README.md

# LLMOE method weights

This directory is an upload-ready collection of **compression-method artifacts only**. It contains no untouched LLaVA, Qwen, InternVL, CLIP, or language-model cache weights.

## Included

- `methods/hidrop/llava-1.5-7b/final`: full HiDrop fine-tuned checkpoints for budgets 32, 64, 128, and 224.
- `methods/hidrop/llava-1.5-7b/stage1_projectors`: matching stage-1 projectors for those budgets.
- `methods/duet/llava-1.5-7b/finetuned`: compression-aware DUET fine-tuned checkpoints for budgets 32, 64, 128, and 224.
- `methods/glimpse_prune`: the learned glimpse-token/VIP modules for LLaVA-1.5-7B and Qwen2.5-VL-7B-Instruct.
- `methods/visionzip/llava-1.5-7b`: LLaVA-1.5-7B efficient-tuning projector checkpoints for budgets 32, 64, 128, and 224.
- `methods/vflowopt`: calibrated strategies and calibration provenance. VFlowOpt is training-free and has no learned tensor checkpoint.

Every full checkpoint includes its method config and the small Hugging Face metadata needed to identify/load it. Training logs, evaluation JSONs, optimizer states, and generic base-model directories are excluded.

## Missing or intentionally excluded

No completed checkpoint was found for QuietPrune, CORE, SoftTopk, or Varest. Nuwa and the other classical pruning methods are training-free. The Qwen VisionZip and Qwen-HiDrop training artifacts are excluded.

See `BUNDLE_MANIFEST.json` for the original source location of every included weight and the complete method inventory. See `SHA256SUMS` for upload-integrity hashes.

## Storage note

The large full-model files are hard-linked to their original files under `/workspace/output`; they are ordinary regular files and can be uploaded directly. Do not edit those staged weights in place, because a hard link shares the same underlying file. The VisionZip projectors were moved into this bundle from `/workspace/app/charles_llmoe` and are independent files. Copying this directory to another filesystem or uploading it materializes normal independent objects.