Mobile and edge AI — SSOT

This page is the single place for how Vox treats Android / iOS / browser relative to desktop Mens training, Ollama, mens coordination, and GPU advertisement. It complements Mens training SSOT, mens SSOT, and unified orchestration.

Non-goals (near term)

• Running Ollama or a full Ollama-compatible daemon on stock consumer phones.
• Running vox mens train with Candle QLoRA or Burn LoRA on the phone (Rust + wgpu/Candle stacks are workstation targets).
• Promising end-to-end LLM LoRA fine-tuning on-device with the same maturity as workstation vox mens train (industry runtimes still steer operators toward train off-device, infer on-device for LLMs).

Industry context (2025–2026)

• On-device LLM inference: Google LiteRT-LM is the cross-platform direction for Android, iOS, web, and desktop with hardware acceleration; see LiteRT-LM and LLM inference (AI Edge). Older MediaPipe-only flows are being superseded; plan migrations against current AI Edge docs.
• LoRA / adapters: Practical path is fine-tune on a workstation or cloud, then ship base + adapter (or converted bundle) -> the device. LiteRT LLM LoRA on-device is still integration-heavy (see discussion in LiteRT issue #1420).
• Web tier: WebGPU helps browser-side compute but is not universal (OS version, browser policy, and security modes can disable it). Treat PWA / WebGPU as an optional tier, not the only mobile story.

Vox tiers

Mobile support boundary (normative)

Mobile support is split across distinct product surfaces. Do not collapse them into one claim.

This SSOT does not define Vox as a replacement for Kotlin or Swift. The recommended product path is:

• Vox for browser-first full-stack app generation.
• Remote phone management for planning, editing, validation, and orchestration against a remote Vox host.
• Native mobile only where thin wrappers or inference SDK integration are the right boundary.

Training pathway for mobile (mobile_edge)

1. On a GPU or CPU workstation, run:

   vox mens train … --deployment-target mobile_edge

   or --preset mobile_edge (implies the same deployment target).

2. The execution planner applies gates: bounded seq_len / rank / batch_size, no --qlora-require-full-proxy-stack, and --device cpu is required so adapters are trained without binding to a desktop-only GPU stack (see planner errors for the exact message).

3. Artifacts (adapter_schema_v3, training_manifest.json) record training_deployment_target and an operator note pointing here and to HF finetune capability matrix. Conversion to LiteRT / Core ML / TFLite is out of tree until a supported exporter exists.

Canonical trainer documentation remains mens-training.md.

Export contract (out of tree)

Training emits artifacts that are consumed by an exporter outside this repository until a first supported exporter lands in-tree.

Inputs (already produced by the Mens pipeline)

• adapter_schema_v3
• training_manifest.json
• training_deployment_target (for example mobile_edge)

Outputs

TBD by the chosen on-device runtime (for example LiteRT bundle layout, Core ML, or vendor-specific packages).

Definition of done (first supported exporter)

• Documented output format(s) and a version pin for the target runtime.
• Reproducible build: same inputs and toolchain version produce artifacts described by a checksum or manifest.
• training_manifest.json (or its successor) records exporter version and output checksums (or equivalent integrity fields).
• Documented validation step (for example a dry-run load in the target runtime, or a future vox mens verify subcommand when one exists).

Further context: HF finetune capability matrix, Mens training SSOT.

Inference profiles (no Ollama on loopback for mobile)

Desktop MCP and CLI default to a local Ollama URL for workstation use only. Mobile apps should set an explicit profile (environment) so routing does not assume localhost:11434.

vox-mcp HTTP inference: local Ollama calls and cloud→Ollama fallback are enabled only when the profile is desktop_ollama or lan_gateway. Other profiles skip Ollama probes and reject ProviderType::Ollama with a clear error unless you switch profile or model.

Registry: Environment variables (SSOT) (VOX_INFERENCE_PROFILE).

Mens and GPU / NPU advertisement

Mens nodes embed TaskCapabilityHints. CUDA and Metal are not sufficient for Android Vulkan phones or NPU classes.

• Legacy: VOX_MESH_ADVERTISE_GPU=1 still sets gpu_cuda (workstation-oriented; unchanged for backward compatibility).
• Additive: VOX_MESH_ADVERTISE_VULKAN, VOX_MESH_ADVERTISE_WEBGPU, VOX_MESH_ADVERTISE_NPU (each 1 / true) set the matching capability flags.
• Class label: VOX_MESH_DEVICE_CLASS — optional free-form hint (server, desktop, mobile, browser, …) stored in TaskCapabilityHints.device_class.

See mens SSOT for the full VOX_MESH_* table.

GPU probing (Mens vs mens)

• Mens training uses probe_gpu for VRAM heuristics. Overrides: VOX_GPU_MODEL, VOX_GPU_VRAM_MB. Windows: wmic; Linux: best-effort nvidia-smi / lspci. Android / iOS: no in-crate probe — the host app should set env overrides or pass capabilities into mens JSON.
• Mens does not require Mens; capability flags come from env + host as above.

Related

• Cross-platform Vox — lanes & Docker matrix (SSOT) — script worker vs app vs mobile; Docker feature matrix.
• Deployment compose SSOT — server/container Compose vs mobile (inference profiles, no phone OCI).
• Orchestration unified SSOT — capability merge rules.
• Environment variables (SSOT).
• vox-mcp API — Ollama fallback is desktop-oriented.

Direct on-device .vox runtime (experimental boundary)

If Vox later explores direct on-device .vox execution, treat it as a reduced, versioned subset and not parity with workstation/server runtime semantics.

Initial unsupported-by-default classes should include:

• actors/workflows/activities
• server/query/mutation function surfaces
• MCP tool declarations in script bodies
• async main in wasm isolation lanes
• host-assumed builtins without mobile/browser-safe shims (for example current std.http.* wasm guardrails)

Use the existing WASI guardrails and diagnostics as a baseline contract source, not as a claim of stock-phone parity.