File size: 13,700 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 |
"""
Hardware-accelerated multi-modal transformer decoder implementation for Helium virtual GPU
"""
from typing import Optional, Union, Dict, Any, TYPE_CHECKING, List, Tuple
from dataclasses import dataclass
import numpy as np
from virtual_gpu_driver.src.ai.tensor_types import TensorDescriptor, DType, Device, Layout
from virtual_gpu_driver.src.stream import Stream as ComputeStream
from virtual_gpu_driver.src.stream import StreamManager as KernelSchedule
from .main import get_device, get_default_device
from .layer_norm import HeliumLayerNorm
from .gelu import HeliumGELU
from .multihead_attention import HeliumMultiHeadAttention
from .core.db_manager import HeliumDBManager
from .broadcast import ModalityType, TensorMetadata
@dataclass
class DecoderConfig:
"""Configuration for multi-modal decoder"""
output_modalities: List[ModalityType]
hidden_dim: int
num_layers: int
num_heads: int
intermediate_size: int
max_seq_len: Dict[ModalityType, int]
vocab_size: Optional[int] = None # For text generation
image_size: Optional[Tuple[int, int]] = None # For image generation
audio_params: Optional[Dict[str, Any]] = None # For audio generation
use_cache: bool = True
dtype: str = "float16"
def validate(self):
"""Validate configuration"""
for modality in self.output_modalities:
if modality == ModalityType.TEXT and not self.vocab_size:
raise ValueError("vocab_size required for text generation")
elif modality == ModalityType.IMAGE and not self.image_size:
raise ValueError("image_size required for image generation")
elif modality == ModalityType.AUDIO and not self.audio_params:
raise ValueError("audio_params required for audio generation")
if TYPE_CHECKING:
from .main import HeliumTensor
class ModalityProjection:
"""Projects hidden states to modality-specific outputs"""
def __init__(
self,
config: DecoderConfig,
modality: ModalityType,
driver=None
):
self.config = config
self.modality = modality
self.driver = driver
if modality == ModalityType.TEXT:
self.proj = self._create_linear(
config.hidden_dim,
config.vocab_size
)
elif modality == ModalityType.IMAGE:
h, w = config.image_size
self.proj = self._create_linear(
config.hidden_dim,
h * w * 3 # RGB channels
)
elif modality == ModalityType.AUDIO:
self.proj = self._create_linear(
config.hidden_dim,
config.audio_params["num_samples"]
)
def _create_linear(self, in_features: int, out_features: int) -> Dict[str, Any]:
"""Create projection layer"""
weight_desc = TensorDescriptor(
shape=(out_features, in_features),
dtype=DType.FLOAT16,
device=Device.VGPU,
layout=Layout.ROW_MAJOR
)
bias_desc = TensorDescriptor(
shape=(out_features,),
dtype=DType.FLOAT16,
device=Device.VGPU,
layout=Layout.ROW_MAJOR
)
return {
'weight': self.driver.allocate_tensor(weight_desc),
'bias': self.driver.allocate_tensor(bias_desc)
}
def forward(
self,
hidden_states: Union[str, "HeliumTensor"]
) -> Union[str, "HeliumTensor"]:
"""Project to modality-specific output space"""
out = self.driver.matmul(hidden_states, self.proj['weight'])
out = self.driver.add(out, self.proj['bias'])
if self.modality == ModalityType.IMAGE:
# Reshape to image format (B, H, W, C)
h, w = self.config.image_size
out = self.driver.reshape(out, (-1, h, w, 3))
elif self.modality == ModalityType.AUDIO:
# Apply audio-specific processing
if self.config.audio_params.get("normalize", True):
out = self.driver.tanh(out)
return out
class HeliumDecoderBlock:
"""
Hardware-accelerated multi-modal transformer decoder block
Implements:
1. Self-attention with causal mask
2. Cross-attention with encoder outputs
3. Feed-forward network
4. Multi-modal output projections
All operations run directly on virtual GPU with modality awareness
"""
def __init__(
self,
config: DecoderConfig,
device_id: Optional[str] = None
):
# Initialize device and stream
self.driver = get_device(device_id) if device_id else get_default_device()
self.device_id = device_id
self.stream = ComputeStream(self.driver)
# Initialize database connection
self.db = HeliumDBManager.get_instance()
# Store configuration
self.config = config
# Architecture parameters
self.hidden_size = config.hidden_dim
self.num_heads = config.num_heads
self.head_dim = config.hidden_dim // config.num_heads
self.intermediate_size = config.intermediate_size
self.dtype = config.dtype
# Initialize layer components
self.self_attention = HeliumMultiHeadAttention(
hidden_size=self.hidden_size,
num_heads=self.num_heads,
device_id=device_id,
dtype=self.dtype
)
self.cross_attention = HeliumMultiHeadAttention(
hidden_size=self.hidden_size,
num_heads=self.num_heads,
device_id=device_id,
dtype=self.dtype
)
# Layer norms
self.ln1 = HeliumLayerNorm(self.hidden_size, device_id=device_id, dtype=self.dtype)
self.ln2 = HeliumLayerNorm(self.hidden_size, device_id=device_id, dtype=self.dtype)
self.ln3 = HeliumLayerNorm(self.hidden_size, device_id=device_id, dtype=self.dtype)
# Feed-forward layers
self.ff1 = self._create_linear(self.hidden_size, self.intermediate_size)
self.ff2 = self._create_linear(self.intermediate_size, self.hidden_size)
