File size: 9,131 Bytes
9abace2 | 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 | """
Stage 7 + 8 + 9: SparseActivation + RoutingNodes + NANDFilter
SparseActivation:
Computes routing weights using a softmax over signal projections,
then applies top-k sparsity gate.
Only top-k experts receive non-zero weight.
RoutingNodes:
Per-expert gating functions. Each expert has:
- A signal affinity vector (which signals it responds to)
- A threshold (minimum signal strength to activate)
- A temperature (sharpness of its gating curve)
Gating output = sigmoid((dot(affinity, signals) - threshold) / temperature)
NANDFilter:
Conflict resolution via NAND logic.
If expert A and expert B are both active AND their combination
is flagged as conflicting, at least one must be suppressed.
NAND(A, B) = NOT(A AND B)
→ If both active + conflict registered → suppress lower-weight one.
Combined output: sparse routing weight vector.
"""
from __future__ import annotations
import math
from dataclasses import dataclass, field
# ==========================================
# Routing Nodes
# ==========================================
@dataclass
class RoutingNode:
expert_name: str
signal_affinity: dict[str, float] # signal_name -> affinity weight
threshold: float = 0.1 # minimum activation
temperature: float = 1.0 # gating sharpness
max_weight: float = 1.0 # ceiling on this expert's weight
def gate(self, signals: dict[str, float]) -> float:
"""
Compute gating score for this expert given signals.
Returns value in [0, max_weight].
"""
# Dot product of affinity and signals
dot = sum(
self.signal_affinity.get(sig, 0.0) * val
for sig, val in signals.items()
)
# Sigmoid gating: σ((dot - threshold) / temperature)
x = (dot - self.threshold) / max(self.temperature, 1e-6)
gate_val = 1.0 / (1.0 + math.exp(-x))
return min(gate_val, self.max_weight)
def build_default_routing_nodes(expert_names: list[str]) -> dict[str, RoutingNode]:
"""
Build default routing nodes.
Each expert gets affinity based on its name keywords.
"""
nodes: dict[str, RoutingNode] = {}
for name in expert_names:
name_lower = name.lower()
# Infer affinity from expert name
affinity: dict[str, float] = {}
if any(k in name_lower for k in ("code", "coder", "dev", "engineer", "impl")):
affinity["code_signal"] = 0.9
affinity["language_signal"] = 0.7
affinity["entity_signal"] = 0.3
if any(k in name_lower for k in ("query", "search", "retrieve", "rag", "lookup")):
affinity["query_signal"] = 0.9
affinity["entity_signal"] = 0.5
if any(k in name_lower for k in ("reason", "logic", "math", "proof", "formal")):
affinity["operator_signal"] = 0.8
affinity["constraint_signal"] = 0.6
affinity["code_signal"] = 0.4
if any(k in name_lower for k in ("constraint", "guard", "safety", "valid")):
affinity["constraint_signal"] = 0.9
affinity["operator_signal"] = 0.5
if any(k in name_lower for k in ("chat", "converse", "respond", "general")):
affinity["query_signal"] = 0.5
affinity["code_signal"] = 0.3
# Default affinity if name gives no hints
if not affinity:
affinity = {sig: 0.3 for sig in [
"code_signal", "query_signal", "constraint_signal",
"operator_signal", "language_signal", "entity_signal"
]}
nodes[name] = RoutingNode(
expert_name=name,
signal_affinity=affinity,
threshold=0.1,
temperature=0.5
)
return nodes
# ==========================================
# NAND Filter
# ==========================================
@dataclass
class NANDConflict:
expert_a: str
expert_b: str
reason: str
class NANDFilter:
"""
Stage 9: Conflict resolution.
NAND(A, B) = if both active + conflict → suppress lower weight.
"""
def __init__(self):
self._conflicts: list[NANDConflict] = []
def register_conflict(self, a: str, b: str, reason: str = "") -> None:
self._conflicts.append(NANDConflict(a, b, reason))
def apply(self, weights: dict[str, float]) -> dict[str, float]:
"""
Apply NAND filtering.
