krrood-compute-resources 26.10.0__tar.gz
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.
- krrood_compute_resources-26.10.0/LICENSE +165 -0
- krrood_compute_resources-26.10.0/PKG-INFO +41 -0
- krrood_compute_resources-26.10.0/README.md +16 -0
- krrood_compute_resources-26.10.0/pyproject.toml +61 -0
- krrood_compute_resources-26.10.0/setup.cfg +4 -0
- krrood_compute_resources-26.10.0/src/krrood_compute_resources/__init__.py +1 -0
- krrood_compute_resources-26.10.0/src/krrood_compute_resources/_version.py +1 -0
- krrood_compute_resources-26.10.0/src/krrood_compute_resources/devices.py +19 -0
- krrood_compute_resources-26.10.0/src/krrood_compute_resources/exceptions.py +75 -0
- krrood_compute_resources-26.10.0/src/krrood_compute_resources/jax_memory.py +42 -0
- krrood_compute_resources-26.10.0/src/krrood_compute_resources/machine.py +235 -0
- krrood_compute_resources-26.10.0/src/krrood_compute_resources/memory.py +181 -0
- krrood_compute_resources-26.10.0/src/krrood_compute_resources/torch_memory.py +39 -0
- krrood_compute_resources-26.10.0/src/krrood_compute_resources.egg-info/PKG-INFO +41 -0
- krrood_compute_resources-26.10.0/src/krrood_compute_resources.egg-info/SOURCES.txt +16 -0
- krrood_compute_resources-26.10.0/src/krrood_compute_resources.egg-info/dependency_links.txt +1 -0
- krrood_compute_resources-26.10.0/src/krrood_compute_resources.egg-info/requires.txt +11 -0
- krrood_compute_resources-26.10.0/src/krrood_compute_resources.egg-info/top_level.txt +1 -0
|
@@ -0,0 +1,165 @@
|
|
|
1
|
+
GNU LESSER GENERAL PUBLIC LICENSE
|
|
2
|
+
Version 3, 29 June 2007
|
|
3
|
+
|
|
4
|
+
Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
|
|
5
|
+
Everyone is permitted to copy and distribute verbatim copies
|
|
6
|
+
of this license document, but changing it is not allowed.
|
|
7
|
+
|
|
8
|
+
|
|
9
|
+
This version of the GNU Lesser General Public License incorporates
|
|
10
|
+
the terms and conditions of version 3 of the GNU General Public
|
|
11
|
+
License, supplemented by the additional permissions listed below.
|
|
12
|
+
|
|
13
|
+
0. Additional Definitions.
|
|
14
|
+
|
|
15
|
+
As used herein, "this License" refers to version 3 of the GNU Lesser
|
|
16
|
+
General Public License, and the "GNU GPL" refers to version 3 of the GNU
|
|
17
|
+
General Public License.
|
|
18
|
+
|
|
19
|
+
"The Library" refers to a covered work governed by this License,
|
|
20
|
+
other than an Application or a Combined Work as defined below.
|
|
21
|
+
|
|
22
|
+
An "Application" is any work that makes use of an interface provided
|
|
23
|
+
by the Library, but which is not otherwise based on the Library.
|
|
24
|
+
Defining a subclass of a class defined by the Library is deemed a mode
|
|
25
|
+
of using an interface provided by the Library.
|
|
26
|
+
|
|
27
|
+
A "Combined Work" is a work produced by combining or linking an
|
|
28
|
+
Application with the Library. The particular version of the Library
|
|
29
|
+
with which the Combined Work was made is also called the "Linked
|
|
30
|
+
Version".
|
|
31
|
+
|
|
32
|
+
The "Minimal Corresponding Source" for a Combined Work means the
|
|
33
|
+
Corresponding Source for the Combined Work, excluding any source code
|
|
34
|
+
for portions of the Combined Work that, considered in isolation, are
|
|
35
|
+
based on the Application, and not on the Linked Version.
|
|
36
|
+
|
|
37
|
+
The "Corresponding Application Code" for a Combined Work means the
|
|
38
|
+
object code and/or source code for the Application, including any data
|
|
39
|
+
and utility programs needed for reproducing the Combined Work from the
|
|
40
|
+
Application, but excluding the System Libraries of the Combined Work.
|
|
41
|
+
|
|
42
|
+
1. Exception to Section 3 of the GNU GPL.
|
|
43
|
+
|
|
44
|
+
You may convey a covered work under sections 3 and 4 of this License
|
|
45
|
+
without being bound by section 3 of the GNU GPL.
|
|
46
|
+
|
|
47
|
+
2. Conveying Modified Versions.
|
|
48
|
+
|
|
49
|
+
If you modify a copy of the Library, and, in your modifications, a
|
|
50
|
+
facility refers to a function or data to be supplied by an Application
|
|
51
|
+
that uses the facility (other than as an argument passed when the
|
|
52
|
+
facility is invoked), then you may convey a copy of the modified
|
|
53
|
+
version:
|
|
54
|
+
|
|
55
|
+
a) under this License, provided that you make a good faith effort to
|
|
56
|
+
ensure that, in the event an Application does not supply the
|
|
57
|
+
function or data, the facility still operates, and performs
|
|
58
|
+
whatever part of its purpose remains meaningful, or
|
|
59
|
+
|
|
60
|
+
b) under the GNU GPL, with none of the additional permissions of
|
|
61
|
+
this License applicable to that copy.
|
|
62
|
+
|
|
63
|
+
3. Object Code Incorporating Material from Library Header Files.
|
|
64
|
+
|
|
65
|
+
The object code form of an Application may incorporate material from
|
|
66
|
+
a header file that is part of the Library. You may convey such object
|
|
67
|
+
code under terms of your choice, provided that, if the incorporated
|
|
68
|
+
material is not limited to numerical parameters, data structure
|
|
69
|
+
layouts and accessors, or small macros, inline functions and templates
|
|
70
|
+
(ten or fewer lines in length), you do both of the following:
|
|
71
|
+
|
|
72
|
+
a) Give prominent notice with each copy of the object code that the
|
|
73
|
+
Library is used in it and that the Library and its use are
|
|
74
|
+
covered by this License.
|
|
75
|
+
|
|
76
|
+
b) Accompany the object code with a copy of the GNU GPL and this license
|
|
77
|
+
document.
|
|
78
|
+
|
|
79
|
+
4. Combined Works.
|
|
80
|
+
|
|
81
|
+
You may convey a Combined Work under terms of your choice that,
|
|
82
|
+
taken together, effectively do not restrict modification of the
|
|
83
|
+
portions of the Library contained in the Combined Work and reverse
|
|
84
|
+
engineering for debugging such modifications, if you also do each of
|
|
85
|
+
the following:
|
|
86
|
+
|
|
87
|
+
a) Give prominent notice with each copy of the Combined Work that
|
|
88
|
+
the Library is used in it and that the Library and its use are
|
|
89
|
+
covered by this License.
