-
Notifications
You must be signed in to change notification settings - Fork 1
/
eval_needle.py
260 lines (238 loc) · 9.37 KB
/
eval_needle.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
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
import argparse
import gc
import sys
import torch
from transformers import AutoTokenizer
from transformers import LlamaForCausalLM
from tqdm import tqdm
from accelerate import Accelerator
import glob
import numpy as np
from tqdm import tqdm
import gc
import matplotlib.pyplot as plt
from matplotlib.colors import LinearSegmentedColormap
import seaborn as sns
import pandas as pd
import random
from easy_context import (
prepare_seq_parallel_inputs,
apply_seq_parallel_monkey_patch,
)
apply_seq_parallel_monkey_patch("zigzag_ring_attn", "llama")
SEED = 24242424
torch.manual_seed(SEED)
random.seed(SEED)
np.random.seed(SEED)
NEEDLE_FORMAT = "\nThe special magic Singapore number is: {}.\n"
PREFIX = "This is a very long story book: <book>"
QUESTION_STR = "</book>.\n Based on the content of the book, Question: What is the special magic Singapore number? Answer: The special magic Singapore number is:"
def eval_forward(accelerator, model, input_ids, pad_id, answer_ids):
# first append labels to input_ids
prompt_length = input_ids.shape[1]
labels_length = answer_ids.shape[1]
input_ids = torch.cat([input_ids, answer_ids], dim=1)
# second pad input_ids to the multiple of accelerator.num_processes
pad_tensor = torch.tensor(
[pad_id]
* (
(accelerator.num_processes * 2)
- input_ids.shape[1] % (accelerator.num_processes * 2)
)
).unsqueeze(0)
input_ids = torch.cat([input_ids, pad_tensor], dim=1)
position_ids = (
torch.arange(input_ids.shape[1]).unsqueeze(0).expand(input_ids.shape[0], -1)
)
prepared = prepare_seq_parallel_inputs(
"zigzag_ring_attn",
input_ids,
position_ids,
None,
accelerator.process_index,
accelerator.num_processes,
accelerator.device,
)
local_input_ids = prepared["local_input_ids"]
local_position_ids = prepared["local_position_ids"]
with torch.inference_mode():
logits = model(
local_input_ids,
position_ids=local_position_ids,
use_cache=False,
).logits
pred = logits.argmax(dim=-1)
# gather all logits using accelerator.gather
def undo_extract_local(gathered_value, world_size, dim=1):
value_chunks = gathered_value.chunk(2 * world_size, dim=dim)
reordered_chunks = [None] * (2 * world_size)
for i in range(world_size):
reordered_chunks[i] = value_chunks[i * 2]
reordered_chunks[2 * world_size - i - 1] = value_chunks[i * 2 + 1]
return torch.cat(reordered_chunks, dim=dim)
correct = False
gathered_logits = accelerator.gather(pred.squeeze(0)).unsqueeze(0)
# undo extract local on the gathered logits
pred = undo_extract_local(gathered_logits, accelerator.num_processes)
pred = pred[:, prompt_length - 1 : prompt_length + labels_length - 1]
# check if the logits are correct, extract argmax id
# compare the predicted_ids with the labels
correct = (pred == answer_ids.to(accelerator.device)).all()
return int(correct)
def load_haystack(args, accelerator, tokenizer):
context = ""
# do not count <s>
while len(tokenizer.encode(context)) - 1 < args.max_context_length:
accelerator.print(f"Current Context Length: {len(tokenizer.encode(context))-1}")
accelerator.print(glob.glob(f"{args.haystack_dir}/*.txt"))
for file in glob.glob(f"{args.haystack_dir}/*.txt"):
with open(file, "r") as f:
accelerator.print(f"Reading {file}")
context += f.read()
if len(tokenizer.encode(context)) - 1 > args.max_context_length:
break
tokenized_haystack = tokenizer.encode(context)
return tokenized_haystack
def construct_prompt(
tokenized_haystack,
tokenized_prefix,
tokenized_postfix,
tokenized_needle,
context_length,
depth,
):
# insert the needle into depth of the haystack
prompt = tokenized_haystack[:context_length]
if depth == 0:
start_index = 0
else:
start_index = int(len(prompt) * depth)
period_tokens = [29889, 869]
# find the closest period token
for i in range(start_index, len(prompt)):
if prompt[i] in period_tokens:
start_index = i + 1
break
prompt = prompt[:start_index] + tokenized_needle + prompt[start_index:]
prompt = tokenized_prefix + prompt + tokenized_postfix
# from transformers import AutoTokenizer
# tk = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")
# tk.decode(prompt)
return prompt
def main(args):
model = args.model
tokenizer = AutoTokenizer.from_pretrained(
args.model,
model_max_length=sys.maxsize,
trust_remote_code=True,
add_bos_token=True,
)
tokenizer.pad_token = tokenizer.eos_token
accelerator = Accelerator(
mixed_precision="bf16",
)
kwargs = {"rope_theta": args.rope_theta} if args.rope_theta is not None else {}
model = LlamaForCausalLM.from_pretrained(
args.model,
torch_dtype=torch.bfloat16,
_attn_implementation="flash_attention_2",
device_map=accelerator.device,
**kwargs,
)
model = accelerator.prepare(model)
model.gradient_checkpointing_enable()
tokenizer.pad_token = tokenizer.eos_token
# remember to remove <s>
accelerator.print("Preparing Haystack...")
