File size: 15,289 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 |
#!/usr/bin/env python3
"""
Agent Token Management System
This module provides comprehensive agent token management for multi-agent training,
including special token handling, embedding management, and integration with
existing tokenization systems.
"""
import os
import json
import logging
from typing import Dict, List, Optional, Tuple, Any, Union
from dataclasses import dataclass
from pathlib import Path
import torch
from transformers import AutoTokenizer, PreTrainedTokenizer
from transformers.tokenization_utils_base import PreTrainedTokenizerBase
logger = logging.getLogger(__name__)
@dataclass
class AgentTokenConfig:
"""Configuration for agent token management"""
agent_prefix: str = "<|agent:"
agent_suffix: str = "|>"
special_tokens: Optional[Dict[str, str]] = None
resize_embeddings: bool = True
save_tokens: bool = True
tokens_file: str = "agent_tokens.json"
class AgentTokenManager:
"""
Manages agent-specific tokens and their integration with tokenizers
"""
def __init__(self, config: AgentTokenConfig):
self.config = config
self.agent_tokens: Dict[str, str] = {}
self.token_ids: Dict[str, int] = {}
self.original_vocab_size: Optional[int] = None
def generate_agent_tokens(self, agents: List[str]) -> List[str]:
"""Generate agent tokens for given agent list"""
tokens = []
for agent in agents:
token = f"{self.config.agent_prefix}{agent}{self.config.agent_suffix}"
tokens.append(token)
self.agent_tokens[agent] = token
logger.info(f"Generated {len(tokens)} agent tokens: {tokens}")
return tokens
def add_agent_tokens_to_tokenizer(self, tokenizer: PreTrainedTokenizer, agents: List[str]) -> Tuple[PreTrainedTokenizer, List[str]]:
"""
Add agent tokens to tokenizer and return updated tokenizer with token list
"""
if not agents:
logger.warning("No agents provided, skipping token addition")
return tokenizer, []
# Generate agent tokens
agent_tokens = self.generate_agent_tokens(agents)
# Check which tokens need to be added
existing_tokens = set(tokenizer.get_vocab().keys())
tokens_to_add = [token for token in agent_tokens if token not in existing_tokens]
if not tokens_to_add:
logger.info("All agent tokens already exist in tokenizer")
return tokenizer, agent_tokens
# Store original vocab size
self.original_vocab_size = len(tokenizer)
# Add special tokens
logger.info(f"Adding {len(tokens_to_add)} new agent tokens to tokenizer")
tokenizer.add_special_tokens({
"additional_special_tokens": tokens_to_add
})
# Update token IDs mapping
for agent, token in self.agent_tokens.items():
if token in tokenizer.get_vocab():
self.token_ids[agent] = tokenizer.convert_tokens_to_ids(token)
logger.info(f"Added agent tokens. New vocab size: {len(tokenizer)}")
return tokenizer, agent_tokens
def resize_model_embeddings(self, model: torch.nn.Module, tokenizer: PreTrainedTokenizer) -> torch.nn.Module:
"""
Resize model embeddings to accommodate new agent tokens
"""
if not self.config.resize_embeddings:
logger.info("Embedding resize disabled, skipping")
return model
if self.original_vocab_size is None:
logger.warning("Original vocab size not set, cannot resize embeddings")
return model
new_vocab_size = len(tokenizer)
if new_vocab_size == self.original_vocab_size:
logger.info("Vocab size unchanged, no embedding resize needed")
return model
logger.info(f"Resizing model embeddings from {self.original_vocab_size} to {new_vocab_size}")
# Resize embeddings
model.resize_token_embeddings(new_vocab_size)
# Initialize new embeddings (copy from unk token or use random initialization)
if hasattr(model, 'get_input_embeddings'):
embeddings = model.get_input_embeddings()
if hasattr(embeddings, 'weight'):
with torch.no_grad():
# Initialize new embeddings with small random values
new_embeddings = embeddings.weight[self.original_vocab_size:]
torch.nn.init.normal_(new_embeddings, mean=0.0, std=0.02)
logger.info("Model embeddings resized successfully")
return model
def format_agent_prompt(self, agent: str, text: str) -> str:
"""Format text with agent token prefix"""
if agent not in self.agent_tokens:
