lepton/website
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Initial website deployment
All checks were successful
Deploy Website / build-and-deploy (push) Successful in 1m44s
Details

Browse Source
This commit is contained in:
Akshay Kolli
2026-07-02 00:28:25 -07:00
commit e240b99cb2
66 changed files with 2774 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

BIN
static/images/2_qwen36_tokens_per_kwh/author_nvidia_mtp_x6_215w_report.png Normal file
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 313 KiB

BIN
static/images/2_qwen36_tokens_per_kwh/fig1_tokens_per_day.png Normal file
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 64 KiB

BIN
static/images/2_qwen36_tokens_per_kwh/fig3_m3_break_even_power.png Normal file
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 80 KiB

BIN
static/images/2_qwen36_tokens_per_kwh/fig4_m3_power_sensitivity.png Normal file
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 150 KiB

49
static/images/2_qwen36_tokens_per_kwh/fig5_autoregressive_decode.svg Normal file
View File

@@ -0,0 +1,49 @@
<svg xmlns="http://www.w3.org/2000/svg" width="1200" height="520" viewBox="0 0 1200 520" role="img" aria-labelledby="title desc">
<title id="title">Autoregressive decoding</title>
<desc id="desc">A simplified diagram showing one target model pass producing one accepted token, repeated sequentially.</desc>
<defs>
<style>
.bg { fill: #f7f4ed; }
.panel { fill: #fffdf7; stroke: #cfc8ba; stroke-width: 2; }
.ink { fill: #181612; font-family: ui-sans-serif, system-ui, -apple-system, BlinkMacSystemFont, "Segoe UI", sans-serif; }
.muted { fill: #6b6254; font-family: ui-monospace, "SFMono-Regular", Consolas, monospace; }
.box-title { fill: #181612; font-family: ui-sans-serif, system-ui, -apple-system, BlinkMacSystemFont, "Segoe UI", sans-serif; text-anchor: middle; }
.box-sub { fill: #6b6254; font-family: ui-monospace, "SFMono-Regular", Consolas, monospace; text-anchor: middle; }
.accent { fill: #0f6f77; }
.arrow-line { stroke: #5c554a; stroke-width: 3; fill: none; stroke-linecap: round; stroke-linejoin: round; }
.arrow-head { fill: #5c554a; }
.dash { stroke: #b8ad9b; stroke-width: 2; stroke-dasharray: 8 8; fill: none; }
</style>
</defs>
<rect class="bg" width="1200" height="520" rx="24"/>
<text class="ink" x="56" y="72" font-size="34" font-weight="650">Autoregressive decode</text>
<text class="muted" x="56" y="112" font-size="18">one target model pass gives one accepted token</text>
<rect class="panel" x="72" y="188" width="220" height="118" rx="18"/>
<text class="box-title" x="182" y="238" font-size="24" font-weight="650">context</text>
<text class="box-sub" x="182" y="274" font-size="17">prompt + KV cache</text>
<path class="arrow-line" d="M292 247 L385 247"/>
<path class="arrow-head" d="M410 247 L385 234 L385 260 Z"/>
<rect class="panel" x="430" y="168" width="280" height="158" rx="20"/>
<text class="box-title" x="570" y="228" font-size="25" font-weight="650">target model</text>
<text class="box-sub" x="570" y="265" font-size="17">full forward pass</text>
<rect class="accent" x="485" y="287" width="170" height="8" rx="4"/>
<path class="arrow-line" d="M710 247 L810 247"/>
<path class="arrow-head" d="M835 247 L810 234 L810 260 Z"/>
<rect class="panel" x="855" y="188" width="166" height="118" rx="18"/>
<text class="box-title" x="938" y="238" font-size="24" font-weight="650">token</text>
<text class="box-sub" x="938" y="274" font-size="17">+1</text>
<path class="arrow-line" d="M1021 247 C1110 247 1102 397 940 417 C650 452 290 427 210 333"/>
<path class="arrow-head" d="M188 309 L214 320 L201 346 Z"/>
<path class="dash" d="M430 344 L710 344"/>
<text class="muted" x="458" y="382" font-size="16">the GPU pays this cost one token at a time</text>
<text class="muted" x="454" y="456" font-size="17">append token, then run again</text>
</svg>
Side by Side

After

Width:  |  Height:  |  Size: 2.8 KiB

73
static/images/2_qwen36_tokens_per_kwh/fig6_mtp_decode.svg Normal file
View File

