-
Notifications
You must be signed in to change notification settings - Fork 5
/
register_lib.py
192 lines (156 loc) · 8.6 KB
/
register_lib.py
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
import torch
import torch._custom_ops
from torch import Tensor
import fast_hadamard_transform_cuda
import quiptools_cuda
my_lib = torch.library.Library("quip_lib", "DEF")
@torch._custom_ops.custom_op("quip_lib::hadamard")
def hadamard(x: Tensor, scale: float) -> Tensor:
raise NotImplementedError()
@torch._custom_ops.impl_abstract("quip_lib::hadamard")
def hadamard_abstract(x: Tensor, scale: float) -> Tensor:
return x
@torch._custom_ops.impl("quip_lib::hadamard", device_types="cuda")
def hadamard_cuda(x: Tensor, scale: float) -> Tensor:
return fast_hadamard_transform_cuda.fast_hadamard_transform(x, scale)
@torch._custom_ops.custom_op("quip_lib::e8p_mm_origorder")
def e8p_mm_origorder(x: Tensor, Qidxs: Tensor, grid: Tensor) -> Tensor:
raise NotImplementedError()
@torch._custom_ops.impl_abstract("quip_lib::e8p_mm_origorder")
def e8p_mm_origorder_abstract(x: Tensor, Qidxs: Tensor, grid: Tensor) -> Tensor:
assert x.dim() == 2
assert Qidxs.dim() == 2
assert x.shape[1] == Qidxs.shape[1] * 8
assert x.device == Qidxs.device
assert x.device == grid.device
result = x.new_empty((x.shape[0], Qidxs.shape[0]), dtype=x.dtype)
return result
@torch._custom_ops.impl("quip_lib::e8p_mm_origorder", device_types="cuda")
def e8p_mm_origorder_cuda(x: Tensor, Qidxs: Tensor, grid: Tensor) -> Tensor:
return quiptools_cuda.e8p_mm_origorder(x, Qidxs, grid)
@torch._custom_ops.custom_op("quip_lib::e8prvq3_mm_origorder")
def e8prvq3_mm_origorder(x: Tensor, Qidxs: Tensor, grid: Tensor, grid2: Tensor, scale: float) -> Tensor:
raise NotImplementedError()
@torch._custom_ops.impl_abstract("quip_lib::e8prvq3_mm_origorder")
def e8prvq3_mm_origorder_abstract(x: Tensor, Qidxs: Tensor, grid: Tensor, grid2: Tensor, scale: float) -> Tensor:
assert x.dim() == 2
assert Qidxs.dim() == 2
assert x.device == Qidxs.device
assert x.device == grid.device
assert x.device == grid2.device
result = x.new_empty((x.shape[0], Qidxs.shape[0]), dtype=x.dtype)
return result
@torch._custom_ops.impl("quip_lib::e8prvq3_mm_origorder", device_types="cuda")
def e8prvq3_mm_origorder_cuda(x: Tensor, Qidxs: Tensor, grid: Tensor, grid2: Tensor, scale: float) -> Tensor:
return quiptools_cuda.e8prvq3_mm_origorder(x, Qidxs, grid, grid2, scale)
@torch._custom_ops.custom_op("quip_lib::e8prvq4_mm_origorder")
def e8prvq4_mm_origorder(x: Tensor, Qidxs: Tensor, grid: Tensor, scale: float) -> Tensor:
raise NotImplementedError()
@torch._custom_ops.impl_abstract("quip_lib::e8prvq4_mm_origorder")
def e8prvq4_mm_origorder_abstract(x: Tensor, Qidxs: Tensor, grid: Tensor, scale: float) -> Tensor:
assert x.dim() == 2
assert Qidxs.dim() == 2
assert x.device == Qidxs.device
assert x.device == grid.device
result = x.new_empty((x.shape[0], Qidxs.shape[0]), dtype=x.dtype)
return result
@torch._custom_ops.impl("quip_lib::e8prvq4_mm_origorder", device_types="cuda")
def e8p_mm_origorder_cuda(x: Tensor, Qidxs: Tensor, grid: Tensor, scale: float) -> Tensor:
return quiptools_cuda.e8prvq4_mm_origorder(x, Qidxs, grid, scale)
@torch._custom_ops.custom_op("quip_lib::d4_mm_origorder")
def d4_mm_origorder(x: Tensor, Qidxs: Tensor, grid: Tensor) -> Tensor:
raise NotImplementedError()
@torch._custom_ops.impl_abstract("quip_lib::d4_mm_origorder")
def d4_mm_origorder_abstract(x: Tensor, Qidxs: Tensor, grid: Tensor) -> Tensor:
assert x.dim() == 2
assert Qidxs.dim() == 2
assert x.device == Qidxs.device
assert x.device == grid.device
result = x.new_empty((x.shape[0], Qidxs.shape[0]), dtype=x.dtype)
return result
@torch._custom_ops.impl("quip_lib::d4_mm_origorder", device_types="cuda")
def d4_mm_origorder_cuda(x: Tensor, Qidxs: Tensor, grid: Tensor) -> Tensor:
return quiptools_cuda.d4_mm_origorder(x, Qidxs, grid)
@torch._custom_ops.custom_op("quip_lib::hi_mm_origorder")
def hi_mm_origorder(x: Tensor, Qidxs: Tensor) -> Tensor:
raise NotImplementedError()
@torch._custom_ops.impl_abstract("quip_lib::hi_mm_origorder")
def hi_mm_origorder_abstract(x: Tensor, Qidxs: Tensor) -> Tensor:
assert x.dim() == 2
assert Qidxs.dim() == 2
assert x.device == Qidxs.device
result = x.new_empty((x.shape[0], Qidxs.shape[0]), dtype=x.dtype)
