Test.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



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):

            f.write(chunk)

    print(f"✅ {save_path} 저장됨")



DATA_PATH = "converted.jsonl"

TOKENIZER_PATH = "ko_unigram.model"



if not os.path.exists(DATA_PATH):

    download_file(

        "https://huggingface.co/datasets/Yuchan5386/SFT/resolve/main/data_shuffled_1.jsonl?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 = 200

batch_size = 128



def text_to_ids(text):

    return sp.encode(text, out_type=int)

def ids_to_text(ids):

    return sp.decode(ids)



def jsonl_stream(file_path):

    with open(file_path, "r", encoding="utf-8") as f:

        for line in f:

            data = json.loads(line)

            conversations = data.get("conversations", [])

            for i in range(0, len(conversations) - 1, 2):

                human_msg = conversations[i]

                gpt_msg   = conversations[i + 1]

                if human_msg.get("from") != "human" or gpt_msg.get("from") != "gpt":

                    continue

                prompt   = human_msg.get("value", "").strip()

                response = gpt_msg.get("value", "").strip()

                full = f"<start> {prompt} <sep> {response} <end>"

                if "<sep>" not in full:

                    continue

                sep_index  = full.index("<sep>")

                input_text = full[:sep_index + len("<sep>")].strip()

                target_text = full[sep_index + len("<sep>"):].strip()



                input_ids  = text_to_ids(input_text)

                target_ids = text_to_ids(target_text + " <end>")



                available_len = max_len - len(input_ids)

                if available_len <= 0:

                    input_ids = input_ids[-max_len:]

                    target_ids = []

                    target_mask = [0] * len(input_ids)

                else:

                    target_ids = target_ids[:available_len]

                    target_mask = [0] * len(input_ids) + [1] * len(target_ids)



                full_input = input_ids + target_ids

                pad_len = max_len - len(full_input)

                full_input += [pad_id] * pad_len

                target_mask += [0] * pad_len



                target_seq = full_input[1:] + [end_id]

                target_seq = target_seq[:max_len]



                masked_target = [

                    t if m == 1 else pad_id

                    for t, m in zip(target_seq, target_mask)

                ]



                yield (

                    tf.convert_to_tensor(full_input, dtype=tf.int32),

                    tf.convert_to_tensor(masked_target, dtype=tf.int32)

                )



dataset = tf.data.Dataset.from_generator(

    lambda: jsonl_stream(DATA_PATH),

    output_signature=(

        tf.TensorSpec(shape=(max_len,), dtype=tf.int32),

        tf.TensorSpec(shape=(max_len,), dtype=tf.int32),

    ),

)

dataset = dataset.shuffle(1000, seed=SEED).batch(batch_size, drop_remainder=True).prefetch(tf.data.AUTOTUNE)



with strategy.scope():

    dist_dataset = strategy.experimental_distribute_dataset(dataset)



class Lo(layers.Layer):

    def __init__(self, d_model):

        super().__init__()

        self.proj = layers.Dense(d_model, use_bias=True, dtype='bfloat16')

        self.p = layers.Dense(128, use_bias=True, dtype='bfloat16')

        

    def call(self, x):

        x = self.proj(x)

        x = tf.nn.gelu(x)

        x = self.p(x)

        return x

        

class LoSoU(layers.Layer):

    def __init__(self, d_model):

        super().__init__()

        self.Q = layers.Dense(128)

        self.K = layers.Dense(128)

        self.V = Lo(d_model)

        self.O = layers.Dense(d_model)

        self.proj = layers.Dense(d_model, use_bias=True)



    def call(self, x):

        residual = x   # 🔹 원본 저장

        q = self.Q(x)

        k = self.K(x)

