-
Notifications
You must be signed in to change notification settings - Fork 0
/
utils.py
272 lines (219 loc) · 12.5 KB
/
utils.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
import argparse
import pprint
import torch
import random
import numpy as np
import os
from datetime import datetime
import logging
from accelerate import dispatch_model, infer_auto_device_map
from accelerate.utils import get_balanced_memory
supported_models = [
'meta-llama/Llama-2-7b-hf',
'meta-llama/Llama-2-13b-hf',
'meta-llama/Llama-2-70b-hf',
'meta-llama/Meta-Llama-3-8B',
'meta-llama/Meta-Llama-3-70B',
'facebook/opt-125m'
]
supported_datasets = ['wikitext2', 'ptb', 'c4']
# These flags disable using TensorFloat-32 tensor cores (to avoid numerical issues)
torch.backends.cuda.matmul.allow_tf32 = False
torch.backends.cudnn.allow_tf32 = False
DEV = torch.device('cuda:0') if torch.cuda.is_available() else torch.device('cpu')
def llama_down_proj_groupsize(model, groupsize):
assert groupsize > 1, 'groupsize should be greater than 1!'
if model.config.intermediate_size % groupsize == 0:
logging.info(f'(Act.) Groupsiz = Down_proj Groupsize: {groupsize}')
return groupsize
group_num = int(model.config.hidden_size/groupsize)
assert groupsize*group_num == model.config.hidden_size, 'Invalid groupsize for llama!'
down_proj_groupsize = model.config.intermediate_size//group_num
assert down_proj_groupsize*group_num == model.config.intermediate_size, 'Invalid groupsize for down_proj!'
logging.info(f'(Act.) Groupsize: {groupsize}, Down_proj Groupsize: {down_proj_groupsize}')
return down_proj_groupsize
def set_seed(seed):
np.random.seed(seed)
torch.random.manual_seed(seed)
random.seed(seed)
# Dump the log both to console and a log file.
def config_logging(log_file, level=logging.INFO):
class LogFormatter(logging.Formatter):
def format(self, record):
if record.levelno == logging.INFO:
self._style._fmt = "%(message)s"
else:
self._style._fmt = "%(levelname)s: %(message)s"
return super().format(record)
console_handler = logging.StreamHandler()
console_handler.setFormatter(LogFormatter())
file_handler = logging.FileHandler(log_file)
file_handler.setFormatter(LogFormatter())
logging.basicConfig(level=level, handlers=[console_handler, file_handler])
def parser_gen():
parser = argparse.ArgumentParser()
# General Arguments
parser.add_argument('--model', type=str, default='meta-llama/Llama-2-7b-hf',
help='Model to load;', choices=supported_models)
parser.add_argument('--seed', type=int, default=0, help='Random Seed for HuggingFace and PyTorch')
parser.add_argument('--eval_dataset', type=str, default='wikitext2',
help='Dataset for Evaluation (default: wikitext2)', choices=supported_datasets,)
parser.add_argument('--hf_token', type=str, default=None)
parser.add_argument('--bsz', type=int, default=32,
help='Batch-size for PPL evaluation (default:32)')
# Rotation Arguments
parser.add_argument('--rotate', action=argparse.BooleanOptionalAction, default=False,
help='''Rotate the moodel. This will include online rotation for down-projection and
out-projection. Note that this does not apply rotation to the K/Q and they will be rotated
if we want to quantize the Keys''')
parser.add_argument('--rotate_mode', type=str, default='hadamard', choices=['hadamard', 'random'])
parser.add_argument('--rotation_seed', type=int, default=-1,
help='Random Seed for generating random matrix!!')
