compressed-tensors 0.10.2a20250611__py3-none-any.whl → 0.10.2a20250613__py3-none-any.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- compressed_tensors/transform/factory/hadamard.py +1 -1
- compressed_tensors/transform/factory/random_hadamard.py +1 -1
- compressed_tensors/transform/utils/hadamard.py +91 -92
- compressed_tensors/transform/utils/hadamards.safetensors +0 -0
- compressed_tensors/utils/offload.py +85 -50
- compressed_tensors/version.py +1 -1
- {compressed_tensors-0.10.2a20250611.dist-info → compressed_tensors-0.10.2a20250613.dist-info}/METADATA +1 -1
- {compressed_tensors-0.10.2a20250611.dist-info → compressed_tensors-0.10.2a20250613.dist-info}/RECORD +11 -10
- {compressed_tensors-0.10.2a20250611.dist-info → compressed_tensors-0.10.2a20250613.dist-info}/WHEEL +0 -0
- {compressed_tensors-0.10.2a20250611.dist-info → compressed_tensors-0.10.2a20250613.dist-info}/licenses/LICENSE +0 -0
- {compressed_tensors-0.10.2a20250611.dist-info → compressed_tensors-0.10.2a20250613.dist-info}/top_level.txt +0 -0
@@ -59,7 +59,7 @@ class HadamardFactory(TransformFactory):
|
|
59
59
|
return HadamardTransform(weight, args)
|
60
60
|
|
61
61
|
def _create_weight(self, size: int, dtype: dtype, device: device) -> Parameter:
|
62
|
-
data = deterministic_hadamard_matrix(size)
|
62
|
+
data = deterministic_hadamard_matrix(size, dtype, device)
|
63
63
|
data = data.to(dtype=dtype, device=device)
|
64
64
|
return Parameter(data, requires_grad=self.scheme.requires_grad)
|
65
65
|
|
@@ -29,6 +29,6 @@ class RandomHadamardFactory(HadamardFactory):
|
|
29
29
|
"""
|
30
30
|
|
31
31
|
def _create_weight(self, size: int, dtype: dtype, device: device) -> Parameter:
|
32
|
-
data = random_hadamard_matrix(size, self.generator)
|
32
|
+
data = random_hadamard_matrix(size, dtype, device, self.generator)
|
33
33
|
data = data.to(dtype=dtype, device=device)
|
34
34
|
return Parameter(data, requires_grad=self.scheme.requires_grad)
|
@@ -13,95 +13,133 @@
|
|
13
13
|
# limitations under the License.
|
14
14
|
|
15
15
|
import math
|
16
|
-
from
|
16
|
+
from pathlib import Path
|
17
|
+
from typing import Optional
|
17
18
|
|
18
|
-
import numpy
|
19
19
|
import torch
|
20
|
+
from safetensors import safe_open
|
20
21
|
|
21
22
|
|
22
|
-
|
23
|
+
REPO_PATH = Path(__file__).parent / "hadamards.safetensors"
|
23
24
|
|
24
|
-
|
25
|
-
|
26
|
-
|
25
|
+
|
26
|
+
__all__ = ["random_hadamard_matrix", "deterministic_hadamard_matrix", "is_pow2"]
|
27
|
+
|
28
|
+
|
29
|
+
# note that hadamard matrix multiplication can be accelerated using a library such as
|
30
|
+
# https://github.com/Dao-AILab/fast-hadamard-transform/tree/master
|
31
|
+
|
32
|
+
|
33
|
+
def deterministic_hadamard_matrix(
|
34
|
+
size: int,
|
35
|
+
dtype: torch.dtype = torch.bfloat16,
|
36
|
+
device: torch.device = torch.device("cpu"),
|
37
|
+
) -> torch.Tensor:
|
27
38
|
"""
|
28
39
|
Construct an n-by-n Hadamard matrix, using Sylvester's construction.
|
29
40
|
`n` must be a power of 2.
