File size: 6,928 Bytes
77d27ba | 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 | """AAM Diffusion LLM — Thinking Toggle
Detects whether input needs deep reasoning (thinking) or quick response
(non-thinking). AAM-specific: simple factual query = 2 anchored steps,
complex reasoning = 5-10 steps + MCTS.
"""
from __future__ import annotations
import math
from dataclasses import dataclass
from enum import Enum
from typing import Optional
import torch
import torch.nn as nn
import torch.nn.functional as F
class ThinkingMode(Enum):
THINKING = "thinking"
NON_THINKING = "non_thinking"
class TaskType(Enum):
SEQUENTIAL = "sequential"
REASONING = "reasoning"
FACTUAL = "factual"
CREATIVE = "creative"
ANOMALY_RESOLUTION = "anomaly_resolution"
@dataclass
class ThinkingAssessment:
mode: ThinkingMode
complexity_score: torch.Tensor
task_type_probs: torch.Tensor
dominant_task: TaskType
depth_multiplier: torch.Tensor
confidence: torch.Tensor
thinking_score: Optional[torch.Tensor] = None
class ThinkingToggle(nn.Module):
"""Thinking/Non-Thinking Toggle for AAM Diffusion LLM."""
NUM_TASK_TYPES = len(TaskType)
def __init__(self, d_model: int, threshold: float = 0.5) -> None:
super().__init__()
self.d_model = d_model
self.threshold = threshold
self.complexity_scorer = nn.Sequential(
nn.Linear(d_model, d_model // 2),
nn.SiLU(),
nn.Linear(d_model // 2, d_model // 4),
nn.SiLU(),
nn.Linear(d_model // 4, 1),
nn.Sigmoid(),
)
self.task_classifier = nn.Sequential(
nn.Linear(d_model, d_model // 2),
nn.SiLU(),
nn.Linear(d_model // 2, self.NUM_TASK_TYPES),
)
self.context_integrator = nn.Sequential(
nn.Linear(1 + self.NUM_TASK_TYPES, d_model // 4),
nn.SiLU(),
nn.Linear(d_model // 4, 1),
nn.Sigmoid(),
)
self.depth_min = 0.3
self.depth_max = 2.0
self.register_buffer("_force_mode_code", torch.tensor(-1, dtype=torch.long), persistent=True)
def forward(self, x: torch.Tensor, force_mode: Optional[ThinkingMode] = None) -> ThinkingAssessment:
if x.dim() != 3:
raise ValueError(f"Input must be 3D [batch, seq, d_model], got {x.dim()}D")
complexity = self.complexity_scorer(x).squeeze(-1)
task_logits = self.task_classifier(x)
task_probs = F.softmax(task_logits, dim=-1)
mean_complexity = complexity.mean(dim=-1, keepdim=True)
mean_task_probs = task_probs.mean(dim=1)
context_input = torch.cat([mean_complexity, mean_task_probs], dim=-1)
thinking_score = self.context_integrator(context_input).squeeze(-1)
# v2.3.0: Use force_mode kwarg if provided (thread-safe, no state mutation).
# Falls back to _get_force_mode() for backward compatibility.
if force_mode is not None:
mode = force_mode
else:
persistent_mode = self._get_force_mode()
if persistent_mode is not None:
mode = persistent_mode
else:
# v1.8.0: Straight-through estimator for differentiable depth_multiplier.
# Forward pass uses hard threshold for control flow (must be non-differentiable),
# but depth_multiplier remains fully differentiable through soft blending.
avg_score_val = thinking_score.mean().item()
mode = ThinkingMode.THINKING if avg_score_val > self.threshold else ThinkingMode.NON_THINKING
overall_task_probs = task_probs.mean(dim=(0, 1))
dominant_task_idx = overall_task_probs.argmax().item()
dominant_task = list(TaskType)[dominant_task_idx]
avg_thinking_score = thinking_score
temperature = 5.0
mode_weight = torch.sigmoid(temperature * (avg_thinking_score - self.threshold))
thinking_depth = self.depth_min + (self.depth_max - self.depth_min) * avg_thinking_score
non_thinking_depth = self.depth_min + 0.2 * avg_thinking_score
depth_multiplier = mode_weight * thinking_depth + (1.0 - mode_weight) * non_thinking_depth
confidence = 1.0 - (avg_thinking_score - self.threshold).abs() / max(self.threshold, 1.0 - self.threshold)
confidence = confidence.clamp(0.0, 1.0)
return ThinkingAssessment(
mode=mode,
complexity_score=complexity,
task_type_probs=task_probs,
dominant_task=dominant_task,
depth_multiplier=depth_multiplier,
confidence=confidence,
thinking_score=thinking_score,
)
def _get_force_mode(self) -> Optional[ThinkingMode]:
"""Decode the persistent buffer back to a ThinkingMode (or None)."""
code = int(self._force_mode_code.item())
if code == -1:
return None
elif code == 0:
return ThinkingMode.NON_THINKING
elif code == 1:
return ThinkingMode.THINKING
else:
# Corrupted value — reset to auto
self._force_mode_code.fill_(-1)
return None
def set_force_mode(self, mode: Optional[ThinkingMode]) -> None:
"""Force mode, bypassing detection. Set None for automatic detection.
The mode is persisted via a registered buffer so it survives
model.state_dict() / model.load_state_dict() round-trips.
"""
if mode is None:
self._force_mode_code.fill_(-1)
elif mode == ThinkingMode.NON_THINKING:
self._force_mode_code.fill_(0)
elif mode == ThinkingMode.THINKING:
self._force_mode_code.fill_(1)
else:
raise ValueError(f"Unknown ThinkingMode: {mode!r}")
def set_threshold(self, threshold: float) -> None:
"""Update complexity threshold.
Args:
threshold: New threshold value (0.0 - 1.0)
"""
if not 0.0 <= threshold <= 1.0:
raise ValueError(f"Threshold must be between 0.0 and 1.0, got {threshold}")
self.threshold = threshold
def get_thinking_mask(self, assessment: ThinkingAssessment, seq_len: int) -> torch.Tensor:
"""Create binary mask marking which tokens need thinking.
Args:
assessment: Assessment result from forward
seq_len: Sequence length
Returns:
Mask [batch, seq] — 1.0 for thinking, 0.0 for non-thinking
"""
mask = (assessment.complexity_score > self.threshold).float()
return mask
def get_depth_schedule(self, assessment: ThinkingAssessment) -> torch.Tensor:
complexity = assessment.complexity_score
depth = self.depth_min + (self.depth_max - self.depth_min) * complexity
if assessment.mode == ThinkingMode.NON_THINKING:
depth = depth.clamp(max=self.depth_min + 0.3)
return depth
|