File size: 13,057 Bytes
dd850a7 |
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 |
import torch
import numpy as np
from typing import List, Dict, Any, Optional
class AttentionProcessor:
@staticmethod
def process_attention_separate(
attention_data: Dict[str, Any],
input_tokens: List[str],
output_tokens: List[str]
) -> List[Dict[str, torch.Tensor]]:
"""
Process attention with separate normalization for input and output.
This preserves the relative importance within each group.
"""
attentions = attention_data['attentions']
input_len_for_attention = attention_data['input_len_for_attention']
output_len = attention_data['output_len']
if not attentions:
return [{'input_attention': torch.zeros(input_len_for_attention),
'output_attention': None} for _ in range(output_len)]
attention_matrices = []
num_steps = len(attentions)
if num_steps == 0:
print("Warning: No attention steps found in output.")
return [{'input_attention': torch.zeros(input_len_for_attention),
'output_attention': None} for _ in range(output_len)]
steps_to_process = min(num_steps, output_len)
for i in range(steps_to_process):
step_attentions = attentions[i]
input_attention_layers = []
output_attention_layers = []
for layer_idx, layer_attn in enumerate(step_attentions):
try:
# Extract attention to input tokens (skip BOS token at position 0)
input_indices = slice(1, 1 + input_len_for_attention)
if layer_attn.shape[3] >= input_indices.stop:
# Get attention from current token (position 0 in generation) to input
input_attn = layer_attn[0, :, 0, input_indices]
input_attention_layers.append(input_attn)
# Extract attention to previous output tokens
if i > 0:
output_indices = slice(1 + input_len_for_attention, 1 + input_len_for_attention + i)
if layer_attn.shape[3] >= output_indices.stop:
output_attn = layer_attn[0, :, 0, output_indices]
output_attention_layers.append(output_attn)
else:
output_attention_layers.append(
torch.zeros((layer_attn.shape[1], i), device=layer_attn.device)
)
else:
input_attention_layers.append(
torch.zeros((layer_attn.shape[1], input_len_for_attention), device=layer_attn.device)
)
if i > 0:
output_attention_layers.append(
torch.zeros((layer_attn.shape[1], i), device=layer_attn.device)
)
except Exception as e:
print(f"Error processing attention at step {i}, layer {layer_idx}: {e}")
input_attention_layers.append(
torch.zeros((layer_attn.shape[1], input_len_for_attention), device=layer_attn.device)
)
if i > 0:
output_attention_layers.append(
torch.zeros((layer_attn.shape[1], i), device=layer_attn.device)
)
# Average across layers and heads
if input_attention_layers:
avg_input_attn = torch.mean(torch.stack(input_attention_layers).float(), dim=[0, 1])
else:
avg_input_attn = torch.zeros(input_len_for_attention)
avg_output_attn = None
if i > 0 and output_attention_layers:
avg_output_attn = torch.mean(torch.stack(output_attention_layers).float(), dim=[0, 1])
elif i > 0:
avg_output_attn = torch.zeros(i)
# Normalize separately with epsilon for numerical stability
epsilon = 1e-8
input_sum = avg_input_attn.sum() + epsilon
normalized_input_attn = avg_input_attn / input_sum
normalized_output_attn = None
if i > 0 and avg_output_attn is not None:
output_sum = avg_output_attn.sum() + epsilon
normalized_output_attn = avg_output_attn / output_sum
attention_matrices.append({
'input_attention': normalized_input_attn.cpu(),
'output_attention': normalized_output_attn.cpu() if normalized_output_attn is not None else None,
'raw_input_attention': avg_input_attn.cpu(), # Keep raw for analysis
'raw_output_attention': avg_output_attn.cpu() if avg_output_attn is not None else None
})
# Fill remaining steps with zeros if needed
while len(attention_matrices) < output_len:
attention_matrices.append({
'input_attention': torch.zeros(input_len_for_attention),
'output_attention': None,
'raw_input_attention': torch.zeros(input_len_for_attention),
'raw_output_attention': None
})
return attention_matrices
@staticmethod
def process_attention_joint(
attention_data: Dict[str, Any],
input_tokens: List[str],
output_tokens: List[str]
) -> List[Dict[str, torch.Tensor]]:
"""
Process attention with joint normalization across input and output.
This preserves the relative importance across all tokens.
