Spaces:
Configuration error
Configuration error
File size: 18,957 Bytes
77bcbf1 | 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 | """
Cascade Core - Symbiotic Adapter.
The heart of Cascade's system-agnostic design. The adapter uses Kleene fixed-point
convergence to interpret ANY signal format and convert it to Events.
"It doesn't hook into your system — it becomes part of it."
"""
import time
import json
import re
from typing import Any, Dict, List, Optional, Callable, Type
from dataclasses import dataclass
from cascade.core.event import Event
@dataclass
class SignalPattern:
"""A learned pattern for interpreting signals."""
pattern_type: str # 'dict', 'string', 'tensor', 'protobuf', etc.
component: str
event_type: str
extractor: Optional[Callable[[Any], Dict[str, Any]]] = None
confidence: float = 0.0
match_count: int = 0
class SymbioticAdapter:
"""
Self-interpreting adapter that converges to any signal format.
The adapter observes signals from the host system and learns how to
interpret them through fixed-point iteration. It starts with naive
interpretations and refines them until stable.
This is the key to Cascade's system-agnostic design:
- No framework-specific hooks required
- No configuration needed
- Feed it ANY signal format, it adapts
Example:
>>> adapter = SymbioticAdapter()
>>>
>>> # Feed it different signal formats
>>> adapter.interpret({"loss": 0.5, "epoch": 10})
>>> adapter.interpret("2024-01-01 12:00:00 ERROR training failed")
>>> adapter.interpret(torch.tensor([0.1, 0.2, 0.3]))
>>>
>>> # It learns patterns and gets better at interpretation
>>> print(adapter.learned_patterns)
"""
def __init__(self):
"""Initialize the symbiotic adapter."""
self._patterns: List[SignalPattern] = []
self._signal_count = 0
self._interpretation_cache: Dict[str, SignalPattern] = {}
# Built-in interpreters for common formats
self._builtin_interpreters = {
dict: self._interpret_dict,
str: self._interpret_string,
list: self._interpret_list,
}
# Regex patterns for log line parsing
self._log_patterns = [
# ISO timestamp with level: "2024-01-01 12:00:00 ERROR message"
re.compile(r'^(\d{4}-\d{2}-\d{2}[T\s]\d{2}:\d{2}:\d{2}(?:\.\d+)?)\s+(\w+)\s+(.*)$'),
# Simple timestamp: "12:00:00.123 component message"
re.compile(r'^(\d{2}:\d{2}:\d{2}(?:\.\d+)?)\s+(\w+)\s+(.*)$'),
# Pipe-delimited: "timestamp|level|component|key:value"
re.compile(r'^([^|]+)\|(\w+)\|(\w+)\|(.*)$'),
]
# Metric extraction patterns - ONLY extract real training metrics
# Be strict to avoid extracting garbage from config lines
self._metric_patterns = [
# Standard training metrics with = or :
re.compile(r'\b(loss|val_loss|train_loss|accuracy|acc|val_acc|lr|learning_rate|epoch|step|iter|iteration|mfu|tokens_per_sec|samples_per_sec|grad_norm|perplexity|ppl)[=:]\s*([+-]?\d+\.?\d*(?:e[+-]?\d+)?)', re.I),
# "iter X: loss=Y" format from nanoGPT
re.compile(r'iter\s+(\d+).*loss[=:]?\s*([+-]?\d+\.?\d*)', re.I),
# "step X loss Y" format
re.compile(r'step\s+(\d+).*loss\s*[=:]?\s*([+-]?\d+\.?\d*)', re.I),
]
def interpret(self, signal: Any) -> Event:
"""
Interpret any signal into a Cascade Event.
Uses Kleene fixed-point iteration to converge on the best interpretation.
