File size: 20,325 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 |
from typing import Optional, Dict, List, Union, Tuple
import numpy as np
from enum import Enum
from dataclasses import dataclass
from .embedding import embedding_lookup, add_positional_encoding
from .positional_encoding import sinusoidal_positional_encoding
from .stack import transformer_stack
from .layer_norm import layer_norm
from .core.db_manager import HeliumDBManager
from .broadcast import ModalityType, TensorMetadata
class EncoderType(Enum):
"""Supported encoder architectures"""
TEXT = "text"
VISION = "vision"
AUDIO = "audio"
MULTIMODAL = "multimodal"
@dataclass
class ModalityConfig:
"""Configuration for specific modalities"""
modality_type: ModalityType
input_channels: int = 1
patch_size: Union[int, Tuple[int, ...]] = 16
sampling_rate: Optional[int] = None
frame_rate: Optional[int] = None
max_seq_len: int = 1024
use_positional: bool = True
use_patch_embed: bool = False
@dataclass
class EncoderConfig:
"""Configuration for TransformerEncoder"""
encoder_type: EncoderType
hidden_dim: int
num_layers: int
num_heads: int
modality_configs: Dict[ModalityType, ModalityConfig]
vocab_size: Optional[int] = None # Only needed for text
dropout_rate: float = 0.1
layer_norm_epsilon: float = 1e-5
initializer_range: float = 0.02
use_cache: bool = True
use_fp16: bool = False
fusion_type: str = "concatenate" # concatenate, add, or learnable
def get_total_sequence_length(self) -> int:
"""Get total sequence length across all modalities"""
return sum(config.max_seq_len for config in self.modality_configs.values())
class EncoderCache:
"""Cache for storing key/value states during inference"""
def __init__(self):
self.layer_states: List[Tuple[np.ndarray, np.ndarray]] = []
self.position_offset: int = 0
def update(self, layer_idx: int, key: np.ndarray, value: np.ndarray):
if layer_idx >= len(self.layer_states):
self.layer_states.append((key, value))
else:
prev_k, prev_v = self.layer_states[layer_idx]
self.layer_states[layer_idx] = (
np.concatenate([prev_k, key], axis=1),
np.concatenate([prev_v, value], axis=1)
)
class ModalityEncoder:
"""Base class for modality-specific encoders"""
def __init__(
self,
config: ModalityConfig,
hidden_dim: int,
driver=None
):
self.config = config
self.hidden_dim = hidden_dim
self.driver = driver
def encode(self, x: np.ndarray) -> Tuple[np.ndarray, TensorMetadata]:
"""Convert input to embeddings with metadata"""
raise NotImplementedError
class VisionEncoder(ModalityEncoder):
"""Vision-specific encoder with patching"""
def encode(self, x: np.ndarray) -> Tuple[np.ndarray, TensorMetadata]:
# Apply patch embedding
if self.config.use_patch_embed:
B, C, H, W = x.shape
P = self.config.patch_size
x = x.reshape(B, C, H//P, P, W//P, P).transpose(0,2,4,1,3,5)
x = x.reshape(B, (H//P)*(W//P), C*P*P)
# Project to hidden dimension
if hasattr(self.driver, 'linear'):
x = self.driver.linear(x, self.hidden_dim)
else:
x = np.random.randn(*x.shape[:-1], self.hidden_dim)
metadata = TensorMetadata(
modality=ModalityType.VISION,
shape=x.shape,
dtype=x.dtype,
channels=self.config.input_channels,
spatial_dims=(H, W) if 'H' in locals() else None
)
return x, metadata
class AudioEncoder(ModalityEncoder):
"""Audio-specific encoder"""
def encode(self, x: np.ndarray) -> Tuple[np.ndarray, TensorMetadata]:
# Apply time-frequency transform if needed
if hasattr(self.driver, 'stft'):
x = self.driver.stft(x)
metadata = TensorMetadata(
modality=ModalityType.AUDIO,
shape=x.shape,
dtype=x.dtype,
channels=self.config.input_channels,
sampling_rate=self.config.sampling_rate
)
return x, metadata
class TextEncoder(ModalityEncoder):
"""Text-specific encoder"""
def __init__(self, config: ModalityConfig, hidden_dim: int,
vocab_size: int, embedding_weights: np.ndarray,
driver=None):
super().__init__(config, hidden_dim, driver)
self.vocab_size = vocab_size
self.embedding_weights = embedding_weights
