File size: 6,329 Bytes
cfbfaf1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import re
from functools import partial

import numpy as np
import torch

from modules import RoPE, shared
from modules.callbacks import Iteratorize
from modules.logging_colors import logger
from modules.text_generation import get_max_prompt_length

try:
    import llama_cpp
except:
    llama_cpp = None

try:
    import llama_cpp_cuda
except:
    llama_cpp_cuda = None

try:
    import llama_cpp_cuda_tensorcores
except:
    llama_cpp_cuda_tensorcores = None


def llama_cpp_lib():
    if shared.args.cpu and llama_cpp is not None:
        return llama_cpp
    elif shared.args.tensorcores and llama_cpp_cuda_tensorcores is not None:
        return llama_cpp_cuda_tensorcores
    elif llama_cpp_cuda is not None:
        return llama_cpp_cuda
    else:
        return llama_cpp


def ban_eos_logits_processor(eos_token, input_ids, logits):
    logits[eos_token] = -float('inf')
    return logits


def custom_token_ban_logits_processor(token_ids, input_ids, logits):
    for token_id in token_ids:
        logits[token_id] = -float('inf')

    return logits


class LlamaCppModel:
    def __init__(self):
        self.initialized = False
        self.grammar_string = ''
        self.grammar = None

    def __del__(self):
        del self.model

    @classmethod
    def from_pretrained(self, path):

        Llama = llama_cpp_lib().Llama
        LlamaCache = llama_cpp_lib().LlamaCache

        result = self()
        cache_capacity = 0
        if shared.args.cache_capacity is not None:
            if 'GiB' in shared.args.cache_capacity:
                cache_capacity = int(re.sub('[a-zA-Z]', '', shared.args.cache_capacity)) * 1000 * 1000 * 1000
            elif 'MiB' in shared.args.cache_capacity:
                cache_capacity = int(re.sub('[a-zA-Z]', '', shared.args.cache_capacity)) * 1000 * 1000
            else:
                cache_capacity = int(shared.args.cache_capacity)

        if cache_capacity > 0:
            logger.info("Cache capacity is " + str(cache_capacity) + " bytes")

        if shared.args.tensor_split is None or shared.args.tensor_split.strip() == '':
            tensor_split_list = None
        else:
            tensor_split_list = [float(x) for x in shared.args.tensor_split.strip().split(",")]

        params = {
            'model_path': str(path),
            'n_ctx': shared.args.n_ctx,
            'n_threads': shared.args.threads or None,
            'n_threads_batch': shared.args.threads_batch or None,
            'n_batch': shared.args.n_batch,
            'use_mmap': not shared.args.no_mmap,
            'use_mlock': shared.args.mlock,
            'mul_mat_q': not shared.args.no_mul_mat_q,
            'numa': shared.args.numa,
            'n_gpu_layers': shared.args.n_gpu_layers,
            'rope_freq_base': RoPE.get_rope_freq_base(shared.args.alpha_value, shared.args.rope_freq_base),
            'tensor_split': tensor_split_list,
            'rope_freq_scale': 1.0 / shared.args.compress_pos_emb,
            'offload_kqv': not shared.args.no_offload_kqv
        }

        result.model = Llama(**params)
        if cache_capacity > 0:
            result.model.set_cache(LlamaCache(capacity_bytes=cache_capacity))

        # This is ugly, but the model and the tokenizer are the same object in this library.
        return result, result

    def encode(self, string):
        if type(string) is str:
            string = string.encode()

        return self.model.tokenize(string)

    def decode(self, ids, **kwargs):
        return self.model.detokenize(ids).decode('utf-8')

    def get_logits(self, tokens):
        self.model.reset()
        self.model.eval(tokens)
        logits = self.model._scores
        logits = np.expand_dims(logits, 0)  # batch dim is expected
        return torch.tensor(logits, dtype=torch.float32)

    def load_grammar(self, string):
        if string != self.grammar_string:
            self.grammar_string = string
            if string.strip() != '':
                self.grammar = llama_cpp_lib().LlamaGrammar.from_string(string)
            else:
                self.grammar = None

    def generate(self, prompt, state, callback=None):
        LogitsProcessorList = llama_cpp_lib().LogitsProcessorList
        prompt = prompt if type(prompt) is str else prompt.decode()

        # Handle truncation
        prompt = self.encode(prompt)
        prompt = prompt[-get_max_prompt_length(state):]
        prompt = self.decode(prompt)

        self.load_grammar(state['grammar_string'])
        logit_processors = LogitsProcessorList()
        if state['ban_eos_token']:
            logit_processors.append(partial(ban_eos_logits_processor, self.model.token_eos()))

        if state['custom_token_bans']:
            to_ban = [int(x) for x in state['custom_token_bans'].split(',')]
            if len(to_ban) > 0:
                logit_processors.append(partial(custom_token_ban_logits_processor, to_ban))

        completion_chunks = self.model.create_completion(
            prompt=prompt,
            max_tokens=state['max_new_tokens'],
            temperature=state['temperature'],
            top_p=state['top_p'],
            min_p=state['min_p'],
            typical_p=state['typical_p'],
            frequency_penalty=state['frequency_penalty'],
            presence_penalty=state['presence_penalty'],
            repeat_penalty=state['repetition_penalty'],
            top_k=state['top_k'],
            stream=True,
            seed=int(state['seed']) if state['seed'] != -1 else None,
            tfs_z=state['tfs'],
            mirostat_mode=int(state['mirostat_mode']),
            mirostat_tau=state['mirostat_tau'],
            mirostat_eta=state['mirostat_eta'],
            logits_processor=logit_processors,
            grammar=self.grammar
        )

        output = ""
        for completion_chunk in completion_chunks:
            if shared.stop_everything:
                break

            text = completion_chunk['choices'][0]['text']
            output += text
            if callback:
                callback(text)

        return output

    def generate_with_streaming(self, *args, **kwargs):
        with Iteratorize(self.generate, args, kwargs, callback=None) as generator:
            reply = ''
            for token in generator:
                reply += token
                yield reply