-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathpred.py
210 lines (196 loc) · 10.1 KB
/
pred.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
import os
from datasets import load_dataset
import torch
import json
from transformers import AutoTokenizer, LlamaTokenizer, AutoModelForCausalLM, AutoConfig
from tqdm import tqdm
import numpy as np
import random
import argparse
import torch.distributed as dist
import torch.multiprocessing as mp
from utils_hh.modify_llama import convert_kvcache_llama_heavy_recent, LlamaAttention_heavy_hitter
import copy
from utils_hh.modeling_llama import LlamaForCausalLM
ENABLE_Heavy_Hitter_FUNCTIONS = {
"llama": convert_kvcache_llama_heavy_recent,
}
def parse_args(args=None):
parser = argparse.ArgumentParser()
parser.add_argument('--model', type=str, default=None, choices=["llama2-7b-chat-4k", "longchat-v1.5-7b-32k", "llama2-13b-chat-4k", "xgen-7b-8k", "internlm-7b-8k", "chatglm2-6b", "chatglm2-6b-32k", "chatglm3-6b-32k", "vicuna-v1.5-7b-16k"])
parser.add_argument('--e', action='store_true', help="Evaluate on LongBench-E")
parser.add_argument('--save', type=str, default=None, help='Path to save results.')
parser.add_argument("--heavy_ratio", type=float, default=0.1)
parser.add_argument("--recent_ratio", type=float, default=0.1)
parser.add_argument("--apval", action='store_true', help="apply value norm")
parser.add_argument("--h2o", action='store_true', help="If true, use the setting of h2o, otherwise ")
parser.add_argument("--sink_len", type=int, help="attention sink length")
return parser.parse_args(args)
# This is the customized building prompt for chat models
def build_chat(tokenizer, prompt, model_name):
if "chatglm3" in model_name:
prompt = tokenizer.build_chat_input(prompt)
elif "chatglm" in model_name:
prompt = tokenizer.build_prompt(prompt)
elif "longchat" in model_name or "vicuna" in 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()
elif "llama2" in model_name:
prompt = f"[INST]{prompt}[/INST]"
elif "xgen" in model_name:
header = (
"A chat between a curious human and an artificial intelligence assistant. "
"The assistant gives helpful, detailed, and polite answers to the human's questions.\n\n"
)
prompt = header + f" ### Human: {prompt}\n###"
elif "internlm" in model_name:
prompt = f"<|User|>:{prompt}<eoh>\n<|Bot|>:"
return prompt
def post_process(response, model_name):
if "xgen" in model_name:
response = response.strip().replace("Assistant:", "")
elif "internlm" in model_name:
response = response.split("<eoa>")[0]
return response
def get_pred(rank, world_size, data, max_length, max_gen, prompt_format, dataset, device, model_name, model2path, out_path, apval,sink_len,heavy_ratio,h2o):
device = torch.device(f'cuda:{rank}')
model, tokenizer = load_model_and_tokenizer(model2path[model_name], model_name, device='cpu')
config = AutoConfig.from_pretrained(model2path[model_name])
# config.heavy_ratio = 0.5
# config.recent_ratio = 0.25
config.heavy_ratio = heavy_ratio
config.recent_ratio = heavy_ratio
config.sink_len=sink_len
config.layer_ratio=[0.4 for _ in range(32)]
config.applyval=apval
config.h2o=h2o
if heavy_ratio< 1.0:
checkpoint = copy.deepcopy(model.state_dict())
model = convert_kvcache_llama_heavy_recent(model, config)
model.load_state_dict(checkpoint)
model.half().eval().cuda()
del checkpoint
torch.cuda.empty_cache()
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 "chatglm3" in model_name:
tokenized_prompt = tokenizer(prompt, truncation=False, return_tensors="pt", add_special_tokens=False).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)
if "chatglm3" in model_name:
input = prompt.to(device)
else:
input = tokenizer(prompt, truncation=False, return_tensors="pt").to(device)
context_length = input.input_ids.shape[-1]
if dataset == "samsum": # prevent illegal output on samsum (model endlessly repeat "\nDialogue"), might be a prompting issue
output = model.generate(
**input,
max_new_tokens=max_gen,
num_beams=1,
do_sample=False,
temperature=1.0,
# past_key_values=cache,
min_length=context_length+1,
eos_token_id=[tokenizer.eos_token_id, tokenizer.encode("\n", add_special_tokens=False)[-1]],
)[0]
else:
output = model.generate(
**input,
max_new_tokens=max_gen,
num_beams=1,
do_sample=False,
temperature=1.0,
)[0]
pred = tokenizer.decode(output[context_length:], skip_special_tokens=True)
pred = post_process(pred, model_name)
for name, m in model.named_modules():
if isinstance(m, LlamaAttention_heavy_hitter):
m._reset_masks()
# torch.cuda.empty_cache()
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')
dist.destroy_process_group()
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):
if "chatglm" in model_name or "internlm" in model_name or "xgen" in model_name:
tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(path, trust_remote_code=True, torch_dtype=torch.bfloat16).to(device)
elif "llama2" in model_name or "vicuna" in model_name:
# replace_llama_attn_with_flash_attn()
tokenizer = LlamaTokenizer.from_pretrained(path)
model = LlamaForCausalLM.from_pretrained(path, torch_dtype=torch.float16).to(device)
model = model.eval()
return model, tokenizer
if __name__ == '__main__':
seed_everything(42)
args = parse_args()
world_size = torch.cuda.device_count()
mp.set_start_method('spawn', force=True)
model2path = json.load(open("config/model2path.json", "r"))
model2maxlen = json.load(open("config/model2maxlen.json", "r"))
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model_name = args.model
# define your 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"]
# datasets = [ "gov_report", "multi_news","trec","lcc"]
# datasets = [ "qasper"]
# datasets=["qasper", "multifieldqa_en", "hotpotqa", "2wikimqa"]
# datasets = [ "qasper", "lcc", "multifieldqa_en", "2wikimqa" ]
# datasets = ["hotpotqa", \
# "trec", "triviaqa", "samsum", "passage_count", "passage_retrieval_en", "repobench-p"]
# datasets = [ "gov_report","qasper", "lcc", "multifieldqa_en", "2wikimqa","multi_news"]
else:
datasets=["narrativeqa","musique","qmsum"]
# 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(f"{args.save}"):
os.makedirs(f"{args.save}")
for dataset in datasets:
if args.e:
data = load_dataset('THUDM/LongBench', f"{dataset}_e", split='test')
if not os.path.exists(f"{args.save}/{model_name}"):
os.makedirs(f"{args.save}/{model_name}")
out_path = f"{args.save}/{model_name}/{dataset}.jsonl"
else:
data = load_dataset('THUDM/LongBench', dataset, split='test')
if not os.path.exists(f"{args.save}/{model_name}"):
os.makedirs(f"{args.save}/{model_name}")
out_path = f"{args.save}/{model_name}/{dataset}.jsonl"
prompt_format = dataset2prompt[dataset]
max_gen = dataset2maxlen[dataset]
data_all = [data_sample for data_sample in data]
data_subsets = [data_all[i::world_size] for i in range(world_size)]
processes = []
for rank in range(world_size):
p = mp.Process(target=get_pred, args=(rank, world_size, data_subsets[rank], max_length, \
max_gen, prompt_format, dataset, device, model_name, model2path, out_path, args.apval, args.sink_len, args.heavy_ratio, args.h2o))
p.start()
processes.append(p)
for p in processes:
p.join()