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
|
# mypy: allow-untyped-defs
from typing import Optional
import torch
import torch.ao.nn.intrinsic as nni
from torch.ao.nn.quantized.modules.utils import (
_hide_packed_params_repr,
_quantize_weight,
)
from torch.ao.nn.sparse.quantized import linear
from torch.ao.nn.sparse.quantized.utils import LinearBlockSparsePattern
__all__ = ["Linear"]
class Linear(torch.nn.Module):
r"""
A dynamically quantized sparse linear module with float tensor as inputs and outputs.
"""
_version = 1
_op_type = "sparse_dynamic"
_FLOAT_MODULE = torch.nn.Linear
def __init__(
self,
in_features,
out_features,
row_block_size,
col_block_size,
bias=True,
dtype=torch.qint8,
):
super().__init__()
if dtype != torch.qint8:
raise NotImplementedError(
"Only QINT8 is supported for Sparse Quantized Linear Dynamic"
)
self.in_features = in_features
self.out_features = out_features
if bias:
bias = torch.zeros(self.out_features, dtype=torch.float)
else:
bias = None
qweight = torch._empty_affine_quantized(
[out_features, in_features], scale=1, zero_point=0, dtype=torch.qint8
)
self._packed_params = linear.LinearPackedParams(
row_block_size=row_block_size, col_block_size=col_block_size, dtype=dtype
)
self._packed_params.set_weight_bias(
qweight, bias, row_block_size, col_block_size
)
def _get_name(self):
return "SparseQuantizedDynamicLinear"
def extra_repr(self):
return f"in_features={self.in_features}, out_features={self.out_features}, qscheme={self.weight().qscheme()}"
def __repr__(self):
return _hide_packed_params_repr(self, linear.LinearPackedParams)
def forward(self, x: torch.Tensor) -> torch.Tensor:
return torch.ops.sparse.qlinear_dynamic(x, self._packed_params._packed_params)
def _save_to_state_dict(self, destination, prefix, keep_vars):
super()._save_to_state_dict(destination, prefix, keep_vars)
destination[prefix + "op_type"] = self._op_type
def _load_from_state_dict(
self,
state_dict,
prefix,
local_metadata,
strict,
missing_keys,
unexpected_keys,
error_msgs,
):
op_type = int(state_dict[prefix + "op_type"])
assert (
op_type == "sparse"
), f"Cannot load from op_type [{op_type}], expecting [{self._op_type}]"
state_dict.pop(prefix + "op_type")
version = local_metadata.get("version", None)
assert version <= self._version
# Is this code valid? In old quantization it seemed to be used to load
# older model
weight = state_dict.pop(prefix + "weight")
bias = state_dict.pop(prefix + "bias")
state_dict.update(
{
prefix + "_packed_params.weight": weight,
prefix + "_packed_params.bias": bias,
}
)
super()._load_from_state_dict(
state_dict,
prefix,
local_metadata,
False,
missing_keys,
unexpected_keys,
error_msgs,
)
def _weight_bias(self):
return self._packed_params._weight_bias()
def weight(self):
return self._weight_bias()[0]
def bias(self):
return self._weight_bias()[1]
def set_weight_bias(
self,
w: torch.Tensor,
b: Optional[torch.Tensor],
row_block_size: Optional[int],
col_block_size: Optional[int],
) -> None:
assert row_block_size is not None and col_block_size is not None
self.out_features = w.shape[0]
self.in_features = w.shape[1]
self._packed_params.set_weight_bias(w, b, row_block_size, col_block_size)
@classmethod
def from_float(cls, mod, use_precomputed_fake_quant=False):
r"""Create a quantized sparse dynamic module from a float module.
We only care about the convert at this stage, no need for observers just yet.
"""
assert type(mod) == cls._FLOAT_MODULE, (
" nnq."
+ cls.__name__
+ ".from_float only works for "
+ cls._FLOAT_MODULE.__name__
)
# TODO: Need to add options to qconfig to avoid the calibration.
# TODO: Add calibration for the sparsity
assert hasattr(mod, "qconfig"), "Input float module must have qconfig defined"
if type(mod) == nni.LinearReLU:
mod = mod[0]
if mod.qconfig is not None and mod.qconfig.weight is not None:
weight_observer = mod.qconfig.weight()
else:
# We have the circular import issues if we import the qconfig in the beginning of this file:
# https://github.com/pytorch/pytorch/pull/24231. The current workaround is to postpone the
# import until we need it.
from torch.ao.quantization.qconfig import default_dynamic_qconfig
weight_observer = default_dynamic_qconfig.weight()
# It is important to multiply by the mask BEFORE calling the `weight_observer`
# TODO (zaf): Mask might not be part of the qconfig (T83295194)
weight = mod.weight
if getattr(mod.qconfig, "mask", False):
weight = mod.qconfig.mask * mod.weight
weight_observer(weight)
dtype = weight_observer.dtype
assert dtype == torch.qint8, "Weight observer must have dtype torch.qint8"
w_sc, w_zp = weight_observer.calculate_qparams()
if isinstance(w_zp, torch.Tensor):
assert not torch.any(w_zp.bool()), "All weight zero points must map to 0"
else:
assert w_zp == 0, "Weight zero point must map to 0"
qweight = _quantize_weight(weight.float(), weight_observer)
row_block_size, col_block_size = LinearBlockSparsePattern.block_size()
qlinear = cls(
mod.in_features,
mod.out_features,
row_block_size,
col_block_size,
dtype=dtype,
)
qlinear.set_weight_bias(qweight, mod.bias, row_block_size, col_block_size)
return qlinear
|
