File size: 10,506 Bytes
585706a
 
 
 
 
c35567a
4bbac72
 
585706a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c35567a
585706a
 
 
 
 
 
 
4bbac72
585706a
 
 
 
 
 
c35567a
585706a
 
 
 
 
 
 
 
c35567a
585706a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c35567a
585706a
 
 
 
 
 
 
 
 
 
 
c35567a
585706a
c35567a
585706a
 
 
 
 
 
 
 
 
 
 
4bbac72
 
 
 
585706a
 
4bbac72
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
585706a
4bbac72
 
 
 
 
 
 
 
 
 
 
 
c35567a
4bbac72
 
585706a
 
c35567a
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
import sentencepiece as spm
import os, numpy as np, tensorflow as tf
from tensorflow.keras import layers
import gradio as gr

# --- 1. ํ™˜๊ฒฝ ์„ค์ • ๋ฐ ๋ชจ๋ธ ๊ตฌ์กฐ ์ •์˜ ---
# ํŒŒ์ผ ์ด๋ฆ„๋งŒ ์‚ฌ์šฉ (ํ˜„์žฌ ์ž‘์—… ๋””๋ ‰ํ† ๋ฆฌ์— ํŒŒ์ผ์ด ์žˆ์–ด์•ผ ํ•จ)
TOKENIZER_PATH = "tokenizer.model" 
sp = spm.SentencePieceProcessor(TOKENIZER_PATH)
pad_id = sp.piece_to_id("<pad>") if sp.piece_to_id("<pad>") != -1 else 0
end_id = sp.piece_to_id("</s>")
vocab_size = sp.get_piece_size()

class TimeMix(layers.Layer):
    def __init__(self, d_model, layer_id, n_layers):
        super().__init__()
        self.d_model = d_model
        ratio = (layer_id / (n_layers - 1)) if n_layers > 1 else 0.5
        decay_speed = np.arange(d_model)
        self.time_decay = tf.Variable(-5 + 8 * (decay_speed / (d_model - 1)) ** (0.7 + 1.3 * ratio), dtype=tf.float32)
        self.time_first = tf.Variable(np.ones(d_model) * np.log(0.3), dtype=tf.float32)
        self.w_proj = layers.Dense(d_model, kernel_initializer='zeros', use_bias=False)
        self.r_proj = layers.Dense(d_model, kernel_initializer='zeros', use_bias=False)
        self.k_proj = layers.Dense(d_model, kernel_initializer='zeros', use_bias=False)
        self.v_proj = layers.Dense(d_model, kernel_initializer='zeros', use_bias=False)
        self.key = layers.Dense(d_model, use_bias=False)
        self.value = layers.Dense(d_model, use_bias=False)
        self.receptance = layers.Dense(d_model, use_bias=False)
        self.output_projection = layers.Dense(d_model, use_bias=False)
        self.tm_w = tf.Variable(1 - (ratio ** 0.5), dtype=tf.float32)
        self.tm_k = tf.Variable(1 - (ratio ** 0.5), dtype=tf.float32)
        self.tm_v = tf.Variable(1 - (ratio ** 0.5), dtype=tf.float32)
        self.tm_r = tf.Variable(1 - (ratio ** 0.2), dtype=tf.float32)

    def call(self, x, state):
        last_x, aa, bb, pp = state
        t_type = x.dtype
        tm_w, tm_k, tm_v, tm_r = tf.cast(self.tm_w, t_type), tf.cast(self.tm_k, t_type), tf.cast(self.tm_v, t_type), tf.cast(self.tm_r, t_type)
        dx = x * tm_w + last_x * (1 - tm_w)
        w = tf.cast(self.time_decay, t_type) + tf.cast(self.w_proj(dx), t_type)
        w = -tf.exp(tf.cast(w, tf.float32))
        r = self.receptance(x * tm_r + last_x * (1 - tm_r)) + self.r_proj(dx)
        k = self.key(x * tm_k + last_x * (1 - tm_k)) + self.k_proj(dx)
        v = self.value(x * tm_v + last_x * (1 - tm_v)) + self.v_proj(dx)
        u = tf.cast(self.time_first, tf.float32)
        kv, vv = tf.cast(k, tf.float32), tf.cast(v, tf.float32)
        ww = u + kv
        p = tf.maximum(pp, ww)
        e1, e2 = tf.exp(pp - p), tf.exp(ww - p)
        wkv = (e1 * aa + e2 * vv) / (e1 * bb + e2 + 1e-12)
        ww_next = w + pp
        p_next = tf.maximum(ww_next, kv)
        e1_next, e2_next = tf.exp(ww_next - p_next), tf.exp(kv - p_next)
        new_state = [x, e1_next * aa + e2_next * vv, e1_next * bb + e2_next, p_next]
        return self.output_projection(tf.nn.sigmoid(r) * tf.cast(wkv, t_type)), new_state

