File size: 4,484 Bytes
3ccaf5a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
Offline quantization of base models to AWQ / GPTQ with a FIXED calibration set,
using llmcompressor (Neural Magic). Output is compressed-tensors format, which
vLLM loads natively.

Must be run from the `sonthh-stepprobe-quant` conda env (vLLM's main env has
incompatible dep pins with llmcompressor).

Usage:
    /home/aiteam1/anaconda3/envs/sonthh-stepprobe-quant/bin/python \
        scripts/quantize_models.py \
        --model deepseek-ai/DeepSeek-R1-Distill-Qwen-7B \
        --method gptq --bits 4 --group-size 128 \
        --output results/quantized_models/r1-qwen-7b_gptq_w4
"""
import argparse
import os
import random


CALIB_N_SAMPLES = 128
CALIB_SEQ_LEN = 512
CALIB_SEED = 42


def load_calib_dataset(tokenizer):
    """Fixed WikiText-2 calibration set, same for every (model, method, bits)."""
    from datasets import load_dataset

    ds = load_dataset("wikitext", "wikitext-2-raw-v1", split="train")
    ds = ds.filter(lambda x: len(x["text"].strip()) > 200)
    ds = ds.shuffle(seed=CALIB_SEED).select(range(min(CALIB_N_SAMPLES, len(ds))))

    def tokenize(example):
        return tokenizer(
            example["text"],
            padding=False,
            max_length=CALIB_SEQ_LEN,
            truncation=True,
            add_special_tokens=False,
        )

    ds = ds.map(tokenize, remove_columns=ds.column_names)
    return ds


def build_recipe(method: str, bits: int, group_size: int):
    """Build an llmcompressor recipe for the given (method, bits)."""
    from compressed_tensors.quantization import (
        QuantizationArgs,
        QuantizationScheme,
        QuantizationStrategy,
        QuantizationType,
    )
    from llmcompressor.modifiers.quantization import GPTQModifier

    weights_args = QuantizationArgs(
        num_bits=bits,
        type=QuantizationType.INT,
        symmetric=True,
        strategy=QuantizationStrategy.GROUP,
        group_size=group_size,
    )
    scheme = QuantizationScheme(targets=["Linear"], weights=weights_args)

    if method == "gptq":
        return GPTQModifier(
            config_groups={"group_0": scheme},
            ignore=["lm_head"],
        )
    elif method == "awq":
        if bits != 4:
            raise ValueError(f"AWQ path only supports 4-bit; got bits={bits}")
        from llmcompressor.modifiers.awq import AWQModifier
        return AWQModifier(
            config_groups={"group_0": scheme},
            ignore=["lm_head"],
        )
    else:
        raise ValueError(f"Unknown method: {method}")


def quantize(model_name: str, method: str, bits: int, group_size: int, output_dir: str):
    from llmcompressor import oneshot
    from transformers import AutoModelForCausalLM, AutoTokenizer

    print(f"Loading base model: {model_name}")
    model = AutoModelForCausalLM.from_pretrained(
        model_name,
        torch_dtype="auto",
        device_map="auto",
        trust_remote_code=True,
    )
    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)

    print(f"Loading calibration set: WikiText-2 ({CALIB_N_SAMPLES} samples, seq_len={CALIB_SEQ_LEN}, seed={CALIB_SEED})")
    ds = load_calib_dataset(tokenizer)

    recipe = build_recipe(method, bits, group_size)
    print(f"Running {method.upper()} oneshot: bits={bits}, group_size={group_size}")

    oneshot(
        model=model,
        dataset=ds,
        recipe=recipe,
        output_dir=output_dir,
        max_seq_length=CALIB_SEQ_LEN,
        num_calibration_samples=CALIB_N_SAMPLES,
    )

    tokenizer.save_pretrained(output_dir)


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--model", required=True, help="HuggingFace model name or local path")
    parser.add_argument("--method", choices=["awq", "gptq"], required=True)
    parser.add_argument("--bits", type=int, required=True)
    parser.add_argument("--group-size", type=int, default=128)
    parser.add_argument("--output", required=True, help="Output directory for quantized model")
    args = parser.parse_args()

    if os.path.isfile(os.path.join(args.output, "config.json")):
        print(f"[SKIP] Already quantized: {args.output}")
        return

    os.makedirs(args.output, exist_ok=True)
    print(f"Quantizing {args.model} -> {args.method} w{args.bits} (group={args.group_size})")
    print(f"Output: {args.output}")

    quantize(args.model, args.method, args.bits, args.group_size, args.output)

    print(f"Done: {args.output}")


if __name__ == "__main__":
    main()