interval/bpe_framework
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Initial commit

Browse Source
This commit is contained in:
Tim O\'Neil 2025-08-21 10:44:50 -07:00
parent 8dac923ee9
commit d83d07a823
22 changed files with 2889 additions and 229 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

112
CMakeLists.txt Executable file
View File

@ -0,0 +1,112 @@
cmake_minimum_required(VERSION 3.14)
project(lm_framework LANGUAGES CXX)
# Set default build type to Release if not specified
if(NOT CMAKE_BUILD_TYPE)
set(CMAKE_BUILD_TYPE Release)
message(STATUS "Build type not specified, defaulting to Release")
endif()
set(CMAKE_CXX_STANDARD 17)
set(CMAKE_CXX_STANDARD_REQUIRED ON)
set(CMAKE_CXX_EXTENSIONS OFF)
# Enable cross-directory linking
if(POLICY CMP0079)
cmake_policy(SET CMP0079 NEW)
endif()
# Include directories
include_directories(
${CMAKE_CURRENT_SOURCE_DIR}/include
)
# Find dependencies
find_package(nlohmann_json 3.9 REQUIRED)
find_package(ICU REQUIRED COMPONENTS uc i18n)
# GoogleTest
include(FetchContent)
FetchContent_Declare(
googletest
GIT_REPOSITORY https://github.com/google/googletest.git
GIT_TAG release-1.11.0
)
FetchContent_MakeAvailable(googletest)
# Add subdirectories
add_subdirectory(src/tokenizer)
add_subdirectory(src/runtime)
# Main library
add_library(lm_core
src/runtime/init.cpp
src/runtime/shutdown.cpp
)
target_link_libraries(lm_core
PRIVATE
lm_tokenizer
nlohmann_json::nlohmann_json
)
# Set optimization flags for the core library
if(CMAKE_COMPILER_IS_GNUCXX OR CMAKE_CXX_COMPILER_ID MATCHES "Clang")
target_compile_options(lm_core PRIVATE -O3)
if(CMAKE_BUILD_TYPE STREQUAL "Release")
target_compile_options(lm_core PRIVATE -DNDEBUG)
endif()
endif()
# Test executables
add_executable(test_bpe src/test_bpe.cpp)
target_link_libraries(test_bpe
PRIVATE
lm_core
GTest::gtest_main
)
add_executable(test_unicode_bpe src/test_unicode_bpe.cpp)
target_link_libraries(test_unicode_bpe
PRIVATE
lm_core
GTest::gtest_main
)
# Alpha prototype executable
add_executable(lm_alpha
src/alpha/repl.cpp
src/alpha/config_io.cpp
)
target_link_libraries(lm_alpha
PRIVATE
lm_core
nlohmann_json::nlohmann_json
)
# Install targets
install(TARGETS lm_core DESTINATION lib)
install(DIRECTORY include/ DESTINATION include)
# Performance testing target
add_executable(performance_test src/performance_test.cpp)
target_link_libraries(performance_test
PRIVATE
lm_core
GTest::gtest_main
)
# Add compiler warning flags
if(CMAKE_COMPILER_IS_GNUCXX OR CMAKE_CXX_COMPILER_ID MATCHES "Clang")
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wall -Wextra -Wpedantic -Werror")
endif()
# Add coverage flags for debug builds
if(CMAKE_BUILD_TYPE STREQUAL "Debug")
if(CMAKE_COMPILER_IS_GNUCXX)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} --coverage")
elseif(CMAKE_CXX_COMPILER_ID MATCHES "Clang")
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fprofile-instr-generate -fcoverage-mapping")
endif()
endif()

