Glossary

Quick reference for terms used throughout the documentation.

Core Concepts

TermDefinition
CRDTConflict-free Replicated Data Type — a data structure that can be merged from any device without conflicts.
LoroThe Rust CRDT library used by Thumper-Run for document state (Map, List, Text, Tree).
DocumentA CRDT container with a unique ID. Settings, app state, and reviews are all documents.
ManifestA YAML file (.thumper.yaml, .model.yaml, .pack.yaml) that defines how to install and run an app or model.
PackA collection of model manifests bundled together for a specific use case.
ModelA neural network weights file (safetensors, GGUF, ONNX, etc.) used by AI apps.
CheckpointA saved model state at a specific training step. Often used interchangeably with "model weights".
OperationA single CRDT change (insert, delete, update). Operations are the unit of sync.
MergeCombining CRDT operations from two devices into a consistent state, automatically and without conflicts.
Conflict-freeCRDTs guarantee all replicas converge to the same state regardless of operation order.
LoroRust CRDT library powering Thumper’s sync engine. Provides JSON-like document structure with automatic conflict resolution.
Manifest LifecycleThe progression from authoring a .thumper.yaml through validation, catalog submission, review, and publication.

Security & Encryption

TermDefinition
E2EEEnd-to-End Encryption — a security property in which only authorized endpoints hold content-decryption keys. Private E2EE is planned, not a current Thumper-Run release guarantee.
BeeKEMA group key agreement protocol based on CGKA that informs Thumper-Run’s planned private multi-device encryption design.
CGKAContinuous Group Key Agreement — the protocol family for evolving group encryption keys. BeeKEM is the planned protocol; MLS and TreeKEM are prior art.
EpochA key-rotation generation in a group encryption protocol. Thumper-Run does not currently guarantee epoch changes for released sync workflows.
Forward SecrecyThe goal that compromise of current keys does not expose earlier content. This remains planned for Thumper-Run sync workflows.
BiscuitA capability-based authorization token used for path-based access control in sync modes 2+.
RauthyThe OIDC identity provider used for optional account login (Google, GitHub).
OIDCOpenID Connect — an identity layer on OAuth 2.0 used for single sign-on.
PKCEProof Key for Code Exchange — prevents authorization code interception in OAuth flows.
Recovery PhraseA mnemonic used for supported identity or key recovery. Its availability and recovery scope depend on the product build and workflow.
CGKA (Continuous Group Key Agreement)A protocol family for managing group encryption keys as members join and leave. It informs the planned BeeKEM design.
EpochA key-rotation boundary in BeeKEM. Thumper-Run’s epoch and forward-secrecy guarantees remain planned.

Models & AI

TermDefinition
VRAMVideo RAM — GPU memory. Determines the largest model you can run.
QuantizationReducing model precision (e.g., fp16 → q4) to use less memory at slight quality cost.
InferenceRunning a model to generate output (images, text, audio) from input.
LoRALow-Rank Adaptation — a small add-on file that modifies model behavior without retraining.
VAEVariational Autoencoder — the encoder/decoder that converts images to/from latent space in diffusion models.
GGUFA quantized model format for CPU/GPU inference via llama.cpp. Common for LLMs.
SafetensorsA safe, fast model format that prevents arbitrary code execution (unlike pickle-based .pt files).
ONNXOpen Neural Network Exchange — a portable format for running models across different frameworks.
FP16 / BF1616-bit floating point formats. FP16 has higher precision but overflows on some AMD GPUs; BF16 is safer.
SafetensorsSafe model file format by HuggingFace. Cannot execute arbitrary code, unlike pickle-based .pt/.pkl files.
MIOpenAMD’s GPU-accelerated library for deep learning primitives. Requires a one-time warm-up (Find-DB) on first run.
comgrAMD’s Compiler for Machine-level GPU IR. Cached at ~/.cache/comgr/ (~650 MB). Used by ROCm for kernel compilation.
Find-DBMIOpen’s benchmark database that stores optimal algorithm choices per GPU. Located at ~/.config/miopen/.
TheRockAMD’s nightly ROCm distribution with newer GPU support (e.g., gfx1150). Ships pre-built PyTorch wheels.
VRAM TierClassification of GPU memory for workload planning: Low (4–6 GB), Medium (8–12 GB), High (16–24 GB), Ultra (32+ GB).

GPU & Hardware

TermDefinition
CUDANVIDIA's GPU compute platform. Required for NVIDIA GPUs.
ROCmAMD's GPU compute platform. Required for AMD discrete and APU GPUs.
MIOpenAMD's deep learning library (similar to cuDNN). Requires a "find database" warmup on first run.
oneAPIIntel's GPU compute platform. Used for Intel Arc GPUs.
MPSMetal Performance Shaders — Apple's GPU backend for M-series chips.
NPUNeural Processing Unit — a dedicated AI accelerator chip (e.g., AMD XDNA, Intel NPU).
Health CheckAn HTTP endpoint the platform pings to confirm an app is ready to serve requests.
setsid()A Unix syscall that creates a new process session, preventing terminal signals from stopping spawned apps.
setsid()Unix system call that creates a new session with no controlling terminal. Used by Thumper to launch GUI apps cleanly.

Platform

TermDefinition
OllamaLocal LLM server with a managed model library. Auto-started by Thumper-Run when apps need text generation. Exposes an OpenAI-compatible API.
llama.cppC++ library for LLM inference on CPU and GPU. Uses GGUF model format. Powers the Thumper-Run agent and embedded inference.
TauriThe desktop app framework. Wraps a native webview with Rust backend IPC.
LeptosThe Rust web framework used for the UI. Compiles to WASM for the browser.
IPCInter-Process Communication — how the frontend (WASM) calls Rust backend commands in Tauri.
WebViewThe native browser engine (WebKitGTK on Linux, WebView2 on Windows) that renders the UI.
AppImageA portable Linux app format. Single file, no installation needed.
Sync TransportThe network layer (WebSocket or HTTP polling) used to exchange CRDT operations between devices.
pack_refsReferences from an app manifest to model pack manifests. Declares which model packs an app can use.
preferred_engineApp manifest field specifying the LLM inference backend: ollama, llamacpp, or openai.

Frequently Asked Questions