OreX/forge-builds
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

initial commit

Browse Source
This commit is contained in:
Bacruru Sakaguchi
2025-10-21 13:37:07 +07:00
commit 9cd16e276a
1574 changed files with 2675557 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
packages_3rdparty/README.md vendored Executable file
View File

@@ -0,0 +1 @@
Please follow the standard of https://github.com/opencv/opencv/tree/315f85d4f484c1e2fa043c73ac3fdd9fc5997ee7/3rdparty when PR or modifying files.

674
packages_3rdparty/comfyui_lora_collection/LICENSE vendored Executable file
View File

@@ -0,0 +1,674 @@
GNU GENERAL PUBLIC LICENSE
Version 3, 29 June 2007
Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
Everyone is permitted to copy and distribute verbatim copies
of this license document, but changing it is not allowed.
Preamble
The GNU General Public License is a free, copyleft license for
software and other kinds of works.
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.
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.
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.
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.
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.
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.
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.
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.
The precise terms and conditions for copying, distribution and
modification follow.
TERMS AND CONDITIONS
0. Definitions.
"This License" refers to version 3 of the GNU General Public License.
"Copyright" also means copyright-like laws that apply to other kinds of
works, such as semiconductor masks.
"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.
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.
A "covered work" means either the unmodified Program or a work based
on the Program.
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 "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.
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.
1. Source Code.
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.
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.
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 "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.
The Corresponding Source need not include anything that users
can regenerate automatically from other parts of the Corresponding
Source.
The Corresponding Source for a work in source code form is that
same work.
2. Basic Permissions.
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.
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.
Conveying under any other circumstances is permitted solely under
the conditions stated below. Sublicensing is not allowed; section 10
makes it unnecessary.
3. Protecting Users' Legal Rights From Anti-Circumvention Law.
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.
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.
4. Conveying Verbatim Copies.
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.
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.
5. Conveying Modified Source Versions.
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:
a) The work must carry prominent notices stating that you modified
it, and giving a relevant date.
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".
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.
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. 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.
14. Revised Versions of this License.
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.
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.
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.
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 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 General Public License for more details.
You should have received a copy of the GNU 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.
If the program does terminal interaction, make it output a short
notice like this when it starts in an interactive mode:
<program> Copyright (C) <year> <name of author>
This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
This is free software, and you are welcome to redistribute it
under certain conditions; type `show c' for details.
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/licenses/why-not-lgpl.html>.

351
packages_3rdparty/comfyui_lora_collection/lora.py vendored Executable file
View File

@@ -0,0 +1,351 @@
"""
This file is part of ComfyUI.
Copyright (C) 2024 Comfy
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.
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.
You should have received a copy of the GNU General Public License
along with this program. If not, see <https://www.gnu.org/licenses/>.
"""
from packages_3rdparty.comfyui_lora_collection import utils
LORA_CLIP_MAP = {
"mlp.fc1": "mlp_fc1",
"mlp.fc2": "mlp_fc2",
"self_attn.k_proj": "self_attn_k_proj",
"self_attn.q_proj": "self_attn_q_proj",
"self_attn.v_proj": "self_attn_v_proj",
"self_attn.out_proj": "self_attn_out_proj",
}
def load_lora(lora, to_load):
# BFL loras for Flux; from ComfyUI: comfy/lora_convert.py
def convert_lora_bfl_control(sd):
import torch
sd_out = {}
for k in sd:
k_to = "diffusion_model.{}".format(k.replace(".lora_B.bias", ".diff_b").replace("_norm.scale", "_norm.scale.set_weight"))
sd_out[k_to] = sd[k]
return sd_out
if "img_in.lora_A.weight" in lora and "single_blocks.0.norm.key_norm.scale" in lora:
lora = convert_lora_bfl_control(lora)
patch_dict = {}
loaded_keys = set()
for x in to_load:
alpha_name = "{}.alpha".format(x)
alpha = None
if alpha_name in lora.keys():
alpha = lora[alpha_name].item()
loaded_keys.add(alpha_name)
dora_scale_name = "{}.dora_scale".format(x)
dora_scale = None
if dora_scale_name in lora.keys():
dora_scale = lora[dora_scale_name]
loaded_keys.add(dora_scale_name)
regular_lora = "{}.lora_up.weight".format(x)
diffusers_lora = "{}_lora.up.weight".format(x)
diffusers2_lora = "{}.lora_B.weight".format(x)
diffusers3_lora = "{}.lora.up.weight".format(x)
transformers_lora = "{}.lora_linear_layer.up.weight".format(x)
A_name = None
if regular_lora in lora.keys():
A_name = regular_lora
B_name = "{}.lora_down.weight".format(x)
mid_name = "{}.lora_mid.weight".format(x)
elif diffusers_lora in lora.keys():
A_name = diffusers_lora
B_name = "{}_lora.down.weight".format(x)
mid_name = None
elif diffusers2_lora in lora.keys():
A_name = diffusers2_lora
B_name = "{}.lora_A.weight".format(x)
mid_name = None
elif diffusers3_lora in lora.keys():
A_name = diffusers3_lora
B_name = "{}.lora.down.weight".format(x)
mid_name = None
elif transformers_lora in lora.keys():
A_name = transformers_lora
B_name ="{}.lora_linear_layer.down.weight".format(x)
mid_name = None
if A_name is not None:
mid = None
if mid_name is not None and mid_name in lora.keys():
mid = lora[mid_name]
loaded_keys.add(mid_name)
patch_dict[to_load[x]] = ("lora", (lora[A_name], lora[B_name], alpha, mid, dora_scale))
loaded_keys.add(A_name)
loaded_keys.add(B_name)
######## loha
hada_w1_a_name = "{}.hada_w1_a".format(x)
hada_w1_b_name = "{}.hada_w1_b".format(x)
hada_w2_a_name = "{}.hada_w2_a".format(x)
hada_w2_b_name = "{}.hada_w2_b".format(x)
hada_t1_name = "{}.hada_t1".format(x)
hada_t2_name = "{}.hada_t2".format(x)
if hada_w1_a_name in lora.keys():
hada_t1 = None
hada_t2 = None
if hada_t1_name in lora.keys():
hada_t1 = lora[hada_t1_name]
hada_t2 = lora[hada_t2_name]
loaded_keys.add(hada_t1_name)
loaded_keys.add(hada_t2_name)
patch_dict[to_load[x]] = ("loha", (lora[hada_w1_a_name], lora[hada_w1_b_name], alpha, lora[hada_w2_a_name], lora[hada_w2_b_name], hada_t1, hada_t2, dora_scale))
loaded_keys.add(hada_w1_a_name)
loaded_keys.add(hada_w1_b_name)
loaded_keys.add(hada_w2_a_name)
loaded_keys.add(hada_w2_b_name)
######## lokr
lokr_w1_name = "{}.lokr_w1".format(x)
lokr_w2_name = "{}.lokr_w2".format(x)
lokr_w1_a_name = "{}.lokr_w1_a".format(x)
lokr_w1_b_name = "{}.lokr_w1_b".format(x)
lokr_t2_name = "{}.lokr_t2".format(x)
lokr_w2_a_name = "{}.lokr_w2_a".format(x)
lokr_w2_b_name = "{}.lokr_w2_b".format(x)
lokr_w1 = None
if lokr_w1_name in lora.keys():
lokr_w1 = lora[lokr_w1_name]
loaded_keys.add(lokr_w1_name)
lokr_w2 = None
if lokr_w2_name in lora.keys():
lokr_w2 = lora[lokr_w2_name]
loaded_keys.add(lokr_w2_name)
lokr_w1_a = None
if lokr_w1_a_name in lora.keys():
lokr_w1_a = lora[lokr_w1_a_name]
loaded_keys.add(lokr_w1_a_name)
lokr_w1_b = None
if lokr_w1_b_name in lora.keys():
lokr_w1_b = lora[lokr_w1_b_name]
loaded_keys.add(lokr_w1_b_name)
lokr_w2_a = None
if lokr_w2_a_name in lora.keys():
lokr_w2_a = lora[lokr_w2_a_name]
loaded_keys.add(lokr_w2_a_name)
lokr_w2_b = None
if lokr_w2_b_name in lora.keys():
lokr_w2_b = lora[lokr_w2_b_name]
loaded_keys.add(lokr_w2_b_name)
lokr_t2 = None
if lokr_t2_name in lora.keys():
lokr_t2 = lora[lokr_t2_name]
loaded_keys.add(lokr_t2_name)
if (lokr_w1 is not None) or (lokr_w2 is not None) or (lokr_w1_a is not None) or (lokr_w2_a is not None):
patch_dict[to_load[x]] = ("lokr", (lokr_w1, lokr_w2, alpha, lokr_w1_a, lokr_w1_b, lokr_w2_a, lokr_w2_b, lokr_t2, dora_scale))
#glora
a1_name = "{}.a1.weight".format(x)
a2_name = "{}.a2.weight".format(x)
b1_name = "{}.b1.weight".format(x)
b2_name = "{}.b2.weight".format(x)
if a1_name in lora:
patch_dict[to_load[x]] = ("glora", (lora[a1_name], lora[a2_name], lora[b1_name], lora[b2_name], alpha, dora_scale))
loaded_keys.add(a1_name)
loaded_keys.add(a2_name)
loaded_keys.add(b1_name)
loaded_keys.add(b2_name)
w_norm_name = "{}.w_norm".format(x)
b_norm_name = "{}.b_norm".format(x)
w_norm = lora.get(w_norm_name, None)
b_norm = lora.get(b_norm_name, None)
if w_norm is not None:
loaded_keys.add(w_norm_name)
patch_dict[to_load[x]] = ("diff", (w_norm,))
if b_norm is not None:
loaded_keys.add(b_norm_name)
patch_dict["{}.bias".format(to_load[x][:-len(".weight")])] = ("diff", (b_norm,))
diff_name = "{}.diff".format(x)
diff_weight = lora.get(diff_name, None)
if diff_weight is not None:
patch_dict[to_load[x]] = ("diff", (diff_weight,))
loaded_keys.add(diff_name)
diff_bias_name = "{}.diff_b".format(x)
diff_bias = lora.get(diff_bias_name, None)
if diff_bias is not None:
patch_dict["{}.bias".format(to_load[x][:-len(".weight")])] = ("diff", (diff_bias,))
loaded_keys.add(diff_bias_name)
set_weight_name = "{}.set_weight".format(x)
set_weight = lora.get(set_weight_name, None)
if set_weight is not None:
patch_dict[to_load[x]] = ("set", (set_weight,))
loaded_keys.add(set_weight_name)
remaining_dict = {x: y for x, y in lora.items() if x not in loaded_keys}
return patch_dict, remaining_dict
def model_lora_keys_clip(model, key_map={}):
sdk = model.state_dict().keys()
text_model_lora_key = "lora_te_text_model_encoder_layers_{}_{}"
clip_l_present = False
for b in range(32): #TODO: clean up
for c in LORA_CLIP_MAP:
k = "clip_h.transformer.text_model.encoder.layers.{}.{}.weight".format(b, c)
if k in sdk:
lora_key = text_model_lora_key.format(b, LORA_CLIP_MAP[c])
key_map[lora_key] = k
lora_key = "lora_te1_text_model_encoder_layers_{}_{}".format(b, LORA_CLIP_MAP[c])
key_map[lora_key] = k
lora_key = "text_encoder.text_model.encoder.layers.{}.{}".format(b, c) #diffusers lora
key_map[lora_key] = k
k = "clip_l.transformer.text_model.encoder.layers.{}.{}.weight".format(b, c)
if k in sdk:
lora_key = text_model_lora_key.format(b, LORA_CLIP_MAP[c])
key_map[lora_key] = k
lora_key = "lora_te1_text_model_encoder_layers_{}_{}".format(b, LORA_CLIP_MAP[c]) #SDXL base
key_map[lora_key] = k
clip_l_present = True
lora_key = "text_encoder.text_model.encoder.layers.{}.{}".format(b, c) #diffusers lora
key_map[lora_key] = k
k = "clip_g.transformer.text_model.encoder.layers.{}.{}.weight".format(b, c)
if k in sdk:
if clip_l_present:
lora_key = "lora_te2_text_model_encoder_layers_{}_{}".format(b, LORA_CLIP_MAP[c]) #SDXL base
key_map[lora_key] = k
lora_key = "text_encoder_2.text_model.encoder.layers.{}.{}".format(b, c) #diffusers lora
key_map[lora_key] = k
else:
lora_key = "lora_te_text_model_encoder_layers_{}_{}".format(b, LORA_CLIP_MAP[c]) #TODO: test if this is correct for SDXL-Refiner
key_map[lora_key] = k
lora_key = "text_encoder.text_model.encoder.layers.{}.{}".format(b, c) #diffusers lora
key_map[lora_key] = k
lora_key = "lora_prior_te_text_model_encoder_layers_{}_{}".format(b, LORA_CLIP_MAP[c]) #cascade lora: TODO put lora key prefix in the model config
key_map[lora_key] = k
for k in sdk:
if k.endswith(".weight"):
if k.startswith("t5xxl.transformer."):#OneTrainer SD3 lora
l_key = k[len("t5xxl.transformer."):-len(".weight")]
lora_key = "lora_te3_{}".format(l_key.replace(".", "_"))
key_map[lora_key] = k
#####
lora_key = "lora_te2_{}".format(l_key.replace(".", "_"))#OneTrainer Flux lora, by Forge
key_map[lora_key] = k
#####
# elif k.startswith("hydit_clip.transformer.bert."): #HunyuanDiT Lora
# l_key = k[len("hydit_clip.transformer.bert."):-len(".weight")]
# lora_key = "lora_te1_{}".format(l_key.replace(".", "_"))
# key_map[lora_key] = k
k = "clip_g.transformer.text_projection.weight"
if k in sdk:
# key_map["lora_prior_te_text_projection"] = k #cascade lora?
key_map["text_encoder.text_projection"] = k #TODO: check if other lora have the text_projection too
key_map["lora_te2_text_projection"] = k #OneTrainer SD3 lora
k = "clip_l.transformer.text_projection.weight"
if k in sdk:
key_map["lora_te1_text_projection"] = k #OneTrainer SD3 lora, not necessary but omits warning
return sdk, key_map
def model_lora_keys_unet(model, key_map={}):
sd = model.state_dict()
sdk = sd.keys()
for k in sdk:
if k.startswith("diffusion_model."):
if k.endswith(".weight"):
key_lora = k[len("diffusion_model."):-len(".weight")].replace(".", "_")
key_map["lora_unet_{}".format(key_lora)] = k
key_map["{}".format(k[:-len(".weight")])] = k #generic lora format without any weird key names
else:
key_map["{}".format(k)] = k #generic lora format for not .weight without any weird key names
diffusers_keys = utils.unet_to_diffusers(model.diffusion_model.config)
for k in diffusers_keys:
if k.endswith(".weight"):
unet_key = "diffusion_model.{}".format(diffusers_keys[k])
key_lora = k[:-len(".weight")].replace(".", "_")
key_map["lora_unet_{}".format(key_lora)] = unet_key
key_map["lycoris_{}".format(key_lora)] = unet_key #simpletuner lycoris format
diffusers_lora_prefix = ["", "unet."]
for p in diffusers_lora_prefix:
diffusers_lora_key = "{}{}".format(p, k[:-len(".weight")].replace(".to_", ".processor.to_"))
if diffusers_lora_key.endswith(".to_out.0"):
diffusers_lora_key = diffusers_lora_key[:-2]
key_map[diffusers_lora_key] = unet_key
# if 'stable-diffusion-3' in model.config.huggingface_repo.lower(): #Diffusers lora SD3
# diffusers_keys = utils.mmdit_to_diffusers(model.diffusion_model.config, output_prefix="diffusion_model.")
# for k in diffusers_keys:
# if k.endswith(".weight"):
# to = diffusers_keys[k]
# key_lora = "transformer.{}".format(k[:-len(".weight")]) #regular diffusers sd3 lora format
# key_map[key_lora] = to
# key_lora = "base_model.model.{}".format(k[:-len(".weight")]) #format for flash-sd3 lora and others?
# key_map[key_lora] = to
# key_lora = "lora_transformer_{}".format(k[:-len(".weight")].replace(".", "_")) #OneTrainer lora
# key_map[key_lora] = to
#
# if isinstance(model, comfy.model_base.AuraFlow): #Diffusers lora AuraFlow
# diffusers_keys = utils.auraflow_to_diffusers(model.diffusion_model.config, output_prefix="diffusion_model.")
# for k in diffusers_keys:
# if k.endswith(".weight"):
# to = diffusers_keys[k]
# key_lora = "transformer.{}".format(k[:-len(".weight")]) #simpletrainer and probably regular diffusers lora format
# key_map[key_lora] = to
#
# if isinstance(model, comfy.model_base.HunyuanDiT):
# for k in sdk:
# if k.startswith("diffusion_model.") and k.endswith(".weight"):
# key_lora = k[len("diffusion_model."):-len(".weight")]
# key_map["base_model.model.{}".format(key_lora)] = k #official hunyuan lora format
if 'flux' in model.config.huggingface_repo.lower(): #Diffusers lora Flux
diffusers_keys = utils.flux_to_diffusers(model.diffusion_model.config, output_prefix="diffusion_model.")
for k in diffusers_keys:
if k.endswith(".weight"):
to = diffusers_keys[k]
key_map["transformer.{}".format(k[:-len(".weight")])] = to # simpletrainer and probably regular diffusers flux lora format
key_map["lycoris_{}".format(k[:-len(".weight")].replace(".", "_"))] = to # simpletrainer lycoris
key_map["lora_transformer_{}".format(k[:-len(".weight")].replace(".", "_"))] = to # onetrainer
return sdk, key_map

