File size: 9,935 Bytes
b182134 | 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 | """AAM Diffusion LLM — Anchored Diffusion Decoder
Replaces the standard softmax → token ID pipeline with:
1. Model predicts continuous vector (NO softmax)
2. 2-3 step anchored diffusion refinement
3. Disambiguation + coherence + Evoformer feedback
4. Final projection to vocabulary
Key Insight (from Losion):
Standard diffusion LLM: starts from NOISE → needs 50-1000 steps
Anchored diffusion: starts from PREDICTED VECTOR (already meaningful) → 2-3 steps only
The predicted vector serves as an "anchor" — it's already in the right
neighborhood of the output space. The decoder just needs to refine it.
AAM-specific: The anchor comes from graph-conditioned denoising, so it's
already shaped by evidence/anomaly/reasoning from the RSVS Knowledge Graph.
"""
from __future__ import annotations
import math
from dataclasses import dataclass
from typing import Dict, Optional, Tuple
import torch
import torch.nn as nn
import torch.nn.functional as F
@dataclass
class AnchoredDecoderConfig:
"""Configuration for Anchored Diffusion Decoder."""
d_model: int = 768
d_vocab: int = 32000
n_refine_steps: int = 3
d_refine: int = 512
use_evoformer_feedback: bool = True
n_feedback_iterations: int = 2
disambiguation_heads: int = 8
class DisambiguationBlock(nn.Module):
"""Resolve between similar tokens based on graph context.
The predicted continuous vector may fall between two tokens with similar
meanings (e.g., "bukti" vs "dugaan"). This block uses local context
and graph-conditioned attention to disambiguate.
"""
def __init__(self, d_model: int, n_heads: int = 8) -> None:
super().__init__()
self.d_model = d_model
self.n_heads = n_heads
self.d_kv = d_model // n_heads
self.q_proj = nn.Linear(d_model, d_model, bias=False)
self.k_proj = nn.Linear(d_model, d_model, bias=False)
self.v_proj = nn.Linear(d_model, d_model, bias=False)
self.out_proj = nn.Linear(d_model, d_model, bias=False)
self.gate = nn.Sequential(
nn.Linear(d_model, 1, bias=False),
nn.Sigmoid(),
)
self.norm = nn.RMSNorm(d_model)
self.scale = math.sqrt(self.d_kv)
def forward(self, x: torch.Tensor, graph_context: Optional[torch.Tensor] = None) -> torch.Tensor:
batch, seq_len, _ = x.shape
q = self.q_proj(x).view(batch, seq_len, self.n_heads, self.d_kv).transpose(1, 2)
# Use graph context as key/value if available, otherwise self-attention
if graph_context is not None:
k = self.k_proj(graph_context)
v = self.v_proj(graph_context)
if k.dim() == 3:
k = k.unsqueeze(1).expand(-1, self.n_heads, -1, -1).reshape(batch, -1, self.d_kv)
v = v.unsqueeze(1).expand(-1, self.n_heads, -1, -1).reshape(batch, -1, self.d_kv)
k = k.unsqueeze(1).transpose(1, 2) if k.dim() == 3 else k
# Simplified: use x for k,v if graph_context shape is tricky
k = self.k_proj(x).view(batch, seq_len, self.n_heads, self.d_kv).transpose(1, 2)
v = self.v_proj(x).view(batch, seq_len, self.n_heads, self.d_kv).transpose(1, 2)
else:
k = self.k_proj(x).view(batch, seq_len, self.n_heads, self.d_kv).transpose(1, 2)
v = self.v_proj(x).view(batch, seq_len, self.n_heads, self.d_kv).transpose(1, 2)
# Causal mask
scores = torch.matmul(q, k.transpose(-2, -1)) / self.scale
mask = torch.triu(torch.ones(seq_len, seq_len, device=x.device, dtype=torch.bool), diagonal=1)
scores = scores.masked_fill(mask.unsqueeze(0).unsqueeze(0), float("-inf"))
attn = F.softmax(scores, dim=-1, dtype=torch.float32).to(x.dtype)
context = torch.matmul(attn, v).transpose(1, 2).contiguous().view(batch, seq_len, self.d_model)
context = self.out_proj(context)
gate = self.gate(x)
refined = x + gate * context
refined = self.norm(refined)
return refined
class CoherenceBlock(nn.Module):
"""Ensure parallel tokens are consistent with each other and the graph.
When predicting multiple tokens in parallel (from the continuous vector
pipeline), each token's vector is predicted independently. This block
ensures they are coherent as a sequence.