For each conflict pair where both are active:
suppress the one with lower weight (set to 0).
"""
result = dict(weights)
for conflict in self._conflicts:
wa = result.get(conflict.expert_a, 0.0)
wb = result.get(conflict.expert_b, 0.0)
# Both active?
if wa > 1e-6 and wb > 1e-6:
# NAND: suppress lower weight
if wa <= wb:
result[conflict.expert_a] = 0.0
else:
result[conflict.expert_b] = 0.0
return result
# ==========================================
# Sparse Activation + Routing Weights
# ==========================================
@dataclass
class RoutingWeights:
weights: dict[str, float] # expert_name -> weight (sum to 1.0)
active_experts: list[str] # experts with weight > 0
top_k: int
raw_scores: dict[str, float] # before normalization
gating_scores: dict[str, float] # from routing nodes
nand_suppressed: list[str] # experts zeroed by NAND
class SparseActivation:
"""
Stage 7: Compute sparse routing weights.
Algorithm:
1. Each RoutingNode produces a gating score per expert
2. Multiply by constraint mask (blocked = 0)
3. Apply top-k sparsity: zero out all but top-k scores
4. Apply NAND filter: resolve conflicts
5. Softmax normalize surviving scores
"""
def __init__(
self,
top_k: int = 2,
routing_nodes: dict[str, RoutingNode] | None = None,
nand_filter: NANDFilter | None = None
):
self.top_k = top_k
self.routing_nodes = routing_nodes or {}
self.nand_filter = nand_filter or NANDFilter()
def compute(
self,
signals: dict[str, float],
expert_names: list[str],
expert_mask: dict[str, bool], # from ConstraintEval
jordan_features: dict[str, float] # from JordanResult
) -> RoutingWeights:
# Ensure routing nodes exist for all experts
for name in expert_names:
if name not in self.routing_nodes:
self.routing_nodes[name] = build_default_routing_nodes([name])[name]
# Step 1: Gating scores from routing nodes
gating_scores: dict[str, float] = {}
for name in expert_names:
node = self.routing_nodes[name]
score = node.gate(signals)
# Scale by Jordan routing confidence (stable graph = more confident)
routing_conf = jordan_features.get("routing_confidence", 1.0)
score = score * (0.5 + 0.5 * routing_conf)
gating_scores[name] = score
# Step 2: Apply constraint mask
masked: dict[str, float] = {
name: (score if expert_mask.get(name, True) else 0.0)
for name, score in gating_scores.items()
}
# Step 3: Top-k sparsity
sorted_experts = sorted(masked.items(), key=lambda x: x[1], reverse=True)
sparse: dict[str, float] = {}
for i, (name, score) in enumerate(sorted_experts):
sparse[name] = score if i < self.top_k else 0.0
# Step 4: NAND filter
after_nand = self.nand_filter.apply(sparse)
nand_suppressed = [
name for name in expert_names
if sparse.get(name, 0.0) > 0 and after_nand.get(name, 0.0) == 0.0
]
# Step 5: Softmax normalize
active = {name: w for name, w in after_nand.items() if w > 1e-6}
normalized = self._softmax(active)
# Zero out everything not in normalized
final_weights: dict[str, float] = {name: 0.0 for name in expert_names}
final_weights.update(normalized)
return RoutingWeights(
weights=final_weights,
active_experts=list(normalized.keys()),
top_k=self.top_k,
raw_scores=gating_scores,
gating_scores=gating_scores,
nand_suppressed=nand_suppressed
)
def _softmax(self, scores: dict[str, float]) -> dict[str, float]:
if not scores:
return {}
vals = list(scores.values())
max_val = max(vals)
exp_vals = {name: math.exp(v - max_val) for name, v in scores.items()}
total = sum(exp_vals.values())
if total == 0:
return {}
return {name: v / total for name, v in exp_vals.items()}
|