-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathpred_longstreaming.py
182 lines (151 loc) · 7.23 KB
/
pred_longstreaming.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
import os
from datasets import load_dataset
import torch
import json
from transformers import AutoTokenizer, LlamaTokenizer, LlamaForCausalLM, AutoModelForCausalLM
from tqdm import tqdm
import numpy as np
import random
import argparse
from longstreaming.stm_utils import load, download_url, load_jsonl
from longstreaming.stm_enable_streaming_llm import enable_streaming_llm
from functools import partial
@torch.no_grad()
def greedy_generate(model, tokenizer, input_ids, past_key_values, max_gen_len):
outputs = model(
input_ids=input_ids,
past_key_values=past_key_values,
use_cache=True,
)
past_key_values = outputs.past_key_values
pred_token_idx = outputs.logits[:, -1, :].argmax(dim=-1).unsqueeze(1)
generated_ids = [pred_token_idx.item()]
generated_text = []
pos = 0
for _ in range(max_gen_len - 1):
outputs = model(
input_ids=pred_token_idx,
past_key_values=past_key_values,
use_cache=True,
)
past_key_values = outputs.past_key_values
pred_token_idx = outputs.logits[:, -1, :].argmax(dim=-1).unsqueeze(1)
generated_ids.append(pred_token_idx.item())
decoded_text = tokenizer.decode(
generated_ids[-1:],
skip_special_tokens=True,
clean_up_tokenization_spaces=True,
spaces_between_special_tokens=False,
).strip()
generated_text.append(decoded_text)
now = len(generated_text) - 1
if now > pos:
pos = now
if pred_token_idx == tokenizer.eos_token_id:
break
final_generated_text = " ".join(generated_text).strip()
return past_key_values, final_generated_text
@torch.no_grad()
def streaming_inference(model, tokenizer, prompts, kv_cache=None, max_gen_len=1000):
all_generated_texts = []
past_key_values = None
for idx, prompt in enumerate(prompts):
input_ids = tokenizer(prompt, return_tensors="pt").input_ids
input_ids = input_ids.to(model.device)
seq_len = input_ids.shape[1]
if kv_cache is not None:
space_needed = seq_len + max_gen_len
past_key_values = kv_cache.evict_for_space(past_key_values, space_needed)
past_key_values, generated_text = greedy_generate(
model, tokenizer, input_ids, past_key_values, max_gen_len=max_gen_len
)
all_generated_texts.append(generated_text)
return all_generated_texts
def parse_args(args=None):
parser = argparse.ArgumentParser()
parser.add_argument('--model', type=str, default="vicuna-v1.5-7b-ls")
parser.add_argument('--e', action='store_true', help="Evaluate on LongBench-E")
return parser.parse_args(args)
# This is the customized building prompt for chat models
def build_chat(tokenizer, prompt, model_name):
from fastchat.model import get_conversation_template
conv = get_conversation_template("vicuna")
conv.append_message(conv.roles[0], prompt)
conv.append_message(conv.roles[1], None)
prompt = conv.get_prompt()
return prompt
def get_pred(data, max_length, max_gen, prompt_format, dataset, device, model_name, model2path, out_path):
model, tokenizer = load_model_and_tokenizer(model2path[model_name], model_name, device)
kv_cache = None
for json_obj in tqdm(data):
prompt = prompt_format.format(**json_obj)
# truncate to fit max_length (we suggest truncate in the middle, since the left and right side may contain crucial instructions)
tokenized_prompt = tokenizer(prompt, truncation=False, return_tensors="pt").input_ids[0]
if len(tokenized_prompt) > max_length:
half = int(max_length/2)
prompt = tokenizer.decode(tokenized_prompt[:half], skip_special_tokens=True)+tokenizer.decode(tokenized_prompt[-half:], skip_special_tokens=True)
if dataset not in ["trec", "triviaqa", "samsum", "lsht", "lcc", "repobench-p"]: # chat models are better off without build prompts on these tasks
prompt = build_chat(tokenizer, prompt, model_name)
prompts = [prompt]
kv_cache = enable_streaming_llm(model, start_size=4, recent_size=2000)
generated_texts = streaming_inference(model, tokenizer, prompts, kv_cache=kv_cache, max_gen_len=max_gen)
pred = generated_texts[0]
with open(out_path, "a", encoding="utf-8") as f:
json.dump({"pred": pred, "answers": json_obj["answers"], "all_classes": json_obj["all_classes"], "length": json_obj["length"]}, f, ensure_ascii=False)
f.write('\n')
def seed_everything(seed):
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
np.random.seed(seed)
random.seed(seed)
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
torch.cuda.manual_seed_all(seed)
def load_model_and_tokenizer(path, model_name, device):
model = AutoModelForCausalLM.from_pretrained(path,torch_dtype=torch.bfloat16).to(device)
tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True, use_fast=False)
model = model.eval()
return model, tokenizer
def load_local_dataset(path):
data = []
with open(path, 'r') as f:
for line in f:
data.append(json.loads(line))
return data
if __name__ == '__main__':
seed_everything(42)
args = parse_args()
model2path = json.load(open("config/model2path.json", "r"))
model2maxlen = json.load(open("config/model2maxlen.json", "r"))
device = torch.device('cuda:0')
model_name = args.model
max_length = model2maxlen[model_name]
if args.e:
datasets = ["qasper", "multifieldqa_en", "hotpotqa", "2wikimqa", "gov_report", "multi_news", \
"trec", "triviaqa", "samsum", "passage_count", "passage_retrieval_en", "lcc", "repobench-p"]
else:
datasets = ["narrativeqa", "qasper", "multifieldqa_en", "multifieldqa_zh", "hotpotqa", "2wikimqa", "musique", \
"dureader", "gov_report", "qmsum", "multi_news", "vcsum", "trec", "triviaqa", "samsum", "lsht", \
"passage_count", "passage_retrieval_en", "passage_retrieval_zh", "lcc", "repobench-p"]
# we design specific prompt format and max generation length for each task, feel free to modify them to optimize model output
dataset2prompt = json.load(open("config/dataset2prompt.json", "r"))
dataset2maxlen = json.load(open("config/dataset2maxlen.json", "r"))
# predict on each dataset
if not os.path.exists("pred"):
os.makedirs("pred")
if not os.path.exists("pred_e"):
os.makedirs("pred_e")
for dataset in datasets:
if args.e:
data = load_dataset('THUDM/LongBench', f"{dataset}_e", split='test')
if not os.path.exists(f"pred_e/{model_name}"):
os.makedirs(f"pred_e/{model_name}")
out_path = f"pred_e/{model_name}/{dataset}.jsonl"
else:
data = load_dataset('THUDM/LongBench', dataset, split='test')
if not os.path.exists(f"pred/{model_name}"):
os.makedirs(f"pred/{model_name}")
out_path = f"pred/{model_name}/{dataset}.jsonl"
prompt_format = dataset2prompt[dataset]
max_gen = dataset2maxlen[dataset]
get_pred(data, max_length, max_gen, prompt_format, dataset, device, model_name, model2path, out_path)