🧠 Mamba Hypernetwork for LLM Personalization

Mamba Hypernetwork sinh LoRA delta cho LLM tại inference time, cho phép cá nhân hóa tức thì mà không cần retrain.

Architecture

Mamba: 1024 dim, 16 state, 4 expand (~12M params)
LoRA: rank 16, inject vào q_proj + v_proj của 8 layers đầu
LLM: Llama-3.2-3B-Instruct

Training

Method: GRPO (Group Relative Policy Optimization)
Reward Model: phammminhhieu/persona-reward-model
Dataset: Personachat TrueCased

Usage

from mamba_ssm import Mamba
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# Load Mamba
config = json.load(open("config.json"))
mamba = MambaHypernetwork(config)
mamba.load_state_dict(torch.load("pytorch_model.bin"))
mamba.eval()

# Load LLM
llm = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.2-3B-Instruct")
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.2-3B-Instruct")

# Sinh delta từ persona
persona = "I love dogs. I am a photographer."
deltas = mamba(persona_ids, persona_mask, history_ids, history_mask)