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
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
|
# mypy: allow-untyped-defs
import typing
import torch
__all__ = [
"ReferenceQuantizedModule",
]
class ReferenceQuantizedModule(torch.nn.Module):
def _init_weight_qparams(self, weight_qparams, device):
if weight_qparams is None:
weight_qparams = {
"qscheme": torch.per_tensor_affine,
"dtype": torch.quint8,
"scale": 1.0,
"zero_point": 0,
}
self.weight_qscheme: torch.qscheme = weight_qparams["qscheme"]
self.weight_dtype = weight_qparams["dtype"]
assert self.weight_qscheme in [
None,
torch.per_tensor_affine,
torch.per_channel_affine,
torch.per_channel_affine_float_qparams,
], f"qscheme: {self.weight_qscheme} is not support in reference quantized {self._get_name()}"
if self.weight_dtype in [
torch.quint8,
torch.qint8,
torch.quint4x2,
torch.qint32,
]:
zero_point_dtype = (
weight_qparams["zero_point"].dtype
if isinstance(weight_qparams["zero_point"], torch.Tensor)
else torch.int
)
w_scale = weight_qparams["scale"]
w_scale_tensor = (
w_scale.detach().clone()
if isinstance(w_scale, torch.Tensor)
else torch.tensor(w_scale, dtype=torch.float, device=device)
)
self.register_buffer("weight_scale", w_scale_tensor)
w_zp = weight_qparams["zero_point"]
w_zp_tensor = (
w_zp.detach().clone()
if isinstance(w_zp, torch.Tensor)
else torch.tensor(w_zp, dtype=zero_point_dtype, device=device)
)
self.register_buffer("weight_zero_point", w_zp_tensor)
if self.weight_qscheme in [
torch.per_channel_affine,
torch.per_channel_affine_float_qparams,
]:
w_axis = weight_qparams["axis"]
w_axis_tensor = (
w_axis.detach().clone()
if isinstance(w_axis, torch.Tensor)
else torch.tensor(w_axis, dtype=torch.int, device=device)
)
self.register_buffer("weight_axis", w_axis_tensor)
else:
# added for TorchScriptability, not used
self.register_buffer(
"weight_axis", torch.tensor(0, dtype=torch.int, device=device)
)
else:
# added for TorchScriptability, and for torch.float
self.register_buffer(
"weight_scale", torch.tensor(1.0, dtype=torch.float, device=device)
)
self.register_buffer(
"weight_zero_point", torch.tensor(0, dtype=torch.int, device=device)
)
self.register_buffer(
"weight_axis", torch.tensor(0, dtype=torch.int, device=device)
)
self.is_decomposed: bool = weight_qparams.get("is_decomposed", False)
# store weight_axis as weight_axis_int due to some constraints of torchdynamo.export
# for capturing `.item` operations
self.weight_axis_int: int = self.weight_axis.item() # type: ignore[operator, assignment]
self.weight_quant_min: typing.Optional[int] = weight_qparams.get(
"quant_min", None
)
self.weight_quant_max: typing.Optional[int] = weight_qparams.get(
"quant_max", None
)
def get_weight(self):
"""
Fake quantize (quantize and dequantize) the weight with
the quantization parameters for weight, this is used to
simulate the numerics for the quantized weight in a quantized
model
"""
# suppress mypy warning
assert isinstance(self.weight_scale, torch.Tensor)
assert isinstance(self.weight_zero_point, torch.Tensor)
if self.is_decomposed:
return _quantize_and_dequantize_weight_decomposed(
self.weight, # type: ignore[arg-type]
self.weight_qscheme,
self.weight_dtype,
self.weight_scale,
self.weight_zero_point,
self.weight_axis_int,
self.weight_quant_min,
self.weight_quant_max,
)
else:
return _quantize_and_dequantize_weight(
self.weight, # type: ignore[arg-type]
self.weight_qscheme,
self.weight_dtype,
self.weight_scale,
self.weight_zero_point,
self.weight_axis_int,
)
def get_quantized_weight(self):
# suppress mypy warning
assert isinstance(self.weight_scale, torch.Tensor)
assert isinstance(self.weight_zero_point, torch.Tensor)
# assert isinstance(self.weight_axis, torch.Tensor)
if self.is_decomposed:
return _quantize_weight_decomposed(
self.weight, # type: ignore[arg-type]
self.weight_qscheme,
