Mo.py

!pip install sentencepiece
import sentencepiece as spm
import os, json, numpy as np, tensorflow as tf
from tensorflow.keras import layers, Model
import requests
from tensorflow import keras
from tensorflow.keras import layers
import tensorflow.keras.backend as K
# ===============================
from tensorflow.keras import mixed_precision
policy = mixed_precision.Policy('mixed_float16')  # fp16
mixed_precision.set_global_policy(policy)


print('1')
tf.get_logger().setLevel("ERROR")
SEED = 42
tf.random.set_seed(SEED)
np.random.seed(SEED)

# TPU 초기화
try:
    resolver = tf.distribute.cluster_resolver.TPUClusterResolver(tpu="local")
    tf.tpu.experimental.initialize_tpu_system(resolver)
    strategy = tf.distribute.TPUStrategy(resolver)
    print("✅ TPU 초기화 완료:", resolver.cluster_spec().as_dict())
    on_tpu = True

except Exception as e:
    print("⚠️ TPU 미사용, GPU/CPU로 진행:", e)
    strategy = tf.distribute.get_strategy()
    on_tpu = False

# Mixed precision
from tensorflow.keras import mixed_precision
policy = mixed_precision.Policy("mixed_bfloat16" if on_tpu else "float32")
mixed_precision.set_global_policy(policy)
print("✅ Mixed precision:", policy)

# =======================
# 1) 파일 다운로드
# =======================
def download_file(url, save_path):
    r = requests.get(url, stream=True)
    r.raise_for_status()
    with open(save_path, "wb") as f:
        for chunk in r.iter_content(8192*2):
            f.write(chunk)
    print(f"✅ {save_path} 저장됨")

DATA_PATH = "corpus.txt"
TOKENIZER_PATH = "ko_unigram.model"

if not os.path.exists(DATA_PATH):
    download_file(
        "https://huggingface.co/datasets/OpenLab-NLP/corpus-ko-22m/resolve/main/shuffled_corpus.txt?download=true",
        DATA_PATH
    )

if not os.path.exists(TOKENIZER_PATH):
    download_file(
        "https://huggingface.co/Yuchan5386/inlam-100m/resolve/main/ko_unigram.model?download=true",
        TOKENIZER_PATH
    )

sp = spm.SentencePieceProcessor(TOKENIZER_PATH)

pad_id = sp.piece_to_id("<pad>") if sp.piece_to_id("<pad>") != -1 else 0
start_id = sp.piece_to_id("<start>")
sep_id = sp.piece_to_id("<sep>")
end_id = sp.piece_to_id("<end>")
unk_id = sp.piece_to_id("<unk>")
vocab_size = sp.get_piece_size()
print(f"✅ Vocabulary size: {vocab_size}")

max_len = 128
batch_size = 96

def text_to_ids(text):
    return sp.encode(text, out_type=int)

def ids_to_text(ids):
    return sp.decode(ids)

def txt_stream(file_path):
    with open(file_path, "r", encoding="utf-8") as f:
        for line in f:
            text = line.strip()
            if not text:
                continue

            ids = text_to_ids(text)
            ids = ids[:max_len - 1]  # 마지막에 <end> 넣기 위해 -1

            full_input = ids + [end_id]
            pad_len = max_len - len(full_input)
            full_input += [pad_id] * pad_len

            # target = next-token shifted sequence
            target = full_input[1:] + [pad_id]
            yield (
                tf.convert_to_tensor(full_input, dtype=tf.int32),
                tf.convert_to_tensor(target, dtype=tf.int32)
            )


steps_per_epoch = 23119910 // batch_size
LIMIT = 23119910

dataset = tf.data.Dataset.from_generator(
    lambda: txt_stream(DATA_PATH),
    output_signature=(
        tf.TensorSpec(shape=(max_len,), dtype=tf.int32),
        tf.TensorSpec(shape=(max_len,), dtype=tf.int32),
    )
)

dataset = dataset.take(LIMIT).shuffle(2000, seed=SEED).batch(batch_size, drop_remainder=True).prefetch(tf.data.AUTOTUNE)

with strategy.scope():
    dist_dataset = strategy.experimental_distribute_dataset(dataset)

class SwiGLU(layers.Layer):
    def __init__(self, d_model, d_ff):
        super().__init__()
        self.proj = layers.Dense(d_ff)
        self.out = layers.Dense(d_model)
    def call(self, x):
        x_proj = self.proj(x)
        x_val, x_gate = tf.split(x_proj, 2, axis=-1)
        return self.out(x_val * tf.nn.silu(x_gate))
        
class Lo(layers.Layer):
    def __init__(self, d_model):
        super().__init__()
        self.d = layers.Dense(64, activation='silu', dtype='float16')  # fp16 연산
        self.w = layers.Dense(d_model, dtype='float16')               # fp16 연산
        self.norm = layers.LayerNormalization(epsilon=1e-5, dtype='float32')  # fp32