760
packages_3rdparty/comfyui_lora_collection/utils.py vendored Executable file
View File

@@ -0,0 +1,760 @@
"""
This file is part of ComfyUI.
Copyright (C) 2024 Comfy
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.
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.
You should have received a copy of the GNU General Public License
along with this program. If not, see <https://www.gnu.org/licenses/>.
"""
import torch
import math
import struct
import numpy as np
from PIL import Image
import itertools
def calculate_parameters(sd, prefix=""):
params = 0
for k in sd.keys():
if k.startswith(prefix):
w = sd[k]
params += w.nelement()
return params
def weight_dtype(sd, prefix=""):
dtypes = {}
for k in sd.keys():
if k.startswith(prefix):
w = sd[k]
dtypes[w.dtype] = dtypes.get(w.dtype, 0) + 1
if len(dtypes) == 0:
return None
return max(dtypes, key=dtypes.get)
def state_dict_key_replace(state_dict, keys_to_replace):
for x in keys_to_replace:
if x in state_dict:
state_dict[keys_to_replace[x]] = state_dict.pop(x)
return state_dict
def state_dict_prefix_replace(state_dict, replace_prefix, filter_keys=False):
if filter_keys:
out = {}
else:
out = state_dict
for rp in replace_prefix:
replace = list(map(lambda a: (a, "{}{}".format(replace_prefix[rp], a[len(rp):])), filter(lambda a: a.startswith(rp), state_dict.keys())))
for x in replace:
w = state_dict.pop(x[0])
out[x[1]] = w
return out
def transformers_convert(sd, prefix_from, prefix_to, number):
keys_to_replace = {
"{}positional_embedding": "{}embeddings.position_embedding.weight",
"{}token_embedding.weight": "{}embeddings.token_embedding.weight",
"{}ln_final.weight": "{}final_layer_norm.weight",
"{}ln_final.bias": "{}final_layer_norm.bias",
}
for k in keys_to_replace:
x = k.format(prefix_from)
if x in sd:
sd[keys_to_replace[k].format(prefix_to)] = sd.pop(x)
resblock_to_replace = {
"ln_1": "layer_norm1",
"ln_2": "layer_norm2",
"mlp.c_fc": "mlp.fc1",
"mlp.c_proj": "mlp.fc2",
"attn.out_proj": "self_attn.out_proj",
}
for resblock in range(number):
for x in resblock_to_replace:
for y in ["weight", "bias"]:
k = "{}transformer.resblocks.{}.{}.{}".format(prefix_from, resblock, x, y)
k_to = "{}encoder.layers.{}.{}.{}".format(prefix_to, resblock, resblock_to_replace[x], y)
if k in sd:
sd[k_to] = sd.pop(k)
for y in ["weight", "bias"]:
k_from = "{}transformer.resblocks.{}.attn.in_proj_{}".format(prefix_from, resblock, y)
if k_from in sd:
weights = sd.pop(k_from)
shape_from = weights.shape[0] // 3
for x in range(3):
p = ["self_attn.q_proj", "self_attn.k_proj", "self_attn.v_proj"]
k_to = "{}encoder.layers.{}.{}.{}".format(prefix_to, resblock, p[x], y)
sd[k_to] = weights[shape_from*x:shape_from*(x + 1)]
return sd
def clip_text_transformers_convert(sd, prefix_from, prefix_to):
sd = transformers_convert(sd, prefix_from, "{}text_model.".format(prefix_to), 32)
tp = "{}text_projection.weight".format(prefix_from)
if tp in sd:
sd["{}text_projection.weight".format(prefix_to)] = sd.pop(tp)
tp = "{}text_projection".format(prefix_from)
if tp in sd:
sd["{}text_projection.weight".format(prefix_to)] = sd.pop(tp).transpose(0, 1).contiguous()
return sd
UNET_MAP_ATTENTIONS = {
"proj_in.weight",
"proj_in.bias",
"proj_out.weight",
"proj_out.bias",
"norm.weight",
"norm.bias",
}
TRANSFORMER_BLOCKS = {
"norm1.weight",
"norm1.bias",
"norm2.weight",
"norm2.bias",
"norm3.weight",
"norm3.bias",
"attn1.to_q.weight",
"attn1.to_k.weight",
"attn1.to_v.weight",
"attn1.to_out.0.weight",
"attn1.to_out.0.bias",
"attn2.to_q.weight",
"attn2.to_k.weight",
"attn2.to_v.weight",
"attn2.to_out.0.weight",
"attn2.to_out.0.bias",
"ff.net.0.proj.weight",
"ff.net.0.proj.bias",
"ff.net.2.weight",
"ff.net.2.bias",
}
UNET_MAP_RESNET = {
"in_layers.2.weight": "conv1.weight",
"in_layers.2.bias": "conv1.bias",
"emb_layers.1.weight": "time_emb_proj.weight",
"emb_layers.1.bias": "time_emb_proj.bias",
"out_layers.3.weight": "conv2.weight",
"out_layers.3.bias": "conv2.bias",
"skip_connection.weight": "conv_shortcut.weight",
"skip_connection.bias": "conv_shortcut.bias",
"in_layers.0.weight": "norm1.weight",
"in_layers.0.bias": "norm1.bias",
"out_layers.0.weight": "norm2.weight",
"out_layers.0.bias": "norm2.bias",
}
UNET_MAP_BASIC = {
("label_emb.0.0.weight", "class_embedding.linear_1.weight"),
("label_emb.0.0.bias", "class_embedding.linear_1.bias"),
("label_emb.0.2.weight", "class_embedding.linear_2.weight"),
("label_emb.0.2.bias", "class_embedding.linear_2.bias"),
("label_emb.0.0.weight", "add_embedding.linear_1.weight"),
("label_emb.0.0.bias", "add_embedding.linear_1.bias"),
("label_emb.0.2.weight", "add_embedding.linear_2.weight"),
("label_emb.0.2.bias", "add_embedding.linear_2.bias"),
("input_blocks.0.0.weight", "conv_in.weight"),
("input_blocks.0.0.bias", "conv_in.bias"),
("out.0.weight", "conv_norm_out.weight"),
("out.0.bias", "conv_norm_out.bias"),
("out.2.weight", "conv_out.weight"),
("out.2.bias", "conv_out.bias"),
("time_embed.0.weight", "time_embedding.linear_1.weight"),
("time_embed.0.bias", "time_embedding.linear_1.bias"),
("time_embed.2.weight", "time_embedding.linear_2.weight"),
("time_embed.2.bias", "time_embedding.linear_2.bias")
}
def unet_to_diffusers(unet_config):
if "num_res_blocks" not in unet_config:
return {}
num_res_blocks = unet_config["num_res_blocks"]
channel_mult = unet_config["channel_mult"]
transformer_depth = unet_config["transformer_depth"][:]
transformer_depth_output = unet_config["transformer_depth_output"][:]
num_blocks = len(channel_mult)
transformers_mid = unet_config.get("transformer_depth_middle", None)
diffusers_unet_map = {}
for x in range(num_blocks):
n = 1 + (num_res_blocks[x] + 1) * x
for i in range(num_res_blocks[x]):
for b in UNET_MAP_RESNET:
diffusers_unet_map["down_blocks.{}.resnets.{}.{}".format(x, i, UNET_MAP_RESNET[b])] = "input_blocks.{}.0.{}".format(n, b)
num_transformers = transformer_depth.pop(0)
if num_transformers > 0:
for b in UNET_MAP_ATTENTIONS:
diffusers_unet_map["down_blocks.{}.attentions.{}.{}".format(x, i, b)] = "input_blocks.{}.1.{}".format(n, b)
for t in range(num_transformers):
for b in TRANSFORMER_BLOCKS:
diffusers_unet_map["down_blocks.{}.attentions.{}.transformer_blocks.{}.{}".format(x, i, t, b)] = "input_blocks.{}.1.transformer_blocks.{}.{}".format(n, t, b)
n += 1
for k in ["weight", "bias"]:
diffusers_unet_map["down_blocks.{}.downsamplers.0.conv.{}".format(x, k)] = "input_blocks.{}.0.op.{}".format(n, k)
i = 0
for b in UNET_MAP_ATTENTIONS:
diffusers_unet_map["mid_block.attentions.{}.{}".format(i, b)] = "middle_block.1.{}".format(b)
for t in range(transformers_mid):
for b in TRANSFORMER_BLOCKS:
diffusers_unet_map["mid_block.attentions.{}.transformer_blocks.{}.{}".format(i, t, b)] = "middle_block.1.transformer_blocks.{}.{}".format(t, b)
for i, n in enumerate([0, 2]):
for b in UNET_MAP_RESNET:
diffusers_unet_map["mid_block.resnets.{}.{}".format(i, UNET_MAP_RESNET[b])] = "middle_block.{}.{}".format(n, b)
num_res_blocks = list(reversed(num_res_blocks))
for x in range(num_blocks):
n = (num_res_blocks[x] + 1) * x
l = num_res_blocks[x] + 1
for i in range(l):
c = 0
for b in UNET_MAP_RESNET:
diffusers_unet_map["up_blocks.{}.resnets.{}.{}".format(x, i, UNET_MAP_RESNET[b])] = "output_blocks.{}.0.{}".format(n, b)
c += 1
num_transformers = transformer_depth_output.pop()
if num_transformers > 0:
c += 1
for b in UNET_MAP_ATTENTIONS:
diffusers_unet_map["up_blocks.{}.attentions.{}.{}".format(x, i, b)] = "output_blocks.{}.1.{}".format(n, b)
for t in range(num_transformers):
for b in TRANSFORMER_BLOCKS:
diffusers_unet_map["up_blocks.{}.attentions.{}.transformer_blocks.{}.{}".format(x, i, t, b)] = "output_blocks.{}.1.transformer_blocks.{}.{}".format(n, t, b)
if i == l - 1:
for k in ["weight", "bias"]:
diffusers_unet_map["up_blocks.{}.upsamplers.0.conv.{}".format(x, k)] = "output_blocks.{}.{}.conv.{}".format(n, c, k)
n += 1
for k in UNET_MAP_BASIC:
diffusers_unet_map[k[1]] = k[0]
return diffusers_unet_map
def swap_scale_shift(weight):
shift, scale = weight.chunk(2, dim=0)
new_weight = torch.cat([scale, shift], dim=0)
return new_weight
MMDIT_MAP_BASIC = {
("context_embedder.bias", "context_embedder.bias"),
("context_embedder.weight", "context_embedder.weight"),
("t_embedder.mlp.0.bias", "time_text_embed.timestep_embedder.linear_1.bias"),
("t_embedder.mlp.0.weight", "time_text_embed.timestep_embedder.linear_1.weight"),
("t_embedder.mlp.2.bias", "time_text_embed.timestep_embedder.linear_2.bias"),
("t_embedder.mlp.2.weight", "time_text_embed.timestep_embedder.linear_2.weight"),
("x_embedder.proj.bias", "pos_embed.proj.bias"),
("x_embedder.proj.weight", "pos_embed.proj.weight"),
("y_embedder.mlp.0.bias", "time_text_embed.text_embedder.linear_1.bias"),
("y_embedder.mlp.0.weight", "time_text_embed.text_embedder.linear_1.weight"),
("y_embedder.mlp.2.bias", "time_text_embed.text_embedder.linear_2.bias"),
("y_embedder.mlp.2.weight", "time_text_embed.text_embedder.linear_2.weight"),
("pos_embed", "pos_embed.pos_embed"),
("final_layer.adaLN_modulation.1.bias", "norm_out.linear.bias", swap_scale_shift),
("final_layer.adaLN_modulation.1.weight", "norm_out.linear.weight", swap_scale_shift),
