Skip to content

Accurate, fast FP4 KV-cache decoding through algorithm-system co-design.

docs: add Overview section and rename figure heading

DD-DuDa48c835a6044c
diff --git a/README.md b/README.md
index 3e6bc79..495612e 100644
--- a/README.md
+++ b/README.md
@@ -7,6 +7,8 @@ BitDecoding-FP4 delivers accurate, high-performance LLM decoding through
native FP4 attention kernels. It integrates into vLLM as an out-of-tree `CUSTOM`
backend.
+## Overview
+
![BitDecoding-FP4 system overview: accuracy-aware FP4 quantization, native FP4 decoding kernels, and a custom vLLM backend with paged FP4 KV storage and separate BF16 sink and recent-token buffers.](imgs/overview.svg)
## Loss-less Accuracy
diff --git a/imgs/overview.svg b/imgs/overview.svg
index a560877..ee25fc4 100644
--- a/imgs/overview.svg
+++ b/imgs/overview.svg
@@ -51,7 +51,7 @@
<!-- Title and supported native FP4 formats. -->
<rect x="56" y="31" width="5" height="55" rx="2.5" fill="url(#accent)"/>
- <text x="78" y="55" font-size="34" font-weight="700" letter-spacing="-1">BitDecoding-FP4</text>
+ <text x="78" y="55" font-size="34" font-weight="700" letter-spacing="-1">Overview</text>
<text x="79" y="85" font-size="20" class="muted">Native FP4 Decoding, from quantization to serving</text>
<text x="1384" y="45" text-anchor="end" font-size="14" font-weight="700" letter-spacing="2" class="muted">ALGORITHM–SYSTEM CO-DESIGN</text>
<rect x="1149" y="61" width="111" height="30" rx="15" fill="#edf3fa"/>