def encode(self, x: np.ndarray) -> Tuple[np.ndarray, TensorMetadata]:
x = embedding_lookup(x, self.embedding_weights, driver=self.driver)
metadata = TensorMetadata(
modality=ModalityType.TEXT,
shape=x.shape,
dtype=x.dtype,
sequence_length=x.shape[1]
)
return x, metadata
class TransformerEncoder:
"""
Multi-modal Transformer Encoder implementation with support for:
- Multiple input modalities (text, vision, audio)
- Cross-modal attention
- Modality-specific processing
- Inference caching
- Mixed precision (FP16/FP32)
- Parallel processing
- Memory optimization
"""
def __init__(
self,
config: EncoderConfig,
embedding_weights: Optional[np.ndarray] = None,
block_weights_list: List[Dict] = None,
driver=None,
scheduler=None
):
"""
Initialize the multi-modal transformer encoder.
Args:
config: Encoder configuration with modality settings
embedding_weights: Optional word embedding matrix for text
block_weights_list: List of weight dictionaries for transformer blocks
driver: Optional hardware driver for optimized computation
scheduler: Optional scheduler for parallel processing
"""
self.validate_inputs(config, embedding_weights, block_weights_list)
self.config = config
self.driver = driver
self.scheduler = scheduler
# Initialize modality-specific encoders
self.encoders = {}
for modality, modal_config in config.modality_configs.items():
if modality == ModalityType.TEXT:
if embedding_weights is None:
raise ValueError("embedding_weights required for text modality")
self.encoders[modality] = TextEncoder(
modal_config,
config.hidden_dim,
config.vocab_size,
self._prepare_weights(embedding_weights),
driver
)
elif modality == ModalityType.VISION:
self.encoders[modality] = VisionEncoder(
modal_config,
config.hidden_dim,
driver
)
elif modality == ModalityType.AUDIO:
self.encoders[modality] = AudioEncoder(
modal_config,
config.hidden_dim,
driver
)
# Initialize transformer blocks
self.block_weights_list = [
self._prepare_weights(weights) for weights in (block_weights_list or [])
]
# Initialize cached computations and fusion layer
self._init_cached_computations()
self._init_fusion_layer()
def _init_cached_computations(self):
"""Initialize cached components for faster inference"""
# Create positional encodings for each modality
self.pos_encodings = {}
dtype = np.float16 if self.config.use_fp16 else np.float32
for modality, modal_config in self.config.modality_configs.items():
if modal_config.use_positional:
self.pos_encodings[modality] = sinusoidal_positional_encoding(
modal_config.max_seq_len,
self.config.hidden_dim,
dtype=dtype
)
# Precompute attention bias if supported
if self.driver and hasattr(self.driver, 'precompute_attention_bias'):
total_seq_len = self.config.get_total_sequence_length()
self.cached_attention_bias = self.driver.precompute_attention_bias(
total_seq_len
)
else:
self.cached_attention_bias = None
def _init_fusion_layer(self):
"""Initialize multi-modal fusion layer"""
if self.config.fusion_type == "learnable":
num_modalities = len(self.config.modality_configs)
if self.driver and hasattr(self.driver, 'create_parameter'):
self.fusion_weights = self.driver.create_parameter(
(num_modalities, 1, 1),
dtype=np.float16 if self.config.use_fp16 else np.float32
)
else:
self.fusion_weights = np.ones((num_modalities, 1, 1)) / num_modalities
else:
self.fusion_weights = None
def _prepare_weights(self, weights: Union[np.ndarray, Dict]) -> Union[np.ndarray, Dict]:
"""Convert weights to appropriate precision"""
if self.config.use_fp16:
if isinstance(weights, np.ndarray):
return weights.astype(np.float16)
return {k: v.astype(np.float16) for k, v in weights.items()}
return weights
def _fuse_modalities(
self,
encoded_states: Dict[ModalityType, np.ndarray],
encoded_metadata: Dict[ModalityType, TensorMetadata]
) -> Tuple[np.ndarray, TensorMetadata]:
"""
Fuse multiple modalities into a single representation
Supports three fusion types:
