01 / DEVICE MEMORY
Execution changes the active state.
Agents, models and memory operate on the GPU. The runtime applies an admitted transition to its authoritative state and prepares the record required to preserve that change.
02 / System architecture
AOTX-1 / ARCHITECTURE
A system on
the device.
One runtime. Defined responsibilities.
AOTX places authoritative state beside the device work that changes it. Around that centre, host clients carry input and results, while disk processes maintain the recoverable record. Each boundary gives an operation a defined path from admission to durable completion.
ARCHITECTURE / PERSISTENCE PATH
01 / STATE JOURNEYILLUSTRATED EXAMPLE
The outward persistence path.
Follow a state change from device memory to the disk journal. Each stage publishes something the next can safely consume; the final acknowledgement establishes the recoverable boundary.
DEVICEAUTHORITATIVE STATE
01EXECUTION
GPU state
Agents · memory · execution
02PUBLICATION
Device records
An ordered account of work
DEVICE BOUNDARY
HOSTSTORAGE CONNECTION
03TRANSFER
Host ring
Published blocks cross the seam
04PERSISTENCE
Disk journal
A lagging, recoverable record
02 / RECORD PUBLICATION
The transition receives an ordered record.
Device producers write a record before publishing its completion. Tick and sequence identifiers retain the order of work, allowing consumers to distinguish a complete record from one still being written.
03 / THE DEVICE BOUNDARY
Published blocks cross into host memory.
The flush transfers records into a pinned host ring in bounded blocks. This gives the disk drain a complete unit to consume while state authority remains with the device.
04 / PERSISTENCE
The journal makes the change recoverable.
The drain writes journal segments and checks their integrity. Committed data supplies the recovery path; durable acknowledgement identifies the completed disk position behind live execution.
Stage 1 of 4: GPU state
Device progress, record publication and durable storage are separate positions in the same ordered history.
DEVICE / JOURNAL
EXECUTION / STATE / CONTROL
01
Device execution
Execution placement
Inference, agent work, memory selection and state transitions execute in CUDA. Host glue loads and launches that work, while clients and disk processes connect it to the rest of the machine. This arrangement keeps the decisions that change runtime state with the device that owns it.
A finite execution step
The tick graph provides ordered, bounded work. Device queues select the admitted batch, and resource checks account for its slots, context and memory. State remains available between executions, so a persistent conversation can wait without holding an active execution slot.Read the execution chapter »
EXECUTION / CUDA
02
Memory and persistence
A recoverable history
GPU memory holds the active state. Published records cross device and host rings to reach the journal; complete CCIR generations bind the saved state to its required assets. The disk representation follows the device, and its acknowledgement establishes how far recovery can proceed.
Capacity as part of operation
The runtime accounts for finite publication capacity before affected work advances. Optional cold-memory offload changes the residence of eligible payloads while preserving GPU ownership of the catalogue and recall decisions. Storage therefore has defined roles in both recovery and selected operating modes.Read the memory chapter »
MEMORY / CONTINUITY
03
Control surfaces
Local and remote control
The console and graphical client expose local runtime operations. A separate gateway supplies authenticated network transport. Across these surfaces, the device applies admission and scope rules to the requested work, maintaining one authority for the state clients observe.
Two service contracts
Supplied-context inference serves an application’s submitted messages for an individual request. Continuing shared service retains conversation identities, scoped memory and operation receipts across connections. The choice determines where history belongs and how a client resumes its work.Read the interface chapter »
ACCESS / CONTROL
04
Recovery
Continue the recorded account
Recovery reconstructs logical state from validated committed records. Earlier decisions retain their saved identities and results, including the evidence selected for interpretation. A restart restores that history before new work begins.
Carry a coherent generation
CCIR binds the checkpoint, replay representation and required assets into one saved unit. A compatible destination reconstructs the state and establishes its local execution resources. Completed operations retain their receipts; operations admitted without a saved result are explicitly interrupted.Read the portability chapter »
JOURNAL / RESTORE
SYSTEM REFERENCE / TERMINOLOGY
05
Architectural terms
- Tick
- One execution of the tick graph. Tick and sequence identifiers order device work and its recorded transitions.
- Authoritative record
- An input, decision or result whose recorded representation is required to reconstruct the system’s logical state.
- Derived view
- A display or export produced from authoritative state, such as the presentation a client uses to observe the runtime.
- Durable generation
- A complete acknowledged file boundary that binds the state and dependencies required for a supported restore.
AOTX-1 / Architecture▰ ▰ ▰