|
|
90
|
+
|
|
91
|
+
b) Accompany the Combined Work with a copy of the GNU GPL and this license
|
|
92
|
+
document.
|
|
93
|
+
|
|
94
|
+
c) For a Combined Work that displays copyright notices during
|
|
95
|
+
execution, include the copyright notice for the Library among
|
|
96
|
+
these notices, as well as a reference directing the user to the
|
|
97
|
+
copies of the GNU GPL and this license document.
|
|
98
|
+
|
|
99
|
+
d) Do one of the following:
|
|
100
|
+
|
|
101
|
+
0) Convey the Minimal Corresponding Source under the terms of this
|
|
102
|
+
License, and the Corresponding Application Code in a form
|
|
103
|
+
suitable for, and under terms that permit, the user to
|
|
104
|
+
recombine or relink the Application with a modified version of
|
|
105
|
+
the Linked Version to produce a modified Combined Work, in the
|
|
106
|
+
manner specified by section 6 of the GNU GPL for conveying
|
|
107
|
+
Corresponding Source.
|
|
108
|
+
|
|
109
|
+
1) Use a suitable shared library mechanism for linking with the
|
|
110
|
+
Library. A suitable mechanism is one that (a) uses at run time
|
|
111
|
+
a copy of the Library already present on the user's computer
|
|
112
|
+
system, and (b) will operate properly with a modified version
|
|
113
|
+
of the Library that is interface-compatible with the Linked
|
|
114
|
+
Version.
|
|
115
|
+
|
|
116
|
+
e) Provide Installation Information, but only if you would otherwise
|
|
117
|
+
be required to provide such information under section 6 of the
|
|
118
|
+
GNU GPL, and only to the extent that such information is
|
|
119
|
+
necessary to install and execute a modified version of the
|
|
120
|
+
Combined Work produced by recombining or relinking the
|
|
121
|
+
Application with a modified version of the Linked Version. (If
|
|
122
|
+
you use option 4d0, the Installation Information must accompany
|
|
123
|
+
the Minimal Corresponding Source and Corresponding Application
|
|
124
|
+
Code. If you use option 4d1, you must provide the Installation
|
|
125
|
+
Information in the manner specified by section 6 of the GNU GPL
|
|
126
|
+
for conveying Corresponding Source.)
|
|
127
|
+
|
|
128
|
+
5. Combined Libraries.
|
|
129
|
+
|
|
130
|
+
You may place library facilities that are a work based on the
|
|
131
|
+
Library side by side in a single library together with other library
|
|
132
|
+
facilities that are not Applications and are not covered by this
|
|
133
|
+
License, and convey such a combined library under terms of your
|
|
134
|
+
choice, if you do both of the following:
|
|
135
|
+
|
|
136
|
+
a) Accompany the combined library with a copy of the same work based
|
|
137
|
+
on the Library, uncombined with any other library facilities,
|
|
138
|
+
conveyed under the terms of this License.
|
|
139
|
+
|
|
140
|
+
b) Give prominent notice with the combined library that part of it
|
|
141
|
+
is a work based on the Library, and explaining where to find the
|
|
142
|
+
accompanying uncombined form of the same work.
|
|
143
|
+
|
|
144
|
+
6. Revised Versions of the GNU Lesser General Public License.
|
|
145
|
+
|
|
146
|
+
The Free Software Foundation may publish revised and/or new versions
|
|
147
|
+
of the GNU Lesser General Public License from time to time. Such new
|
|
148
|
+
versions will be similar in spirit to the present version, but may
|
|
149
|
+
differ in detail to address new problems or concerns.
|
|
150
|
+
|
|
151
|
+
Each version is given a distinguishing version number. If the
|
|
152
|
+
Library as you received it specifies that a certain numbered version
|
|
153
|
+
of the GNU Lesser General Public License "or any later version"
|
|
154
|
+
applies to it, you have the option of following the terms and
|
|
155
|
+
conditions either of that published version or of any later version
|
|
156
|
+
published by the Free Software Foundation. If the Library as you
|
|
157
|
+
received it does not specify a version number of the GNU Lesser
|
|
158
|
+
General Public License, you may choose any version of the GNU Lesser
|
|
159
|
+
General Public License ever published by the Free Software Foundation.
|
|
160
|
+
|
|
161
|
+
If the Library as you received it specifies that a proxy can decide
|
|
162
|
+
whether future versions of the GNU Lesser General Public License shall
|
|
163
|
+
apply, that proxy's public statement of acceptance of any version is
|
|
164
|
+
permanent authorization for you to choose that version for the
|
|
165
|
+
Library.
|
|
@@ -0,0 +1,41 @@
|
|
|
1
|
+
Metadata-Version: 2.4
|
|
2
|
+
Name: krrood_compute_resources
|
|
3
|
+
Version: 26.10.0
|
|
4
|
+
Summary: The processor cores, memory and GPUs a computation runs on, and how a machine is divided among processes.
|
|
5
|
+
Author-email: Tom Schierenbeck <tom_sch@uni-bremen.de>
|
|
6
|
+
License-Expression: LGPL-3.0-only
|
|
7
|
+
Project-URL: Source, https://github.com/cram2/cognitive_robot_abstract_machine
|
|
8
|
+
Project-URL: Bug Tracker, https://github.com/cram2/cognitive_robot_abstract_machine/issues
|
|
9
|
+
Project-URL: Changelog, https://github.com/cram2/cognitive_robot_abstract_machine/blob/main/CHANGELOG.md
|
|
10
|
+
Keywords: compute resources,memory,gpu,scheduling
|
|
11
|
+
Classifier: Programming Language :: Python :: 3
|
|
12
|
+
Classifier: Operating System :: POSIX :: Linux
|
|
13
|
+
Requires-Python: <3.13,>=3.12
|
|
14
|
+
Description-Content-Type: text/markdown
|
|
15
|
+
License-File: LICENSE
|
|
16
|
+
Requires-Dist: krrood>=26.10.0
|
|
17
|
+
Requires-Dist: typing_extensions
|
|
18
|
+
Provides-Extra: jax
|
|
19
|
+
Requires-Dist: jax>=0.6.0; extra == "jax"
|
|
20
|
+
Provides-Extra: torch
|
|
21
|
+
Requires-Dist: torch>=2.4; extra == "torch"
|
|
22
|
+
Provides-Extra: dev
|
|
23
|
+
Requires-Dist: pytest<8.0.0,>=7.0.0; extra == "dev"
|
|
24
|
+
Dynamic: license-file
|
|
25
|
+
|
|
26
|
+
# krrood_compute_resources
|
|
27
|
+
|
|
28
|
+
The processor cores, memory and GPUs a computation runs on, and how a machine is divided
|
|
29
|
+
among processes that run side by side.
