File size: 11,515 Bytes
7a0c684 |
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 |
from typing import Optional, List, Dict, Union, Tuple
import numpy as np
from dataclasses import dataclass
from enum import Enum
import warnings
from .block import TransformerBlock
from .core.db_manager import HeliumDBManager
import json
import hashlib
from contextlib import contextmanager
import time
class ExecutionStrategy(Enum):
"""Execution strategies for transformer stack"""
SEQUENTIAL = "sequential" # Process blocks one by one
PIPELINED = "pipelined" # Pipeline blocks across multiple devices
PARALLEL = "parallel" # Process blocks in parallel where possible
@dataclass
class StackConfig:
"""Configuration for transformer stack"""
num_layers: int
hidden_dim: int
num_heads: int
intermediate_size: int
max_sequence_length: int
dropout_rate: float = 0.1
layer_norm_epsilon: float = 1e-5
use_cache: bool = True
use_checkpointing: bool = False
execution_strategy: ExecutionStrategy = ExecutionStrategy.SEQUENTIAL
dtype: np.dtype = np.float32
gradient_checkpointing_steps: int = 2
max_batch_size: Optional[int] = None
class TransformerStackCache:
"""Cache manager for transformer stack computations"""
def __init__(self, config: StackConfig):
self.config = config
self.db = HeliumDBManager.get_instance()
def _compute_cache_key(
self,
layer_idx: int,
input_shape: Tuple,
block_config: Dict
) -> str:
"""Compute cache key for layer outputs"""
cache_data = {
'layer_idx': layer_idx,
'input_shape': input_shape,
'block_config': block_config,
'dtype': str(self.config.dtype)
}
return hashlib.sha256(json.dumps(cache_data).encode()).hexdigest()
def get(self, key: str) -> Optional[np.ndarray]:
"""Get cached computation result"""
return self.db.get_activation(key)
def set(self, key: str, value: np.ndarray, metadata: Dict):
"""Cache computation result"""
self.db.set_activation(key, value, metadata)
class ResourceManager:
"""Manages hardware resources and scheduling"""
def __init__(self, driver=None):
self.driver = driver
self.available_devices = self._get_available_devices()
self.device_queues = {device: [] for device in self.available_devices}
def _get_available_devices(self) -> List[str]:
"""Get list of available compute devices"""
if self.driver and hasattr(self.driver, 'list_devices'):
return self.driver.list_devices()
return ['cpu']
@contextmanager
def acquire_device(self, preferred_device: Optional[str] = None):
"""Acquire a compute device"""
device = self._select_device(preferred_device)
try:
yield device
finally:
self._release_device(device)
def _select_device(self, preferred_device: Optional[str] = None) -> str:
"""Select best available device"""
if preferred_device and preferred_device in self.available_devices:
return preferred_device
# Select device with shortest queue
return min(
self.device_queues.items(),
key=lambda x: len(x[1])
)[0]
def _release_device(self, device: str):
"""Release device back to pool"""
if device in self.device_queues:
self.device_queues[device].pop(0) if self.device_queues[device] else None
class TransformerStack:
"""
Optimized transformer stack implementation with support for:
- Multiple execution strategies
- Hardware acceleration
- Gradient checkpointing
- Mixed precision
- Memory optimization
"""
def __init__(
self,
config: StackConfig,
weights_list: List[Dict],
driver = None
):
"""
Initialize transformer stack
Args:
config: Stack configuration
weights_list: List of block weights
driver: Optional hardware driver
"""
self.config = config
self.weights_list = weights_list
self.driver = driver
self._validate_config()
self._setup_components()
def _validate_config(self):
"""Validate configuration parameters"""
if len(self.weights_list) != self.config.num_layers:
raise ValueError(
f"Expected {self.config.num_layers} weight dicts, got {len(self.weights_list)}"
)
if self.config.num_heads <= 0:
raise ValueError(f"Invalid number of heads: {self.config.num_heads}")
if self.config.hidden_dim % self.config.num_heads != 0:
raise ValueError(
f"Hidden dimension {self.config.hidden_dim} must be divisible "
f"by number of heads {self.config.num_heads}"
)
def _setup_components(self):
"""Setup stack components"""
# Initialize blocks
self.blocks = [
TransformerBlock(