Args:
signal: Any signal from the host system
Returns:
Event: The interpreted event
"""
self._signal_count += 1
# Get signal type
signal_type = type(signal)
# Try cached pattern first
cache_key = self._get_cache_key(signal)
if cache_key in self._interpretation_cache:
pattern = self._interpretation_cache[cache_key]
pattern.match_count += 1
return self._apply_pattern(signal, pattern)
# Try built-in interpreter
if signal_type in self._builtin_interpreters:
event = self._builtin_interpreters[signal_type](signal)
self._learn_pattern(signal, event)
return event
# Try tensor-like objects (duck typing)
if hasattr(signal, 'numpy') or hasattr(signal, 'detach'):
event = self._interpret_tensor(signal)
self._learn_pattern(signal, event)
return event
# Try protobuf-like objects
if hasattr(signal, 'SerializeToString'):
event = self._interpret_protobuf(signal)
self._learn_pattern(signal, event)
return event
# Fallback: convert to string and interpret
event = self._interpret_string(str(signal))
return event
def _interpret_dict(self, signal: Dict[str, Any]) -> Event:
"""Interpret a dictionary signal."""
# Extract common fields
timestamp = signal.get('timestamp', signal.get('time', time.time()))
if isinstance(timestamp, str):
try:
from datetime import datetime
timestamp = datetime.fromisoformat(timestamp).timestamp()
except:
timestamp = time.time()
component = signal.get('component', signal.get('source', 'unknown'))
event_type = signal.get('event_type', signal.get('type', 'state_change'))
# Everything else goes in data
reserved = {'timestamp', 'time', 'component', 'source', 'event_type', 'type'}
data = {k: v for k, v in signal.items() if k not in reserved}
return Event(
timestamp=timestamp,
component=component,
event_type=event_type,
data=data,
source_signal=signal,
)
def _interpret_string(self, signal: str) -> Event:
"""Interpret a string signal (log line, message, etc.)."""
signal = signal.strip()
# Try each log pattern
for pattern in self._log_patterns:
match = pattern.match(signal)
if match:
groups = match.groups()
if len(groups) >= 3:
timestamp_str, level_or_component, rest = groups[0], groups[1], groups[-1]
# Parse timestamp
try:
from datetime import datetime
timestamp = datetime.fromisoformat(timestamp_str.replace(' ', 'T')).timestamp()
except:
timestamp = time.time()
# Extract metrics from the rest
data = self._extract_metrics(rest)
data['raw_message'] = rest
# Determine event type from keywords
event_type = self._infer_event_type(signal)
return Event(
timestamp=timestamp,
component=level_or_component.lower(),
event_type=event_type,
data=data,
source_signal=signal,
)
# Fallback: extract what we can with smarter component detection
data = self._extract_metrics(signal)
data['raw_message'] = signal
# Infer component from content
component = self._infer_component(signal)
return Event(
timestamp=time.time(),
component=component,
event_type=self._infer_event_type(signal),
data=data,
source_signal=signal,
)
def _interpret_list(self, signal: List[Any]) -> Event:
"""Interpret a list signal."""
# Convert to dict with indices
data = {f'item_{i}': v for i, v in enumerate(signal)}
data['length'] = len(signal)
# Check if it looks like numeric data
if all(isinstance(x, (int, float)) for x in signal):
data['mean'] = sum(signal) / len(signal) if signal else 0
data['min'] = min(signal) if signal else 0
data['max'] = max(signal) if signal else 0
return Event(
timestamp=time.time(),
component='data',
event_type='list_signal',
data=data,
source_signal=signal,
)
def _interpret_tensor(self, signal: Any) -> Event:
"""Interpret a tensor-like signal (PyTorch, NumPy, etc.)."""
# Try to get numpy array
try:
if hasattr(signal, 'detach'):
arr = signal.detach().cpu().numpy()
elif hasattr(signal, 'numpy'):
arr = signal.numpy()
else:
arr = signal
data = {
'shape': list(arr.shape) if hasattr(arr, 'shape') else [],
'dtype': str(arr.dtype) if hasattr(arr, 'dtype') else 'unknown',
'mean': float(arr.mean()) if hasattr(arr, 'mean') else 0,
'std': float(arr.std()) if hasattr(arr, 'std') else 0,
'min': float(arr.min()) if hasattr(arr, 'min') else 0,
'max': float(arr.max()) if hasattr(arr, 'max') else 0,
}
# Check for NaN/Inf (common in gradient explosions)
if hasattr(arr, 'isnan'):
data['has_nan'] = bool(arr.isnan().any())
if hasattr(arr, 'isinf'):
data['has_inf'] = bool(arr.isinf().any())
except Exception as e:
data = {'error': str(e), 'type': str(type(signal))}
return Event(
timestamp=time.time(),
component='tensor',
event_type='tensor_signal',
data=data,
source_signal=None, # Don't store tensor to save memory
)
def _interpret_protobuf(self, signal: Any) -> Event:
"""Interpret a protobuf-like signal."""
