---
license: llama2
base_model: meta-llama/Llama-2-7b-hf
tags:
  - rl-mpq
  - mixed-precision
  - quantization
library_name: transformers
---

# LLAMA-2-7B — Conservative

Subfolder of [`AvoCahDoe/llama-2-7b-rlmpq`](https://huggingface.co/AvoCahDoe/llama-2-7b-rlmpq) (all RL-MPQ scenarios for this model).

Fake-quantized weights for **Conservative** applied to
[meta-llama/Llama-2-7b-hf](https://huggingface.co/meta-llama/Llama-2-7b-hf) (per-layer asymmetric group-wise quant, group_size=128).

## Load

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "AvoCahDoe/llama-2-7b-rlmpq"
model = AutoModelForCausalLM.from_pretrained(repo, subfolder="Conservative", torch_dtype="float16")
tokenizer = AutoTokenizer.from_pretrained(repo, subfolder="Conservative")
```

## Metrics

| Metric | Value |
|--------|-------|
| Scenario | Conservative |
| Average bits per weight | 5.125 |
| Compression ratio (vs FP16) | 3.122x |
| WikiText-2 perplexity | 5.0276 |
| Layers | 32 |
| Bit distribution | {'4': 23, '8': 9} |

## Policy (bits per layer)

```
[8, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 8, 8, 8, 8, 8, 8, 8, 8]
```

Generated: 2026-06-11T15:30:00.790589