"""
def __init__(self, d_model: int, d_refine: int = 512) -> None:
super().__init__()
self.d_model = d_model
self.coherence_mlp = nn.Sequential(
nn.Linear(d_model, d_refine, bias=False),
nn.SiLU(),
nn.Linear(d_refine, d_model, bias=False),
)
self.gate = nn.Sequential(
nn.Linear(d_model, 1, bias=False),
nn.Sigmoid(),
)
self.norm = nn.RMSNorm(d_model)
def forward(self, x: torch.Tensor) -> torch.Tensor:
mlp_out = self.coherence_mlp(x)
gate = self.gate(x)
refined = x + gate * mlp_out
refined = self.norm(refined)
return refined
class AnchoredDiffusionDecoder(nn.Module):
"""Anchored Diffusion Decoder — the core output pipeline for AAM v2.0.
Replaces the standard softmax → token ID pipeline with:
1. Model predicts continuous vector (NO softmax)
2. 2-3 step anchored diffusion refinement
3. Disambiguation + coherence + Evoformer feedback
4. Final projection to vocabulary
The key innovation: the predicted vector is ALREADY meaningful (it's
the model's best prediction after graph-conditioned denoising). The
decoder doesn't need to find the output from scratch — it just refines.
"""
def __init__(self, config: Optional[AnchoredDecoderConfig] = None) -> None:
super().__init__()
self.config = config or AnchoredDecoderConfig()
self.d_model = self.config.d_model
self.d_vocab = self.config.d_vocab
self.n_refine_steps = self.config.n_refine_steps
self.disambiguation = DisambiguationBlock(
d_model=self.d_model,
n_heads=self.config.disambiguation_heads,
)
self.coherence_blocks = nn.ModuleList([
CoherenceBlock(d_model=self.d_model, d_refine=self.config.d_refine)
for _ in range(self.n_refine_steps)
])
if self.config.use_evoformer_feedback:
self.feedback_proj = nn.Sequential(
nn.Linear(self.d_model, self.d_model, bias=False),
nn.SiLU(),
nn.Linear(self.d_model, self.d_model, bias=False),
)
self.feedback_gate = nn.Sequential(
nn.Linear(self.d_model, 1, bias=False),
nn.Sigmoid(),
)
self.feedback_norm = nn.RMSNorm(self.d_model)
self.vocab_proj = nn.Linear(self.d_model, self.d_vocab, bias=False)
self.pre_proj_norm = nn.RMSNorm(self.d_model)
def forward(
self,
predicted_vectors: torch.Tensor,
context: Optional[torch.Tensor] = None,
) -> Tuple[torch.Tensor, Dict[str, object]]:
x = predicted_vectors
info = {"n_refine_steps": self.n_refine_steps}
if self.config.use_evoformer_feedback:
for fb_iter in range(self.config.n_feedback_iterations):
disambiguated = self.disambiguation(x, context)
refined = disambiguated
for step in range(self.n_refine_steps):
refined = self.coherence_blocks[step](refined)
feedback = self.feedback_proj(refined - x)
gate = self.feedback_gate(x)
x = self.feedback_norm(x + gate * feedback)
info["feedback_iterations"] = self.config.n_feedback_iterations
else:
x = self.disambiguation(x, context)
for step in range(self.n_refine_steps):
x = self.coherence_blocks[step](x)
x = self.pre_proj_norm(x)
logits = self.vocab_proj(x)
delta = (x - predicted_vectors).norm(dim=-1).mean().item()
info["refinement_delta"] = delta
return logits, info
def predict_continuous(self, hidden_states: torch.Tensor) -> torch.Tensor:
"""Produce continuous prediction vectors (NO softmax)."""
return hidden_states
class ContinuousOutputHead(nn.Module):
"""Continuous output head that produces prediction vectors without softmax.
Replaces the standard nn.Linear → softmax pipeline with:
nn.Linear → continuous vector → AnchoredDiffusionDecoder → logits
"""
def __init__(
self,
d_model: int,
d_vocab: int = 32000,
decoder_config: Optional[AnchoredDecoderConfig] = None,
) -> None:
super().__init__()
self.d_model = d_model
self.d_vocab = d_vocab
self.predict_proj = nn.Sequential(
nn.Linear(d_model, d_model, bias=False),
nn.SiLU(),
nn.Linear(d_model, d_model, bias=False),
)
if decoder_config is None:
decoder_config = AnchoredDecoderConfig(d_model=d_model, d_vocab=d_vocab)
else:
decoder_config.d_model = d_model
decoder_config.d_vocab = d_vocab
self.decoder = AnchoredDiffusionDecoder(decoder_config)
def forward(
self,
hidden_states: torch.Tensor,
use_diffusion: bool = True,
context: Optional[torch.Tensor] = None,
) -> Tuple[torch.Tensor, Dict[str, object]]:
pred_vectors = self.predict_proj(hidden_states)
if use_diffusion:
return self.decoder(pred_vectors, context=context)
else:
logits = self.decoder.vocab_proj(self.decoder.pre_proj_norm(pred_vectors))
return logits, {"mode": "standard"}
def get_continuous_vectors(self, hidden_states: torch.Tensor) -> torch.Tensor:
return self.predict_proj(hidden_states)
|