try:
# Try to convert to dict
if hasattr(signal, 'DESCRIPTOR'):
from google.protobuf.json_format import MessageToDict
data = MessageToDict(signal)
else:
data = {'raw': str(signal)}
except:
data = {'raw': str(signal)}
return Event(
timestamp=time.time(),
component='protobuf',
event_type='protobuf_signal',
data=data,
source_signal=None,
)
def _extract_metrics(self, text: str) -> Dict[str, Any]:
"""Extract numeric metrics from text - STRICT, only real training metrics."""
metrics = {}
# nanoGPT format: "iter 0: loss=4.2176, time 46.76ms, mfu 0.62%"
nano_match = re.search(r'iter\s+(\d+).*loss[=:]?\s*([\d.]+)', text, re.I)
if nano_match:
metrics['iter'] = int(nano_match.group(1))
metrics['loss'] = float(nano_match.group(2))
# Diffusers/tqdm format: "step_loss=0.1234" or "step_loss: 0.1234"
step_loss_match = re.search(r'step_loss[=:]\s*([\d.e+-]+)', text, re.I)
if step_loss_match:
metrics['loss'] = float(step_loss_match.group(1))
# train_loss format from accelerator.log
train_loss_match = re.search(r'train_loss[=:]\s*([\d.e+-]+)', text, re.I)
if train_loss_match:
metrics['loss'] = float(train_loss_match.group(1))
# tqdm progress format: " 5%|█ | 5/100 [00:30<09:30, step_loss=0.234, lr=1e-5]"
tqdm_match = re.search(r'(\d+)%\|.*\|\s*(\d+)/(\d+)', text)
if tqdm_match:
metrics['progress_pct'] = int(tqdm_match.group(1))
metrics['step'] = int(tqdm_match.group(2))
metrics['total_steps'] = int(tqdm_match.group(3))
# Generic loss patterns
generic_loss = re.search(r'\bloss[=:]\s*([\d.e+-]+)', text, re.I)
if generic_loss and 'loss' not in metrics:
metrics['loss'] = float(generic_loss.group(1))
# mfu extraction
mfu_match = re.search(r'mfu\s*[=:]?\s*([\d.]+)%?', text, re.I)
if mfu_match:
metrics['mfu'] = float(mfu_match.group(1))
# time extraction (ms)
time_match = re.search(r'time\s*[=:]?\s*([\d.]+)\s*ms', text, re.I)
if time_match:
metrics['time_ms'] = float(time_match.group(1))
# learning rate - multiple formats
lr_match = re.search(r'\b(?:lr|learning_rate)\s*[=:]\s*([\d.e+-]+)', text, re.I)
if lr_match:
metrics['lr'] = float(lr_match.group(1))
# epoch/step for other frameworks
epoch_match = re.search(r'\bepoch\s*[=:]\s*(\d+)', text, re.I)
if epoch_match:
metrics['epoch'] = int(epoch_match.group(1))
step_match = re.search(r'\bstep\s*[=:]\s*(\d+)', text, re.I)
if step_match and 'step' not in metrics:
metrics['step'] = int(step_match.group(1))
# global_step from diffusers
global_step_match = re.search(r'global_step[=:]\s*(\d+)', text, re.I)
if global_step_match:
metrics['step'] = int(global_step_match.group(1))
return metrics
def _infer_event_type(self, text: str) -> str:
"""Infer event type from text content."""