|
30
41
|
|
42
|
+
Adapated from https://github.com/scipy/scipy/blob/v1.15.2/scipy/linalg/_special_matrices.py # noqa: E501
|
43
|
+
|
31
44
|
:param size: order of the matrix, must be a power of 2
|
45
|
+
:param dtype: data type of matrix
|
46
|
+
:param device: device to construct matrix on
|
32
47
|
:return: hadamard matrix of size `size`
|
33
48
|
"""
|
34
49
|
if size <= 0:
|
35
50
|
raise ValueError("Cannot construct deterministic hadamard of size <= 0")
|
36
51
|
|
37
|
-
log2 = int(math.
|
52
|
+
log2 = int(math.log2(size))
|
38
53
|
if size != 2**log2:
|
39
54
|
raise ValueError("Cannot construct deterministic hadamard of size != 2^n")
|
40
55
|
|
41
|
-
H =
|
56
|
+
H = torch.tensor([[1]], dtype=dtype, device=device)
|
42
57
|
|
43
58
|
# Sylvester's construction
|
44
|
-
for
|
45
|
-
H =
|
46
|
-
|
47
|
-
return torch.from_numpy(H / math.sqrt(size))
|
59
|
+
for _ in range(log2):
|
60
|
+
H = torch.vstack((torch.hstack((H, H)), torch.hstack((H, -H))))
|
48
61
|
|
49
|
-
|
50
|
-
# adapted from:
|
51
|
-
# https://github.com/facebookresearch/SpinQuant/blob/main/utils/hadamard_utils.py
|
52
|
-
|
53
|
-
# TODO: the following library exists for online rotations and should be considered
|
54
|
-
# in the future:
|
55
|
-
# https://github.com/Dao-AILab/fast-hadamard-transform/tree/master
|
62
|
+
return H / math.sqrt(size)
|
56
63
|
|
57
64
|
|
58
65
|
def random_hadamard_matrix(
|
59
|
-
size: int,
|
66
|
+
size: int,
|
67
|
+
dtype: torch.dtype = torch.bfloat16,
|
68
|
+
device: torch.device = torch.device("cpu"),
|
69
|
+
gen: Optional[torch.Generator] = None,
|
60
70
|
) -> torch.Tensor:
|
61
71
|
"""
|
62
|
-
Produces a randomly generated Hadamard matrix.
|
63
|
-
|
64
|
-
|
72
|
+
Produces a randomly generated Hadamard matrix. Differs from
|
73
|
+
`deterministic_hadamard_matrix` in that this function supports non powers of 2
|
74
|
+
and randomization using a seeded generator
|
75
|
+
|
76
|
+
Adapated from https://github.com/facebookresearch/SpinQuant/blob/main/utils/hadamard_utils.py # noqa: E501
|
77
|
+
Known matrices were retrieved from N. J. A. Sloane's Library of Hadamard Matrices http://www.neilsloane.com/hadamard/ # noqa: E501
|
65
78
|
|
66
79
|
:param size: The dimension of the hamadard matrix
|
80
|
+
:param dtype: data type of matrix
|
81
|
+
:param device: device to construct matrix on
|
67
82
|
:param gen: Optional generator random values
|
68
83
|
:return: randomly generated hadamard matrix
|
69
84
|
"""
|
70
|
-
|
71
|
-
Q =
|
85
|
+
Q = torch.randint(low=0, high=2, size=(size,), generator=gen, dtype=dtype) # cpu
|
86
|
+
Q = Q.to(device=device)
|
72
87
|
Q = Q * 2 - 1
|
73
88
|
Q = torch.diag(Q)
|
74
89
|
return _matmul_hadU(Q) / math.sqrt(size)
|
75
90
