Repository navigation
loraspectra: Fast Rust CLI to audit spectral rank, bake rsLoRA scaling, compress, and merge Unsloth LoRA .safetensors #17570
SalyyS1
started this conversation in
Show and tell
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Hi everyone,
For anyone training LoRA, rsLoRA, or DoRA adapters with Unsloth and deploying them to vLLM,
llama.cpp/ Ollama, or Hugging Face PEFT, we built loraspectra, a zero-dependency Rust CLI that audits, compresses, bakes scaling, and merges.safetensorsLoRA checkpoints on CPU in milliseconds.What it solves for Unsloth users
rsLoRA(rank_patternscaling directly into--bake-alpha --rslora):Downstream inference engines like
llama.cppGGUF converters and vLLM multi-LoRA runners sometimes apply uniformalpha_pattern. Runningloraspectra compress adapter_model.safetensors -o ready.safetensors --bake-alpha --rslorafoldsloraspectra audit):See exact Shannon effective rank (
R_EFF), stable rank (R_STABLE), 99% energy rank (K_99%), and Marchenko-Pastur noise thresholds (K_MP) for everyq_proj,k_proj,v_proj,up_proj,down_projlayer in ~2ms without loading PyTorch or CUDA.ortho-svd):Prune noise singular values with
--energy 0.99or--policy mp, or merge multiple Unsloth LoRA adapters (including Gemma-3 VLMlanguage_model.model.prefixes) without destructive cross-talk.Repository & binaries: https://github.com/SalyyS1/loraspectra
If you find
loraspectrauseful alongside Unsloth, a star on SalyyS1/loraspectra is greatly appreciated!All reactions