-
Notifications
You must be signed in to change notification settings - Fork 2
/
eval.py
285 lines (239 loc) · 9.08 KB
/
eval.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
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
import argparse
import torch
from tqdm import tqdm
from lm_eval.models.huggingface import HFLM
import lm_eval
import torch.nn as nn
from pathlib import Path
import json
from datautils import get_loaders
from transformers import AutoTokenizer,LlamaConfig,LlamaForCausalLM, AutoModelForCausalLM
from safetensors import safe_open
from utils import load_json, save_json
from qat.replace_module import replace_with_learnable_binarylinear
def _parse_eval_task(task_str):
optional_tasks = ['ppl', 'boolq', 'piqa', 'hellaswag', 'winogrande', 'arc_easy', 'arc_challenge', 'openbookqa', 'mmlu', 'storycloze_2016', 'storycloze', 'storycloze_2018']
tasks = [s.strip() for s in task_str.split(',')]
parsed_tasks = []
for task in tasks:
if task in optional_tasks:
parsed_tasks.append(task)
else:
print(f'Wrong task name: {task} in your input: {task_str}. The optional tasks are: {", ".join(optional_tasks)}')
return parsed_tasks
def load_open_src(model_name):
model = AutoModelForCausalLM.from_pretrained(model_name).to('cuda')
tokenizer = AutoTokenizer.from_pretrained(model_name, padding_side="right", use_fast=False)
return model, tokenizer
def load_ckpts(model_size, ckpt_dir, ckpt_type, exist_extra_para):
assert model_size in ["130M", "1.3B", "7B"]
assert ckpt_type in ['torch', 'hf_st', 'lightning']
ckpt_dir = Path(ckpt_dir)
with Path(f'FBI-LLM_configs/FBI-LLM_llama2_{model_size}.json').open('r') as r_f:
config = json.load(r_f)
llama_config = LlamaConfig(**config)
model = LlamaForCausalLM(llama_config).to('cuda')
tokenizer = AutoTokenizer.from_pretrained('huggyllama/llama-7b', padding_side="right", use_fast=False)
if exist_extra_para:
model = replace_with_learnable_binarylinear(model, scaling_pattern = "column", keep_parts = ["lm_head"])
weight_dict = {}
if ckpt_type == 'torch':
ckpt_plist = [p for p in ckpt_dir.iterdir() if p.suffix == '.bin']
for p in ckpt_plist:
_weight_dict = torch.load(p)
for k,v in _weight_dict.items():
if 'self_attn.rotary_emb.inv_freq' not in k:
weight_dict[k] = v
elif ckpt_type == 'lightning':
ckpt_plist = [p for p in (ckpt_dir/'fabric_ckpt').iterdir() if p.suffix == '.distcp']
pass # TODO: add lightning ckpt load
elif ckpt_type == 'hf_st':
ckpt_plist = [p for p in ckpt_dir.iterdir() if p.suffix == '.safetensors']
for p in ckpt_plist:
with safe_open(p, framework="pt", device="cpu") as f:
weight_dict.update({key: f.get_tensor(key) for key in f.keys()})
model.load_state_dict(weight_dict)
for param in model.parameters():
param.data = param.data.to(torch.float16)
return model, tokenizer
@torch.no_grad()
def evaluate_ckpt_task(model, tokenizer, tasks, num_fewshot, batch_size, max_length):
task_manager = lm_eval.tasks.TaskManager()
eval_lm = HFLM(model, tokenizer=tokenizer, batch_size=batch_size, max_length=max_length)
result = lm_eval.simple_evaluate(eval_lm, tasks = tasks, num_fewshot = num_fewshot, task_manager=task_manager)
result = result["results"]
print(result)
return result
@torch.no_grad()
def evaluate_ckpt_ppl(model, tokenizer, ppl_datasets, max_length, limit = -1):
results = {}
for dataset in ppl_datasets:
_, testloader = get_loaders(dataset, tokenizer)
testenc = testloader.input_ids
nsamples = testenc.numel() // max_length
use_cache = model.config.use_cache
model.config.use_cache = False
model.eval()
nlls = []
for i in tqdm(range(nsamples)):
batch = testenc[:, (i * max_length) : ((i + 1) * max_length)].to(model.device)
outputs = model.model(batch)
hidden_states = outputs[0] # .to(model.lm_head.weight.device)
logits = model.lm_head(hidden_states) # .contiguous()
shift_logits = logits[:, :-1, :] # .contiguous()
shift_labels = testenc[:, (i * max_length) : ((i + 1) * max_length)][:, 1:].to(model.device)
