File size: 15,265 Bytes
c8b77b5 |
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 |
#!/usr/bin/env python3
"""
Phi-3.5-MoE Integration Module
This module integrates Microsoft's Phi-3.5-MoE-instruct model as a baseline
for expanding our current MoE (Mixture of Experts) framework.
Based on: https://huggingface.co/microsoft/Phi-3.5-MoE-instruct
- 41.9B parameters
- 128K context length
- Multilingual support
- Strong reasoning capabilities (code, math, logic)
"""
import torch
import logging
from typing import Dict, List, Optional, Any, Union
from transformers import (
AutoModelForCausalLM,
AutoTokenizer,
pipeline,
Pipeline
)
from pathlib import Path
import json
import asyncio
from dataclasses import dataclass
logger = logging.getLogger(__name__)
@dataclass
class Phi35MoEConfig:
"""Configuration for Phi-3.5-MoE integration"""
model_name: str = "microsoft/Phi-3.5-MoE-instruct"
device_map: str = "auto"
torch_dtype: str = "auto"
trust_remote_code: bool = False
max_new_tokens: int = 500
temperature: float = 0.7
top_p: float = 0.9
do_sample: bool = True
context_length: int = 128000 # 128K context length
vocabulary_size: int = 32064
class Phi35MoEExpert:
"""Individual expert using Phi-3.5-MoE model"""
def __init__(self, config: Phi35MoEConfig, expert_id: str, specialization: str):
self.config = config
self.expert_id = expert_id
self.specialization = specialization
self.model = None
self.tokenizer = None
self.pipeline = None
self.is_loaded = False
async def load_model(self):
"""Load the Phi-3.5-MoE model and tokenizer"""
try:
logger.info(f"Loading Phi-3.5-MoE model for expert {self.expert_id}")
# Load model
self.model = AutoModelForCausalLM.from_pretrained(
self.config.model_name,
device_map=self.config.device_map,
torch_dtype=self.config.torch_dtype,
trust_remote_code=self.config.trust_remote_code,
)
# Load tokenizer
self.tokenizer = AutoTokenizer.from_pretrained(self.config.model_name)
# Create pipeline
self.pipeline = pipeline(
"text-generation",
model=self.model,
tokenizer=self.tokenizer,
device_map=self.config.device_map,
)
self.is_loaded = True
logger.info(f"Successfully loaded Phi-3.5-MoE model for expert {self.expert_id}")
except Exception as e:
logger.error(f"Failed to load Phi-3.5-MoE model for expert {self.expert_id}: {e}")
raise
async def generate_response(
self,
messages: List[Dict[str, str]],
**generation_kwargs
) -> Dict[str, Any]:
"""Generate response using Phi-3.5-MoE model"""
if not self.is_loaded:
await self.load_model()
try:
# Set default generation parameters
generation_args = {
"max_new_tokens": self.config.max_new_tokens,
"temperature": self.config.temperature,
"top_p": self.config.top_p,
"do_sample": self.config.do_sample,
"return_full_text": False,
**generation_kwargs
}
# Generate response
response = self.pipeline(messages, **generation_args)
return {
"expert_id": self.expert_id,
"specialization": self.specialization,
"response": response[0]["generated_text"] if response else "",
"model": self.config.model_name,
"success": True
}
except Exception as e:
logger.error(f"Error generating response for expert {self.expert_id}: {e}")
return {
"expert_id": self.expert_id,
"specialization": self.specialization,
"response": "",
"error": str(e),
"success": False
}
class Phi35MoERouter:
"""Router for selecting appropriate Phi-3.5-MoE experts"""
def __init__(self):
self.expert_specializations = {
"code": ["programming", "software", "development", "coding", "algorithm", "python", "javascript", "java", "function", "code"],
"math": ["mathematics", "calculation", "equation", "formula", "statistics", "derivative", "integral", "algebra", "calculus", "math", "solve", "calculate"],
"reasoning": ["logic", "analysis", "reasoning", "problem-solving", "critical", "explain", "why", "how", "because"],
"multilingual": ["translation", "language", "multilingual", "localization", "translate", "spanish", "french", "german"],