|
|
76
91
|
|
77
|
-
def
|
78
|
-
|
79
|
-
|
80
|
-
hadK, K = None, None
|
81
|
-
if n % 20 == 0:
|
82
|
-
assert _is_pow2(n // 20)
|
83
|
-
K = 20
|
84
|
-
hadK = _get_had20().T if transpose else _get_had20()
|
85
|
-
elif n % 12 == 0:
|
86
|
-
assert _is_pow2(n // 12)
|
87
|
-
K = 12
|
88
|
-
hadK = _get_had12().T if transpose else _get_had12()
|
89
|
-
else:
|
90
|
-
assert _is_pow2(n)
|
91
|
-
K = 1
|
92
|
+
def is_pow2(n: int) -> bool:
|
93
|
+
"""
|
94
|
+
Check if a number is a power of 2
|
92
95
|
|
93
|
-
|
96
|
+
:param n: number to check
|
97
|
+
:return: True iff `n` is a power of 2
|
98
|
+
"""
|
99
|
+
return n > 0 and (n & (n - 1) == 0)
|
100
|
+
|
101
|
+
|
102
|
+
def _fetch_hadamard_divisor(
|
103
|
+
n: int,
|
104
|
+
dtype: torch.dtype,
|
105
|
+
device: torch.device = torch.device("cpu"),
|
106
|
+
file_path: str = REPO_PATH,
|
107
|
+
) -> Optional[torch.Tensor]:
|
108
|
+
"""
|
109
|
+
Fetch a known hadamard matrix from the given file path. The returned matrix will
|
110
|
+
be of of size `k` such that `n / k` is a power of two. Return None if no such
|
111
|
+
matrix exists.
|
94
112
|
|
113
|
+
Note: This function reopens the safetensors file every time it is called.
|
114
|
+
This is technically inefficient, but a very small runtime cost and simpler
|
115
|
+
than forcing callers to manage the file open context
|
116
|
+
|
117
|
+
:param n: size of known hadamard matrix
|
118
|
+
:return: a known hadamard matrix of size `n` if one exists, else None
|
119
|
+
"""
|
120
|
+
with safe_open(file_path, framework="pt", device=str(device)) as file:
|
121
|
+
divisors = sorted((int(key) for key in file.keys()), reverse=True)
|
122
|
+
for divisor in divisors:
|
123
|
+
if n % divisor == 0 and is_pow2(n // divisor):
|
124
|
+
return file.get_tensor(str(divisor)).to(dtype=dtype)
|
125
|
+
|
126
|
+
return None
|
127
|
+
|
128
|
+
|
129
|
+
def _matmul_hadU(X: torch.Tensor) -> torch.Tensor:
|
130
|
+
size = X.size(0)
|
131
|
+
dtype = X.dtype
|
132
|
+
device = X.device
|
95
133
|
|
96
|
-
def _matmul_hadU(X, transpose=False) -> torch.Tensor:
|
97
|
-
n = X.shape[-1]
|
98
134
|
# Check if we have the determined hadamard matrix
|
99
|
-
hadK
|
135
|
+
hadK = _fetch_hadamard_divisor(size, dtype, device=device)
|
136
|
+
if hadK is None:
|
137
|
+
raise ValueError(f"Cannot construct random hadamard matrix of size {size}")
|
138
|
+
K = hadK.size(0)
|
139
|
+
|
100
140
|
# Reshape diag matrix with randomized -1/+1
|
101
|
-
input = X.clone().view(-1,
|
141
|
+
input = X.clone().view(-1, size, 1)
|
102
142
|
output = input.clone()
|
103
|
-
|
104
|
-
# for cases when hadK is not predetermined, determine hadamard matrix
|
105
143
|
while input.shape[1] > K:
|
106
144
|
