File size: 22,334 Bytes
0747201
8f4d405
 
0747201
 
8f4d405
 
927854c
8f4d405
 
 
 
927854c
 
0747201
8f4d405
927854c
 
 
 
 
0747201
8f4d405
927854c
0747201
 
 
 
 
927854c
0747201
 
 
 
 
927854c
 
0747201
 
927854c
0747201
 
927854c
 
0747201
 
 
 
927854c
0747201
 
f608c70
0747201
 
8f4d405
 
 
0747201
927854c
 
 
 
 
 
 
 
8f4d405
0747201
8f4d405
5787d0a
0747201
 
 
 
 
 
 
 
 
927854c
5787d0a
0747201
 
5787d0a
0747201
5787d0a
 
0747201
5787d0a
8f4d405
0747201
5787d0a
 
 
0747201
5787d0a
 
0747201
5787d0a
8f4d405
0747201
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8f4d405
0747201
 
927854c
0747201
 
927854c
0747201
 
 
 
 
 
 
 
8d4bf4a
0747201
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a9135e0
0747201
 
 
 
927854c
0747201
 
 
 
927854c
8f4d405
 
0747201
 
 
 
 
 
 
 
 
 
 
 
 
927854c
0747201
 
927854c
0747201
 
 
 
043d8e9
a67b5f4
0747201
927854c
0747201
 
 
44b41ba
 
a9135e0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ca77f38
0747201
a9135e0
ca77f38
 
 
 
 
a9135e0
 
 
 
 
 
 
 
 
 
 
 
ca77f38
 
a9135e0
 
 
 
0747201
927854c
0747201
 
927854c
 
8f4d405
0747201
927854c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0747201
 
 
 
 
927854c
 
 
8f4d405
 
927854c
8f4d405
 
 
 
 
 
 
 
 
 
0747201
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8f4d405
 
0747201
927854c
0747201
 
927854c
0747201
 
 
 
 
 
 
 
 
 
 
 
927854c
 
 
0747201
927854c
 
 
 
0747201
 
 
 
 
 
 
 
 
 
 
 
927854c
 
8f4d405
927854c
 
 
 
 
 
 
8f4d405
927854c
 
 
5787d0a
927854c
 
8f4d405
927854c
 
8f4d405
927854c
 
 
 
 
 
 
8f4d405
927854c
 
 
 
 
 
 
 
 
8f4d405
 
 
 
 
 
 
 
927854c
8f4d405
927854c
 
 
 
 
8f4d405
 
 
 
 
 
 
 
 
 
 
 
 
 
927854c
8f4d405
 
 
 
927854c
bd329bc
8f4d405
 
 
 
927854c
8f4d405
 
 
 
927854c
8f4d405
 
927854c
8f4d405
 
 
 
927854c
bd329bc
 
 
927854c
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
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
# llm_router.py - ZeroGPU Chat API (RunPod)
import logging
import asyncio
import aiohttp
import time
from typing import Dict, Optional
from .models_config import LLM_CONFIG
from .config import get_settings

logger = logging.getLogger(__name__)

class LLMRouter:
    def __init__(self, hf_token=None, use_local_models: bool = False):
        """
        Initialize LLM Router with ZeroGPU Chat API (RunPod).
        
        Args:
            hf_token: Not used (kept for backward compatibility)
            use_local_models: Must be False (local models disabled)
        """
        if use_local_models:
            raise ValueError("Local models are disabled. Only ZeroGPU Chat API is supported.")
        
        self.settings = get_settings()
        self.base_url = self.settings.zerogpu_base_url.rstrip('/')
        self.access_token = None
        self.refresh_token = None
        self.token_expires_at = 0
        self.session = None
        
        # Validate base URL
        if not self.settings.zerogpu_base_url:
            raise ValueError(
                "ZEROGPU_BASE_URL is required. "
                "Set it in environment variables or .env file"
            )
        
        # Validate credentials
        if not self.settings.zerogpu_email or not self.settings.zerogpu_password:
            raise ValueError(
                "ZEROGPU_EMAIL and ZEROGPU_PASSWORD are required. "
                "Set them in environment variables or .env file"
            )
        
        logger.info("ZeroGPU Chat API client initializing")
        logger.info(f"Base URL: {self.base_url}")
        
        # Initialize session and authenticate
        try:
            # Authentication will happen on first request if needed
            logger.info("ZeroGPU Chat API client initialized (authentication on first request)")
        except Exception as e:
            logger.error(f"Failed to initialize ZeroGPU Chat API client: {e}")
            raise RuntimeError(f"Could not initialize ZeroGPU Chat API client: {e}") from e
    
    async def route_inference(self, task_type: str, prompt: str, **kwargs):
        """
        Route inference to ZeroGPU Chat API.
        
