File size: 11,325 Bytes
4b7d6aa
 
dd20db5
4b7d6aa
 
 
 
18f8fd1
4b7d6aa
 
 
 
 
 
 
 
 
 
45891c7
725a466
45891c7
 
4b7d6aa
45891c7
 
4b7d6aa
45891c7
4b7d6aa
 
 
 
45891c7
4b7d6aa
 
 
45891c7
 
 
 
 
 
 
 
4b7d6aa
45891c7
 
 
 
4b7d6aa
 
 
45891c7
 
 
 
 
 
 
 
 
 
 
 
 
4b7d6aa
45891c7
 
61063bc
45891c7
61063bc
92fd13b
45891c7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4b7d6aa
45891c7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
92fd13b
45891c7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
bc62c3d
 
4b7d6aa
 
45891c7
4b7d6aa
 
45891c7
 
 
 
 
c71e1ae
7b2a773
 
92fd13b
dd20db5
 
92fd13b
dd20db5
 
 
92fd13b
 
 
45891c7
dd20db5
45891c7
92fd13b
af1487b
45891c7
c71e1ae
4b7d6aa
 
 
 
45891c7
4b7d6aa
 
af1487b
4b7d6aa
 
 
45891c7
 
4b7d6aa
45891c7
4b7d6aa
45891c7
 
 
 
 
 
4b7d6aa
45891c7
 
 
 
4b7d6aa
 
 
45891c7
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
235
236
237
238
239
240
241
242
243
import os
import re
import sys
import torch
import torch.nn as nn
import torch.nn.functional as F
from dataclasses import make_dataclass
from tokenizers import Tokenizer, decoders
from flask import Flask, request, jsonify
from flask_cors import CORS

# ==============================================================================
# 1. CONFIGURATION
# ==============================================================================
BASE_DIR = os.path.dirname(os.path.abspath(__file__))
MODEL_DIR = os.path.join(BASE_DIR, "saved_models")
TOKEN_PATH = os.path.join(MODEL_DIR, "bpe-tokenizer.json")

# Update this to your latest best model name!
MODEL_PATH = os.path.join(MODEL_DIR, "gemma_3_1b_distill_qwen_2.5_0.5b_distill_aai_mini+_new_260322.pth")

DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
SEQ_LENGTH = 256

# Updated to match the new 384-dimension architecture
MODEL_CONFIG_PARAMS = {
    'MODEL_DIM': 384, 'NUM_HEADS': 6, 'NUM_LAYERS': 6, 'DROPOUT': 0.1, 'SEQ_LENGTH': SEQ_LENGTH
}
print(f"INFO: Using device: {DEVICE}")

# ==============================================================================
# 2. AI MODEL ARCHITECTURE
# ==============================================================================
class DecoderBlock(nn.Module):
    def __init__(self, d_model, nhead, dropout):
        super().__init__()
        self.d_model, self.nhead = d_model, nhead
        assert d_model % nhead == 0, "d_model must be divisible by nhead"
        self.ln1, self.qkv_proj = nn.LayerNorm(d_model), nn.Linear(d_model, d_model * 3)
        self.out_proj, self.dropout1 = nn.Linear(d_model, d_model), nn.Dropout(dropout)
        self.ln2, self.ffwd = nn.LayerNorm(d_model), nn.Sequential(
            nn.Linear(d_model, d_model * 4), nn.GELU(),
            nn.Linear(d_model * 4, d_model), nn.Dropout(dropout))
    def forward(self, x):
        B, T, D = x.shape; x_norm = self.ln1(x); q, k, v = self.qkv_proj(x_norm).chunk(3, dim=-1)
        head_dim = D // self.nhead; q,k,v = (t.view(B, T, self.nhead, head_dim).transpose(1, 2) for t in (q,k,v))
        attn_output = F.scaled_dot_product_attention(q, k, v, is_causal=True).transpose(1, 2).contiguous().view(B, T, D)
        x = x + self.dropout1(self.out_proj(attn_output)); x = x + self.ffwd(self.ln2(x)); return x

class GPTTransformerLM(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.tok_emb = nn.Embedding(config.VOCAB_SIZE, config.MODEL_DIM)
        self.pos_emb = nn.Embedding(config.SEQ_LENGTH, config.MODEL_DIM)
        self.dropout = nn.Dropout(config.DROPOUT)
        self.layers = nn.ModuleList([DecoderBlock(config.MODEL_DIM, config.NUM_HEADS, config.DROPOUT) for _ in range(config.NUM_LAYERS)])
        self.ln_f = nn.LayerNorm(config.MODEL_DIM)
        self.head = nn.Linear(config.MODEL_DIM, config.VOCAB_SIZE, bias=False)
        
