Yaongi/Yaongi-430M-A166M-PoC

HybriKo-430M-A166M-PoC

⚠️ PoC only. 실제 사용 불가.

Total	430.8M
Active	166.2M
Layers	52

Colab

# ========================================
# 새 Colab - 바로 실행
# ========================================

!pip install huggingface_hub sentencepiece torch -q

from huggingface_hub import snapshot_download
import sys
import torch
import sentencepiece as spm

# 다운로드
local_dir = snapshot_download("Yaongi/Yaongi-430M-A166M-PoC")
print(f"✅ 다운로드: {local_dir}")

# 모델 코드 로드
sys.path.insert(0, local_dir)
from model import HybriKo52, CFG

# 토크나이저
sp = spm.SentencePieceProcessor()
sp.load(f"{local_dir}/tokenizer.model")

# 모델 로드
device = "cuda" if torch.cuda.is_available() else "cpu"
model = HybriKo52(CFG).to(device)
model.load_state_dict(torch.load(f"{local_dir}/model.pt", map_location=device))
model.eval()
print(f"✅ 모델 로드 ({device})")

# 생성
import torch.nn.functional as F

@torch.no_grad()
def generate(prompt, max_new=50, temp=0.5, rep_penalty=1.2):
    ids = torch.tensor([[2] + sp.encode(prompt)]).to(device)
    for _ in range(max_new):
        logits = model(ids)['logits'][:, -1, :] / temp
        for t in set(ids[0].tolist()): logits[0, t] /= rep_penalty
        next_id = torch.multinomial(F.softmax(logits, -1), 1)
        if next_id.item() == 1: break
        ids = torch.cat([ids, next_id], 1)
    return sp.decode(ids[0].tolist())

# 테스트
print(generate("인공지능은"))
print(generate("대한민국"))
print(generate("서울은", max_new=150, temp=0.7, rep_penalty=1.15))
print(generate("한국어는", max_new=150))