    def call(self, x):
        p = self.d(x)
        p = self.w(p)
        p = self.norm(p)  # fp32
        return tf.cast(p, x.dtype) + x  # 다시 fp16로 맞춰서 Add


class Block(layers.Layer):
    def __init__(self, d_model):
        super().__init__()
        self.mha = layers.MultiHeadAttention(8, 384//8)
        self.glu = SwiGLU(d_model, 1048)
        self.lo = Lo(d_model)

    def call(self, x):
        x = self.mha(x)
        x = self.glu(x)
        x = self.lo(x)
        return x

class LaSLM(tf.keras.Model):
    def __init__(self, vocab_size, max_seq_len, d_model, n_layers, dropout_rate=0.1):
        super().__init__()
        self.token_embedding = layers.Embedding(vocab_size, d_model, dtype=policy.compute_dtype)
        self.pos_embedding = layers.Embedding(max_seq_len, d_model, dtype=policy.compute_dtype)
        self.blocks = [Block(d_model) for _ in range(n_layers)]
        self.ln_f = layers.LayerNormalization(epsilon=1e-5, dtype='float32')  # ln_f는 fp32

    def call(self, x, training=False):
        batch_size, seq_len = tf.shape(x)[0], tf.shape(x)[1]
        positions = tf.range(seq_len)[tf.newaxis, :]
        x = self.token_embedding(x) + self.pos_embedding(positions)
        for block in self.blocks:
            x = block(x)
        x = self.ln_f(x)
        embedding_matrix = tf.cast(self.token_embedding.embeddings, x.dtype)
        logits = tf.matmul(x, embedding_matrix, transpose_b=True)
        return tf.cast(logits, tf.float32)  # loss 계산을 위해 fp32로 변환

def smoothed_loss_keras(y_true, y_pred, eps=0.1):
    y_true = tf.cast(y_true, tf.int32)
    mask = tf.cast(tf.not_equal(y_true, pad_id), tf.float32)
    vocab = tf.shape(y_pred)[-1]
    y_true_oh = tf.one_hot(y_true, depth=vocab, dtype=tf.float32)
    y_true_ls = (1.0 - eps) * y_true_oh + eps / tf.cast(vocab, tf.float32)
    log_probs = tf.nn.log_softmax(y_pred, axis=-1)
    per_tok = -tf.reduce_sum(y_true_ls * log_probs, axis=-1)
    per_tok = per_tok * mask
    return tf.reduce_sum(per_tok) / (tf.reduce_sum(mask) + 1e-8)


with strategy.scope():
    model = LaSLM(vocab_size=vocab_size, max_seq_len=max_len, d_model=384, n_layers=3)
    dummy_input = tf.zeros((batch_size, max_len), dtype=tf.int32)
    _ = model(dummy_input, training=False)
    model.summary()

    optimizer = tf.keras.optimizers.Adam(1e-4, beta_1=0.9, beta_2=0.95, epsilon=1e-8, clipnorm=1.0)
    model.compile(optimizer=optimizer, loss=smoothed_loss_keras)

    # 학습
    history = model.fit(dist_dataset, epochs=1, steps_per_epoch=steps_per_epoch, verbose=1)
model.save_weights("tf_model.weights.h5")
print("✅ 모델 가중치 저장 완료!")

def generate_text_topp(model, prompt, max_len=500, max_gen=500, p=0.9, temperature=0.8, min_len=20):
    model_input = text_to_ids(f"<start> {prompt}")
    model_input = model_input[:max_len]
    generated = list(model_input)
    for step in range(max_gen):
        if len(generated) > max_len:
            input_seq = generated[-max_len:]
        else:
            input_seq = generated
        input_padded = np.pad(input_seq, (0, max_len - len(input_seq)), constant_values=pad_id)
        input_tensor = tf.convert_to_tensor([input_padded])
        logits = model(input_tensor, training=False)
        next_token_logits = logits[0, len(input_seq) - 1].numpy()
        next_token_logits[end_id] -= 5.0
        next_token_logits[pad_id] -= 10.0
        probs = tf.nn.softmax(next_token_logits / temperature).numpy()
        sorted_indices = np.argsort(probs)[::-1]
        sorted_probs = probs[sorted_indices]
        cumulative_probs = np.cumsum(sorted_probs)
        cutoff = np.searchsorted(cumulative_probs, p)
        top_indices = sorted_indices[:cutoff + 1]
        top_probs = sorted_probs[:cutoff + 1]
        top_probs /= np.sum(top_probs)
        next_token_id = np.random.choice(top_indices, p=top_probs)
        if next_token_id == end_id and len(generated) >= min_len:
            break
        generated.append(int(next_token_id))
    return ids_to_text(generated)

print("\n\n===== 생성 결과 =====")  
print(generate_text_topp(model, "지난 2년 동안 출연연이 국가가 필요한 연구를", p=0.9))