File size: 7,789 Bytes
96ba80e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Union
import tqdm
import numpy as np
import pdb
import math
from torch import Tensor

CLIPMIN = 1e-5


class SimpleRMSNorm(torch.nn.Module):
    """
    This class implements the Root Mean Square Normalization (RMSN) layer.
    We use the implementation from LLAMARMSNorm here:
    https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/modeling_llama.py#L75
    """

    def __init__(self, mean_dim, eps=1e-5):
        super().__init__()
        self.eps = eps
        self.mean_dim = mean_dim

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        input_dtype = x.dtype
        if x.dtype == torch.float16 or x.dtype == torch.bfloat16:
            x = x.to(torch.float32)
        variance = x.pow(2).sum(-1, keepdim=True) / self.mean_dim
        x = x * torch.rsqrt(variance + self.eps)
        return x.to(input_dtype)



def activation_quant(x: Tensor,quant_rate=1.0):
    """Per token quantization to 8bits. No grouping is needed for quantization

    Args:
        x (Tensor): _description_

    Returns:
        _type_: _description_
    """
    scale = 127.0 / x.abs().max(dim=-1, keepdim=True).values.clamp_(min=1e-5)
    y = (x * scale).round().clamp_(-128, 127) / scale

    x_quant =  x + (y - x).detach()
    return x_quant



def round_ste(x: torch.Tensor):
    """
    Implement Straight-Through Estimator for rounding operation.
    """
    return (x.round() - x).detach() + x

def clamp(value, min_value, max_value):
    return max(min_value, min(value, max_value))


class UniformAffineQuantizer(nn.Module):
    def __init__(
        self,
        n_bits: int = 8,
        symmetric: bool = False,
        per_channel_axes=[],
        metric="minmax",
        dynamic=False,
        dynamic_method="per_cluster",
        group_size=None,
        shape=None,
        lwc=False,
        disable_zero_point=False,
        is_weight_quant=False,
        **kwargs, 
    ):
        """
        support cluster quantize
        dynamic_method support per_token and per_cluster
        """
        super().__init__()
        self.symmetric = symmetric
        self.disable_zero_point = disable_zero_point
        assert 1 <= n_bits <= 16, "bitwidth not supported"
        self.n_bits = n_bits
        if self.disable_zero_point:
            self.qmin = -(2 ** (n_bits - 1))
            self.qmax = 2 ** (n_bits - 1) - 1
        else:
            self.qmin = 0
            self.qmax = 2 ** (n_bits) - 1
        self.per_channel_axes = per_channel_axes
        self.metric = metric
        self.cluster_counts = None
        self.cluster_dim = None

        self.scale = None
        self.zero_point = None
        self.round_zero_point = None

        self.cached_xmin = None
        self.cached_xmax = None
        self.dynamic = dynamic
        self.dynamic_method = dynamic_method
        self.deficiency = 0
        self.lwc = lwc
        self.is_weight_quant = is_weight_quant
        self.shape = shape
        
        init_value = 4.             # inti value of learnable weight clipping
        if lwc:
            if group_size:
                dim1 = int(self.shape[0]*math.ceil(self.shape[1]/group_size))
                self.deficiency = shape[-1]%group_size
                if self.deficiency > 0:
                    self.deficiency = group_size - self.deficiency
                    assert self.symmetric   # support for mlc-llm symmetric quantization
            else:
                dim1 = self.shape[0]
            self.upbound_factor = nn.Parameter(torch.ones((dim1,1))*init_value)
            self.lowbound_factor = nn.Parameter(torch.ones((dim1,1))*init_value)
        self.sigmoid = nn.Sigmoid()

        self.enable = True
        self.group_size = group_size

    def change_n_bits(self, n_bits):
        self.n_bits = n_bits
        if self.disable_zero_point:
            self.qmin = -(2 ** (n_bits - 1))
            self.qmax = 2 ** (n_bits - 1) - 1
        else:
            self.qmin = 0
            self.qmax = 2 ** (n_bits) - 1