        # Tie weights
        self.tok_emb.weight = self.head.weight 
        self.register_buffer('pos', torch.arange(config.SEQ_LENGTH, dtype=torch.long))

    def forward(self, x):
        B, T = x.size()
        h = self.tok_emb(x) + self.pos_emb(self.pos[:T])
        h = self.dropout(h)
        for layer in self.layers: h = layer(x=h)
        return self.head(self.ln_f(h))

# ==============================================================================
# 3. NANO-REASONING GENERATION LOGIC
# ==============================================================================
@torch.inference_mode()
def generate_full_response(model, tokenizer, prompt, max_new_tokens=256, temp=0.45, top_k=50, top_p=0.9, repetition_penalty=1.15, nano_samples=25):
    model.eval()
    prompt_ids = tokenizer.encode(prompt).ids
    gen_ids = prompt_ids[:] 
    
    x = torch.tensor(prompt_ids, dtype=torch.long, device=DEVICE).unsqueeze(0)
    
    # --- PROTECTED TOKENS ---
    protected_tokens = set()
    if getattr(tokenizer, 'token_to_id', None):
        eos = tokenizer.token_to_id('[EOS]')
        if eos is not None: protected_tokens.add(eos)
    for char in ['\n', '\n\n', '.', ',', '!', '?', ' ', ' \n', '"', "'", ':', ';']:
        protected_tokens.update(tokenizer.encode(char).ids)
    
    bf16_supported = (torch.cuda.is_available() and torch.cuda.is_bf16_supported())
    autocast_dtype = torch.bfloat16 if bf16_supported else torch.float16
    autocast_kwargs = {'device_type': DEVICE.type, 'dtype': autocast_dtype, 'enabled': (DEVICE.type == 'cuda')}
    
    hard_limit = max_new_tokens + 100

    for step in range(hard_limit):
        x_cond = x if x.size(1) <= model.config.SEQ_LENGTH else x[:, -model.config.SEQ_LENGTH:]
        
        with torch.autocast(**autocast_kwargs):
            logits = model(x_cond)[:, -1, :] 
        
        # 1. DYNAMIC Frequency Penalty
        if repetition_penalty != 1.0:
            penalty_window = gen_ids[-128:] 
            for token_id in set(penalty_window):
                if token_id not in protected_tokens:
                    count = penalty_window.count(token_id)
                    dynamic_penalty = repetition_penalty + (count * 0.03)
                    if logits[0, token_id] < 0:
                        logits[0, token_id] *= dynamic_penalty
                    else:
                        logits[0, token_id] /= dynamic_penalty
        
        # 2. Temperature & Sampling
        if temp > 0.0:
            logits = logits / temp
            if top_k > 0:
                v, _ = torch.topk(logits, min(top_k, logits.size(-1)))
                logits[logits < v[:, [-1]]] = -float('Inf')
            if top_p > 0.0 and top_p < 1.0:
                sorted_logits, sorted_indices = torch.sort(logits, descending=True)
                cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
                sorted_indices_to_remove = cumulative_probs > top_p
                sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
                sorted_indices_to_remove[..., 0] = 0
                indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
                logits[indices_to_remove] = -float('Inf')
                
            probs = F.softmax(logits, dim=-1)
            
            # --- NANO-LOOKAHEAD REASONING ---
            simulations = torch.multinomial(probs, num_samples=nano_samples, replacement=True)
            unique_ids, counts = torch.unique(simulations[0], return_counts=True)
            
            if len(unique_ids) == 1:
                next_id = unique_ids[0].item()
            else:
                sorted_indices = torch.argsort(counts, descending=True)
                top_2_ids = unique_ids[sorted_indices[:2]]
                best_score = -1.0
                best_id = top_2_ids[0].item() 
                