"""
attentions = attention_data['attentions']
input_len_for_attention = attention_data['input_len_for_attention']
output_len = attention_data['output_len']
if not attentions:
return [{'input_attention': torch.zeros(input_len_for_attention),
'output_attention': None} for _ in range(output_len)]
attention_matrices = []
num_steps = len(attentions)
if num_steps == 0:
print("Warning: No attention steps found in output.")
return [{'input_attention': torch.zeros(input_len_for_attention),
'output_attention': None} for _ in range(output_len)]
steps_to_process = min(num_steps, output_len)
for i in range(steps_to_process):
step_attentions = attentions[i]
input_attention_layers = []
output_attention_layers = []
for layer_idx, layer_attn in enumerate(step_attentions):
try:
# Extract attention to input tokens
input_indices = slice(1, 1 + input_len_for_attention)
if layer_attn.shape[3] >= input_indices.stop:
input_attn = layer_attn[0, :, 0, input_indices]
input_attention_layers.append(input_attn)
# Extract attention to previous output tokens
if i > 0:
output_indices = slice(1 + input_len_for_attention, 1 + input_len_for_attention + i)
if layer_attn.shape[3] >= output_indices.stop:
output_attn = layer_attn[0, :, 0, output_indices]
output_attention_layers.append(output_attn)
else:
output_attention_layers.append(
torch.zeros((layer_attn.shape[1], i), device=layer_attn.device)
)
else:
input_attention_layers.append(
torch.zeros((layer_attn.shape[1], input_len_for_attention), device=layer_attn.device)
)
if i > 0:
output_attention_layers.append(
torch.zeros((layer_attn.shape[1], i), device=layer_attn.device)
)
except Exception as e:
print(f"Error processing attention at step {i}, layer {layer_idx}: {e}")
input_attention_layers.append(
torch.zeros((layer_attn.shape[1], input_len_for_attention), device=layer_attn.device)
)
if i > 0:
output_attention_layers.append(
torch.zeros((layer_attn.shape[1], i), device=layer_attn.device)
)
# Average across layers and heads
if input_attention_layers:
avg_input_attn = torch.mean(torch.stack(input_attention_layers).float(), dim=[0, 1])
else:
avg_input_attn = torch.zeros(input_len_for_attention)
avg_output_attn = None
if i > 0 and output_attention_layers:
avg_output_attn = torch.mean(torch.stack(output_attention_layers).float(), dim=[0, 1])
elif i > 0:
avg_output_attn = torch.zeros(i)
# Joint normalization
epsilon = 1e-8
if i > 0 and avg_output_attn is not None:
# Concatenate and normalize together
combined_attn = torch.cat([avg_input_attn, avg_output_attn])
sum_attn = combined_attn.sum() + epsilon
normalized_combined = combined_attn / sum_attn
normalized_input_attn = normalized_combined[:input_len_for_attention]
normalized_output_attn = normalized_combined[input_len_for_attention:]
else:
# Only input attention available
sum_attn = avg_input_attn.sum() + epsilon
normalized_input_attn = avg_input_attn / sum_attn
normalized_output_attn = None
attention_matrices.append({
'input_attention': normalized_input_attn.cpu(),
'output_attention': normalized_output_attn.cpu() if normalized_output_attn is not None else None
})
# Fill remaining steps with zeros if needed
while len(attention_matrices) < output_len:
attention_matrices.append({
'input_attention': torch.zeros(input_len_for_attention),
'output_attention': None
})
return attention_matrices
@staticmethod
def extract_attention_for_step(
attention_data: Dict[str, Any],
step: int,
input_len: int
) -> Dict[str, torch.Tensor]:
"""
Extract attention weights for a specific generation step.
Optimized to only process the needed step.
"""
attentions = attention_data['attentions']
if step >= len(attentions):
return {
'input_attention': torch.zeros(input_len),
'output_attention': None
}
step_attentions = attentions[step]
input_attention_layers = []
output_attention_layers = []
for layer_attn in step_attentions:
# Extract input attention
input_indices = slice(1, 1 + input_len)
if layer_attn.shape[3] >= input_indices.stop:
input_attn = layer_attn[0, :, 0, input_indices]
input_attention_layers.append(input_attn)
# Extract output attention if there are previous outputs
if step > 0:
output_indices = slice(1 + input_len, 1 + input_len + step)
if layer_attn.shape[3] >= output_indices.stop:
output_attn = layer_attn[0, :, 0, output_indices]
output_attention_layers.append(output_attn)
# Average and normalize
if input_attention_layers:
avg_input = torch.mean(torch.stack(input_attention_layers).float(), dim=[0, 1])
normalized_input = avg_input / (avg_input.sum() + 1e-8)
else:
normalized_input = torch.zeros(input_len)
normalized_output = None
if step > 0 and output_attention_layers:
avg_output = torch.mean(torch.stack(output_attention_layers).float(), dim=[0, 1])
normalized_output = avg_output / (avg_output.sum() + 1e-8)
return {
'input_attention': normalized_input.cpu(),
'output_attention': normalized_output.cpu() if normalized_output is not None else None
} |