("final_layer.linear.bias", "proj_out.bias"),
("final_layer.linear.weight", "proj_out.weight"),
}
MMDIT_MAP_BLOCK = {
("context_block.adaLN_modulation.1.bias", "norm1_context.linear.bias"),
("context_block.adaLN_modulation.1.weight", "norm1_context.linear.weight"),
("context_block.attn.proj.bias", "attn.to_add_out.bias"),
("context_block.attn.proj.weight", "attn.to_add_out.weight"),
("context_block.mlp.fc1.bias", "ff_context.net.0.proj.bias"),
("context_block.mlp.fc1.weight", "ff_context.net.0.proj.weight"),
("context_block.mlp.fc2.bias", "ff_context.net.2.bias"),
("context_block.mlp.fc2.weight", "ff_context.net.2.weight"),
("x_block.adaLN_modulation.1.bias", "norm1.linear.bias"),
("x_block.adaLN_modulation.1.weight", "norm1.linear.weight"),
("x_block.attn.proj.bias", "attn.to_out.0.bias"),
("x_block.attn.proj.weight", "attn.to_out.0.weight"),
("x_block.mlp.fc1.bias", "ff.net.0.proj.bias"),
("x_block.mlp.fc1.weight", "ff.net.0.proj.weight"),
("x_block.mlp.fc2.bias", "ff.net.2.bias"),
("x_block.mlp.fc2.weight", "ff.net.2.weight"),
}
def mmdit_to_diffusers(mmdit_config, output_prefix=""):
key_map = {}
depth = mmdit_config.get("depth", 0)
num_blocks = mmdit_config.get("num_blocks", depth)
for i in range(num_blocks):
block_from = "transformer_blocks.{}".format(i)
block_to = "{}joint_blocks.{}".format(output_prefix, i)
offset = depth * 64
for end in ("weight", "bias"):
k = "{}.attn.".format(block_from)
qkv = "{}.x_block.attn.qkv.{}".format(block_to, end)
key_map["{}to_q.{}".format(k, end)] = (qkv, (0, 0, offset))
key_map["{}to_k.{}".format(k, end)] = (qkv, (0, offset, offset))
key_map["{}to_v.{}".format(k, end)] = (qkv, (0, offset * 2, offset))
qkv = "{}.context_block.attn.qkv.{}".format(block_to, end)
key_map["{}add_q_proj.{}".format(k, end)] = (qkv, (0, 0, offset))
key_map["{}add_k_proj.{}".format(k, end)] = (qkv, (0, offset, offset))
key_map["{}add_v_proj.{}".format(k, end)] = (qkv, (0, offset * 2, offset))
for k in MMDIT_MAP_BLOCK:
key_map["{}.{}".format(block_from, k[1])] = "{}.{}".format(block_to, k[0])
map_basic = MMDIT_MAP_BASIC.copy()
map_basic.add(("joint_blocks.{}.context_block.adaLN_modulation.1.bias".format(depth - 1), "transformer_blocks.{}.norm1_context.linear.bias".format(depth - 1), swap_scale_shift))
map_basic.add(("joint_blocks.{}.context_block.adaLN_modulation.1.weight".format(depth - 1), "transformer_blocks.{}.norm1_context.linear.weight".format(depth - 1), swap_scale_shift))
for k in map_basic:
if len(k) > 2:
key_map[k[1]] = ("{}{}".format(output_prefix, k[0]), None, k[2])
else:
key_map[k[1]] = "{}{}".format(output_prefix, k[0])
return key_map
def auraflow_to_diffusers(mmdit_config, output_prefix=""):
n_double_layers = mmdit_config.get("n_double_layers", 0)
n_layers = mmdit_config.get("n_layers", 0)
key_map = {}
for i in range(n_layers):
if i < n_double_layers:
index = i
prefix_from = "joint_transformer_blocks"
prefix_to = "{}double_layers".format(output_prefix)
block_map = {
"attn.to_q.weight": "attn.w2q.weight",
"attn.to_k.weight": "attn.w2k.weight",
"attn.to_v.weight": "attn.w2v.weight",
"attn.to_out.0.weight": "attn.w2o.weight",
"attn.add_q_proj.weight": "attn.w1q.weight",
"attn.add_k_proj.weight": "attn.w1k.weight",
"attn.add_v_proj.weight": "attn.w1v.weight",
"attn.to_add_out.weight": "attn.w1o.weight",
"ff.linear_1.weight": "mlpX.c_fc1.weight",
"ff.linear_2.weight": "mlpX.c_fc2.weight",
"ff.out_projection.weight": "mlpX.c_proj.weight",
"ff_context.linear_1.weight": "mlpC.c_fc1.weight",
"ff_context.linear_2.weight": "mlpC.c_fc2.weight",
"ff_context.out_projection.weight": "mlpC.c_proj.weight",
"norm1.linear.weight": "modX.1.weight",
"norm1_context.linear.weight": "modC.1.weight",
}
else:
index = i - n_double_layers
prefix_from = "single_transformer_blocks"
prefix_to = "{}single_layers".format(output_prefix)
block_map = {
"attn.to_q.weight": "attn.w1q.weight",
"attn.to_k.weight": "attn.w1k.weight",
"attn.to_v.weight": "attn.w1v.weight",
"attn.to_out.0.weight": "attn.w1o.weight",
"norm1.linear.weight": "modCX.1.weight",
"ff.linear_1.weight": "mlp.c_fc1.weight",
"ff.linear_2.weight": "mlp.c_fc2.weight",
"ff.out_projection.weight": "mlp.c_proj.weight"
}
for k in block_map:
key_map["{}.{}.{}".format(prefix_from, index, k)] = "{}.{}.{}".format(prefix_to, index, block_map[k])
MAP_BASIC = {
("positional_encoding", "pos_embed.pos_embed"),
("register_tokens", "register_tokens"),
("t_embedder.mlp.0.weight", "time_step_proj.linear_1.weight"),
("t_embedder.mlp.0.bias", "time_step_proj.linear_1.bias"),
("t_embedder.mlp.2.weight", "time_step_proj.linear_2.weight"),
("t_embedder.mlp.2.bias", "time_step_proj.linear_2.bias"),
("cond_seq_linear.weight", "context_embedder.weight"),
("init_x_linear.weight", "pos_embed.proj.weight"),
("init_x_linear.bias", "pos_embed.proj.bias"),
("final_linear.weight", "proj_out.weight"),
("modF.1.weight", "norm_out.linear.weight", swap_scale_shift),
}
for k in MAP_BASIC:
if len(k) > 2:
key_map[k[1]] = ("{}{}".format(output_prefix, k[0]), None, k[2])
else:
key_map[k[1]] = "{}{}".format(output_prefix, k[0])
return key_map
def flux_to_diffusers(mmdit_config, output_prefix=""):
n_double_layers = mmdit_config.get("depth", 0)
n_single_layers = mmdit_config.get("depth_single_blocks", 0)
hidden_size = mmdit_config.get("hidden_size", 0)
key_map = {}
for index in range(n_double_layers):
prefix_from = "transformer_blocks.{}".format(index)
prefix_to = "{}double_blocks.{}".format(output_prefix, index)
for end in ("weight", "bias"):
k = "{}.attn.".format(prefix_from)
qkv = "{}.img_attn.qkv.{}".format(prefix_to, end)
key_map["{}to_q.{}".format(k, end)] = (qkv, (0, 0, hidden_size))
key_map["{}to_k.{}".format(k, end)] = (qkv, (0, hidden_size, hidden_size))
key_map["{}to_v.{}".format(k, end)] = (qkv, (0, hidden_size * 2, hidden_size))
k = "{}.attn.".format(prefix_from)
qkv = "{}.txt_attn.qkv.{}".format(prefix_to, end)
key_map["{}add_q_proj.{}".format(k, end)] = (qkv, (0, 0, hidden_size))
key_map["{}add_k_proj.{}".format(k, end)] = (qkv, (0, hidden_size, hidden_size))
key_map["{}add_v_proj.{}".format(k, end)] = (qkv, (0, hidden_size * 2, hidden_size))
block_map = {
"attn.to_out.0.weight": "img_attn.proj.weight",
"attn.to_out.0.bias": "img_attn.proj.bias",
"norm1.linear.weight": "img_mod.lin.weight",
"norm1.linear.bias": "img_mod.lin.bias",
"norm1_context.linear.weight": "txt_mod.lin.weight",
"norm1_context.linear.bias": "txt_mod.lin.bias",
"attn.to_add_out.weight": "txt_attn.proj.weight",
"attn.to_add_out.bias": "txt_attn.proj.bias",
"ff.net.0.proj.weight": "img_mlp.0.weight",
"ff.net.0.proj.bias": "img_mlp.0.bias",
"ff.net.2.weight": "img_mlp.2.weight",
"ff.net.2.bias": "img_mlp.2.bias",
"ff_context.net.0.proj.weight": "txt_mlp.0.weight",
"ff_context.net.0.proj.bias": "txt_mlp.0.bias",
"ff_context.net.2.weight": "txt_mlp.2.weight",
"ff_context.net.2.bias": "txt_mlp.2.bias",
"attn.norm_q.weight": "img_attn.norm.query_norm.scale",
"attn.norm_k.weight": "img_attn.norm.key_norm.scale",
"attn.norm_added_q.weight": "txt_attn.norm.query_norm.scale",
"attn.norm_added_k.weight": "txt_attn.norm.key_norm.scale",
}
for k in block_map:
key_map["{}.{}".format(prefix_from, k)] = "{}.{}".format(prefix_to, block_map[k])
for index in range(n_single_layers):
prefix_from = "single_transformer_blocks.{}".format(index)
prefix_to = "{}single_blocks.{}".format(output_prefix, index)
for end in ("weight", "bias"):
k = "{}.attn.".format(prefix_from)
qkv = "{}.linear1.{}".format(prefix_to, end)
key_map["{}to_q.{}".format(k, end)] = (qkv, (0, 0, hidden_size))
key_map["{}to_k.{}".format(k, end)] = (qkv, (0, hidden_size, hidden_size))
key_map["{}to_v.{}".format(k, end)] = (qkv, (0, hidden_size * 2, hidden_size))
key_map["{}.proj_mlp.{}".format(prefix_from, end)] = (qkv, (0, hidden_size * 3, hidden_size * 4))
block_map = {
"norm.linear.weight": "modulation.lin.weight",
"norm.linear.bias": "modulation.lin.bias",
"proj_out.weight": "linear2.weight",
"proj_out.bias": "linear2.bias",
"attn.norm_q.weight": "norm.query_norm.scale",
"attn.norm_k.weight": "norm.key_norm.scale",
}
for k in block_map:
key_map["{}.{}".format(prefix_from, k)] = "{}.{}".format(prefix_to, block_map[k])
MAP_BASIC = {
("final_layer.linear.bias", "proj_out.bias"),
("final_layer.linear.weight", "proj_out.weight"),
("img_in.bias", "x_embedder.bias"),
("img_in.weight", "x_embedder.weight"),
("time_in.in_layer.bias", "time_text_embed.timestep_embedder.linear_1.bias"),
("time_in.in_layer.weight", "time_text_embed.timestep_embedder.linear_1.weight"),
("time_in.out_layer.bias", "time_text_embed.timestep_embedder.linear_2.bias"),
("time_in.out_layer.weight", "time_text_embed.timestep_embedder.linear_2.weight"),
("txt_in.bias", "context_embedder.bias"),
("txt_in.weight", "context_embedder.weight"),
("vector_in.in_layer.bias", "time_text_embed.text_embedder.linear_1.bias"),
("vector_in.in_layer.weight", "time_text_embed.text_embedder.linear_1.weight"),
("vector_in.out_layer.bias", "time_text_embed.text_embedder.linear_2.bias"),
("vector_in.out_layer.weight", "time_text_embed.text_embedder.linear_2.weight"),
("guidance_in.in_layer.bias", "time_text_embed.guidance_embedder.linear_1.bias"),
