100% Local Browser SVD (0 Uploads) Pure Rust CLI (0 Dependencies) 1,050x–2,680x Faster QR-Core SVD hf://org/repo Remote Inspection PEFT • Kohya • ComfyUI • FLUX • Diffusers

loraspectra v0.6.0

Spectral SVD Auditor, Eckart-Young Dynamic Rank Compressor, and Orthogonal Subspace Merger for .safetensors LoRA Adapters — powered by exact QR-Core SVD (O(r²d + r³)) in your browser or native Rust CLI without PyTorch or GPU VRAM.

View on GitHub → Prebuilt Binaries & Releases → Companion: mcp-slim →

Interactive .safetensors Spectral SVD & Rank-Collapse Inspector

🔒 100% Local Web File API — Your weights never leave your device

Drop any .safetensors LoRA file here, or click to select from disk

Decodes F32, IEEE-754 F16, and BF16 2D (lora_A, lora_B) pairs + .alpha scalars across PEFT, Kohya/ComfyUI, FLUX, Diffusers & OneTrainer

Ready — drop a local .safetensors adapter or click a 1-click preset above.

Native CLI Quickstart (Zero Dependencies)

cargo install --git https://github.com/SalyyS1/loraspectra

# Inspect or audit remote HuggingFace LoRAs directly (auto-cached in ${TMPDIR:-/tmp}/loraspectra-cache/)
loraspectra info hf://XLabs-AI/flux-RealismLora
loraspectra audit hf://org/llama3-lora/adapter_model.safetensors

# Compress a rank-collapsed LoRA at 99% Frobenius energy and emit PEFT rank_pattern config
loraspectra compress adapter_model.safetensors adapter_slim.safetensors --energy 0.99 --peft-config adapter_config.json

# Analyze attn vs mlp subspace conflict and get an exact merge command recommendation
loraspectra interference hf://org/code-lora hf://org/math-lora

Empirical Benchmarks

Eckart-Young Adaptive Energy Rank Compression Benchmark QR-Core SVD Speedup and Orthogonal Subspace Merge Benchmark