Llama 2 7B — RL-MPQ Quantized (Thesis Release)

Quantized variant of meta-llama/Llama-2-7b-hf using RL-MPQ (Reinforcement Learning Mixed-Precision Quantization): per-layer bit-width policies trained with PPO, validated on WikiText-2 perplexity.

This repo ships five compression scenarios as subfolders — from near-FP16 fidelity to aggressive survival mode — so you can pick the bits-vs-quality trade-off for your thesis experiments.

Scenarios

• High_Fidelity/ — load with subfolder="High_Fidelity"
• Conservative/ — load with subfolder="Conservative"
• Balanced/ — load with subfolder="Balanced"
• Aggressive/ — load with subfolder="Aggressive"
• Extreme_Survival/ — load with subfolder="Extreme_Survival"

Results (WikiText-2)

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "AvoCahDoe/llama-2-7b-rlmpq"
scenario = "Balanced"  # High_Fidelity | Conservative | Aggressive | Extreme_Survival

model = AutoModelForCausalLM.from_pretrained(repo, subfolder=scenario, torch_dtype="float16")
tokenizer = AutoTokenizer.from_pretrained(repo, subfolder=scenario)

Important: Always pass subfolder=<scenario>. Root config.json describes the collection; weights and tokenizer live inside each scenario folder.

Method (thesis summary)

1. Phase 3 — PPO agent selects per-layer bit widths under scenario-specific reward targets.
2. Phase 4 — Policies replayed on real weights; WikiText-2 PPL measures quality retention.
3. Export — Fake-quantized FP16 checkpoints (compatible with Hugging Face Transformers).

Citation

@misc{rlmpq2026,
  title  = {RL-MPQ: Reinforcement Learning Mixed-Precision Quantization},
  author = {AvoCahDoe},
  year   = {2026},
  url    = {https://huggingface.co/AvoCahDoe/llama-2-7b-rlmpq}
}

Part of the RL-NMP-Model-Quantasation thesis framework.