885
LICENSE
View File

@ -1,232 +1,661 @@
GNU GENERAL PUBLIC LICENSE GNU AFFERO GENERAL PUBLIC LICENSE
Version 3, 29 June 2007 Version 3, 19 November 2007
Copyright © 2007 Free Software Foundation, Inc. <https://fsf.org/> Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
Everyone is permitted to copy and distribute verbatim copies
Everyone is permitted to copy and distribute verbatim copies of this license document, but changing it is not allowed. of this license document, but changing it is not allowed.
Preamble Preamble
The GNU General Public License is a free, copyleft license for software and other kinds of works. The GNU Affero General Public License is a free, copyleft license for
software and other kinds of works, specifically designed to ensure
The licenses for most software and other practical works are designed to take away your freedom to share and change the works. By contrast, the GNU General Public License is intended to guarantee your freedom to share and change all versions of a program--to make sure it remains free software for all its users. We, the Free Software Foundation, use the GNU General Public License for most of our software; it applies also to any other work released this way by its authors. You can apply it to your programs, too. cooperation with the community in the case of network server software.
When we speak of free software, we are referring to freedom, not price. Our General Public Licenses are designed to make sure that you have the freedom to distribute copies of free software (and charge for them if you wish), that you receive source code or can get it if you want it, that you can change the software or use pieces of it in new free programs, and that you know you can do these things. The licenses for most software and other practical works are designed
to take away your freedom to share and change the works. By contrast,
To protect your rights, we need to prevent others from denying you these rights or asking you to surrender the rights. Therefore, you have certain responsibilities if you distribute copies of the software, or if you modify it: responsibilities to respect the freedom of others. our General Public Licenses are intended to guarantee your freedom to
share and change all versions of a program--to make sure it remains free
For example, if you distribute copies of such a program, whether gratis or for a fee, you must pass on to the recipients the same freedoms that you received. You must make sure that they, too, receive or can get the source code. And you must show them these terms so they know their rights. software for all its users.
Developers that use the GNU GPL protect your rights with two steps: (1) assert copyright on the software, and (2) offer you this License giving you legal permission to copy, distribute and/or modify it. When we speak of free software, we are referring to freedom, not
price. Our General Public Licenses are designed to make sure that you
For the developers' and authors' protection, the GPL clearly explains that there is no warranty for this free software. For both users' and authors' sake, the GPL requires that modified versions be marked as changed, so that their problems will not be attributed erroneously to authors of previous versions. have the freedom to distribute copies of free software (and charge for
them if you wish), that you receive source code or can get it if you
Some devices are designed to deny users access to install or run modified versions of the software inside them, although the manufacturer can do so. This is fundamentally incompatible with the aim of protecting users' freedom to change the software. The systematic pattern of such abuse occurs in the area of products for individuals to use, which is precisely where it is most unacceptable. Therefore, we have designed this version of the GPL to prohibit the practice for those products. If such problems arise substantially in other domains, we stand ready to extend this provision to those domains in future versions of the GPL, as needed to protect the freedom of users. want it, that you can change the software or use pieces of it in new
free programs, and that you know you can do these things.
Finally, every program is threatened constantly by software patents. States should not allow patents to restrict development and use of software on general-purpose computers, but in those that do, we wish to avoid the special danger that patents applied to a free program could make it effectively proprietary. To prevent this, the GPL assures that patents cannot be used to render the program non-free.
Developers that use our General Public Licenses protect your rights
The precise terms and conditions for copying, distribution and modification follow. with two steps: (1) assert copyright on the software, and (2) offer
you this License which gives you legal permission to copy, distribute
TERMS AND CONDITIONS and/or modify the software.
0. Definitions. A secondary benefit of defending all users' freedom is that
improvements made in alternate versions of the program, if they
“This License” refers to version 3 of the GNU General Public License. receive widespread use, become available for other developers to
incorporate. Many developers of free software are heartened and
“Copyright” also means copyright-like laws that apply to other kinds of works, such as semiconductor masks. encouraged by the resulting cooperation. However, in the case of
software used on network servers, this result may fail to come about.
“The Program” refers to any copyrightable work licensed under this License. Each licensee is addressed as “you”. “Licensees” and “recipients” may be individuals or organizations. The GNU General Public License permits making a modified version and
letting the public access it on a server without ever releasing its
To “modify” a work means to copy from or adapt all or part of the work in a fashion requiring copyright permission, other than the making of an exact copy. The resulting work is called a “modified version” of the earlier work or a work “based on” the earlier work. source code to the public.
A “covered work” means either the unmodified Program or a work based on the Program. The GNU Affero General Public License is designed specifically to
ensure that, in such cases, the modified source code becomes available
To “propagate” a work means to do anything with it that, without permission, would make you directly or secondarily liable for infringement under applicable copyright law, except executing it on a computer or modifying a private copy. Propagation includes copying, distribution (with or without modification), making available to the public, and in some countries other activities as well. to the community. It requires the operator of a network server to
provide the source code of the modified version running there to the
To “convey” a work means any kind of propagation that enables other parties to make or receive copies. Mere interaction with a user through a computer network, with no transfer of a copy, is not conveying. users of that server. Therefore, public use of a modified version, on
a publicly accessible server, gives the public access to the source
An interactive user interface displays “Appropriate Legal Notices” to the extent that it includes a convenient and prominently visible feature that (1) displays an appropriate copyright notice, and (2) tells the user that there is no warranty for the work (except to the extent that warranties are provided), that licensees may convey the work under this License, and how to view a copy of this License. If the interface presents a list of user commands or options, such as a menu, a prominent item in the list meets this criterion. code of the modified version.
1. Source Code. An older license, called the Affero General Public License and
The “source code” for a work means the preferred form of the work for making modifications to it. “Object code” means any non-source form of a work. published by Affero, was designed to accomplish similar goals. This is
a different license, not a version of the Affero GPL, but Affero has
A “Standard Interface” means an interface that either is an official standard defined by a recognized standards body, or, in the case of interfaces specified for a particular programming language, one that is widely used among developers working in that language. released a new version of the Affero GPL which permits relicensing under
this license.
The “System Libraries” of an executable work include anything, other than the work as a whole, that (a) is included in the normal form of packaging a Major Component, but which is not part of that Major Component, and (b) serves only to enable use of the work with that Major Component, or to implement a Standard Interface for which an implementation is available to the public in source code form. A “Major Component”, in this context, means a major essential component (kernel, window system, and so on) of the specific operating system (if any) on which the executable work runs, or a compiler used to produce the work, or an object code interpreter used to run it.
The precise terms and conditions for copying, distribution and
The “Corresponding Source” for a work in object code form means all the source code needed to generate, install, and (for an executable work) run the object code and to modify the work, including scripts to control those activities. However, it does not include the work's System Libraries, or general-purpose tools or generally available free programs which are used unmodified in performing those activities but which are not part of the work. For example, Corresponding Source includes interface definition files associated with source files for the work, and the source code for shared libraries and dynamically linked subprograms that the work is specifically designed to require, such as by intimate data communication or control flow between those subprograms and other parts of the work. modification follow.
The Corresponding Source need not include anything that users can regenerate automatically from other parts of the Corresponding Source. TERMS AND CONDITIONS
The Corresponding Source for a work in source code form is that same work. 0. Definitions.
2. Basic Permissions. "This License" refers to version 3 of the GNU Affero General Public License.
All rights granted under this License are granted for the term of copyright on the Program, and are irrevocable provided the stated conditions are met. This License explicitly affirms your unlimited permission to run the unmodified Program. The output from running a covered work is covered by this License only if the output, given its content, constitutes a covered work. This License acknowledges your rights of fair use or other equivalent, as provided by copyright law.
"Copyright" also means copyright-like laws that apply to other kinds of
You may make, run and propagate covered works that you do not convey, without conditions so long as your license otherwise remains in force. You may convey covered works to others for the sole purpose of having them make modifications exclusively for you, or provide you with facilities for running those works, provided that you comply with the terms of this License in conveying all material for which you do not control copyright. Those thus making or running the covered works for you must do so exclusively on your behalf, under your direction and control, on terms that prohibit them from making any copies of your copyrighted material outside their relationship with you. works, such as semiconductor masks.
Conveying under any other circumstances is permitted solely under the conditions stated below. Sublicensing is not allowed; section 10 makes it unnecessary. "The Program" refers to any copyrightable work licensed under this
License. Each licensee is addressed as "you". "Licensees" and
3. Protecting Users' Legal Rights From Anti-Circumvention Law. "recipients" may be individuals or organizations.
No covered work shall be deemed part of an effective technological measure under any applicable law fulfilling obligations under article 11 of the WIPO copyright treaty adopted on 20 December 1996, or similar laws prohibiting or restricting circumvention of such measures.
To "modify" a work means to copy from or adapt all or part of the work
When you convey a covered work, you waive any legal power to forbid circumvention of technological measures to the extent such circumvention is effected by exercising rights under this License with respect to the covered work, and you disclaim any intention to limit operation or modification of the work as a means of enforcing, against the work's users, your or third parties' legal rights to forbid circumvention of technological measures. in a fashion requiring copyright permission, other than the making of an
exact copy. The resulting work is called a "modified version" of the
4. Conveying Verbatim Copies. earlier work or a work "based on" the earlier work.
You may convey verbatim copies of the Program's source code as you receive it, in any medium, provided that you conspicuously and appropriately publish on each copy an appropriate copyright notice; keep intact all notices stating that this License and any non-permissive terms added in accord with section 7 apply to the code; keep intact all notices of the absence of any warranty; and give all recipients a copy of this License along with the Program.
A "covered work" means either the unmodified Program or a work based
You may charge any price or no price for each copy that you convey, and you may offer support or warranty protection for a fee. on the Program.
5. Conveying Modified Source Versions. To "propagate" a work means to do anything with it that, without
You may convey a work based on the Program, or the modifications to produce it from the Program, in the form of source code under the terms of section 4, provided that you also meet all of these conditions: permission, would make you directly or secondarily liable for
infringement under applicable copyright law, except executing it on a
a) The work must carry prominent notices stating that you modified it, and giving a relevant date. computer or modifying a private copy. Propagation includes copying,
distribution (with or without modification), making available to the
b) The work must carry prominent notices stating that it is released under this License and any conditions added under section 7. This requirement modifies the requirement in section 4 to “keep intact all notices”. public, and in some countries other activities as well.
c) You must license the entire work, as a whole, under this License to anyone who comes into possession of a copy. This License will therefore apply, along with any applicable section 7 additional terms, to the whole of the work, and all its parts, regardless of how they are packaged. This License gives no permission to license the work in any other way, but it does not invalidate such permission if you have separately received it. To "convey" a work means any kind of propagation that enables other
parties to make or receive copies. Mere interaction with a user through
d) If the work has interactive user interfaces, each must display Appropriate Legal Notices; however, if the Program has interactive interfaces that do not display Appropriate Legal Notices, your work need not make them do so. a computer network, with no transfer of a copy, is not conveying.
A compilation of a covered work with other separate and independent works, which are not by their nature extensions of the covered work, and which are not combined with it such as to form a larger program, in or on a volume of a storage or distribution medium, is called an “aggregate” if the compilation and its resulting copyright are not used to limit the access or legal rights of the compilation's users beyond what the individual works permit. Inclusion of a covered work in an aggregate does not cause this License to apply to the other parts of the aggregate. An interactive user interface displays "Appropriate Legal Notices"
to the extent that it includes a convenient and prominently visible
6. Conveying Non-Source Forms. feature that (1) displays an appropriate copyright notice, and (2)
You may convey a covered work in object code form under the terms of sections 4 and 5, provided that you also convey the machine-readable Corresponding Source under the terms of this License, in one of these ways: tells the user that there is no warranty for the work (except to the
extent that warranties are provided), that licensees may convey the
a) Convey the object code in, or embodied in, a physical product (including a physical distribution medium), accompanied by the Corresponding Source fixed on a durable physical medium customarily used for software interchange. work under this License, and how to view a copy of this License. If
the interface presents a list of user commands or options, such as a
b) Convey the object code in, or embodied in, a physical product (including a physical distribution medium), accompanied by a written offer, valid for at least three years and valid for as long as you offer spare parts or customer support for that product model, to give anyone who possesses the object code either (1) a copy of the Corresponding Source for all the software in the product that is covered by this License, on a durable physical medium customarily used for software interchange, for a price no more than your reasonable cost of physically performing this conveying of source, or (2) access to copy the Corresponding Source from a network server at no charge. menu, a prominent item in the list meets this criterion.
c) Convey individual copies of the object code with a copy of the written offer to provide the Corresponding Source. This alternative is allowed only occasionally and noncommercially, and only if you received the object code with such an offer, in accord with subsection 6b. 1. Source Code.
d) Convey the object code by offering access from a designated place (gratis or for a charge), and offer equivalent access to the Corresponding Source in the same way through the same place at no further charge. You need not require recipients to copy the Corresponding Source along with the object code. If the place to copy the object code is a network server, the Corresponding Source may be on a different server (operated by you or a third party) that supports equivalent copying facilities, provided you maintain clear directions next to the object code saying where to find the Corresponding Source. Regardless of what server hosts the Corresponding Source, you remain obligated to ensure that it is available for as long as needed to satisfy these requirements. The "source code" for a work means the preferred form of the work
for making modifications to it. "Object code" means any non-source
e) Convey the object code using peer-to-peer transmission, provided you inform other peers where the object code and Corresponding Source of the work are being offered to the general public at no charge under subsection 6d. form of a work.
A separable portion of the object code, whose source code is excluded from the Corresponding Source as a System Library, need not be included in conveying the object code work. A "Standard Interface" means an interface that either is an official
standard defined by a recognized standards body, or, in the case of
A “User Product” is either (1) a “consumer product”, which means any tangible personal property which is normally used for personal, family, or household purposes, or (2) anything designed or sold for incorporation into a dwelling. In determining whether a product is a consumer product, doubtful cases shall be resolved in favor of coverage. For a particular product received by a particular user, “normally used” refers to a typical or common use of that class of product, regardless of the status of the particular user or of the way in which the particular user actually uses, or expects or is expected to use, the product. A product is a consumer product regardless of whether the product has substantial commercial, industrial or non-consumer uses, unless such uses represent the only significant mode of use of the product. interfaces specified for a particular programming language, one that
is widely used among developers working in that language.
“Installation Information” for a User Product means any methods, procedures, authorization keys, or other information required to install and execute modified versions of a covered work in that User Product from a modified version of its Corresponding Source. The information must suffice to ensure that the continued functioning of the modified object code is in no case prevented or interfered with solely because modification has been made.
The "System Libraries" of an executable work include anything, other
If you convey an object code work under this section in, or with, or specifically for use in, a User Product, and the conveying occurs as part of a transaction in which the right of possession and use of the User Product is transferred to the recipient in perpetuity or for a fixed term (regardless of how the transaction is characterized), the Corresponding Source conveyed under this section must be accompanied by the Installation Information. But this requirement does not apply if neither you nor any third party retains the ability to install modified object code on the User Product (for example, the work has been installed in ROM). than the work as a whole, that (a) is included in the normal form of
