-
Notifications
You must be signed in to change notification settings - Fork 1
/
dist_latency_inference_on_euler.py
159 lines (129 loc) · 7.24 KB
/
dist_latency_inference_on_euler.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
import argparse
from pipeline_parallel.dist_pp_utils import get_pp_inference_module
from utils.dist_args_utils import *
from utils.dist_inference_utils import *
from comm.comm_utils import *
from coordinator.lsf.lsf_coordinate_client_deprecated import CoordinatorInferenceClient
from coordinator.lsf.lsf_job_scheduler import alias_to_model_name
def sync_setting(args, pipeline, device, return_msg=None):
num_return_sequences_tensor = torch.zeros(1, dtype=torch.int64, device=device)
generate_token_length_tensor = torch.zeros(1, dtype=torch.int64, device=device)
temperature_tensor = torch.zeros(1, dtype=torch.float32, device=device)
top_p_tensor = torch.zeros(1, dtype=torch.float32, device=device)
do_sample_tensor = torch.zeros(1, dtype=torch.uint8, device=device)
if get_pipeline_parallel_rank() == 0:
generate_token_length = return_msg['task_api']['parameters']['max_new_tokens']
do_sample = return_msg['task_api']['parameters']['do_sample']
temperature = return_msg['task_api']['parameters']['temperature']
top_p = return_msg['task_api']['parameters']['top_p']
num_return_sequences = return_msg['task_api']['parameters']['num_return_sequences']
num_return_sequences_tensor[:] = num_return_sequences
generate_token_length_tensor[:] = generate_token_length
temperature_tensor[:] = temperature
top_p_tensor[:] = top_p
do_sample_tensor[:] = do_sample
pipeline.comm.broadcast(num_return_sequences_tensor, src=0)
pipeline.num_completions = num_return_sequences_tensor.item()
pipeline.comm.broadcast(generate_token_length_tensor, src=0)
pipeline.generate_seq_length = generate_token_length_tensor.item()
pipeline.comm.broadcast(temperature_tensor, src=0)
args.temperature = temperature_tensor.item()
pipeline.comm.broadcast(top_p_tensor, src=0)
args.top_p = top_p_tensor.item()
pipeline.comm.broadcast(do_sample_tensor, src=0)
if do_sample_tensor.item() == 0:
args.temperature = 0
pipeline.change_buffer_size()
def main():
parser = argparse.ArgumentParser(description='Inference Runner with coordinator.')
add_device_arguments(parser)
add_torch_distributed_inference_w_euler_coordinator_arguments(parser)
add_inference_arguments(parser)
add_inference_details_arguments(parser)
add_global_coordinator_arguments(parser)
parser.add_argument('--seed', type=int, default=1, metavar='S',
help='random seed (default: 1)')
parser.add_argument('--profiling', type=str, default='tidy_profiling', metavar='S',
help='enable which profiling? default: tidy mode')
parser.add_argument('--trace-postfix', type=str, default='default', metavar='S',
help='postfix of the tracing file name.')
args = parser.parse_args()
print_arguments(args)
torch.manual_seed(args.seed)
if args.use_cuda:
assert (torch.cuda.is_available())
device = torch.device('cuda', args.cuda_id)
else:
device = torch.device('cpu')
model_name_abbr = args.model_name.split('/')[-1]
print("model name abbr: ", model_name_abbr)
print("model name: ", alias_to_model_name(model_name_abbr))
coord_client = CoordinatorInferenceClient(args, alias_to_model_name(model_name_abbr))
prime_ip, rank, port = coord_client.notify_inference_join()
print("<====Coordinator assigned prime-IP:", prime_ip, " and my assigned rank", rank, "====>")
init_inference_communicators_with_coordinator(args, prime_ip, rank, port=port)
pipeline = get_pp_inference_module(args, device, rank=rank)
tokenizer = get_tokenizer(args)
tokenizer.model_max_length = args.input_seq_length
input_ids = torch.ones([args.batch_size, args.input_seq_length]).long().cuda()
attention_mask = torch.ones([args.batch_size, args.input_seq_length]).long().cuda()
coord_client.notify_inference_heartbeat()
last_timestamp = time.time()
while True:
if get_pipeline_parallel_rank() == 0:
current_timestamp = time.time()
if current_timestamp - last_timestamp >= args.heartbeats_timelimit:
coord_client.notify_inference_heartbeat()
last_timestamp = current_timestamp
return_msg = coord_client.load_input_job_from_dfs()
# print("<<<<<<<<<<<<<<Return_msg Dict>>>>>>>>>>>>")
# print(return_msg)
if return_msg is not None:
print(f"Handel request: <{return_msg['_id']}>")
sync_setting(args, pipeline, device, return_msg)
pipeline.update_processors(args)
#####
inputs = tokenizer(return_msg['task_api']['inputs'], return_tensors='pt',
padding='max_length', truncation=True, )
input_ids = inputs['input_ids'].long().to(device)
attention_mask = inputs['attention_mask'].long().to(device)
pipeline.comm.broadcast(input_ids, src=0)
pipeline.comm.broadcast(attention_mask, src=0)
output_ids_list = []
pipeline.inference_batch(input_ids, output_ids_list, attention_mask=attention_mask)
return_full_text = return_msg['task_api']['parameters']['return_full_text']
results = []
for i in range(pipeline.num_completions):
token_len = torch.zeros([1], dtype=torch.int64).cuda()
pipeline.comm.recv(token_len, src=pipeline.pipeline_group_size - 1)
result = torch.empty((1, token_len.item()), dtype=torch.long).cuda()
pipeline.comm.recv(result, src=pipeline.pipeline_group_size - 1)
if return_full_text:
results.append(return_msg['task_api']['inputs'] + tokenizer.decode(result[0]))
else:
results.append(tokenizer.decode(result[0]))
return_msg['task_api']['outputs'] = results
coord_client.save_output_job_to_dfs(return_msg)
elif get_pipeline_parallel_rank() == pipeline.pipeline_group_size - 1:
while True:
sync_setting(args, pipeline, device)
pipeline.update_processors(args)
pipeline.comm.broadcast(input_ids, src=0)
pipeline.comm.broadcast(attention_mask, src=0)
output_ids_list = []
pipeline.inference_batch(input_ids, output_ids_list, attention_mask=attention_mask)
for i in range(pipeline.num_completions):
result = output_ids_list[i]
token_len = torch.tensor(result['token_ids'].size(1), dtype=torch.long).cuda()
pipeline.comm.send(token_len, dst=0)
pipeline.comm.send(result['token_ids'].cuda(), dst=0)
# s.send((json.dumps(output_ids_list)).encode())
else:
while True:
sync_setting(args, pipeline, device)
pipeline.update_processors(args)
pipeline.comm.broadcast(input_ids, src=0)
pipeline.comm.broadcast(attention_mask, src=0)
pipeline.inference_batch(input_ids, attention_mask=attention_mask)
if __name__ == '__main__':
main()