|
|
30
|
+
|
|
31
|
+
- `krrood_compute_resources.devices`: `Device`, the kinds of device a computation runs on.
|
|
32
|
+
- `krrood_compute_resources.memory`: `DeviceMemory`, the free memory of a device;
|
|
33
|
+
`HostMemory`, the main memory of a Linux host within its control group; and
|
|
34
|
+
`BudgetedMemory`, the share of a device's memory one process may use.
|
|
35
|
+
- `krrood_compute_resources.machine`: `Machine`, the cores and memory of the machine a
|
|
36
|
+
process runs on; `MachineDivision`, which divides it into `MachineShare`s; and the
|
|
37
|
+
environment and core affinity that keep a process within its share.
|
|
38
|
+
- `krrood_compute_resources.jax_memory`, with the extra `krrood_compute_resources[jax]`:
|
|
39
|
+
`JaxDeviceMemory`, the memory of an accelerator jax stores arrays on.
|
|
40
|
+
- `krrood_compute_resources.torch_memory`, with the extra `krrood_compute_resources[torch]`:
|
|
41
|
+
`TorchDeviceMemory`, the memory of a CUDA device torch stores tensors on.
|
|
@@ -0,0 +1,16 @@
|
|
|
1
|
+
# krrood_compute_resources
|
|
2
|
+
|
|
3
|
+
The processor cores, memory and GPUs a computation runs on, and how a machine is divided
|
|
4
|
+
among processes that run side by side.
|
|
5
|
+
|
|
6
|
+
- `krrood_compute_resources.devices`: `Device`, the kinds of device a computation runs on.
|
|
7
|
+
- `krrood_compute_resources.memory`: `DeviceMemory`, the free memory of a device;
|
|
8
|
+
`HostMemory`, the main memory of a Linux host within its control group; and
|
|
9
|
+
`BudgetedMemory`, the share of a device's memory one process may use.
|
|
10
|
+
- `krrood_compute_resources.machine`: `Machine`, the cores and memory of the machine a
|
|
11
|
+
process runs on; `MachineDivision`, which divides it into `MachineShare`s; and the
|
|
12
|
+
environment and core affinity that keep a process within its share.
|
|
13
|
+
- `krrood_compute_resources.jax_memory`, with the extra `krrood_compute_resources[jax]`:
|
|
14
|
+
`JaxDeviceMemory`, the memory of an accelerator jax stores arrays on.
|
|
15
|
+
- `krrood_compute_resources.torch_memory`, with the extra `krrood_compute_resources[torch]`:
|
|
16
|
+
`TorchDeviceMemory`, the memory of a CUDA device torch stores tensors on.
|
|
@@ -0,0 +1,61 @@
|
|
|
1
|
+
[build-system]
|
|
2
|
+
requires = ["setuptools>=68", "wheel"]
|
|
3
|
+
build-backend = "setuptools.build_meta"
|
|
4
|
+
|
|
5
|
+
[project]
|
|
6
|
+
name = "krrood_compute_resources"
|
|
7
|
+
authors = [
|
|
8
|
+
{ name = "Tom Schierenbeck", email = "tom_sch@uni-bremen.de" },
|
|
9
|
+
]
|
|
10
|
+
description = "The processor cores, memory and GPUs a computation runs on, and how a machine is divided among processes."
|
|
11
|
+
readme = "README.md"
|
|
12
|
+
requires-python = ">=3.12,<3.13"
|
|
13
|
+
classifiers = [
|
|
14
|
+
"Programming Language :: Python :: 3",
|
|
15
|
+
"Operating System :: POSIX :: Linux",
|
|
16
|
+
]
|
|
17
|
+
keywords = ["compute resources", "memory", "gpu", "scheduling"]
|
|
18
|
+
license = "LGPL-3.0-only"
|
|
19
|
+
dynamic = ["version"]
|
|
20
|
+
dependencies = [
|
|
21
|
+
"krrood>=26.10.0",
|
|
22
|
+
"typing_extensions",
|
|
23
|
+
]
|
|
24
|
+
|
|
25
|
+
[project.optional-dependencies]
|
|
26
|
+
jax = [
|
|
27
|
+
"jax>=0.6.0",
|
|
28
|
+
]
|
|
29
|
+
torch = [
|
|
30
|
+
"torch>=2.4",
|
|
31
|
+
]
|
|
32
|
+
dev = [
|
|
33
|
+
"pytest>=7.0.0,<8.0.0",
|
|
34
|
+
]
|
|
35
|
+
|
|
36
|
+
[project.urls]
|
|
37
|
+
"Source" = 'https://github.com/cram2/cognitive_robot_abstract_machine'
|
|
38
|
+
"Bug Tracker" = "https://github.com/cram2/cognitive_robot_abstract_machine/issues"
|
|
39
|
+
"Changelog" = "https://github.com/cram2/cognitive_robot_abstract_machine/blob/main/CHANGELOG.md"
|
|
40
|
+
|
|
41
|
+
[tool.setuptools.packages.find]
|
|
42
|
+
where = ["src"]
|
|
43
|
+
include = ["krrood_compute_resources*"]
|
|
44
|
+
|
|
45
|
+
[tool.setuptools.dynamic]
|
|
46
|
+
version = { attr = "krrood_compute_resources._version.__version__" }
|
|
47
|
+
|
|
48
|
+
[tool.krrood.core-boundary]
|
|
49
|
+
# The modules a plain `pip install krrood_compute_resources` must be able to load; the
|
|
50
|
+
# memory of jax and torch devices needs the extras of the same name.
|
|
51
|
+
core = [
|
|
52
|
+
"krrood_compute_resources._version",
|
|
53
|
+
"krrood_compute_resources.devices",
|
|
54
|
+
"krrood_compute_resources.exceptions",
|
|
55
|
+
"krrood_compute_resources.machine",
|
|
56
|
+
"krrood_compute_resources.memory",
|
|
57
|
+
]
|
|
58
|
+
dependencies = [
|
|
59
|
+
"krrood",
|
|
60
|
+
"typing-extensions",
|
|
61
|
+
]
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
from ._version import __version__
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
__version__ = "26.10.0"
|
|
@@ -0,0 +1,19 @@
|
|
|
1
|
+
"""
|
|
2
|
+
The kinds of device a computation runs on.
|
|
3
|
+
"""
|
|
4
|
+
|
|
5
|
+
from __future__ import annotations
|
|
6
|
+
|
|
7
|
+
# %% imports
|
|
8
|
+
from enum import StrEnum
|
|
9
|
+
|
|
10
|
+
# %% devices
|
|
11
|
+
|
|
12
|
+
|
|
13
|
+
class Device(StrEnum):
|
|
14
|
+
"""
|
|
15
|
+
The kind of device a computation runs on, valued as jax names its platform.
