File size: 10,738 Bytes
b3a3b15 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 |
import os
import yaml
import pandas as pd
import matplotlib.pyplot as plt
import numpy as np
MODELS = [
"facebook/opt-125m", # 1
"facebook/opt-350m", # 2
"facebook/opt-1.3b", # 3
"facebook/opt-2.7b", # 4
"facebook/opt-6.7b", # 5
"facebook/opt-13b", # 6
"facebook/opt-30b", # 7
"facebook/opt-66b", # 8
"SparseLLM/ReluLLaMA-7B", # 9
"SparseLLM/prosparse-llama-2-7b", # 10
"meta-llama/Llama-2-7b-hf", # 11
"meta-llama/Llama-2-13b-hf", # 12
"-",
"meta-llama/Llama-3.1-8B", # 14
"meta-llama/Llama-3.1-70B", # 15
]
CONFIGS = {
'facebook/opt-175b':{
'num_layer': 95,
'd':12288,
'h': 96,
'neurons': 49152
},
'facebook/opt-66b':{
'num_layer': 64,
'd':9216,
'h': 72,
'neurons': 36864,
'layer_imp': "results/attn_importance/opt-66b_attn_importance.json"
},
'facebook/opt-30b':{
'num_layer': 48,
'd':7168,
'h': 56,
'neurons': 28672,
'layer_imp': "results/attn_importance/opt-30b_attn_importance.json"
},
'facebook/opt-13b':{
'num_layer': 40,
'd':5120,
'h': 40,
'neurons': 20480,
'layer_imp': "results/attn_importance/opt-13b_attn_importance.json"
},
'facebook/opt-6.7b':{
'num_layer': 32,
'd':4096,
'h': 32,
'neurons': 16384,
'layer_imp': "results/attn_importance/opt-6.7b_attn_importance.json"
},
'facebook/opt-2.7b':{
'num_layer': 32,
'd':2560,
'h': 32,
'neurons': 10240
},
'facebook/opt-1.3b':{
'num_layer': 24,
'd':2048,
'h': 32,
'neurons': 8192
},
'facebook/opt-350m':{
'num_layer': 24,
'd':1024,
'h': 16,
'neurons': 4096
},
'facebook/opt-125m':{
'num_layer': 12,
'd':768,
'h': 12,
'neurons': 3072
},
'SparseLLM/ReluLLaMA-7B':{
'num_layer': 32,
'd': 4096,
'h': 32,
'neurons': 11008
},
'SparseLLM/prosparse-llama-2-7b':{
'num_layer': 32,
'd': 4096,
'h': 32,
'neurons': 11008
},
'meta-llama/Llama-2-7b-hf':{
'num_layer': 32,
'd': 4096,
'h': 32,
'neurons': 11008,
'layer_imp': "results/attn_importance/Llama-2-7b-hf_attn_importance.json"
},
'meta-llama/Llama-2-13b-hf':{
'num_layer': 40,
'd': 5120,
'h': 40,
'neurons': 13824,
'layer_imp': "results/attn_importance/Llama-2-13b-hf_attn_importance.json"
},
'meta-llama/Llama-3.1-8B':{
'num_layer': 32,
'd': 4096,
'h': 32,
'neurons': 14336,
'layer_imp': "results/attn_importance/Llama-3.1-8B_attn_importance.json"
},
'meta-llama/Llama-3.1-70B':{
'num_layer': 80,
'd': 8192,
'h': 64,
'neurons': 28672,
'layer_imp': "results/attn_importance/Llama-3.1-70B_attn_importance.json"
}
}
# This class is used to store the activation thresholds for each layer of the model. Used for evaluation purposes.
class ActivationThresholds:
def __init__(self, num_layers, attn_th=0.0, mlp_th = 0.0):
self.activation_threshold = {}
self.mlp_threshold = {}
for i in range(num_layers):
self.activation_threshold[i] = attn_th
self.mlp_threshold[i] = mlp_th
def set_threshold(self, layer_idx, threshold):
self.activation_threshold[layer_idx] = threshold
def get_threshold(self, layer_idx):
return self.activation_threshold[layer_idx]
def save_thresholds(self, file_path):
with open(file_path, 'w') as file:
documents = yaml.dump(self.activation_threshold, file)
# def load_thresholds(self, file_path):
# with open(file_path, 'r') as file:
# self.activation_threshold = yaml.load(file, Loader=yaml.FullLoader)
def load_thresholds(self, sparsity_map):
"""
Load the activation thresholds from a given sparsity map.
Parameters:
sparsity_map (dict): A dictionary containing layer indices and their corresponding activation thresholds.
"""
for layer_idx, threshold in sparsity_map.items():
self.activation_threshold[layer_idx] = threshold
@classmethod
def from_file(cls, file_path):
"""Class method to create an instance from a YAML file."""
with open(file_path, 'r') as file:
activation_threshold = yaml.load(file, Loader=yaml.FullLoader)
# Create an instance and set its activation_threshold
instance = cls()
instance.activation_threshold = activation_threshold
return instance
def build_mlp_topk_lookup(data_path: str, batch_size: int, delta: int = 128) -> dict:
"""
Creates a lookup table from the CSV file 'mlp_act_batch_{batch_size}_stats.csv' in the given directory.
