compressed-tensors-nightly 0.8.0.20241202__py3-none-any.whl → 0.8.0.20241203__py3-none-any.whl
Sign up to get free protection for your applications and to get access to all the features.
- compressed_tensors/quantization/quant_args.py +16 -3
- compressed_tensors/quantization/quant_scheme.py +1 -0
- compressed_tensors/utils/helpers.py +33 -1
- {compressed_tensors_nightly-0.8.0.20241202.dist-info → compressed_tensors_nightly-0.8.0.20241203.dist-info}/METADATA +1 -1
- {compressed_tensors_nightly-0.8.0.20241202.dist-info → compressed_tensors_nightly-0.8.0.20241203.dist-info}/RECORD +8 -8
- {compressed_tensors_nightly-0.8.0.20241202.dist-info → compressed_tensors_nightly-0.8.0.20241203.dist-info}/LICENSE +0 -0
- {compressed_tensors_nightly-0.8.0.20241202.dist-info → compressed_tensors_nightly-0.8.0.20241203.dist-info}/WHEEL +0 -0
- {compressed_tensors_nightly-0.8.0.20241202.dist-info → compressed_tensors_nightly-0.8.0.20241203.dist-info}/top_level.txt +0 -0
@@ -17,6 +17,7 @@ from enum import Enum
|
|
17
17
|
from typing import Any, Dict, Optional, Union
|
18
18
|
|
19
19
|
import torch
|
20
|
+
from compressed_tensors.utils import Aliasable
|
20
21
|
from pydantic import BaseModel, Field, field_validator, model_validator
|
21
22
|
|
22
23
|
|
@@ -53,17 +54,29 @@ class QuantizationStrategy(str, Enum):
|
|
53
54
|
TOKEN = "token"
|
54
55
|
|
55
56
|
|
56
|
-
class ActivationOrdering(str, Enum):
|
57
|
+
class ActivationOrdering(Aliasable, str, Enum):
|
57
58
|
"""
|
58
59
|
Enum storing strategies for activation ordering
|
59
60
|
|
60
61
|
Group: reorder groups and weight\n
|
61
|
-
Weight: only reorder weight, not groups. Slightly lower
|
62
|
-
|
62
|
+
Weight: only reorder weight, not groups. Slightly lower accuracy but also lower
|
63
|
+
latency when compared to group actorder\n
|
64
|
+
Dynamic: alias for Group\n
|
65
|
+
Static: alias for Weight\n
|
63
66
|
"""
|
64
67
|
|
65
68
|
GROUP = "group"
|
66
69
|
WEIGHT = "weight"
|
70
|
+
# aliases
|
71
|
+
DYNAMIC = "dynamic"
|
72
|
+
STATIC = "static"
|
73
|
+
|
74
|
+
@staticmethod
|
75
|
+
def get_aliases() -> Dict[str, str]:
|
76
|
+
return {
|
77
|
+
"dynamic": "group",
|
78
|
+
"static": "weight",
|
79
|
+
}
|
67
80
|
|
68
81
|
|
69
82
|
class QuantizationArgs(BaseModel, use_enum_values=True):
|
@@ -12,7 +12,7 @@
|
|
12
12
|
# See the License for the specific language governing permissions and
|
13
13
|
# limitations under the License.
|
14
14
|
|
15
|
-
from typing import Any, Optional
|
15
|
+
from typing import Any, Dict, Optional
|
16
16
|
|
17
17
|
import torch
|
18
18
|
from transformers import AutoConfig
|
@@ -24,6 +24,7 @@ __all__ = [
|
|
24
24
|
"tensor_follows_mask_structure",
|
25
25
|
"replace_module",
|
26
26
|
"is_compressed_tensors_config",
|
27
|
+
"Aliasable",
|
27
28
|
]
|
28
29
|
|
29
30
|
FSDP_WRAPPER_NAME = "_fsdp_wrapped_module"
|
@@ -119,3 +120,34 @@ def is_compressed_tensors_config(compression_config: Any) -> bool:
|
|
119
120
|
return isinstance(compression_config, CompressedTensorsConfig)
|
120
121
|
except ImportError:
|
121
122
|
return False
|
123
|
+
|
124
|
+
|
125
|
+
class Aliasable:
|
126
|
+
"""
|
127
|
+
A mixin for enums to allow aliasing of enum members
|
128
|
+
|
129
|
+
Example:
|
130
|
+
>>> class MyClass(Aliasable, int, Enum):
|
131
|
+
>>> ...
|
132
|
+
"""
|
133
|
+
|
134
|
+
@staticmethod
|
135
|
+
def get_aliases() -> Dict[str, str]:
|
136
|
+
raise NotImplementedError()
|
137
|
+
|
138
|
+
def __eq__(self, other):
|
139
|
+
if isinstance(other, self.__class__):
|
140
|
+
aliases = self.get_aliases()
|
141
|
+
return self.value == other.value or (
|
142
|
+
aliases.get(self.value, self.value)
|
143
|
+
== aliases.get(other.value, other.value)
|
144
|
+
)
|
145
|
+
else:
|
146
|
+
aliases = self.get_aliases()
|
147
|
+
self_value = aliases.get(self.value, self.value)
|
148
|
+
other_value = aliases.get(other, other)
|
149
|
+
return self_value == other_value
|
150
|
+
|
151
|
+
def __hash__(self):
|
152
|
+
canonical_value = self.aliases.get(self.value, self.value)
|
153
|
+
return hash(canonical_value)
|
@@ -1,6 +1,6 @@
|
|
1
1
|
Metadata-Version: 2.1
|
2
2
|
Name: compressed-tensors-nightly
|
3
|
-
Version: 0.8.0.
