ '--d': `${gpu * 0.25}s` }}
/>
))}
</div>
<p className="moeroute__count">2 of 32 firing</p>
</div>
);
})}
</div>

<div className="moeroute__note">
<div className="moeroute__cell">
<span className="moeroute__k">In memory</span>
<span className="moeroute__v">235B params</span>
<span className="moeroute__d">
All 128 experts per layer must be resident, which is why the model is big
</span>
</div>
<div className="moeroute__cell">
<span className="moeroute__k">Active per token</span>
<span className="moeroute__v">22B params</span>
<span className="moeroute__d">
Only the 8 chosen experts do arithmetic, so it runs like a much smaller model
</span>
</div>
<div className="moeroute__cell">
<span className="moeroute__k">What this costs you</span>
<span className="moeroute__v">a network hop</span>
<span className="moeroute__d">
With expert parallelism the token travels to whichever GPU owns its expert, then the
answer travels back
</span>
</div>
</div>
</div>
</div>
<figcaption>
Counts are from the model config: 128 experts per layer, 8 per token, 94 layers. Splitting 128
experts over 4 GPUs gives 32 each, so on average 2 experts per GPU fire for any given token.
That average is the catch, because routing is not guaranteed to be even.
</figcaption>
</figure>
);
}
Loading
Loading