packaging a Major Component, but which is not part of that Major
The requirement to provide Installation Information does not include a requirement to continue to provide support service, warranty, or updates for a work that has been modified or installed by the recipient, or for the User Product in which it has been modified or installed. Access to a network may be denied when the modification itself materially and adversely affects the operation of the network or violates the rules and protocols for communication across the network. Component, and (b) serves only to enable use of the work with that
Major Component, or to implement a Standard Interface for which an
Corresponding Source conveyed, and Installation Information provided, in accord with this section must be in a format that is publicly documented (and with an implementation available to the public in source code form), and must require no special password or key for unpacking, reading or copying. implementation is available to the public in source code form. A
"Major Component", in this context, means a major essential component
7. Additional Terms. (kernel, window system, and so on) of the specific operating system
“Additional permissions” are terms that supplement the terms of this License by making exceptions from one or more of its conditions. Additional permissions that are applicable to the entire Program shall be treated as though they were included in this License, to the extent that they are valid under applicable law. If additional permissions apply only to part of the Program, that part may be used separately under those permissions, but the entire Program remains governed by this License without regard to the additional permissions. (if any) on which the executable work runs, or a compiler used to
produce the work, or an object code interpreter used to run it.
When you convey a copy of a covered work, you may at your option remove any additional permissions from that copy, or from any part of it. (Additional permissions may be written to require their own removal in certain cases when you modify the work.) You may place additional permissions on material, added by you to a covered work, for which you have or can give appropriate copyright permission.
The "Corresponding Source" for a work in object code form means all
Notwithstanding any other provision of this License, for material you add to a covered work, you may (if authorized by the copyright holders of that material) supplement the terms of this License with terms: the source code needed to generate, install, and (for an executable
work) run the object code and to modify the work, including scripts to
a) Disclaiming warranty or limiting liability differently from the terms of sections 15 and 16 of this License; or control those activities. However, it does not include the work's
System Libraries, or general-purpose tools or generally available free
b) Requiring preservation of specified reasonable legal notices or author attributions in that material or in the Appropriate Legal Notices displayed by works containing it; or programs which are used unmodified in performing those activities but
which are not part of the work. For example, Corresponding Source
c) Prohibiting misrepresentation of the origin of that material, or requiring that modified versions of such material be marked in reasonable ways as different from the original version; or includes interface definition files associated with source files for
the work, and the source code for shared libraries and dynamically
d) Limiting the use for publicity purposes of names of licensors or authors of the material; or linked subprograms that the work is specifically designed to require,
such as by intimate data communication or control flow between those
e) Declining to grant rights under trademark law for use of some trade names, trademarks, or service marks; or subprograms and other parts of the work.
f) Requiring indemnification of licensors and authors of that material by anyone who conveys the material (or modified versions of it) with contractual assumptions of liability to the recipient, for any liability that these contractual assumptions directly impose on those licensors and authors. The Corresponding Source need not include anything that users
can regenerate automatically from other parts of the Corresponding
All other non-permissive additional terms are considered “further restrictions” within the meaning of section 10. If the Program as you received it, or any part of it, contains a notice stating that it is governed by this License along with a term that is a further restriction, you may remove that term. If a license document contains a further restriction but permits relicensing or conveying under this License, you may add to a covered work material governed by the terms of that license document, provided that the further restriction does not survive such relicensing or conveying. Source.
If you add terms to a covered work in accord with this section, you must place, in the relevant source files, a statement of the additional terms that apply to those files, or a notice indicating where to find the applicable terms. The Corresponding Source for a work in source code form is that
same work.
Additional terms, permissive or non-permissive, may be stated in the form of a separately written license, or stated as exceptions; the above requirements apply either way.
2. Basic Permissions.
8. Termination.
You may not propagate or modify a covered work except as expressly provided under this License. Any attempt otherwise to propagate or modify it is void, and will automatically terminate your rights under this License (including any patent licenses granted under the third paragraph of section 11). All rights granted under this License are granted for the term of
copyright on the Program, and are irrevocable provided the stated
However, if you cease all violation of this License, then your license from a particular copyright holder is reinstated (a) provisionally, unless and until the copyright holder explicitly and finally terminates your license, and (b) permanently, if the copyright holder fails to notify you of the violation by some reasonable means prior to 60 days after the cessation. conditions are met. This License explicitly affirms your unlimited
permission to run the unmodified Program. The output from running a
Moreover, your license from a particular copyright holder is reinstated permanently if the copyright holder notifies you of the violation by some reasonable means, this is the first time you have received notice of violation of this License (for any work) from that copyright holder, and you cure the violation prior to 30 days after your receipt of the notice. covered work is covered by this License only if the output, given its
content, constitutes a covered work. This License acknowledges your
Termination of your rights under this section does not terminate the licenses of parties who have received copies or rights from you under this License. If your rights have been terminated and not permanently reinstated, you do not qualify to receive new licenses for the same material under section 10. rights of fair use or other equivalent, as provided by copyright law.
9. Acceptance Not Required for Having Copies. You may make, run and propagate covered works that you do not
You are not required to accept this License in order to receive or run a copy of the Program. Ancillary propagation of a covered work occurring solely as a consequence of using peer-to-peer transmission to receive a copy likewise does not require acceptance. However, nothing other than this License grants you permission to propagate or modify any covered work. These actions infringe copyright if you do not accept this License. Therefore, by modifying or propagating a covered work, you indicate your acceptance of this License to do so. convey, without conditions so long as your license otherwise remains
in force. You may convey covered works to others for the sole purpose
10. Automatic Licensing of Downstream Recipients. of having them make modifications exclusively for you, or provide you
Each time you convey a covered work, the recipient automatically receives a license from the original licensors, to run, modify and propagate that work, subject to this License. You are not responsible for enforcing compliance by third parties with this License. with facilities for running those works, provided that you comply with
the terms of this License in conveying all material for which you do
An “entity transaction” is a transaction transferring control of an organization, or substantially all assets of one, or subdividing an organization, or merging organizations. If propagation of a covered work results from an entity transaction, each party to that transaction who receives a copy of the work also receives whatever licenses to the work the party's predecessor in interest had or could give under the previous paragraph, plus a right to possession of the Corresponding Source of the work from the predecessor in interest, if the predecessor has it or can get it with reasonable efforts. not control copyright. Those thus making or running the covered works
for you must do so exclusively on your behalf, under your direction
You may not impose any further restrictions on the exercise of the rights granted or affirmed under this License. For example, you may not impose a license fee, royalty, or other charge for exercise of rights granted under this License, and you may not initiate litigation (including a cross-claim or counterclaim in a lawsuit) alleging that any patent claim is infringed by making, using, selling, offering for sale, or importing the Program or any portion of it. and control, on terms that prohibit them from making any copies of
your copyrighted material outside their relationship with you.
11. Patents.
A “contributor” is a copyright holder who authorizes use under this License of the Program or a work on which the Program is based. The work thus licensed is called the contributor's “contributor version”. Conveying under any other circumstances is permitted solely under
the conditions stated below. Sublicensing is not allowed; section 10
A contributor's “essential patent claims” are all patent claims owned or controlled by the contributor, whether already acquired or hereafter acquired, that would be infringed by some manner, permitted by this License, of making, using, or selling its contributor version, but do not include claims that would be infringed only as a consequence of further modification of the contributor version. For purposes of this definition, “control” includes the right to grant patent sublicenses in a manner consistent with the requirements of this License. makes it unnecessary.
Each contributor grants you a non-exclusive, worldwide, royalty-free patent license under the contributor's essential patent claims, to make, use, sell, offer for sale, import and otherwise run, modify and propagate the contents of its contributor version. 3. Protecting Users' Legal Rights From Anti-Circumvention Law.
In the following three paragraphs, a “patent license” is any express agreement or commitment, however denominated, not to enforce a patent (such as an express permission to practice a patent or covenant not to sue for patent infringement). To “grant” such a patent license to a party means to make such an agreement or commitment not to enforce a patent against the party. No covered work shall be deemed part of an effective technological
measure under any applicable law fulfilling obligations under article
If you convey a covered work, knowingly relying on a patent license, and the Corresponding Source of the work is not available for anyone to copy, free of charge and under the terms of this License, through a publicly available network server or other readily accessible means, then you must either (1) cause the Corresponding Source to be so available, or (2) arrange to deprive yourself of the benefit of the patent license for this particular work, or (3) arrange, in a manner consistent with the requirements of this License, to extend the patent license to downstream recipients. “Knowingly relying” means you have actual knowledge that, but for the patent license, your conveying the covered work in a country, or your recipient's use of the covered work in a country, would infringe one or more identifiable patents in that country that you have reason to believe are valid. 11 of the WIPO copyright treaty adopted on 20 December 1996, or
similar laws prohibiting or restricting circumvention of such
If, pursuant to or in connection with a single transaction or arrangement, you convey, or propagate by procuring conveyance of, a covered work, and grant a patent license to some of the parties receiving the covered work authorizing them to use, propagate, modify or convey a specific copy of the covered work, then the patent license you grant is automatically extended to all recipients of the covered work and works based on it. measures.
A patent license is “discriminatory” if it does not include within the scope of its coverage, prohibits the exercise of, or is conditioned on the non-exercise of one or more of the rights that are specifically granted under this License. You may not convey a covered work if you are a party to an arrangement with a third party that is in the business of distributing software, under which you make payment to the third party based on the extent of your activity of conveying the work, and under which the third party grants, to any of the parties who would receive the covered work from you, a discriminatory patent license (a) in connection with copies of the covered work conveyed by you (or copies made from those copies), or (b) primarily for and in connection with specific products or compilations that contain the covered work, unless you entered into that arrangement, or that patent license was granted, prior to 28 March 2007. When you convey a covered work, you waive any legal power to forbid
circumvention of technological measures to the extent such circumvention
Nothing in this License shall be construed as excluding or limiting any implied license or other defenses to infringement that may otherwise be available to you under applicable patent law. is effected by exercising rights under this License with respect to
the covered work, and you disclaim any intention to limit operation or
12. No Surrender of Others' Freedom. modification of the work as a means of enforcing, against the work's
If conditions are imposed on you (whether by court order, agreement or otherwise) that contradict the conditions of this License, they do not excuse you from the conditions of this License. If you cannot convey a covered work so as to satisfy simultaneously your obligations under this License and any other pertinent obligations, then as a consequence you may not convey it at all. For example, if you agree to terms that obligate you to collect a royalty for further conveying from those to whom you convey the Program, the only way you could satisfy both those terms and this License would be to refrain entirely from conveying the Program. users, your or third parties' legal rights to forbid circumvention of
technological measures.
13. Use with the GNU Affero General Public License.
Notwithstanding any other provision of this License, you have permission to link or combine any covered work with a work licensed under version 3 of the GNU Affero General Public License into a single combined work, and to convey the resulting work. The terms of this License will continue to apply to the part which is the covered work, but the special requirements of the GNU Affero General Public License, section 13, concerning interaction through a network will apply to the combination as such. 4. Conveying Verbatim Copies.
14. Revised Versions of this License. You may convey verbatim copies of the Program's source code as you
The Free Software Foundation may publish revised and/or new versions of the GNU General Public License from time to time. Such new versions will be similar in spirit to the present version, but may differ in detail to address new problems or concerns. receive it, in any medium, provided that you conspicuously and
appropriately publish on each copy an appropriate copyright notice;
Each version is given a distinguishing version number. If the Program specifies that a certain numbered version of the GNU General Public License “or any later version” applies to it, you have the option of following the terms and conditions either of that numbered version or of any later version published by the Free Software Foundation. If the Program does not specify a version number of the GNU General Public License, you may choose any version ever published by the Free Software Foundation. keep intact all notices stating that this License and any
non-permissive terms added in accord with section 7 apply to the code;
If the Program specifies that a proxy can decide which future versions of the GNU General Public License can be used, that proxy's public statement of acceptance of a version permanently authorizes you to choose that version for the Program. keep intact all notices of the absence of any warranty; and give all
recipients a copy of this License along with the Program.
Later license versions may give you additional or different permissions. However, no additional obligations are imposed on any author or copyright holder as a result of your choosing to follow a later version.
You may charge any price or no price for each copy that you convey,
15. Disclaimer of Warranty. and you may offer support or warranty protection for a fee.
THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM “AS IS” WITHOUT WARRANTY OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
5. Conveying Modified Source Versions.
16. Limitation of Liability.
IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH DAMAGES. You may convey a work based on the Program, or the modifications to
produce it from the Program, in the form of source code under the
17. Interpretation of Sections 15 and 16. terms of section 4, provided that you also meet all of these conditions:
If the disclaimer of warranty and limitation of liability provided above cannot be given local legal effect according to their terms, reviewing courts shall apply local law that most closely approximates an absolute waiver of all civil liability in connection with the Program, unless a warranty or assumption of liability accompanies a copy of the Program in return for a fee.
a) The work must carry prominent notices stating that you modified
END OF TERMS AND CONDITIONS it, and giving a relevant date.
How to Apply These Terms to Your New Programs b) The work must carry prominent notices stating that it is
released under this License and any conditions added under section
If you develop a new program, and you want it to be of the greatest possible use to the public, the best way to achieve this is to make it free software which everyone can redistribute and change under these terms. 7. This requirement modifies the requirement in section 4 to
"keep intact all notices".
To do so, attach the following notices to the program. It is safest to attach them to the start of each source file to most effectively state the exclusion of warranty; and each file should have at least the “copyright” line and a pointer to where the full notice is found.
c) You must license the entire work, as a whole, under this
bpe_framework License to anyone who comes into possession of a copy. This
Copyright (C) 2025 interval License will therefore apply, along with any applicable section 7
additional terms, to the whole of the work, and all its parts,
This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version. regardless of how they are packaged. This License gives no
permission to license the work in any other way, but it does not
This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details. invalidate such permission if you have separately received it.
You should have received a copy of the GNU General Public License along with this program. If not, see <https://www.gnu.org/licenses/>. d) If the work has interactive user interfaces, each must display
Appropriate Legal Notices; however, if the Program has interactive
interfaces that do not display Appropriate Legal Notices, your
work need not make them do so.
A compilation of a covered work with other separate and independent
works, which are not by their nature extensions of the covered work,
and which are not combined with it such as to form a larger program,
in or on a volume of a storage or distribution medium, is called an
"aggregate" if the compilation and its resulting copyright are not
used to limit the access or legal rights of the compilation's users
beyond what the individual works permit. Inclusion of a covered work
in an aggregate does not cause this License to apply to the other
parts of the aggregate.
6. Conveying Non-Source Forms.
You may convey a covered work in object code form under the terms
of sections 4 and 5, provided that you also convey the
machine-readable Corresponding Source under the terms of this License,
in one of these ways:
a) Convey the object code in, or embodied in, a physical product
(including a physical distribution medium), accompanied by the
Corresponding Source fixed on a durable physical medium
customarily used for software interchange.
b) Convey the object code in, or embodied in, a physical product
(including a physical distribution medium), accompanied by a
written offer, valid for at least three years and valid for as
