-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathD2-qwen.py
177 lines (134 loc) · 7.47 KB
/
D2-qwen.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
import argparse
import os
import torch
import torch.backends.cudnn as cudnn
from config import cfg, process_args
from dataset import make_dataset, make_data_loader, process_dataset, make_batchnorm_stats, make_calibration_dataloader
from metric import make_metric, make_logger
from model import make_prune_model
from module import to_device, process_control, makedir_exist_ok, check_dense_model, save_calib_info
from deepspeed.profiling.flops_profiler import FlopsProfiler
from utils import run_lm_eval, ppl_eval_sharing
from transformers import AutoTokenizer, AutoModelForCausalLM
from tqdm import tqdm
from model.merge_qwen import Merge_QwenMoE
cudnn.benchmark = False
parser = argparse.ArgumentParser(description='cfg')
parser.add_argument('--base_model_path', type=str, default="Qwen/Qwen2-57B-A14B", help='Path to base model')
parser.add_argument('--expert_freq_path', type=str, default="cache/QwenMoE_wikitext_20000_expert_frequencies.json", help='Path to expert frequencies')
parser.add_argument('--fisher_path', type=str, default="Model/fisher_QwenMoE.pt", help='Path to fisher info')
parser.add_argument('--svd_scale_path', type=str, default="Model/SVD_scale_QwenMoE_0-31_512.pt", help='Path to svd scale')
parser.add_argument('--result_path', type=str, default="result", help='Path to result')
parser.add_argument("--pp_ratio", type=float, default=0.2)
parser.add_argument("--delta_ratio", type=float, default=1)
parser.add_argument("--share_ratio", type=float, default=1)
parser.add_argument("--share_V", action='store_true', default=False)
parser.add_argument("--share_U", action='store_true', default=False)
parser.add_argument("--merge_method", type=str, default="fisher")
for k in cfg:
exec('parser.add_argument(\'--{0}\', default=cfg[\'{0}\'], type=type(cfg[\'{0}\']))'.format(k))
parser.add_argument('--control_name', default=None, type=str)
parser.add_argument('--output_dir', default=None, type=str)
args = vars(parser.parse_args())
process_args(args)
def main():
process_control()
seeds = list(range(cfg['init_seed'], cfg['init_seed'] + cfg['num_experiments']))
for i in range(cfg['num_experiments']):
model_tag_list = [str(seeds[i]), cfg['control_name']]
cfg['model_tag'] = '_'.join([x for x in model_tag_list if x])
runExperiment()
return
def runExperiment():
cfg['seed'] = int(cfg['model_tag'].split('_')[0])
cfg['prune_ratio'] = args['pp_ratio']
torch.manual_seed(cfg['seed'])
torch.cuda.manual_seed(cfg['seed'])
result_path = os.path.join('output', 'result')
makedir_exist_ok(result_path)
if check_dense_model() is None:
pass
cfg['epoch'] = 0
cfg['data_name'] = 'wikitext'
dataset = make_dataset(cfg['data_name'], cfg['subset_name'])
cfg['model_name'] = 'mixtral'
cfg['skip_layers'] = []
cfg['test_stage'] = False
cfg['no_probe_process'] = False
cfg['merge_model'] = True
cfg['shared_infer'] = False
def merge_model(base_model_path, expert_freq_path, svd_scale_path, fisher_path, delta_ratio, share_ratio, share_V, share_U, merge_method):
import json
with open(expert_freq_path, 'r') as f:
expert_freq = json.load(f)
svd_scale = torch.load(svd_scale_path, map_location='cpu')
fisher_info = torch.load(fisher_path, map_location="cpu")
model = AutoModelForCausalLM.from_pretrained(base_model_path,
device_map="auto",
trust_remote_code=True,
torch_dtype=torch.bfloat16)
for i in tqdm(range(len(model.model.layers)), desc="Merging layers"):
try:
Merge_MoE_Block = Merge_QwenMoE(model.config, share_ratio=share_ratio,
delta_ratio=delta_ratio, expert_freq=expert_freq[str(i)],
delta_share_V=share_V, delta_share_U=share_U,
merge_method=merge_method, shared_infer=cfg['shared_infer']).to(model.model.layers[i].mlp.gate.weight.device)
Merge_MoE_Block.merge_experts(model.model.layers[i].mlp, svd_scale=svd_scale[i], hessian = fisher_info[i], scale_type='svdllm')
model.model.layers[i].mlp = Merge_MoE_Block
except ValueError as e:
print(f"Warning: SVD failed for layer {i}, skipping this layer")
continue
return model
model = merge_model(args['base_model_path'], args['expert_freq_path'], args['svd_scale_path'], args['fisher_path'],
args['delta_ratio'], args['share_ratio'], args['share_V'], args['share_U'], args['merge_method'])
tokenizer = AutoTokenizer.from_pretrained(args['base_model_path'], use_fast=False)
if tokenizer.pad_token is None:
tokenizer.pad_token = tokenizer.eos_token
# -----------------------------------------------------------------------------------------
cfg['tokenizer'] = tokenizer
cfg['model_name'] = 'llama-2-7b'
cfg['model_type'] = 'qwen'
test_logger = make_logger(os.path.join('output', 'runs', 'test_{}'.format(cfg['model_tag'])))
dataset = process_dataset(dataset, tokenizer)
if cfg['model_name'] in ['cnn', 'resnet18', 'wresnet28x2']:
model = make_batchnorm_stats(dataset['train'], model, cfg['model_name'])
model = make_prune_model(model)
if cfg['merge_model']:
for i in range(len(model.model.model.layers)):
model.model.model.layers[i].mlp.update_Wmean()
if 'calib' in cfg['prune_method']:
print('Running Calibration ...', flush=True)
cfg['calibration_stage'] = True
cfg['calibration_dataset'] = 'wikitest'
calibration_data_loader = make_calibration_dataloader(tokenizer)
run_calibration(model, calibration_data_loader['train'])
save_calib_info(model)
if 'flapratio' in cfg['prune_method']:
from model import HiddenRepresentationPruning
pruning_module = HiddenRepresentationPruning(cfg, 'flapratio')
pruning_module.flap_ratio(model, test_logger)
cfg['calibration_stage'] = False
print('Calibration Done...', flush=True)
save_dir = f"{args['result_path']}/qwen_delta-{args['delta_ratio']}-pp_ratio-{args['pp_ratio']}-shareV-{args['share_V']}"
os.makedirs(save_dir, exist_ok=True)
result_str = ppl_eval_sharing(model, tokenizer, experiment_name=f"D2-qwen", datasets=['wikitext2', 'ptb', 'c4'], params_only=False, batch_size=8)
with open(f"{save_dir}/ppl_eval_sharing.txt", "w") as f:
f.write(result_str)
run_lm_eval(model, tokenizer, batch_size=8, task_names=["openbookqa", "arc_easy", "winogrande", "hellaswag",
"arc_challenge", "piqa", "mathqa"], output_dir=save_dir)
return
def run_calibration(model, data_loader):
with torch.no_grad():
model.eval()
for i, input in enumerate(data_loader):
# now, the wikitext and c4 datsets used for calibration are clm tasks
# input_size = input['labels'].size(0)
input = {'input_ids': input['input_ids'], 'attention_mask': input['attention_mask'],
'labels': input['labels']}
input = to_device(input, "cuda")
output = model(**input)
# input_ = {'target': input['labels']}
# output_ = {'target': output['logits'], 'loss': output['loss']}
return
if __name__ == "__main__":
main()