chrishayuk/larql

Health v2 · maintenance only

Flags: young

Adoption not part of the score

1142 stars · 205 forks observed · 2026-09-03

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

LARQL decompiles transformer models into a queryable 'vindex' (vector index) format and provides LQL, a SQL-like query language, to browse, edit, and recompile model knowledge without a GPU. It also serves models over HTTP and gRPC with support for MoE expert sharding and HuggingFace publishing.

Use cases

• query transformer model weights like a graph database
• inspect what a model knows about an entity
• edit model knowledge by inserting edges
• run LLM inference on CPU without a GPU
• shard MoE experts across CPU-only machines
• research mechanistic interpretability of transformers
• publish model slices to HuggingFace

When to choose

• you want to browse or edit LLM knowledge as structured data
• you need CPU-only inference or distributed MoE sharding
• you're doing interpretability research on transformer internals

When to avoid

• you need a battle-tested production inference server
• you want standard fine-tuning or training workflows
• you need broad model-format support beyond supported transformer families

Facets

cli-tool · maturity active

llm-inference machine-learning database search-engine cli http-server rpc large-language-models machine-learning databases developer-tools artificial-intelligence windows cross-platform cli query-language lql vindex mechanistic-interpretability model-editing transformer-weights moe-sharding cpu-inference huggingface macos linux

Member repositories

For agents

markdown · JSON · MCP: product_card(name="chrishayuk/larql")