"general": ["general", "conversation", "assistance", "help", "hello", "hi", "what", "who", "when", "where"]
}
def route_query(self, query: str, available_experts: List[Phi35MoEExpert]) -> List[Phi35MoEExpert]:
"""Route query to appropriate experts based on content analysis"""
query_lower = query.lower()
selected_experts = []
# Score each expert based on query content
expert_scores = {}
for expert in available_experts:
score = 0
for keyword in self.expert_specializations.get(expert.specialization, []):
if keyword in query_lower:
score += 1
expert_scores[expert.expert_id] = score
# Select experts with highest scores
sorted_experts = sorted(expert_scores.items(), key=lambda x: x[1], reverse=True)
# Always include at least one expert (preferably general)
if sorted_experts and sorted_experts[0][1] > 0:
# Select experts with score > 0
for expert_id, score in sorted_experts:
if score > 0:
expert = next((e for e in available_experts if e.expert_id == expert_id), None)
if expert:
selected_experts.append(expert)
else:
# Fallback to general expert
general_expert = next((e for e in available_experts if e.specialization == "general"), None)
if general_expert:
selected_experts.append(general_expert)
return selected_experts[:3] # Limit to top 3 experts
class EnhancedMoEFramework:
"""Enhanced MoE framework using Phi-3.5-MoE as baseline"""
def __init__(self, config: Optional[Phi35MoEConfig] = None):
self.config = config or Phi35MoEConfig()
self.experts: Dict[str, Phi35MoEExpert] = {}
self.router = Phi35MoERouter()
self.is_initialized = False
async def initialize_experts(self, expert_configs: List[Dict[str, str]]):
"""Initialize multiple Phi-3.5-MoE experts with different specializations"""
logger.info("Initializing Phi-3.5-MoE experts...")
for expert_config in expert_configs:
expert_id = expert_config["expert_id"]
specialization = expert_config["specialization"]
expert = Phi35MoEExpert(
config=self.config,
expert_id=expert_id,
specialization=specialization
)
self.experts[expert_id] = expert
self.is_initialized = True
logger.info(f"Initialized {len(self.experts)} Phi-3.5-MoE experts")
async def process_query(
self,
query: str,
system_message: Optional[str] = None,
use_multiple_experts: bool = True
) -> Dict[str, Any]:
"""Process query using the enhanced MoE framework"""
if not self.is_initialized:
raise RuntimeError("MoE framework not initialized. Call initialize_experts() first.")
# Prepare messages
messages = []
if system_message:
messages.append({"role": "system", "content": system_message})
messages.append({"role": "user", "content": query})
if use_multiple_experts:
# Route to multiple experts
selected_experts = self.router.route_query(query, list(self.experts.values()))
# Generate responses from selected experts
tasks = []
for expert in selected_experts:
task = expert.generate_response(messages)
tasks.append(task)
# Wait for all responses
responses = await asyncio.gather(*tasks, return_exceptions=True)
# Process responses
successful_responses = []
for response in responses:
if isinstance(response, dict) and response.get("success", False):
successful_responses.append(response)
# Combine responses
combined_response = self._combine_responses(successful_responses)
return {
"query": query,
"responses": successful_responses,
"combined_response": combined_response,
"num_experts_used": len(successful_responses),
"success": len(successful_responses) > 0
}
else:
# Use single expert (general)
general_expert = self.experts.get("general")
if not general_expert:
general_expert = list(self.experts.values())[0] # Fallback to first expert
response = await general_expert.generate_response(messages)
return {
"query": query,
"response": response,
"success": response.get("success", False)
}
def _combine_responses(self, responses: List[Dict[str, Any]]) -> str:
"""Combine multiple expert responses into a coherent answer"""
if not responses:
return "No responses available."