        Args:
            task_type: Type of task (general_reasoning, intent_classification, etc.)
            prompt: Input prompt
            **kwargs: Additional parameters (max_tokens, temperature, etc.)
        
        Returns:
            Generated text response
        """
        logger.info(f"Routing inference to ZeroGPU Chat API for task: {task_type}")
        
        try:
            # Ensure authenticated
            await self._ensure_authenticated()
            
            # Map internal task types to API task types
            api_task = self._map_task_type(task_type)
            
            # Pass original task type for model config lookup
            kwargs['original_task_type'] = task_type
            
            # Handle embedding generation (may need special handling)
            if task_type == "embedding_generation":
                logger.warning("Embedding generation via ZeroGPU API may require special implementation")
                result = await self._call_zerogpu_api(api_task, prompt, **kwargs)
            else:
                result = await self._call_zerogpu_api(api_task, prompt, **kwargs)
            
            if result is None:
                logger.error(f"ZeroGPU Chat API returned None for task: {task_type}")
                raise RuntimeError(f"Inference failed for task: {task_type}")
            
            logger.info(f"Inference complete for {task_type} (ZeroGPU Chat API)")
            return result
            
        except Exception as e:
            logger.error(f"ZeroGPU Chat API inference failed: {e}", exc_info=True)
            raise RuntimeError(
                f"Inference failed for task: {task_type}. "
                f"ZeroGPU Chat API error: {e}"
            ) from e
    
    async def _ensure_authenticated(self):
        """Ensure we have a valid access token, login if needed."""
        # Check if token is expired (with 60 second buffer)
        if self.access_token and time.time() < (self.token_expires_at - 60):
            return
        
        # Create session if needed
        if self.session is None:
            self.session = aiohttp.ClientSession()
        
        # Login to get tokens
        await self._login()
    
    async def _login(self):
        """Login to ZeroGPU Chat API and get access/refresh tokens."""
        try:
            login_url = f"{self.base_url}/login"
            login_data = {
                "email": self.settings.zerogpu_email,
                "password": self.settings.zerogpu_password
            }
            
            async with self.session.post(login_url, json=login_data) as response:
                if response.status == 401:
                    raise ValueError("Invalid email or password for ZeroGPU Chat API")
                response.raise_for_status()
                data = await response.json()
                
                self.access_token = data.get("access_token")
                self.refresh_token = data.get("refresh_token")
                
                # Access tokens typically expire in 15 minutes (900 seconds)
                self.token_expires_at = time.time() + 900
                
                logger.info("Successfully authenticated with ZeroGPU Chat API")
                
        except aiohttp.ClientError as e:
            logger.error(f"Failed to login to ZeroGPU Chat API: {e}")
            raise RuntimeError(f"Authentication failed: {e}") from e
    
    async def _refresh_token(self):
        """Refresh access token using refresh token."""
        try:
            refresh_url = f"{self.base_url}/refresh"
            headers = {"X-Refresh-Token": self.refresh_token}
            
            async with self.session.post(refresh_url, headers=headers) as response:
                if response.status == 401:
                    # Refresh token expired, need to login again
                    await self._login()
                    return
                
                response.raise_for_status()
                data = await response.json()
                
                self.access_token = data.get("access_token")
                self.refresh_token = data.get("refresh_token")
                self.token_expires_at = time.time() + 900
                
                logger.info("Successfully refreshed ZeroGPU Chat API token")
                
        except aiohttp.ClientError as e:
            logger.error(f"Failed to refresh token: {e}")
            # Try login as fallback
            await self._login()
    
    def _map_task_type(self, internal_task: str) -> str:
        """Map internal task types to ZeroGPU Chat API task types."""
        task_mapping = {
            "general_reasoning": "general",
            "response_synthesis": "general",
            "intent_classification": "classification",
            "safety_check": "classification",
            "embedding_generation": "embedding"
        }
        return task_mapping.get(internal_task, "general")
    
    async def _call_zerogpu_api(self, task: str, prompt: str, **kwargs) -> Optional[str]:
        """Call ZeroGPU Chat API for inference."""
        if not self.session:
            self.session = aiohttp.ClientSession()
        