|
|
16
|
+
"""
|
|
17
|
+
|
|
18
|
+
CENTRAL_PROCESSOR = "cpu"
|
|
19
|
+
GRAPHICS_PROCESSOR = "gpu"
|
|
@@ -0,0 +1,75 @@
|
|
|
1
|
+
"""
|
|
2
|
+
Exceptions of the compute resources model.
|
|
3
|
+
"""
|
|
4
|
+
|
|
5
|
+
from __future__ import annotations
|
|
6
|
+
|
|
7
|
+
# %% imports
|
|
8
|
+
from dataclasses import dataclass
|
|
9
|
+
|
|
10
|
+
from krrood.exceptions import DataclassException
|
|
11
|
+
|
|
12
|
+
# %% memory
|
|
13
|
+
|
|
14
|
+
|
|
15
|
+
@dataclass(kw_only=True)
|
|
16
|
+
class MissingMemoryFieldError(DataclassException, LookupError):
|
|
17
|
+
"""
|
|
18
|
+
Exception raised when a Linux memory file lacks a field that is read from it.
|
|
19
|
+
"""
|
|
20
|
+
|
|
21
|
+
memory_file: str
|
|
22
|
+
"""
|
|
23
|
+
The memory file that was read.
|
|
24
|
+
"""
|
|
25
|
+
|
|
26
|
+
memory_field: str
|
|
27
|
+
"""
|
|
28
|
+
The memory field that was not found in it.
|
|
29
|
+
"""
|
|
30
|
+
|
|
31
|
+
def error_message(self) -> str:
|
|
32
|
+
return f"{self.memory_file} has no field {self.memory_field}."
|
|
33
|
+
|
|
34
|
+
def suggest_correction(self) -> str:
|
|
35
|
+
return "Run on a Linux kernel that reports this field."
|
|
36
|
+
|
|
37
|
+
|
|
38
|
+
# %% division of a machine
|
|
39
|
+
|
|
40
|
+
|
|
41
|
+
@dataclass(kw_only=True)
|
|
42
|
+
class EmptyMachineDivisionError(DataclassException, ValueError):
|
|
43
|
+
"""
|
|
44
|
+
Exception raised when a machine is divided into no shares at all.
|
|
45
|
+
"""
|
|
46
|
+
|
|
47
|
+
def error_message(self) -> str:
|
|
48
|
+
return "A machine division needs at least one share."
|
|
49
|
+
|
|
50
|
+
def suggest_correction(self) -> str:
|
|
51
|
+
return "Give at least one share of the GPU or of the processor cores."
|
|
52
|
+
|
|
53
|
+
|
|
54
|
+
@dataclass(kw_only=True)
|
|
55
|
+
class MissingGraphicsProcessorError(DataclassException, LookupError):
|
|
56
|
+
"""
|
|
57
|
+
Exception raised when shares of the GPU are asked of a machine without one.
|
|
58
|
+
"""
|
|
59
|
+
|
|
60
|
+
graphics_processor_shares: int
|
|
61
|
+
"""
|
|
62
|
+
The number of shares of the GPU that were asked for.
|
|
63
|
+
"""
|
|
64
|
+
|
|
65
|
+
def error_message(self) -> str:
|
|
66
|
+
return (
|
|
67
|
+
f"{self.graphics_processor_shares} shares of the GPU were asked for, but "
|
|
68
|
+
f"the machine has no GPU."
|
|
69
|
+
)
|
|
70
|
+
|
|
71
|
+
def suggest_correction(self) -> str:
|
|
72
|
+
return (
|
|
73
|
+
"Give the machine the memory of its GPU, or divide it into shares of the "
|
|
74
|
+
"processor cores only."
|
|
75
|
+
)
|
|
@@ -0,0 +1,42 @@
|
|
|
1
|
+
"""
|
|
2
|
+
The memory of the accelerators jax stores arrays on.
|
|
3
|
+
"""
|
|
4
|
+
|
|
5
|
+
from __future__ import annotations
|
|
6
|
+
|
|
7
|
+
# %% imports
|
|
8
|
+
from dataclasses import dataclass
|
|
9
|
+
from enum import StrEnum
|
|
10
|
+
|
|
11
|
+
import jax
|
|
12
|
+
|
|
13
|
+
from krrood_compute_resources.memory import DeviceMemory
|
|
14
|
+
|
|
15
|
+
# %% memory of a jax device
|
|
16
|
+
|
|
17
|
+
|
|
18
|
+
class JaxMemoryStatistic(StrEnum):
|
|
19
|
+
"""
|
|
20
|
+
The entries of :meth:`jax.Device.memory_stats` that are read, each in bytes.
|
|
21
|
+
"""
|
|
22
|
+
|
|
23
|
+
LIMIT = "bytes_limit"
|
|
24
|
+
IN_USE = "bytes_in_use"
|
|
25
|
+
|
|
26
|
+
|
|
27
|
+
@dataclass(kw_only=True)
|
|
28
|
+
class JaxDeviceMemory(DeviceMemory):
|
|
29
|
+
"""
|
|
30
|
+
The memory of an accelerator jax stores arrays on.
|
|
31
|
+
"""
|
|
32
|
+
|
|
33
|
+
device: jax.Device
|
|
34
|
+
"""
|
|
35
|
+
The accelerator.
|
|
36
|
+
"""
|
|
37
|
+
|
|
38
|
+
def free_bytes(self) -> int:
|
|
39
|
+
statistics = self.device.memory_stats()
|
|
40
|
+
return (
|
|
41
|
+
statistics[JaxMemoryStatistic.LIMIT] - statistics[JaxMemoryStatistic.IN_USE]
|
|
42
|
+
)
|
|
@@ -0,0 +1,235 @@
|
|
|
1
|
+
"""
|
|
2
|
+
The machine a process runs on, and its division into shares that processes running side
|
|
3
|
+
by side each keep to.
|
|
4
|
+
"""
|
|
5
|
+
|
|
6
|
+
from __future__ import annotations
|
|
7
|
+
|
|
8
|
+
# %% imports
|
|
9
|
+
import os
|
|
10
|
+
from dataclasses import dataclass
|
|
11
|
+
from enum import StrEnum
|
|
12
|
+
|
|
13
|
+
from krrood_compute_resources.devices import Device
|
|
14
|
+
from krrood_compute_resources.exceptions import (
|
|
15
|
+
EmptyMachineDivisionError,
|
|
16
|
+
MissingGraphicsProcessorError,
|
|
17
|
+
)
|
|
18
|
+
from krrood_compute_resources.memory import DeviceMemory, HostMemory
|
|
19
|
+
|
|
20
|
+
# %% environment of a process
|
|
21
|
+
|
|
22
|
+
|
|
23
|
+
class ThreadVariable(StrEnum):
|
|
24
|
+
"""
|
|
25
|
+
The environment variables that cap the threads of the numeric libraries.