("guidance_in.in_layer.weight", "time_text_embed.guidance_embedder.linear_1.weight"),
("guidance_in.out_layer.bias", "time_text_embed.guidance_embedder.linear_2.bias"),
("guidance_in.out_layer.weight", "time_text_embed.guidance_embedder.linear_2.weight"),
("final_layer.adaLN_modulation.1.bias", "norm_out.linear.bias", swap_scale_shift),
("final_layer.adaLN_modulation.1.weight", "norm_out.linear.weight", swap_scale_shift),
}
for k in MAP_BASIC:
if len(k) > 2:
key_map[k[1]] = ("{}{}".format(output_prefix, k[0]), None, k[2])
else:
key_map[k[1]] = "{}{}".format(output_prefix, k[0])
return key_map
def repeat_to_batch_size(tensor, batch_size, dim=0):
if tensor.shape[dim] > batch_size:
return tensor.narrow(dim, 0, batch_size)
elif tensor.shape[dim] < batch_size:
return tensor.repeat(dim * [1] + [math.ceil(batch_size / tensor.shape[dim])] + [1] * (len(tensor.shape) - 1 - dim)).narrow(dim, 0, batch_size)
return tensor
def resize_to_batch_size(tensor, batch_size):
in_batch_size = tensor.shape[0]
if in_batch_size == batch_size:
return tensor
if batch_size <= 1:
return tensor[:batch_size]
output = torch.empty([batch_size] + list(tensor.shape)[1:], dtype=tensor.dtype, device=tensor.device)
if batch_size < in_batch_size:
scale = (in_batch_size - 1) / (batch_size - 1)
for i in range(batch_size):
output[i] = tensor[min(round(i * scale), in_batch_size - 1)]
else:
scale = in_batch_size / batch_size
for i in range(batch_size):
output[i] = tensor[min(math.floor((i + 0.5) * scale), in_batch_size - 1)]
return output
def convert_sd_to(state_dict, dtype):
keys = list(state_dict.keys())
for k in keys:
state_dict[k] = state_dict[k].to(dtype)
return state_dict
def safetensors_header(safetensors_path, max_size=100*1024*1024):
with open(safetensors_path, "rb") as f:
header = f.read(8)
length_of_header = struct.unpack('<Q', header)[0]
if length_of_header > max_size:
return None
return f.read(length_of_header)
def set_attr(obj, attr, value):
attrs = attr.split(".")
for name in attrs[:-1]:
obj = getattr(obj, name)
prev = getattr(obj, attrs[-1])
setattr(obj, attrs[-1], value)
return prev
def set_attr_param(obj, attr, value):
return set_attr(obj, attr, torch.nn.Parameter(value, requires_grad=False))
def copy_to_param(obj, attr, value):
# inplace update tensor instead of replacing it
attrs = attr.split(".")
for name in attrs[:-1]:
obj = getattr(obj, name)
prev = getattr(obj, attrs[-1])
prev.data.copy_(value)
def get_attr(obj, attr):
attrs = attr.split(".")
for name in attrs:
obj = getattr(obj, name)
return obj
def bislerp(samples, width, height):
def slerp(b1, b2, r):
'''slerps batches b1, b2 according to ratio r, batches should be flat e.g. NxC'''
c = b1.shape[-1]
#norms
b1_norms = torch.norm(b1, dim=-1, keepdim=True)
b2_norms = torch.norm(b2, dim=-1, keepdim=True)
#normalize
b1_normalized = b1 / b1_norms
b2_normalized = b2 / b2_norms
#zero when norms are zero
b1_normalized[b1_norms.expand(-1,c) == 0.0] = 0.0
b2_normalized[b2_norms.expand(-1,c) == 0.0] = 0.0
#slerp
dot = (b1_normalized*b2_normalized).sum(1)
omega = torch.acos(dot)
so = torch.sin(omega)
#technically not mathematically correct, but more pleasing?
res = (torch.sin((1.0-r.squeeze(1))*omega)/so).unsqueeze(1)*b1_normalized + (torch.sin(r.squeeze(1)*omega)/so).unsqueeze(1) * b2_normalized
res *= (b1_norms * (1.0-r) + b2_norms * r).expand(-1,c)
#edge cases for same or polar opposites
res[dot > 1 - 1e-5] = b1[dot > 1 - 1e-5]
res[dot < 1e-5 - 1] = (b1 * (1.0-r) + b2 * r)[dot < 1e-5 - 1]
return res
def generate_bilinear_data(length_old, length_new, device):
coords_1 = torch.arange(length_old, dtype=torch.float32, device=device).reshape((1,1,1,-1))
coords_1 = torch.nn.functional.interpolate(coords_1, size=(1, length_new), mode="bilinear")
ratios = coords_1 - coords_1.floor()
coords_1 = coords_1.to(torch.int64)
coords_2 = torch.arange(length_old, dtype=torch.float32, device=device).reshape((1,1,1,-1)) + 1
coords_2[:,:,:,-1] -= 1
coords_2 = torch.nn.functional.interpolate(coords_2, size=(1, length_new), mode="bilinear")
coords_2 = coords_2.to(torch.int64)
return ratios, coords_1, coords_2
orig_dtype = samples.dtype
samples = samples.float()
n,c,h,w = samples.shape
h_new, w_new = (height, width)
#linear w
ratios, coords_1, coords_2 = generate_bilinear_data(w, w_new, samples.device)
coords_1 = coords_1.expand((n, c, h, -1))
coords_2 = coords_2.expand((n, c, h, -1))
ratios = ratios.expand((n, 1, h, -1))
pass_1 = samples.gather(-1,coords_1).movedim(1, -1).reshape((-1,c))
pass_2 = samples.gather(-1,coords_2).movedim(1, -1).reshape((-1,c))
ratios = ratios.movedim(1, -1).reshape((-1,1))
result = slerp(pass_1, pass_2, ratios)
result = result.reshape(n, h, w_new, c).movedim(-1, 1)
#linear h
ratios, coords_1, coords_2 = generate_bilinear_data(h, h_new, samples.device)
coords_1 = coords_1.reshape((1,1,-1,1)).expand((n, c, -1, w_new))
coords_2 = coords_2.reshape((1,1,-1,1)).expand((n, c, -1, w_new))
ratios = ratios.reshape((1,1,-1,1)).expand((n, 1, -1, w_new))
pass_1 = result.gather(-2,coords_1).movedim(1, -1).reshape((-1,c))
pass_2 = result.gather(-2,coords_2).movedim(1, -1).reshape((-1,c))
ratios = ratios.movedim(1, -1).reshape((-1,1))
result = slerp(pass_1, pass_2, ratios)
result = result.reshape(n, h_new, w_new, c).movedim(-1, 1)
return result.to(orig_dtype)
def lanczos(samples, width, height):
images = [Image.fromarray(np.clip(255. * image.movedim(0, -1).cpu().numpy(), 0, 255).astype(np.uint8)) for image in samples]
images = [image.resize((width, height), resample=Image.Resampling.LANCZOS) for image in images]
images = [torch.from_numpy(np.array(image).astype(np.float32) / 255.0).movedim(-1, 0) for image in images]
result = torch.stack(images)
return result.to(samples.device, samples.dtype)
def common_upscale(samples, width, height, upscale_method, crop):
if crop == "center":
old_width = samples.shape[3]
old_height = samples.shape[2]
old_aspect = old_width / old_height
new_aspect = width / height
x = 0
y = 0
if old_aspect > new_aspect:
x = round((old_width - old_width * (new_aspect / old_aspect)) / 2)
elif old_aspect < new_aspect:
y = round((old_height - old_height * (old_aspect / new_aspect)) / 2)
s = samples[:,:,y:old_height-y,x:old_width-x]
else:
s = samples
if upscale_method == "bislerp":
return bislerp(s, width, height)
elif upscale_method == "lanczos":
return lanczos(s, width, height)
else:
return torch.nn.functional.interpolate(s, size=(height, width), mode=upscale_method)
def get_tiled_scale_steps(width, height, tile_x, tile_y, overlap):
return math.ceil((height / (tile_y - overlap))) * math.ceil((width / (tile_x - overlap)))
@torch.inference_mode()
def tiled_scale_multidim(samples, function, tile=(64, 64), overlap = 8, upscale_amount = 4, out_channels = 3, output_device="cpu", pbar = None):
dims = len(tile)
output = torch.empty([samples.shape[0], out_channels] + list(map(lambda a: round(a * upscale_amount), samples.shape[2:])), device=output_device)
for b in range(samples.shape[0]):
s = samples[b:b+1]
out = torch.zeros([s.shape[0], out_channels] + list(map(lambda a: round(a * upscale_amount), s.shape[2:])), device=output_device)
out_div = torch.zeros([s.shape[0], out_channels] + list(map(lambda a: round(a * upscale_amount), s.shape[2:])), device=output_device)
for it in itertools.product(*map(lambda a: range(0, a[0], a[1] - overlap), zip(s.shape[2:], tile))):
s_in = s
upscaled = []
for d in range(dims):
pos = max(0, min(s.shape[d + 2] - overlap, it[d]))
l = min(tile[d], s.shape[d + 2] - pos)
s_in = s_in.narrow(d + 2, pos, l)
upscaled.append(round(pos * upscale_amount))
ps = function(s_in).to(output_device)
mask = torch.ones_like(ps)
feather = round(overlap * upscale_amount)
for t in range(feather):
for d in range(2, dims + 2):
m = mask.narrow(d, t, 1)
m *= ((1.0/feather) * (t + 1))
m = mask.narrow(d, mask.shape[d] -1 -t, 1)
m *= ((1.0/feather) * (t + 1))
o = out
o_d = out_div
for d in range(dims):
o = o.narrow(d + 2, upscaled[d], mask.shape[d + 2])
o_d = o_d.narrow(d + 2, upscaled[d], mask.shape[d + 2])
o += ps * mask
o_d += mask
if pbar is not None:
pbar.update(1)
output[b:b+1] = out/out_div
return output
def tiled_scale(samples, function, tile_x=64, tile_y=64, overlap = 8, upscale_amount = 4, out_channels = 3, output_device="cpu", pbar = None):
return tiled_scale_multidim(samples, function, (tile_y, tile_x), overlap, upscale_amount, out_channels, output_device, pbar)
PROGRESS_BAR_ENABLED = True
def set_progress_bar_enabled(enabled):
global PROGRESS_BAR_ENABLED
PROGRESS_BAR_ENABLED = enabled
PROGRESS_BAR_HOOK = None
def set_progress_bar_global_hook(function):
global PROGRESS_BAR_HOOK
PROGRESS_BAR_HOOK = function
class ProgressBar:
def __init__(self, total):
global PROGRESS_BAR_HOOK
self.total = total
self.current = 0
self.hook = PROGRESS_BAR_HOOK
def update_absolute(self, value, total=None, preview=None):
if total is not None:
self.total = total
if value > self.total:
value = self.total
self.current = value
if self.hook is not None:
self.hook(self.current, self.total, preview)
def update(self, value):
self.update_absolute(self.current + value)

