Copy-ready prompt
Create a clean, vertical technical workflow infographic with a light gray background, employing a minimalist modern product chart style. Use rounded white cards, thin colored outlines, simple vector line icons, dark blue text, and dark blue connecting arrows. The overall composition is a centered, top-to-bottom flowchart containing 7 numbered main steps, and 2 parallel cache group panels branching from step 4 to step 5. A dark, thick return arrow loops from the bottom to the top on the far left. Use a clean sans-serif font, maintain ample spacing, and use soft, light colors without gradients or shadows to ensure clarity for the presentation slides. Place the step 1 card at the top center with a blue outline and a code/chat icon on the left. Title text: "1."{argument name="step 1 title" default="Chat completion request"} The subtitle is: "conversation_id + cache_salt + new suffix message". Below this is the Step 2 card, with a blue outline and document/list icon. Title: "2. Front-end Conversation Ledger". Subtitle: "Renting Same ID + Tracking Committed Messages". Below this is the Step 3 card, with a cyan outline and a database icon with a magnifying glass. Title: "3. Precise Conversation Cache Lookup". Subtitle: "conversation_id Committed Round Status". Below this is the Step 4 card, with a purple outline and branch scheduler icon. Title: "4. Scheduler Cache Attachment". Subtitle: "Setting num_computed_tokens + Attaching Committed Status". From Step 4, it branches down into two side-by-side group panels. Left group panel: a light green rounded container, titled "Full-attention KV Cache Group". Inside, two inner cards are stacked. The first inner card has a green block grid icon, titled "Committed Block Reference", and subtitled "Shared Aligned Full KV". The first internal card, the second one below, has a green layered paper icon, titled "Tail COW Copy," with the subtitle "Copying Unaligned KV Tail." A small footer is added to the bottom of the green panel: "Paging K/V Tensors for Transformer Layers." The right-hand group panel is a light purple rounded container titled "Mamba Terminal State Cache Group." It contains two stacked internal cards. The first card has a purple database/network icon, titled "Committed Terminal State," with the subtitle "Exact State at Committed Length." The second card below has a purple wavy line icon, titled "Requesting Ownership of Terminal Copy," with the subtitle "Copying SSM + Dialogue State." A small footer is added to the bottom of the purple panel: "Aligned Mode Terminal State Placement." The outputs of the two group panels are merged into the centered Step 5 card, with a blue outline and microchip icon. The title is: "5. Hybrid Model Execution." The subtitle is: "Run Only Uncached Suffixes." At the bottom of this card, two capsule-shaped labels are placed side-by-side: "Transformer Layer" and "Mamba Layer." Step 6 is placed below. Step 1: A card with a blue outline and a glittering icon. Title: "6. Decoding Assistant Token". Subtitle: "Streaming Response Token by Token". Below this is Step 7: A card with a warm yellow-orange outline and a checked database icon. Title: "7. Round of Commit Completion". Subtitle: "Publish Suspended State or Discard on Failure". Add a thick, dark blue looping arrow on the far left, leading from the upper left to Step 1 and back up from Step 7 at the bottom. Loop along this left side, placing stacked comment text near the lower half: "Reuse the committed dialog header on the next request". Add two upward dashed publish arrows from Step 7 toward the cache group panels: a green dashed arrow on the left pointing to the green cache panel, labeled "Publish New State"; and a purple dashed arrow on the right pointing to the purple cache panel, also labeled "Publish New State". Maintain the total number of 7 numbered main cards, 2 cache group panels, 4 internal cache cards, and 2 capsule labels. Maintain a vertical scale similar to the conference slides architecture diagram.
Prompt breakdown
Create a clean, vertical technical workflow infographic with a light gray background, employing a minimalist modern product chart style.
Use rounded white cards, thin colored outlines, simple vector line icons, dark blue text, and dark blue connecting arrows.
The overall composition is a centered, top-to-bottom flowchart containing 7 numbered main steps, and 2 parallel cache group panels branching from step 4 to step 5.
A dark, thick return arrow loops from the bottom to the top on the far left.