self.gelu = HeliumGELU(device_id=device_id)
# Initialize modality-specific output projections
self.output_projections = {
modality: ModalityProjection(config, modality, self.driver)
for modality in config.output_modalities
}
# Operation scheduling
self.schedule = KernelSchedule(self.driver)
# Track allocated tensors
self._temp_tensors = {}
self._counter = 0
# Initialize layer components
self.self_attention = HeliumMultiHeadAttention(
hidden_size=hidden_size,
num_heads=num_heads,
device_id=device_id,
dtype=dtype
)
self.cross_attention = HeliumMultiHeadAttention(
hidden_size=hidden_size,
num_heads=num_heads,
device_id=device_id,
dtype=dtype
)
# Layer norms
self.ln1 = HeliumLayerNorm(hidden_size, device_id=device_id, dtype=dtype)
self.ln2 = HeliumLayerNorm(hidden_size, device_id=device_id, dtype=dtype)
self.ln3 = HeliumLayerNorm(hidden_size, device_id=device_id, dtype=dtype)
# Feed-forward layers
self.ff1 = self._create_linear(hidden_size, intermediate_size)
self.ff2 = self._create_linear(intermediate_size, hidden_size)
self.gelu = HeliumGELU(device_id=device_id)
# Operation scheduling
self.schedule = KernelSchedule(self.driver)
# Track allocated tensors
self._temp_tensors = {}
self._counter = 0
def _create_linear(self, in_features: int, out_features: int) -> Dict[str, Any]:
"""Create a linear layer's weight tensors"""
weight_desc = TensorDescriptor(
shape=(out_features, in_features),
dtype=getattr(DType, self.dtype.upper()),
device=Device.VGPU,
layout=Layout.ROW_MAJOR
)
bias_desc = TensorDescriptor(
shape=(out_features,),
dtype=getattr(DType, self.dtype.upper()),
device=Device.VGPU,
layout=Layout.ROW_MAJOR
)
return {
'weight': self.driver.allocate_tensor(weight_desc),
'bias': self.driver.allocate_tensor(bias_desc)
}
def _get_temp_tensor(self, shape: tuple) -> str:
"""Allocate a temporary tensor"""
tensor_id = f"decoder_temp_{self._counter}"
self._counter += 1
desc = TensorDescriptor(
shape=shape,
dtype=getattr(DType, self.dtype.upper()),
device=Device.VGPU,
layout=Layout.ROW_MAJOR
)
self._temp_tensors[tensor_id] = self.driver.allocate_tensor(desc)
return tensor_id
def _free_temp_tensor(self, tensor_id: str):
"""Free a temporary tensor"""
if tensor_id in self._temp_tensors:
self.driver.free_tensor(self._temp_tensors[tensor_id])
del self._temp_tensors[tensor_id]
def __del__(self):
"""Clean up temporary tensors"""
for tensor_id in list(self._temp_tensors.keys()):
self._free_temp_tensor(tensor_id)
def forward(
self,
hidden_states: Union[str, "HeliumTensor"],
target_modality: ModalityType,
encoder_hidden_states: Optional[Union[str, "HeliumTensor"]] = None,
attention_mask: Optional[Union[str, "HeliumTensor"]] = None,
encoder_attention_mask: Optional[Union[str, "HeliumTensor"]] = None,
metadata: Optional[TensorMetadata] = None
) -> Union[str, "HeliumTensor"]:
"""
Forward pass of decoder block
Args:
hidden_states: Input tensor (B, S, H)
encoder_hidden_states: Optional encoder output (B, S_enc, H)
attention_mask: Optional attention mask for self-attention
encoder_attention_mask: Optional mask for encoder-decoder attention
Returns:
Output tensor (B, S, H)
"""
residual = hidden_states
# Self attention branch
with self.stream:
# Layer norm 1
hidden_states = self.ln1(hidden_states)
# Self attention
hidden_states = self.self_attention(
hidden_states,
attention_mask=attention_mask,
causal_mask=True # Always use causal mask in decoder
)
# Residual connection
hidden_states = self.driver.add(hidden_states, residual)
# Cross attention branch (if encoder present)
if encoder_hidden_states is not None:
residual = hidden_states
with self.stream:
# Layer norm 2
hidden_states = self.ln2(hidden_states)
# Cross attention
hidden_states = self.cross_attention(
hidden_states,
encoder_hidden_states,
attention_mask=encoder_attention_mask
)
# Residual connection
hidden_states = self.driver.add(hidden_states, residual)
# Feed-forward branch
residual = hidden_states
with self.stream:
# Layer norm 3
hidden_states = self.ln3(hidden_states)
# Feed-forward
hidden_states = self.driver.matmul(
hidden_states,
self.ff1['weight']
)
hidden_states = self.driver.add(hidden_states, self.ff1['bias'])
hidden_states = self.gelu(hidden_states)
hidden_states = self.driver.matmul(
hidden_states,
self.ff2['weight']
)
hidden_states = self.driver.add(hidden_states, self.ff2['bias'])
# Final residual
hidden_states = self.driver.add(hidden_states, residual)
# Project to target modality
if target_modality not in self.output_projections:
raise ValueError(f"No projection available for modality {target_modality}")
output = self.output_projections[target_modality].forward(hidden_states)
# Update metadata if provided
if metadata is not None:
metadata.modality = target_modality
if target_modality == ModalityType.IMAGE:
h, w = self.config.image_size
metadata.spatial_dims = (h, w)
metadata.channels = 3
elif target_modality == ModalityType.AUDIO:
metadata.sampling_rate = self.config.audio_params.get("sampling_rate")
elif target_modality == ModalityType.TEXT:
metadata.sequence_length = output.shape[1]
return output
|