File size: 6,451 Bytes
1215123
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from __future__ import annotations

import argparse
from math import prod
from pathlib import Path
import ctypes
import logging
import numpy as np


import gguf
from gguf.constants import GGMLQuantizationType


logger = logging.getLogger(__name__)


c_float_p = ctypes.POINTER(ctypes.c_float)


class ggml_init_params(ctypes.Structure):
    _fields_ = [
        ("mem_size", ctypes.c_size_t),
        ("mem_buffer", ctypes.c_void_p),
        ("no_alloc", ctypes.c_bool),
    ]


class GGMLQuants:
    libggml: ctypes.CDLL

    def __init__(self, libggml: Path):
        self.libggml = ctypes.CDLL(str(libggml), winmode=0)
        # self.libggml = ctypes.WinDLL(str(libggml), winmode=0)
        self.libggml.ggml_quantize_chunk.restype = ctypes.c_size_t
        # enum ggml_type   type,
        #    const float * src,
        #           void * dst,
        #        int64_t   start,
        #        int64_t   nrows,
        #        int64_t   n_per_row,
        #    const float * imatrix) {
        self.libggml.ggml_quantize_chunk.argtypes = (
            ctypes.c_int,
            ctypes.POINTER(ctypes.c_float),
            ctypes.c_void_p,
            ctypes.c_int64,
            ctypes.c_int64,
            ctypes.c_int64,
            ctypes.POINTER(ctypes.c_float),
        )

        self.libggml.ggml_quantize_requires_imatrix.restype = ctypes.c_bool
        self.libggml.ggml_quantize_requires_imatrix.argtypes = (ctypes.c_int,)

        for t in (
            "q4_0", "q4_1", "q5_0", "q5_1", "q8_0",
            "q2_K", "q3_K", "q4_K", "q5_K", "q6_K",
            "tq1_0", "tq2_0",
            "iq2_xxs", "iq2_xs", "iq2_s", "iq3_xxs", "iq3_s", "iq1_s", "iq1_m",
            "iq4_nl", "iq4_xs",
        ):
            dequant_func: ctypes._NamedFuncPointer = getattr(self.libggml, "dequantize_row_" + t)
            dequant_func.restype = None
            dequant_func.argtypes = (ctypes.c_void_p, ctypes.POINTER(ctypes.c_float), ctypes.c_int64)

        self.libggml.ggml_fp16_to_fp32_row.restype = None
        self.libggml.ggml_fp16_to_fp32_row.argtypes = (ctypes.POINTER(ctypes.c_uint16), ctypes.POINTER(ctypes.c_float), ctypes.c_int64)
        self.libggml.ggml_bf16_to_fp32_row.restype = None
        self.libggml.ggml_bf16_to_fp32_row.argtypes = (ctypes.POINTER(ctypes.c_uint16), ctypes.POINTER(ctypes.c_float), ctypes.c_int64)

        self.libggml.ggml_init.argtypes = (ggml_init_params,)

        self.libggml.ggml_init(ggml_init_params(1 * 1024 * 1024, 0, False))

    def dequantize(self, tensor: np.ndarray, qtype: GGMLQuantizationType) -> np.ndarray:
        result = np.zeros(gguf.quant_shape_from_byte_shape(tensor.shape, qtype), dtype=np.float32, order="C")
        if qtype == GGMLQuantizationType.F32:
            # no-op
            result = tensor.view(np.float32)
        elif qtype == GGMLQuantizationType.F16:
            self.libggml.ggml_fp16_to_fp32_row(tensor.ctypes.data_as(ctypes.POINTER(ctypes.c_uint16)), result.ctypes.data_as(c_float_p), result.size)
        elif qtype == GGMLQuantizationType.BF16:
            self.libggml.ggml_bf16_to_fp32_row(tensor.ctypes.data_as(ctypes.POINTER(ctypes.c_uint16)), result.ctypes.data_as(c_float_p), result.size)
        else:
            lw_qname = qtype.name.lower()
            if lw_qname[-1] == "k":
                lw_qname = lw_qname[:-1] + "K"
            dequant_func: ctypes._NamedFuncPointer = getattr(self.libggml, "dequantize_row_" + lw_qname)
            dequant_func(tensor.ctypes.data_as(ctypes.c_void_p), result.ctypes.data_as(c_float_p), result.size)
        return result

    def quantize(self, data: np.ndarray, qtype: GGMLQuantizationType) -> np.ndarray:
        result = np.zeros(gguf.quant_shape_to_byte_shape(data.shape, qtype), dtype=np.uint8, order="C")
        if self.libggml.ggml_quantize_requires_imatrix(qtype.value):
            # TODO: is a column-wise sum of squares appropriate?
            qw = np.sum((data * data).reshape((-1, data.shape[-1])), axis=0).ctypes.data_as(c_float_p)
        else:
            qw = ctypes.cast(0, c_float_p)
        result_size = self.libggml.ggml_quantize_chunk(qtype.value, data.ctypes.data_as(c_float_p), result.ctypes.data_as(ctypes.c_void_p), 0, prod(data.shape[:-1]), data.shape[-1], qw)
        assert result.size == result_size
        return result


def create_sample(ggml_quants: GGMLQuants, hidden_size, qtype: GGMLQuantizationType) -> np.ndarray:
    gguf_writer = gguf.GGUFWriter(f"Quant_{qtype.name}_{hidden_size}.gguf", "llama")

    # Create a sample tensor
    for size in [768, 1024, 2048, 5120, 18944]:
        tensor = np.random.randn(size, hidden_size).astype(np.float32)
        shape_str = "x".join(map(str, tensor.shape))
        
        gguf_writer.add_tensor(f"tensor_{qtype.name}_{shape_str}", ggml_quants.quantize(tensor, qtype), raw_dtype=qtype)

    gguf_writer.write_header_to_file()
    gguf_writer.write_kv_data_to_file()
    gguf_writer.write_tensors_to_file()

    gguf_writer.close()


if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="Test Python (de)quantization against the reference C implementation")
    parser.add_argument("--libggml", type=Path, default="libggml.so", help="The path to libggml.so")
    parser.add_argument("--hidden_size", type=int, default=256, help="The hidden size of the sample tensor")
    parser.add_argument("--seed", type=int, default=0, help="The hidden size of the sample tensor")

    np.random.seed(0)

    args = parser.parse_args()

    logging.basicConfig(level=logging.DEBUG)

    ggml_quants = GGMLQuants(args.libggml)

    qtypes = [
        GGMLQuantizationType.IQ1_M,
        GGMLQuantizationType.IQ1_S,
        GGMLQuantizationType.IQ2_S,
        GGMLQuantizationType.IQ2_XS,
        GGMLQuantizationType.IQ2_XXS,
        GGMLQuantizationType.IQ3_S,
        GGMLQuantizationType.IQ3_XXS,
        GGMLQuantizationType.IQ4_NL,
        GGMLQuantizationType.IQ4_XS,
        GGMLQuantizationType.Q2_K,
        GGMLQuantizationType.Q3_K,
        GGMLQuantizationType.Q4_K,
        GGMLQuantizationType.Q5_K,
        GGMLQuantizationType.Q6_K,
        GGMLQuantizationType.Q4_0,
        GGMLQuantizationType.Q5_0,
        GGMLQuantizationType.Q8_0,
    ]

    for qtype in qtypes:
        create_sample(ggml_quants, args.hidden_size, qtype)