The lookup maps each layer to a top-k value calculated as:
top_k = ceil((average_activation + std_activation + delta))
Parameters:
data_path (str): The path to the directory containing the CSV files.
batch_size (int): The batch size to use in the file name and for filtering.
delta (float): The additional value added to the activations before rounding.
Returns:
dict: A mapping from layer id to computed top-k value.
Raises:
FileNotFoundError: If the CSV file does not exist in the provided directory.
"""
file_name = f"mlp_act_batch_{batch_size}_stats.csv"
full_path = os.path.join(data_path, file_name)
if not os.path.exists(full_path):
raise FileNotFoundError(f"File not found: {full_path}")
# Read the CSV file into a DataFrame
df = pd.read_csv(full_path)
def calc_top_k(avg, std, delta):
raw_value = avg + std + delta
# return int(np.ceil(raw_value /128) * 128) # Round up to the nearest multiple of 128
return int(np.ceil(raw_value))
mlp_lookup = {
row["layer"]: calc_top_k(row["average_activation"], row["std_activation"], delta)
for _, row in df.iterrows() if row["batch_size"] == batch_size
}
return mlp_lookup
def _update_hf_mlp_topk(model, mlp_lookup):
"""
Updates the top-k values in the model's MLP layers using the provided lookup table.
Parameters:
model (HybridModel): The model to update.
mlp_lookup (dict): The lookup table mapping layer id to top-k value.
delta (int): The additional value added to the activations before rounding.
"""
for layer_idx, top_k in mlp_lookup.items():
model.model.decoder.layers[layer_idx].mlp_act= int(top_k)
def identify_model_type(model_name):
"""
Identifies if the given model name is an OPT model or a Llama model.
Args:
model_name (str): The name of the model.
Returns:
str: "OPT" if the model is an OPT model, "Llama" if it's a Llama model, or "Unknown".
"""
if "opt" in model_name.lower():
return "OPT"
elif "llama" in model_name.lower():
return "Llama"
else:
return "Unknown"
OPT_MODELS = [
"facebook/opt-125m", # 1
"facebook/opt-350m", # 2
"facebook/opt-1.3b", # 3
"facebook/opt-2.7b", # 4
"facebook/opt-6.7b", # 5
"facebook/opt-13b", # 6
"facebook/opt-30b", # 7
"facebook/opt-66b" # 8
]
OPT_CONFIGS = {
'facebook/opt-175b':{
'num_layer': 95,
'sp_config': None,
'd':12288,
'h': 96,
},
'facebook/opt-66b':{
'num_layer': 64,
'd':9216,
'h': 72,
},
'facebook/opt-30b':{
'num_layer': 48,
'd':7168,
'h': 56,
},
'facebook/opt-13b':{
'num_layer': 40,
'd':5120,
'h': 40,
},
'facebook/opt-6.7b':{
'num_layer': 32,
'd':4096,
'h': 32,
},
'facebook/opt-2.7b':{
'num_layer': 32,
'd':2560,
'h': 32,
},
'facebook/opt-1.3b':{
'num_layer': 24,
'd':2048,
'h': 32,
},
'facebook/opt-350m':{
'num_layer': 24,
'd':1024,
'h': 16,
},
'facebook/opt-125m':{
'num_layer': 12,
'd':768,
'h': 12,
},
}
'''
import seaborn as sns
def plot_average_activation(directory_path, model_name):
# Read all the .csv files in the specified directory and concatenate them into a single DataFrame
files = [os.path.join(directory_path, f) for f in os.listdir(directory_path) if f.endswith('.csv')]
df = pd.concat([pd.read_csv(f) for f in files])
# Set the seaborn style for better aesthetics
sns.set(style="whitegrid")
total_neurons = OPT_CONFIGS[model_name]["d"] * 4
# Compute the average activation percentage and standard deviation percentage
df['average_activation_percentage'] = (df['average_activation'] / total_neurons) * 100
df['std_activation_percentage'] = (df['std_activation'] / total_neurons) * 100
# Create a color palette with a different color for each batch size
palette = sns.color_palette("husl", df['batch_size'].nunique())
# Initialize the matplotlib figure
plt.figure(figsize=(12, 8))
# Loop over each batch size and plot the average_activation_percentage with error bars
for i, (batch_size, group) in enumerate(df.groupby('batch_size')):
plt.errorbar(
group['layer'],
group['average_activation_percentage'],
yerr=group['std_activation_percentage'],
label=f'Batch Size {batch_size}',
capsize=3,
marker='o',
linestyle='-',
color=palette[i]
)
# Set y-axis ticks at every 10% increment
plt.yticks(range(0, 101, 10)) # Y-ticks from 0% to 100% in steps of 10%
# Shaded region
plt.axhspan(80, plt.ylim()[1], facecolor='gray', alpha=0.1)
# Set the labels and title of the plot
plt.xlabel('Layer', fontsize=14)
plt.ylabel('Average Activation Percentage (%)', fontsize=14)
plt.title(f'Model: {model_name} Average Activation Percentage vs Layer for Different Batch Sizes', fontsize=16)
# Show legend
plt.legend(title='Batch Size', fontsize=12, title_fontsize=12)
# Tight layout for better spacing
plt.tight_layout()
# Save the image
plt.savefig('average_activation_analysis.png')
# Display the plot
plt.show()
'''
|