long as you offer spare parts or customer support for that product
model, to give anyone who possesses the object code either (1) a
copy of the Corresponding Source for all the software in the
product that is covered by this License, on a durable physical
medium customarily used for software interchange, for a price no
more than your reasonable cost of physically performing this
conveying of source, or (2) access to copy the
Corresponding Source from a network server at no charge.
c) Convey individual copies of the object code with a copy of the
written offer to provide the Corresponding Source. This
alternative is allowed only occasionally and noncommercially, and
only if you received the object code with such an offer, in accord
with subsection 6b.
d) Convey the object code by offering access from a designated
place (gratis or for a charge), and offer equivalent access to the
Corresponding Source in the same way through the same place at no
further charge. You need not require recipients to copy the
Corresponding Source along with the object code. If the place to
copy the object code is a network server, the Corresponding Source
may be on a different server (operated by you or a third party)
that supports equivalent copying facilities, provided you maintain
clear directions next to the object code saying where to find the
Corresponding Source. Regardless of what server hosts the
Corresponding Source, you remain obligated to ensure that it is
available for as long as needed to satisfy these requirements.
e) Convey the object code using peer-to-peer transmission, provided
you inform other peers where the object code and Corresponding
Source of the work are being offered to the general public at no
charge under subsection 6d.
A separable portion of the object code, whose source code is excluded
from the Corresponding Source as a System Library, need not be
included in conveying the object code work.
A "User Product" is either (1) a "consumer product", which means any
tangible personal property which is normally used for personal, family,
or household purposes, or (2) anything designed or sold for incorporation
into a dwelling. In determining whether a product is a consumer product,
doubtful cases shall be resolved in favor of coverage. For a particular
product received by a particular user, "normally used" refers to a
typical or common use of that class of product, regardless of the status
of the particular user or of the way in which the particular user
actually uses, or expects or is expected to use, the product. A product
is a consumer product regardless of whether the product has substantial
commercial, industrial or non-consumer uses, unless such uses represent
the only significant mode of use of the product.
"Installation Information" for a User Product means any methods,
procedures, authorization keys, or other information required to install
and execute modified versions of a covered work in that User Product from
a modified version of its Corresponding Source. The information must
suffice to ensure that the continued functioning of the modified object
code is in no case prevented or interfered with solely because
modification has been made.
If you convey an object code work under this section in, or with, or
specifically for use in, a User Product, and the conveying occurs as
part of a transaction in which the right of possession and use of the
User Product is transferred to the recipient in perpetuity or for a
fixed term (regardless of how the transaction is characterized), the
Corresponding Source conveyed under this section must be accompanied
by the Installation Information. But this requirement does not apply
if neither you nor any third party retains the ability to install
modified object code on the User Product (for example, the work has
been installed in ROM).
The requirement to provide Installation Information does not include a
requirement to continue to provide support service, warranty, or updates
for a work that has been modified or installed by the recipient, or for
the User Product in which it has been modified or installed. Access to a
network may be denied when the modification itself materially and
adversely affects the operation of the network or violates the rules and
protocols for communication across the network.
Corresponding Source conveyed, and Installation Information provided,
in accord with this section must be in a format that is publicly
documented (and with an implementation available to the public in
source code form), and must require no special password or key for
unpacking, reading or copying.
7. Additional Terms.
"Additional permissions" are terms that supplement the terms of this
License by making exceptions from one or more of its conditions.
Additional permissions that are applicable to the entire Program shall
be treated as though they were included in this License, to the extent
that they are valid under applicable law. If additional permissions
apply only to part of the Program, that part may be used separately
under those permissions, but the entire Program remains governed by
this License without regard to the additional permissions.
When you convey a copy of a covered work, you may at your option
remove any additional permissions from that copy, or from any part of
it. (Additional permissions may be written to require their own
removal in certain cases when you modify the work.) You may place
additional permissions on material, added by you to a covered work,
for which you have or can give appropriate copyright permission.
Notwithstanding any other provision of this License, for material you
add to a covered work, you may (if authorized by the copyright holders of
that material) supplement the terms of this License with terms:
a) Disclaiming warranty or limiting liability differently from the
terms of sections 15 and 16 of this License; or
b) Requiring preservation of specified reasonable legal notices or
author attributions in that material or in the Appropriate Legal
Notices displayed by works containing it; or
c) Prohibiting misrepresentation of the origin of that material, or
requiring that modified versions of such material be marked in
reasonable ways as different from the original version; or
d) Limiting the use for publicity purposes of names of licensors or
authors of the material; or
e) Declining to grant rights under trademark law for use of some
trade names, trademarks, or service marks; or
f) Requiring indemnification of licensors and authors of that
material by anyone who conveys the material (or modified versions of
it) with contractual assumptions of liability to the recipient, for
any liability that these contractual assumptions directly impose on
those licensors and authors.
All other non-permissive additional terms are considered "further
restrictions" within the meaning of section 10. If the Program as you
received it, or any part of it, contains a notice stating that it is
governed by this License along with a term that is a further
restriction, you may remove that term. If a license document contains
a further restriction but permits relicensing or conveying under this
License, you may add to a covered work material governed by the terms
of that license document, provided that the further restriction does
not survive such relicensing or conveying.
If you add terms to a covered work in accord with this section, you
must place, in the relevant source files, a statement of the
additional terms that apply to those files, or a notice indicating
where to find the applicable terms.
Additional terms, permissive or non-permissive, may be stated in the
form of a separately written license, or stated as exceptions;
the above requirements apply either way.
8. Termination.
You may not propagate or modify a covered work except as expressly
provided under this License. Any attempt otherwise to propagate or
modify it is void, and will automatically terminate your rights under
this License (including any patent licenses granted under the third
paragraph of section 11).
However, if you cease all violation of this License, then your
license from a particular copyright holder is reinstated (a)
provisionally, unless and until the copyright holder explicitly and
finally terminates your license, and (b) permanently, if the copyright
holder fails to notify you of the violation by some reasonable means
prior to 60 days after the cessation.
Moreover, your license from a particular copyright holder is
reinstated permanently if the copyright holder notifies you of the
violation by some reasonable means, this is the first time you have
received notice of violation of this License (for any work) from that
copyright holder, and you cure the violation prior to 30 days after
your receipt of the notice.
Termination of your rights under this section does not terminate the
licenses of parties who have received copies or rights from you under
this License. If your rights have been terminated and not permanently
reinstated, you do not qualify to receive new licenses for the same
material under section 10.
9. Acceptance Not Required for Having Copies.
You are not required to accept this License in order to receive or
run a copy of the Program. Ancillary propagation of a covered work
occurring solely as a consequence of using peer-to-peer transmission
to receive a copy likewise does not require acceptance. However,
nothing other than this License grants you permission to propagate or
modify any covered work. These actions infringe copyright if you do
not accept this License. Therefore, by modifying or propagating a
covered work, you indicate your acceptance of this License to do so.
10. Automatic Licensing of Downstream Recipients.
Each time you convey a covered work, the recipient automatically
receives a license from the original licensors, to run, modify and
propagate that work, subject to this License. You are not responsible
for enforcing compliance by third parties with this License.
An "entity transaction" is a transaction transferring control of an
organization, or substantially all assets of one, or subdividing an
organization, or merging organizations. If propagation of a covered
work results from an entity transaction, each party to that
transaction who receives a copy of the work also receives whatever
licenses to the work the party's predecessor in interest had or could
give under the previous paragraph, plus a right to possession of the
Corresponding Source of the work from the predecessor in interest, if
the predecessor has it or can get it with reasonable efforts.
You may not impose any further restrictions on the exercise of the
rights granted or affirmed under this License. For example, you may
not impose a license fee, royalty, or other charge for exercise of
rights granted under this License, and you may not initiate litigation
(including a cross-claim or counterclaim in a lawsuit) alleging that
any patent claim is infringed by making, using, selling, offering for
sale, or importing the Program or any portion of it.
11. Patents.
A "contributor" is a copyright holder who authorizes use under this
License of the Program or a work on which the Program is based. The
work thus licensed is called the contributor's "contributor version".
A contributor's "essential patent claims" are all patent claims
owned or controlled by the contributor, whether already acquired or
hereafter acquired, that would be infringed by some manner, permitted
by this License, of making, using, or selling its contributor version,
but do not include claims that would be infringed only as a
consequence of further modification of the contributor version. For
purposes of this definition, "control" includes the right to grant
patent sublicenses in a manner consistent with the requirements of
this License.
Each contributor grants you a non-exclusive, worldwide, royalty-free
patent license under the contributor's essential patent claims, to
make, use, sell, offer for sale, import and otherwise run, modify and
propagate the contents of its contributor version.
In the following three paragraphs, a "patent license" is any express
agreement or commitment, however denominated, not to enforce a patent
(such as an express permission to practice a patent or covenant not to
sue for patent infringement). To "grant" such a patent license to a
party means to make such an agreement or commitment not to enforce a
patent against the party.
If you convey a covered work, knowingly relying on a patent license,
and the Corresponding Source of the work is not available for anyone
to copy, free of charge and under the terms of this License, through a
publicly available network server or other readily accessible means,
then you must either (1) cause the Corresponding Source to be so
available, or (2) arrange to deprive yourself of the benefit of the
patent license for this particular work, or (3) arrange, in a manner
consistent with the requirements of this License, to extend the patent
license to downstream recipients. "Knowingly relying" means you have
actual knowledge that, but for the patent license, your conveying the
covered work in a country, or your recipient's use of the covered work
in a country, would infringe one or more identifiable patents in that
country that you have reason to believe are valid.
If, pursuant to or in connection with a single transaction or
arrangement, you convey, or propagate by procuring conveyance of, a
covered work, and grant a patent license to some of the parties
receiving the covered work authorizing them to use, propagate, modify
or convey a specific copy of the covered work, then the patent license
you grant is automatically extended to all recipients of the covered
work and works based on it.
A patent license is "discriminatory" if it does not include within
the scope of its coverage, prohibits the exercise of, or is
conditioned on the non-exercise of one or more of the rights that are
specifically granted under this License. You may not convey a covered
work if you are a party to an arrangement with a third party that is
in the business of distributing software, under which you make payment
to the third party based on the extent of your activity of conveying
the work, and under which the third party grants, to any of the
parties who would receive the covered work from you, a discriminatory
patent license (a) in connection with copies of the covered work
conveyed by you (or copies made from those copies), or (b) primarily
for and in connection with specific products or compilations that
contain the covered work, unless you entered into that arrangement,
or that patent license was granted, prior to 28 March 2007.
Nothing in this License shall be construed as excluding or limiting
any implied license or other defenses to infringement that may
otherwise be available to you under applicable patent law.
12. No Surrender of Others' Freedom.
If conditions are imposed on you (whether by court order, agreement or
otherwise) that contradict the conditions of this License, they do not
excuse you from the conditions of this License. If you cannot convey a
covered work so as to satisfy simultaneously your obligations under this
License and any other pertinent obligations, then as a consequence you may
not convey it at all. For example, if you agree to terms that obligate you
to collect a royalty for further conveying from those to whom you convey
the Program, the only way you could satisfy both those terms and this
License would be to refrain entirely from conveying the Program.
13. Remote Network Interaction; Use with the GNU General Public License.
Notwithstanding any other provision of this License, if you modify the
Program, your modified version must prominently offer all users
interacting with it remotely through a computer network (if your version
supports such interaction) an opportunity to receive the Corresponding
Source of your version by providing access to the Corresponding Source
from a network server at no charge, through some standard or customary
means of facilitating copying of software. This Corresponding Source
shall include the Corresponding Source for any work covered by version 3
of the GNU General Public License that is incorporated pursuant to the
following paragraph.
Notwithstanding any other provision of this License, you have
permission to link or combine any covered work with a work licensed
under version 3 of the GNU General Public License into a single
combined work, and to convey the resulting work. The terms of this
License will continue to apply to the part which is the covered work,
but the work with which it is combined will remain governed by version
3 of the GNU General Public License.
14. Revised Versions of this License.
The Free Software Foundation may publish revised and/or new versions of
the GNU Affero General Public License from time to time. Such new versions
will be similar in spirit to the present version, but may differ in detail to
address new problems or concerns.
Each version is given a distinguishing version number. If the
Program specifies that a certain numbered version of the GNU Affero General
Public License "or any later version" applies to it, you have the
option of following the terms and conditions either of that numbered
version or of any later version published by the Free Software
Foundation. If the Program does not specify a version number of the
GNU Affero General Public License, you may choose any version ever published
by the Free Software Foundation.
If the Program specifies that a proxy can decide which future
versions of the GNU Affero General Public License can be used, that proxy's
public statement of acceptance of a version permanently authorizes you
to choose that version for the Program.
Later license versions may give you additional or different
permissions. However, no additional obligations are imposed on any
author or copyright holder as a result of your choosing to follow a
later version.
15. Disclaimer of Warranty.
THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
16. Limitation of Liability.
IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
SUCH DAMAGES.
17. Interpretation of Sections 15 and 16.
If the disclaimer of warranty and limitation of liability provided
above cannot be given local legal effect according to their terms,
reviewing courts shall apply local law that most closely approximates
an absolute waiver of all civil liability in connection with the
Program, unless a warranty or assumption of liability accompanies a
copy of the Program in return for a fee.
END OF TERMS AND CONDITIONS
How to Apply These Terms to Your New Programs
If you develop a new program, and you want it to be of the greatest
possible use to the public, the best way to achieve this is to make it
free software which everyone can redistribute and change under these terms.
To do so, attach the following notices to the program. It is safest
to attach them to the start of each source file to most effectively
state the exclusion of warranty; and each file should have at least
the "copyright" line and a pointer to where the full notice is found.
<one line to give the program's name and a brief idea of what it does.>
Copyright (C) <year> <name of author>
This program is free software: you can redistribute it and/or modify
it under the terms of the GNU Affero General Public License as published
by the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU Affero General Public License for more details.
You should have received a copy of the GNU Affero General Public License
along with this program. If not, see <https://www.gnu.org/licenses/>.
Also add information on how to contact you by electronic and paper mail. Also add information on how to contact you by electronic and paper mail.
If the program does terminal interaction, make it output a short notice like this when it starts in an interactive mode: If your software can interact with users remotely through a computer
network, you should also make sure that it provides a way for users to
get its source. For example, if your program is a web application, its
interface could display a "Source" link that leads users to an archive
of the code. There are many ways you could offer source, and different
solutions will be better for different programs; see section 13 for the
specific requirements.
bpe_framework Copyright (C) 2025 interval You should also get your employer (if you work as a programmer) or school,
This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. if any, to sign a "copyright disclaimer" for the program, if necessary.
This is free software, and you are welcome to redistribute it under certain conditions; type `show c' for details. For more information on this, and how to apply and follow the GNU AGPL, see
<https://www.gnu.org/licenses/>.
The hypothetical commands `show w' and `show c' should show the appropriate parts of the General Public License. Of course, your program's commands might be different; for a GUI interface, you would use an “about box”.
You should also get your employer (if you work as a programmer) or school, if any, to sign a “copyright disclaimer” for the program, if necessary. For more information on this, and how to apply and follow the GNU GPL, see <https://www.gnu.org/licenses/>.
The GNU General Public License does not permit incorporating your program into proprietary programs. If your program is a subroutine library, you may consider it more useful to permit linking proprietary applications with the library. If this is what you want to do, use the GNU Lesser General Public License instead of this License. But first, please read <https://www.gnu.org/philosophy/why-not-lgpl.html>.