@@ -0,0 +1,73 @@
<svg xmlns="http://www.w3.org/2000/svg" width="1200" height="560" viewBox="0 0 1200 560" role="img" aria-labelledby="title desc">
<title id="title">MTP speculative decoding</title>
<desc id="desc">A simplified diagram showing a multi token prediction path proposing draft tokens and the target model verifying them together.</desc>
<defs>
<style>
.bg { fill: #f7f4ed; }
.panel { fill: #fffdf7; stroke: #cfc8ba; stroke-width: 2; }
.ink { fill: #181612; font-family: ui-sans-serif, system-ui, -apple-system, BlinkMacSystemFont, "Segoe UI", sans-serif; }
.muted { fill: #6b6254; font-family: ui-monospace, "SFMono-Regular", Consolas, monospace; }
.box-title { fill: #181612; font-family: ui-sans-serif, system-ui, -apple-system, BlinkMacSystemFont, "Segoe UI", sans-serif; text-anchor: middle; }
.box-sub { fill: #6b6254; font-family: ui-monospace, "SFMono-Regular", Consolas, monospace; text-anchor: middle; }
.accent { fill: #0f6f77; }
.warm { fill: #c06922; }
.cool { fill: #315f9f; }
.arrow-line { stroke: #5c554a; stroke-width: 3; fill: none; stroke-linecap: round; stroke-linejoin: round; }
.arrow-head { fill: #5c554a; }
.thin { stroke: #b8ad9b; stroke-width: 2; fill: none; }
</style>
</defs>
<rect class="bg" width="1200" height="560" rx="24"/>
<text class="ink" x="56" y="72" font-size="34" font-weight="650">MTP speculative decode</text>
<text class="muted" x="56" y="112" font-size="18">draft several tokens, verify them, keep the accepted prefix</text>
<rect class="panel" x="72" y="218" width="220" height="118" rx="18"/>
<text class="box-title" x="182" y="268" font-size="24" font-weight="650">context</text>
<text class="box-sub" x="182" y="304" font-size="17">prompt + KV cache</text>
<path class="arrow-line" d="M292 277 C342 277 342 196 395 196"/>
<path class="arrow-head" d="M420 196 L395 183 L395 209 Z"/>
<path class="arrow-line" d="M292 277 C342 277 342 374 395 374"/>
<path class="arrow-head" d="M420 374 L395 361 L395 387 Z"/>
<rect class="panel" x="440" y="138" width="270" height="116" rx="20"/>
<text class="box-title" x="575" y="188" font-size="24" font-weight="650">MTP lookahead</text>
<text class="box-sub" x="575" y="224" font-size="17">propose draft tokens</text>
<rect class="panel" x="440" y="316" width="270" height="116" rx="20"/>
<text class="box-title" x="575" y="366" font-size="24" font-weight="650">target model</text>
<text class="box-sub" x="575" y="402" font-size="17">verify in one pass</text>
<path class="arrow-line" d="M710 196 L738 196"/>
<path class="arrow-head" d="M763 196 L738 183 L738 209 Z"/>
<g transform="translate(782 170)">
<rect class="warm" x="0" y="0" width="54" height="54" rx="12"/>
<rect class="warm" x="66" y="0" width="54" height="54" rx="12"/>
<rect class="warm" x="132" y="0" width="54" height="54" rx="12"/>
<rect class="warm" x="198" y="0" width="54" height="54" rx="12"/>
<rect class="warm" x="264" y="0" width="54" height="54" rx="12"/>
<rect class="warm" x="330" y="0" width="54" height="54" rx="12"/>
</g>
<text class="muted" x="874" y="252" font-size="17">draft block</text>
<path class="arrow-line" d="M710 374 L748 374"/>
<path class="arrow-head" d="M773 374 L748 361 L748 387 Z"/>
<path class="arrow-line" d="M1030 232 L1030 340"/>
<path class="arrow-head" d="M1030 365 L1017 340 L1043 340 Z"/>
<rect class="panel" x="790" y="356" width="344" height="102" rx="18"/>
<rect class="accent" x="820" y="384" width="50" height="50" rx="11"/>
<rect class="accent" x="880" y="384" width="50" height="50" rx="11"/>
<rect class="accent" x="940" y="384" width="50" height="50" rx="11"/>
<rect class="cool" x="1000" y="384" width="50" height="50" rx="11"/>
<rect class="cool" x="1060" y="384" width="50" height="50" rx="11"/>
<path class="thin" d="M820 448 L990 448"/>
<text class="muted" x="820" y="492" font-size="16">accepted prefix, retry on mismatch</text>
<text class="muted" x="720" y="530" font-size="16">best case, one pass advances several tokens</text>
</svg>
Side by Side

After

Width:  |  Height:  |  Size: 4.0 KiB