MeshLLM runtime integration¶
MeshLLM is a separate upstream inference runtime. PQWERAI does not fork it into the node; providers run it as their own service and the project supplies a supervisor and client that enforce pins and boundaries.
| Aspect | Implementation |
|---|---|
| Version | MeshLLM 0.78.x with a pinned CUDA bundle and pinned GGUF model (SHA-256) |
| Start mode | --local-model-only; analytics, default plugins and prompt disk caching disabled; isolated data dir |
| Supply chain | No downloader or auto-updater; publisher/product/native-file hashes verified before execution |
| API | OpenAI-compatible HTTP on literal loopback (127.0.0.1); remote operators need separately authenticated tunnels |
| Usage | Gateway signs usage bound to request/output commitments; duplicates and non-execution are reconciled |
| Resources | Per-job token/time/memory limits; GPU/CPU/power/temperature admission checks before work (not hard limits) |
| Low-VRAM hosting | Experimental, disabled by default: CPU-only, VRAM-capped partial offload or full-GPU profiles for allowlisted GGUF models |
Not yet established¶
- Mesh peer discovery and distributed routing across independent providers.
- Accountable hardware measurement; receipts remain provider claims.
- Canary, upgrade and rollback policy for runtime versions across operators.
Keep AI hosting off the CPU cores you use for mining or validation.
Sources at 72057fe3ec: ai/README.md · testnet/release/network/GUIDES.md
Found a problem? Report a documentation issue · Build testnet-rc1+72057fe3ec