input = input.view(input.shape[0], input.shape[1] // 2, 2, input.shape[2])
|
107
145
|
output = output.view(input.shape)
|
@@ -109,53 +147,14 @@ def _matmul_hadU(X, transpose=False) -> torch.Tensor:
|
|
109
147
|
output[:, :, 1, :] = input[:, :, 0, :] - input[:, :, 1, :]
|
110
148
|
output = output.view(input.shape[0], input.shape[1], -1)
|
111
149
|
(input, output) = (output, input)
|
150
|
+
assert input.shape[1] == K
|
112
151
|
del output
|
113
152
|
|
114
|
-
#
|
115
|
-
#
|
116
|
-
|
117
|
-
|
118
|
-
|
119
|
-
# input = torch.bmm(
|
120
|
-
# hadK.repeat(len(input), 1, 1).to(input.device).to(input.dtype), input)
|
121
|
-
# Use bcast instead
|
122
|
-
|
123
|
-
# for cases when hadK is pre-determined
|
124
|
-
input = hadK.view(1, K, K).to(input) @ input
|
153
|
+
# Do not explicitly repeat - OOM
|
154
|
+
# input = torch.bmm(
|
155
|
+
# hadK.repeat(len(input), 1, 1).to(input.device).to(input.dtype), input)
|
156
|
+
# Use bcast instead
|
157
|
+
input = hadK.view(1, K, K).to(input) @ input
|
125
158
|
|
126
159
|
# normalize
|
127
160
|
return input.view(X.shape)
|
128
|
-
|
129
|
-
|
130
|
-
def _is_pow2(n: int) -> bool:
|
131
|
-
return (n & (n - 1) == 0) and (n > 0)
|
132
|
-
|
133
|
-
|
134
|
-
def _reshape_bits(packed_bits: numpy.ndarray, original_size: int) -> numpy.ndarray:
|
135
|
-
had_unpacked = numpy.unpackbits(packed_bits)
|
136
|
-
had_unpacked = [1 if x == 1 else -1 for x in had_unpacked]
|
137
|
-
had_unpacked = numpy.array(had_unpacked).reshape((original_size, original_size))
|
138
|
-
return had_unpacked
|
139
|
-
|
140
|
-
|
141
|
-
# http://www.neilsloane.com/hadamard/index.html
|
142
|
-
def _get_had12() -> torch.Tensor:
|
143
|
-
# fmt: off
|
144
|
-
had_12 = numpy.array([128, 13, 29, 232, 235, 71, 218,
|
145
|
-
62, 209, 246, 139, 180, 157, 168, 237, 199, 106, 59], dtype=numpy.uint8)
|
146
|
-
# fmt: on
|
147
|
-
# TODO: just unpack during apply
|
148
|
-
had_12_unpacked = _reshape_bits(had_12, original_size=12)
|
149
|
-
return torch.tensor(had_12_unpacked)
|
150
|
-
|
151
|
-
|
152
|
-
def _get_had20() -> torch.Tensor:
|
153
|
-
# fmt: off
|
154
|
-
had_20 = numpy.array([128, 0, 13, 133, 121, 236, 43, 203, 97, 94, 155, 10, 252,
|
155
|
-
216, 87, 230, 194, 191, 54, 21, 249, 176, 171, 205, 133, 222, 108, 42, 243,
|
156
|
-
97, 215, 155, 10, 188, 216, 149, 230, 200, 175, 54, 133, 121, 188, 43,
|
157
|
-
205, 225, 94, 107, 10, 243], dtype=numpy.uint8)
|
158
|
-
# fmt: on
|
159
|
-
# TODO: just unpack during apply
|
160
|
-
had_20_unpacked = _reshape_bits(had_20, original_size=20)
|
161
|
-
return torch.tensor(had_20_unpacked)
|
Binary file
|
@@ -14,27 +14,30 @@
|
|
14
14
|
"""
|
15
15
|
Utilities associated with offloading functionality provided by `accelerate`.