164
README.md
View File

@ -1,3 +1,165 @@
# bpe_framework # bpe_framework
Large Language Model for Agentic AI
Build: cmake -DCMAKE_POLICY_VERSION_MINIMUM=3.5 ..
#### The test_bpe application does the following:
1. Includes necessary headers and defines the main function.
2. Creates an instance of the BPETokenizer.
3. Defines a training corpus (a vector of strings).
4. Trains the tokenizer on the corpus with a specified vocabulary size (500 in this case).
5. Tests the tokenizer by encoding a sample string ("the quick brown fox").
6. Decodes the tokens back to a string and prints the original, tokens, and decoded string.
7. Saves the tokenizer to a file ("bpe_model.txt").
8. Loads the tokenizer from the file and verifies the loaded tokenizer's vocabulary size.
The purpose of this test is to verify that the BPE tokenizer can be trained, encode, decode, and serialize/deserialize correctly.
Let's break down the code step by step.
test_bpe Application Overview
#### The test_bpe application is a comprehensive test program that validates the functionality of the BPE tokenizer implementation in the LM Framework. Here's how it works:
1. Initialization
Creates an instance of BPETokenizer
Defines a training corpus with sample English text
2. Training Process
Calls tokenizer.train(corpus, 500) to train the tokenizer
The training process:
Initializes with byte-level vocabulary (0-255)
Analyzes word frequencies in the corpus
Iteratively merges the most frequent character pairs
Builds a vocabulary of 500 tokens (as specified)
3. Encoding Test
Encodes the test string "the quick brown fox"
The encoding process:
Splits text into words
Converts each character to its initial token ID
Applies learned BPE merges to combine tokens
Returns a sequence of integer token IDs
4. Decoding Test
Decodes the token IDs back to text
The decoding process:
Converts each token ID back to its string representation
Concatenates the strings to reconstruct the original text
5. Serialization Test
Saves the trained tokenizer to "bpe_model.txt"
The serialization process:
Writes vocabulary size and token-ID mappings
Records all learned merge rules
6. Deserialization Test
Loads the tokenizer from "bpe_model.txt"
Verifies the loaded tokenizer has the same vocabulary size
Confirms the tokenizer can perform encoding/decoding
Expected Output
text
Training tokenizer...
Vocabulary size: 500
Original: the quick brown fox
Tokens: [list of token IDs]
Decoded: the quick brown fox
Successfully loaded tokenizer
Loaded vocabulary size: 500
Key Validations
Training Completes without errors
Encoding/Decoding Round-Trip preserves the original text
Serialization/Deserialization maintains tokenizer state
Vocabulary Size matches the specified target (500)
Token IDs are consistent between sessions
# BPE Tokenizer Performance Test Suite
## Overview
This performance test application is a comprehensive benchmarking tool designed to evaluate the efficiency and scalability of the Byte Pair Encoding (BPE) tokenizer implementation. The test suite measures critical performance metrics including training time, memory usage, encoding/decoding speed, and serialization performance across various configurations.
## Key Features
### 1. Corpus Generation
- Automatically generates realistic test corpora using common AI/ML terminology
- Configurable sentence count and word range parameters
- Creates diverse text samples that mimic real-world language patterns
### 2. Multi-Dimensional Testing
- Tests multiple corpus sizes (100, 1000, 5000 sentences)
- Evaluates different vocabulary sizes (500, 1000, 2000 tokens)
- Measures performance across various workload scenarios
### 3. Comprehensive Performance Metrics
- **Training Time**: Measures how long it takes to build the BPE vocabulary from a corpus
- **Memory Usage**: Tracks peak memory consumption during training (Linux-specific)
- **Encoding Speed**: Calculates processing time per token during text encoding
- **Round-Trip Verification**: Ensures encoding/decoding preserves original content
- **Serialization Performance**: Measures model save/load operations
### 4. Validation Checks
- Verifies encoding/decoding consistency
- Detects potential data corruption issues
- Validates vocabulary construction
## Usage Scenarios
This performance test is ideal for:
- Benchmarking different BPE implementations
- Evaluating hardware suitability for language processing tasks
- Identifying performance bottlenecks in tokenization pipelines
- Testing scalability of tokenizer implementations
- Comparing optimization techniques
## Technical Implementation
The test suite utilizes:
- High-resolution timing with `<chrono>` for precise measurements
- Linux-specific memory tracking via `/proc/self/status`
- Randomized corpus generation with configurable parameters
- Exception handling for robust testing
- Automatic cleanup of temporary files
## Output Metrics
The application provides detailed performance reports including:
- Training duration in milliseconds
- Peak memory usage in megabytes
- Encoding speed in microseconds per token
- Serialization/deserialization times
- Vocabulary size validation
- Round-trip integrity verification
This test framework serves as an essential tool for developers and researchers working with BPE tokenizers, providing quantitative data to guide optimization efforts and implementation choices.
Large Language Model for Agentic AI

0
configs/alpha_config.json Normal file
View File

54
include/lm/runtime/init.hpp Executable file
View File

@ -0,0 +1,54 @@
// Runtime Initialization Header File
//Here's the complete `include/lm/runtime/init.hpp` file:
//```cpp
#pragma once
#include <string>
#include <nlohmann/json.hpp>
#include <filesystem>
namespace lm::runtime {
class SystemState {
public:
// Singleton access
static SystemState& get_instance();
// Initialize from JSON config
void initialize(const std::filesystem::path& config_path);
// Configuration accessors
const nlohmann::json& config() const noexcept;
std::string get_string(const std::string& key) const;
int get_int(const std::string& key, int default_val = 0) const;
// Subsystem states
bool is_tokenizer_ready() const noexcept;
bool is_model_loaded() const noexcept;
private:
SystemState() = default; // Private constructor
nlohmann::json config_;
bool tokenizer_ready_ = false;
bool model_loaded_ = false;
};
} // namespace lm::runtime
/*```
This header provides the interface for the framework initialization system with:
1. **Singleton pattern** for global system state access
2. **JSON configuration** loading and access methods
3. **Subsystem state tracking** for tokenizer and model
4. **Type-safe configuration access** with default values
The implementation (in the corresponding `.cpp` file) handles:
- JSON configuration parsing and validation
- Subsystem initialization sequencing
- Error handling for malformed configurations
- State management across the framework
This initialization system provides a centralized way to configure and manage the LM framework components.*/

22
include/lm/runtime/shutdown.hpp Executable file
View File

@ -0,0 +1,22 @@
#pragma once
#include <nlohmann/json.hpp>
#include <filesystem>
#include <chrono>
namespace lm::runtime {
class ShutdownHandler {
public:
// Serialize state to JSON
static void save_state(
const std::filesystem::path& output_path,
bool include_model_weights = false
);
// Cleanup hooks
static void register_cleanup(void (*func)());
static void execute_cleanup();
};
} // namespace lm::runtime

