Skip to content

Architecture

Fabrizio Degni edited this page Nov 15, 2026 · 6 revisions

Architecture and execution contracts

Target architecture revision 2.0. NodeLift connects useful memory, compute and services through cloud orchestration and heterogeneous peer workers. Its purpose is to expand access to demanding workloads and extend existing device capability and useful life. The mission defines that destination.

Cloud authority and peer execution

The global Hive manages account and device identity, network policy, regional discovery and a unified resource directory. Regional hubs admit workloads, manage queues and leases, select compatible workers, account for contribution and coordinate recovery. Workers execute assigned work within their operating-system and owner policies.

Computers running macOS, Windows or Linux, workstations, smartphones, tablets and eligible consoles all participate as peers. A Mac is a computer worker. Device type does not create a privileged coordinator or a mandatory cloud-management bridge. A workload may assign a worker a target-verification or partition role according to its capabilities.

flowchart TB
    C[Client and control console] --> G[Global cloud Hive]
    G -->|identity policy and discovery| A[Regional hub A]
    G -->|identity policy and discovery| B[Regional hub B]
    A <-->|approved handover and residency policy| B
    A -->|queues placement leases and recovery| W1[Computer workers including Mac]
    A -->|same peer work contract| W2[Workstation workers]
    A -->|same peer work contract| W3[Smartphone and tablet workers]
    B -->|same peer work contract| W4[Eligible console workers]
    B -->|same peer work contract| W5[Other compatible peer workers]
    W1 <-->|workload data and model activations| W2
    W3 <-->|authorised task data| W2
Loading

Regional placement follows home-region and data-residency rules, local-time operating windows, supported operators, numerical compatibility, available memory and measured link costs. Control-plane authority and workload data paths are separate responsibilities. The interactive architecture explorer provides the four target-system diagrams.

Four execution routes

Route Resource purpose Result and failure contract
Required model partitions Place nonempty model partitions, weights and state across compatible workers to extend resident capacity beyond one device Required state needs a compatible checkpoint and sufficient capacity for restoration, or an explicit interrupted state
Independent requests Route a complete independent job to a suitable worker that can execute its model Fenced ownership protects result acceptance; eligible independent work can be retried
Verified draft assistance Assign optional candidate generation and target verification by workload capability Only target-verified output commits; optional assistance can be dropped according to its cost and fallback contract
Complementary services Compose retrieval, embeddings, preprocessing and specialised compute through adapters Each service declares inputs, data policy, resource budgets, dependencies and verifiable results

Independent jobs and draft assistance have different memory contracts from required model partitions. Separate full models do not become one shared-memory model. A hardware abstraction adapter must expose an executable, OS-approved backend and validate the workload it accepts.

Resource accounting and placement

Physical system RAM, shared CPU/GPU memory, dedicated accelerator memory and application headroom retain explicit scopes. Shared physical memory is counted once. Model placement also accounts for runtime overhead and KV state. Installed hardware inventory alone does not determine usable model capacity.

Compute rates carry precision, workload profile, backend, timing scope, coverage and observation time. A sum of qualified isolated worker rates describes a resource envelope. Actual delivered service uses a common observation window and its own workload scope. GFLOP/s cannot be converted directly to LLM token throughput. See the compute metrics contract.

Placement compares local and remote execution costs, including transfers, queue, setup, computation, verification and recovery. It keeps work on a local peer when remote execution offers no useful benefit. The console distinguishes available resources, measured useful contribution and bottlenecks at worker, regional hub and Hive level.

Continuity and platform boundaries

Persistent device identity is separate from scoped work authorisation. Fenced leases identify the current owner of each attempt, and result acceptance rejects stale work. Independent retry, cloud-control recovery and required model-state recovery remain separate contracts. Regional handover preserves authorisation, residency and compatible state.

Mobile and console execution follows application quotas, thermal conditions, foreground and background rules, platform signing and distribution requirements. A worker pauses or stops when its operating-system lifecycle or owner policy requires it. The architecture does not depend on unrestricted console daemons, arbitrary system drivers or universal acceleration.

The detailed architecture, cloud roles and capability census define these design contracts. For actual available versions and platform requirements, consult Downloads and the mobile package guide separately from this target-system overview.

Direct Hive LLM pilot

Hive now owns bounded complete-model generation without a mandatory Mac coordinator. See the LLM engine, setup, physical Linux measurements and remaining mobile gates. Complete-model placement does not pool device memories.

Clone this wiki locally