File size: 4,281 Bytes
5ed9e85
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import re
from functools import partial

import torch

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

import llama_cpp

if torch.cuda.is_available() and not torch.version.hip:
    try:
        import llama_cpp_cuda
    except:
        llama_cpp_cuda = None
else:
    llama_cpp_cuda = None


def llama_cpp_lib():
    if shared.args.cpu or llama_cpp_cuda is None:
        return llama_cpp
    else:
        return llama_cpp_cuda


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


class LlamaCppModel:
    def __init__(self):
        self.initialized = False

    def __del__(self):
        self.model.__del__()

    @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)

        logger.info("Cache capacity is " + str(cache_capacity) + " bytes")
        params = {
            'model_path': str(path),
            'n_ctx': shared.args.n_ctx,
            'seed': int(shared.args.llama_cpp_seed),
            'n_threads': shared.args.threads or None,
            'n_batch': shared.args.n_batch,
            'use_mmap': not shared.args.no_mmap,
            'use_mlock': shared.args.mlock,
            'low_vram': shared.args.low_vram,
            'n_gpu_layers': shared.args.n_gpu_layers,
            'rope_freq_base': 10000 * shared.args.alpha_value ** (64 / 63.),
            'rope_freq_scale': 1.0 / shared.args.compress_pos_emb,
            'n_gqa': shared.args.n_gqa or None,
            'rms_norm_eps': shared.args.rms_norm_eps or None,
        }

        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, tokens):
        return self.model.detokenize(tokens)

    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).decode('utf-8')

        completion_chunks = self.model.create_completion(
            prompt=prompt,
            max_tokens=state['max_new_tokens'],
            temperature=state['temperature'],
            top_p=state['top_p'],
            top_k=state['top_k'],
            repeat_penalty=state['repetition_penalty'],
            tfs_z=state['tfs'],
            mirostat_mode=int(state['mirostat_mode']),
            mirostat_tau=state['mirostat_tau'],
            mirostat_eta=state['mirostat_eta'],
            stream=True,
            logits_processor=LogitsProcessorList([
                partial(ban_eos_logits_processor, self.model.token_eos()),
            ]) if state['ban_eos_token'] else None,
        )

        output = ""
        for completion_chunk in completion_chunks:
            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