21
packages_3rdparty/gguf/LICENSE vendored Executable file
View File

@@ -0,0 +1,21 @@
MIT License
Copyright (c) 2023 Georgi Gerganov
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.

2
packages_3rdparty/gguf/README.md vendored Executable file
View File

@@ -0,0 +1,2 @@
This is Forge's implementation of GGUF - the difference is that it supports pytorch quant/dequant
Codes are based on LLama.cpp's GGUF - the difference is that it supports quant

9
packages_3rdparty/gguf/__init__.py vendored Executable file
View File

@@ -0,0 +1,9 @@
from .constants import *
from .lazy import *
from .gguf_reader import *
from .gguf_writer import *
from .quants import *
from .tensor_mapping import *
from .vocab import *
from .utility import *
from .metadata import *

1357
packages_3rdparty/gguf/constants.py vendored Executable file
View File

File diff suppressed because it is too large Load Diff

317
packages_3rdparty/gguf/gguf_reader.py vendored Executable file
View File

@@ -0,0 +1,317 @@
#
# GGUF file reading/modification support. For API usage information,
# please see the files scripts/ for some fairly simple examples.
#
from __future__ import annotations
import logging
import os
from collections import OrderedDict
from typing import Any, Literal, NamedTuple, TypeVar, Union
import numpy as np
import numpy.typing as npt
from .quants import quant_shape_to_byte_shape
if __name__ == "__main__":
import sys
from pathlib import Path
# Allow running file in package as a script.
sys.path.insert(0, str(Path(__file__).parent.parent))
from gguf.constants import (
GGML_QUANT_SIZES,
GGUF_DEFAULT_ALIGNMENT,
GGUF_MAGIC,
GGUF_VERSION,
GGMLQuantizationType,
GGUFValueType,
)
logger = logging.getLogger(__name__)
READER_SUPPORTED_VERSIONS = [2, GGUF_VERSION]
class ReaderField(NamedTuple):
# Offset to start of this field.
offset: int
# Name of the field (not necessarily from file data).
name: str
# Data parts. Some types have multiple components, such as strings
# that consist of a length followed by the string data.
parts: list[npt.NDArray[Any]] = []
# Indexes into parts that we can call the actual data. For example
# an array of strings will be populated with indexes to the actual
# string data.
data: list[int] = [-1]
types: list[GGUFValueType] = []
class ReaderTensor(NamedTuple):
name: str
tensor_type: GGMLQuantizationType
shape: npt.NDArray[np.uint32]
n_elements: int
n_bytes: int
data_offset: int
data: npt.NDArray[Any]
field: ReaderField
class GGUFReader:
# I - same as host, S - swapped
byte_order: Literal['I', 'S'] = 'I'
alignment: int = GGUF_DEFAULT_ALIGNMENT
data_offset: int
# Note: Internal helper, API may change.
gguf_scalar_to_np: dict[GGUFValueType, type[np.generic]] = {
GGUFValueType.UINT8: np.uint8,
GGUFValueType.INT8: np.int8,
GGUFValueType.UINT16: np.uint16,
GGUFValueType.INT16: np.int16,
GGUFValueType.UINT32: np.uint32,
GGUFValueType.INT32: np.int32,
GGUFValueType.FLOAT32: np.float32,
GGUFValueType.UINT64: np.uint64,
GGUFValueType.INT64: np.int64,
GGUFValueType.FLOAT64: np.float64,
GGUFValueType.BOOL: np.bool_,
}
def __init__(self, path: os.PathLike[str] | str, mode: Literal['r', 'r+', 'c'] = 'r'):
self.data = np.memmap(path, mode = mode)
offs = 0
# Check for GGUF magic
if self._get(offs, np.uint32, override_order = '<')[0] != GGUF_MAGIC:
raise ValueError('GGUF magic invalid')
offs += 4
# Check GGUF version
temp_version = self._get(offs, np.uint32)
if temp_version[0] & 65535 == 0:
# If we get 0 here that means it's (probably) a GGUF file created for
# the opposite byte order of the machine this script is running on.
self.byte_order = 'S'
temp_version = temp_version.newbyteorder(self.byte_order)
version = temp_version[0]
if version not in READER_SUPPORTED_VERSIONS:
raise ValueError(f'Sorry, file appears to be version {version} which we cannot handle')
self.fields: OrderedDict[str, ReaderField] = OrderedDict()
self.tensors: list[ReaderTensor] = []
offs += self._push_field(ReaderField(offs, 'GGUF.version', [temp_version], [0], [GGUFValueType.UINT32]))
# Check tensor count and kv count
temp_counts = self._get(offs, np.uint64, 2)
offs += self._push_field(ReaderField(offs, 'GGUF.tensor_count', [temp_counts[:1]], [0], [GGUFValueType.UINT64]))
offs += self._push_field(ReaderField(offs, 'GGUF.kv_count', [temp_counts[1:]], [0], [GGUFValueType.UINT64]))
tensor_count, kv_count = temp_counts
offs = self._build_fields(offs, kv_count)
# Build Tensor Info Fields
offs, tensors_fields = self._build_tensor_info(offs, tensor_count)
new_align = self.fields.get('general.alignment')
if new_align is not None:
if new_align.types != [GGUFValueType.UINT32]:
raise ValueError('Bad type for general.alignment field')
self.alignment = new_align.parts[-1][0]
padding = offs % self.alignment
if padding != 0:
offs += self.alignment - padding
self.data_offset = offs
self._build_tensors(offs, tensors_fields)
_DT = TypeVar('_DT', bound = npt.DTypeLike)
# Fetch a key/value metadata field by key.
def get_field(self, key: str) -> Union[ReaderField, None]:
return self.fields.get(key, None)
# Fetch a tensor from the list by index.
def get_tensor(self, idx: int) -> ReaderTensor:
return self.tensors[idx]
def _get(
self, offset: int, dtype: npt.DTypeLike, count: int = 1, override_order: None | Literal['I', 'S', '<'] = None,
) -> npt.NDArray[Any]:
count = int(count)
itemsize = int(np.empty([], dtype = dtype).itemsize)
end_offs = offset + itemsize * count
return (
self.data[offset:end_offs]
.view(dtype = dtype)[:count]
.newbyteorder(override_order or self.byte_order)
)
def _push_field(self, field: ReaderField, skip_sum: bool = False) -> int:
if field.name in self.fields:
# TODO: add option to generate error on duplicate keys
# raise KeyError(f'Duplicate {field.name} already in list at offset {field.offset}')
logger.warning(f'Duplicate key {field.name} at offset {field.offset}')
self.fields[field.name + '_{}'.format(field.offset)] = field
else:
self.fields[field.name] = field
return 0 if skip_sum else sum(int(part.nbytes) for part in field.parts)
def _get_str(self, offset: int) -> tuple[npt.NDArray[np.uint64], npt.NDArray[np.uint8]]:
slen = self._get(offset, np.uint64)
return slen, self._get(offset + 8, np.uint8, slen[0])
def _get_field_parts(
self, orig_offs: int, raw_type: int,
) -> tuple[int, list[npt.NDArray[Any]], list[int], list[GGUFValueType]]:
offs = orig_offs
types: list[GGUFValueType] = []
gtype = GGUFValueType(raw_type)
types.append(gtype)
# Handle strings.
if gtype == GGUFValueType.STRING:
sparts: list[npt.NDArray[Any]] = list(self._get_str(offs))
size = sum(int(part.nbytes) for part in sparts)
return size, sparts, [1], types
# Check if it's a simple scalar type.
nptype = self.gguf_scalar_to_np.get(gtype)
if nptype is not None:
val = self._get(offs, nptype)
return int(val.nbytes), [val], [0], types
# Handle arrays.
if gtype == GGUFValueType.ARRAY:
raw_itype = self._get(offs, np.uint32)
offs += int(raw_itype.nbytes)
alen = self._get(offs, np.uint64)
offs += int(alen.nbytes)
aparts: list[npt.NDArray[Any]] = [raw_itype, alen]
data_idxs: list[int] = []
for idx in range(alen[0]):
curr_size, curr_parts, curr_idxs, curr_types = self._get_field_parts(offs, raw_itype[0])
if idx == 0:
types += curr_types
idxs_offs = len(aparts)
aparts += curr_parts
data_idxs += (idx + idxs_offs for idx in curr_idxs)
offs += curr_size
return offs - orig_offs, aparts, data_idxs, types
# We can't deal with this one.
raise ValueError('Unknown/unhandled field type {gtype}')
def _get_tensor_info_field(self, orig_offs: int) -> ReaderField:
offs = orig_offs
# Get Tensor Name
name_len, name_data = self._get_str(offs)
offs += int(name_len.nbytes + name_data.nbytes)
# Get Tensor Dimensions Count
n_dims = self._get(offs, np.uint32)
offs += int(n_dims.nbytes)
# Get Tensor Dimension Array
dims = self._get(offs, np.uint64, n_dims[0])
offs += int(dims.nbytes)
# Get Tensor Encoding Scheme Type
raw_dtype = self._get(offs, np.uint32)
offs += int(raw_dtype.nbytes)
# Get Tensor Offset
offset_tensor = self._get(offs, np.uint64)
offs += int(offset_tensor.nbytes)
return ReaderField(
orig_offs,
str(bytes(name_data), encoding = 'utf-8'),
[name_len, name_data, n_dims, dims, raw_dtype, offset_tensor],
[1, 3, 4, 5],
)
def _build_fields(self, offs: int, count: int) -> int:
for _ in range(count):
orig_offs = offs
kv_klen, kv_kdata = self._get_str(offs)
offs += int(kv_klen.nbytes + kv_kdata.nbytes)
raw_kv_type = self._get(offs, np.uint32)
offs += int(raw_kv_type.nbytes)
parts: list[npt.NDArray[Any]] = [kv_klen, kv_kdata, raw_kv_type]
idxs_offs = len(parts)
field_size, field_parts, field_idxs, field_types = self._get_field_parts(offs, raw_kv_type[0])
parts += field_parts
self._push_field(ReaderField(
orig_offs,
str(bytes(kv_kdata), encoding = 'utf-8'),
parts,
[idx + idxs_offs for idx in field_idxs],
field_types,
), skip_sum = True)
offs += field_size
return offs
def _build_tensor_info(self, offs: int, count: int) -> tuple[int, list[ReaderField]]:
tensor_fields = []
for _ in range(count):
field = self._get_tensor_info_field(offs)
offs += sum(int(part.nbytes) for part in field.parts)
tensor_fields.append(field)
return offs, tensor_fields
def _build_tensors(self, start_offs: int, fields: list[ReaderField]) -> None:
tensors = []
tensor_names = set() # keep track of name to prevent duplicated tensors
for field in fields:
_name_len, name_data, _n_dims, dims, raw_dtype, offset_tensor = field.parts
# check if there's any tensor having same name already in the list
tensor_name = str(bytes(name_data), encoding = 'utf-8')
if tensor_name in tensor_names:
raise ValueError(f'Found duplicated tensor with name {tensor_name}')
tensor_names.add(tensor_name)
ggml_type = GGMLQuantizationType(raw_dtype[0])
n_elems = int(np.prod(dims))
np_dims = tuple(reversed(dims.tolist()))
block_size, type_size = GGML_QUANT_SIZES[ggml_type]
n_bytes = n_elems * type_size // block_size
data_offs = int(start_offs + offset_tensor[0])
item_type: npt.DTypeLike
if ggml_type == GGMLQuantizationType.F16:
item_count = n_elems
item_type = np.float16
elif ggml_type == GGMLQuantizationType.F32:
item_count = n_elems
item_type = np.float32
elif ggml_type == GGMLQuantizationType.F64:
item_count = n_elems
item_type = np.float64
elif ggml_type == GGMLQuantizationType.I8:
item_count = n_elems
item_type = np.int8
elif ggml_type == GGMLQuantizationType.I16:
item_count = n_elems
item_type = np.int16
elif ggml_type == GGMLQuantizationType.I32:
item_count = n_elems
item_type = np.int32
elif ggml_type == GGMLQuantizationType.I64:
item_count = n_elems
item_type = np.int64
else:
item_count = n_bytes
item_type = np.uint8
np_dims = quant_shape_to_byte_shape(np_dims, ggml_type)
tensors.append(ReaderTensor(
name = tensor_name,
tensor_type = ggml_type,
shape = dims,
n_elements = n_elems,
n_bytes = n_bytes,
data_offset = data_offs,
data = self._get(data_offs, item_type, item_count).reshape(np_dims),
field = field,
))
self.tensors = tensors