|
|
26
|
+
"""
|
|
27
|
+
|
|
28
|
+
OPEN_MULTI_PROCESSING = "OMP_NUM_THREADS"
|
|
29
|
+
OPEN_BASIC_LINEAR_ALGEBRA_SUBPROGRAMS = "OPENBLAS_NUM_THREADS"
|
|
30
|
+
MATH_KERNEL_LIBRARY = "MKL_NUM_THREADS"
|
|
31
|
+
|
|
32
|
+
|
|
33
|
+
class GraphicsProcessorVariable(StrEnum):
|
|
34
|
+
"""
|
|
35
|
+
The environment variables that configure how a process uses the GPU.
|
|
36
|
+
"""
|
|
37
|
+
|
|
38
|
+
JAX_MEMORY_PREALLOCATION = "XLA_PYTHON_CLIENT_PREALLOCATE"
|
|
39
|
+
"""
|
|
40
|
+
Whether jax claims most of the GPU memory when it starts.
|
|
41
|
+
"""
|
|
42
|
+
|
|
43
|
+
VISIBLE_DEVICES = "CUDA_VISIBLE_DEVICES"
|
|
44
|
+
"""
|
|
45
|
+
The GPUs a process sees.
|
|
46
|
+
"""
|
|
47
|
+
|
|
48
|
+
|
|
49
|
+
# %% machine
|
|
50
|
+
|
|
51
|
+
|
|
52
|
+
@dataclass(kw_only=True)
|
|
53
|
+
class Machine:
|
|
54
|
+
"""
|
|
55
|
+
The processor cores and memory of the machine a process runs on.
|
|
56
|
+
"""
|
|
57
|
+
|
|
58
|
+
processor_cores: set[int]
|
|
59
|
+
"""
|
|
60
|
+
The processor cores the process may run on.
|
|
61
|
+
"""
|
|
62
|
+
|
|
63
|
+
host_memory: DeviceMemory
|
|
64
|
+
"""
|
|
65
|
+
The main memory of the machine.
|
|
66
|
+
"""
|
|
67
|
+
|
|
68
|
+
graphics_processor_memory: DeviceMemory | None = None
|
|
69
|
+
"""
|
|
70
|
+
The memory of the GPU, ``None`` on a machine without one.
|
|
71
|
+
"""
|
|
72
|
+
|
|
73
|
+
@classmethod
|
|
74
|
+
def from_current_process(
|
|
75
|
+
cls, *, graphics_processor_memory: DeviceMemory | None = None
|
|
76
|
+
) -> Machine:
|
|
77
|
+
"""
|
|
78
|
+
:param graphics_processor_memory: The memory of the GPU, ``None`` without one.
|
|
79
|
+
:return: The machine as the current process sees it: the cores it may run on
|
|
80
|
+
and the memory of the host within its control group.
|
|
81
|
+
"""
|
|
82
|
+
return cls(
|
|
83
|
+
processor_cores=os.sched_getaffinity(0),
|
|
84
|
+
host_memory=HostMemory(),
|
|
85
|
+
graphics_processor_memory=graphics_processor_memory,
|
|
86
|
+
)
|
|
87
|
+
|
|
88
|
+
|
|
89
|
+
# %% shares of a machine
|
|
90
|
+
|
|
91
|
+
|
|
92
|
+
@dataclass(kw_only=True)
|
|
93
|
+
class MachineShare:
|
|
94
|
+
"""
|
|
95
|
+
A share of a machine that one process uses at a time.
|
|
96
|
+
"""
|
|
97
|
+
|
|
98
|
+
device: Device
|
|
99
|
+
"""
|
|
100
|
+
The device the process computes on.
|
|
101
|
+
"""
|
|
102
|
+
|
|
103
|
+
processor_cores: set[int]
|
|
104
|
+
"""
|
|
105
|
+
The processor cores of the share.
|
|
106
|
+
"""
|
|
107
|
+
|
|
108
|
+
memory_budget_bytes: int
|
|
109
|
+
"""
|
|
110
|
+
The memory of its device the process may allocate.
|
|
111
|
+
"""
|
|
112
|
+
|
|
113
|
+
host_memory_limit_bytes: int
|
|
114
|
+
"""
|
|
115
|
+
The resident host memory above which the process is stopped.
|
|
116
|
+
"""
|
|
117
|
+
|
|
118
|
+
def process_environment(self, *, environment: dict[str, str]) -> dict[str, str]:
|
|
119
|
+
"""
|
|
120
|
+
:param environment: The environment the process would otherwise get.
|
|
121
|
+
:return: That environment, with the threads of the numeric libraries capped at
|
|
122
|
+
the cores of the share, and with the GPU hidden from a share on the CPU and
|
|
123
|
+
shared without preallocation on a share of the GPU.
|
|
124
|
+
"""
|
|
125
|
+
confined = dict(environment)
|
|
126
|
+
for variable in ThreadVariable:
|
|
127
|
+
confined[variable] = str(len(self.processor_cores))
|
|
128
|
+
if self.device is Device.GRAPHICS_PROCESSOR:
|
|
129
|
+
confined[GraphicsProcessorVariable.JAX_MEMORY_PREALLOCATION] = "false"
|
|
130
|
+
else:
|
|
131
|
+
confined[GraphicsProcessorVariable.VISIBLE_DEVICES] = ""
|
|
132
|
+
return confined
|
|
133
|
+
|
|
134
|
+
def confine_current_process(self) -> None:
|
|
135
|
+
"""
|
|
136
|
+
Restrict the current process to the processor cores of the share.
|
|
137
|
+
"""
|
|
138
|
+
os.sched_setaffinity(0, self.processor_cores)
|
|
139
|
+
|
|
140
|
+
|
|
141
|
+
@dataclass(kw_only=True)
|
|
142
|
+
class MachineDivision:
|
|
143
|
+
"""
|
|
144
|
+
The division of a machine into shares of the GPU and shares of the processor cores
|
|
145
|
+
alone.
|
|
146
|
+
"""
|
|
147
|
+
|
|
148
|
+
graphics_processor_shares: int = 3
|
|
149
|
+
"""
|
|
150
|
+
The number of shares that compute on the GPU.
|
|
151
|
+
"""
|
|
152
|
+
|
|
153
|
+
processor_core_shares: int = 7
|
|
154
|
+
"""
|
|
155
|
+
The number of shares that compute on the processor cores alone.