text_lower = text.lower()
# Training iteration logs (highest priority)
if re.search(r'iter\s+\d+.*loss', text_lower):
return 'training_step'
if re.search(r'step\s+\d+.*loss', text_lower):
return 'training_step'
if any(kw in text_lower for kw in ['error', 'exception', 'failed', 'crash']):
return 'error'
if any(kw in text_lower for kw in ['warning', 'warn']):
return 'warning'
if any(kw in text_lower for kw in ['gradient', 'backward']):
return 'training'
if 'loss' in text_lower and 'val' in text_lower:
return 'validation'
if any(kw in text_lower for kw in ['inference', 'predict', 'forward']):
return 'inference'
if any(kw in text_lower for kw in ['epoch', 'step', 'iteration', 'iter']):
return 'progress'
if any(kw in text_lower for kw in ['nan', 'inf', 'explode', 'overflow']):
return 'anomaly'
if any(kw in text_lower for kw in ['save', 'checkpoint', 'load', 'saving']):
return 'checkpoint'
if any(kw in text_lower for kw in ['config', 'setting', 'parameter', 'device', 'gpu', 'cuda']):
return 'config'
if any(kw in text_lower for kw in ['initializ', 'loading model', 'compiling']):
return 'init'
return 'state_change'
def _infer_component(self, text: str) -> str:
"""Infer component from text content - NO MORE 'unknown'."""
text_lower = text.lower()
# Training/optimizer related
if any(kw in text_lower for kw in ['iter', 'step', 'epoch', 'batch']):
return 'trainer'
if any(kw in text_lower for kw in ['loss', 'backward', 'gradient']):
return 'loss'
if any(kw in text_lower for kw in ['optim', 'adam', 'sgd', 'lr', 'learning']):
return 'optimizer'
if any(kw in text_lower for kw in ['model', 'layer', 'param', 'weight']):
return 'model'
if any(kw in text_lower for kw in ['data', 'batch', 'loader', 'dataset']):
return 'data'
if any(kw in text_lower for kw in ['cuda', 'gpu', 'device', 'memory']):
return 'device'
if any(kw in text_lower for kw in ['checkpoint', 'save', 'load']):
return 'checkpoint'
if any(kw in text_lower for kw in ['config', 'setting', 'override']):
return 'config'
if any(kw in text_lower for kw in ['eval', 'valid', 'test']):
return 'evaluator'
if any(kw in text_lower for kw in ['token', 'vocab', 'embed']):
return 'tokenizer'
return 'system' # Generic fallback, not "unknown"
def _get_cache_key(self, signal: Any) -> str:
"""Generate a cache key for a signal's structure."""
if isinstance(signal, dict):
# Key based on dict keys
return f"dict:{':'.join(sorted(signal.keys()))}"
elif isinstance(signal, str):
# Key based on first word
first_word = signal.split()[0] if signal.split() else ''
return f"str:{first_word[:20]}"
else:
return f"type:{type(signal).__name__}"
def _learn_pattern(self, signal: Any, event: Event) -> None:
"""Learn a pattern from a successful interpretation."""
cache_key = self._get_cache_key(signal)
pattern = SignalPattern(
pattern_type=type(signal).__name__,
component=event.component,
event_type=event.event_type,
confidence=0.5,
match_count=1,
)
self._interpretation_cache[cache_key] = pattern
self._patterns.append(pattern)
def _apply_pattern(self, signal: Any, pattern: SignalPattern) -> Event:
"""Apply a learned pattern to interpret a signal."""
# Re-interpret with learned hints - use direct interpreters to avoid recursion
if isinstance(signal, dict):
event = self._interpret_dict(signal)
# Apply learned component/type if more confident
if pattern.confidence > 0.7:
return Event(
timestamp=event.timestamp,
component=pattern.component,
event_type=pattern.event_type,
data=event.data,
source_signal=signal,
)
return event
elif isinstance(signal, str):
return self._interpret_string(signal)
elif isinstance(signal, list):
return self._interpret_list(signal)
else:
# Fallback: interpret as string without recursion
return self._interpret_string(str(signal))
@property
def learned_patterns(self) -> List[SignalPattern]:
"""Get all learned signal patterns."""
return sorted(self._patterns, key=lambda p: p.match_count, reverse=True)
@property
def signal_count(self) -> int:
"""Total number of signals interpreted."""
return self._signal_count
def __repr__(self) -> str:
return f"<SymbioticAdapter | {self._signal_count} signals, {len(self._patterns)} patterns>"
|