logger.warning(f"Agent '{agent}' not found in token mappings")
return text
agent_token = self.agent_tokens[agent]
return f"{agent_token}\n{text}"
def extract_agent_from_text(self, text: str) -> Optional[str]:
"""Extract agent name from text if it starts with agent token"""
for agent, token in self.agent_tokens.items():
if text.startswith(token):
return agent
return None
def get_agent_token_id(self, agent: str) -> Optional[int]:
"""Get token ID for agent token"""
return self.token_ids.get(agent)
def save_agent_tokens(self, output_dir: str) -> str:
"""Save agent tokens to file"""
if not self.config.save_tokens:
return ""
os.makedirs(output_dir, exist_ok=True)
tokens_file = os.path.join(output_dir, self.config.tokens_file)
tokens_data = {
"agent_tokens": self.agent_tokens,
"token_ids": self.token_ids,
"config": {
"agent_prefix": self.config.agent_prefix,
"agent_suffix": self.config.agent_suffix,
"original_vocab_size": self.original_vocab_size
}
}
with open(tokens_file, 'w') as f:
json.dump(tokens_data, f, indent=2)
logger.info(f"Saved agent tokens to {tokens_file}")
return tokens_file
def load_agent_tokens(self, tokens_file: str) -> bool:
"""Load agent tokens from file"""
if not os.path.isfile(tokens_file):
logger.warning(f"Agent tokens file not found: {tokens_file}")
return False
try:
with open(tokens_file, 'r') as f:
tokens_data = json.load(f)
self.agent_tokens = tokens_data.get("agent_tokens", {})
self.token_ids = tokens_data.get("token_ids", {})
config_data = tokens_data.get("config", {})
self.original_vocab_size = config_data.get("original_vocab_size")
logger.info(f"Loaded {len(self.agent_tokens)} agent tokens from {tokens_file}")
return True
except Exception as e:
logger.error(f"Failed to load agent tokens: {e}")
return False
def get_agent_statistics(self) -> Dict[str, Any]:
"""Get statistics about agent tokens"""
return {
"total_agents": len(self.agent_tokens),
"agents": list(self.agent_tokens.keys()),
"token_ids": self.token_ids,
"original_vocab_size": self.original_vocab_size,
"config": {
"agent_prefix": self.config.agent_prefix,
"agent_suffix": self.config.agent_suffix
}
}
class AgentTokenizer:
"""
Enhanced tokenizer wrapper that integrates agent token management
"""
def __init__(self, tokenizer: PreTrainedTokenizer, agent_manager: AgentTokenManager):
self.tokenizer = tokenizer
self.agent_manager = agent_manager
def tokenize_agent_text(self, agent: str, text: str, **kwargs) -> Dict[str, Any]:
"""Tokenize text with agent prefix"""
formatted_text = self.agent_manager.format_agent_prompt(agent, text)
return self.tokenizer(formatted_text, **kwargs)
def decode_agent_tokens(self, token_ids: Union[List[int], torch.Tensor], **kwargs) -> str:
"""Decode token IDs back to text"""
return self.tokenizer.decode(token_ids, **kwargs)
def get_agent_attention_mask(self, input_ids: torch.Tensor, agent: str) -> torch.Tensor:
"""Get attention mask with special handling for agent tokens"""
attention_mask = torch.ones_like(input_ids)
# Find agent token position
agent_token_id = self.agent_manager.get_agent_token_id(agent)
if agent_token_id is not None:
# Ensure agent token is attended to
agent_positions = (input_ids == agent_token_id)
attention_mask[agent_positions] = 1
return attention_mask
def __getattr__(self, name):
"""Delegate unknown attributes to underlying tokenizer"""
return getattr(self.tokenizer, name)
class AgentTokenValidator:
"""Validator for agent token configurations"""
@staticmethod
def validate_agent_tokens(agents: List[str], config: AgentTokenConfig) -> Dict[str, Any]:
"""Validate agent token configuration"""
validation_result = {
"valid": True,
"errors": [],
"warnings": [],
"tokens": {}
}
if not agents:
validation_result["warnings"].append("No agents provided")
return validation_result
# Check for duplicate agents
if len(agents) != len(set(agents)):
validation_result["errors"].append("Duplicate agents found")
validation_result["valid"] = False
# Generate and validate tokens
manager = AgentTokenManager(config)
tokens = manager.generate_agent_tokens(agents)
# Check for token conflicts
token_set = set(tokens)
