03 / inferenceRevised
PromptPrefillDecodeTokens
One request, 2 stages
Prompt processing and token generation have separate roles. The saved context passes from prefill to decode.
Ready
Reference: Kepler circle-packing sheet
Compute / inferenceExplain it
Explain it
with a drawing
6 new and revised studies, built in Blender. Clear at rest, with a short sequence you can play. Choose a drawing to see how it works.
07 / inferenceRevised
PromptPrefillDecodeTokens
The context handoff
A section view of the same journey. The small stack represents saved attention context, not a third serving node.
Ready
Reference: Annotated orbital diagram
10 / computeNew study
Shared poolSelected group
A field of capacity
A compact allocation study. Select a small group from the shared field. The circles represent resource units, not a chip layout.
Ready
Reference: Kepler circle-packing sheet
11 / computeNew study
StackedSeparated
Inside the working set
Layers separate to reveal a working set. An abstract memory study for a feature section, with no capacity or performance claim.
Ready
Reference: Layered Polaron symbols
12 / computeNew study
Worker linksNetwork switch
Connected capacity
A network switch joins the workers. Each pulse shows one link, not a request visiting every machine.
Ready
Reference: Annotated orbital diagram
13 / inferenceNew study
TogetherIndividual sequences
Requests in a batch
Several request sequences share a processing step. The rows separate so the individual sequences remain visible.
Ready
Reference: Geometric symbol family
7 earlier studies
01 / computeEarlier study
AssembledHighlight
The node board
Earlier hardware study, preserved for comparison.
Ready
02 / computeEarlier study
StaticInspect
The cluster
Illustrative rack arrangement. This is not live capacity.
Ready
04 / computeEarlier study
StaticInspect
The engineering view
Earlier orthographic hardware view, with the revised gallery controls.
Ready
05 / computeEarlier study
AssembledOpen
Pull out a node
Earlier rack assembly with a single open state.
Ready
06 / computeEarlier study
SuspendedPlaced
The training job
Conceptual container allocation. The slab is a visual metaphor, not hardware or a billing measurement.
Ready
08 / inferenceEarlier study
TogetherSeparated
A sharded model
Conceptual partitioning across accelerators. The layer and device counts are illustrative.
Ready
09 / inferenceEarlier study
OriginalSwitched
Change the endpoint
Earlier client and endpoint metaphor. Compatibility still depends on the selected model and client features.
Ready
What these drawings explain
The inference pair separates prefill, which processes the prompt, from decode, which generates tokens. Saved attention context passes between the stages. The pictures describe roles and a representative request, not our exact deployment or a fixed machine count.
The new forms take cues from the circular packing, layered symbols and annotated line diagrams in the linked Cosmos references. They are original studies.