885
packages_3rdparty/gguf/gguf_writer.py vendored Executable file
View File

@@ -0,0 +1,885 @@
from __future__ import annotations
import logging
import os
import shutil
import struct
import tempfile
from dataclasses import dataclass
from enum import Enum, auto
from math import prod
from pathlib import Path
from io import BufferedWriter
from typing import IO, Any, Sequence, Mapping
from string import ascii_letters, digits
import numpy as np
from .constants import (
GGUF_DEFAULT_ALIGNMENT,
GGUF_MAGIC,
GGUF_VERSION,
GGMLQuantizationType,
GGUFEndian,
GGUFValueType,
Keys,
RopeScalingType,
PoolingType,
TokenType,
)
from .quants import quant_shape_from_byte_shape
logger = logging.getLogger(__name__)
SHARD_NAME_FORMAT = "{:s}-{:05d}-of-{:05d}.gguf"
@dataclass
class TensorInfo:
shape: Sequence[int]
dtype: GGMLQuantizationType
nbytes: int
tensor: np.ndarray[Any, Any] | None = None
@dataclass
class GGUFValue:
value: Any
type: GGUFValueType
class WriterState(Enum):
NO_FILE = auto()
EMPTY = auto()
HEADER = auto()
KV_DATA = auto()
TI_DATA = auto()
WEIGHTS = auto()
class GGUFWriter:
fout: list[BufferedWriter] | None
path: Path | None
temp_file: tempfile.SpooledTemporaryFile[bytes] | None
tensors: list[dict[str, TensorInfo]]
kv_data: list[dict[str, GGUFValue]]
state: WriterState
_simple_value_packing = {
GGUFValueType.UINT8: "B",
GGUFValueType.INT8: "b",
GGUFValueType.UINT16: "H",
GGUFValueType.INT16: "h",
GGUFValueType.UINT32: "I",
GGUFValueType.INT32: "i",
GGUFValueType.FLOAT32: "f",
GGUFValueType.UINT64: "Q",
GGUFValueType.INT64: "q",
GGUFValueType.FLOAT64: "d",
GGUFValueType.BOOL: "?",
}
def __init__(
self, path: os.PathLike[str] | str | None, arch: str, use_temp_file: bool = False, endianess: GGUFEndian = GGUFEndian.LITTLE,
split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False, small_first_shard: bool = False
):
self.fout = None
self.path = Path(path) if path else None
self.arch = arch
self.endianess = endianess
self.data_alignment = GGUF_DEFAULT_ALIGNMENT
self.use_temp_file = use_temp_file
self.temp_file = None
self.tensors = [{}]
self.kv_data = [{}]
self.split_max_tensors = split_max_tensors
self.split_max_size = split_max_size
self.dry_run = dry_run
self.small_first_shard = small_first_shard
logger.info("gguf: This GGUF file is for {0} Endian only".format(
"Big" if self.endianess == GGUFEndian.BIG else "Little",
))
self.state = WriterState.NO_FILE
if self.small_first_shard:
self.tensors.append({})
self.add_architecture()
def get_total_parameter_count(self) -> tuple[int, int, int, int]:
total_params = 0
shared_params = 0
expert_params = 0
expert_sum = 0
n_expert_tensors = 0
last_lora_a: tuple[str, TensorInfo] | None = None
for tensors in self.tensors:
for name, info in tensors.items():
shape = info.shape
if name.endswith(".lora_a"):
last_lora_a = (name, info)
continue
elif name.endswith(".lora_b"):
if last_lora_a is None or last_lora_a[0] != name[:-1] + "a":
# Bail when the LoRA pair can't be found trivially
logger.warning("can't measure LoRA size correctly, tensor order is unusual")
return 0, 0, 0, 0
else:
shape = (*shape[:-1], last_lora_a[1].shape[-1])
size = prod(shape)
if "_exps." in name:
expert_params += (size // shape[-3])
expert_sum += shape[-3]
n_expert_tensors += 1
else:
shared_params += size
total_params += size
# Hopefully this should work even for variable-expert-count models
expert_count = (expert_sum // n_expert_tensors) if n_expert_tensors > 0 else 0
# Negate the total to signal it's likely not exact
if last_lora_a is not None:
total_params = -total_params
# NOTE: keep the output in the same order as accepted by 'size_label' in gguf-py/gguf/utility.py
return total_params, shared_params, expert_params, expert_count
def format_shard_names(self, path: Path) -> list[Path]:
if len(self.tensors) == 1:
return [path]
return [path.with_name(SHARD_NAME_FORMAT.format(path.stem, i + 1, len(self.tensors))) for i in range(len(self.tensors))]
def open_output_file(self, path: Path | None = None) -> None:
if self.state is WriterState.EMPTY and self.fout is not None and (path is None or path == self.path):
# allow calling this multiple times as long as the path is the same
return
if self.state is not WriterState.NO_FILE:
raise ValueError(f'Expected output file to be not yet opened, got {self.state}')
if path is not None:
self.path = path
if self.path is not None:
filenames = self.print_plan()
self.fout = [open(filename, "wb") for filename in filenames]
self.state = WriterState.EMPTY
def print_plan(self) -> list[Path]:
logger.info("Writing the following files:")
assert self.path is not None
filenames = self.format_shard_names(self.path)
assert len(filenames) == len(self.tensors)
for name, tensors in zip(filenames, self.tensors):
logger.info(f"{name}: n_tensors = {len(tensors)}, total_size = {GGUFWriter.format_n_bytes_to_str(sum(ti.nbytes for ti in tensors.values()))}")
if self.dry_run:
logger.info("Dry run, not writing files")
for name in filenames:
print(name) # noqa: NP100
exit()
return filenames
def add_shard_kv_data(self) -> None:
if len(self.tensors) == 1:
return
total_tensors = sum(len(t) for t in self.tensors)
assert self.fout is not None
total_splits = len(self.fout)
self.kv_data.extend({} for _ in range(len(self.kv_data), total_splits))
for i, kv_data in enumerate(self.kv_data):
kv_data[Keys.Split.LLM_KV_SPLIT_NO] = GGUFValue(i, GGUFValueType.UINT16)
kv_data[Keys.Split.LLM_KV_SPLIT_COUNT] = GGUFValue(total_splits, GGUFValueType.UINT16)
kv_data[Keys.Split.LLM_KV_SPLIT_TENSORS_COUNT] = GGUFValue(total_tensors, GGUFValueType.INT32)
def write_header_to_file(self, path: Path | None = None) -> None:
if len(self.tensors) == 1 and (self.split_max_tensors != 0 or self.split_max_size != 0):
logger.warning("Model fails split requirements, not splitting")
self.open_output_file(path)
if self.state is not WriterState.EMPTY:
raise ValueError(f'Expected output file to be empty, got {self.state}')
assert self.fout is not None
assert len(self.fout) == len(self.tensors)
assert len(self.kv_data) == 1
self.add_shard_kv_data()
for fout, tensors, kv_data in zip(self.fout, self.tensors, self.kv_data):
fout.write(self._pack("<I", GGUF_MAGIC, skip_pack_prefix = True))
fout.write(self._pack("I", GGUF_VERSION))
fout.write(self._pack("Q", len(tensors)))
fout.write(self._pack("Q", len(kv_data)))
fout.flush()
self.state = WriterState.HEADER
def write_kv_data_to_file(self) -> None:
if self.state is not WriterState.HEADER:
raise ValueError(f'Expected output file to contain the header, got {self.state}')
assert self.fout is not None
for fout, kv_data in zip(self.fout, self.kv_data):
kv_bytes = bytearray()
for key, val in kv_data.items():
kv_bytes += self._pack_val(key, GGUFValueType.STRING, add_vtype=False)
kv_bytes += self._pack_val(val.value, val.type, add_vtype=True)
fout.write(kv_bytes)
self.flush()
self.state = WriterState.KV_DATA
def write_ti_data_to_file(self) -> None:
if self.state is not WriterState.KV_DATA:
raise ValueError(f'Expected output file to contain KV data, got {self.state}')
assert self.fout is not None
for fout, tensors in zip(self.fout, self.tensors):
ti_data = bytearray()
offset_tensor = 0
for name, ti in tensors.items():
ti_data += self._pack_val(name, GGUFValueType.STRING, add_vtype=False)
n_dims = len(ti.shape)
ti_data += self._pack("I", n_dims)
for j in range(n_dims):
ti_data += self._pack("Q", ti.shape[n_dims - 1 - j])
ti_data += self._pack("I", ti.dtype)
ti_data += self._pack("Q", offset_tensor)
offset_tensor += GGUFWriter.ggml_pad(ti.nbytes, self.data_alignment)
fout.write(ti_data)
fout.flush()
self.state = WriterState.TI_DATA
def add_key_value(self, key: str, val: Any, vtype: GGUFValueType) -> None:
if any(key in kv_data for kv_data in self.kv_data):
raise ValueError(f'Duplicated key name {key!r}')
self.kv_data[0][key] = GGUFValue(value=val, type=vtype)
def add_uint8(self, key: str, val: int) -> None:
self.add_key_value(key,val, GGUFValueType.UINT8)
def add_int8(self, key: str, val: int) -> None:
self.add_key_value(key, val, GGUFValueType.INT8)
def add_uint16(self, key: str, val: int) -> None:
self.add_key_value(key, val, GGUFValueType.UINT16)
def add_int16(self, key: str, val: int) -> None:
self.add_key_value(key, val, GGUFValueType.INT16)
def add_uint32(self, key: str, val: int) -> None:
self.add_key_value(key, val, GGUFValueType.UINT32)
def add_int32(self, key: str, val: int) -> None:
self.add_key_value(key, val, GGUFValueType.INT32)
def add_float32(self, key: str, val: float) -> None:
self.add_key_value(key, val, GGUFValueType.FLOAT32)
def add_uint64(self, key: str, val: int) -> None:
self.add_key_value(key, val, GGUFValueType.UINT64)
def add_int64(self, key: str, val: int) -> None:
self.add_key_value(key, val, GGUFValueType.INT64)
def add_float64(self, key: str, val: float) -> None:
self.add_key_value(key, val, GGUFValueType.FLOAT64)
def add_bool(self, key: str, val: bool) -> None:
self.add_key_value(key, val, GGUFValueType.BOOL)
def add_string(self, key: str, val: str) -> None:
if not val:
return
self.add_key_value(key, val, GGUFValueType.STRING)
def add_array(self, key: str, val: Sequence[Any]) -> None:
if len(val) == 0:
return
self.add_key_value(key, val, GGUFValueType.ARRAY)
@staticmethod
def ggml_pad(x: int, n: int) -> int:
return ((x + n - 1) // n) * n
def add_tensor_info(
self, name: str, tensor_shape: Sequence[int], tensor_dtype: np.dtype,
tensor_nbytes: int, raw_dtype: GGMLQuantizationType | None = None,
) -> None:
if self.state is not WriterState.NO_FILE:
raise ValueError(f'Expected output file to be not yet opened, got {self.state}')
if any(name in tensors for tensors in self.tensors):
raise ValueError(f'Duplicated tensor name {name!r}')
if raw_dtype is None:
if tensor_dtype == np.float16:
dtype = GGMLQuantizationType.F16
elif tensor_dtype == np.float32:
dtype = GGMLQuantizationType.F32
elif tensor_dtype == np.float64:
dtype = GGMLQuantizationType.F64
elif tensor_dtype == np.int8:
dtype = GGMLQuantizationType.I8
elif tensor_dtype == np.int16:
dtype = GGMLQuantizationType.I16
elif tensor_dtype == np.int32:
dtype = GGMLQuantizationType.I32
elif tensor_dtype == np.int64:
dtype = GGMLQuantizationType.I64
else:
raise ValueError("Only F16, F32, F64, I8, I16, I32, I64 tensors are supported for now")
else:
dtype = raw_dtype
if tensor_dtype == np.uint8:
tensor_shape = quant_shape_from_byte_shape(tensor_shape, raw_dtype)
# make sure there is at least one tensor before splitting
if len(self.tensors[-1]) > 0:
if ( # split when over tensor limit
self.split_max_tensors != 0
and len(self.tensors[-1]) >= self.split_max_tensors
) or ( # split when over size limit
self.split_max_size != 0
and sum(ti.nbytes for ti in self.tensors[-1].values()) + tensor_nbytes > self.split_max_size
):
self.tensors.append({})
self.tensors[-1][name] = TensorInfo(shape=tensor_shape, dtype=dtype, nbytes=tensor_nbytes)
def add_tensor(
self, name: str, tensor: np.ndarray[Any, Any], raw_shape: Sequence[int] | None = None,
raw_dtype: GGMLQuantizationType | None = None,
) -> None:
if self.endianess == GGUFEndian.BIG:
tensor.byteswap(inplace=True)
if self.use_temp_file and self.temp_file is None:
fp = tempfile.SpooledTemporaryFile(mode="w+b", max_size=256 * 1024 * 1024)
fp.seek(0)
self.temp_file = fp
shape: Sequence[int] = raw_shape if raw_shape is not None else tensor.shape
self.add_tensor_info(name, shape, tensor.dtype, tensor.nbytes, raw_dtype=raw_dtype)
if self.temp_file is None:
self.tensors[-1][name].tensor = tensor
return
tensor.tofile(self.temp_file)
self.write_padding(self.temp_file, tensor.nbytes)
def write_padding(self, fp: IO[bytes], n: int, align: int | None = None) -> None:
pad = GGUFWriter.ggml_pad(n, align if align is not None else self.data_alignment) - n
if pad != 0:
fp.write(bytes([0] * pad))
def write_tensor_data(self, tensor: np.ndarray[Any, Any]) -> None:
if self.state is not WriterState.TI_DATA and self.state is not WriterState.WEIGHTS:
raise ValueError(f'Expected output file to contain tensor info or weights, got {self.state}')
assert self.fout is not None
if self.endianess == GGUFEndian.BIG:
tensor.byteswap(inplace=True)
file_id = -1
for i, tensors in enumerate(self.tensors):
if len(tensors) > 0:
file_id = i
break
fout = self.fout[file_id]
# pop the first tensor info
# TODO: cleaner way to get the first key
first_tensor_name = [name for name, _ in zip(self.tensors[file_id].keys(), range(1))][0]
ti = self.tensors[file_id].pop(first_tensor_name)
assert ti.nbytes == tensor.nbytes
self.write_padding(fout, fout.tell())
tensor.tofile(fout)
self.write_padding(fout, tensor.nbytes)
self.state = WriterState.WEIGHTS
def write_tensors_to_file(self, *, progress: bool = False) -> None:
self.write_ti_data_to_file()
assert self.fout is not None
for fout in self.fout:
self.write_padding(fout, fout.tell())
if self.temp_file is None:
shard_bar = None
bar = None
if progress:
from tqdm import tqdm
total_bytes = sum(ti.nbytes for t in self.tensors for ti in t.values())
if len(self.fout) > 1:
shard_bar = tqdm(desc=f"Shard (0/{len(self.fout)})", total=None, unit="byte", unit_scale=True)
bar = tqdm(desc="Writing", total=total_bytes, unit="byte", unit_scale=True)
for i, (fout, tensors) in enumerate(zip(self.fout, self.tensors)):
if shard_bar is not None:
shard_bar.set_description(f"Shard ({i + 1}/{len(self.fout)})")
total = sum(ti.nbytes for ti in tensors.values())
shard_bar.reset(total=(total if total > 0 else None))
# relying on the fact that Python dicts preserve insertion order (since 3.7)
for ti in tensors.values():
assert ti.tensor is not None # can only iterate once over the tensors
assert ti.tensor.nbytes == ti.nbytes
ti.tensor.tofile(fout)
if shard_bar is not None:
shard_bar.update(ti.nbytes)
if bar is not None:
bar.update(ti.nbytes)
self.write_padding(fout, ti.nbytes)
ti.tensor = None
else:
self.temp_file.seek(0)
shutil.copyfileobj(self.temp_file, self.fout[0 if not self.small_first_shard else 1])
self.flush()
self.temp_file.close()
self.state = WriterState.WEIGHTS
def flush(self) -> None:
assert self.fout is not None
for fout in self.fout:
fout.flush()
def close(self) -> None:
if self.fout is not None:
for fout in self.fout:
fout.close()
self.fout = None
def add_type(self, type_name: str) -> None:
self.add_string(Keys.General.TYPE, type_name)
def add_architecture(self) -> None:
self.add_string(Keys.General.ARCHITECTURE, self.arch)
def add_quantization_version(self, quantization_version: int) -> None:
self.add_uint32(Keys.General.QUANTIZATION_VERSION, quantization_version)
def add_custom_alignment(self, alignment: int) -> None:
self.data_alignment = alignment
self.add_uint32(Keys.General.ALIGNMENT, alignment)