return result
@torch._custom_ops.impl("quip_lib::hi_mm_origorder", device_types="cuda")
def hi_mm_origorder_cuda(x: Tensor, Qidxs: Tensor) -> Tensor:
return quiptools_cuda.hi_mm_origorder(x, Qidxs)
@torch._custom_ops.custom_op("quip_lib::decompress_e8p_origorder")
def decompress_e8p_origorder(Qidxs: Tensor, grid: Tensor) -> Tensor:
raise NotImplementedError()
@torch._custom_ops.impl_abstract("quip_lib::decompress_e8p_origorder")
def decompress_e8p_origorder_abstract(Qidxs: Tensor, grid: Tensor) -> Tensor:
assert Qidxs.dim() == 2
result = Qidxs.new_empty((Qidxs.shape[0], Qidxs.shape[1] * 8), dtype=torch.float16)
return result
@torch._custom_ops.impl("quip_lib::decompress_e8p_origorder", device_types="cuda")
def decompress_e8p_origorder_cuda(Qidxs: Tensor, grid: Tensor) -> Tensor:
x_dequant = torch.empty((Qidxs.shape[0], Qidxs.shape[1] * 8),
dtype=torch.float16, device=Qidxs.device)
quiptools_cuda.decompress_e8p_origorder(Qidxs, grid, x_dequant)
return x_dequant
@torch._custom_ops.custom_op("quip_lib::decompress_e8prvq3_origorder")
def decompress_e8prvq3_origorder(Qidxs: Tensor, grid: Tensor, grid2: Tensor, scale: float) -> Tensor:
raise NotImplementedError()
@torch._custom_ops.impl_abstract("quip_lib::decompress_e8prvq3_origorder")
def decompress_e8prvq3_origorder_abstract(Qidxs: Tensor, grid: Tensor, grid2: Tensor, scale: float) -> Tensor:
assert Qidxs.dim() == 2
result = Qidxs.new_empty((Qidxs.shape[0], Qidxs.shape[1] * 32 // 3), dtype=torch.float16)
return result
@torch._custom_ops.impl("quip_lib::decompress_e8prvq3_origorder", device_types="cuda")
def decompress_e8prvq3_origorder_cuda(Qidxs: Tensor, grid: Tensor, grid2: Tensor, scale: float) -> Tensor:
x_dequant = torch.empty((Qidxs.shape[0], Qidxs.shape[1] * 32 // 3),
dtype=torch.float16, device=Qidxs.device)
quiptools_cuda.decompress_e8prvq3_origorder(Qidxs, grid, grid2, x_dequant, scale)
return x_dequant
@torch._custom_ops.custom_op("quip_lib::decompress_e8prvq4_origorder")
def decompress_e8prvq4_origorder(Qidxs: Tensor, grid: Tensor, scale: float) -> Tensor:
raise NotImplementedError()
@torch._custom_ops.impl_abstract("quip_lib::decompress_e8prvq4_origorder")
def decompress_e8prvq4_origorder_abstract(Qidxs: Tensor, grid: Tensor, scale: float) -> Tensor:
assert Qidxs.dim() == 2
result = Qidxs.new_empty((Qidxs.shape[0], Qidxs.shape[1] * 8), dtype=torch.float16)
return result
@torch._custom_ops.impl("quip_lib::decompress_e8prvq4_origorder", device_types="cuda")
def decompress_e8prvq4_origorder_cuda(Qidxs: Tensor, grid: Tensor, scale: float) -> Tensor:
x_dequant = torch.empty((Qidxs.shape[0], Qidxs.shape[1] * 8),
dtype=torch.float16, device=Qidxs.device)
quiptools_cuda.decompress_e8prvq4_origorder(Qidxs, grid, x_dequant, scale)
return x_dequant
@torch._custom_ops.custom_op("quip_lib::decompress_d4_origorder")
def decompress_d4_origorder(Qidxs: Tensor, grid: Tensor) -> Tensor:
raise NotImplementedError()
@torch._custom_ops.impl_abstract("quip_lib::decompress_d4_origorder")
def decompress_d4_origorder_abstract(Qidxs: Tensor, grid: Tensor) -> Tensor:
assert Qidxs.dim() == 2
result = Qidxs.new_empty((Qidxs.shape[0], Qidxs.shape[1] * 4), dtype=torch.float16)
return result
@torch._custom_ops.impl("quip_lib::decompress_d4_origorder", device_types="cuda")
def decompress_d4_origorder_cuda(Qidxs: Tensor, grid: Tensor) -> Tensor:
x_dequant = torch.empty((Qidxs.shape[0], Qidxs.shape[1] * 4),
dtype=torch.float16, device=Qidxs.device)
quiptools_cuda.decompress_d4_origorder(Qidxs, grid, x_dequant)
return x_dequant
@torch._custom_ops.custom_op("quip_lib::decompress_hi_origorder")
def decompress_hi_origorder(Qidxs: Tensor) -> Tensor:
raise NotImplementedError()
@torch._custom_ops.impl_abstract("quip_lib::decompress_hi_origorder")
def decompress_hi_origorder_abstract(Qidxs: Tensor) -> Tensor:
assert Qidxs.dim() == 2
result = Qidxs.new_empty((Qidxs.shape[0], Qidxs.shape[1] * 8), dtype=torch.float16)
return result
@torch._custom_ops.impl("quip_lib::decompress_hi_origorder", device_types="cuda")
def decompress_hi_origorder_cuda(Qidxs: Tensor) -> Tensor:
x_dequant = torch.empty((Qidxs.shape[0], Qidxs.shape[1] * 8),
dtype=torch.float16, device=Qidxs.device)
quiptools_cuda.decompress_hi_origorder(Qidxs, x_dequant)
return x_dequant