TaH2-1.7B-max2
Qwen3-1.7B with adaptive latent iteration, at most two iterations per token.
Uses duo attention, a learned iteration decider (threshold 0.5), and stop_prob_mix.
Fine-tuned on the 1.7B subset of TaH2 AMTeam Tool.
Code: thu-nics/TaH · Paper: arXiv:2609.35748.
Install the TaH code, then load the backbone and both TaH modules:
import torch
from huggingface_hub import snapshot_download
from transformers import AutoTokenizer
from tah.model.tah_model import TaHForCausalLM
path = snapshot_download("nics-efc/TaH2-1.7B-max2")
tokenizer = AutoTokenizer.from_pretrained(path)
model = TaHForCausalLM.from_pretrained(
path, dtype=torch.bfloat16, device_map="cuda").eval()