1. concatenate: Concatenate along sequence dimension
2. add: Element-wise addition (requires same shape)
3. learnable: Weighted sum using learned weights
"""
modalities = list(encoded_states.keys())
if len(modalities) == 1:
return encoded_states[modalities[0]], encoded_metadata[modalities[0]]
if self.config.fusion_type == "concatenate":
# Concatenate along sequence dimension
fused = np.concatenate(
[encoded_states[m] for m in modalities],
axis=1
)
elif self.config.fusion_type == "add":
# Verify shapes match
shapes = [encoded_states[m].shape for m in modalities]
if not all(s == shapes[0] for s in shapes):
raise ValueError(
f"All modalities must have same shape for addition fusion. Got {shapes}"
)
fused = sum(encoded_states[m] for m in modalities)
elif self.config.fusion_type == "learnable":
# Apply learned weights
weighted = [
encoded_states[m] * self.fusion_weights[i]
for i, m in enumerate(modalities)
]
fused = sum(weighted)
else:
raise ValueError(f"Unknown fusion type: {self.config.fusion_type}")
# Create metadata for fused representation
fused_metadata = TensorMetadata(
modality=ModalityType.LATENT,
shape=fused.shape,
dtype=fused.dtype,
channels=sum(m.channels for m in encoded_metadata.values()),
sequence_length=fused.shape[1]
)
return fused, fused_metadata
@staticmethod
def validate_inputs(
config: EncoderConfig,
embedding_weights: np.ndarray,
block_weights_list: List[Dict]
):
"""Validate input parameters and weights"""
if embedding_weights.shape != (config.vocab_size, config.hidden_dim):
raise ValueError(
f"Embedding weights shape {embedding_weights.shape} doesn't match "
f"config (vocab_size={config.vocab_size}, hidden_dim={config.hidden_dim})"
)
if len(block_weights_list) != config.num_layers:
raise ValueError(
f"Expected {config.num_layers} transformer blocks, got {len(block_weights_list)}"
)
def create_attention_mask(
self,
input_shape: Tuple[int, int],
past_length: int = 0
) -> np.ndarray:
"""Create causal attention mask for autoregressive inference"""
batch_size, seq_length = input_shape
mask = np.ones((batch_size, 1, seq_length, seq_length + past_length))
# Create causal mask for autoregressive generation
if past_length > 0:
mask[:, :, :, :past_length] = 1.0
return mask
def forward(
self,
inputs: Dict[ModalityType, np.ndarray],
attention_mask: Optional[np.ndarray] = None,
past_cache: Optional[EncoderCache] = None,
return_cache: bool = False
) -> Union[np.ndarray, Tuple[np.ndarray, EncoderCache]]:
"""
Forward pass of the multi-modal encoder
Args:
inputs: Dictionary mapping modality types to input arrays
attention_mask: Optional attention mask
past_cache: Optional cached key/value states
return_cache: Whether to return updated cache
Returns:
Encoded representations, optionally with cache
"""
# Encode each modality
encoded_states = {}
encoded_metadata = {}
max_seq_len = 0
for modality, x in inputs.items():
if modality not in self.encoders:
raise ValueError(f"No encoder configured for modality {modality}")
# Encode input
states, metadata = self.encoders[modality].encode(x)
encoded_states[modality] = states
encoded_metadata[modality] = metadata
max_seq_len = max(max_seq_len, states.shape[1])
# Pad sequences to same length
for modality in encoded_states:
states = encoded_states[modality]
if states.shape[1] < max_seq_len:
pad_len = max_seq_len - states.shape[1]
encoded_states[modality] = np.pad(
states,
((0, 0), (0, pad_len), (0, 0)),
mode='constant'
)
# Add positional encodings
for modality, states in encoded_states.items():
if modality in self.pos_encodings:
pos_enc = self.pos_encodings[modality][:states.shape[1]]
encoded_states[modality] = states + pos_enc
# Create attention mask if not provided
if attention_mask is None:
attention_mask = self.create_attention_mask(
(encoded_states[list(encoded_states.keys())[0]].shape[0], max_seq_len),
past_length=past_cache.position_offset if past_cache else 0
)
"""
Forward pass through the transformer encoder.
