90 lines
2.9 KiB
Markdown
90 lines
2.9 KiB
Markdown
|
---
|
|||
|
|
license: llama2
|
|||
|
|
base_model: meta-llama/Llama-2-7b-hf
|
|||
|
|
pipeline_tag: text-generation
|
|||
|
|
language:
|
|||
|
|
- en
|
|||
|
|
tags:
|
|||
|
|
- llama
|
|||
|
|
- text-generation
|
|||
|
|
- rl-mpq
|
|||
|
|
- mixed-precision
|
|||
|
|
- quantization
|
|||
|
|
- fake-quantization
|
|||
|
|
- balanced
|
|||
|
|
- llama-2
|
|||
|
|
library_name: transformers
|
|||
|
|
datasets:
|
|||
|
|
- wikitext
|
|||
|
|
widget:
|
|||
|
|
- text: "The capital of France is"
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
# Llama 2 7B — RL-MPQ Balanced
|
|||
|
|
|
|||
|
|
Standalone **RL-MPQ** (Reinforcement Learning Mixed-Precision Quantization) checkpoint for the
|
|||
|
|
**Balanced** scenario — a quantized variant of
|
|||
|
|
[meta-llama/Llama-2-7b-hf](https://huggingface.co/meta-llama/Llama-2-7b-hf).
|
|||
|
|
|
|||
|
|
| Field | Value |
|
|||
|
|
|-------|-------|
|
|||
|
|
| **Base model** | [meta-llama/Llama-2-7b-hf](https://huggingface.co/meta-llama/Llama-2-7b-hf) |
|
|||
|
|
| **Scenario** | Balanced |
|
|||
|
|
| **Avg bits / weight** | 4.375 |
|
|||
|
|
| **Compression vs FP16** | 3.6571× |
|
|||
|
|
| **WikiText-2 PPL** | 5.0437 |
|
|||
|
|
| **Layers** | 32 |
|
|||
|
|
| **Bit distribution** | `{'4': 29, '8': 3}` |
|
|||
|
|
| **Format** | Fake-quant FP16 + `rlmpq_policy.json` |
|
|||
|
|
|
|||
|
|
**Collection:** [RL-MPQ — Llama 2 7B](https://huggingface.co/collections/AvoCahDoe/rl-mpq-llama-2-7b-6a2ae4f8c590727304e3f634) — all five scenarios for Llama 2 7B.
|
|||
|
|
|
|||
|
|
## Usage
|
|||
|
|
|
|||
|
|
```python
|
|||
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|||
|
|
|
|||
|
|
repo = "AvoCahDoe/llama-2-7b-rlmpq-balanced"
|
|||
|
|
|
|||
|
|
model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="float16")
|
|||
|
|
tokenizer = AutoTokenizer.from_pretrained(repo)
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
## Other Llama 2 7B scenarios
|
|||
|
|
|
|||
|
|
| Scenario | Avg bits | Compression | WikiText-2 PPL |
|
|||
|
|
|----------|----------|-------------|----------------|
|
|||
|
|
| [High Fidelity](https://huggingface.co/AvoCahDoe/llama-2-7b-rlmpq-high-fidelity) | 6.5 | 2.4615x | 4.9808 |
|
|||
|
|
| [Conservative](https://huggingface.co/AvoCahDoe/llama-2-7b-rlmpq-conservative) | 5.125 | 3.122x | 5.0276 |
|
|||
|
|
| [Aggressive](https://huggingface.co/AvoCahDoe/llama-2-7b-rlmpq-aggressive) | 3.5938 | 4.4522x | 5.2614 |
|
|||
|
|
| [Extreme Survival](https://huggingface.co/AvoCahDoe/llama-2-7b-rlmpq-extreme-survival) | 2.9688 | 5.3895x | 10.9577 |
|
|||
|
|
|
|||
|
|
Grouped archive (all scenarios in one repo):
|
|||
|
|
[AvoCahDoe/llama-2-7b-rlmpq](https://huggingface.co/AvoCahDoe/llama-2-7b-rlmpq)
|
|||
|
|
|
|||
|
|
## Method
|
|||
|
|
|
|||
|
|
1. **Phase 3** — PPO agent assigns per-layer bit widths under the Balanced reward target.
|
|||
|
|
2. **Phase 4** — Policy replayed on real weights; WikiText-2 perplexity validates quality.
|
|||
|
|
3. **Export** — Fake-quantized FP16 weights compatible with Hugging Face Transformers.
|
|||
|
|
|
|||
|
|
## Files
|
|||
|
|
|
|||
|
|
| File | Description |
|
|||
|
|
|------|-------------|
|
|||
|
|
| `config.json` | Llama architecture + RL-MPQ metadata |
|
|||
|
|
| `model.safetensors` | Fake-quantized weights |
|
|||
|
|
| `rlmpq_policy.json` | Per-layer bit-width policy |
|
|||
|
|
| `rlmpq_metrics.json` | Validation & PPL summary |
|
|||
|
|
|
|||
|
|
## Citation
|
|||
|
|
|
|||
|
|
```bibtex
|
|||
|
|
@misc{rlmpq_llama_2_7b_balanced_2026,
|
|||
|
|
title = {RL-MPQ Balanced: Llama 2 7B Mixed-Precision Quantization},
|
|||
|
|
author = {AvoCahDoe},
|
|||
|
|
year = {2026},
|
|||
|
|
url = {https://huggingface.co/AvoCahDoe/llama-2-7b-rlmpq-balanced}
|
|||
|
|
}
|
|||
|
|
```
|