File size: 8,478 Bytes
72268ee
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from model import ExLlama, ExLlamaCache, ExLlamaConfig
from lora import ExLlamaLora
from tokenizer import ExLlamaTokenizer
from generator import ExLlamaGenerator
import argparse
import torch
import sys
import os
import glob
import model_init

# Simple interactive chatbot script

torch.set_grad_enabled(False)
torch.cuda._lazy_init()

# Parse arguments

parser = argparse.ArgumentParser(description = "Simple chatbot example for ExLlama")

model_init.add_args(parser)

parser.add_argument("-lora", "--lora", type = str, help = "Path to LoRA binary to use during benchmark")
parser.add_argument("-loracfg", "--lora_config", type = str, help = "Path to LoRA config to use during benchmark")
parser.add_argument("-ld", "--lora_dir", type = str, help = "Path to LoRA config and binary. to use during benchmark")

parser.add_argument("-p", "--prompt", type = str, help = "Prompt file")
parser.add_argument("-un", "--username", type = str, help = "Display name of user", default = "User")
parser.add_argument("-bn", "--botname", type = str, help = "Display name of chatbot", default = "Chatbort")
parser.add_argument("-bf", "--botfirst", action = "store_true", help = "Start chat on bot's turn")

parser.add_argument("-nnl", "--no_newline", action = "store_true", help = "Do not break bot's response on newline (allow multi-paragraph responses)")
parser.add_argument("-temp", "--temperature", type = float, help = "Temperature", default = 0.95)
parser.add_argument("-topk", "--top_k", type = int, help = "Top-K", default = 20)
parser.add_argument("-topp", "--top_p", type = float, help = "Top-P", default = 0.65)
parser.add_argument("-minp", "--min_p", type = float, help = "Min-P", default = 0.00)
parser.add_argument("-repp",  "--repetition_penalty", type = float, help = "Repetition penalty", default = 1.15)
parser.add_argument("-repps", "--repetition_penalty_sustain", type = int, help = "Past length for repetition penalty", default = 256)
parser.add_argument("-beams", "--beams", type = int, help = "Number of beams for beam search", default = 1)
parser.add_argument("-beamlen", "--beam_length", type = int, help = "Number of future tokens to consider", default = 1)

args = parser.parse_args()
model_init.post_parse(args)
model_init.get_model_files(args)

# Paths

if args.lora_dir is not None:
    args.lora_config = os.path.join(args.lora_dir, "adapter_config.json")
    args.lora = os.path.join(args.lora_dir, "adapter_model.bin")

# Some feedback

print(f" -- Sequence length: {args.length}")
print(f" -- Temperature: {args.temperature:.2f}")
print(f" -- Top-K: {args.top_k}")
print(f" -- Top-P: {args.top_p:.2f}")
print(f" -- Min-P: {args.min_p:.2f}")
print(f" -- Repetition penalty: {args.repetition_penalty:.2f}")
print(f" -- Beams: {args.beams} x {args.beam_length}")

print_opts = []
if args.no_newline: print_opts.append("no_newline")
if args.botfirst: print_opts.append("botfirst")

model_init.print_options(args, print_opts)

# Globals

model_init.set_globals(args)

# Load prompt file

username = args.username
bot_name = args.botname

if args.prompt is not None:
    with open(args.prompt, "r") as f:
        past = f.read()
        past = past.replace("{username}", username)
        past = past.replace("{bot_name}", bot_name)
        past = past.strip() + "\n"
else:
    past = f"{bot_name}: Hello, {username}\n"

# past += "User: Hi. Please say \"Shhhhhh\"?\n"
# args.botfirst = True

# Instantiate model and generator

config = model_init.make_config(args)

model = ExLlama(config)
cache = ExLlamaCache(model)
tokenizer = ExLlamaTokenizer(args.tokenizer)

model_init.print_stats(model)

# Load LoRA

lora = None
if args.lora:
    print(f" -- LoRA config: {args.lora_config}")
    print(f" -- Loading LoRA: {args.lora}")
    if args.lora_config is None:
        print(f" ## Error: please specify lora path to adapter_config.json")
        sys.exit()
    lora = ExLlamaLora(model, args.lora_config, args.lora)
    if lora.bias_ignored:
        print(f" !! Warning: LoRA zero bias ignored")