|
16
16
|
|
17
|
-
|
|
18
|
-
| Operation
|
19
|
-
|
|
20
|
-
| Add
|
21
|
-
| Check
|
22
|
-
| Onload
|
23
|
-
| Update
|
24
|
-
| Delete
|
25
|
-
|
|
17
|
+
| ------------------------------------------------------------------------------------------------------ | # noqa: E501
|
18
|
+
| Operation | Without offloading support | With offloading support | # noqa: E501
|
19
|
+
| ---------- | -------------------------------------- | ------------------------------------------------ | # noqa: E501
|
20
|
+
| Add | module.register_parameter(name, param) | register_offload_parameter(module, name, param) | # noqa: E501
|
21
|
+
| Check | N/A | has_offloaded_params(module) | # noqa: E501
|
22
|
+
| Onload | N/A | with align_module_device(module) | # noqa: E501
|
23
|
+
| Update | module.name.data.copy_(new_data) | update_offload_parameter(module, name, new_data) | # noqa: E501
|
24
|
+
| Delete | del module.name | delete_offload_parameter(module, name) | # noqa: E501
|
25
|
+
| Add Module | module.register_module(name, child) | register_offload_module(name, child) | # noqa: E501
|
26
|
+
| Del Module | del module.name | delete_offload_module(module, name) | # noqa: E501
|
27
|
+
| ------------------------------------------------------------------------------------------------------ | # noqa: E501
|
26
28
|
"""
|
27
29
|
|
28
30
|
import contextlib
|
29
31
|
import warnings
|
30
32
|
from functools import wraps
|
31
|
-
from
|
33
|
+
from operator import attrgetter
|
34
|
+
from typing import Any, Callable, Dict, Iterable, Literal, Optional, Tuple, Union
|
32
35
|
|
33
36
|
import torch
|
37
|
+
from compressed_tensors.utils import patch_attr
|
34
38
|
|
35
39
|
|
36
40
|
try:
|
37
|
-
from accelerate import dispatch_model
|
38
41
|
from accelerate.hooks import (
|
39
42
|
AlignDevicesHook,
|
40
43
|
add_hook_to_module,
|
@@ -45,10 +48,12 @@ try:
|
|
45
48
|
from accelerate.utils import (
|
46
49
|
OffloadedWeightsLoader,
|
47
50
|
PrefixedDataset,
|
51
|
+
find_tied_parameters,
|
48
52
|
set_module_tensor_to_device,
|
49
53
|
)
|
50
54
|
|
51
55
|
_has_accelerate = True
|
56
|
+
|
52
57
|
except ImportError:
|
53
58
|
_has_accelerate = False
|
54
59
|
AlignDevicesHook = None
|
@@ -58,8 +63,8 @@ except ImportError:
|
|
58
63
|
PrefixedDataset = None
|
59
64
|
set_module_tensor_to_device = None
|
60
65
|
named_module_tensors = None
|
61
|
-
dispatch_model = None
|
62
66
|
attach_align_device_hook = None
|
67
|
+
find_tied_parameters = None
|
63
68
|
|
64
69
|
|
65
70
|
__all__ = [
|
@@ -78,14 +83,14 @@ __all__ = [
|
|
78
83
|
"align_module_device",
|
79
84
|
"register_offload_module",
|
80
85
|
"delete_offload_module",
|
81
|
-
"
|
86
|
+
"offloaded_dispatch",
|
87
|
+
"disable_offloading",
|
82
88
|
]
|
83
89
|
|
84
90
|
|
85
91
|
def check_accelerate(fallback: Any):
|
86
92
|
def decorator(func: Callable[[Any], Any]):
|
87
93
|
if not _has_accelerate:
|
88
|
-
|
89
94
|
if fallback == "error":
|
90
95
|
|
91
96
|
@wraps(func)
|
@@ -211,7 +216,7 @@ def register_offload_parameter(
|
|
211
216
|
def update_offload_parameter(
|
212
217
|
module: torch.nn.Module,
|
213
218
|
name: str,
|
214
|
-
data:
|
219
|
+
data: torch.Tensor,
|
215
220
|
offload_device: Optional[Union[torch.device, Literal["disk"]]] = None,
|
216
221
|
):
|
217
222
|
"""
|
@@ -224,7 +229,7 @@ def update_offload_parameter(
|
|
224
229
|
