Spaces:
Sleeping
Sleeping
File size: 12,366 Bytes
055a9c8 |
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 |
#!/usr/bin/env python3
"""
Mamba Encoder Swarm - Integration with Existing Mamba Implementation
Uses your existing Mamba components as building blocks for the swarm architecture
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import List, Optional, Tuple
# Import your existing Mamba components
from core.config import MambaConfig
from core.model import MambaModel
from core.mamba import MambaLayer, RMSNorm
from core.embedding import MambaEmbedding
class SwarmRouter(nn.Module):
"""
Routes input tokens to different encoder instances
This is the NEW component that enables the swarm architecture
"""
def __init__(self, d_model: int, num_encoders: int, routing_strategy: str = "learned"):
super().__init__()
self.d_model = d_model
self.num_encoders = num_encoders
self.routing_strategy = routing_strategy
if routing_strategy == "learned":
# Neural router that learns optimal token distribution
self.router_network = nn.Sequential(
nn.Linear(d_model, d_model // 2),
nn.SiLU(),
nn.Linear(d_model // 2, num_encoders),
nn.Softmax(dim=-1)
)
# Load balancing coefficient
self.load_balance_coef = 0.01
def forward(self, x: torch.Tensor) -> Tuple[List[torch.Tensor], torch.Tensor, torch.Tensor]:
"""
Route tokens to encoder instances
Args:
x: [batch, seq_len, d_model]
Returns:
encoder_inputs: List of inputs for each encoder
routing_weights: Weights for aggregation [batch, seq_len, num_encoders]
load_balance_loss: Loss term for training
"""
batch_size, seq_len, d_model = x.shape
if self.routing_strategy == "learned":
# Learn routing patterns
routing_logits = self.router_network(x) # [batch, seq_len, num_encoders]
routing_weights = F.gumbel_softmax(routing_logits, tau=1.0, hard=False)
# Load balancing loss to encourage equal usage
avg_routing = routing_weights.mean(dim=[0, 1])
load_balance_loss = self.load_balance_coef * torch.var(avg_routing)
else: # Round-robin for simplicity
seq_indices = torch.arange(seq_len, device=x.device)
encoder_ids = seq_indices % self.num_encoders
routing_weights = F.one_hot(encoder_ids, self.num_encoders).float()
routing_weights = routing_weights.unsqueeze(0).expand(batch_size, -1, -1)
load_balance_loss = torch.tensor(0.0, device=x.device)
# Create weighted inputs for each encoder
encoder_inputs = []
for i in range(self.num_encoders):
weight = routing_weights[:, :, i:i+1] # [batch, seq_len, 1]
encoder_input = x * weight
encoder_inputs.append(encoder_input)
return encoder_inputs, routing_weights, load_balance_loss
class SwarmAggregator(nn.Module):
"""
Aggregates outputs from all encoder instances
This is the NEW component that combines swarm outputs
"""
def __init__(self, d_model: int, num_encoders: int):
super().__init__()
self.d_model = d_model
self.num_encoders = num_encoders
# Attention-based aggregation
self.attention = nn.MultiheadAttention(
embed_dim=d_model,
num_heads=8,
batch_first=True
)
# Output processing
self.norm = RMSNorm(d_model)
self.output_proj = nn.Linear(d_model, d_model)
def forward(self, encoder_outputs: List[torch.Tensor], routing_weights: torch.Tensor) -> torch.Tensor:
"""
Aggregate encoder outputs using learned attention
Args:
encoder_outputs: List of [batch, seq_len, d_model] tensors
routing_weights: [batch, seq_len, num_encoders]
Returns:
aggregated: [batch, seq_len, d_model]
"""
batch_size, seq_len, d_model = encoder_outputs[0].shape
# Stack and weight encoder outputs
stacked = torch.stack(encoder_outputs, dim=2) # [batch, seq_len, num_encoders, d_model]
routing_expanded = routing_weights.unsqueeze(-1) # [batch, seq_len, num_encoders, 1]
weighted = stacked * routing_expanded
# Initial aggregation
initial = weighted.sum(dim=2) # [batch, seq_len, d_model]
# Attention-based refinement
encoder_sequence = stacked.view(batch_size, seq_len * self.num_encoders, d_model)
refined, _ = self.attention(initial, encoder_sequence, encoder_sequence)
# Final processing
output = self.output_proj(refined)
output = self.norm(output + initial) # Residual connection
return output
class MambaEncoderSwarmModel(nn.Module):
"""
Complete Swarm Model using your existing Mamba components
Architecture:
1. Use your MambaEmbedding for input processing
2. NEW: Router distributes tokens to encoder swarm
3. Use your MambaLayer instances as shared encoders
4. NEW: Aggregator combines encoder outputs
5. Use your MambaLayer instances for decoder
6. Use your existing LM head for output
"""
def __init__(self, config: MambaConfig, num_encoders: int = 8, routing_strategy: str = "learned"):
super().__init__()
self.config = config
self.num_encoders = num_encoders
# Use your existing embedding
self.embedding = MambaEmbedding(config)
# NEW: Swarm components
self.router = SwarmRouter(config.d_model, num_encoders, routing_strategy)