class ChannelMix(layers.Layer):
    def __init__(self, d_model, layer_id, n_layers):
        super().__init__()
        ratio = (layer_id / (n_layers - 1)) if n_layers > 1 else 0.5
        self.time_mix_k = tf.Variable(1 - (ratio ** 0.5), dtype=tf.float32)
        self.time_mix_r = tf.Variable(1 - (ratio ** 0.5), dtype=tf.float32)
        self.key = layers.Dense(int(d_model * 4.25), use_bias=False)
        self.receptance = layers.Dense(d_model, use_bias=False)
        self.value = layers.Dense(d_model, use_bias=False)

    def call(self, x, last_x):
        t_type = x.dtype
        tm_k, tm_r = tf.cast(self.time_mix_k, t_type), tf.cast(self.time_mix_r, t_type)
        k = self.key(x * tm_k + last_x * (1 - tm_k))
        r = self.receptance(x * tm_r + last_x * (1 - tm_r))      
        kv = self.value(tf.square(tf.nn.relu(k)))
        return tf.nn.sigmoid(r) * kv, x

class Block(layers.Layer):
    def __init__(self, d_model, layer_id, n_layers):
        super().__init__()
        self.ln = layers.LayerNormalization(epsilon=1e-5)
        self.time_mix = TimeMix(d_model, layer_id, n_layers)
        self.channel_mix = ChannelMix(d_model, layer_id, n_layers)
    def call(self, x, state):
        ln_x = self.ln(x)
        tm_out, tm_state = self.time_mix(ln_x, state[:4])
        x = x + tm_out
        cm_out, cm_last_x = self.channel_mix(ln_x, state[4])
        x = x + cm_out
        return x, tm_state + [cm_last_x]

class Head(tf.keras.Model):
    def __init__(self, vocab_size):
        super().__init__()
        self.lm_head = layers.Dense(vocab_size, use_bias=False, name="output_head")
    def call(self, x):
        return tf.cast(self.lm_head(x), tf.float32)
    
class LM(tf.keras.Model):
    def __init__(self, d_model, n_layers):
        super().__init__()
        self.token_embedding = layers.Embedding(vocab_size, d_model)
        self.blocks = [Block(d_model, i, n_layers) for i in range(n_layers)]
        self.ln_f = layers.LayerNormalization(epsilon=1e-5, dtype=tf.float32)
    def call(self, x, states):
        x = self.token_embedding(x)
        new_states = []
        for i, block in enumerate(self.blocks):
            x, b_state = block(x, states[i*5 : (i+1)*5])
            new_states.extend(b_state)
        return self.ln_f(x), new_states

# --- 2. ์ดˆ๊ธฐํ™” ๋ฐ ๊ฐ€์ค‘์น˜ ๋กœ๋“œ ---
d_model, n_layers = 512, 10
blocklm = LM(d_model, n_layers)
head = Head(vocab_size)

def get_init_state():
    return [tf.zeros((1, 1, d_model)) if i%5!=3 else tf.ones((1, 1, d_model))*-1e30 for i in range(n_layers*5)]

# Dummy call
_o, _s = blocklm(tf.constant([[0]]), get_init_state())
_ = head(_o)

blocklm.load_weights("blocklm.weights.h5")
head.load_weights("head.weights.h5")

# --- 3. ์ถ”๋ก  ์—”์ง„ ---
class InferenceEngine:
    def __init__(self, model, head, sp):
        self.model = model
        self.head = head
        self.sp = sp
        self.pad_id = sp.piece_to_id("<pad>") if sp.piece_to_id("<pad>") != -1 else 0
        self.eos_id = sp.piece_to_id("</s>") if sp.piece_to_id("</s>") != -1 else sp.piece_to_id("[EOS]")

    def apply_repetition_penalty(self, logits, generated_ids, penalty, window=64):
        if not generated_ids: return logits
        recent_ids = set(generated_ids[-window:])
        for token_id in recent_ids:
            if logits[token_id] > 0: logits[token_id] /= penalty
            else: logits[token_id] *= penalty
        return logits

    def sample(self, logits, temp, top_k, top_p):
        if temp <= 0: return np.argmax(logits)
        logits = logits / temp
        if top_k > 0:
            indices_to_remove = logits < np.sort(logits)[-min(top_k, logits.shape[-1])]
            logits[indices_to_remove] = -float('inf')
        probs = tf.nn.softmax(logits).numpy()
        sorted_indices = np.argsort(probs)[::-1]
        sorted_probs = probs[sorted_indices]
        cumulative_probs = np.cumsum(sorted_probs)
        idx_to_remove = cumulative_probs > top_p
        if np.any(idx_to_remove):
            cutoff_idx = max(1, np.where(idx_to_remove)[0][0] + 1)
            probs[sorted_indices[cutoff_idx:]] = 0
            if np.sum(probs) > 0: probs /= np.sum(probs)
            else: probs[sorted_indices[0]] = 1.0
        return np.random.choice(len(probs), p=probs)