:param offload_device: device on which weight will be offloaded to. If None is
|
225
230
|
provided, then infer device from parameters on module
|
226
231
|
"""
|
227
|
-
param = getattr(module, name)
|
232
|
+
param: torch.nn.Parameter = getattr(module, name)
|
228
233
|
if param.data.shape != data.shape:
|
229
234
|
warnings.warn(
|
230
235
|
f"Shape of parameter being updated {param.data.shape} does not match shape "
|
@@ -232,7 +237,7 @@ def update_offload_parameter(
|
|
232
237
|
)
|
233
238
|
|
234
239
|
# copy data into onloaded parameter if applicable
|
235
|
-
if param.device != torch.device("meta"):
|
240
|
+
if param.device != torch.device("meta") and data is not param.data:
|
236
241
|
param.data.copy_(data)
|
237
242
|
|
238
243
|
# update offload dict
|
@@ -479,46 +484,76 @@ def delete_offload_module(base: torch.nn.Module, name: str):
|
|
479
484
|
|
480
485
|
|
481
486
|
@check_accelerate(fallback="error")
|
482
|
-
def
|
483
|
-
module: torch.nn.Module,
|
487
|
+
def offloaded_dispatch(
|
488
|
+
module: torch.nn.Module,
|
489
|
+
execution_device: torch.device,
|
490
|
+
offload_device: Union[torch.device, Literal["disk"]] = torch.device("cpu"),
|
484
491
|
) -> torch.nn.Module:
|
485
492
|
"""
|
486
|
-
|
493
|
+
Unlike `dispatch_model`, this function forces a module (and its submodules) to
|
494
|
+
offload all parameters and replace them with meta tensors, utiliizing the
|
495
|
+
`AlignDevicesHook` to control onloading and offloading.
|
487
496
|
|
488
497
|
:param module: module containing parameters to offload
|
489
|
-
:param execution_device:
|
490
|
-
:
|
491
|
-
|
492
|
-
|
493
|
-
|
494
|
-
|
495
|
-
|
496
|
-
|
497
|
-
|
498
|
-
|
499
|
-
|
500
|
-
|
501
|
-
|
502
|
-
|
503
|
-
|
504
|
-
|
505
|
-
|
506
|
-
|
507
|
-
|
508
|
-
|
509
|
-
|
498
|
+
:param execution_device: device that modules will onload and execute on
|
499
|
+
:param offload_device: device that module parameters will offload to
|
500
|
+
:return: module with offloading device hooks
|
501
|
+
"""
|
502
|
+
if offload_device == "disk":
|
503
|
+
raise NotImplementedError("Disk offloading is not currently supported")
|
504
|
+
|
505
|
+
# create weights map
|
506
|
+
state_dict = module.state_dict()
|
507
|
+
state_dict = {key: val.to(offload_device) for key, val in state_dict.items()}
|
508
|
+
weights_map = OffloadedWeightsLoader(state_dict=state_dict, device=offload_device)
|
509
|
+
|
510
|
+
# create tied params map
|
511
|
+
tied_params = find_tied_parameters(module)
|
512
|
+
tied_params_map = {}
|
513
|
+
for group in tied_params:
|
514
|
+
for param_name in group:
|
515
|
+
data_ptr = attrgetter(param_name)(module).data_ptr()
|
516
|
+
tied_params_map[data_ptr] = {}
|
517
|
+
|
518
|
+
# recursively attaches hooks to all submodules
|
519
|
+
attach_align_device_hook(
|
520
|
+
module,
|
521
|
+
execution_device=execution_device,
|
522
|
+
offload=True,
|
523
|
+
weights_map=weights_map,
|
524
|
+
tied_params_map=tied_params_map,
|
525
|
+
)
|
526
|
+
return module
|
510
527
|
|
511
|
-
else:
|
512
|
-
for submodule_name, submodule in module.named_children():
|
513
|
-
name.append(submodule_name)
|
514
|
-
collect_device_map(name, submodule)
|
515
|
-
name.pop()
|
516
528
|
|
517
|
-
|
529
|
+
@contextlib.contextmanager
|
530
|
+
def disable_offloading():
|
531
|
+
"""
|
532
|
+
Keep modules onloaded and disable offloading until this context exits.