                for candidate_id in top_2_ids:
                    candidate_prob = probs[0, candidate_id].item()
                    if x.size(1) < model.config.SEQ_LENGTH:
                        temp_x = torch.cat((x, torch.tensor([[candidate_id]], device=DEVICE)), dim=1)
                        temp_x_cond = temp_x if temp_x.size(1) <= model.config.SEQ_LENGTH else temp_x[:, -model.config.SEQ_LENGTH:]
                        with torch.autocast(**autocast_kwargs):
                            next_logits = model(temp_x_cond)[:, -1, :]
                            next_max_prob = torch.max(F.softmax(next_logits / temp, dim=-1)).item()
                        score = candidate_prob * next_max_prob
                        if score > best_score:
                            best_score = score
                            best_id = candidate_id.item()
                next_id = best_id
            # ---------------------------
        else: 
            next_id = torch.argmax(logits, dim=-1).item()
            
        gen_ids.append(next_id)
        x = torch.cat((x, torch.tensor([[next_id]], device=DEVICE)), dim=1)
        
        # --- STOP CONDITIONS ---
        # Stop on double newline
        tail_text = tokenizer.decode(gen_ids[-8:])
        if "\n\n" in tail_text.replace(" ", "").replace("\r", ""):
            break
        # Stop on User turn hallucination
        if "\nUser:" in tail_text or "User:" in tail_text.replace("\n", "")[-6:]:
            break
        # Stop on EOS token
        if next_id == getattr(tokenizer, 'token_to_id', lambda x: -1)('[EOS]'): 
            break
        # Graceful Sentence Cutoff
        if step >= max_new_tokens:
            new_text = tokenizer.decode([next_id])
            if any(punct in new_text for punct in ['.', '!', '?', '\n']):
                break
                
    # 🚀 FIXED: Return the full, unbroken string (Prompt + Continuation)
    return tokenizer.decode(gen_ids)

# ==============================================================================
# 4. LOAD MODEL & TOKENIZER ON STARTUP
# ==============================================================================
try:
    print("INFO: Loading tokenizer...")
    tokenizer = Tokenizer.from_file(TOKEN_PATH)
    vocab_size = tokenizer.get_vocab_size()
    MODEL_CONFIG_PARAMS['VOCAB_SIZE'] = vocab_size
    print(f"INFO: Tokenizer loaded. Vocab size: {vocab_size}")
    
    TuneConfig = make_dataclass("TuneConfig", [(k, type(v)) for k, v in MODEL_CONFIG_PARAMS.items()])
    setattr(sys.modules['__main__'], 'TuneConfig', TuneConfig)
    
    model_config_instance = TuneConfig(**MODEL_CONFIG_PARAMS)
    model = GPTTransformerLM(model_config_instance).to(DEVICE)
    
    print(f"INFO: Loading checkpoint dictionary from: {MODEL_PATH}")
    checkpoint = torch.load(MODEL_PATH, map_location=DEVICE, weights_only=False)
    
    state_dict = checkpoint
    if 'model_state_dict' in checkpoint: state_dict = checkpoint['model_state_dict']
    elif 'state_dict' in checkpoint: state_dict = checkpoint['state_dict']
    
    cleaned_state_dict = {k.replace('_orig_mod.', ''): v for k, v in state_dict.items()}
    model.load_state_dict(cleaned_state_dict, strict=False)
    
    model.eval()
    print("✅ INFO: Backend Nano-Reasoning Model loaded and ready for API requests!")
    
except Exception as e:
    print(f"❌ FATAL ERROR during model loading: {e}"); exit()

# ==============================================================================
# 5. FLASK SERVER APPLICATION
# ==============================================================================
app = Flask(__name__)
CORS(app) 

@app.route("/generate", methods=["POST"])
def handle_generation():
    data = request.get_json()
    prompt = data.get("prompt")
    if not prompt: return jsonify({"error": "Missing 'prompt'"}), 400
    
    try:
        print(f"INFO: Received prompt: '{prompt}'")
        # Generate clean response using hidden Nano-Reasoning math
        generated_text = generate_full_response(model, tokenizer, prompt)
        print(f"INFO: Generated response: '{generated_text}'")
        
        return jsonify({"response": generated_text})
    except Exception as e:
        import traceback
        traceback.print_exc()
        print(f"❌ ERROR during generation: {e}")
        return jsonify({"error": "Failed to generate response"}), 500

@app.route("/")
def health_check():
    return "Nano-Reasoning Backend server is running. Use the /generate endpoint to interact with the AI."