---
license: llama2
base_model:
  - meta-llama/Llama-2-13b-hf
pipeline_tag: text-generation
tags:
  - grasprune
  - structured-pruning
  - pruning
  - llama-2
---

# Llama-2-13B — original GRASPrune, pruning ratio 0.5

This is a structurally pruned derivative of [`meta-llama/Llama-2-13b-hf`](https://huggingface.co/meta-llama/Llama-2-13b-hf).

## Model provenance

**This checkpoint was produced by the original GRASPrune pruning pipeline in `methods/GRASPrune`, not by `new_method_2`.** The local `new_method_2` code was used only for downstream evaluation and did not produce or alter the pruned weights.

- Method: [GRASPrune: Global Gating for Budgeted Structured Pruning of Large Language Models](https://arxiv.org/abs/2604.19398)
- Original implementation: [ZiY-Wang/GRASPrune](https://github.com/ZiY-Wang/GRASPrune)
- Upstream source revision recorded with this release: `666f4f04ec9e5e3218904df4d0935841bbd570fc`
- Base model: `meta-llama/Llama-2-13b-hf`
- Target pruning ratio: `0.5`
- GRASPrune retained budget: `keep_ratio=0.5`
- Export dtype/format: `torch.bfloat16`, materialized state dict
- Rescale compensation: `rescale_alpha=0.5`

The 50% ratio is the target reduction of GRASPrune's global prunable structural budget across FFN intermediate channels and attention KV groups. It is not a uniform per-layer sparsity ratio or a file-size reduction ratio.

## Files and loading

This is not a standard Transformers `save_pretrained` directory and cannot be loaded directly with `AutoModelForCausalLM.from_pretrained()`.

- `pruned_state_dict.safetensors`: materialized pruned weights
- `meta.json`: layer-specific shapes required to rebuild the architecture
- `layer_mask_report.csv`: layer-wise retention report
- `provenance.json`: release provenance and artifact identity
- `SHA256SUMS`: checksum for the weight file

Clone the original GRASPrune repository, run from its root, and use its `rebuild.py` loader. Loading also requires authorized access to the gated Llama-2-13B base model:

```python
import os
import torch
from huggingface_hub import snapshot_download
from rebuild import load_pruned_model

checkpoint_dir = snapshot_download("LiamCarter/grasprune_llama2-13b_ratio0.5")
model, tokenizer, meta = load_pruned_model(
    model_id="meta-llama/Llama-2-13b-hf",
    state_dict_path=os.path.join(checkpoint_dir, "pruned_state_dict.safetensors"),
    meta_path=os.path.join(checkpoint_dir, "meta.json"),
    torch_dtype=torch.bfloat16,
    device="cuda:0",
    local_only=False,
)
```

## Local evaluation

The weights were evaluated read-only with the local `new_method_2` evaluation flow. Percentages:

| Common-sense 5-task macro avg | ICL 0-shot | ICL 1-shot | ICL 4-shot | ICL 8-shot |
|---:|---:|---:|---:|---:|
| 49.636 | 22.36 | 18.91 | 21.96 | 28.53 |

These are local benchmark results, not upstream GRASPrune claims.

## Intended use, limitations, and license

This checkpoint is intended for structured-pruning and benchmark research. It has not been validated for production deployment, safety, factual reliability, bias, multilingual robustness, or long-context behavior. Performance can be substantially lower than the parent model, especially at this pruning ratio.

This is a derivative of Llama 2 and remains subject to the [Llama 2 Community License](https://ai.meta.com/llama/license/) and [Acceptable Use Policy](https://ai.meta.com/llama/use-policy/). Llama 2 is licensed under the LLAMA 2 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved. Users must separately obtain access to the gated base model.