52
include/lm/tokenizer/bpe_tokenizer.hpp Executable file
View File

@ -0,0 +1,52 @@
#pragma once
#include <string>
#include <vector>
#include <unordered_map>
#include <map>
#include <memory>
#include <utility>
#include <cstdint> // For uint16_t
#include <queue>
#include <functional>
namespace lm {
using TokenID = uint16_t; // Support for 65k vocabulary
class BPETokenizer {
public:
BPETokenizer();
~BPETokenizer();
// Training methods
void train(const std::vector<std::string>& corpus, size_t vocab_size = 30000);
void train_from_file(const std::string& filename, size_t vocab_size = 30000);
// Tokenization methods
std::vector<TokenID> encode(const std::string& text) const;
std::string decode(const std::vector<TokenID>& tokens) const;
// Serialization
bool save(const std::string& filename) const;
bool load(const std::string& filename);
// Vocabulary access
size_t vocab_size() const;
std::string id_to_token(TokenID id) const;
TokenID token_to_id(const std::string& token) const;
// Configuration
void set_unknown_token(const std::string& token);
void add_special_token(const std::string& token);
// Unicode-specific methods
void set_normalization(bool enabled);
void set_byte_fallback(bool enabled);
private:
struct Impl;
std::unique_ptr<Impl> pimpl_;
};
} // namespace lm

59
include/lm/tokenizer/unicode_utils.hpp Executable file
View File

@ -0,0 +1,59 @@
//# Unicode Utilities Header File
#pragma once
#include <string>
#include <vector>
#include <cstdint>
namespace lm::unicode {
// Unicode character representation
struct CodePoint {
uint32_t value;
std::string utf8; // UTF-8 representation
};
// Check if a code point is whitespace
bool is_whitespace(uint32_t codepoint);
// Check if a code point is punctuation
bool is_punctuation(uint32_t codepoint);
// Check if a code point is a control character
bool is_control(uint32_t codepoint);
// Normalize Unicode text (NFC normalization)
std::string normalize(const std::string& text);
// Split text into Unicode code points
std::vector<CodePoint> to_code_points(const std::string& text);
// Convert code points back to UTF-8 string
std::string from_code_points(const std::vector<CodePoint>& code_points);
// Unicode-aware string split (handles Unicode whitespace)
std::vector<std::string> unicode_split(const std::string& text);
// Unicode-aware character boundaries
std::vector<std::string> split_on_character_boundaries(const std::string& text);
} // namespace lm::unicode
/*```
This header provides a comprehensive set of Unicode utilities for the BPE tokenizer, including:
1. **Code point representation** with both numeric value and UTF-8 string
2. **Character classification** functions (whitespace, punctuation, control)
3. **Text normalization** using Unicode NFC form
4. **UTF-8 encoding/decoding** utilities
5. **Unicode-aware text splitting** for proper tokenization
These utilities enable the tokenizer to properly handle:
- Multi-byte UTF-8 sequences
- Unicode whitespace characters
- Text normalization for consistent processing
- Proper character boundary detection
- Support for various writing systems (Latin, CJK, Arabic, etc.)
The implementation in the corresponding `.cpp` file uses the ICU library for robust Unicode handling.*/

48
src/CMakeLists.txt Normal file
View File

@ -0,0 +1,48 @@
# Tokenizer library
add_library(lm_tokenizer
bpe_tokenizer.cpp
unicode_utils.cpp
)
target_include_directories(lm_tokenizer
PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/../../include
)
target_link_libraries(lm_tokenizer
PRIVATE
ICU::uc
ICU::i18n
)
# CPU-specific optimization flags
if(CMAKE_COMPILER_IS_GNUCXX OR CMAKE_CXX_COMPILER_ID MATCHES "Clang")
# Enable aggressive optimizations
target_compile_options(lm_tokenizer PRIVATE -O3 -march=native)
# Enable SSE4.2 instructions if available
target_compile_options(lm_tokenizer PRIVATE -msse4.2)
# Enable link-time optimization
if(CMAKE_VERSION VERSION_GREATER_EQUAL 3.9)
target_compile_options(lm_tokenizer PRIVATE -flto)
set_target_properties(lm_tokenizer PROPERTIES INTERPROCEDURAL_OPTIMIZATION TRUE)
endif()
# Enable specific optimizations for GCC
if(CMAKE_COMPILER_IS_GNUCXX)
target_compile_options(lm_tokenizer PRIVATE -ftree-vectorize -funroll-loops)
endif()
# Enable specific optimizations for Clang
if(CMAKE_CXX_COMPILER_ID MATCHES "Clang")
target_compile_options(lm_tokenizer PRIVATE -Rpass=.* -Rpass-missed=.* -Rpass-analysis=.*)
endif()
endif()
# Add profiling support
if(PROFILE)
if(CMAKE_COMPILER_IS_GNUCXX)
target_compile_options(lm_tokenizer PRIVATE -pg)
target_link_options(lm_tokenizer PRIVATE -pg)
endif()
endif()

32
src/alpha/CMakeLists.txt Normal file
View File

@ -0,0 +1,32 @@
cmake_minimum_required(VERSION 3.6)
project(lm_framework)
# Alpha prototype executable
add_executable(lm_alpha
repl.cpp
config_io.cpp
)
target_include_directories(lm_alpha
PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/../../include
)
target_link_libraries(lm_alpha
PRIVATE
lm_core
nlohmann_json::nlohmann_json
)
# Copy config file to build directory
configure_file(
${CMAKE_CURRENT_SOURCE_DIR}/../../configs/alpha_config.json
${CMAKE_CURRENT_BINARY_DIR}/alpha_config.json
COPYONLY
)
# Create run script for easy execution
file(GENERATE OUTPUT run_alpha.sh
CONTENT "#!/bin/bash\ncd ${CMAKE_CURRENT_BINARY_DIR} && ./lm_alpha alpha_config.json"
FILE_PERMISSIONS OWNER_READ OWNER_WRITE OWNER_EXECUTE
)

49
src/alpha/config_io.cpp Normal file
View File

@ -0,0 +1,49 @@
#include "lm/runtime/init.hpp"
#include <nlohmann/json.hpp>
#include <fstream>
#include <stdexcept>
nlohmann::json load_config(const std::string& path) {
try {
std::ifstream file(path);
if (!file.is_open()) {
throw std::runtime_error("Cannot open config file: " + path);
}
nlohmann::json config;
file >> config;
return config;
} catch (const std::exception& e) {
// Fallback to default config if file doesn't exist or is invalid
return nlohmann::json{
{"alpha", {
{"prompt", "> "},
{"save_on_exit", true}
}},
{"tokenizer", {
{"type", "bpe"},
{"vocab_size", 100},
{"dummy_data", true}
}},
{"model", {
{"layers", 2},
{"dim", 64}
}}
};
}
}
void save_config(const nlohmann::json& config, const std::string& path) {
try {
std::ofstream file(path);
if (!file.is_open()) {
throw std::runtime_error("Cannot open file for writing: " + path);
}
file << config.dump(2); // Pretty print with 2-space indentation
} catch (const std::exception& e) {
throw std::runtime_error("Failed to save config: " + std::string(e.what()));
}
}

44
src/alpha/repl.cpp Normal file
View File

@ -0,0 +1,44 @@
#include <iostream>
#include <string>
#include "lm/tokenizer/bpe_tokenizer.hpp"
void run_repl() {
lm::BPETokenizer tokenizer;
// Simple training for the alpha
std::vector<std::string> corpus = {
"hello world", "test input", "simple example"
};
tokenizer.train(corpus, 100);
std::cout << "LM Framework Alpha\n> ";
std::string input;
while (std::getline(std::cin, input)) {
if (input == "/exit") break;
try {
auto tokens = tokenizer.encode(input);
std::cout << "Tokens: ";
for (auto token : tokens) {
std::cout << token << " ";
}
std::cout << "\n> ";
} catch (const std::exception& e) {
std::cout << "Error: " << e.what() << "\n> ";
}
}
std::cout << "Saving session...\n";
tokenizer.save("alpha_session.bpe");
}
int main() {
try {
run_repl();
} catch (const std::exception& e) {
std::cerr << "Error: " << e.what() << "\n";
return 1;
}
return 0;
}

169
src/performance_test.cpp Normal file
View File

@ -0,0 +1,169 @@
#include "lm/tokenizer/bpe_tokenizer.hpp"
#include <iostream>
#include <vector>
#include <chrono>
#include <fstream>
#include <random>
#include <algorithm>
#include <sstream> // Add this include for std::istringstream
// Generate random text for testing
std::vector<std::string> generate_test_corpus(size_t num_sentences, size_t min_words, size_t max_words) {
std::vector<std::string> common_words = {
"the", "quick", "brown", "fox", "jumps", "over", "lazy", "dog",
"artificial", "intelligence", "machine", "learning", "deep", "neural", "network",
"language", "model", "transformer", "attention", "mechanism", "tokenization",
"byte", "pair", "encoding", "subword", "vocabulary", "training", "inference"
};
std::vector<std::string> corpus;
std::random_device rd;
std::mt19937 gen(rd());
std::uniform_int_distribution<> word_count_dist(min_words, max_words);
std::uniform_int_distribution<> word_index_dist(0, common_words.size() - 1);
for (size_t i = 0; i < num_sentences; ++i) {
int word_count = word_count_dist(gen);
std::string sentence;
for (int j = 0; j < word_count; ++j) {
if (!sentence.empty()) {
sentence += " ";
}
sentence += common_words[word_index_dist(gen)];
}
corpus.push_back(sentence);
}
return corpus;
}
// Measure memory usage (Linux specific)
size_t get_peak_memory_usage() {
#ifdef __linux__
std::ifstream status("/proc/self/status");
std::string line;
while (std::getline(status, line)) {
if (line.compare(0, 6, "VmPeak") == 0) {
std::istringstream iss(line);
std::string key;
size_t value;
std::string unit;
iss >> key >> value >> unit;
if (unit == "kB") {
return value * 1024; // Convert to bytes
}
}
}
#endif
return 0;
}
void run_performance_test() {
std::cout << "=== BPE Tokenizer Performance Test ===\n";
// Test different corpus sizes
std::vector<size_t> corpus_sizes = {100, 1000, 5000};
std::vector<size_t> vocab_sizes = {500, 1000, 2000};
for (size_t corpus_size : corpus_sizes) {
for (size_t vocab_size : vocab_sizes) {
std::cout << "\n--- Test Configuration: " << corpus_size
<< " sentences, " << vocab_size << " vocabulary ---\n";
// Generate test corpus
auto corpus = generate_test_corpus(corpus_size, 5, 15);
// Measure training performance
auto start_time = std::chrono::high_resolution_clock::now();
size_t start_memory = get_peak_memory_usage();
lm::BPETokenizer tokenizer;
try {
tokenizer.train(corpus, vocab_size);
auto end_time = std::chrono::high_resolution_clock::now();
size_t end_memory = get_peak_memory_usage();
auto duration = std::chrono::duration_cast<std::chrono::milliseconds>(
end_time - start_time);
size_t memory_used = (end_memory - start_memory) / (1024 * 1024);
std::cout << "Training time: " << duration.count() << " ms\n";
std::cout << "Peak memory used: " << memory_used << " MB\n";
std::cout << "Final vocabulary size: " << tokenizer.vocab_size() << "\n";
// Measure encoding performance
std::vector<std::string> test_texts = {
"the quick brown fox jumps over the lazy dog",
"artificial intelligence and machine learning",
"transformer language model with attention mechanism"
};
auto encode_start = std::chrono::high_resolution_clock::now();
size_t total_tokens = 0;
for (const auto& text : test_texts) {
auto tokens = tokenizer.encode(text);
total_tokens += tokens.size();
// Verify round-trip
std::string decoded = tokenizer.decode(tokens);
if (text != decoded) {
std::cout << "WARNING: Round-trip mismatch!\n";
std::cout << "Original: " << text << "\n";
std::cout << "Decoded: " << decoded << "\n";
}
}
auto encode_end = std::chrono::high_resolution_clock::now();
auto encode_duration = std::chrono::duration_cast<std::chrono::microseconds>(
encode_end - encode_start);
double encode_time_per_token = static_cast<double>(encode_duration.count()) / total_tokens;
std::cout << "Encoding performance: " << encode_time_per_token << " μs/token\n";
std::cout << "Total tokens processed: " << total_tokens << "\n";
} catch (const std::exception& e) {
std::cout << "Error during training: " << e.what() << "\n";
}
}
}
// Test serialization performance
std::cout << "\n--- Serialization Performance Test ---\n";
auto corpus = generate_test_corpus(1000, 5, 15);
lm::BPETokenizer tokenizer;
tokenizer.train(corpus, 1000);
auto start_time = std::chrono::high_resolution_clock::now();
tokenizer.save("test_model.bpe");
auto save_time = std::chrono::duration_cast<std::chrono::microseconds>(
std::chrono::high_resolution_clock::now() - start_time);
start_time = std::chrono::high_resolution_clock::now();
lm::BPETokenizer loaded_tokenizer;
loaded_tokenizer.load("test_model.bpe");
auto load_time = std::chrono::duration_cast<std::chrono::microseconds>(
std::chrono::high_resolution_clock::now() - start_time);
std::cout << "Model save time: " << save_time.count() << " μs\n";
std::cout << "Model load time: " << load_time.count() << " μs\n";
// Clean up
remove("test_model.bpe");
}
int main() {
try {
run_performance_test();
std::cout << "\n=== Performance Test Completed ===\n";
} catch (const std::exception& e) {
std::cerr << "Performance test failed: " << e.what() << "\n";
return 1;
}
return 0;
}

