File size: 4,809 Bytes
d754e91
 
87a0e23
d754e91
 
90c428d
d754e91
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
90c428d
 
 
87a0e23
a1771a7
 
 
 
 
 
 
 
 
 
 
90c428d
49b832f
d754e91
90c428d
 
 
d754e91
90c428d
a1771a7
 
d754e91
 
90c428d
 
d754e91
 
 
 
90c428d
 
d754e91
 
90c428d
4b2400e
 
 
90c428d
4b2400e
49b832f
90c428d
 
 
966795b
90c428d
 
 
4b2400e
90c428d
 
d754e91
90c428d
d754e91
 
90c428d
a1771a7
 
87a0e23
 
 
90c428d
 
 
 
 
 
 
 
 
 
 
 
 
 
a1771a7
 
90c428d
 
 
 
 
 
 
 
 
 
 
d754e91
90c428d
 
 
d754e91
a1771a7
 
d754e91
 
90c428d
 
 
d754e91
 
 
 
90c428d
 
 
 
d754e91
87a0e23
90c428d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7b14813
87a0e23
a1771a7
 
 
 
 
9279c83
 
 
 
 
 
 
 
87a0e23
90c428d
 
 
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
import os
import sys
import gc

import torch
from transformers import LlamaForCausalLM, LlamaTokenizer
from peft import PeftModel

from .globals import Global


def get_device():
    if torch.cuda.is_available():
        return "cuda"
    else:
        return "cpu"

    try:
        if torch.backends.mps.is_available():
            return "mps"
    except:  # noqa: E722
        pass


def get_new_base_model(base_model_name):
    if Global.ui_dev_mode:
        return

    if Global.new_base_model_that_is_ready_to_be_used:
        if Global.name_of_new_base_model_that_is_ready_to_be_used == base_model_name:
            model = Global.new_base_model_that_is_ready_to_be_used
            Global.new_base_model_that_is_ready_to_be_used = None
            Global.name_of_new_base_model_that_is_ready_to_be_used = None
            return model
        else:
            Global.new_base_model_that_is_ready_to_be_used = None
            Global.name_of_new_base_model_that_is_ready_to_be_used = None
            clear_cache()

    device = get_device()

    if device == "cuda":
        model = LlamaForCausalLM.from_pretrained(
            base_model_name,
            load_in_8bit=Global.load_8bit,
            torch_dtype=torch.float16,
            # device_map="auto",
            # ? https://github.com/tloen/alpaca-lora/issues/21
            device_map={'': 0},
        )
    elif device == "mps":
        model = LlamaForCausalLM.from_pretrained(
            base_model_name,
            device_map={"": device},
            torch_dtype=torch.float16,
        )
    else:
        model = LlamaForCausalLM.from_pretrained(
            base_model_name, device_map={"": device}, low_cpu_mem_usage=True
        )

    model.config.pad_token_id = get_tokenizer(base_model_name).pad_token_id = 0
    model.config.bos_token_id = 1
    model.config.eos_token_id = 2

    return model


def get_tokenizer(base_model_name):
    if Global.ui_dev_mode:
        return

    loaded_tokenizer = Global.loaded_tokenizers.get(base_model_name)
    if loaded_tokenizer:
        return loaded_tokenizer

    tokenizer = LlamaTokenizer.from_pretrained(base_model_name)
    Global.loaded_tokenizers.set(base_model_name, tokenizer)

    return tokenizer


def get_model(
        base_model_name,
        peft_model_name=None):
    if Global.ui_dev_mode:
        return

    if peft_model_name == "None":
        peft_model_name = None

    model_key = base_model_name
    if peft_model_name:
        model_key = f"{base_model_name}//{peft_model_name}"

    loaded_model = Global.loaded_models.get(model_key)
    if loaded_model:
        return loaded_model

    peft_model_name_or_path = peft_model_name

    lora_models_directory_path = os.path.join(Global.data_dir, "lora_models")
    possible_lora_model_path = os.path.join(
        lora_models_directory_path, peft_model_name)
    if os.path.isdir(possible_lora_model_path):
        peft_model_name_or_path = possible_lora_model_path

    Global.loaded_models.prepare_to_set()
    clear_cache()

    model = get_new_base_model(base_model_name)

    if peft_model_name:
        device = get_device()

        if device == "cuda":
            model = PeftModel.from_pretrained(
                model,
                peft_model_name_or_path,
                torch_dtype=torch.float16,
                # ? https://github.com/tloen/alpaca-lora/issues/21
                device_map={'': 0},
            )
        elif device == "mps":
            model = PeftModel.from_pretrained(
                model,
                peft_model_name_or_path,
                device_map={"": device},
                torch_dtype=torch.float16,
            )
        else:
            model = PeftModel.from_pretrained(
                model,
                peft_model_name_or_path,
                device_map={"": device},
            )

    model.config.pad_token_id = get_tokenizer(base_model_name).pad_token_id = 0
    model.config.bos_token_id = 1
    model.config.eos_token_id = 2

    if not Global.load_8bit:
        model.half()  # seems to fix bugs for some users.

    model.eval()
    if torch.__version__ >= "2" and sys.platform != "win32":
        model = torch.compile(model)

    Global.loaded_models.set(model_key, model)
    clear_cache()

    return model


def prepare_base_model(base_model_name=Global.default_base_model_name):
    Global.new_base_model_that_is_ready_to_be_used = get_new_base_model(base_model_name)
    Global.name_of_new_base_model_that_is_ready_to_be_used = base_model_name


def clear_cache():
    gc.collect()

    # if not shared.args.cpu: # will not be running on CPUs anyway
    with torch.no_grad():
        torch.cuda.empty_cache()


def unload_models():
    Global.loaded_models.clear()
    Global.loaded_tokenizers.clear()
    clear_cache()