Skip to content

Release: ruvllm v2.0.2 with BitNet Integration #157

@ruvnet

Description

@ruvnet

Release Summary

Published new versions of crates to crates.io with BitNet integration for Craftsman Ultra 30b 1-bit model support.

Published Crates

Crate Version Link
ruvllm 2.0.2 https://crates.io/crates/ruvllm
ruvector-sona 0.1.5 https://crates.io/crates/ruvector-sona
ruvector-temporal-tensor 2.0.1 https://crates.io/crates/ruvector-temporal-tensor
ruvector-crv 0.1.1 https://crates.io/crates/ruvector-crv
rvlite 0.3.0 https://crates.io/crates/rvlite
ruvector-core 2.0.1 https://crates.io/crates/ruvector-core
ruvector-gnn 2.0.1 https://crates.io/crates/ruvector-gnn
ruvector-graph 2.0.1 https://crates.io/crates/ruvector-graph
ruvector-mincut 2.0.1 https://crates.io/crates/ruvector-mincut
ruvector-raft 2.0.1 https://crates.io/crates/ruvector-raft
ruvector-cluster 2.0.1 https://crates.io/crates/ruvector-cluster
ruvector-replication 2.0.1 https://crates.io/crates/ruvector-replication

BitNet Integration Features (ruvllm v2.0.2)

From PR #151:

  • TL1 Kernels: High-performance ternary linear kernels

    • AVX2 SIMD implementation (tl1_avx2.rs)
    • WASM portable implementation (tl1_wasm.rs)
    • Generic kernel interface (tl1_kernel.rs)
  • Ternary Tensor Quantization

    • 1.58-bit quantization (ternary: -1, 0, +1)
    • Efficient bitpacked storage
    • Fast dequantization paths
  • RLM (Reasoning Language Model) Components

    • rlm_embedder.rs - Embedding layer with ternary weights
    • rlm_refiner.rs - Refinement passes for improved accuracy
  • Expert Cache with MoE Support

    • expert_cache.rs - Mixture-of-Experts caching
    • Dynamic expert loading/unloading
  • GGUF Export

    • gguf_export.rs - Export to GGUF format for llama.cpp compatibility
  • Evaluation & Tracing

    • eval.rs - Model evaluation utilities
    • trace.rs - Inference tracing and debugging

Other Updates

  • ruvector-sona v0.1.5: Added Debug implementation for SonaEngine
  • ruvector-crv v0.1.1: Added README for crates.io documentation
  • rvlite v0.3.0: Standalone vector database with 22 WASM module integrations
  • Workspace version: Bumped to 2.0.2

Related PRs

Documentation

  • ADR-017: Craftsman Ultra 30b 1-bit BitNet integration
  • DDD: BitNet quantizer module design
  • Research: Craftsman Ultra 30b 1-bit analysis

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions