-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathdemo_APE.py
166 lines (147 loc) · 7.82 KB
/
demo_APE.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
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
import numpy as np
import random
import argparse
def parse_args(args=None):
parser = argparse.ArgumentParser()
parser.add_argument('--model', type=str, default=None, choices=["llama3-8b-instruct", "llama3.1-8b-instruct", "mistral-7b-instruct-v0.3", "gemma2-9b-it"])
parser.add_argument("--temperature", type=float, default=1.0)
parser.add_argument("--scale", type=float, default=1.0)
return parser.parse_args(args)
def load_model_and_tokenizer(model_name, device):
if model_name == "llama3-8b-instruct":
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3-8B-Instruct")
model = AutoModelForCausalLM.from_pretrained("meta-llama/Meta-Llama-3-8B-Instruct", torch_dtype=torch.bfloat16).to(device)
elif model_name == "llama3.1-8b-instruct":
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")
model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.1-8B-Instruct", torch_dtype=torch.bfloat16).to(device)
elif model_name == "mistral-7b-instruct-v0.3":
tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-Instruct-v0.3")
model = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-Instruct-v0.3", torch_dtype=torch.bfloat16).to(device)
elif model_name == "gemma2-9b-it":
tokenizer = AutoTokenizer.from_pretrained("google/gemma-2-9b-it")
model = AutoModelForCausalLM.from_pretrained("google/gemma-2-9b-it", torch_dtype=torch.bfloat16).to(device)
return tokenizer, model
def build_prefix(model_name, prompt):
if "llama" in model_name:
prompt = f"<|begin_of_text|>\n<|start_header_id|>user<|end_header_id|>\n{prompt}"
elif "mistral" in model_name:
prompt = f"<s>[INST]{prompt}"
elif "gemma" in model_name:
prompt = f"<bos><start_of_turn>user\n{prompt}"
return prompt
def build_suffix(model_name, prompt):
if "llama" in model_name:
prompt = f"{prompt}\n<|eot_id|>\n<|start_header_id|>assistant<|end_header_id|>"
elif "mistral" in model_name:
prompt = f"{prompt}[/INST]"
elif "gemma" in model_name:
prompt = f"{prompt}<end_of_turn>\n<start_of_turn>model\n"
return prompt
def enable_attention_prefill_prefix(model_name, model):
if "llama" in args.model:
from src.ape_llama import enable_llama_attention_prefill_prefix
enable_llama_attention_prefill_prefix(model)
elif "mistral" in model_name:
from src.ape_mistral import enable_mistral_attention_prefill_prefix
enable_mistral_attention_prefill_prefix(model)
elif "gemma" in model_name:
from src.ape_gemma import enable_gemma_attention_prefill_prefix
enable_gemma_attention_prefill_prefix(model)
def enable_attention_prefill_context(model_name, model):
if "llama" in args.model:
from src.ape_llama import enable_llama_attention_prefill_context
enable_llama_attention_prefill_context(model)
elif "mistral" in model_name:
from src.ape_mistral import enable_mistral_attention_prefill_context
enable_mistral_attention_prefill_context(model)
elif "gemma" in model_name:
from src.ape_gemma import enable_gemma_attention_prefill_context
enable_gemma_attention_prefill_context(model)
def enable_attention_prefill_query(model_name, model):
if "llama" in args.model:
from src.ape_llama import enable_llama_attention_prefill_query
enable_llama_attention_prefill_query(model)
elif "mistral" in model_name:
from src.ape_mistral import enable_mistral_attention_prefill_query
enable_mistral_attention_prefill_query(model)
elif "gemma" in model_name:
from src.ape_gemma import enable_gemma_attention_prefill_query
enable_gemma_attention_prefill_query(model)
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 generate(args):
prefix = ""
contexts = [
"My friends and I enjoy eating out at restaurants together. However, we also enjoy cooking and making food as a group as well."
"Many of my friends like to play soccer and volleyball. We also enjoy watching movies and going to museums and galleries.",
]
query = "Question: what are three ideas for a social with a large groups of friends in New York City.\nAnswer:"
device = torch.device(f'cuda:0')
tokenizer, model = load_model_and_tokenizer(args.model, device)
model = model.eval()
tokenizer.pad_token_id = tokenizer.eos_token_id
prefix = build_prefix(args.model, prefix)
query = build_suffix(args.model, query)
with torch.no_grad():
prefix_input_ids = tokenizer(prefix, truncation=False, return_tensors="pt", add_special_tokens=False).input_ids
query_input_ids = tokenizer(query, truncation=False, return_tensors="pt").input_ids
len_prefix = prefix_input_ids.shape[1]
len_query = query_input_ids.shape[1]
context_input_ids = tokenizer(contexts, return_tensors='pt', truncation=True, max_length=8192-len_prefix-len_query-256, padding=True, add_special_tokens=False).input_ids
context_mask = (context_input_ids != tokenizer.pad_token_id).reshape(-1)
enable_attention_prefill_prefix(args.model, model)
past_key_values = None
outputs = model(
prefix_input_ids.to(model.device),
past_key_values=past_key_values,
use_cache=True,
)
past_key_values = []
for past_key_value in outputs.past_key_values:
bsz, _ = context_input_ids.shape
past_key = past_key_value[0].repeat(bsz, 1, 1, 1)
past_value = past_key_value[1].repeat(bsz, 1, 1, 1)
past_position = past_key_value[2]
past_key_values.append((past_key, past_value, past_position))
enable_attention_prefill_context(args.model, model)
outputs = model(
context_input_ids.to(model.device),
past_key_values=past_key_values,
use_cache=True,
)
past_key_values = []
for past_key_value in outputs.past_key_values:
bsz, num_heads, seq_len, _ = past_key_value[0].size()
past_key = torch.cat([past_key_value[0][:1, :, :len_prefix, :],
past_key_value[0][:, :, len_prefix:, :].transpose(1, 2).flatten(0, 1)[context_mask].unsqueeze(0).transpose(1, 2)], dim=2)
past_value = torch.cat([past_key_value[1][:1, :, :len_prefix, :],
past_key_value[1][:, :, len_prefix:, :].transpose(1, 2).flatten(0, 1)[context_mask].unsqueeze(0).transpose(1, 2)], dim=2)
past_position = torch.cat([past_key_value[2][:, :len_prefix],
past_key_value[2][:, len_prefix:].repeat(bsz, 1).flatten()[context_mask].unsqueeze(0)], dim=1)
past_key_values.append((past_key, past_value, past_position, len(contexts)))
context_input_ids = context_input_ids.flatten()[context_mask].unsqueeze(0)
input_ids = torch.cat([prefix_input_ids, context_input_ids, query_input_ids], dim=-1)
context_length = input_ids.shape[-1]
enable_attention_prefill_query(args.model, model, args.temperature, args.scale)
output = model.generate(
input_ids=input_ids.to(model.device),
max_new_tokens=256,
num_beams=1,
do_sample=False,
temperature=1.0,
past_key_values=past_key_values,
)[0]
pred = tokenizer.decode(output[context_length:], skip_special_tokens=True)
print(pred)
if __name__ == '__main__':
args = parse_args()
seed_everything(42)
generate(args)