|
533
|
+
Affects modules which have been hooked with accelerate's `AlignDevicesHook`
|
534
|
+
"""
|
535
|
+
original_pre_forward = AlignDevicesHook.pre_forward
|
536
|
+
onloaded_modules: Dict[torch.nn.Module, Tuple[AlignDevicesHook, bool]] = dict()
|
537
|
+
|
538
|
+
# onload once and disable any future onloading/offloading steps
|
539
|
+
def keep_onload_pre_forward(self: AlignDevicesHook, module, *args, **kwargs):
|
540
|
+
ret = original_pre_forward(self, module, *args, **kwargs)
|
541
|
+
if module not in onloaded_modules:
|
542
|
+
onloaded_modules[module] = (self, self.offload)
|
543
|
+
self.offload = False
|
544
|
+
return ret
|
545
|
+
|
546
|
+
# use the patched pre_forward function within the context
|
547
|
+
with patch_attr(AlignDevicesHook, "pre_forward", keep_onload_pre_forward):
|
548
|
+
yield
|
518
549
|
|
519
|
-
|
520
|
-
|
521
|
-
)
|
550
|
+
# manually offload all modules that were onloaded
|
551
|
+
# update any parameters which may have changed
|
552
|
+
for module, (hook, offload) in onloaded_modules.items():
|
553
|
+
hook.offload = offload
|
554
|
+
for name, param in module.named_parameters():
|
555
|
+
update_offload_parameter(module, name, param.data)
|
556
|
+
hook.post_forward(module, None)
|
522
557
|
|
523
558
|
|
524
559
|
""" Upstreamed Functions """
|
compressed_tensors/version.py
CHANGED
@@ -1,6 +1,6 @@
|
|
1
1
|
Metadata-Version: 2.4
|
2
2
|
Name: compressed-tensors
|
3
|
-
Version: 0.10.
|
3
|
+
Version: 0.10.2a20250613
|
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.
|
{compressed_tensors-0.10.2a20250611.dist-info → compressed_tensors-0.10.2a20250613.dist-info}/RECORD
RENAMED
@@ -1,6 +1,6 @@
|
|
1
1
|
compressed_tensors/__init__.py,sha256=UtKmifNeBCSE2TZSAfduVNNzHY-3V7bLjZ7n7RuXLOE,812
|
2
2
|
compressed_tensors/base.py,sha256=73HYH7HY7O2roC89yG_piPFnZwrBfn_i7HmKl90SKc0,875
|
3
|
-
compressed_tensors/version.py,sha256=
|
3
|
+
compressed_tensors/version.py,sha256=W2dIoBkBkOOTKcVgQ7KVgwm7EtQxgrkm_57h8wJ40X0,523
|
4
4
|
compressed_tensors/compressors/__init__.py,sha256=smSygTSfcfuujRrAXDc6uZm4L_ccV1tWZewqVnOb4lM,825
|
5
5
|
compressed_tensors/compressors/base.py,sha256=nvWsv4xEw1Tkxkxth6TmHplDYXfBeP22xWxOsZERyDY,7204
|
6
6
|
compressed_tensors/compressors/helpers.py,sha256=OK6qxX9j3bHwF9JfIYSGMgBJe2PWjlTA3byXKCJaTIQ,5431
|
@@ -45,21 +45,22 @@ compressed_tensors/transform/transform_config.py,sha256=6JA8VFcoz4EGHOev6thj51Ou
|
|
45
45
|
compressed_tensors/transform/transform_scheme.py,sha256=c7NAuLDL0itFgUfBMNShegMI9bzKL7s4LR3QJTHsXLs,1733
|
46
46
|
compressed_tensors/transform/factory/__init__.py,sha256=fH6rjBYAxuwrTzBTlTjTgCYNyh6TCvCqajCz4Im4YrA,617
|
47
47
|