loss_fct = nn.CrossEntropyLoss()
loss = loss_fct(
shift_logits.view(-1, shift_logits.size(-1)),
shift_labels.view(-1),
)
neg_log_likelihood = loss.float() * max_length
nlls.append(neg_log_likelihood)
if i == limit:
break
ppl = torch.exp(torch.stack(nlls).sum() / (len(nlls) * max_length))
print(dataset, ppl.item())
model.config.use_cache = use_cache
results[dataset] = ppl.item()
return results
def eval_ckpt(ckpt_dir, args):
tasks = _parse_eval_task(args.task)
print('tasks', tasks)
eval_ppl = 'ppl' in tasks
down_stream_tasks = [t for t in tasks if t != 'ppl']
ckpt_dir = Path(ckpt_dir)
model, tokenizer = load_ckpts(
model_size = args.model_size,
ckpt_dir = ckpt_dir,
ckpt_type = args.ckpt_type,
exist_extra_para = args.exist_extra_para
)
# model = to_regular_linear(model)
res = {}
if eval_ppl:
ppl_res = evaluate_ckpt_ppl(
model = model,
tokenizer = tokenizer,
ppl_datasets = ['wikitext2', 'ptb', 'c4'],
max_length = 2048
)
res.update(ppl_res)
if len(down_stream_tasks) > 0:
task_res = evaluate_ckpt_task(
model = model,
tokenizer = tokenizer,
tasks = down_stream_tasks,
num_fewshot = 0,
batch_size = args.batch_size,
max_length = 2048
)
res.update(task_res)
return res
def evaluate_qat(args):
save_dir = Path('eval_result')
save_dir.mkdir(exist_ok=True, parents=True)
src_dir = Path(args.path)
ckpt_ids = [i.strip() for i in args.ckpt_ids.split(',')]
ckpt_ids = sorted(ckpt_ids, key=lambda x: int(x))
save_p = save_dir / f"{src_dir.name}_{'-'.join(ckpt_ids)}.json"
if save_p.exists():
result = load_json(save_p)
else:
result = {}
for cid in ckpt_ids:
ckpt_name = f'ckpt-{cid}'
print(ckpt_name)
ckpt_dir = src_dir / ckpt_name
res = eval_ckpt(ckpt_dir, args)
if ckpt_name not in result:
result[ckpt_name] = res
else:
result[ckpt_name].update(res)
save_json(result, save_p)
def evaluate_open_src(args):
save_dir = Path('eval_result')
save_dir.mkdir(exist_ok=True, parents=True)
tasks = _parse_eval_task(args.task)
print('tasks', tasks)
eval_ppl = 'ppl' in tasks
down_stream_tasks = [t for t in tasks if t != 'ppl']
save_p = save_dir / f"{'_'.join(args.path.split('/'))}.json"
if save_p.exists():
result = load_json(save_p)
else:
result = {}
model, tokenizer = load_open_src(args.path)
res = {}
if eval_ppl:
ppl_res = evaluate_ckpt_ppl(
model = model,
tokenizer = tokenizer,
ppl_datasets = ['wikitext2', 'ptb', 'c4'],
max_length = 2048
)
res.update(ppl_res)
if len(down_stream_tasks) > 0:
task_res = evaluate_ckpt_task(
model = model,
tokenizer = tokenizer,
tasks = down_stream_tasks,
num_fewshot = 0,
batch_size = args.batch_size,
max_length = 2048
)
res.update(task_res)
if '_'.join(args.path.split('/')) not in result:
result['_'.join(args.path.split('/'))] = res
else:
result['_'.join(args.path.split('/'))].update(res)
save_json(result, save_p)
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="Model Training Script")
parser.add_argument(
"--path",
type=str,
help="Saved model path",
)
parser.add_argument(
"--eval_open_src",
action="store_true",
help="If evaluating open source LLMs outside of FBI-LLMs, please specify this argument."
)
parser.add_argument(
"--task",
type=str,
default="ppl,boolq,piqa,hellaswag,winogrande,arc_easy,arc_challenge,openbookqa",
help="evaluate tasks",
)
parser.add_argument(
"--ckpt_ids",
type=str,
help='The checkpoints to evaluate'
)
parser.add_argument(
"--model_size",
type=str,
help="model size",
)
parser.add_argument(
"--ckpt_type",
type=str,
default='torch',
help="The saving type of checkpoints"
)
parser.add_argument(
"--exist_extra_para",
action="store_true",
help="Are there any additional parameters for the model to be evaluated. If evaluating FBI-LLM, please specify this argument. If evaluating other open source LLMs, do not specify this argument"
)
parser.add_argument(
"--batch_size",
type=int,
default=16,
help="batch_size for evaluation"
)
args = parser.parse_args()
if args.eval_open_src:
evaluate_open_src(args)
else:
evaluate_qat(args)