File size: 10,790 Bytes
7a564e3 | 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 | """
MoE Decoder — Mixture of Experts text decoder for VL-JEPA.
Takes a predicted embedding from the JEPA predictor and autoregressively generates
text output. Each transformer block's FFN is replaced with a MoE layer containing
task-specialized experts.
Only invoked when selective decoding detects a semantic shift — NOT on every frame.
v2 additions:
- apply_lora() / clear_lora(): Inject/remove per-camera LoRA adapters
generated by the HyperNetwork via HyperMother orchestrator.
- LoRA targets Q, V attention projections only (NOT MoE FFN gating).
"""
from __future__ import annotations
from typing import Optional
import torch
import torch.nn as nn
import torch.nn.functional as F
from model.transformer import TransformerBlock
from model.moe import MoELayer
class MoEDecoder(nn.Module):
"""
MoE Decoder — generates text from predicted embedding.
Architecture:
Predicted embedding → prepend as first token
→ Token embedding + positional embedding
→ N × TransformerBlock(causal, MoE FFN)
→ LayerNorm → Linear head → logits
The predicted embedding from JEPA predictor is used as the initial "thought"
token, and the decoder autoregressively generates the text output.
Args:
hidden_dim: Transformer dimension (768)
embed_dim: Input embedding dimension from predictor (1536)
vocab_size: BPE vocabulary size (8192)
num_heads: Number of attention heads (12)
num_blocks: Number of transformer blocks (6)
num_experts: Number of experts per MoE layer (5)
top_k: Active experts per token (2)
max_seq_len: Maximum output sequence length (512)
dropout: Dropout rate
"""
def __init__(
self,
hidden_dim: int = 768,
embed_dim: int = 1536,
vocab_size: int = 8192,
num_heads: int = 12,
num_blocks: int = 6,
num_experts: int = 5,
top_k: int = 2,
max_seq_len: int = 512,
dropout: float = 0.1,
):
super().__init__()
self.hidden_dim = hidden_dim
self.vocab_size = vocab_size
self.max_seq_len = max_seq_len
# Project predictor embedding to decoder dimension
self.embed_proj = nn.Linear(embed_dim, hidden_dim)
# Token and position embeddings
self.token_embed = nn.Embedding(vocab_size, hidden_dim, padding_idx=0)
self.pos_embed = nn.Parameter(torch.randn(1, max_seq_len + 1, hidden_dim) * 0.02) # +1 for embedding token
self.embed_dropout = nn.Dropout(dropout)
# Transformer blocks with MoE FFN (causal attention for autoregressive generation)
self.blocks = nn.ModuleList()
self.moe_layers: list[MoELayer] = []
self._feature_gates_enabled = False
for _ in range(num_blocks):
moe = MoELayer(hidden_dim, num_experts, top_k, dropout=dropout)
self.moe_layers.append(moe)
block = TransformerBlock(hidden_dim, num_heads, dropout, mode="causal", ffn=moe)
self.blocks.append(block)
# Control signal for v2 feature gating (set externally by HyperMother)
self._control_signal: Optional[torch.Tensor] = None
self.norm = nn.LayerNorm(hidden_dim)
# Output head: hidden_dim → vocab_size
self.lm_head = nn.Linear(hidden_dim, vocab_size, bias=False)
def forward(
self,
pred_embedding: torch.Tensor,
target_ids: torch.Tensor | None = None,
) -> tuple[torch.Tensor, torch.Tensor | None]:
"""
Args:
pred_embedding: [batch, embed_dim] — predicted embedding from JEPA predictor
target_ids: [batch, seq_len] — target token IDs for training (teacher forcing)
Returns:
logits: [batch, seq_len+1, vocab_size] — token predictions
loss: scalar loss if target_ids provided, else None
"""
B = pred_embedding.shape[0]
# Project embedding to decoder dimension: [B, embed_dim] → [B, 1, hidden_dim]
embed_token = self.embed_proj(pred_embedding).unsqueeze(1)
if target_ids is not None:
# Training: teacher forcing
token_embeds = self.token_embed(target_ids) # [B, T, hidden_dim]
x = torch.cat([embed_token, token_embeds], dim=1) # [B, 1+T, hidden_dim]
else:
# Inference: start with just the embedding token
x = embed_token # [B, 1, hidden_dim]
T = x.shape[1]
x = x + self.pos_embed[:, :T, :]
x = self.embed_dropout(x)
# Pass through MoE transformer blocks
for block in self.blocks:
x = block(x, control_signal=self._control_signal)
x = self.norm(x)
logits = self.lm_head(x) # [B, T, vocab_size]
loss = None
if target_ids is not None:
# Shift logits and targets for next-token prediction
# Logits from positions [0, ..., T-1] predict tokens at [1, ..., T]
# Position 0 (embedding token) predicts first text token
shift_logits = logits[:, :-1, :].contiguous() # [B, T, vocab]
shift_targets = target_ids.contiguous() # [B, T]
loss = F.cross_entropy(
shift_logits.view(-1, self.vocab_size),
shift_targets.view(-1),
ignore_index=0, # ignore padding
)
return logits, loss
@torch.no_grad()
def generate(
self,
pred_embedding: torch.Tensor,
max_new_tokens: int = 256,
temperature: float = 0.0,
eos_id: int = 2,
) -> torch.Tensor:
"""
Autoregressive text generation.