def add_file_type(self, ftype: int) -> None:
self.add_uint32(Keys.General.FILE_TYPE, ftype)
def add_name(self, name: str) -> None:
self.add_string(Keys.General.NAME, name)
def add_author(self, author: str) -> None:
self.add_string(Keys.General.AUTHOR, author)
def add_version(self, version: str) -> None:
self.add_string(Keys.General.VERSION, version)
def add_organization(self, organization: str) -> None:
self.add_string(Keys.General.ORGANIZATION, organization)
def add_finetune(self, finetune: str) -> None:
self.add_string(Keys.General.FINETUNE, finetune)
def add_basename(self, basename: str) -> None:
self.add_string(Keys.General.BASENAME, basename)
def add_description(self, description: str) -> None:
self.add_string(Keys.General.DESCRIPTION, description)
def add_quantized_by(self, quantized: str) -> None:
self.add_string(Keys.General.QUANTIZED_BY, quantized)
def add_size_label(self, size_label: str) -> None:
self.add_string(Keys.General.SIZE_LABEL, size_label)
def add_license(self, license: str) -> None:
self.add_string(Keys.General.LICENSE, license)
def add_license_name(self, license: str) -> None:
self.add_string(Keys.General.LICENSE_NAME, license)
def add_license_link(self, license: str) -> None:
self.add_string(Keys.General.LICENSE_LINK, license)
def add_url(self, url: str) -> None:
self.add_string(Keys.General.URL, url)
def add_doi(self, doi: str) -> None:
self.add_string(Keys.General.DOI, doi)
def add_uuid(self, uuid: str) -> None:
self.add_string(Keys.General.UUID, uuid)
def add_repo_url(self, repo_url: str) -> None:
self.add_string(Keys.General.REPO_URL, repo_url)
def add_source_url(self, url: str) -> None:
self.add_string(Keys.General.SOURCE_URL, url)
def add_source_doi(self, doi: str) -> None:
self.add_string(Keys.General.SOURCE_DOI, doi)
def add_source_uuid(self, uuid: str) -> None:
self.add_string(Keys.General.SOURCE_UUID, uuid)
def add_source_repo_url(self, repo_url: str) -> None:
self.add_string(Keys.General.SOURCE_REPO_URL, repo_url)
def add_base_model_count(self, source_count: int) -> None:
self.add_uint32(Keys.General.BASE_MODEL_COUNT, source_count)
def add_base_model_name(self, source_id: int, name: str) -> None:
self.add_string(Keys.General.BASE_MODEL_NAME.format(id=source_id), name)
def add_base_model_author(self, source_id: int, author: str) -> None:
self.add_string(Keys.General.BASE_MODEL_AUTHOR.format(id=source_id), author)
def add_base_model_version(self, source_id: int, version: str) -> None:
self.add_string(Keys.General.BASE_MODEL_VERSION.format(id=source_id), version)
def add_base_model_organization(self, source_id: int, organization: str) -> None:
self.add_string(Keys.General.BASE_MODEL_ORGANIZATION.format(id=source_id), organization)
def add_base_model_url(self, source_id: int, url: str) -> None:
self.add_string(Keys.General.BASE_MODEL_URL.format(id=source_id), url)
def add_base_model_doi(self, source_id: int, doi: str) -> None:
self.add_string(Keys.General.BASE_MODEL_DOI.format(id=source_id), doi)
def add_base_model_uuid(self, source_id: int, uuid: str) -> None:
self.add_string(Keys.General.BASE_MODEL_UUID.format(id=source_id), uuid)
def add_base_model_repo_url(self, source_id: int, repo_url: str) -> None:
self.add_string(Keys.General.BASE_MODEL_REPO_URL.format(id=source_id), repo_url)
def add_tags(self, tags: Sequence[str]) -> None:
self.add_array(Keys.General.TAGS, tags)
def add_languages(self, languages: Sequence[str]) -> None:
self.add_array(Keys.General.LANGUAGES, languages)
def add_datasets(self, datasets: Sequence[str]) -> None:
self.add_array(Keys.General.DATASETS, datasets)
def add_tensor_data_layout(self, layout: str) -> None:
self.add_string(Keys.LLM.TENSOR_DATA_LAYOUT.format(arch=self.arch), layout)
def add_vocab_size(self, size: int) -> None:
self.add_uint32(Keys.LLM.VOCAB_SIZE.format(arch=self.arch), size)
def add_context_length(self, length: int) -> None:
self.add_uint32(Keys.LLM.CONTEXT_LENGTH.format(arch=self.arch), length)
def add_embedding_length(self, length: int) -> None:
self.add_uint32(Keys.LLM.EMBEDDING_LENGTH.format(arch=self.arch), length)
def add_block_count(self, length: int) -> None:
self.add_uint32(Keys.LLM.BLOCK_COUNT.format(arch=self.arch), length)
def add_leading_dense_block_count(self, length: int) -> None:
self.add_uint32(Keys.LLM.LEADING_DENSE_BLOCK_COUNT.format(arch=self.arch), length)
def add_feed_forward_length(self, length: int | Sequence[int]) -> None:
if isinstance(length, int):
self.add_uint32(Keys.LLM.FEED_FORWARD_LENGTH.format(arch=self.arch), length)
else:
self.add_array(Keys.LLM.FEED_FORWARD_LENGTH.format(arch=self.arch), length)
def add_expert_feed_forward_length(self, length: int) -> None:
self.add_uint32(Keys.LLM.EXPERT_FEED_FORWARD_LENGTH.format(arch=self.arch), length)
def add_expert_shared_feed_forward_length(self, length: int) -> None:
self.add_uint32(Keys.LLM.EXPERT_SHARED_FEED_FORWARD_LENGTH.format(arch=self.arch), length)
def add_parallel_residual(self, use: bool) -> None:
self.add_bool(Keys.LLM.USE_PARALLEL_RESIDUAL.format(arch=self.arch), use)
def add_decoder_start_token_id(self, id: int) -> None:
self.add_uint32(Keys.LLM.DECODER_START_TOKEN_ID.format(arch=self.arch), id)
def add_head_count(self, count: int | Sequence[int]) -> None:
if isinstance(count, int):
self.add_uint32(Keys.Attention.HEAD_COUNT.format(arch=self.arch), count)
else:
self.add_array(Keys.Attention.HEAD_COUNT.format(arch=self.arch), count)
def add_head_count_kv(self, count: int | Sequence[int]) -> None:
if isinstance(count, int):
self.add_uint32(Keys.Attention.HEAD_COUNT_KV.format(arch=self.arch), count)
else:
self.add_array(Keys.Attention.HEAD_COUNT_KV.format(arch=self.arch), count)
def add_key_length(self, length: int) -> None:
self.add_uint32(Keys.Attention.KEY_LENGTH.format(arch=self.arch), length)
def add_value_length(self, length: int) -> None:
self.add_uint32(Keys.Attention.VALUE_LENGTH.format(arch=self.arch), length)
def add_max_alibi_bias(self, bias: float) -> None:
self.add_float32(Keys.Attention.MAX_ALIBI_BIAS.format(arch=self.arch), bias)
def add_clamp_kqv(self, value: float) -> None:
self.add_float32(Keys.Attention.CLAMP_KQV.format(arch=self.arch), value)
def add_logit_scale(self, value: float) -> None:
self.add_float32(Keys.LLM.LOGIT_SCALE.format(arch=self.arch), value)
def add_attn_logit_softcapping(self, value: float) -> None:
self.add_float32(Keys.LLM.ATTN_LOGIT_SOFTCAPPING.format(arch=self.arch), value)
def add_final_logit_softcapping(self, value: float) -> None:
self.add_float32(Keys.LLM.FINAL_LOGIT_SOFTCAPPING.format(arch=self.arch), value)
def add_expert_count(self, count: int) -> None:
self.add_uint32(Keys.LLM.EXPERT_COUNT.format(arch=self.arch), count)
def add_expert_used_count(self, count: int) -> None:
self.add_uint32(Keys.LLM.EXPERT_USED_COUNT.format(arch=self.arch), count)
def add_expert_shared_count(self, count: int) -> None:
self.add_uint32(Keys.LLM.EXPERT_SHARED_COUNT.format(arch=self.arch), count)
def add_expert_weights_scale(self, value: float) -> None:
self.add_float32(Keys.LLM.EXPERT_WEIGHTS_SCALE.format(arch=self.arch), value)
def add_layer_norm_eps(self, value: float) -> None:
self.add_float32(Keys.Attention.LAYERNORM_EPS.format(arch=self.arch), value)
def add_layer_norm_rms_eps(self, value: float) -> None:
self.add_float32(Keys.Attention.LAYERNORM_RMS_EPS.format(arch=self.arch), value)
def add_causal_attention(self, value: bool) -> None:
self.add_bool(Keys.Attention.CAUSAL.format(arch=self.arch), value)
def add_q_lora_rank(self, length: int) -> None:
self.add_uint32(Keys.Attention.Q_LORA_RANK.format(arch=self.arch), length)
def add_kv_lora_rank(self, length: int) -> None:
self.add_uint32(Keys.Attention.KV_LORA_RANK.format(arch=self.arch), length)
def add_relative_attn_buckets_count(self, value: int) -> None:
self.add_uint32(Keys.Attention.REL_BUCKETS_COUNT.format(arch=self.arch), value)
def add_sliding_window(self, value: int) -> None:
self.add_uint32(Keys.Attention.SLIDING_WINDOW.format(arch=self.arch), value)
def add_pooling_type(self, value: PoolingType) -> None:
self.add_uint32(Keys.LLM.POOLING_TYPE.format(arch=self.arch), value.value)
def add_rope_dimension_count(self, count: int) -> None:
self.add_uint32(Keys.Rope.DIMENSION_COUNT.format(arch=self.arch), count)
def add_rope_freq_base(self, value: float) -> None:
self.add_float32(Keys.Rope.FREQ_BASE.format(arch=self.arch), value)
def add_rope_scaling_type(self, value: RopeScalingType) -> None:
self.add_string(Keys.Rope.SCALING_TYPE.format(arch=self.arch), value.value)
def add_rope_scaling_factor(self, value: float) -> None:
self.add_float32(Keys.Rope.SCALING_FACTOR.format(arch=self.arch), value)
def add_rope_scaling_attn_factors(self, value: float) -> None:
self.add_float32(Keys.Rope.SCALING_ATTN_FACTOR.format(arch=self.arch), value)
def add_rope_scaling_orig_ctx_len(self, value: int) -> None:
self.add_uint32(Keys.Rope.SCALING_ORIG_CTX_LEN.format(arch=self.arch), value)
def add_rope_scaling_finetuned(self, value: bool) -> None:
self.add_bool(Keys.Rope.SCALING_FINETUNED.format(arch=self.arch), value)
def add_rope_scaling_yarn_log_mul(self, value: float) -> None:
self.add_float32(Keys.Rope.SCALING_YARN_LOG_MUL.format(arch=self.arch), value)
def add_ssm_conv_kernel(self, value: int) -> None:
self.add_uint32(Keys.SSM.CONV_KERNEL.format(arch=self.arch), value)
def add_ssm_inner_size(self, value: int) -> None:
self.add_uint32(Keys.SSM.INNER_SIZE.format(arch=self.arch), value)
def add_ssm_state_size(self, value: int) -> None:
self.add_uint32(Keys.SSM.STATE_SIZE.format(arch=self.arch), value)
def add_ssm_time_step_rank(self, value: int) -> None:
self.add_uint32(Keys.SSM.TIME_STEP_RANK.format(arch=self.arch), value)
def add_tokenizer_model(self, model: str) -> None:
self.add_string(Keys.Tokenizer.MODEL, model)
def add_tokenizer_pre(self, pre: str) -> None:
self.add_string(Keys.Tokenizer.PRE, pre)
def add_token_list(self, tokens: Sequence[str] | Sequence[bytes] | Sequence[bytearray]) -> None:
self.add_array(Keys.Tokenizer.LIST, tokens)
def add_token_merges(self, merges: Sequence[str] | Sequence[bytes] | Sequence[bytearray]) -> None:
self.add_array(Keys.Tokenizer.MERGES, merges)
def add_token_types(self, types: Sequence[TokenType] | Sequence[int]) -> None:
self.add_array(Keys.Tokenizer.TOKEN_TYPE, types)
def add_token_type_count(self, value: int) -> None:
self.add_uint32(Keys.Tokenizer.TOKEN_TYPE_COUNT, value)
def add_token_scores(self, scores: Sequence[float]) -> None:
self.add_array(Keys.Tokenizer.SCORES, scores)
def add_bos_token_id(self, id: int) -> None:
self.add_uint32(Keys.Tokenizer.BOS_ID, id)
def add_eos_token_id(self, id: int) -> None:
self.add_uint32(Keys.Tokenizer.EOS_ID, id)
def add_unk_token_id(self, id: int) -> None:
self.add_uint32(Keys.Tokenizer.UNK_ID, id)
def add_sep_token_id(self, id: int) -> None:
self.add_uint32(Keys.Tokenizer.SEP_ID, id)
def add_pad_token_id(self, id: int) -> None:
self.add_uint32(Keys.Tokenizer.PAD_ID, id)
def add_cls_token_id(self, id: int) -> None:
self.add_uint32(Keys.Tokenizer.CLS_ID, id)
def add_mask_token_id(self, id: int) -> None:
self.add_uint32(Keys.Tokenizer.MASK_ID, id)
def add_add_bos_token(self, value: bool) -> None:
self.add_bool(Keys.Tokenizer.ADD_BOS, value)
def add_add_eos_token(self, value: bool) -> None:
self.add_bool(Keys.Tokenizer.ADD_EOS, value)
def add_add_space_prefix(self, value: bool) -> None:
self.add_bool(Keys.Tokenizer.ADD_PREFIX, value)
def add_remove_extra_whitespaces(self, value: bool) -> None:
self.add_bool(Keys.Tokenizer.REMOVE_EXTRA_WS, value)
def add_precompiled_charsmap(self, charsmap: Sequence[bytes]) -> None:
self.add_array(Keys.Tokenizer.PRECOMPILED_CHARSMAP, charsmap)
def add_chat_template(self, value: str | Sequence[Mapping[str, str]]) -> None:
if not isinstance(value, str):
template_default = None
template_names = set()
for choice in value:
name = choice.get('name', '')
template = choice.get('template')
# Allowing non-alphanumerical characters in template name is probably not a good idea, so filter it
name = ''.join((c if c in ascii_letters + digits else '_' for c in name))
if name and template is not None:
if name == 'default':
template_default = template
else:
template_names.add(name)
self.add_string(Keys.Tokenizer.CHAT_TEMPLATE_N.format(name=name), template)
if template_names:
self.add_array(Keys.Tokenizer.CHAT_TEMPLATES, list(template_names))
if template_default is None:
return
value = template_default
self.add_string(Keys.Tokenizer.CHAT_TEMPLATE, value)
def add_prefix_token_id(self, id: int) -> None:
self.add_uint32(Keys.Tokenizer.PREFIX_ID, id)
def add_suffix_token_id(self, id: int) -> None:
self.add_uint32(Keys.Tokenizer.SUFFIX_ID, id)
def add_middle_token_id(self, id: int) -> None:
self.add_uint32(Keys.Tokenizer.MIDDLE_ID, id)
def add_eot_token_id(self, id: int) -> None:
self.add_uint32(Keys.Tokenizer.EOT_ID, id)
def add_eom_token_id(self, id: int) -> None:
self.add_uint32(Keys.Tokenizer.EOM_ID, id)
def _pack(self, fmt: str, value: Any, skip_pack_prefix: bool = False) -> bytes:
pack_prefix = ''
if not skip_pack_prefix:
pack_prefix = '<' if self.endianess == GGUFEndian.LITTLE else '>'
return struct.pack(f'{pack_prefix}{fmt}', value)
def _pack_val(self, val: Any, vtype: GGUFValueType, add_vtype: bool) -> bytes:
kv_data = bytearray()
if add_vtype:
kv_data += self._pack("I", vtype)
pack_fmt = self._simple_value_packing.get(vtype)
if pack_fmt is not None:
kv_data += self._pack(pack_fmt, val, skip_pack_prefix = vtype == GGUFValueType.BOOL)
elif vtype == GGUFValueType.STRING:
encoded_val = val.encode("utf-8") if isinstance(val, str) else val
kv_data += self._pack("Q", len(encoded_val))
kv_data += encoded_val
elif vtype == GGUFValueType.ARRAY:
if not isinstance(val, Sequence):
raise ValueError("Invalid GGUF metadata array, expecting sequence")
if len(val) == 0:
raise ValueError("Invalid GGUF metadata array. Empty array")
if isinstance(val, bytes):
ltype = GGUFValueType.UINT8
else:
ltype = GGUFValueType.get_type(val[0])
if not all(GGUFValueType.get_type(i) is ltype for i in val[1:]):
raise ValueError("All items in a GGUF array should be of the same type")
kv_data += self._pack("I", ltype)
kv_data += self._pack("Q", len(val))
for item in val:
kv_data += self._pack_val(item, ltype, add_vtype=False)
else:
raise ValueError("Invalid GGUF metadata value type or value")
return kv_data
@staticmethod
def format_n_bytes_to_str(num: int) -> str:
if num == 0:
return "negligible - metadata only"
fnum = float(num)
for unit in ("", "K", "M", "G"):
if abs(fnum) < 1000.0:
return f"{fnum:3.1f}{unit}"
fnum /= 1000.0
return f"{fnum:.1f}T - over 1TB, split recommended"