self.weight_dtype,
self.weight_scale,
self.weight_zero_point,
self.weight_axis_int,
self.weight_quant_min,
self.weight_quant_max,
)
else:
return _quantize_weight(
self.weight, # type: ignore[arg-type]
self.weight_qscheme,
self.weight_dtype,
self.weight_scale,
self.weight_zero_point,
self.weight_axis_int,
)
def _save_to_state_dict(self, destination, prefix, keep_vars):
super()._save_to_state_dict(destination, prefix, keep_vars)
_save_weight_qparams(
destination,
prefix,
self.weight_qscheme,
self.weight_dtype,
self.weight_scale,
self.weight_zero_point,
self.weight_axis,
)
def _load_from_state_dict(
self,
state_dict,
prefix,
local_metadata,
strict,
missing_keys,
unexpected_keys,
error_msgs,
):
for key in _get_weight_qparam_keys(state_dict, prefix):
setattr(self, key, state_dict[prefix + key])
state_dict.pop(prefix + key)
super()._load_from_state_dict(
state_dict,
prefix,
local_metadata,
False,
missing_keys,
unexpected_keys,
error_msgs,
)
def _quantize_weight_decomposed(
weight: torch.Tensor,
weight_qscheme: torch.qscheme,
weight_dtype: torch.dtype,
weight_scale: torch.Tensor,
weight_zero_point: torch.Tensor,
weight_axis: int,
weight_quant_min: typing.Optional[int],
weight_quant_max: typing.Optional[int],
) -> torch.Tensor:
_DTYPE_TO_QVALUE_BOUNDS = {
torch.uint8: (0, 255),
torch.int8: (-128, 127),
torch.int32: (-(2**31), 2**31 - 1),
}
# TODO: add an util function for converting qdtype to dtype
_QDTYPE_TO_UNDERLYING_INT_REPR_DTYPE = {
torch.quint8: torch.uint8,
torch.qint8: torch.int8,
torch.qint32: torch.int32,
}
if weight_qscheme == torch.per_tensor_affine:
if weight_dtype in [torch.quint8, torch.qint8, torch.qint32]:
weight_dtype_ = _QDTYPE_TO_UNDERLYING_INT_REPR_DTYPE[weight_dtype]
if weight_quant_min is None or weight_quant_max is None:
weight_quant_min, weight_quant_max = _DTYPE_TO_QVALUE_BOUNDS[
weight_dtype_
]
weight = torch.ops.quantized_decomposed.quantize_per_tensor(
weight,
weight_scale,
weight_zero_point,
weight_quant_min,
weight_quant_max,
weight_dtype_,
)
return weight
elif weight_qscheme in [
torch.per_channel_affine,
torch.per_channel_affine_float_qparams,
]:
# TODO: torch.quint4x2 is not supported
if weight_dtype in [torch.quint8, torch.qint8, torch.qint32]:
weight_dtype_ = _QDTYPE_TO_UNDERLYING_INT_REPR_DTYPE[weight_dtype]
if weight_quant_min is None or weight_quant_max is None:
weight_quant_min, weight_quant_max = _DTYPE_TO_QVALUE_BOUNDS[
weight_dtype_
]
weight = torch.ops.quantized_decomposed.quantize_per_channel(
weight,
weight_scale,
weight_zero_point,
weight_axis,
weight_quant_min,
weight_quant_max,
weight_dtype_,
) # type: ignore[arg-type]
return weight
raise ValueError(f"Unsupported dtype and qscheme: {weight_dtype}, {weight_qscheme}")
def _dequantize_weight_decomposed(
weight: torch.Tensor,
weight_qscheme: torch.qscheme,
weight_dtype: torch.dtype,
weight_scale: torch.Tensor,
weight_zero_point: torch.Tensor,
weight_axis: int,
weight_quant_min: typing.Optional[int],
weight_quant_max: typing.Optional[int],
) -> torch.Tensor:
# TODO: get the quant_min and quant_max from activation_post_process
_DTYPE_TO_QVALUE_BOUNDS = {
torch.uint8: (0, 255),
torch.int8: (-128, 127),
torch.int32: (-(2**31), 2**31 - 1),
}
# TODO: add an util function for converting qdtype to dtype
_QDTYPE_TO_UNDERLYING_INT_REPR_DTYPE = {
torch.quint8: torch.uint8,
torch.qint8: torch.int8,
torch.qint32: torch.int32,
}
weight_dtype_ = _QDTYPE_TO_UNDERLYING_INT_REPR_DTYPE[weight_dtype]
if weight_quant_min is None or weight_quant_max is None:
weight_quant_min, weight_quant_max = _DTYPE_TO_QVALUE_BOUNDS[weight_dtype_]
if weight_qscheme == torch.per_tensor_affine:
if weight_dtype in [torch.quint8, torch.qint8, torch.qint32]:
weight = torch.ops.quantized_decomposed.dequantize_per_tensor(
weight,
weight_scale,
weight_zero_point,
weight_quant_min,
weight_quant_max,