Args:
pred_embedding: [batch, embed_dim] — predicted embedding
max_new_tokens: Maximum tokens to generate
temperature: Sampling temperature
eos_id: End of sequence token ID
Returns:
[batch, generated_len] — generated token IDs
"""
B = pred_embedding.shape[0]
device = pred_embedding.device
# Start with just the embedding token
embed_token = self.embed_proj(pred_embedding).unsqueeze(1) # [B, 1, hidden_dim]
generated = torch.zeros(B, 0, dtype=torch.long, device=device)
x = embed_token
for _ in range(max_new_tokens):
T = x.shape[1]
pos_x = x + self.pos_embed[:, :T, :]
h = pos_x
for block in self.blocks:
h = block(h, control_signal=self._control_signal)
h = self.norm(h)
# Get logits for last position only
next_logits = self.lm_head(h[:, -1, :]) # [B, vocab]
if temperature <= 0:
next_token = next_logits.argmax(dim=-1, keepdim=True)
else:
probs = F.softmax(next_logits / temperature, dim=-1)
next_token = torch.multinomial(probs, num_samples=1) # [B, 1]
generated = torch.cat([generated, next_token], dim=1)
# Check for EOS
if (next_token == eos_id).all():
break
# Append token embedding for next step
next_embed = self.token_embed(next_token) # [B, 1, hidden_dim]
x = torch.cat([x, next_embed], dim=1)
return generated
def get_all_load_balancing_data(self) -> list[tuple[torch.Tensor, torch.Tensor]]:
"""Collect load balancing data from all MoE layers."""
data = []
for moe in self.moe_layers:
lb_data = moe.get_load_balancing_data()
if lb_data is not None:
data.append(lb_data)
return data
# -- Feature control (v2) -----------------------------------------------
def enable_feature_gates(self, control_dim: int = 256) -> None:
"""
Add FeatureControlGate to all transformer blocks.
Called once during model initialization when v2 features are enabled.
Safe to call multiple times — gates are only added if not already present.
"""
from model.feature_control import FeatureControlGate
for block in self.blocks:
if block.feature_gate is None:
block.feature_gate = FeatureControlGate(
dim=self.hidden_dim,
control_dim=control_dim,
)
self._feature_gates_enabled = True
def set_control_signal(self, control_signal: Optional[torch.Tensor]) -> None:
"""
Set the control signal for feature gating.
Called by HyperMother before running decoder forward/generate.
Args:
control_signal: [B, control_dim] from ConditionEncoder, or None to clear
"""
self._control_signal = control_signal
# -- LoRA injection (v2) -------------------------------------------------
def apply_lora(
self,
lora_layers: list[dict[str, "nn.Module"]],
) -> None:
"""
Inject LoRA layers into all transformer blocks' attention modules.
Args:
lora_layers: List of dicts (one per block), mapping target name
("q", "v") to LoRALayer instances. Length must match
self.blocks.
"""
assert len(lora_layers) == len(self.blocks), (
f"Expected {len(self.blocks)} LoRA layer dicts, got {len(lora_layers)}"
)
for block, block_loras in zip(self.blocks, lora_layers):
block.attn.set_lora(
lora_q=block_loras.get("q"),
lora_v=block_loras.get("v"),
)
def clear_lora(self) -> None:
"""Remove all LoRA layers from all transformer blocks."""
for block in self.blocks:
block.attn.clear_lora()
@property
def has_lora(self) -> bool:
"""Whether any block currently has LoRA active."""
return any(block.attn.has_lora for block in self.blocks)
def apply_lora_from_flat(
self,
flat_params: torch.Tensor,
config: "LoRAConfig",
) -> None:
"""
Convenience: create and inject LoRA layers from a flat parameter vector.
This is the primary interface used by HyperMother — the HyperNetwork
outputs a flat tensor, and this method handles reshaping and injection.
Args:
flat_params: Flat parameter tensor from HyperNetwork
config: LoRA configuration
"""
from model.lora import LoRAInjector
injector = LoRAInjector(config, len(self.blocks), self.hidden_dim)
lora_layers = injector.create_lora_layers(
flat_params, device=next(self.parameters()).device
)
self.apply_lora(lora_layers)
|