# Shared encoder (using your MambaLayer)
# All encoder instances will use this same layer (weight sharing!)
self.shared_encoder_layer = MambaLayer(config)
# NEW: Aggregator
self.aggregator = SwarmAggregator(config.d_model, num_encoders)
# Decoder layers (using your MambaLayer)
self.decoder_layers = nn.ModuleList([
MambaLayer(config) for _ in range(config.n_layers)
])
# Use your existing components
self.norm_f = RMSNorm(config.d_model)
self.lm_head = nn.Linear(config.d_model, config.vocab_size, bias=False)
# Initialize weights
self.apply(self._init_weights)
def _init_weights(self, module):
if isinstance(module, nn.Linear):
nn.init.normal_(module.weight, mean=0.0, std=0.02)
if module.bias is not None:
nn.init.zeros_(module.bias)
elif isinstance(module, nn.Embedding):
nn.init.normal_(module.weight, mean=0.0, std=0.02)
def forward(self, input_ids: torch.Tensor, targets: torch.Tensor = None):
"""
Forward pass through swarm architecture
Args:
input_ids: [batch, seq_len]
targets: [batch, seq_len] (optional, for training)
Returns:
if targets is None: logits [batch, seq_len, vocab_size]
else: (logits, loss, load_balance_loss)
"""
# 1. Embedding (using your existing component)
x = self.embedding(input_ids) # [batch, seq_len, d_model]
# 2. Route to encoder swarm
encoder_inputs, routing_weights, load_balance_loss = self.router(x)
# 3. Process through shared encoder instances
encoder_outputs = []
for encoder_input in encoder_inputs:
# Each instance uses the SAME shared_encoder_layer (weight sharing!)
encoder_output = self.shared_encoder_layer(encoder_input)
encoder_outputs.append(encoder_output)
# 4. Aggregate encoder outputs
x = self.aggregator(encoder_outputs, routing_weights)
# 5. Process through decoder (using your existing layers)
for decoder_layer in self.decoder_layers:
x = decoder_layer(x)
# 6. Final processing (using your existing components)
x = self.norm_f(x)
logits = self.lm_head(x) # [batch, seq_len, vocab_size]
if targets is not None:
# Compute loss
loss = F.cross_entropy(
logits.view(-1, logits.size(-1)),
targets.view(-1),
ignore_index=-100
)
return logits, loss, load_balance_loss
return logits
def generate(self, input_ids: torch.Tensor, max_new_tokens: int = 100,
temperature: float = 1.0, top_k: int = None):
"""Generate using swarm architecture"""
self.eval()
for _ in range(max_new_tokens):
with torch.no_grad():
logits = self.forward(input_ids)
logits = logits[:, -1, :] / temperature
if top_k is not None:
v, _ = torch.topk(logits, top_k)
logits[logits < v[:, [-1]]] = -float('Inf')
probs = F.softmax(logits, dim=-1)
next_token = torch.multinomial(probs, num_samples=1)
input_ids = torch.cat([input_ids, next_token], dim=1)
return input_ids
def get_num_params(self):
"""Get number of parameters"""
return sum(p.numel() for p in self.parameters() if p.requires_grad)
def create_swarm_from_existing_config(config: MambaConfig, num_encoders: int = 8) -> MambaEncoderSwarmModel:
"""
Create swarm model using your existing configuration
"""
swarm_model = MambaEncoderSwarmModel(config, num_encoders, routing_strategy="learned")
num_params = swarm_model.get_num_params()
print(f"π Swarm model created with {num_params:,} parameters ({num_params/1e6:.1f}M)")
print(f"π Using {num_encoders} encoder instances with shared weights")
return swarm_model
def compare_architectures(config: MambaConfig):
"""
Compare standard Mamba vs Swarm architecture
"""
print("π Architecture Comparison")
print("=" * 50)
# Standard model (your existing)
standard_model = MambaModel(config)
standard_params = standard_model.get_num_params()
# Swarm model (new architecture)
swarm_model = create_swarm_from_existing_config(config, num_encoders=8)
swarm_params = swarm_model.get_num_params()
print(f"π Standard Mamba: {standard_params:,} parameters ({standard_params/1e6:.1f}M)")
print(f"π₯ Swarm Mamba: {swarm_params:,} parameters ({swarm_params/1e6:.1f}M)")
print(f"π‘ Parameter overhead: {((swarm_params - standard_params) / standard_params * 100):.1f}%")
return standard_model, swarm_model
if __name__ == "__main__":
# Test with your existing config
from core.config import MambaConfig
# Create a test config
config = MambaConfig(
vocab_size=50257,
d_model=512,
n_layers=8,
d_state=16,
d_conv=4,
bias=False
)
print("π§ͺ Testing Swarm Integration")
print("=" * 40)
# Compare architectures
standard_model, swarm_model = compare_architectures(config)
# Test forward pass
batch_size, seq_len = 2, 32
input_ids = torch.randint(0, config.vocab_size, (batch_size, seq_len))
# Test standard model
with torch.no_grad():
standard_logits = standard_model(input_ids)
print(f"β
Standard model output: {standard_logits.shape}")
# Test swarm model
with torch.no_grad():
swarm_logits = swarm_model(input_ids)
print(f"β
Swarm model output: {swarm_logits.shape}")
print(f"\nπ Both architectures working! Ready to train the swarm.") |