Summary

LARQL enables direct querying and manipulation of transformer neural network weights by decompiling models into a queryable vindex format and using the Lazarus Query Language (LQL). This approach treats model knowledge as a graph database, allowing users to browse, edit, and recompile weights without traditional fine-tuning or GPU requirements for basic operations, targeting researchers and power users.

How It Works

LARQL's core concept is "the model IS the database." It decompiles transformer weights into a vindex, where gate vectors become KNN, embeddings token lookups, and down projections edge labels. LQL provides a SQL-like interface for querying, browsing, and mutating this vindex, offering direct interaction with learned knowledge.

Quick Start & Requirements

• Install: Requires Rust toolchain; build with cargo build --release.
• Model Extraction: larql extract-index <model> -o <vindex> [--f16] [--level <browse|inference|all>]. Browse-only vindexes are ~3 GB (f16), inference-enabled ~6 GB (f16).
• Prerequisites: Openblas for Linux builds. Metal GPU backend available (--features metal). No GPU needed for browse/query.
• Docs: docs/lql-guide.md.

Highlighted Details

• Vindex: Query-optimized weight format with streaming extraction, zero-copy mmap loading, and patch overlays.
• LQL: Comprehensive query language for lifecycle, browsing, inference, mutation, and patch operations.
• Patches: Lightweight .vlp files for incremental, read-only knowledge edits.
• Vindexfile: Declarative build system for vindexes.
• Tiered Context: Enables effectively infinite context without large KV caches.
• Residual Stream Trace: Captures and queries full inference trajectories.
• Model Support: Broad support for Gemma, Llama, Mistral, Mixtral, Qwen, Phi, DeepSeek, GPT-OSS, GPT-2 families (safetensors, GGUF, MLX).

Maintenance & Community

No explicit details on maintainers, community channels, or roadmap were found in the provided README.

Licensing & Compatibility

• License: Apache-2.0.
• Compatibility: Permissive for commercial use and closed-source projects.

Limitations & Caveats

• Build requires Rust toolchain and Openblas on Linux.
• MXFP4-quantized MoE models may yield noisy DESCRIBE/WALK results; INFER is recommended.
• Some vindex ecosystem specifications are noted as not fully implemented.