Args:
input_ids: Input token IDs of shape (batch_size, seq_len)
attention_mask: Optional attention mask
past_cache: Optional past key/value cache for inference
return_cache: Whether to return updated cache
Returns:
output: Encoded representations
cache: Updated cache if return_cache is True
"""
batch_size, seq_length = input_ids.shape
if seq_length > self.config.max_seq_len:
raise ValueError(
f"Input sequence length {seq_length} exceeds maximum "
f"sequence length {self.config.max_seq_len}"
)
# Fuse modalities
hidden_states, fused_metadata = self._fuse_modalities(
encoded_states,
encoded_metadata
)
# Initialize cache for current forward pass
current_cache = EncoderCache() if self.config.use_cache else None
if current_cache:
current_cache.modality_metadata = fused_metadata
# Process through transformer stack with modality-aware attention
hidden_states = transformer_stack(
hidden_states,
self.block_weights_list,
self.config.num_heads,
attention_mask=attention_mask,
past_cache=past_cache,
current_cache=current_cache,
driver=self.driver,
scheduler=self.scheduler,
metadata=fused_metadata
)
if return_cache:
return hidden_states, current_cache
return hidden_states
def generate(
self,
input_ids: np.ndarray,
max_length: int,
temperature: float = 1.0,
top_k: int = 50,
top_p: float = 0.95
) -> np.ndarray:
"""
Generate sequences autoregressively.
Args:
input_ids: Initial input tokens
max_length: Maximum sequence length to generate
temperature: Sampling temperature
top_k: Number of top tokens to sample from
top_p: Cumulative probability threshold for nucleus sampling
Returns:
generated_ids: Generated token sequences
"""
batch_size = input_ids.shape[0]
generated_ids = [list(seq) for seq in input_ids]
cache = EncoderCache()
for _ in range(max_length - input_ids.shape[1]):
# Forward pass with caching
outputs, cache = self.forward(
input_ids,
past_cache=cache,
return_cache=True
)
# Get next token logits
next_token_logits = outputs[:, -1, :]
# Apply temperature
next_token_logits = next_token_logits / temperature
# Apply top-k filtering
if top_k > 0:
indices_to_remove = next_token_logits < np.partition(
next_token_logits, -top_k, axis=-1
)[:, -top_k:].min(axis=-1, keepdims=True)
next_token_logits[indices_to_remove] = -float('inf')
# Apply top-p (nucleus) filtering
if top_p < 1.0:
sorted_logits = np.sort(next_token_logits, axis=-1)[:, ::-1]
cumsum_probs = np.cumsum(np.exp(sorted_logits), axis=-1)
mask = cumsum_probs > top_p
mask[:, 1:] = mask[:, :-1].copy()
mask[:, 0] = 0
indices_to_remove = next_token_logits < np.min(
sorted_logits[mask],
axis=-1,
keepdims=True
)
next_token_logits[indices_to_remove] = -float('inf')
# Sample next tokens
probs = np.exp(next_token_logits)
probs = probs / np.sum(probs, axis=-1, keepdims=True)
next_tokens = np.array([
np.random.choice(self.config.vocab_size, p=p)
for p in probs
])
# Update generated sequences
for i in range(batch_size):
generated_ids[i].append(next_tokens[i])
# Update input_ids for next iteration
input_ids = next_tokens[:, np.newaxis]
return np.array(generated_ids)
|