LOG-postech/elsa — reverse-engineered prompt

Build me a clean Python research repo for pruning large language models with ELSA, a surrogate free ADMM method for pushing models to very high sparsity without collapse. I want it to run from a simple command line script where I can choose a Hugging Face model, calibration dataset, sparsity ratio, sparsity pattern like unstructured or 2:4, ADMM training settings, precision, seed, and whether to run zero shot evaluation.

Please include setup instructions, requirements, example commands for single GPU and multi GPU with accelerate FSDP, and config examples. The code should support saving the pruned model when I pass an output path, and optional W&B logging. Keep the interface close to a research paper release, easy to reproduce, and include figures plus a README explaining the method, arguments, inference acceleration note, acknowledgements, and citation.

Use PyTorch and Hugging Face tooling, and look up current docs online if needed.