18
src/runtime/CMakeLists.txt Executable file
View File

@ -0,0 +1,18 @@
project(lm_framework)
cmake_minimum_required(VERSION 3.6)
add_library(lm_runtime
init.cpp
shutdown.cpp
state_utils.cpp # Add this line
)
target_include_directories(lm_runtime
PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/../../include
)
target_link_libraries(lm_runtime
PRIVATE nlohmann_json::nlohmann_json
)

123
src/runtime/init.cpp Executable file
View File

@ -0,0 +1,123 @@
/*# Runtime Initialization Implementation File
Here's the complete `src/runtime/init.cpp` file:
```cpp*/
#include "lm/runtime/init.hpp"
#include <fstream>
#include <stdexcept>
namespace lm::runtime {
namespace {
// Private implementation details
SystemState* g_instance = nullptr;
bool initialize_tokenizer(const nlohmann::json& config) {
// TODO: Implement actual tokenizer initialization
// For now, just check if tokenizer config exists
return config.contains("tokenizer");
}
bool initialize_model(const nlohmann::json& config) {
// TODO: Implement actual model initialization
// For now, just check if model config exists
return config.contains("model");
}
} // anonymous namespace
SystemState& SystemState::get_instance() {
if (!g_instance) {
g_instance = new SystemState();
}
return *g_instance;
}
void SystemState::initialize(const std::filesystem::path& config_path) {
try {
// Load JSON config
std::ifstream f(config_path);
if (!f.is_open()) {
throw std::runtime_error("Cannot open config file: " + config_path.string());
}
config_ = nlohmann::json::parse(f);
// Validate required fields
if (!config_.contains("tokenizer") || !config_.contains("model")) {
throw std::runtime_error("Invalid config: missing required sections");
}
// Initialize subsystems
tokenizer_ready_ = initialize_tokenizer(config_["tokenizer"]);
model_loaded_ = initialize_model(config_["model"]);
if (!tokenizer_ready_) {
throw std::runtime_error("Tokenizer initialization failed");
}
if (!model_loaded_) {
throw std::runtime_error("Model initialization failed");
}
} catch (const std::exception& e) {
throw std::runtime_error("Initialization failed: " + std::string(e.what()));
}
}
const nlohmann::json& SystemState::config() const noexcept {
return config_;
}
std::string SystemState::get_string(const std::string& key) const {
if (!config_.contains(key)) {
throw std::runtime_error("Config key not found: " + key);
}
if (!config_[key].is_string()) {
throw std::runtime_error("Config value is not a string: " + key);
}
return config_[key].get<std::string>();
}
int SystemState::get_int(const std::string& key, int default_val) const {
if (!config_.contains(key)) {
return default_val;
}
if (!config_[key].is_number()) {
throw std::runtime_error("Config value is not a number: " + key);
}
return config_[key].get<int>();
}
bool SystemState::is_tokenizer_ready() const noexcept {
return tokenizer_ready_;
}
bool SystemState::is_model_loaded() const noexcept {
return model_loaded_;
}
} // namespace lm::runtime
/*```
This implementation provides:
1. **Singleton pattern** with thread-safe initialization
2. **JSON configuration loading** with error handling
3. **Subsystem initialization** stubs for tokenizer and model
4. **Type-safe configuration access** with proper error reporting
5. **State tracking** for framework components
Key features:
- **Robust error handling** with descriptive error messages
- **Config validation** to ensure required sections are present
- **Graceful fallbacks** for optional configuration values
- **Exception safety** with proper resource cleanup
The implementation follows the RAII pattern and provides a solid foundation for the framework's initialization system. The tokenizer and model initialization functions are currently stubbed but can be expanded with actual implementation as the framework develops.*/

159
src/runtime/shutdown.cpp Normal file
View File

@ -0,0 +1,159 @@
#include "lm/runtime/shutdown.hpp"
#include "lm/runtime/init.hpp"
#include "lm/tokenizer/bpe_tokenizer.hpp"
#include <fstream>
#include <vector>
#include <mutex>
#include <sstream>
#include <iostream>
namespace lm::runtime {
namespace {
std::vector<void (*)()> cleanup_functions;
std::mutex cleanup_mutex;
}
// Serialize tokenizer state to JSON
nlohmann::json serialize_tokenizer_state() {
auto& system_state = SystemState::get_instance();
nlohmann::json tokenizer_state;
// Get tokenizer configuration from system state
try {
const auto& config = system_state.config();
if (config.contains("tokenizer")) {
tokenizer_state = config["tokenizer"];
}
// Add runtime information
tokenizer_state["runtime"] = {
{"initialized", system_state.is_tokenizer_ready()},
{"timestamp", std::chrono::system_clock::now().time_since_epoch().count()}
};
} catch (const std::exception& e) {
tokenizer_state["error"] = std::string("Failed to serialize tokenizer state: ") + e.what();
}
return tokenizer_state;
}
// Serialize model state to JSON
nlohmann::json serialize_model_state(bool include_weights) {
auto& system_state = SystemState::get_instance();
nlohmann::json model_state;
try {
const auto& config = system_state.config();
if (config.contains("model")) {
model_state = config["model"];
}
// Add runtime information
model_state["runtime"] = {
{"loaded", system_state.is_model_loaded()},
{"timestamp", std::chrono::system_clock::now().time_since_epoch().count()}
};
if (include_weights) {
// Placeholder for actual weight serialization
model_state["weights"] = {
{"serialized", false},
{"message", "Weight serialization not yet implemented"}
};
}
} catch (const std::exception& e) {
model_state["error"] = std::string("Failed to serialize model state: ") + e.what();
}
return model_state;
}
// Serialize threading state to JSON
nlohmann::json serialize_thread_pool_stats() {
nlohmann::json threading_state;
try {
// Placeholder for actual thread pool statistics
// This would normally come from ThreadPool::get_stats()
threading_state = {
{"active_threads", 0},
{"queued_tasks", 0},
{"completed_tasks", 0},
{"thread_pool_initialized", false}
};
} catch (const std::exception& e) {
threading_state["error"] = std::string("Failed to serialize threading state: ") + e.what();
}
return threading_state;
}
void ShutdownHandler::save_state(
const std::filesystem::path& output_path,
bool include_model_weights)
{
try {
nlohmann::json state;
// Capture framework state
auto& system_state = SystemState::get_instance();
// Add system configuration
state["config"] = system_state.config();
// Add component states
state["tokenizer"] = serialize_tokenizer_state();
state["model"] = serialize_model_state(include_model_weights);
state["threading"] = serialize_thread_pool_stats();
// Add shutdown metadata
state["metadata"] = {
{"shutdown_time", std::chrono::system_clock::now().time_since_epoch().count()},
{"include_weights", include_model_weights},
{"version", "0.1.0"},
{"format_version", 1}
};
// Write to file
std::ofstream file(output_path);
if (!file.is_open()) {
throw std::runtime_error("Cannot open file for writing: " + output_path.string());
}
file << state.dump(2); // Pretty print with 2-space indentation
file.close();
std::cout << "Framework state saved to: " << output_path << std::endl;
} catch (const std::exception& e) {
throw std::runtime_error("Failed to save state: " + std::string(e.what()));
}
}
void ShutdownHandler::register_cleanup(void (*func)()) {
std::lock_guard<std::mutex> lock(cleanup_mutex);
cleanup_functions.push_back(func);
}
void ShutdownHandler::execute_cleanup() {
std::lock_guard<std::mutex> lock(cleanup_mutex);
// Execute cleanup functions in reverse order (LIFO)
for (auto it = cleanup_functions.rbegin(); it != cleanup_functions.rend(); ++it) {
try {
(*it)();
} catch (const std::exception& e) {
// Log error but continue with other cleanup functions
std::cerr << "Cleanup function error: " << e.what() << std::endl;
}
}
cleanup_functions.clear();
}
} // namespace lm::runtime

81
src/runtime/state_utils.cpp Normal file
View File

@ -0,0 +1,81 @@
#include "lm/runtime/shutdown.hpp"
#include "lm/runtime/init.hpp"
#include <iomanip>
#include <ctime>
namespace lm::runtime {
// Helper function to format timestamp
std::string format_timestamp(int64_t timestamp_ns) {
std::time_t time = timestamp_ns / 1000000000;
std::tm* tm = std::localtime(&time);
if (tm) {
std::ostringstream oss;
oss << std::put_time(tm, "%Y-%m-%d %H:%M:%S");
return oss.str();
}
return "invalid_timestamp";
}
// Generate a comprehensive state report
std::string generate_state_report(const nlohmann::json& state) {
std::ostringstream report;
report << "=== LM Framework State Report ===\n\n";
// Basic information
if (state.contains("metadata")) {
const auto& metadata = state["metadata"];
report << "Shutdown Time: ";
if (metadata.contains("shutdown_time")) {
report << format_timestamp(metadata["shutdown_time"].get<int64_t>());
} else {
report << "unknown";
}
report << "\nVersion: " << metadata.value("version", "unknown") << "\n\n";
}
// Tokenizer state
if (state.contains("tokenizer")) {
const auto& tokenizer = state["tokenizer"];
report << "Tokenizer:\n";
report << " Initialized: " << tokenizer.value("runtime/initialized", false) << "\n";
if (tokenizer.contains("type")) {
report << " Type: " << tokenizer["type"] << "\n";
}
if (tokenizer.contains("vocab_size")) {
report << " Vocab Size: " << tokenizer["vocab_size"] << "\n";
}
report << "\n";
}
// Model state
if (state.contains("model")) {
const auto& model = state["model"];
report << "Model:\n";
report << " Loaded: " << model.value("runtime/loaded", false) << "\n";
if (model.contains("layers")) {
report << " Layers: " << model["layers"] << "\n";
}
if (model.contains("dim")) {
report << " Dimension: " << model["dim"] << "\n";
}
report << "\n";
}
// Threading state
if (state.contains("threading")) {
const auto& threading = state["threading"];
report << "Threading:\n";
report << " Active Threads: " << threading.value("active_threads", 0) << "\n";
report << " Queued Tasks: " << threading.value("queued_tasks", 0) << "\n";
report << "\n";
}
return report.str();
}
} // namespace lm::runtime

51
src/test_bpe.cpp Normal file
View File

@ -0,0 +1,51 @@
#include "lm/tokenizer/bpe_tokenizer.hpp"
#include <iostream>
#include <vector>
int main() {
lm::BPETokenizer tokenizer;
// Training corpus
std::vector<std::string> corpus = {
"the quick brown fox jumps over the lazy dog",
"artificial intelligence is transforming the world",
"C++ is a powerful programming language",
"machine learning models require large amounts of data"
};
try {
// Train the tokenizer
std::cout << "Training tokenizer..." << std::endl;
tokenizer.train(corpus, 500);
std::cout << "Vocabulary size: " << tokenizer.vocab_size() << std::endl;
// Test encoding/decoding
std::string test_text = "the quick brown fox";
auto tokens = tokenizer.encode(test_text);
std::string decoded = tokenizer.decode(tokens);
std::cout << "Original: " << test_text << std::endl;
std::cout << "Tokens: ";
for (auto token : tokens) {
std::cout << token << " ";
}
std::cout << std::endl;
std::cout << "Decoded: " << decoded << std::endl;
// Save and load test
tokenizer.save("bpe_model.txt");
lm::BPETokenizer loaded_tokenizer;
if (loaded_tokenizer.load("bpe_model.txt")) {
std::cout << "Successfully loaded tokenizer" << std::endl;
std::cout << "Loaded vocabulary size: " << loaded_tokenizer.vocab_size() << std::endl;
}
} catch (const std::exception& e) {
std::cerr << "Error: " << e.what() << std::endl;
return 1;
}
return 0;
}

77
src/test_unicode_bpe.cpp Normal file
View File

@ -0,0 +1,77 @@
#include "lm/tokenizer/bpe_tokenizer.hpp"
#include <iostream>
#include <vector>
int main() {
lm::BPETokenizer tokenizer;
// Training corpus with Unicode text
std::vector<std::string> corpus = {
"the quick brown fox jumps over the lazy dog",
"artificial intelligence is transforming the world",
"C++ is a powerful programming language",
"machine learning models require large amounts of data",
"你好世界", // Hello world in Chinese
"こんにちは世界", // Hello world in Japanese
"안녕하세요 세계", // Hello world in Korean
"مرحبا بالعالم", // Hello world in Arabic
"Γειά σου Κόσμε", // Hello world in Greek
"Привет мир", // Hello world in Russian
"नमस्ते दुनिया" // Hello world in Hindi
};
try {
// Train the tokenizer
std::cout << "Training tokenizer with Unicode text..." << std::endl;
tokenizer.train(corpus, 1000);
std::cout << "Vocabulary size: " << tokenizer.vocab_size() << std::endl;
// Test encoding/decoding with various scripts
std::vector<std::string> test_texts = {
"hello world",
"你好世界",
"こんにちは世界",
"مرحبا بالعالم",
"Привет мир"
};
for (const auto& test_text : test_texts) {
auto tokens = tokenizer.encode(test_text);
std::string decoded = tokenizer.decode(tokens);
std::cout << "\nOriginal: " << test_text << std::endl;
std::cout << "Tokens: ";
for (auto token : tokens) {
std::cout << token << " ";
}
std::cout << std::endl;
std::cout << "Decoded: " << decoded << std::endl;
std::cout << "Match: " << (test_text == decoded ? "YES" : "NO") << std::endl;
}
// Save and load test
tokenizer.save("unicode_bpe_model.txt");
lm::BPETokenizer loaded_tokenizer;
if (loaded_tokenizer.load("unicode_bpe_model.txt")) {
std::cout << "\nSuccessfully loaded Unicode tokenizer" << std::endl;
std::cout << "Loaded vocabulary size: " << loaded_tokenizer.vocab_size() << std::endl;
// Test with the loaded tokenizer
std::string test_text = "你好世界";
auto tokens = loaded_tokenizer.encode(test_text);
std::string decoded = loaded_tokenizer.decode(tokens);
std::cout << "Loaded tokenizer test:" << std::endl;
std::cout << "Original: " << test_text << std::endl;
std::cout << "Decoded: " << decoded << std::endl;
std::cout << "Match: " << (test_text == decoded ? "YES" : "NO") << std::endl;
}
} catch (const std::exception& e) {
std::cerr << "Error: " << e.what() << std::endl;
return 1;
}
return 0;
}

16
src/tokenizer/CMakeLists.txt Executable file
View File

@ -0,0 +1,16 @@
# Tokenizer library
add_library(lm_tokenizer
bpe_tokenizer.cpp
unicode_utils.cpp
)
target_include_directories(lm_tokenizer
PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/../../include
)
target_link_libraries(lm_tokenizer
PRIVATE
ICU::uc
ICU::i18n
)

758
src/tokenizer/bpe_tokenizer.cpp Executable file
View File

@ -0,0 +1,758 @@
#include "lm/tokenizer/bpe_tokenizer.hpp"
#include "lm/tokenizer/unicode_utils.hpp"
#include <fstream>
#include <sstream>
#include <queue>
#include <algorithm>
#include <stdexcept>
#include <iostream>
#include <sys/resource.h>
// Add CPU-specific optimizations
#ifdef __SSE4_2__
#include <nmmintrin.h> // For SSE4.2 intrinsics
#endif
namespace lm {
// Custom hash function for pair<TokenID, TokenID>
struct PairHash {
size_t operator()(const std::pair<TokenID, TokenID>& p) const {
return (static_cast<size_t>(p.first) << 16) | p.second;
}
};
// Memory tracking function
size_t get_peak_memory_usage() {
#ifdef __linux__
std::ifstream status("/proc/self/status");
std::string line;
while (std::getline(status, line)) {
if (line.compare(0, 6, "VmPeak") == 0) {
std::istringstream iss(line);
std::string key;
size_t value;
std::string unit;
iss >> key >> value >> unit;
if (unit == "kB") {
return value * 1024; // Convert to bytes
}
}
}
#endif
return 0;
}
// String interning class
class StringInternPool {
std::unordered_map<std::string, std::shared_ptr<const std::string>> pool;
public:
std::shared_ptr<const std::string> intern(const std::string& str) {
auto it = pool.find(str);
if (it != pool.end()) {
return it->second;
}
auto shared_str = std::make_shared<std::string>(str);
pool[str] = shared_str;
return shared_str;
}
void clear() {
pool.clear();
}
};
struct BPETokenizer::Impl {
std::unordered_map<std::string, TokenID> vocab;
std::unordered_map<TokenID, std::string> inv_vocab;
std::unordered_map<std::pair<TokenID, TokenID>, TokenID, PairHash> merges;
std::unordered_map<std::string, TokenID> special_tokens;
std::string unknown_token = "<unk>";
TokenID unknown_token_id = 0;
TokenID next_token_id = 0;
bool normalization_enabled = true;
bool byte_fallback_enabled = true;
StringInternPool string_pool;
// Helper functions
std::vector<std::string> split_text(const std::string& text) const;
std::vector<TokenID> word_to_token_ids(const std::string& word) const;
void initialize_vocab();
void count_word_frequencies(const std::vector<std::string>& words,
std::unordered_map<std::string, int>& word_counts) const;
void get_pair_counts(const std::unordered_map<std::string, int>& word_counts,
std::unordered_map<std::pair<TokenID, TokenID>, int, PairHash>& pair_counts) const;
void perform_merge(const std::pair<TokenID, TokenID>& pair, TokenID new_token_id,
std::unordered_map<std::string, int>& word_counts);
// CPU Optimization: Batch processing
void process_string_batch(const std::vector<std::string>& batch);
};
BPETokenizer::BPETokenizer() : pimpl_(new Impl) {
pimpl_->initialize_vocab();
}
BPETokenizer::~BPETokenizer() = default;
void BPETokenizer::Impl::initialize_vocab() {
vocab.reserve(65536); // More realistic initial allocation
inv_vocab.reserve(65536);
// Add bytes with optimized insertion
for (int i = 0; i < 256; i++) {
std::string token(1, static_cast<char>(i));
vocab.emplace(token, next_token_id);
inv_vocab.emplace(next_token_id++, std::move(token));
}
// Add unknown token
auto [it, inserted] = vocab.emplace(unknown_token, next_token_id);
if (inserted) {
inv_vocab.emplace(next_token_id, unknown_token);
special_tokens.emplace(unknown_token, next_token_id);
unknown_token_id = next_token_id++;
} else {
unknown_token_id = it->second;
}
}
std::vector<std::string> BPETokenizer::Impl::split_text(const std::string& text) const {
if (normalization_enabled) {
std::string normalized = unicode::normalize(text);
return unicode::unicode_split(normalized);
} else {
std::vector<std::string> words;
std::istringstream iss(text);
std::string word;
while (iss >> word) {
words.push_back(word);
}
return words;
}
}
std::vector<TokenID> BPETokenizer::Impl::word_to_token_ids(const std::string& word) const {
// Hot path optimization: use local references to avoid repeated pointer dereferencing
const auto& local_vocab = vocab;
const TokenID local_unknown_id = unknown_token_id;
const bool local_byte_fallback = byte_fallback_enabled;
const bool local_normalization = normalization_enabled;
std::vector<TokenID> tokens;
if (local_normalization) {
std::string normalized = unicode::normalize(word);
auto characters = unicode::split_on_character_boundaries(normalized);
for (const auto& character : characters) {
auto it = local_vocab.find(character);
if (it != local_vocab.end()) {
tokens.push_back(it->second);
} else if (local_byte_fallback) {
// If character not found, try to split into bytes
for (unsigned char c : character) {
std::string byte_str(1, static_cast<char>(c));
auto byte_it = local_vocab.find(byte_str);
if (byte_it != local_vocab.end()) {
tokens.push_back(byte_it->second);
} else {
tokens.push_back(local_unknown_id);
}
}
} else {
tokens.push_back(local_unknown_id);
}
}
} else {
// Non-Unicode mode: treat as ASCII
for (char c : word) {
std::string token(1, c);
auto it = local_vocab.find(token);
if (it != local_vocab.end()) {
tokens.push_back(it->second);
} else {
tokens.push_back(local_unknown_id);
}
}
}
return tokens;
}
void BPETokenizer::Impl::count_word_frequencies(
const std::vector<std::string>& words,
std::unordered_map<std::string, int>& word_counts) const {
// Use a local reference to avoid repeated map lookups
const auto& local_vocab = vocab;
