forked from avnermay/Sequoia
-
Notifications
You must be signed in to change notification settings - Fork 0
/
tree_search.py
232 lines (201 loc) · 8.59 KB
/
tree_search.py
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
import argparse
from copy import deepcopy
import json
import matplotlib.pyplot as plt
import networkx as nx
import numpy as np
import torch
from tqdm import tqdm
def get_args():
parser = argparse.ArgumentParser()
parser.add_argument('--output_path', type=str, default='/work/avner/results/sequoia/growmaps/growmaps')
parser.add_argument('--acceptance_rates', type=str, default=None,
help='Json file or pytorch file with acceptance rates as function of tree branch width')
parser.add_argument('--target_model_speeds', type=str, default=None,
help='Json file with target model speeds as function of # of tokens')
parser.add_argument('--draft_model_speeds', type=str, default=None,
help='Json file with draft model speeds as function of # of tokens (only tokens=1 is used)')
parser.add_argument('--max_depth', type=int, default=16)
args = parser.parse_args()
return args
def dynamic_program(alpha, max_budget, max_depth, max_branch):
T = torch.zeros((max_budget + 1, max_depth + 1, max_branch + 1)).fill_(-torch.inf)
branch_map = {}
for l in range(1, max_depth + 1):
for b in range(0, max_branch + 1):
if b == 0:
T[1, l, b] = 1.0
branch_map[(1,l,b)] = []
for m in tqdm(range(2, max_budget + 1)):
for l in range(2, max_depth + 1):
T[m, l, 1] = 1 + alpha[1] * T[m-1, l-1].max()
if T[m, l, 1] > 0:
branch_map[(m,l,1)] = [(m-1, l-1, T[m-1, l-1].argmax(dim=0).item())]
for b in range(2, max_branch + 1):
max_value = -torch.inf
for y in range(1, m):
new_value = T[y, l, b-1] + alpha[b] * T[m-y, l-1].max()
if new_value > max_value:
max_value = new_value
new_y = y
max_value = max(max_value, new_value)
T[m, l, b] = max_value
if max_value >= 0:
new_branch = T[m-new_y, l-1].argmax(dim=0).item()
new_list :list = deepcopy(branch_map[(new_y, l, b-1)])
new_list.append((m-new_y, l-1, new_branch))
branch_map[(m,l,b)] = new_list
results = T.max(dim=2).values
return results, T, branch_map
def get_optimal_tree_size_and_depth(draft_inference_time, target_verify_time, valid_budget, results):
best_speedup = 0.0
best_tree_size = 1
best_tree_depth = 1
best_acc_length = 1
c = draft_inference_time / target_verify_time[0]
t = target_verify_time / target_verify_time[0]
for i, b in enumerate(valid_budget):
for d, ac_len in enumerate(results[b]):
if ac_len < 0:
continue
x = ac_len / (c * d + t[i])
if x > best_speedup:
best_speedup = x
best_tree_size, best_tree_depth = b, d
best_acc_length = ac_len
best_time_per_token = target_verify_time[0] / best_speedup
print(f'{best_tree_size=}, {best_tree_depth=}, {best_time_per_token=}, {best_acc_length=}, {best_speedup=}')
return best_tree_size, best_tree_depth
def get_grow_map(branch_map, tree_size, tree_depth, max_branches):
m, l, b = tree_size, tree_depth, max_branches
positions = [0]
states = [(m,l,b)]
active = [True]
depth = [0]
Successors = [[]]
attention_mask = torch.zeros(m,m).long()
parents = [-1]
expand_lists = []
expand_branches = []
num_nodes = 1
while True:
expand = []
expand_branch = []
for i, act in enumerate(active):
if act:
if parents[i] != -1:
attention_mask[i] = attention_mask[parents[i]]
attention_mask[i, i] = 1
expand.append(i)
active[i] = False
(x,y,z) = states[i]
expand_branch.append(z)
positions.extend(list(range(num_nodes, num_nodes + z)))
Successors[i].extend(list(range(num_nodes, num_nodes + z)))
Successors.extend([[] for _ in range(z)])
parents.extend([i for _ in range(z)])
depth.extend([depth[i] + 1 for _ in range(z)])
states.extend(branch_map[(x,y,z)])
assert len(branch_map[(x,y,z)]) == z
