File size: 4,054 Bytes
4c13408
 
 
92d0750
4c13408
92d0750
 
4c13408
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from typing import Dict, Any
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel, PeftConfig
import torch
import time

class EndpointHandler:
    def __init__(self, path="samadeniyi/lora_lesson_plan_model"):
        try:
            config = PeftConfig.from_pretrained(path)
        except TypeError as e:
            print(f"Error while loading config: {e}")
            # Manually filter out any unsupported config parameters (e.g., 'layer_replication')
            config_dict = PeftConfig.from_pretrained(path).__dict__
            if "layer_replication" in config_dict:
                del config_dict["layer_replication"]
            config = PeftConfig(**config_dict)

        # Define 4-bit quantization configuration (this is necessary for low-memory usage)
        bnb_config = BitsAndBytesConfig(
            load_in_4bit=True,
            bnb_4bit_use_double_quant=True,
            bnb_4bit_quant_type="nf4",
            bnb_4bit_compute_dtype=torch.float16,
        )

        # Load the model using 4-bit quantization and optimized settings
        self.model = AutoModelForCausalLM.from_pretrained(
            config.base_model_name_or_path,
            return_dict=True,
            device_map={"": 0},  # Map to CUDA device 0
            trust_remote_code=True,
            quantization_config=bnb_config,
        )

        # Load tokenizer and ensure it matches the model
        self.tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path)
        self.tokenizer.pad_token = self.tokenizer.eos_token

        # Apply PEFT (Parameter-Efficient Fine-Tuning) to the model
        self.model = PeftModel.from_pretrained(self.model, path)

    def __call__(self, data: Any) -> Dict[str, Any]:
        """
        Args:
            data :obj:`dict`:. The object should contain {"instruction": "some text", "input": "some text"}:
                - "instruction": The instruction describing what to generate.
                - "input": Context to guide the generation.

        Returns:
            A :obj:`dict` containing {"generated_text": "the generated lesson plan", "time": "..."}:
                - "generated_text": The generated text based on the input.
                - "time": The time taken to generate the output.
        """

        # Parse input data
        inputs = data.pop("inputs", data)
        instruction = inputs.get("instruction", "")
        input_context = inputs.get("input", "")

        # Create the lesson plan prompt based on your preparation format
        lesson_prompt = f"""Below is an instruction that describes how to create a lesson plan, paired with an input that provides further context. Write a response that appropriately completes the request.

### Instruction:
{instruction}

### Input:
{input_context}

### Response:
"""

        # Tokenize the prompt
        batch = self.tokenizer(
            lesson_prompt,
            padding=True,
            truncation=True,
            return_tensors='pt'
        )
        batch = batch.to('cuda:0')

        # Configure generation settings
        generation_config = self.model.generation_config
        generation_config.top_p = 0.7
        generation_config.temperature = 0.7
        generation_config.max_new_tokens = 256
        generation_config.num_return_sequences = 1
        generation_config.pad_token_id = self.tokenizer.eos_token_id
        generation_config.eos_token_id = self.tokenizer.eos_token_id

        # Time the prediction
        start = time.time()
        with torch.cuda.amp.autocast():
            output_tokens = self.model.generate(
                input_ids=batch.input_ids,
                generation_config=generation_config,
            )
        end = time.time()

        # Decode generated tokens into text
        generated_text = self.tokenizer.decode(output_tokens[0], skip_special_tokens=True)

        # Return the generated text and the time taken
        return {"generated_text": generated_text, "time": f"{(end-start):.2f} s"}