if len(responses) == 1:
return responses[0]["response"]
# Combine responses with expert attribution
combined = "Based on analysis from multiple experts:\n\n"
for i, response in enumerate(responses, 1):
expert_id = response.get("expert_id", f"Expert {i}")
specialization = response.get("specialization", "general")
expert_response = response.get("response", "")
combined += f"**{expert_id} ({specialization}):**\n{expert_response}\n\n"
return combined.strip()
async def benchmark_performance(self, test_queries: List[str]) -> Dict[str, Any]:
"""Benchmark the enhanced MoE framework performance"""
logger.info("Starting Phi-3.5-MoE framework benchmark...")
results = {
"total_queries": len(test_queries),
"successful_queries": 0,
"failed_queries": 0,
"average_response_time": 0,
"expert_usage_stats": {},
"detailed_results": []
}
total_time = 0
for i, query in enumerate(test_queries):
start_time = asyncio.get_event_loop().time()
try:
result = await self.process_query(query)
end_time = asyncio.get_event_loop().time()
response_time = end_time - start_time
total_time += response_time
if result.get("success", False):
results["successful_queries"] += 1
else:
results["failed_queries"] += 1
# Track expert usage
if "responses" in result:
for response in result["responses"]:
expert_id = response.get("expert_id", "unknown")
results["expert_usage_stats"][expert_id] = results["expert_usage_stats"].get(expert_id, 0) + 1
results["detailed_results"].append({
"query": query,
"success": result.get("success", False),
"response_time": response_time,
"num_experts_used": result.get("num_experts_used", 1)
})
except Exception as e:
logger.error(f"Error processing query {i}: {e}")
results["failed_queries"] += 1
results["detailed_results"].append({
"query": query,
"success": False,
"error": str(e),
"response_time": 0
})
results["average_response_time"] = total_time / len(test_queries) if test_queries else 0
results["success_rate"] = results["successful_queries"] / results["total_queries"] if test_queries else 0
logger.info(f"Benchmark completed. Success rate: {results['success_rate']:.2%}")
return results
# Example usage and testing
async def main():
"""Example usage of the enhanced MoE framework"""
# Initialize configuration
config = Phi35MoEConfig(
max_new_tokens=300,
temperature=0.7
)
# Initialize framework
moe_framework = EnhancedMoEFramework(config)
# Define expert configurations
expert_configs = [
{"expert_id": "code_expert", "specialization": "code"},
{"expert_id": "math_expert", "specialization": "math"},
{"expert_id": "reasoning_expert", "specialization": "reasoning"},
{"expert_id": "general_expert", "specialization": "general"}
]
# Initialize experts
await moe_framework.initialize_experts(expert_configs)
# Test queries
test_queries = [
"How do I implement a binary search algorithm in Python?",
"What is the derivative of x^2 + 3x + 1?",
"Explain the logical reasoning behind the Monty Hall problem",
"Hello, how are you today?"
]
# Process queries
for query in test_queries:
print(f"\n{'='*50}")
print(f"Query: {query}")
print(f"{'='*50}")
result = await moe_framework.process_query(query)
if result.get("success", False):
if "combined_response" in result:
print(f"Combined Response:\n{result['combined_response']}")
else:
print(f"Response:\n{result['response']['response']}")
else:
print("Failed to generate response")
# Run benchmark
print(f"\n{'='*50}")
print("Running Performance Benchmark...")
print(f"{'='*50}")
benchmark_results = await moe_framework.benchmark_performance(test_queries)
print(f"Success Rate: {benchmark_results['success_rate']:.2%}")
print(f"Average Response Time: {benchmark_results['average_response_time']:.2f}s")
print(f"Expert Usage Stats: {benchmark_results['expert_usage_stats']}")
if __name__ == "__main__":
asyncio.run(main())
|