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
|
from typing import Optional
import torch
from torch.ao.nn.quantized.modules.utils import _quantize_weight, hide_packed_params_repr
__all__ = ['LinearPackedParams', 'Linear']
# TODO (zaf): Inherit from `quantized.LinearPackedParams` (T83294430)
class LinearPackedParams(torch.nn.Module):
_version = 1
def __init__(self, row_block_size=1, col_block_size=4, dtype=torch.qint8):
super().__init__()
if dtype != torch.qint8:
raise NotImplementedError("Linear prepacking only supports QINT8")
self.dtype = dtype
wq = torch._empty_affine_quantized([1, 1], scale=1.0, zero_point=0, dtype=torch.qint8)
self.set_weight_bias(wq, None, row_block_size, col_block_size)
def _get_name(self):
return "SparseQuantizedLinearPackedParams"
@torch.jit.export
def set_weight_bias(self, weight: torch.Tensor, bias: 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._packed_params = torch.ops.sparse.qlinear_prepack(weight, bias, row_block_size, col_block_size)
@torch.jit.export
def _weight_bias(self):
(weight, bias, block_sizes) = torch.ops.sparse.qlinear_unpack(self._packed_params)
return (weight, bias, block_sizes[0], block_sizes[1])
def forward(self, x):
return x
def _save_to_state_dict(self, destination, prefix, keep_vars):
super()._save_to_state_dict(destination, prefix, keep_vars)
destination[prefix + 'dtype'] = self.dtype
destination[prefix + '_packed_params'] = self._weight_bias()
def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict,
missing_keys, unexpected_keys, error_msgs):
version = local_metadata.get('version', None)
assert version <= self._version
self.dtype = state_dict.pop(prefix + 'dtype')
weight, bias, row_block_size, col_block_size = state_dict.pop(prefix + '_packed_params')
self.set_weight_bias(weight, bias, row_block_size, col_block_size)
super()._load_from_state_dict(state_dict, prefix, local_metadata, False,
missing_keys, unexpected_keys, error_msgs)
@torch.jit.export
def __getstate__(self):
return self._packed_params, self.training, self.dtype
@torch.jit.export
def __setstate__(self, state):
(self._packed_params, self.training, self.dtype) = state
def __repr__(self):
return self._weight_bias().__repr__()
# TODO (zaf): Inherit from `quantized.Linear` (T83294430)
class Linear(torch.nn.Module):
r"""
A quantized sparse linear module with quantized tensor as inputs and outputs.
"""
_version = 1
_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")
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 = 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)
self.scale = 1.0
self.zero_point = 0
@classmethod
def _get_name(cls):
return 'SparseQuantizedLinear'
def extra_repr(self):
return 'in_features={}, out_features={}, scale={}, zero_point={}, qscheme={}'.format(
self.in_features, self.out_features, self.scale, self.zero_point, self.weight().qscheme()
)
def __repr__(self):
return hide_packed_params_repr(self, LinearPackedParams)
def forward(self, x: torch.Tensor) -> torch.Tensor:
return torch.ops.sparse.qlinear(x, self._packed_params._packed_params, self.scale, self.zero_point)
def _save_to_state_dict(self, destination, prefix, keep_vars):
super()._save_to_state_dict(destination, prefix, keep_vars)
destination[prefix + 'scale'] = torch.tensor(self.scale)
destination[prefix + 'zero_point'] = torch.tensor(self.zero_point)
def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict,
missing_keys, unexpected_keys, error_msgs):
self.scale = float(state_dict[prefix + 'scale'])
state_dict.pop(prefix + 'scale')
self.zero_point = int(state_dict[prefix + 'zero_point'])
state_dict.pop(prefix + 'zero_point')
op_type = int(state_dict[prefix + 'op_type'])
state_dict.pop(prefix + 'op_type')
version = local_metadata.get('version', None)
assert version <= self._version
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._packed_params.set_weight_bias(w, b, row_block_size, col_block_size)
@classmethod
def from_float(cls, mod):
r"""Create a quantized sparse module from a float module.
We only care about the convert at this stage, no need for observers just yet.
TODO(zaf): Need to add the sparse params to the qconfig
"""
assert type(mod) == cls._FLOAT_MODULE, cls._get_name() + \
'.from_float only works for ' + cls._FLOAT_MODULE.__name__
assert hasattr(mod, 'sparse_params'), \
('Expecting the Linear to have `sparse_params`. Make sure you have provided arguments '
'in the `sparsifier.squash_mask(params_to_save=("sparse_block_shape",))` method.')
sparse_block_shape = mod.sparse_params.get('sparse_block_shape', None) # type: ignore[operator, union-attr]
assert isinstance(sparse_block_shape, (tuple, list))
assert len(sparse_block_shape) == 2
# 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'
activation_post_process = mod.activation_post_process
weight_post_process = mod.qconfig.weight() # type: ignore[operator, union-attr]
# Assumption is that the weight is already sparsified by the
# `sparsifier.convert`
weight = mod.weight
weight_post_process(weight)
dtype = weight_post_process.dtype
act_scale, act_zp = activation_post_process.calculate_qparams() # type: ignore[operator, union-attr]
assert dtype == torch.qint8, 'Weight observer must have dtype torch.qint8'
w_sc, w_zp = weight_post_process.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_post_process)
row_block_size = mod.sparse_params['sparse_block_shape'][0] # type: ignore[index]
col_block_size = mod.sparse_params['sparse_block_shape'][1] # type: ignore[index]
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) # type: ignore[arg-type]
qlinear.scale = float(act_scale)
qlinear.zero_point = int(act_zp)
return qlinear
|
