File size: 5,268 Bytes
167596f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import copy
import os
from functools import lru_cache

import pipmaster as pm  # Pipmaster for dynamic library install

# install specific modules
if not pm.is_installed("transformers"):
    pm.install("transformers")
if not pm.is_installed("torch"):
    pm.install("torch")
if not pm.is_installed("numpy"):
    pm.install("numpy")

from transformers import AutoTokenizer, AutoModelForCausalLM
from tenacity import (
    retry,
    stop_after_attempt,
    wait_exponential,
    retry_if_exception_type,
)
from lightrag.exceptions import (
    APIConnectionError,
    RateLimitError,
    APITimeoutError,
)
import torch
import numpy as np

os.environ["TOKENIZERS_PARALLELISM"] = "false"


@lru_cache(maxsize=1)
def initialize_hf_model(model_name):
    hf_tokenizer = AutoTokenizer.from_pretrained(
        model_name, device_map="auto", trust_remote_code=True
    )
    hf_model = AutoModelForCausalLM.from_pretrained(
        model_name, device_map="auto", trust_remote_code=True
    )
    if hf_tokenizer.pad_token is None:
        hf_tokenizer.pad_token = hf_tokenizer.eos_token

    return hf_model, hf_tokenizer


@retry(
    stop=stop_after_attempt(3),
    wait=wait_exponential(multiplier=1, min=4, max=10),
    retry=retry_if_exception_type(
        (RateLimitError, APIConnectionError, APITimeoutError)
    ),
)
async def hf_model_if_cache(
    model,
    prompt,
    system_prompt=None,
    history_messages=[],
    enable_cot: bool = False,
    **kwargs,
) -> str:
    if enable_cot:
        from lightrag.utils import logger

        logger.debug(
            "enable_cot=True is not supported for Hugging Face local models and will be ignored."
        )
    model_name = model
    hf_model, hf_tokenizer = initialize_hf_model(model_name)
    messages = []
    if system_prompt:
        messages.append({"role": "system", "content": system_prompt})
    messages.extend(history_messages)
    messages.append({"role": "user", "content": prompt})
    kwargs.pop("hashing_kv", None)
    input_prompt = ""
    try:
        input_prompt = hf_tokenizer.apply_chat_template(
            messages, tokenize=False, add_generation_prompt=True
        )
    except Exception:
        try:
            ori_message = copy.deepcopy(messages)
            if messages[0]["role"] == "system":
                messages[1]["content"] = (
                    "<system>"
                    + messages[0]["content"]
                    + "</system>\n"
                    + messages[1]["content"]
                )
                messages = messages[1:]
                input_prompt = hf_tokenizer.apply_chat_template(
                    messages, tokenize=False, add_generation_prompt=True
                )
        except Exception:
            len_message = len(ori_message)
            for msgid in range(len_message):
                input_prompt = (
                    input_prompt
                    + "<"
                    + ori_message[msgid]["role"]
                    + ">"
                    + ori_message[msgid]["content"]
                    + "</"
                    + ori_message[msgid]["role"]
                    + ">\n"
                )

    input_ids = hf_tokenizer(
        input_prompt, return_tensors="pt", padding=True, truncation=True
    ).to("cuda")
    inputs = {k: v.to(hf_model.device) for k, v in input_ids.items()}
    output = hf_model.generate(
        **input_ids, max_new_tokens=512, num_return_sequences=1, early_stopping=True
    )
    response_text = hf_tokenizer.decode(
        output[0][len(inputs["input_ids"][0]) :], skip_special_tokens=True
    )

    return response_text


async def hf_model_complete(
    prompt,
    system_prompt=None,
    history_messages=[],
    keyword_extraction=False,
    enable_cot: bool = False,
    **kwargs,
) -> str:
    kwargs.pop("keyword_extraction", None)
    model_name = kwargs["hashing_kv"].global_config["llm_model_name"]
    result = await hf_model_if_cache(
        model_name,
        prompt,
        system_prompt=system_prompt,
        history_messages=history_messages,
        enable_cot=enable_cot,
        **kwargs,
    )
    return result


async def hf_embed(texts: list[str], tokenizer, embed_model) -> np.ndarray:
    # Detect the appropriate device
    if torch.cuda.is_available():
        device = next(embed_model.parameters()).device  # Use CUDA if available
    elif torch.backends.mps.is_available():
        device = torch.device("mps")  # Use MPS for Apple Silicon
    else:
        device = torch.device("cpu")  # Fallback to CPU

    # Move the model to the detected device
    embed_model = embed_model.to(device)

    # Tokenize the input texts and move them to the same device
    encoded_texts = tokenizer(
        texts, return_tensors="pt", padding=True, truncation=True
    ).to(device)

    # Perform inference
    with torch.no_grad():
        outputs = embed_model(
            input_ids=encoded_texts["input_ids"],
            attention_mask=encoded_texts["attention_mask"],
        )
        embeddings = outputs.last_hidden_state.mean(dim=1)

    # Convert embeddings to NumPy
    if embeddings.dtype == torch.bfloat16:
        return embeddings.detach().to(torch.float32).cpu().numpy()
    else:
        return embeddings.detach().cpu().numpy()