|
|
156
|
+
"""
|
|
157
|
+
|
|
158
|
+
host_memory_fraction: float = 0.8
|
|
159
|
+
"""
|
|
160
|
+
The fraction of the free host memory that is divided among the shares.
|
|
161
|
+
"""
|
|
162
|
+
|
|
163
|
+
host_memory_limit_shares: float = 2.0
|
|
164
|
+
"""
|
|
165
|
+
The number of host memory shares a process may hold before it is stopped.
|
|
166
|
+
"""
|
|
167
|
+
|
|
168
|
+
graphics_processor_memory_fraction: float = 0.9
|
|
169
|
+
"""
|
|
170
|
+
The fraction of the free GPU memory that is divided among the GPU shares, the rest
|
|
171
|
+
kept free since the memory a process needs is often only estimated.
|
|
172
|
+
"""
|
|
173
|
+
|
|
174
|
+
def __post_init__(self) -> None:
|
|
175
|
+
"""
|
|
176
|
+
:raises EmptyMachineDivisionError: If the division has no shares.
|
|
177
|
+
"""
|
|
178
|
+
if self.graphics_processor_shares + self.processor_core_shares == 0:
|
|
179
|
+
raise EmptyMachineDivisionError()
|
|
180
|
+
|
|
181
|
+
def shares_device(self, *, device: Device) -> bool:
|
|
182
|
+
"""
|
|
183
|
+
:param device: A kind of device.
|
|
184
|
+
:return: Whether the division has a share that computes on it.
|
|
185
|
+
"""
|
|
186
|
+
if device is Device.GRAPHICS_PROCESSOR:
|
|
187
|
+
return self.graphics_processor_shares > 0
|
|
188
|
+
return self.processor_core_shares > 0
|
|
189
|
+
|
|
190
|
+
def shares(self, *, machine: Machine) -> list[MachineShare]:
|
|
191
|
+
"""
|
|
192
|
+
:param machine: The machine to divide.
|
|
193
|
+
:return: The shares of the GPU, then the shares of the processor cores, each
|
|
194
|
+
with its own equal part of the cores and of the free memory of its device.
|
|
195
|
+
:raises MissingGraphicsProcessorError: If shares of the GPU are asked of a
|
|
196
|
+
machine without one.
|
|
197
|
+
"""
|
|
198
|
+
if self.graphics_processor_shares and machine.graphics_processor_memory is None:
|
|
199
|
+
raise MissingGraphicsProcessorError(
|
|
200
|
+
graphics_processor_shares=self.graphics_processor_shares
|
|
201
|
+
)
|
|
202
|
+
count = self.graphics_processor_shares + self.processor_core_shares
|
|
203
|
+
cores = sorted(machine.processor_cores)
|
|
204
|
+
cores_per_share = len(cores) // count
|
|
205
|
+
host_share = int(
|
|
206
|
+
self.host_memory_fraction * machine.host_memory.free_bytes() / count
|
|
207
|
+
)
|
|
208
|
+
graphics_processor_share = (
|
|
209
|
+
int(
|
|
210
|
+
self.graphics_processor_memory_fraction
|
|
211
|
+
* machine.graphics_processor_memory.free_bytes()
|
|
212
|
+
/ self.graphics_processor_shares
|
|
213
|
+
)
|
|
214
|
+
if self.graphics_processor_shares
|
|
215
|
+
else 0
|
|
216
|
+
)
|
|
217
|
+
return [
|
|
218
|
+
MachineShare(
|
|
219
|
+
device=(
|
|
220
|
+
Device.GRAPHICS_PROCESSOR
|
|
221
|
+
if index < self.graphics_processor_shares
|
|
222
|
+
else Device.CENTRAL_PROCESSOR
|
|
223
|
+
),
|
|
224
|
+
processor_cores=set(
|
|
225
|
+
cores[index * cores_per_share : (index + 1) * cores_per_share]
|
|
226
|
+
),
|
|
227
|
+
memory_budget_bytes=(
|
|
228
|
+
graphics_processor_share
|
|
229
|
+
if index < self.graphics_processor_shares
|
|
230
|
+
else host_share
|
|
231
|
+
),
|
|
232
|
+
host_memory_limit_bytes=int(self.host_memory_limit_shares * host_share),
|
|
233
|
+
)
|
|
234
|
+
for index in range(count)
|
|
235
|
+
]
|
|
@@ -0,0 +1,181 @@
|
|
|
1
|
+
"""
|
|
2
|
+
The memory of the devices arrays are stored on: how much is free, how much is in use, and
|
|
3
|
+
the most that was in use at once.
|
|
4
|
+
"""
|
|
5
|
+
|
|
6
|
+
from __future__ import annotations
|
|
7
|
+
|
|
8
|
+
# %% imports
|
|
9
|
+
from abc import ABC, abstractmethod
|
|
10
|
+
from dataclasses import dataclass
|
|
11
|
+
from enum import StrEnum
|
|
12
|
+
|
|
13
|
+
from krrood.patterns.classproperty import classproperty
|
|
14
|
+
from krrood.patterns.path_enumeration import PathEnumeration
|
|
15
|
+
|
|
16
|
+
from krrood_compute_resources.exceptions import MissingMemoryFieldError
|
|
17
|
+
|
|
18
|
+
# %% linux memory files
|
|
19
|
+
|
|
20
|
+
|
|
21
|
+
class LinuxMemoryFile(PathEnumeration):
|
|
22
|
+
"""
|
|
23
|
+
The files Linux reports memory in.
|
|
24
|
+
"""
|
|
25
|
+
|
|
26
|
+
MEMORY_INFORMATION = "/proc/meminfo"
|
|
27
|
+
PROCESS_STATUS = "/proc/self/status"
|
|
28
|
+
CLEAR_REFERENCES = "/proc/self/clear_refs"
|
|
29
|
+
CONTROL_GROUP_LIMIT = "/sys/fs/cgroup/memory.max"
|
|
30
|
+
CONTROL_GROUP_USAGE = "/sys/fs/cgroup/memory.current"
|
|
31
|
+
|
|
32
|
+
|
|
33
|
+
class LinuxMemoryField(StrEnum):
|
|
34
|
+
"""
|
|
35
|
+
The fields of :attr:`LinuxMemoryFile.MEMORY_INFORMATION` and
|
|
36
|
+
:attr:`LinuxMemoryFile.PROCESS_STATUS` that are read, each in kibibytes.
|
|
37
|
+
"""
|
|
38
|
+
|
|
39
|
+
AVAILABLE = "MemAvailable"
|
|
40
|
+
RESIDENT = "VmRSS"
|
|
41
|
+
PEAK_RESIDENT = "VmHWM"
|
|
42
|
+
|
|
43
|
+
|
|
44
|
+
class LinuxMemoryValue(StrEnum):
|
|
45
|
+
"""
|
|
46
|
+
Fixed values of the Linux memory files.