    def fake_quant(self, x, scale, round_zero_point):
        if self.deficiency > 0:
            pad_zeros = torch.zeros((x.shape[0],self.deficiency),dtype=x.dtype,device=x.device)
            x = torch.cat((x,pad_zeros),dim=1)
        
        if self.group_size:
            assert len(x.shape)==2, "only support linear layer now"
            dim1, dim2 = x.shape
            x = x.reshape(-1, self.group_size)

        x = round_ste(x / scale)
        if round_zero_point is not None:
            x = x.add(round_zero_point)
        x = x.clamp(self.qmin, self.qmax)
        if round_zero_point is not None:
            x = x.sub(round_zero_point)
        x = x.mul(scale)
        if self.group_size:
            x = x.reshape(dim1, dim2)
        if self.deficiency > 0:
            x = x[:,:-self.deficiency]

        return x
    

    def forward(self, x: torch.Tensor,quant_rate=1.0):


        if self.n_bits >= 16 or not self.enable:
            return x
        if self.metric == "fix0to1":
            return x.mul_(2**self.n_bits-1).round_().div_(2**self.n_bits-1)

        if self.dynamic_method == "per_token" or self.dynamic_method == "per_channel":
            self.per_token_dynamic_calibration(x)
        else:
            raise NotImplementedError()   
        
        # import ipdb;ipdb.set_trace()
        scale_dim = self.scale.shape[0]
        if self.group_size:
            scale_quant_dim_size =  clamp(math.ceil(self.scale.shape[0] * quant_rate),0,scale_dim)
        else:
            scale_quant_dim_size = scale_dim

        if quant_rate < 0.99:
            x_dim_size = x.shape[-1]
            quant_dim_size =  clamp(math.ceil(x.shape[-1] * quant_rate),0,x_dim_size)
            quant_x = self.fake_quant(x[...,:quant_dim_size], self.scale[:scale_quant_dim_size], self.round_zero_point[:scale_quant_dim_size])
            non_quant_x = x[..., quant_dim_size:]
            x = torch.cat((quant_x, non_quant_x), dim=-1)
        else:
            x = self.fake_quant(x, self.scale, self.round_zero_point)

        return x
    
    def quantize(self, x: torch.Tensor):
        return self.forward(x)
    
    def ready(self):
        return True

    def per_token_dynamic_calibration(self, x):
        if self.group_size:
            if self.deficiency == 0:
                x = x.reshape(-1,self.group_size)
            else:
                pad_zeros = torch.zeros((x.shape[0],self.deficiency),dtype=x.dtype,device=x.device)
                x = torch.cat((x,pad_zeros),dim=1)
                x = x.reshape(-1,self.group_size)
        reduce_shape = [-1]
        xmin = x.amin(reduce_shape, keepdim=True)
        xmax =  x.amax(reduce_shape, keepdim=True)
        if self.lwc:
            xmax = self.sigmoid(self.upbound_factor)*xmax
            xmin = self.sigmoid(self.lowbound_factor)*xmin
        if self.symmetric:
            abs_max = torch.max(xmax.abs(),xmin.abs())
            scale = abs_max / (2**(self.n_bits-1)-1)
            self.scale = scale.clamp(min=CLIPMIN, max=1e4)
            zero_point = (2**(self.n_bits-1)-1)*torch.ones_like(self.scale)
        else:
            range = xmax - xmin
            scale = range / (2**self.n_bits-1)
            self.scale = scale.clamp(min=CLIPMIN, max=1e4)
            zero_point = -(xmin) / (self.scale)
        if self.disable_zero_point:
            self.round_zero_point = None
        else:
            self.round_zero_point = zero_point.clamp(min=-1e4, max=1e4).round()
        
    def register_scales_and_zeros(self):
        self.register_buffer('scales', self.scale)
        self.register_buffer('zeros', self.round_zero_point)
        del self.scale
        del self.round_zero_point