        # Store original task type for model config lookup
        original_task = kwargs.pop('original_task_type', None)
        
        # Get model config for defaults
        model_config = self._select_model(original_task or 'general_reasoning')
        
        # Build request payload according to API documentation
        payload = {
            "message": prompt,
            "task": task,
            "max_tokens": kwargs.get('max_tokens', model_config.get('max_tokens', 512)),
            "temperature": kwargs.get('temperature', model_config.get('temperature', 0.7)),
            "top_p": kwargs.get('top_p', model_config.get('top_p', 0.9)),
        }
        
        # Add optional parameters
        if 'context' in kwargs and kwargs['context']:
            # Convert context to API format if needed
            context = kwargs['context']
            if isinstance(context, list) and len(context) > 0:
                # Convert to API format: list of dicts with role, content, timestamp
                api_context = []
                for item in context[:50]:  # Max 50 messages
                    if isinstance(item, (list, tuple)) and len(item) >= 2:
                        # Format: [user_msg, assistant_msg]
                        api_context.append({
                            "role": "user",
                            "content": str(item[0]),
                            "timestamp": kwargs.get('timestamp', time.time())
                        })
                        api_context.append({
                            "role": "assistant",
                            "content": str(item[1]),
                            "timestamp": kwargs.get('timestamp', time.time())
                        })
                    elif isinstance(item, dict):
                        api_context.append(item)
                payload["context"] = api_context
        
        if 'system_prompt' in kwargs and kwargs['system_prompt']:
            payload["system_prompt"] = kwargs['system_prompt']
        if 'repetition_penalty' in kwargs:
            payload["repetition_penalty"] = kwargs['repetition_penalty']
        
        # Prepare headers
        headers = {
            "Authorization": f"Bearer {self.access_token}",
            "Content-Type": "application/json"
        }
        
        try:
            chat_url = f"{self.base_url}/chat"
            
            async with self.session.post(chat_url, json=payload, headers=headers) as response:
                # Handle token expiration
                if response.status == 401:
                    logger.info("Token expired, refreshing...")
                    await self._refresh_token()
                    headers["Authorization"] = f"Bearer {self.access_token}"
                    # Retry request
                    async with self.session.post(chat_url, json=payload, headers=headers) as retry_response:
                        retry_response.raise_for_status()
                        data = await retry_response.json()
                        return data.get("response")
                
                response.raise_for_status()
                data = await response.json()
                
                # Extract response from API
                result = data.get("response")
                if result:
                    logger.info(f"ZeroGPU Chat API generated response (length: {len(result)})")
                    return result
                else:
                    logger.error("ZeroGPU Chat API returned empty response")
                    return None
                    
        except aiohttp.ClientError as e:
            logger.error(f"Error calling ZeroGPU Chat API: {e}", exc_info=True)
            raise
    
    def _calculate_safe_max_tokens(self, prompt: str, requested_max_tokens: int) -> int:
        """
        Calculate safe max_tokens based on input token count and model context window.
        
        Args:
            prompt: Input prompt text
            requested_max_tokens: Desired max_tokens value
            
        Returns:
            int: Adjusted max_tokens that fits within context window
        """
        # Estimate input tokens (rough: 1 token ≈ 4 characters)
        # For more accuracy, you could use tiktoken if available
        input_tokens = len(prompt) // 4
        
        # Get model context window from settings
        context_window = self.settings.zerogpu_model_context_window
        
        logger.debug(
            f"Calculating safe max_tokens: input ~{input_tokens} tokens, "
            f"context_window={context_window}, requested={requested_max_tokens}"
        )
        