|
3
|
+
Version: 0.8.0.20241203
|
4
4
|
Summary: Library for utilization of compressed safetensors of neural network models
|
5
5
|
Home-page: https://github.com/neuralmagic/compressed-tensors
|
6
6
|
Author: Neuralmagic, Inc.
|
@@ -23,9 +23,9 @@ compressed_tensors/config/sparse_bitmask.py,sha256=pZUboRNZTu6NajGOQEFExoPknak5y
|
|
23
23
|
compressed_tensors/linear/__init__.py,sha256=fH6rjBYAxuwrTzBTlTjTgCYNyh6TCvCqajCz4Im4YrA,617
|
24
24
|
compressed_tensors/linear/compressed_linear.py,sha256=MJa-UfoKhIkdUWRD1shrXXri2cOwR5GK0a4t4bNYosM,3268
|
25
25
|
compressed_tensors/quantization/__init__.py,sha256=83J5bPB7PavN2TfCoW7_vEDhfYpm4TDrqYO9vdSQ5bk,760
|
26
|
-
compressed_tensors/quantization/quant_args.py,sha256=
|
26
|
+
compressed_tensors/quantization/quant_args.py,sha256=jwC__lSmuiJ2qSJYYZGgWgQNbZu6YhhS0e-qugrTNXE,9058
|
27
27
|
compressed_tensors/quantization/quant_config.py,sha256=K6kOZ6LDXpFlqsVzR4NEATV6y6Ea83rJWnNyVlvw-pI,10379
|
28
|
-
compressed_tensors/quantization/quant_scheme.py,sha256=
|
28
|
+
compressed_tensors/quantization/quant_scheme.py,sha256=eQ0JrRZ80GX69fpwW87VzPzzhajhk4mUaJScjk82OY4,6010
|
29
29
|
compressed_tensors/quantization/lifecycle/__init__.py,sha256=_uItzFWusyV74Zco_pHLOTdE9a83cL-R-ZdyQrBkIyw,772
|
30
30
|
compressed_tensors/quantization/lifecycle/apply.py,sha256=jCUSgeOBtagE5IhgIbyYMZ4kv8Rm20VGJ4IxXZ5HAnw,15066
|
31
31
|
compressed_tensors/quantization/lifecycle/compressed.py,sha256=Fj9n66IN0EWsOAkBHg3O0GlOQpxstqjCcs0ttzMXrJ0,2296
|
@@ -37,14 +37,14 @@ compressed_tensors/quantization/utils/helpers.py,sha256=DBP-sGRpGAY01K0LFE7qqonN
|
|
37
37
|
compressed_tensors/registry/__init__.py,sha256=FwLSNYqfIrb5JD_6OK_MT4_svvKTN_nEhpgQlQvGbjI,658
|
38
38
|
compressed_tensors/registry/registry.py,sha256=vRcjVB1ITfSbfYUaGndBBmqhip_5vsS62weorVg0iXo,11896
|
39
39
|
compressed_tensors/utils/__init__.py,sha256=gS4gSU2pwcAbsKj-6YMaqhm25udFy6ISYaWBf-myRSM,808
|
40
|
-
compressed_tensors/utils/helpers.py,sha256=
|
40
|
+
compressed_tensors/utils/helpers.py,sha256=T3p0TbhWbQIRjL6Up2Z7UhZO5jpR6WxBhYPPvrhE6lE,5018
|
41
41
|
compressed_tensors/utils/offload.py,sha256=d9q8LNe8HyF8tOjgjA7QGLD3HRysmNp0d8eBbdqBgIM,4089
|
42
42
|
compressed_tensors/utils/permutations_24.py,sha256=kx6fsfDHebx94zsSzhXGyCyuC9sVyah6BUUir_StT28,2530
|
43
43
|
compressed_tensors/utils/permute.py,sha256=V6tJLKo3Syccj-viv4F7ZKZgJeCB-hl-dK8RKI_kBwI,2355
|
44
44
|
compressed_tensors/utils/safetensors_load.py,sha256=m08ANVuTBxQdoa6LufDgcNJ7wCLDJolyZljB8VEybAU,8578
|
45
45
|
compressed_tensors/utils/semi_structured_conversions.py,sha256=XKNffPum54kPASgqKzgKvyeqWPAkair2XEQXjkp7ho8,13489
|
46
|
-
compressed_tensors_nightly-0.8.0.
|
47
|
-
compressed_tensors_nightly-0.8.0.
|
48
|
-
compressed_tensors_nightly-0.8.0.
|
49
|
-
compressed_tensors_nightly-0.8.0.
|
50
|
-
compressed_tensors_nightly-0.8.0.
|
46
|
+
compressed_tensors_nightly-0.8.0.20241203.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
|
47
|
+
compressed_tensors_nightly-0.8.0.20241203.dist-info/METADATA,sha256=BHqRph3AkWyQPLlLwjoTembURpiG2fGN_CRF2dWDtk0,6799
|
48
|
+
compressed_tensors_nightly-0.8.0.20241203.dist-info/WHEEL,sha256=tZoeGjtWxWRfdplE7E3d45VPlLNQnvbKiYnx7gwAy8A,92
|
49
|
+
compressed_tensors_nightly-0.8.0.20241203.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
|
50
|
+
compressed_tensors_nightly-0.8.0.20241203.dist-info/RECORD,,
|
File without changes
|
File without changes
|