rlt-loop-pipelines.gif and rlt-loop-pipelines.png are original code-drawn assets for
vLLM-RLT. No source artwork, logos, stock imagery, or external fonts are bundled.
The four design colors are white, navy, blue, and yellow (antialiasing adds shades).
Implementation/documentation baseline: v0.2.0,
c08b9d79834bb8b4ff283b15bae2c14c559f6b52.
- Stage/loop ownership and shared runtime
- Wavefront chunked prefill
- Experimental pipelined self-speculation
The README layout reference is vLLM-Omni's current README (read November 1, 2026): centered header/navigation, About overview, architecture illustration, Getting Started, contributor/community sections. Its current overview asset is static, not an existing GIF to copy.
The outer Stage boundary owns deployment/resources. StageEngine coordinates Scheduler selection/progress and Executor submission; arrows are a simplified work-flow, not a direct Scheduler-to-runner call graph. Shared model runtime and state labels summarize resources, not additional exclusive owners.
The two phases are separate execution examples, NOT a claim that speculative verification joins WCPB mixed recurrent batches. Four chunks/inputs and four loops illustrate state transitions, not measured time, throughput or capacity. Yellow marks current work; checks mark earlier completed recurrence. WCPB advances ready chunks at different depths. Decode shows the all-accepted speculative path: processing input x_i ultimately yields target output x_(i+1). Hidden states survive submissions. Rejection discards dependent tokens; no quality equivalence to block mode is asserted. Block is still the default. Pipeline is experimental and requires greedy sampling and LAST_EXITED KV; backend/model configuration limits remain in the linked documentation. Neither PD topology nor cross-host support is depicted.
Requires Python 3 and Pillow, with DejaVu Sans TTF available at the standard Linux
paths used by generate_pipelines.py (adjust FONT/BOLD constants on other platforms).
python docs/assets/runtime/generate_pipelines.py --output-dir docs/assetsOutputs: animated GIF, static WCPB poster, and a decode QA frame. The renderer uses no network or GPU. GIF frame coalescing is intentional: repeated frames are stored as longer durations to keep the README asset small. The loop is 20 seconds. Validation: load every GIF frame with Pillow, decode with FFmpeg, visually inspect both phase posters, check duration/dimensions/infinite-loop metadata.