hidden_size=self.config.hidden_dim,
num_heads=self.config.num_heads,
intermediate_size=self.config.intermediate_size,
weights=weights,
dropout_rate=self.config.dropout_rate,
layer_norm_epsilon=self.config.layer_norm_epsilon,
dtype=self.config.dtype,
driver=self.driver
)
for weights in self.weights_list
]
# Initialize cache
self.cache = TransformerStackCache(self.config)
# Initialize resource manager
self.resource_manager = ResourceManager(self.driver)
def _execute_sequential(
self,
x: np.ndarray,
mask: Optional[np.ndarray] = None,
use_cache: bool = True
) -> np.ndarray:
"""Execute blocks sequentially"""
current_state = x
for i, block in enumerate(self.blocks):
if use_cache:
cache_key = self.cache._compute_cache_key(
i, current_state.shape, block.get_config()
)
cached_result = self.cache.get(cache_key)
if cached_result is not None:
current_state = cached_result
continue
with self.resource_manager.acquire_device() as device:
current_state = block(
current_state,
mask=mask,
device=device
)
if use_cache:
self.cache.set(
cache_key,
current_state,
{'layer_idx': i, 'shape': current_state.shape}
)
return current_state
def _execute_pipelined(
self,
x: np.ndarray,
mask: Optional[np.ndarray] = None
) -> np.ndarray:
"""Execute blocks in a pipelined fashion"""
batch_size = x.shape[0]
num_chunks = min(
batch_size,
len(self.resource_manager.available_devices)
)
chunk_size = batch_size // num_chunks
# Split input into chunks
chunks = np.array_split(x, num_chunks)
results = []
# Process chunks in pipeline
for i, chunk in enumerate(chunks):
current_state = chunk
for j, block in enumerate(self.blocks):
with self.resource_manager.acquire_device() as device:
current_state = block(
current_state,
mask=mask[i*chunk_size:(i+1)*chunk_size] if mask is not None else None,
device=device
)
results.append(current_state)
# Concatenate results
return np.concatenate(results, axis=0)
def _execute_parallel(
self,
x: np.ndarray,
mask: Optional[np.ndarray] = None
) -> np.ndarray:
"""Execute blocks in parallel where possible"""
if not self.driver or not hasattr(self.driver, 'parallel_execute'):
warnings.warn("Parallel execution not supported, falling back to sequential")
return self._execute_sequential(x, mask)
return self.driver.parallel_execute(
self.blocks,
x,
mask,
self.config.num_layers
)
def forward(
self,
x: np.ndarray,
mask: Optional[np.ndarray] = None,
use_cache: bool = True
) -> np.ndarray:
"""
Forward pass through transformer stack
Args:
x: Input tensor of shape (batch_size, seq_len, hidden_dim)
mask: Optional attention mask
use_cache: Whether to use computation caching
Returns:
Output tensor of shape (batch_size, seq_len, hidden_dim)
"""
# Input validation
if x.ndim != 3:
raise ValueError(f"Expected 3D input tensor, got shape {x.shape}")
if x.shape[2] != self.config.hidden_dim:
raise ValueError(
f"Expected hidden dimension {self.config.hidden_dim}, got {x.shape[2]}"
)
if (
self.config.max_sequence_length and
x.shape[1] > self.config.max_sequence_length
):
raise ValueError(
f"Input sequence length {x.shape[1]} exceeds maximum "
f"allowed length {self.config.max_sequence_length}"
)
# Choose execution strategy
if self.config.execution_strategy == ExecutionStrategy.PIPELINED:
return self._execute_pipelined(x, mask)
elif self.config.execution_strategy == ExecutionStrategy.PARALLEL:
return self._execute_parallel(x, mask)
else:
return self._execute_sequential(x, mask, use_cache)
def __call__(
self,
x: np.ndarray,
mask: Optional[np.ndarray] = None,
use_cache: bool = True
) -> np.ndarray:
"""Callable interface"""
return self.forward(x, mask, use_cache)
# Legacy function for backward compatibility
def transformer_stack(
x: np.ndarray,
weights_list: List[Dict],
num_heads: int,
mask: Optional[np.ndarray] = None,
driver = None,
scheduler = None
) -> np.ndarray:
"""Legacy transformer stack interface"""
warnings.warn(
"transformer_stack function is deprecated, use TransformerStack class instead",
DeprecationWarning
)
config = StackConfig(
num_layers=len(weights_list),
hidden_dim=x.shape[2],
num_heads=num_heads,
intermediate_size=4 * x.shape[2], # Standard size
max_sequence_length=x.shape[1]
)
stack = TransformerStack(config, weights_list, driver)
return stack.forward(x, mask)
|