        V = self.V(x)



        g_q = tf.nn.sigmoid(q)

        g_k = tf.nn.sigmoid(k)



        score = g_q * g_k

        score = tf.cumsum(score, axis=1)  # (B, L, D)

        x = score * V



        out = self.proj(x)  # 🔹 residual과 같은 차원으로 통일



        a, b = tf.split(out, 2, axis=-1)

        out = self.O(tf.nn.silu(a) * b)



        return out + residual   # ✅ 잔차 연결 안정화



      

class Block(layers.Layer):

    def __init__(self, d_model, r, hyper_n, num_heads, num_groups):

        super().__init__()

        self.losou = [LoSoU(d_model) for _ in range(hyper_n)]



    def call(self, x):

        for losou in self.losou:

            x = losou(x)     

        return x

 

class Sequen(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)

        self.pos_embedding = layers.Embedding(max_seq_len, d_model)

        self.blocks = [Block(d_model, r=204, hyper_n=3, num_heads=8, num_groups=2) for _ in range(n_layers)]



        # ✅ 마지막도 RMSNorm으로

        self.ln_f = layers.LayerNormalization(epsilon=1e-5, dtype="bfloat16")



    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, :]  # (1, seq_len)



        x = self.token_embedding(x) + self.pos_embedding(positions)  # (batch, seq_len, d_model)

        for block in self.blocks:

            x = block(x)

            

        x = self.ln_f(x)  # (batch, seq_len, d_model)



        # ✅ embedding weight tying

        embedding_matrix = self.token_embedding.embeddings

        logits = tf.matmul(x, embedding_matrix, transpose_b=True)  # (batch, seq_len, vocab_size)

        return tf.cast(logits, tf.float32)



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) * mask

    return tf.reduce_sum(per_tok) / (tf.reduce_sum(mask) + 1e-8)



def masked_accuracy(y_true, y_pred):

    mask = tf.cast(tf.not_equal(y_true, pad_id), tf.float32)

    pred_id = tf.argmax(y_pred, axis=-1, output_type=tf.int32)

    acc = tf.cast(tf.equal(y_true, pred_id), tf.float32) * mask

    return tf.reduce_sum(acc) / (tf.reduce_sum(mask) + 1e-8)

    

# =======================

# 모델 생성 & 학습

# =======================

with strategy.scope():

    model = Sequen(vocab_size, max_seq_len=max_len, d_model=384, n_layers=12, dropout_rate=0.1)

    dummy_input = tf.zeros((batch_size, max_len), dtype=tf.int32)

    _ = model(dummy_input, training=False)

    model.summary()



    optimizer = tf.keras.optimizers.Adam(3e-4, beta_1=0.9, beta_2=0.95, epsilon=1e-8, clipnorm=1.0)

    model.compile(optimizer=optimizer, loss=smoothed_loss_keras, metrics=[masked_accuracy])

    history = model.fit(dist_dataset, epochs=1, verbose=1)



# =======================

# 가중치 저장

# =======================

model.save_weights("Sequen.weights.h5")

print("✅ 모델 가중치 저장 완료!")



# =======================

@tf.function(input_signature=[

    tf.TensorSpec(shape=(1, None), dtype=tf.int32),  # input_ids

    tf.TensorSpec(shape=(vocab_size,), dtype=tf.int32),  # token_counts

    tf.TensorSpec(shape=(), dtype=tf.int32),  # current_length

    tf.TensorSpec(shape=(), dtype=tf.float32),  # temperature

    tf.TensorSpec(shape=(), dtype=tf.float32),  # repetition_penalty

    tf.TensorSpec(shape=(), dtype=tf.float32),  # top_p

    tf.TensorSpec(shape=(), dtype=tf.int32),  # top_k

    tf.TensorSpec(shape=(), dtype=tf.int32),  # min_len

    tf.TensorSpec(shape=(), dtype=tf.int32),  # step

])

def generate_step(input_ids, token_counts, current_length, temperature, repetition_penalty, top_p, top_k, min_len, step):

    pad_len = max_len - tf.shape(input_ids)[1]

    input_padded = tf.pad(input_ids, [[0,0],[0,pad_len]], constant_values=pad_id)

    logits = model(input_padded, training=False)

    next_logits = logits[0, current_length - 1]



    penalty = tf.pow(repetition_penalty, tf.cast(token_counts, tf.float32))

    next_logits = next_logits / penalty



    # 최소 길이와 pad 마스킹

    if current_length < min_len:

        next_logits = tf.tensor_scatter_nd_update(next_logits, [[end_id]], [-1e9])

    next_logits = tf.tensor_scatter_nd_update(next_logits, [[pad_id]], [-1e9])