parser.add_argument('--fp32_had', action=argparse.BooleanOptionalAction, default=False,
help='Apply Hadamard rotation in FP32 (default: False)')
# Activation Quantization Arguments
parser.add_argument('--a_bits', type=int, default=16,
help='''Number of bits for inputs of the Linear layers. This will be
for all the linear layers in the model (including down-projection and out-projection)''')
parser.add_argument('--a_groupsize', type=int, default=-1,
help='Groupsize for activation quantization. Note that this should be the same as w_groupsize')
parser.add_argument('--a_asym', action=argparse.BooleanOptionalAction, default=False,
help='ASymmetric Activation quantization (default: False)')
parser.add_argument('--a_clip_ratio', type=float, default=1.0,
help='Clip ratio for activation quantization. new_max = max * clip_ratio')
# Weight Quantization Arguments
parser.add_argument('--w_bits', type=int, default=16,
help='Number of bits for weights of the Linear layers')
parser.add_argument('--w_groupsize', type=int, default=-1,
help='Groupsize for weight quantization. Note that this should be the same as a_groupsize')
parser.add_argument('--w_asym', action=argparse.BooleanOptionalAction, default=False,
help='ASymmetric weight quantization (default: False)')
parser.add_argument('--w_rtn', action=argparse.BooleanOptionalAction, default=False,
help='Quantize the weights using RtN. If the w_bits < 16 and this flag is not set, we use GPTQ')
parser.add_argument('--w_clip', action=argparse.BooleanOptionalAction, default=False,
help='''Clipping the weight quantization!
We do not support arguments for clipping and we find the best clip ratio during the weight quantization''')
parser.add_argument('--nsamples', type=int, default=128,
help='Number of calibration data samples for GPTQ.')
parser.add_argument('--cal_dataset', type=str, default='wikitext2',
help='calibration data samples for GPTQ.', choices=supported_datasets)
parser.add_argument('--percdamp', type=float, default=.01,
help='Percent of the average Hessian diagonal to use for dampening.')
parser.add_argument('--act_order', action=argparse.BooleanOptionalAction, default=False,
help='act-order in GPTQ')
# General Quantization Arguments
parser.add_argument('--int8_down_proj', action=argparse.BooleanOptionalAction, default=False,
help='Use INT8 for Down Projection! If this set, both weights and activations of this layer will be in INT8')
# KV-Cache Quantization Arguments
parser.add_argument('--v_bits', type=int, default=16,
help='''Number of bits for V-cache quantization.
Note that quantizing the V-cache does not need any other rotation''')
parser.add_argument('--v_groupsize', type=int, default=-1)
parser.add_argument('--v_asym', action=argparse.BooleanOptionalAction, default=False,
help='ASymmetric V-cache quantization')
parser.add_argument('--v_clip_ratio', type=float, default=1.0,
help='Clip ratio for v-cache quantization. new_max = max * clip_ratio')
parser.add_argument('--k_bits', type=int, default=16,
help='''Number of bits for K-cache quantization.
Note that quantizing the K-cache needs another rotation for the keys/queries''')
parser.add_argument('--k_groupsize', type=int, default=-1)
parser.add_argument('--k_asym', action=argparse.BooleanOptionalAction, default=False,
help='ASymmetric K-cache quantization')
parser.add_argument('--k_pre_rope', action=argparse.BooleanOptionalAction, default=False,
help='Pre-RoPE quantization for K-cache (not Supported yet!)')
parser.add_argument('--k_clip_ratio', type=float, default=1.0,
help='Clip ratio for k-cache quantization. new_max = max * clip_ratio')
# Save/Load Quantized Model Arguments
parser.add_argument('--load_qmodel_path', type=str, default=None,
help='Load the quantized model from the specified path!')
parser.add_argument('--save_qmodel_path', type=str, default=None,
help='Save the quantized model to the specified path!')