num_nodes = num_nodes + z
if len(expand) == 0:
break
expand_lists.append(expand)
expand_branches.append(expand_branch)
active.extend([True for _ in range(sum(expand_branch))])
assert num_nodes == m
assert len(positions) == m
assert len(depth) == m
grow_map = {
'roots': expand_lists,
'branches': expand_branches,
'Successors':Successors,
'mask': attention_mask,
'depth': torch.LongTensor(depth),
'size': num_nodes
}
return grow_map
class Node:
def __init__(self):
self.children = []
self.path = []
self.name = ''
def create_tree(all_nodes):
for i, node in enumerate(all_nodes):
node.name = str(i)
tree = {}
for node in all_nodes:
tree[node.name] = [n.name for n in node.children]
return tree
def convert_expand_branches(expand_branches):
root = Node()
all_nodes = [root]
curr_layer_nodes = [root] # root node
for d, layer_child_nums in enumerate(expand_branches):
next_layer_nodes = []
for node, num_children in zip(curr_layer_nodes, layer_child_nums):
for i in range(num_children):
child = Node()
child.path = node.path + [i]
node.children.append(child)
all_nodes.append(child)
next_layer_nodes.append(child)
curr_layer_nodes = next_layer_nodes
return all_nodes
def convert_eagle_format(eagle_format):
root = Node()
all_nodes = {str([]): root}
for node_path in eagle_format:
parent = all_nodes[str(node_path[:-1])]
child = Node()
child.path=node_path
parent.children.append(child)
all_nodes[str(node_path)] = child
return [node for node in all_nodes.values()]
def plot_tree(tree):
# Convert the tree structure to a directed graph
G = nx.DiGraph(tree)
# Compute the hierarchical layout for the tree
pos = nx.drawing.nx_pydot.graphviz_layout(G, prog='dot')
# Draw the tree
plt.figure(figsize=(8, 6))
nx.draw(G, pos, with_labels=False, node_size=50, node_color='lightblue', font_size=12, font_weight='bold')
plt.show()
if __name__ == '__main__':
args = get_args()
# alpha = get_alpha(args.eval_results_json)
if args.acceptance_rates.endswith('.json'):
with open(args.acceptance_rates, 'r') as f:
acc_rate_dict = json.load(f)
alpha = np.array(acc_rate_dict['acceptance_rates'])
else:
alpha = torch.load(args.acceptance_rates).numpy()
with open(args.target_model_speeds, 'r') as f:
target_model_speeds_dict = json.load(f)
target_verify_time = np.array(target_model_speeds_dict['avg_forward_pass_times'])
valid_budget = target_model_speeds_dict['decode_lengths']
with open(args.draft_model_speeds, 'r') as f:
draft_model_speeds_dict = json.load(f)
assert draft_model_speeds_dict['decode_lengths'][0] == 1, (
'We need the time for the target model to process 1 token')
draft_inference_time = draft_model_speeds_dict['avg_forward_pass_times'][0]
max_depth = args.max_depth
max_budget = valid_budget[-1]
max_branch = alpha.shape[0] - 1
results, T, branch_map = dynamic_program(alpha, max_budget, max_depth, max_branch)
best_tree_size, best_tree_depth = get_optimal_tree_size_and_depth(draft_inference_time, target_verify_time, valid_budget, results)
max_branches = T[best_tree_size, best_tree_depth].argmax(dim=0).item()
grow_map = get_grow_map(branch_map, best_tree_size, best_tree_depth, max_branches)
# Create the tree representation string in the same format as the `mc_sim_7b_63` eagle tree.
all_nodes = convert_expand_branches(grow_map['branches'])
full_str = '[' + ', '.join([str(n.path) for n in all_nodes[1:]]) + ']'
print(full_str)
# Code for plotting the tree
# tree = create_tree(all_nodes)
# plot_tree(tree)
# For convenience, we save as both a pytorch file and as a json file.
torch.save(grow_map, args.output_path + '.pt')
with open(args.output_path + '.json', 'w') as f:
grow_map['mask'] = grow_map['mask'].numpy().tolist()
grow_map['depth'] = grow_map['depth'].numpy().tolist()
json.dump(grow_map, f)