|
|
47
|
+
"""
|
|
48
|
+
|
|
49
|
+
UNLIMITED = "max"
|
|
50
|
+
"""
|
|
51
|
+
The content of :attr:`LinuxMemoryFile.CONTROL_GROUP_LIMIT` without a limit.
|
|
52
|
+
"""
|
|
53
|
+
|
|
54
|
+
RESET_PEAK_RESIDENT = "5"
|
|
55
|
+
"""
|
|
56
|
+
Written to :attr:`LinuxMemoryFile.CLEAR_REFERENCES`, resets the peak resident memory.
|
|
57
|
+
"""
|
|
58
|
+
|
|
59
|
+
|
|
60
|
+
# %% memory of a device
|
|
61
|
+
|
|
62
|
+
|
|
63
|
+
class DeviceMemory(ABC):
|
|
64
|
+
"""
|
|
65
|
+
The memory of a device that arrays are stored on.
|
|
66
|
+
"""
|
|
67
|
+
|
|
68
|
+
@abstractmethod
|
|
69
|
+
def free_bytes(self) -> int:
|
|
70
|
+
"""
|
|
71
|
+
:return: The number of bytes that can still be allocated on the device.
|
|
72
|
+
"""
|
|
73
|
+
|
|
74
|
+
|
|
75
|
+
class PeakTrackingMemory(DeviceMemory, ABC):
|
|
76
|
+
"""
|
|
77
|
+
Device memory that records the most memory in use since its peak was last reset.
|
|
78
|
+
"""
|
|
79
|
+
|
|
80
|
+
@abstractmethod
|
|
81
|
+
def in_use_bytes(self) -> int:
|
|
82
|
+
"""
|
|
83
|
+
:return: The number of bytes in use now.
|
|
84
|
+
"""
|
|
85
|
+
|
|
86
|
+
@abstractmethod
|
|
87
|
+
def peak_bytes(self) -> int:
|
|
88
|
+
"""
|
|
89
|
+
:return: The most bytes in use at once since :meth:`reset_peak`.
|
|
90
|
+
"""
|
|
91
|
+
|
|
92
|
+
@abstractmethod
|
|
93
|
+
def reset_peak(self) -> None:
|
|
94
|
+
"""
|
|
95
|
+
Start recording the peak anew from the memory in use now.
|
|
96
|
+
"""
|
|
97
|
+
|
|
98
|
+
|
|
99
|
+
@dataclass(kw_only=True)
|
|
100
|
+
class HostMemory(PeakTrackingMemory):
|
|
101
|
+
"""
|
|
102
|
+
The main memory of a Linux host, as seen by this process and its control group.
|
|
103
|
+
"""
|
|
104
|
+
|
|
105
|
+
@classproperty
|
|
106
|
+
def bytes_per_kibibyte(cls) -> int:
|
|
107
|
+
"""
|
|
108
|
+
:return: The unit the Linux memory fields are given in, in bytes.
|
|
109
|
+
"""
|
|
110
|
+
return 1024
|
|
111
|
+
|
|
112
|
+
def free_bytes(self) -> int:
|
|
113
|
+
available = self.field_bytes(
|
|
114
|
+
memory_file=LinuxMemoryFile.MEMORY_INFORMATION,
|
|
115
|
+
memory_field=LinuxMemoryField.AVAILABLE,
|
|
116
|
+
)
|
|
117
|
+
if not LinuxMemoryFile.CONTROL_GROUP_LIMIT.exists():
|
|
118
|
+
return available
|
|
119
|
+
limit = LinuxMemoryFile.CONTROL_GROUP_LIMIT.read_text().strip()
|
|
120
|
+
if limit == LinuxMemoryValue.UNLIMITED:
|
|
121
|
+
return available
|
|
122
|
+
usage = int(LinuxMemoryFile.CONTROL_GROUP_USAGE.read_text())
|
|
123
|
+
return min(available, int(limit) - usage)
|
|
124
|
+
|
|
125
|
+
def in_use_bytes(self) -> int:
|
|
126
|
+
return self.field_bytes(
|
|
127
|
+
memory_file=LinuxMemoryFile.PROCESS_STATUS,
|
|
128
|
+
memory_field=LinuxMemoryField.RESIDENT,
|
|
129
|
+
)
|
|
130
|
+
|
|
131
|
+
def peak_bytes(self) -> int:
|
|
132
|
+
return self.field_bytes(
|
|
133
|
+
memory_file=LinuxMemoryFile.PROCESS_STATUS,
|
|
134
|
+
memory_field=LinuxMemoryField.PEAK_RESIDENT,
|
|
135
|
+
)
|
|
136
|
+
|
|
137
|
+
def reset_peak(self) -> None:
|
|
138
|
+
LinuxMemoryFile.CLEAR_REFERENCES.write_text(
|
|
139
|
+
LinuxMemoryValue.RESET_PEAK_RESIDENT
|
|
140
|
+
)
|
|
141
|
+
|
|
142
|
+
def field_bytes(
|
|
143
|
+
self, *, memory_file: LinuxMemoryFile, memory_field: LinuxMemoryField
|
|
144
|
+
) -> int:
|
|
145
|
+
"""
|
|
146
|
+
:param memory_file: The file that holds the field, one ``name: value kB`` line
|
|
147
|
+
per field.
|
|
148
|
+
:param memory_field: The field to read.
|
|
149
|
+
:return: The value of the field, in bytes.
|
|
150
|
+
"""
|
|
151
|
+
for line in memory_file.read_text().splitlines():
|
|
152
|
+
name, value = line.split(":", 1)
|
|
153
|
+
if name == memory_field:
|
|
154
|
+
return int(value.split()[0]) * self.bytes_per_kibibyte
|
|
155
|
+
raise MissingMemoryFieldError(
|
|
156
|
+
memory_file=str(memory_file), memory_field=memory_field
|
|
157
|
+
)
|
|
158
|
+
|
|
159
|
+
|
|
160
|
+
# %% memory budget
|
|
161
|
+
|
|
162
|
+
|
|
163
|
+
@dataclass(kw_only=True)
|
|
164
|
+
class BudgetedMemory(DeviceMemory):
|
|
165
|
+
"""
|
|
166
|
+
The memory of a device of which a process may use at most a fixed amount, since other
|
|
167
|
+
processes share the device.
|
|
168
|
+
"""
|
|
169
|
+
|
|
170
|
+
memory: DeviceMemory
|
|
171
|
+
"""
|
|
172
|
+
The memory of the device.
