File size: 4,808 Bytes
38098b4
 
efb4fe4
38098b4
 
 
 
 
 
 
efb4fe4
 
38098b4
 
efb4fe4
38098b4
efb4fe4
 
 
 
 
 
 
 
 
 
 
 
 
38098b4
efb4fe4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ff08af5
efb4fe4
 
 
 
 
 
ff08af5
efb4fe4
 
 
 
 
 
ff08af5
efb4fe4
 
 
 
ff08af5
efb4fe4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ff08af5
efb4fe4
ff08af5
efb4fe4
 
 
 
 
 
 
 
 
 
ff08af5
efb4fe4
 
ff08af5
38098b4
 
efb4fe4
38098b4
 
efb4fe4
38098b4
efb4fe4
ff08af5
efb4fe4
ff08af5
38098b4
efb4fe4
 
 
 
38098b4
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os
import torch
from transformers import pipeline, BitsAndBytesConfig
from dotenv import load_dotenv
from pathlib import Path

env_path = Path(__file__).resolve().parent.parent / ".env"
load_dotenv(dotenv_path=env_path)

MODEL_ID = os.getenv("MODEL_ID", "Qwen/Qwen2.5-0.5B-Instruct")
QUANTIZATION = os.getenv("QUANTIZATION", "auto")
USE_DOUBLE_QUANT = os.getenv("USE_DOUBLE_QUANT", "true").lower() == "true"

_pipe = None
_current_model = None

def _log(msg: str):
    print(f"[ModelLoader] {msg}")

def _has_gpu() -> bool:
    return torch.cuda.is_available()

def _gpu_name() -> str:
    if _has_gpu():
        return torch.cuda.get_device_name(0)
    return "None"

def _gpu_memory_gb() -> float:
    if _has_gpu():
        try:
            return torch.cuda.get_device_properties(0).total_mem / 1e9
        except:
            return 0
    return 0

def _select_quantization() -> str:
    """Auto-select quantization tier based on MODEL_ID and hardware."""
    user_mode = QUANTIZATION.lower()

    if user_mode == "none":
        return "none"

    if user_mode != "auto":
        return user_mode

    # Auto-detect: GPU with enough VRAM for requested model
    if "7B" in MODEL_ID:
        if _has_gpu() and _gpu_memory_gb() >= 5.5:
            _log(f"7B model detected, GPU {_gpu_name()} ({_gpu_memory_gb():.1f}GB) — using 4-bit")
            return "4bit"
        _log("7B model requested but no GPU with 5.5GB+ VRAM — falling back to 1.5B 8-bit")
        return "cpu_fallback_8bit"

    if "1.5B" in MODEL_ID:
        if _has_gpu():
            _log(f"1.5B model detected, GPU available — using 8-bit")
            return "8bit"
        _log("1.5B model detected, CPU only — using bfloat16")
        return "none"

    return "none"

def _build_model_kwargs(quant_mode: str) -> dict:
    """Build pipeline kwargs based on quantization mode."""
    kwargs = {
        "trust_remote_code": True,
    }

    if quant_mode == "4bit":
        kwargs["device_map"] = "auto"
        kwargs["quantization_config"] = BitsAndBytesConfig(
            load_in_4bit=True,
            bnb_4bit_compute_dtype=torch.bfloat16,
            bnb_4bit_use_double_quant=USE_DOUBLE_QUANT,
            bnb_4bit_quant_type="nf4",
        )
        _log("[OK] 4-bit quantization enabled (NF4, double quant)")

    elif quant_mode == "8bit":
        kwargs["device_map"] = "auto"
        kwargs["quantization_config"] = BitsAndBytesConfig(
            load_in_8bit=True,
        )
        _log("[OK] 8-bit quantization enabled")

    elif quant_mode == "cpu_fallback_8bit":
        kwargs["device_map"] = "auto"
        kwargs["quantization_config"] = BitsAndBytesConfig(
            load_in_8bit=True,
        )
        _log("[OK] CPU fallback 8-bit for 1.5B model")

    else:
        kwargs["torch_dtype"] = torch.bfloat16
        kwargs["device_map"] = "auto"
        _log(f"[OK] Loading {MODEL_ID} in bfloat16 (CPU-friendly)")

    return kwargs

def get_pipe():
    global _pipe, _current_model

    if _pipe is not None:
        return _pipe

    actual_model_id = MODEL_ID
    quant_mode = _select_quantization()

    # Handle CPU fallback for 7B → 1.5B
    if quant_mode == "cpu_fallback_8bit":
        actual_model_id = "Qwen/Qwen2.5-1.5B-Instruct"
        _log(f"[FALLBACK] loading {actual_model_id} instead of {MODEL_ID}")

    _log(f"Loading {actual_model_id} (quantization: {quant_mode})")
    _log(f"   Hardware: GPU={_gpu_name()}, VRAM={_gpu_memory_gb():.1f}GB, CUDA={_has_gpu()}")

    try:
        kwargs = _build_model_kwargs(quant_mode)
        _pipe = pipeline(
            "text-generation",
            model=actual_model_id,
            **kwargs
        )
        _current_model = actual_model_id
        _log("[DONE] Model loaded successfully!")
    except ImportError as e:
        if "bitsandbytes" in str(e):
            _log("[ERROR] bitsandbytes not installed. Falling back to CPU bfloat16.")
            _pipe = pipeline(
                "text-generation",
                model=actual_model_id,
                torch_dtype=torch.bfloat16,
                device_map="auto",
                trust_remote_code=True,
            )
            _current_model = actual_model_id
            _log("[DONE] Model loaded with CPU fallback")
        else:
            _log(f"[ERROR] Model load failed: {e}")
            _pipe = None
    except Exception as e:
        _log(f"❌ Model load failed: {e}")
        _pipe = None

    return _pipe

def generate_text(messages, temperature=0.3, max_new_tokens=2000):
    pipe = get_pipe()
    if pipe is None:
        return None
    outputs = pipe(
        messages,
        max_new_tokens=max_new_tokens,
        do_sample=True,
        temperature=temperature,
        top_p=0.9
    )
    return outputs[0]["generated_text"][-1]["content"]