if len(token_set) != len(tokens):
validation_result["errors"].append("Duplicate tokens generated")
validation_result["valid"] = False
# Check token length
for agent, token in zip(agents, tokens):
if len(token) > 50: # Reasonable limit
validation_result["warnings"].append(f"Long token for agent '{agent}': {token}")
validation_result["tokens"] = dict(zip(agents, tokens))
return validation_result
@staticmethod
def validate_tokenizer_compatibility(tokenizer: PreTrainedTokenizer, agents: List[str], config: AgentTokenConfig) -> Dict[str, Any]:
"""Validate tokenizer compatibility with agent tokens"""
validation_result = {
"compatible": True,
"errors": [],
"warnings": [],
"existing_tokens": [],
"new_tokens": []
}
if not agents:
return validation_result
# Generate tokens
manager = AgentTokenManager(config)
tokens = manager.generate_agent_tokens(agents)
# Check existing vocabulary
vocab = tokenizer.get_vocab()
for agent, token in zip(agents, tokens):
if token in vocab:
validation_result["existing_tokens"].append(agent)
else:
validation_result["new_tokens"].append(agent)
# Check for potential conflicts
for token in tokens:
if token in vocab:
# Check if it's already a special token
if hasattr(tokenizer, 'special_tokens_map'):
special_tokens = tokenizer.special_tokens_map
if token not in special_tokens.values():
validation_result["warnings"].append(f"Token '{token}' exists in vocab but not as special token")
return validation_result
# Integration with existing MoE framework
class MoEAgentTokenIntegration:
"""
Integration layer between agent tokens and MoE framework
"""
def __init__(self, agent_manager: AgentTokenManager):
self.agent_manager = agent_manager
self.agent_to_expert_mapping: Dict[str, str] = {}
def map_agent_to_expert(self, agent: str, expert: str):
"""Map agent to MoE expert specialization"""
self.agent_to_expert_mapping[agent] = expert
logger.info(f"Mapped agent '{agent}' to expert '{expert}'")
def get_expert_for_agent(self, agent: str) -> Optional[str]:
"""Get expert specialization for agent"""
return self.agent_to_expert_mapping.get(agent)
def format_moe_prompt(self, agent: str, text: str, expert: Optional[str] = None) -> str:
"""Format prompt for MoE framework with agent and expert context"""
# Start with agent token
formatted_text = self.agent_manager.format_agent_prompt(agent, text)
# Add expert context if available
if expert:
expert_context = f"\n<|expert:{expert}|>\n"
formatted_text = formatted_text.replace("\n", expert_context, 1)
return formatted_text
def extract_agent_and_expert(self, text: str) -> Tuple[Optional[str], Optional[str]]:
"""Extract both agent and expert from formatted text"""
agent = self.agent_manager.extract_agent_from_text(text)
# Extract expert if present
expert = None
if "<|expert:" in text and "|>" in text:
start = text.find("<|expert:") + 9
end = text.find("|>", start)
if end > start:
expert = text[start:end]
return agent, expert
# Example usage and testing
if __name__ == "__main__":
# Configure logging
logging.basicConfig(level=logging.INFO)
# Example configuration
config = AgentTokenConfig(
agent_prefix="<|agent:",
agent_suffix="|>",
resize_embeddings=True
)
# Example agents
agents = ["SWE", "SQE", "DevOps", "Architect", "Security"]
# Create agent manager
manager = AgentTokenManager(config)
# Generate tokens
tokens = manager.generate_agent_tokens(agents)
print(f"Generated tokens: {tokens}")
# Example tokenizer (would be loaded from actual model)
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("microsoft/Phi-3.5-MoE-instruct")
# Add tokens to tokenizer
updated_tokenizer, agent_tokens = manager.add_agent_tokens_to_tokenizer(tokenizer, agents)
print(f"Updated tokenizer vocab size: {len(updated_tokenizer)}")
print(f"Agent token IDs: {manager.token_ids}")
# Test formatting
test_text = "How do I implement a binary search?"
formatted = manager.format_agent_prompt("SWE", test_text)
print(f"Formatted prompt: {formatted}")
# Test extraction
extracted_agent = manager.extract_agent_from_text(formatted)
print(f"Extracted agent: {extracted_agent}")
|