for (const auto& word : words) {
// Use string_view to avoid copies where possible
auto it = word_counts.find(word);
if (it != word_counts.end()) {
it->second++;
} else {
word_counts[word] = 1;
}
}
}
/*void BPETokenizer::Impl::perform_merge(const std::pair<TokenID, TokenID>& pair, TokenID new_token_id,
std::unordered_map<std::string, int>& word_counts) {
std::string new_token = inv_vocab.at(pair.first) + inv_vocab.at(pair.second);
// Use move semantics to avoid copies
auto shared_token = string_pool.intern(std::move(new_token));
// Add new token to vocabulary
vocab[*shared_token] = new_token_id;
inv_vocab[new_token_id] = *shared_token;
merges[pair] = new_token_id;
// Update word counts more efficiently
std::vector<std::string> to_remove;
std::unordered_map<std::string, int> new_counts;
for (auto& [word, count] : word_counts) {
std::string new_word;
new_word.reserve(word.size()); // Pre-allocate
size_t pos = 0;
while (pos < word.size()) {
if (pos < word.size() - 1 &&
word.substr(pos, inv_vocab.at(pair.first).size()) == inv_vocab.at(pair.first) &&
word.substr(pos + inv_vocab.at(pair.first).size(), inv_vocab.at(pair.second).size()) == inv_vocab.at(pair.second)) {
new_word += *shared_token;
pos += inv_vocab.at(pair.first).size() + inv_vocab.at(pair.second).size();
} else {
new_word += word[pos++];
}
}
if (new_word != word) {
to_remove.push_back(word);
new_counts[new_word] += count;
}
}
// Remove old words and add new ones
for (const auto& word : to_remove) {
word_counts.erase(word);
}
for (auto& [word, count] : new_counts) {
word_counts[word] += count;
}
}*/
void BPETokenizer::Impl::perform_merge(const std::pair<TokenID, TokenID>& pair, TokenID new_token_id,
std::unordered_map<std::string, int>& word_counts) {
std::string new_token = inv_vocab.at(pair.first) + inv_vocab.at(pair.second);
auto shared_token = string_pool.intern(new_token);
// Add new token to vocabulary
vocab[*shared_token] = new_token_id;
inv_vocab[new_token_id] = *shared_token;
// Record the merge
merges[pair] = new_token_id;
// Update word counts with new merges
std::unordered_map<std::string, int> new_word_counts;
for (const auto& [word, count] : word_counts) {
std::string new_word = word;
size_t pos = 0;
while ((pos = new_word.find(inv_vocab.at(pair.first) + inv_vocab.at(pair.second), pos)) != std::string::npos) {
new_word.replace(pos, 2, *shared_token);
pos += shared_token->length();
}
new_word_counts[new_word] += count;
}
word_counts = std::move(new_word_counts);
}
// CPU Optimization: Batch processing for better cache utilization
void BPETokenizer::Impl::process_string_batch(const std::vector<std::string>& batch) {
const size_t batch_size = batch.size();
#ifdef __SSE4_2__
// Predefined patterns to detect
const std::string url_pattern = "http";
const std::string email_pattern = "@";
const std::string number_pattern = "0123456789";
// Precompute pattern lengths
const int url_len = url_pattern.size();
const int email_len = email_pattern.size();
const int number_len = number_pattern.size();
// Load patterns into SSE registers
__m128i url_pattern_vec = _mm_loadu_si128(reinterpret_cast<const __m128i*>(url_pattern.c_str()));
__m128i email_pattern_vec = _mm_loadu_si128(reinterpret_cast<const __m128i*>(email_pattern.c_str()));
__m128i number_pattern_vec = _mm_loadu_si128(reinterpret_cast<const __m128i*>(number_pattern.c_str()));
for (size_t i = 0; i < batch_size; i++) {
const std::string& str = batch[i];
const size_t len = str.size();
const char* data = str.data();
int url_count = 0;
int email_count = 0;
int number_count = 0;
// Process 16 characters at a time
size_t j = 0;
for (; j + 16 <= len; j += 16) {
__m128i chunk = _mm_loadu_si128(reinterpret_cast<const __m128i*>(data + j));
// Search for URL pattern using SSE4.2 string instructions
int url_idx = _mm_cmpestri(
url_pattern_vec, url_len,
chunk, 16,
_SIDD_CMP_EQUAL_ORDERED | _SIDD_UBYTE_OPS
);
if (url_idx < 16) {
url_count++;
// Optional: Add special URL token if not already in vocabulary
auto it = vocab.find("<URL>");
if (it == vocab.end()) {
vocab["<URL>"] = next_token_id;
inv_vocab[next_token_id] = "<URL>";
next_token_id++;
}
}
// Search for email pattern
int email_idx = _mm_cmpestri(
email_pattern_vec, email_len,
chunk, 16,
_SIDD_CMP_EQUAL_ORDERED | _SIDD_UBYTE_OPS
);
if (email_idx < 16) {
email_count++;
// Optional: Add special email token
auto it = vocab.find("<EMAIL>");
if (it == vocab.end()) {
vocab["<EMAIL>"] = next_token_id;
inv_vocab[next_token_id] = "<EMAIL>";
next_token_id++;
}
}
// Search for number patterns using range comparison
__m128i digit_lo = _mm_cmpgt_epi8(chunk, _mm_set1_epi8('0' - 1));
__m128i digit_hi = _mm_cmplt_epi8(chunk, _mm_set1_epi8('9' + 1));
__m128i digit_mask = _mm_and_si128(digit_lo, digit_hi);
int digit_count = _mm_popcnt_u32(_mm_movemask_epi8(digit_mask));
if (digit_count >= 3) { // Threshold for considering it a number
number_count++;
// Optional: Add special number token
auto it = vocab.find("<NUMBER>");
if (it == vocab.end()) {
vocab["<NUMBER>"] = next_token_id;
inv_vocab[next_token_id] = "<NUMBER>";
next_token_id++;
}
}
// Additional pattern detection can be added here
}
// Process remaining characters with traditional methods
for (; j < len; j++) {
// Check for URL pattern
if (j + url_len <= len &&
strncmp(data + j, url_pattern.c_str(), url_len) == 0) {
url_count++;
j += url_len - 1; // Skip ahead
}
// Check for email pattern
if (j + email_len <= len &&
strncmp(data + j, email_pattern.c_str(), email_len) == 0) {
email_count++;
}
// Check for number pattern
if (data[j] >= '0' && data[j] <= '9') {
number_count++;
}
}
// Use pattern counts to make tokenization decisions
if (url_count > 0) {
// Apply special tokenization for URLs
}
if (email_count > 0) {
// Apply special tokenization for emails
}
if (number_count > 5) { // Threshold for considering as numeric content
// Apply special tokenization for numbers
}
}
#else
// Fallback implementation without SSE
const std::string url_pattern = "http";
const std::string email_pattern = "@";
for (const auto& str : batch) {
int url_count = 0;
int email_count = 0;
int number_count = 0;
// Simple pattern matching without SSE
for (size_t j = 0; j < str.size(); j++) {
// Check for URL pattern
if (j + url_pattern.size() <= str.size() &&
str.substr(j, url_pattern.size()) == url_pattern) {
url_count++;
j += url_pattern.size() - 1;
}
// Check for email pattern
if (j + email_pattern.size() <= str.size() &&
str.substr(j, email_pattern.size()) == email_pattern) {
email_count++;
}
// Check for numbers
if (str[j] >= '0' && str[j] <= '9') {
number_count++;
}
}
// Same logic as above for tokenization decisions
if (url_count > 0) {
auto it = vocab.find("<URL>");
if (it == vocab.end()) {
vocab["<URL>"] = next_token_id;
inv_vocab[next_token_id] = "<URL>";
next_token_id++;
}
}
// Similar handling for other patterns
}
#endif
}
void BPETokenizer::train(const std::vector<std::string>& corpus, size_t vocab_size) {
size_t start_memory = get_peak_memory_usage();
if (corpus.empty()) {
throw std::invalid_argument("Corpus cannot be empty");
}
// Split text into words
std::vector<std::string> words;
words.reserve(corpus.size() * 10); // Estimate average words per sentence
for (const auto& text : corpus) {
auto text_words = pimpl_->split_text(text);
words.insert(words.end(), text_words.begin(), text_words.end());
}
// Count word frequencies using strings
std::unordered_map<std::string, int> word_counts;
word_counts.reserve(words.size());
pimpl_->count_word_frequencies(words, word_counts);
// Track token frequencies for pruning
std::unordered_map<TokenID, size_t> token_frequencies;
// Initialize token frequencies
for (const auto& [word, count] : word_counts) {
auto tokens = pimpl_->word_to_token_ids(word);
for (TokenID token : tokens) {
token_frequencies[token] += count;
}
}
// BPE training algorithm with safety limit
int iteration = 0;
int max_iterations = 10000;
// Pruning function - remove infrequent tokens
auto prune_infrequent_tokens = [&](size_t frequency_threshold = 2) {
std::vector<TokenID> tokens_to_remove;
// Identify tokens to remove (excluding special tokens)
for (const auto& [token_id, freq] : token_frequencies) {
if (freq < frequency_threshold) {
// Check if this is a special token
std::string token_text = pimpl_->inv_vocab.at(token_id);
if (pimpl_->special_tokens.find(token_text) == pimpl_->special_tokens.end()) {
tokens_to_remove.push_back(token_id);
}
}
}
// Remove tokens from vocabulary
for (TokenID token_id : tokens_to_remove) {
std::string token_text = pimpl_->inv_vocab.at(token_id);
// Remove from vocabulary mappings
pimpl_->vocab.erase(token_text);
pimpl_->inv_vocab.erase(token_id);
token_frequencies.erase(token_id);
// Update word counts to use subword components instead of removed tokens
std::unordered_map<std::string, int> updated_word_counts;
for (const auto& [word, count] : word_counts) {
std::string updated_word = word;
size_t pos = 0;
// Replace all occurrences of the token text with its byte representation
while ((pos = updated_word.find(token_text, pos)) != std::string::npos) {
// Replace with byte fallback
std::string replacement;
for (unsigned char c : token_text) {
std::string byte_str(1, static_cast<char>(c));
replacement += byte_str;
}
updated_word.replace(pos, token_text.size(), replacement);
pos += replacement.size();
}
updated_word_counts[updated_word] += count;
}
// Update the word_counts with the modified words
word_counts = std::move(updated_word_counts);
}
// Recalculate token frequencies after pruning
token_frequencies.clear();
for (const auto& [word, count] : word_counts) {
auto tokens = pimpl_->word_to_token_ids(word);
for (TokenID token : tokens) {
token_frequencies[token] += count;
}
}
};
while (pimpl_->vocab.size() < vocab_size && iteration < max_iterations) {
// Count pairs
std::unordered_map<std::pair<TokenID, TokenID>, int, PairHash> pair_counts;
pimpl_->get_pair_counts(word_counts, pair_counts);
if (pair_counts.empty()) {
std::cout << "No more pairs to merge. Stopping early." << std::endl;
break;
}
// Find most frequent pair
auto max_pair = std::max_element(
pair_counts.begin(), pair_counts.end(),
[](const auto& a, const auto& b) { return a.second < b.second; }
);
// Debug output
if (iteration % 100 == 0) {
std::cout << "Iteration " << iteration
<< ", Vocab size: " << pimpl_->vocab.size()
<< ", Most frequent pair: (" << max_pair->first.first
<< "," << max_pair->first.second << ") count: "
<< max_pair->second << std::endl;
}
// Perform merge
pimpl_->perform_merge(max_pair->first, pimpl_->next_token_id, word_counts);
pimpl_->next_token_id++;
// Update token frequencies
token_frequencies.clear();
for (const auto& [word, count] : word_counts) {
auto tokens = pimpl_->word_to_token_ids(word);
for (TokenID token : tokens) {
token_frequencies[token] += count;
}
}
// Periodically prune infrequent tokens
if (iteration % 500 == 0 && iteration > 0) {
size_t pre_prune_size = pimpl_->vocab.size();
prune_infrequent_tokens(2); // Remove tokens with frequency < 2
std::cout << "Pruned " << (pre_prune_size - pimpl_->vocab.size())
<< " infrequent tokens. New vocab size: "
<< pimpl_->vocab.size() << std::endl;
}
// Periodically check memory usage
if (iteration % 500 == 0) {
size_t current_memory = get_peak_memory_usage();
std::cout << "Memory after " << iteration << " iterations: "
<< (current_memory - start_memory) / (1024 * 1024) << "MB\n";
}
iteration++;
}
// Final pruning after training completes
size_t pre_prune_size = pimpl_->vocab.size();
prune_infrequent_tokens(3); // Remove tokens with frequency < 3
std::cout << "Final pruning: Removed " << (pre_prune_size - pimpl_->vocab.size())
<< " tokens. Final vocab size: " << pimpl_->vocab.size() << std::endl;
if (iteration >= max_iterations) {
std::cout << "Reached maximum iterations. Stopping training." << std::endl;
}
size_t end_memory = get_peak_memory_usage();
std::cout << "Training completed in " << iteration << " iterations\n";
std::cout << "Peak memory used: " << (end_memory - start_memory) / (1024 * 1024) << "MB\n";
std::cout << "Final vocabulary size: " << pimpl_->vocab.size() << std::endl;
// Clear the string intern pool to free memory
pimpl_->string_pool.clear();
}
std::string BPETokenizer::decode(const std::vector<TokenID>& tokens) const {
std::string text;
for (TokenID token_id : tokens) {
if (pimpl_->inv_vocab.find(token_id) != pimpl_->inv_vocab.end()) {
text += pimpl_->inv_vocab.at(token_id);
} else {
text += pimpl_->unknown_token;
}
}
return text;
}
bool BPETokenizer::save(const std::string& filename) const {
std::ofstream file(filename);
if (!file.is_open()) {
return false;
}
// Save vocabulary
file << pimpl_->vocab.size() << "\n";
for (const auto& [token, id] : pimpl_->vocab) {
file << id << " " << token << "\n";
}
// Save merges
file << pimpl_->merges.size() << "\n";
for (const auto& [pair, new_id] : pimpl_->merges) {
file << pair.first << " " << pair.second << " " << new_id << "\n";
}
return true;
}
bool BPETokenizer::load(const std::string& filename) {
std::ifstream file(filename);
if (!file.is_open()) {
return false;
}
// Clear existing data
pimpl_->vocab.clear();
pimpl_->inv_vocab.clear();
pimpl_->merges.clear();
// Load vocabulary
size_t vocab_size;
file >> vocab_size;
for (size_t i = 0; i < vocab_size; i++) {
TokenID id;
std::string token;
file >> id;
std::getline(file, token);
// Remove leading space
if (!token.empty() && token[0] == ' ') {
token = token.substr(1);
}
pimpl_->vocab[token] = id;
pimpl_->inv_vocab[id] = token;
}
// Load merges
size_t merge_count;
file >> merge_count;
for (size_t i = 0; i < merge_count; i++) {
TokenID first, second, new_id;
file >> first >> second >> new_id;
pimpl_->merges[{first, second}] = new_id;
}
return true;
}
size_t BPETokenizer::vocab_size() const {
return pimpl_->vocab.size();
}
std::string BPETokenizer::id_to_token(TokenID id) const {
if (pimpl_->inv_vocab.find(id) != pimpl_->inv_vocab.end()) {
return pimpl_->inv_vocab.at(id);
}
return pimpl_->unknown_token;
}
TokenID BPETokenizer::token_to_id(const std::string& token) const {
if (pimpl_->vocab.find(token) != pimpl_->vocab.end()) {
return pimpl_->vocab.at(token);
}
return pimpl_->unknown_token_id;
}
void BPETokenizer::set_unknown_token(const std::string& token) {
pimpl_->unknown_token = token;
// If token already exists in vocab, use its ID
if (pimpl_->vocab.find(token) != pimpl_->vocab.end()) {
pimpl_->unknown_token_id = pimpl_->vocab.at(token);
} else {
// Add to vocab
pimpl_->vocab[token] = pimpl_->next_token_id;
pimpl_->inv_vocab[pimpl_->next_token_id] = token;
pimpl_->unknown_token_id = pimpl_->next_token_id;
pimpl_->next_token_id++;
}
}
void BPETokenizer::add_special_token(const std::string& token) {
if (pimpl_->vocab.find(token) == pimpl_->vocab.end()) {
pimpl_->vocab[token] = pimpl_->next_token_id;
pimpl_->inv_vocab[pimpl_->next_token_id] = token;
pimpl_->special_tokens[token] = pimpl_->next_token_id;
pimpl_->next_token_id++;
}
}
void BPETokenizer::set_normalization(bool enabled) {
pimpl_->normalization_enabled = enabled;
}
void BPETokenizer::set_byte_fallback(bool enabled) {
pimpl_->byte_fallback_enabled = enabled;
}
void BPETokenizer::Impl::get_pair_counts(
const std::unordered_map<std::string, int>& word_counts,
std::unordered_map<std::pair<TokenID, TokenID>, int, PairHash>& pair_counts) const {
// Pre-allocate memory for better performance
pair_counts.reserve(word_counts.size() * 10);
for (const auto& [word, count] : word_counts) {
auto tokens = word_to_token_ids(word);
for (size_t i = 0; i < tokens.size() - 1; i++) {
auto pair = std::make_pair(tokens[i], tokens[i+1]);
pair_counts[pair] += count;
}
}
}
} // namespace lm

145
src/tokenizer/unicode_utils.cpp Executable file
View File

@ -0,0 +1,145 @@
/*# Unicode Utilities Implementation File
Here's the complete `src/tokenizer/unicode_utils.cpp` file:
```cpp*/
#include "lm/tokenizer/unicode_utils.hpp"
#include <unicode/uchar.h>
#include <unicode/unistr.h>
#include <unicode/normlzr.h>
#include <unicode/ustring.h>
#include <stdexcept>
#include <algorithm>
namespace lm::unicode {
bool is_whitespace(uint32_t codepoint) {
return u_isUWhiteSpace(codepoint);
}
bool is_punctuation(uint32_t codepoint) {
return u_ispunct(codepoint);
}
bool is_control(uint32_t codepoint) {
return u_iscntrl(codepoint);
}
std::string normalize(const std::string& text) {
try {
icu::UnicodeString unicode_str = icu::UnicodeString::fromUTF8(text);
icu::UnicodeString normalized;
UErrorCode status = U_ZERO_ERROR;
icu::Normalizer::normalize(unicode_str, UNORM_NFC, 0, normalized, status);
if (U_FAILURE(status)) {
throw std::runtime_error("Unicode normalization failed");
}
std::string result;
normalized.toUTF8String(result);
return result;
} catch (const std::exception& e) {
throw std::runtime_error("Unicode normalization error: " + std::string(e.what()));
}
}
std::vector<CodePoint> to_code_points(const std::string& text) {
std::vector<CodePoint> code_points;
for (size_t i = 0; i < text.size(); ) {
CodePoint cp;
uint32_t codepoint;
int offset = 0;
// Decode UTF-8
U8_NEXT(text.c_str(), i, text.size(), codepoint);
if (codepoint == U_SENTINEL) {
// Handle invalid UTF-8 gracefully instead of throwing
// Use replacement character (U+FFFD) for invalid sequences
cp.value = 0xFFFD;
cp.utf8 = "<EFBFBD>"; // Replacement character
code_points.push_back(cp);
// Skip this byte and continue
i++;
continue;
}
// Get the UTF-8 bytes for this code point
char utf8_buf[5] = {0};
U8_APPEND_UNSAFE(utf8_buf, offset, codepoint);
cp.value = codepoint;
cp.utf8 = std::string(utf8_buf, offset);
code_points.push_back(cp);
i += offset;
}
return code_points;
}
std::string from_code_points(const std::vector<CodePoint>& code_points) {
std::string result;
for (const auto& cp : code_points) {
result += cp.utf8;
}
return result;
}
std::vector<std::string> unicode_split(const std::string& text) {
std::vector<std::string> words;
std::string current_word;
auto code_points = to_code_points(text);
for (const auto& cp : code_points) {
if (is_whitespace(cp.value)) {
if (!current_word.empty()) {
words.push_back(current_word);
current_word.clear();
}
} else {
current_word += cp.utf8;
}
}
if (!current_word.empty()) {
words.push_back(current_word);
}
return words;
}
std::vector<std::string> split_on_character_boundaries(const std::string& text) {
std::vector<std::string> characters;
auto code_points = to_code_points(text);
for (const auto& cp : code_points) {
characters.push_back(cp.utf8);
}
return characters;
}
} // namespace lm::unicode
/*```
This implementation provides comprehensive Unicode support for the BPE tokenizer with:
1. **Character classification** using ICU library functions
2. **Unicode normalization** to NFC form for consistent processing
3. **UTF-8 encoding/decoding** with proper error handling
4. **Unicode-aware text splitting** based on character boundaries
5. **Robust error handling** for invalid UTF-8 sequences
Key features:
- **ICU library integration** for professional-grade Unicode handling
- **Proper UTF-8 sequence processing** with boundary detection
- **Character classification** for whitespace, punctuation, and control characters
- **Normalization** to ensure consistent representation of equivalent Unicode sequences
- **Exception safety** with proper error reporting
This implementation enables the BPE tokenizer to properly handle multilingual text, including complex scripts and emoji, while maintaining compatibility with ASCII text.*/