# Generator

generator = ExLlamaGenerator(model, tokenizer, cache)
generator.settings = ExLlamaGenerator.Settings()
generator.settings.temperature = args.temperature
generator.settings.top_k = args.top_k
generator.settings.top_p = args.top_p
generator.settings.min_p = args.min_p
generator.settings.token_repetition_penalty_max = args.repetition_penalty
generator.settings.token_repetition_penalty_sustain = args.repetition_penalty_sustain
generator.settings.token_repetition_penalty_decay = generator.settings.token_repetition_penalty_sustain // 2
generator.settings.beams = args.beams
generator.settings.beam_length = args.beam_length

generator.lora = lora

break_on_newline = not args.no_newline

# Be nice to Chatbort

min_response_tokens = 4
max_response_tokens = 256
extra_prune = 256

print(past, end = "")
ids = tokenizer.encode(past)
generator.gen_begin(ids)

next_userprompt = username + ": "

first_round = True

while True:

    res_line = bot_name + ":"
    res_tokens = tokenizer.encode(res_line)
    num_res_tokens = res_tokens.shape[-1]  # Decode from here

    if first_round and args.botfirst: in_tokens = res_tokens

    else:

        # Read and format input

        in_line = input(next_userprompt)
        in_line = username + ": " + in_line.strip() + "\n"

        next_userprompt = username + ": "

        # No need for this, really, unless we were logging the chat. The actual history we work on is kept in the
        # tokenized sequence in the generator and the state in the cache.

        past += in_line

        # SentencePiece doesn't tokenize spaces separately so we can't know from individual tokens if they start a new word
        # or not. Instead, repeatedly decode the generated response as it's being built, starting from the last newline,
        # and print out the differences between consecutive decodings to stream out the response.

        in_tokens = tokenizer.encode(in_line)
        in_tokens = torch.cat((in_tokens, res_tokens), dim = 1)

    # If we're approaching the context limit, prune some whole lines from the start of the context. Also prune a
    # little extra so we don't end up rebuilding the cache on every line when up against the limit.

    expect_tokens = in_tokens.shape[-1] + max_response_tokens
    max_tokens = config.max_seq_len - expect_tokens
    if generator.gen_num_tokens() >= max_tokens:
        generator.gen_prune_to(config.max_seq_len - expect_tokens - extra_prune, tokenizer.newline_token_id)

    # Feed in the user input and "{bot_name}:", tokenized

    generator.gen_feed_tokens(in_tokens)

    # Generate with streaming

    print(res_line, end = "")
    sys.stdout.flush()

    generator.begin_beam_search()

    for i in range(max_response_tokens):

        # Disallowing the end condition tokens seems like a clean way to force longer replies.

        if i < min_response_tokens:
            generator.disallow_tokens([tokenizer.newline_token_id, tokenizer.eos_token_id])
        else:
            generator.disallow_tokens(None)

        # Get a token

        gen_token = generator.beam_search()

        # If token is EOS, replace it with newline before continuing

        if gen_token.item() == tokenizer.eos_token_id:
            generator.replace_last_token(tokenizer.newline_token_id)

        # Decode the current line and print any characters added

        num_res_tokens += 1
        text = tokenizer.decode(generator.sequence_actual[:, -num_res_tokens:][0])
        new_text = text[len(res_line):]

        skip_space = res_line.endswith("\n") and new_text.startswith(" ")  # Bit prettier console output
        res_line += new_text
        if skip_space: new_text = new_text[1:]

        print(new_text, end="")  # (character streaming output is here)
        sys.stdout.flush()

        # End conditions

        if break_on_newline and gen_token.item() == tokenizer.newline_token_id: break
        if gen_token.item() == tokenizer.eos_token_id: break

        # Some models will not (or will inconsistently) emit EOS tokens but in a chat sequence will often begin
        # generating for the user instead. Try to catch this and roll back a few tokens to begin the user round.

        if res_line.endswith(f"{username}:"):
            plen = tokenizer.encode(f"{username}:").shape[-1]
            generator.gen_rewind(plen)
            next_userprompt = " "
            break

    generator.end_beam_search()

    past += res_line
    first_round = False