compressed_tensors/transform/factory/base.py,sha256=yVrYWEnrr2RFWE5AjSNeXzO9aXc443dTNMVSxuLztz8,5940
|
48
|
-
compressed_tensors/transform/factory/hadamard.py,sha256=
|
48
|
+
compressed_tensors/transform/factory/hadamard.py,sha256=zkq6w8uJXRLokUXajAkFb2fJrH0K3SL6qrR2dARrAr8,3139
|
49
49
|
compressed_tensors/transform/factory/matrix_multiply.py,sha256=0g4sYC_tOmCjOomae2gl54UTXiFdl0mCCkmbqIRX8yw,3613
|
50
|
-
compressed_tensors/transform/factory/random_hadamard.py,sha256=
|
50
|
+
compressed_tensors/transform/factory/random_hadamard.py,sha256=TFInxbHslqREOFFiy_mpR88eEYXQnslxXmyh-ZbN-MU,1499
|
51
51
|
compressed_tensors/transform/utils/__init__.py,sha256=fH6rjBYAxuwrTzBTlTjTgCYNyh6TCvCqajCz4Im4YrA,617
|
52
|
-
compressed_tensors/transform/utils/hadamard.py,sha256=
|
52
|
+
compressed_tensors/transform/utils/hadamard.py,sha256=U27Kvo-eDebKcVt8oXTSIAaQ5DvPQj9tDv2hdXHCPPQ,5584
|
53
|
+
compressed_tensors/transform/utils/hadamards.safetensors,sha256=mFd1GzNodGG-ifA1IoH-0nHYzfraCOvrq_dX2zFI1B4,1436901
|
53
54
|
compressed_tensors/transform/utils/utils.py,sha256=PRPTYwPs2nnNaQMq2GEbC4QYKHFKlZwaRyPgdDhl66g,2992
|
54
55
|
compressed_tensors/utils/__init__.py,sha256=gS4gSU2pwcAbsKj-6YMaqhm25udFy6ISYaWBf-myRSM,808
|
55
56
|
compressed_tensors/utils/helpers.py,sha256=cPg-ikdeA92aIGwBONg8GmPNvcGlFhozyJVwsRiXBTA,11981
|
56
|
-
compressed_tensors/utils/offload.py,sha256=
|
57
|
+
compressed_tensors/utils/offload.py,sha256=57TvfCPUYG81q0yyCOWRABaIsg5qIuOrXMI1mpRCLMM,22172
|
57
58
|
compressed_tensors/utils/permutations_24.py,sha256=kx6fsfDHebx94zsSzhXGyCyuC9sVyah6BUUir_StT28,2530
|
58
59
|
compressed_tensors/utils/permute.py,sha256=V6tJLKo3Syccj-viv4F7ZKZgJeCB-hl-dK8RKI_kBwI,2355
|
59
60
|
compressed_tensors/utils/safetensors_load.py,sha256=DMfZBuUbA6qp_BG_zIWT3ckiEE33K9ob34s-OgzReO4,12057
|
60
61
|
compressed_tensors/utils/semi_structured_conversions.py,sha256=XKNffPum54kPASgqKzgKvyeqWPAkair2XEQXjkp7ho8,13489
|
61
|
-
compressed_tensors-0.10.
|
62
|
-
compressed_tensors-0.10.
|
63
|
-
compressed_tensors-0.10.
|
64
|
-
compressed_tensors-0.10.
|
65
|
-
compressed_tensors-0.10.
|
62
|
+
compressed_tensors-0.10.2a20250613.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
|
63
|
+
compressed_tensors-0.10.2a20250613.dist-info/METADATA,sha256=E2m2_QlCeFFDbi6cfe4Uf13f7xaF-84jVzfuzywN2No,7005
|
64
|
+
compressed_tensors-0.10.2a20250613.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
|
65
|
+
compressed_tensors-0.10.2a20250613.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
|
66
|
+
compressed_tensors-0.10.2a20250613.dist-info/RECORD,,
|
{compressed_tensors-0.10.2a20250611.dist-info → compressed_tensors-0.10.2a20250613.dist-info}/WHEEL
RENAMED
File without changes
|
File without changes
|
File without changes
|