        # Reserve minimum 100 tokens for safety margin
        available_tokens = context_window - input_tokens - 100
        
        # Use the smaller of requested or available
        safe_max_tokens = min(requested_max_tokens, available_tokens)
        
        # Ensure minimum of 50 tokens for output
        safe_max_tokens = max(50, safe_max_tokens)
        
        if safe_max_tokens < requested_max_tokens:
            logger.warning(
                f"Reduced max_tokens from {requested_max_tokens} to {safe_max_tokens} "
                f"(input: ~{input_tokens} tokens, context window: {context_window} tokens, "
                f"available: {available_tokens} tokens)"
            )
        
        return safe_max_tokens
            
    def _format_prompt(self, prompt: str, task_type: str, model_config: dict) -> str:
        """
        Format prompt for ZeroGPU Chat API.
        Can be customized based on model requirements.
        """
        formatted_prompt = prompt
        
        # Add math directive for mathematical problems if needed
        if self._is_math_query(prompt):
            math_directive = "Please reason step by step, and put your final answer within \\boxed{}."
            formatted_prompt = f"{formatted_prompt}\n\n{math_directive}"
        
        return formatted_prompt
    
    def _is_math_query(self, prompt: str) -> bool:
        """Detect if query is mathematical"""
        math_keywords = [
            "solve", "calculate", "compute", "equation", "formula", 
            "mathematical", "algebra", "geometry", "calculus", "integral",
            "derivative", "theorem", "proof", "problem"
        ]
        prompt_lower = prompt.lower()
        return any(keyword in prompt_lower for keyword in math_keywords)
    
    def _clean_reasoning_tags(self, text: str) -> str:
        """Clean up reasoning tags from response if present"""
        if not text:
            return text
        # Remove common reasoning tags if present
        text = text.replace("`<think>`", "").replace("`</think>`", "")
        text = text.replace("`<think>`", "").replace("`</think>`", "")
        text = text.strip()
        return text
    
    def _select_model(self, task_type: str) -> dict:
        """Select model configuration based on task type"""
        model_map = {
            "intent_classification": LLM_CONFIG["models"]["classification_specialist"],
            "embedding_generation": LLM_CONFIG["models"]["embedding_specialist"],
            "safety_check": LLM_CONFIG["models"]["safety_checker"],
            "general_reasoning": LLM_CONFIG["models"]["reasoning_primary"],
            "response_synthesis": LLM_CONFIG["models"]["reasoning_primary"]
        }
        return model_map.get(task_type, LLM_CONFIG["models"]["reasoning_primary"])
    
    async def get_available_models(self):
        """Get list of available models from ZeroGPU Chat API"""
        try:
            await self._ensure_authenticated()
            if not self.session:
                self.session = aiohttp.ClientSession()
            
            tasks_url = f"{self.base_url}/tasks"
            headers = {"Authorization": f"Bearer {self.access_token}"}
            
            async with self.session.get(tasks_url, headers=headers) as response:
                if response.status == 401:
                    await self._refresh_token()
                    headers["Authorization"] = f"Bearer {self.access_token}"
                    async with self.session.get(tasks_url, headers=headers) as retry_response:
                        retry_response.raise_for_status()
                        data = await retry_response.json()
                else:
                    response.raise_for_status()
                    data = await response.json()
                
                tasks = data.get("tasks", {})
                models = [f"ZeroGPU Chat API - {task}: {info.get('model', 'N/A')}" 
                         for task, info in tasks.items()]
                return models if models else ["ZeroGPU Chat API"]
        except Exception as e:
            logger.error(f"Failed to get available models: {e}")
            return ["ZeroGPU Chat API"]
    
    async def health_check(self):
        """Perform health check on ZeroGPU Chat API"""
        try:
            if not self.session:
                self.session = aiohttp.ClientSession()
            
            # Check health endpoint (no auth required)
            health_url = f"{self.base_url}/health"
            async with self.session.get(health_url) as response:
                response.raise_for_status()
                data = await response.json()
                
                return {
                    "provider": "zerogpu_chat_api",
                    "status": "healthy" if data.get("status") == "healthy" else "unhealthy",
                    "models_ready": data.get("models_ready", False),
                    "base_url": self.base_url
                }
        except Exception as e:
            logger.error(f"Health check failed: {e}")
            return {
                "provider": "zerogpu_chat_api",
                "status": "unhealthy",
                "error": str(e)
            }
    
    async def __aenter__(self):
        """Async context manager entry"""
        if not self.session:
            self.session = aiohttp.ClientSession()
        return self
    
    async def __aexit__(self, exc_type, exc_val, exc_tb):
        """Async context manager exit"""
        if self.session:
            await self.session.close()
            self.session = None
    
    def prepare_context_for_llm(self, raw_context: Dict, max_tokens: Optional[int] = None, 
                                user_input: Optional[str] = None) -> str:
        """
        Smart context windowing with user input priority.
        User input is NEVER truncated - context is reduced to fit.
        
