File size: 4,275 Bytes
7d52396
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import sys
import time
import warnings
from pathlib import Path
from typing import Optional

import lightning as L
import torch

# support running without installing as a package
wd = Path(__file__).parent.parent.resolve()
sys.path.append(str(wd))

from generate import generate
from lit_llama import Tokenizer, LLaMA
from lit_llama.lora import lora
from lit_llama.utils import lazy_load, llama_model_lookup
from scripts.prepare_alpaca import generate_prompt

lora_r = 8
lora_alpha = 16
lora_dropout = 0.05


def main(
    prompt: str = "What food do lamas eat?",
    input: str = "",
    lora_path: Path = Path("out/lora/alpaca/lit-llama-lora-finetuned.pth"),
    pretrained_path: Path = Path("checkpoints/lit-llama/7B/lit-llama.pth"),
    tokenizer_path: Path = Path("checkpoints/lit-llama/tokenizer.model"),
    quantize: Optional[str] = None,
    max_new_tokens: int = 100,
    top_k: int = 200,
    temperature: float = 0.8,
) -> None:
    """Generates a response based on a given instruction and an optional input.
    This script will only work with checkpoints from the instruction-tuned LoRA model.
    See `finetune_lora.py`.

    Args:
        prompt: The prompt/instruction (Alpaca style).
        lora_path: Path to the checkpoint with trained LoRA weights, which are the output of
            `finetune_lora.py`.
        input: Optional input (Alpaca style).
        pretrained_path: The path to the checkpoint with pretrained LLaMA weights.
        tokenizer_path: The tokenizer path to load.
        quantize: Whether to quantize the model and using which method:
            ``"llm.int8"``: LLM.int8() mode,
            ``"gptq.int4"``: GPTQ 4-bit mode.
        max_new_tokens: The number of generation steps to take.
        top_k: The number of top most probable tokens to consider in the sampling process.
        temperature: A value controlling the randomness of the sampling process. Higher values result in more random
            samples.
    """
    assert lora_path.is_file()
    assert pretrained_path.is_file()
    assert tokenizer_path.is_file()

    if quantize is not None:
        raise NotImplementedError("Quantization in LoRA is not supported yet")

    precision = "bf16-true" if torch.cuda.is_available() and torch.cuda.is_bf16_supported() else "32-true"
    fabric = L.Fabric(devices=1, precision=precision)

    print("Loading model ...", file=sys.stderr)
    t0 = time.time()

    with lazy_load(pretrained_path) as pretrained_checkpoint, lazy_load(lora_path) as lora_checkpoint:
        name = llama_model_lookup(pretrained_checkpoint)

        with fabric.init_module(empty_init=True), lora(r=lora_r, alpha=lora_alpha, dropout=lora_dropout, enabled=True):
            model = LLaMA.from_name(name)

            # 1. Load the pretrained weights
            model.load_state_dict(pretrained_checkpoint, strict=False)
            # 2. Load the fine-tuned lora weights
            model.load_state_dict(lora_checkpoint, strict=False)

    print(f"Time to load model: {time.time() - t0:.02f} seconds.", file=sys.stderr)

    model.eval()
    model = fabric.setup(model)

    tokenizer = Tokenizer(tokenizer_path)
    sample = {"instruction": prompt, "input": input}
    prompt = generate_prompt(sample)
    encoded = tokenizer.encode(prompt, bos=True, eos=False, device=model.device)

    t0 = time.perf_counter()
    output = generate(
        model,
        idx=encoded,
        max_new_tokens=max_new_tokens,
        temperature=temperature,
        top_k=top_k,
        eos_id=tokenizer.eos_id
    )
    t = time.perf_counter() - t0

    output = tokenizer.decode(output)
    output = output.split("### Response:")[1].strip()
    print(output)

    print(f"\n\nTime for inference: {t:.02f} sec total, {max_new_tokens / t:.02f} tokens/sec", file=sys.stderr)
    if fabric.device.type == "cuda":
        print(f"Memory used: {torch.cuda.max_memory_reserved() / 1e9:.02f} GB", file=sys.stderr)


if __name__ == "__main__":
    from jsonargparse import CLI

    torch.set_float32_matmul_precision("high")
    warnings.filterwarnings(
        # Triggered internally at ../aten/src/ATen/EmptyTensor.cpp:31
        "ignore", 
        message="ComplexHalf support is experimental and many operators don't support it yet"
    )
    CLI(main)