# WandB Arguments
parser.add_argument('--wandb', action=argparse.BooleanOptionalAction, default=False)
parser.add_argument('--wandb_id', type=str, default=None)
parser.add_argument('--wandb_project', type=str, default=None)
#Experiments Arguments
parser.add_argument('--save_name', type=str, default=None, help='The path to save experiment data, '
'including quantized models, dumped layer inputs, etc. The data will be saved in experiments/[model]/save_name. Default: [datetime].')
parser.add_argument('--capture_layer_io', action=argparse.BooleanOptionalAction, default=False,
help='Capture the input and output of the specified decoder layer and dump into a file')
parser.add_argument('--layer_idx', type=int, default=10, help='Which decoder layer to capture')
# LM Eval Arguments
parser.add_argument("--lm_eval", action="store_true", help="Evaluate the model on LM Eval tasks.")
parser.add_argument(
'--tasks',
nargs='+',
default=["piqa", "hellaswag", "arc_easy", "arc_challenge", "winogrande", "lambada"],
)
parser.add_argument('--lm_eval_batch_size', type=int, default=128, help='Batch size for evaluating with lm eval harness.')
parser.add_argument(
"--distribute",
action="store_true",
help="Distribute the model on multiple GPUs for evaluation.",
)
args = parser.parse_args()
if args.lm_eval:
from lm_eval import tasks
from lm_eval import utils as lm_eval_utils
from lm_eval.tasks import initialize_tasks
initialize_tasks()
for task in args.tasks:
if task not in lm_eval_utils.MultiChoice(tasks.ALL_TASKS):
raise ValueError(f"Invalid task: {task}")
# quant_type = f'w{args.w_bits}a{args.a_bits}_{args.rotate_mode}'
if args.save_name is None:
args.save_name = datetime.now().strftime("%Y%m%d_%H%M%S")
setattr(args, 'save_path',
os.path.join(os.path.dirname(os.path.abspath(__file__)), 'experiments', args.model, args.save_name))
os.makedirs(args.save_path, exist_ok=True)
config_logging(os.path.join(args.save_path, f'{args.save_name}.log'))
assert args.a_groupsize == args.w_groupsize, 'a_groupsize should be the same as w_groupsize!'
assert args.k_pre_rope == False, 'Pre-RoPE quantization is not supported yet!'
if args.model == 'facebook/opt-125m' or args.model == 'facebook/opt-1.3b':
logging.warning('Warning: OPT-125M/1.3B is only for debugging purposes!!')
if args.wandb:
assert args.wandb_id is not None and args.wandb_project is not None, 'WandB ID/project is not provided!'
logging.info('Arguments: ')
logging.info(pprint.pformat(vars(args)))
logging.info('--' * 30)
return args
def cleanup_memory(verbos=True) -> None:
"""Run GC and clear GPU memory."""
import gc
import inspect
caller_name = ''
try:
caller_name = f' (from {inspect.stack()[1].function})'
except (ValueError, KeyError):
pass
def total_reserved_mem() -> int:
return sum(torch.cuda.memory_reserved(device=i) for i in range(torch.cuda.device_count()))
memory_before = total_reserved_mem()
# gc.collect and empty cache are necessary to clean up GPU memory if the model was distributed
gc.collect()
if torch.cuda.is_available():
torch.cuda.empty_cache()
memory_after = total_reserved_mem()
if verbos:
logging.info(
f"GPU memory{caller_name}: {memory_before / (1024 ** 3):.2f} -> {memory_after / (1024 ** 3):.2f} GB"
f" ({(memory_after - memory_before) / (1024 ** 3):.2f} GB)"
)
def distribute_model(model) -> None:
"""Distribute the model across available GPUs. NB: only implemented for Llama-2."""
no_split_module_classes = ['LlamaDecoderLayer']
max_memory = get_balanced_memory(
model,
no_split_module_classes=no_split_module_classes,
)
device_map = infer_auto_device_map(
model, max_memory=max_memory, no_split_module_classes=no_split_module_classes
)
dispatch_model(
model,
device_map=device_map,
offload_buffers=True,
offload_dir="offload",
state_dict=model.state_dict(),
)
cleanup_memory()