        Args:
            raw_context: Context dictionary
            max_tokens: Optional override (uses config default if None)
            user_input: Optional explicit user input (takes priority over raw_context['user_input'])
        """
        # Use config budget if not provided
        if max_tokens is None:
            max_tokens = self.settings.context_preparation_budget
        
        # Get user input (explicit parameter takes priority)
        actual_user_input = user_input or raw_context.get('user_input', '')
        
        # Calculate user input tokens (simple estimation: 1 token ≈ 4 chars)
        user_input_tokens = len(actual_user_input) // 4
        
        # Ensure user input fits within dedicated budget
        user_input_max = self.settings.user_input_max_tokens
        if user_input_tokens > user_input_max:
            logger.warning(f"User input ({user_input_tokens} tokens) exceeds max ({user_input_max}), truncating")
            max_chars = user_input_max * 4
            actual_user_input = actual_user_input[:max_chars - 3] + "..."
            user_input_tokens = user_input_max
        
        # Reserve space for user input (it has highest priority)
        remaining_tokens = max_tokens - user_input_tokens
        if remaining_tokens < 0:
            logger.warning(f"User input ({user_input_tokens} tokens) exceeds total budget ({max_tokens})")
            remaining_tokens = 0
        
        logger.info(f"Token allocation: User input={user_input_tokens}, Context budget={remaining_tokens}, Total={max_tokens}")
        
        # Priority order for context elements (user input already handled)
        priority_elements = [
            ('recent_interactions', 0.8),
            ('user_preferences', 0.6),
            ('session_summary', 0.4),
            ('historical_context', 0.2)
        ]
        
        formatted_context = []
        total_tokens = user_input_tokens  # Start with user input tokens
        
        # Add user input first (unconditionally, never truncated)
        if actual_user_input:
            formatted_context.append(f"=== USER INPUT ===\n{actual_user_input}")
        
        # Now add context elements within remaining budget
        for element, priority in priority_elements:
            element_key_map = {
                'recent_interactions': raw_context.get('interaction_contexts', []),
                'user_preferences': raw_context.get('preferences', {}),
                'session_summary': raw_context.get('session_context', {}),
                'historical_context': raw_context.get('user_context', '')
            }
            
            content = element_key_map.get(element, '')
            
            # Convert to string if needed
            if isinstance(content, dict):
                content = str(content)
            elif isinstance(content, list):
                content = "\n".join([str(item) for item in content[:10]])
            
            if not content:
                continue
            
            # Estimate tokens (simple: 1 token ≈ 4 chars)
            tokens = len(content) // 4
            
            if total_tokens + tokens <= max_tokens:
                formatted_context.append(f"=== {element.upper()} ===\n{content}")
                total_tokens += tokens
            elif priority > 0.5 and remaining_tokens > 0:  # Critical elements - truncate if needed
                available = max_tokens - total_tokens
                if available > 100:  # Only truncate if we have meaningful space
                    truncated = self._truncate_to_tokens(content, available)
                    formatted_context.append(f"=== {element.upper()} (TRUNCATED) ===\n{truncated}")
                    total_tokens += available
                break
        
        logger.info(f"Context prepared: {total_tokens}/{max_tokens} tokens (user input: {user_input_tokens}, context: {total_tokens - user_input_tokens})")
        return "\n\n".join(formatted_context)
    
    def _truncate_to_tokens(self, content: str, max_tokens: int) -> str:
        """Truncate content to fit within token limit"""
        # Simple character-based truncation (1 token ≈ 4 chars)
        max_chars = max_tokens * 4
        if len(content) <= max_chars:
            return content
        return content[:max_chars - 3] + "..."