Diffusion Language Model KV-Cache Quantization Diagram

Overview

A three-panel pipeline diagram of KV-cache quantization for a diffusion language model. It is an editable PowerPoint diagram adapted from the original paper figure under CC BY 4.0.

What the diagram shows

(a) Process of block diffusion decoding: a prompt + gold-answer token strip and a legend (prompt, mask, decoded and quantized tokens, token confidence); an initial assignment trajectory of hatched mask cells being decoded over intermediate steps t = 1 … 0 with confidence mini-bars against a 0.5 threshold → an iterative-refinement loop (Build x → Score s[i, t] → Min-Cost Max-Flow → Update Trajectory) → a refined trajectory with a 0.8 threshold and a new decoding order. (b) Offline policy construction: a candidates-protect heat map (per group × tokens, with a score colour bar) feeds a layer-wise token-protection policy table (Layer / Policy / Num, with ● block-init attention, ▲ quant error and ◆ attention × error markers). (c) Online KV cache quantization under a bit budget: 1. KV-cache group for each layer, 2. protect important tokens (top-k, outlined cells), 3. quantize (mixed precision, magenta cells), then a diffusion LM decoder → output sequence. Every box, arrow, label and math term is a native shape you can recolor, resize and relabel for your own slides, posters or papers.

File format

The download contains .pptx (1 slide, 17.54 × 9.56 in; opens in PowerPoint, Keynote, WPS and LibreOffice Impress; Google Slides works too, but complex freeforms may shift slightly), .svg and .drawio. Includes preview.png (full resolution) and a 1200 px web preview.

Shapes and editing

The file has 631 native editable shapes: 343 auto-shapes, 71 text boxes, 217 freeforms, 0 lines/connectors. No embedded images. Everything is vector. You can edit all text and math (t = 1, t = 0, s[i, t]), every token cell of the strips and trajectories, the hatched mask cells, the confidence mini-bars with dashed thresholds, the curved step arrows, the iterative-refinement loop boxes and arrows, the heat-map cells and score colour bar, the policy table cells and ● ▲ ◆ markers, the quantization rows with outlined protected cells, the decoder and output boxes, the grey block arrows and all dashed panels.

Embedded images and fonts

This file has no embedded images; everything is vector. Body text uses Calibri and formulas use Times New Roman. Calibri ships with Microsoft Office and Times New Roman is preinstalled on Windows and macOS (LibreOffice substitutes the metric-compatible Carlito and Liberation Serif). Sub- and superscripts are real formatted text runs, not equation objects. The SVG keeps text as text; the draw.io file is a shape-by-shape conversion and may need small nudges. This is a faithful redraw for reuse, not the authors' original source file.