weight_dtype_,
)
return weight
elif weight_qscheme in [
torch.per_channel_affine,
torch.per_channel_affine_float_qparams,
]:
# TODO: torch.quint4x2 is not supported
if weight_dtype in [torch.quint8, torch.qint8, torch.qint32]:
weight = torch.ops.quantized_decomposed.dequantize_per_channel(
weight,
weight_scale,
weight_zero_point,
weight_axis,
weight_quant_min,
weight_quant_max,
weight_dtype_,
) # type: ignore[arg-type]
return weight
raise ValueError(f"Unsupported dtype and qscheme: {weight_dtype}, {weight_qscheme}")
def _quantize_weight(
weight: torch.Tensor,
weight_qscheme: torch.qscheme,
weight_dtype: torch.dtype,
weight_scale: torch.Tensor,
weight_zero_point: torch.Tensor,
weight_axis_int: int,
) -> torch.Tensor:
if weight_dtype == torch.float16:
weight = weight.to(weight_dtype)
return weight
if weight_qscheme == torch.per_tensor_affine:
if weight_dtype in [torch.quint8, torch.qint8, torch.qint32]:
weight = torch.quantize_per_tensor(
weight, weight_scale, weight_zero_point, weight_dtype
)
return weight
elif weight_qscheme in [
torch.per_channel_affine,
torch.per_channel_affine_float_qparams,
]:
if weight_dtype in [torch.quint8, torch.qint8, torch.quint4x2, torch.qint32]:
weight = torch.quantize_per_channel(
weight, weight_scale, weight_zero_point, weight_axis_int, weight_dtype
) # type: ignore[arg-type]
return weight
raise ValueError(f"Unsupported dtype and qscheme: {weight_dtype}, {weight_qscheme}")
def _quantize_and_dequantize_weight_decomposed(
weight: torch.Tensor,
weight_qscheme: torch.qscheme,
weight_dtype: torch.dtype,
weight_scale: torch.Tensor,
weight_zero_point: torch.Tensor,
weight_axis_int: int,
weight_quant_min: typing.Optional[int],
weight_quant_max: typing.Optional[int],
) -> torch.Tensor:
"""Quantize and then dequantize the weight based on
the quantization parameters
"""
if weight_qscheme in [
torch.per_tensor_affine,
torch.per_channel_affine,
torch.per_channel_affine_float_qparams,
]:
weight_quant = _quantize_weight_decomposed(
weight,
weight_qscheme,
weight_dtype,
weight_scale,
weight_zero_point,
weight_axis_int,
weight_quant_min,
weight_quant_max,
)
weight_dequant = _dequantize_weight_decomposed(
weight_quant,
weight_qscheme,
weight_dtype,
weight_scale,
weight_zero_point,
weight_axis_int,
weight_quant_min,
weight_quant_max,
)
else:
weight_dequant = weight
return weight_dequant
def _quantize_and_dequantize_weight(
weight: torch.Tensor,
weight_qscheme: torch.qscheme,
weight_dtype: torch.dtype,
weight_scale: torch.Tensor,
weight_zero_point: torch.Tensor,
weight_axis_int: int,
) -> torch.Tensor:
"""Quantize and then dequantize the weight based on
the quantization parameters
"""
if weight_qscheme in [
torch.per_tensor_affine,
torch.per_channel_affine,
torch.per_channel_affine_float_qparams,
]:
weight_quant = _quantize_weight(
weight,
weight_qscheme,
weight_dtype,
weight_scale,
weight_zero_point,
weight_axis_int,
)
weight_dequant = weight_quant.dequantize()
else:
weight_dequant = weight
return weight_dequant
def _save_weight_qparams(
destination,
prefix,
weight_qscheme,
weight_dtype,
weight_scale,
weight_zero_point,
weight_axis,
):
destination[prefix + "weight_qscheme"] = weight_qscheme
destination[prefix + "weight_dtype"] = weight_dtype
if weight_qscheme is not None:
destination[prefix + "weight_scale"] = weight_scale
destination[prefix + "weight_zero_point"] = weight_zero_point
if weight_qscheme == torch.per_channel_affine:
destination[prefix + "weight_axis"] = weight_axis
def _get_weight_qparam_keys(state_dict: typing.Dict[str, typing.Any], prefix: str):
keys = ["weight_qscheme", "weight_dtype"]
weight_qscheme = state_dict[prefix + "weight_qscheme"]
if weight_qscheme is not None:
keys.append("weight_scale")
keys.append("weight_zero_point")
if weight_qscheme == torch.quantize_per_channel:
keys.append("weight_axis")
return keys
|