tokenized_haystack = load_haystack(args, accelerator, tokenizer)[1:]
tokenized_prefix = tokenizer.encode(PREFIX)
accelerator.print("Starting Evaluation...")
random_number_list = [
int(np.random.randint(10**args.rnd_number_digits))
for i in range(args.num_samples)
]
print(random_number_list)
all_accuries = []
for context_length in tqdm(
range(
args.min_context_length, args.max_context_length + 1, args.context_interval
)
):
for depth in np.arange(0, 1 + args.depth_interval, args.depth_interval):
accuracies = []
for random_number in random_number_list:
needle_str = NEEDLE_FORMAT.format(random_number)
question_str = QUESTION_STR
asnwer_str = str(random_number)
tokenized_needle = tokenizer.encode(needle_str)[1:]
tokenized_postfix = tokenizer.encode(question_str)[1:]
tokenizer_answer = tokenizer.encode(asnwer_str)[1:]
prompt = construct_prompt(
tokenized_haystack,
tokenized_prefix,
tokenized_postfix,
tokenized_needle,
context_length,
depth,
)
input_ids = torch.tensor([prompt])
answer_ids = torch.tensor([tokenizer_answer])
correct = eval_forward(
accelerator, model, input_ids, tokenizer.pad_token_id, answer_ids
)
gc.collect()
torch.cuda.empty_cache()
if accelerator.is_main_process:
accuracies.append(correct)
if accelerator.is_main_process:
result = {
"Context Length": context_length,
"Document Depth": round(depth * 100, -1),
"Score": sum(accuracies) / len(accuracies),
}
accelerator.print(result)
all_accuries.append(result)
if accelerator.is_main_process:
df = pd.DataFrame(all_accuries)
cmap = LinearSegmentedColormap.from_list(
"custom_cmap", ["#F0496E", "#EBB839", "#0CD79F"]
)
pivot_table = pd.pivot_table(
df,
values="Score",
index=["Document Depth", "Context Length"],
aggfunc="mean",
).reset_index() # This will aggregate
pivot_table = pivot_table.pivot(
index="Document Depth", columns="Context Length", values="Score"
)
# Create the heatmap with better aesthetics
plt.figure(figsize=(17.5, 8)) # Can adjust these dimensions as needed
sns.heatmap(
pivot_table,
# annot=True,
fmt="g",
cmap=cmap,
cbar_kws={"label": "Score"},
)
# More aesthetics
plt.xlabel("Token Limit") # X-axis label
plt.ylabel("Depth Percent") # Y-axis label
plt.xticks(rotation=45) # Rotates the x-axis labels to prevent overlap
plt.yticks(rotation=0) # Ensures the y-axis labels are horizontal
plt.tight_layout() # Fits everything neatly into the figure area
# save
model_name = args.model.split("/")[-1]
plt.savefig(f"data/heatmap_{model_name}.png".format(model_name))
if __name__ == "__main__":
args = argparse.ArgumentParser()
args.add_argument("--model", type=str, default="PY007/Llama2-7B-64K")
args.add_argument("--max_context_length", type=int, default=100000)
args.add_argument("--min_context_length", type=int, default=1000)
args.add_argument("--context_interval", type=int, default=1000)
args.add_argument("--depth_interval", type=float, default=0.1)
args.add_argument("--num_samples", type=int, default=10)
args.add_argument("--rope_theta", type=float, default=None)
args.add_argument("--rnd_number_digits", type=int, default=7)
args.add_argument("--haystack_dir", type=str, required=True)
main(args.parse_args())