|
|
173
|
+
"""
|
|
174
|
+
|
|
175
|
+
budget_bytes: int
|
|
176
|
+
"""
|
|
177
|
+
The most bytes the process may allocate.
|
|
178
|
+
"""
|
|
179
|
+
|
|
180
|
+
def free_bytes(self) -> int:
|
|
181
|
+
return min(self.memory.free_bytes(), self.budget_bytes)
|
|
@@ -0,0 +1,39 @@
|
|
|
1
|
+
"""
|
|
2
|
+
The memory of the CUDA devices torch stores tensors on.
|
|
3
|
+
"""
|
|
4
|
+
|
|
5
|
+
from __future__ import annotations
|
|
6
|
+
|
|
7
|
+
# %% imports
|
|
8
|
+
from dataclasses import dataclass
|
|
9
|
+
|
|
10
|
+
import torch
|
|
11
|
+
|
|
12
|
+
from krrood_compute_resources.memory import PeakTrackingMemory
|
|
13
|
+
|
|
14
|
+
# %% memory of a torch device
|
|
15
|
+
|
|
16
|
+
|
|
17
|
+
@dataclass(kw_only=True)
|
|
18
|
+
class TorchDeviceMemory(PeakTrackingMemory):
|
|
19
|
+
"""
|
|
20
|
+
The memory of a CUDA device torch stores tensors on.
|
|
21
|
+
"""
|
|
22
|
+
|
|
23
|
+
device: torch.device
|
|
24
|
+
"""
|
|
25
|
+
The CUDA device.
|
|
26
|
+
"""
|
|
27
|
+
|
|
28
|
+
def free_bytes(self) -> int:
|
|
29
|
+
free, _ = torch.cuda.mem_get_info(self.device)
|
|
30
|
+
return free
|
|
31
|
+
|
|
32
|
+
def in_use_bytes(self) -> int:
|
|
33
|
+
return torch.cuda.memory_allocated(self.device)
|
|
34
|
+
|
|
35
|
+
def peak_bytes(self) -> int:
|
|
36
|
+
return torch.cuda.max_memory_allocated(self.device)
|
|
37
|
+
|
|
38
|
+
def reset_peak(self) -> None:
|
|
39
|
+
torch.cuda.reset_peak_memory_stats(self.device)
|
|
@@ -0,0 +1,41 @@
|
|
|
1
|
+
Metadata-Version: 2.4
|
|
2
|
+
Name: krrood_compute_resources
|
|
3
|
+
Version: 26.10.0
|
|
4
|
+
Summary: The processor cores, memory and GPUs a computation runs on, and how a machine is divided among processes.
|
|
5
|
+
Author-email: Tom Schierenbeck <tom_sch@uni-bremen.de>
|
|
6
|
+
License-Expression: LGPL-3.0-only
|
|
7
|
+
Project-URL: Source, https://github.com/cram2/cognitive_robot_abstract_machine
|
|
8
|
+
Project-URL: Bug Tracker, https://github.com/cram2/cognitive_robot_abstract_machine/issues
|
|
9
|
+
Project-URL: Changelog, https://github.com/cram2/cognitive_robot_abstract_machine/blob/main/CHANGELOG.md
|
|
10
|
+
Keywords: compute resources,memory,gpu,scheduling
|
|
11
|
+
Classifier: Programming Language :: Python :: 3
|
|
12
|
+
Classifier: Operating System :: POSIX :: Linux
|
|
13
|
+
Requires-Python: <3.13,>=3.12
|
|
14
|
+
Description-Content-Type: text/markdown
|
|
15
|
+
License-File: LICENSE
|
|
16
|
+
Requires-Dist: krrood>=26.10.0
|
|
17
|
+
Requires-Dist: typing_extensions
|
|
18
|
+
Provides-Extra: jax
|
|
19
|
+
Requires-Dist: jax>=0.6.0; extra == "jax"
|
|
20
|
+
Provides-Extra: torch
|
|
21
|
+
Requires-Dist: torch>=2.4; extra == "torch"
|
|
22
|
+
Provides-Extra: dev
|
|
23
|
+
Requires-Dist: pytest<8.0.0,>=7.0.0; extra == "dev"
|
|
24
|
+
Dynamic: license-file
|
|
25
|
+
|
|
26
|
+
# krrood_compute_resources
|
|
27
|
+
|
|
28
|
+
The processor cores, memory and GPUs a computation runs on, and how a machine is divided
|
|
29
|
+
among processes that run side by side.
|
|
30
|
+
|
|
31
|
+
- `krrood_compute_resources.devices`: `Device`, the kinds of device a computation runs on.
|
|
32
|
+
- `krrood_compute_resources.memory`: `DeviceMemory`, the free memory of a device;
|
|
33
|
+
`HostMemory`, the main memory of a Linux host within its control group; and
|
|
34
|
+
`BudgetedMemory`, the share of a device's memory one process may use.
|
|
35
|
+
- `krrood_compute_resources.machine`: `Machine`, the cores and memory of the machine a
|
|
36
|
+
process runs on; `MachineDivision`, which divides it into `MachineShare`s; and the
|
|
37
|
+
environment and core affinity that keep a process within its share.
|
|
38
|
+
- `krrood_compute_resources.jax_memory`, with the extra `krrood_compute_resources[jax]`:
|
|
39
|
+
`JaxDeviceMemory`, the memory of an accelerator jax stores arrays on.
|
|
40
|
+
- `krrood_compute_resources.torch_memory`, with the extra `krrood_compute_resources[torch]`:
|
|
41
|
+
`TorchDeviceMemory`, the memory of a CUDA device torch stores tensors on.
|
|
@@ -0,0 +1,16 @@
|
|
|
1
|
+
LICENSE
|
|
2
|
+
README.md
|
|
3
|
+
pyproject.toml
|
|
4
|
+
src/krrood_compute_resources/__init__.py
|
|
5
|
+
src/krrood_compute_resources/_version.py
|
|
6
|
+
src/krrood_compute_resources/devices.py
|
|
7
|
+
src/krrood_compute_resources/exceptions.py
|
|
8
|
+
src/krrood_compute_resources/jax_memory.py
|
|
9
|
+
src/krrood_compute_resources/machine.py
|
|
10
|
+
src/krrood_compute_resources/memory.py
|
|
11
|
+
src/krrood_compute_resources/torch_memory.py
|
|
12
|
+
src/krrood_compute_resources.egg-info/PKG-INFO
|
|
13
|
+
src/krrood_compute_resources.egg-info/SOURCES.txt
|
|
14
|
+
src/krrood_compute_resources.egg-info/dependency_links.txt
|
|
15
|
+
src/krrood_compute_resources.egg-info/requires.txt
|
|
16
|
+
src/krrood_compute_resources.egg-info/top_level.txt
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
krrood_compute_resources
|