    @tf.function(reduce_retracing=True)
    def model_step(self, token_id, states):
        out, next_states = self.model(token_id, states)
        logits = self.head(out)
        return logits, next_states

    def generate(self, prompt, max_new_tokens, temp, top_k, top_p, penalty):
        input_ids = self.sp.encode(prompt)
        states = get_init_state()
        generated = []
        if len(input_ids) > 1:
            for i in range(len(input_ids) - 1):
                _, states = self.model_step(tf.constant([[input_ids[i]]]), states)
        curr_token_id = input_ids[-1]
        prev_text = ""
        for _ in range(max_new_tokens):
            logits_out, states = self.model_step(tf.constant([[curr_token_id]]), states)
            logits = logits_out[0, 0].numpy()
            logits = self.apply_repetition_penalty(logits, input_ids + generated, penalty)
            logits[self.pad_id] = -float('inf')
            next_id = int(self.sample(logits, temp, top_k, top_p))
            if next_id == self.eos_id: break
            generated.append(next_id)
            full_text = self.sp.decode(generated)
            new_part = full_text[len(prev_text):]
            if new_part:
                yield new_part
                prev_text = full_text
            curr_token_id = next_id

engine = InferenceEngine(blocklm, head, sp)

# --- 4. Gradio UI (๋‹จ์ˆœ ํ…์ŠคํŠธ ์ž…์ถœ๋ ฅ ๋ฐฉ์‹) ---
with gr.Blocks(title="RWKV Text Generator") as demo:
    gr.Markdown("## ๐Ÿ–‹๏ธ Dynamic RWKV Text Generation")
    gr.Markdown("์งˆ๋ฌธ์„ ์ž…๋ ฅํ•˜๊ณ  Generate๋ฅผ ๋ˆ„๋ฅด๋ฉด ๋‹ต๋ณ€์ด ์•„๋ž˜ ํ…์ŠคํŠธ ๋ฐ•์Šค์— ์‹ค์‹œ๊ฐ„์œผ๋กœ ์ƒ์„ฑ๋ฉ๋‹ˆ๋‹ค.")
    
    with gr.Row():
        with gr.Column():
            input_text = gr.Textbox(lines=5, label="Input Prompt", placeholder="์—ฌ๊ธฐ์— ์งˆ๋ฌธ์ด๋‚˜ ๋ฌธ์žฅ์„ ์ž…๋ ฅํ•˜์„ธ์š”...")
            with gr.Row():
                temp_slider = gr.Slider(0, 2, value=0.7, label="Temperature")
                top_p_slider = gr.Slider(0, 1, value=0.92, label="Top-P")
            with gr.Row():
                penalty_slider = gr.Slider(1, 2, value=1.2, label="Penalty")
                max_tokens = gr.Slider(1, 2048, value=512, step=1, label="Max Tokens")
            
            submit_btn = gr.Button("Generate", variant="primary")
            clear_btn = gr.Button("Clear")

        with gr.Column():
            output_text = gr.Textbox(lines=15, label="Generated Output", interactive=False)

    def run_generation(prompt, tokens, temp, top_p, penalty):
        if not prompt.strip():
            return "ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ž…๋ ฅํ•ด์ฃผ์„ธ์š”."
        
        full_prompt = f"Question: {prompt}\nAnswer:"
        current_output = ""
        
        for chunk in engine.generate(full_prompt, int(tokens), temp, 40, top_p, penalty):
            current_output += chunk
            yield current_output

    # ๋ฒ„ํŠผ ํด๋ฆญ ๋ฐ ์—”ํ„ฐ ํ‚ค ์ž…๋ ฅ ์ด๋ฒคํŠธ
    submit_btn.click(
        fn=run_generation, 
        inputs=[input_text, max_tokens, temp_slider, top_p_slider, penalty_slider], 
        outputs=output_text
    )
    
    clear_btn.click(lambda: ("", ""), outputs=[input_text, output_text])

if __name__ == "__main__":
    demo.queue().launch()