    # top-k 필터링

    if top_k > 0:

        kth_val = tf.math.top_k(next_logits, k=top_k).values[-1]

        mask = next_logits < kth_val

        next_logits = tf.where(mask, -1e9, next_logits)



    # top-p (nucleus) 필터링 + temperature

    next_logits = next_logits / temperature

    probs = tf.nn.softmax(next_logits)

    sorted_probs, sorted_idx = tf.math.top_k(probs, k=vocab_size)

    cum_probs = tf.cumsum(sorted_probs)

    cutoff_mask = cum_probs <= top_p

    cutoff_idx = tf.reduce_sum(tf.cast(cutoff_mask, tf.int32)) + 1

    cutoff_idx = tf.minimum(cutoff_idx, vocab_size)

    filtered_idx = sorted_idx[:cutoff_idx]

    filtered_probs = sorted_probs[:cutoff_idx]

    filtered_probs = filtered_probs / tf.reduce_sum(filtered_probs)



    # 🔹 50%는 argmax, 50%는 샘플링

    rand_val = tf.random.uniform([], 0.1, 1)

    def sample():

        sampled_id = tf.random.categorical(tf.math.log([filtered_probs]), 1)[0,0]

        return filtered_idx[sampled_id]

    def argmax():

        return filtered_idx[tf.argmax(filtered_probs)]

    sampled_id = tf.cond(rand_val < 0, argmax, sample)

    sampled_id = tf.cast(sampled_id, tf.int32)



    # token_counts 업데이트

    token_counts = tf.tensor_scatter_nd_add(token_counts, [[sampled_id]], [1])

    return sampled_id, token_counts





# =====================

# 스트리밍 생성기 (CPU 최적화 버전)

# =====================

def generate_text_streaming(model, prompt, max_len=115, max_gen=100,

                            temperature=0.75, min_len=20,

                            repetition_penalty=1.2, top_p=0.9, top_k=50):

    model_input = text_to_ids(f"<start> {prompt} <sep>")

    model_input = model_input[:max_len]

    generated = list(model_input)

    start_output_idx = len(model_input)



    # TF 변수로 토큰 카운트 관리

    token_counts_np = np.zeros(vocab_size, dtype=np.int32)

    for t in generated:

        token_counts_np[t] += 1

    token_counts = tf.Variable(token_counts_np, dtype=tf.int32)



    prev_decoded = ""



    for step in range(max_gen):

        input_tensor = tf.expand_dims(generated, axis=0)  # [1, seq_len]



        sampled_id, token_counts = generate_step(

            input_tensor,

            token_counts,

            tf.constant(len(generated), dtype=tf.int32),

            tf.constant(temperature, dtype=tf.float32),

            tf.constant(repetition_penalty, dtype=tf.float32),

            tf.constant(top_p, dtype=tf.float32),

            tf.constant(top_k, dtype=tf.int32),

            tf.constant(min_len, dtype=tf.int32),

            tf.constant(step, dtype=tf.int32)

        )



        sampled_id = int(sampled_id.numpy())

        generated.append(sampled_id)



        # 디코딩은 출력 시점에만

        if len(generated) > start_output_idx:

            decoded_full = sp.decode(generated[start_output_idx:])

            decoded_full = decoded_full.replace("▁", " ").strip()

            for t in ["<start>", "<sep>", "<end>"]:

                decoded_full = decoded_full.replace(t, "")

            decoded_full = decoded_full.lstrip(",!?.는은 ")



            new_output = decoded_full[len(prev_decoded):]

            if new_output:

                yield new_output

                prev_decoded = decoded_full



            # 종료 조건

            if len(generated) >= min_len and (sampled_id == end_id or decoded_full.endswith(('.', '!', '?'))):

                break







for token in generate_text_streaming(

    model, '안녕하세요',

    max_len=max_len,

    max_gen=115,

    temperature=0.8,

    min_len=10,

    repetition_penalty=1.1,

    top_p=0.9,

    top_k=32

):

    print(token, end="", flush=True)  




이 학습 코드가 왜 NaN을 뱉는지 설명해