213
packages_3rdparty/gguf/lazy.py vendored Executable file
View File

@@ -0,0 +1,213 @@
from __future__ import annotations
from abc import ABC, ABCMeta, abstractmethod
import logging
from typing import Any, Callable
import numpy as np
from numpy.typing import DTypeLike
logger = logging.getLogger(__name__)
class LazyMeta(ABCMeta):
def __new__(cls, name: str, bases: tuple[type, ...], namespace: dict[str, Any], **kwargs):
def __getattr__(self, name: str) -> Any:
meta_attr = getattr(self._meta, name)
if callable(meta_attr):
return type(self)._wrap_fn(
(lambda s, *args, **kwargs: getattr(s, name)(*args, **kwargs)),
use_self=self,
)
elif isinstance(meta_attr, self._tensor_type):
# e.g. self.T with torch.Tensor should still be wrapped
return type(self)._wrap_fn(lambda s: getattr(s, name))(self)
else:
# no need to wrap non-tensor properties,
# and they likely don't depend on the actual contents of the tensor
return meta_attr
namespace["__getattr__"] = __getattr__
# need to make a builder for the wrapped wrapper to copy the name,
# or else it fails with very cryptic error messages,
# because somehow the same string would end up in every closures
def mk_wrap(op_name: str, *, meta_noop: bool = False):
# need to wrap the wrapper to get self
def wrapped_special_op(self, *args, **kwargs):
return type(self)._wrap_fn(
getattr(type(self)._tensor_type, op_name),
meta_noop=meta_noop,
)(self, *args, **kwargs)
return wrapped_special_op
# special methods bypass __getattr__, so they need to be added manually
# ref: https://docs.python.org/3/reference/datamodel.html#special-lookup
# NOTE: doing this from a metaclass is very convenient
# TODO: make this even more comprehensive
for binary_op in (
"lt", "le", "eq", "ne", "ge", "gt", "not"
"abs", "add", "and", "floordiv", "invert", "lshift", "mod", "mul", "matmul",
"neg", "or", "pos", "pow", "rshift", "sub", "truediv", "xor",
"iadd", "iand", "ifloordiv", "ilshift", "imod", "imul", "ior", "irshift", "isub", "ixor",
"radd", "rand", "rfloordiv", "rmul", "ror", "rpow", "rsub", "rtruediv", "rxor",
):
attr_name = f"__{binary_op}__"
# the result of these operators usually has the same shape and dtype as the input,
# so evaluation on the meta tensor can be skipped.
namespace[attr_name] = mk_wrap(attr_name, meta_noop=True)
for special_op in (
"getitem", "setitem", "len",
):
attr_name = f"__{special_op}__"
namespace[attr_name] = mk_wrap(attr_name, meta_noop=False)
return super().__new__(cls, name, bases, namespace, **kwargs)
# Tree of lazy tensors
class LazyBase(ABC, metaclass=LazyMeta):
_tensor_type: type
_meta: Any
_data: Any | None
_args: tuple
_kwargs: dict[str, Any]
_func: Callable[[Any], Any] | None
def __init__(self, *, meta: Any, data: Any | None = None, args: tuple = (), kwargs: dict[str, Any] | None = None, func: Callable[[Any], Any] | None = None):
super().__init__()
self._meta = meta
self._data = data
self._args = args
self._kwargs = kwargs if kwargs is not None else {}
self._func = func
assert self._func is not None or self._data is not None
def __init_subclass__(cls) -> None:
if "_tensor_type" not in cls.__dict__:
raise TypeError(f"property '_tensor_type' must be defined for {cls!r}")
return super().__init_subclass__()
@staticmethod
def _recurse_apply(o: Any, fn: Callable[[Any], Any]) -> Any:
# TODO: dict and set
if isinstance(o, (list, tuple)):
L = []
for item in o:
L.append(LazyBase._recurse_apply(item, fn))
if isinstance(o, tuple):
L = tuple(L)
return L
elif isinstance(o, LazyBase):
return fn(o)
else:
return o
@classmethod
def _wrap_fn(cls, fn: Callable, *, use_self: LazyBase | None = None, meta_noop: bool | DTypeLike | tuple[DTypeLike, Callable[[tuple[int, ...]], tuple[int, ...]]] = False) -> Callable[[Any], Any]:
def wrapped_fn(*args, **kwargs):
if kwargs is None:
kwargs = {}
args = ((use_self,) if use_self is not None else ()) + args
meta_args = LazyBase._recurse_apply(args, lambda t: t._meta)
# TODO: maybe handle tensors in kwargs too
if isinstance(meta_noop, bool) and not meta_noop:
try:
res = fn(*meta_args, **kwargs)
except NotImplementedError:
# running some operations on PyTorch's Meta tensors can cause this exception
res = None
else:
# some operators don't need to actually run on the meta tensors
assert len(args) > 0
res = args[0]
assert isinstance(res, cls)
res = res._meta
# allow operations to override the dtype and shape
if meta_noop is not True:
if isinstance(meta_noop, tuple):
dtype, shape = meta_noop
assert callable(shape)
res = cls.meta_with_dtype_and_shape(dtype, shape(res.shape))
else:
res = cls.meta_with_dtype_and_shape(meta_noop, res.shape)
if isinstance(res, cls._tensor_type):
return cls(meta=cls.eager_to_meta(res), args=args, kwargs=kwargs, func=fn)
else:
del res # not needed
# non-tensor return likely relies on the contents of the args
# (e.g. the result of torch.equal)
eager_args = cls.to_eager(args)
return fn(*eager_args, **kwargs)
return wrapped_fn
@classmethod
def to_eager(cls, t: Any) -> Any:
def simple_to_eager(_t: LazyBase) -> Any:
if _t._data is not None:
return _t._data
# NOTE: there's a recursion limit in Python (usually 1000)
assert _t._func is not None
_t._args = cls._recurse_apply(_t._args, simple_to_eager)
_t._data = _t._func(*_t._args, **_t._kwargs)
# sanity check
assert _t._data is not None
assert _t._data.dtype == _t._meta.dtype
assert _t._data.shape == _t._meta.shape
return _t._data
# recurse into lists and/or tuples, keeping their structure
return cls._recurse_apply(t, simple_to_eager)
@classmethod
def eager_to_meta(cls, t: Any) -> Any:
return cls.meta_with_dtype_and_shape(t.dtype, t.shape)
# must be overridden, meta tensor init is backend-specific
@classmethod
@abstractmethod
def meta_with_dtype_and_shape(cls, dtype: Any, shape: Any) -> Any: pass
@classmethod
def from_eager(cls, t: Any) -> Any:
if type(t) is cls:
# already lazy
return t
elif isinstance(t, cls._tensor_type):
return cls(meta=cls.eager_to_meta(t), data=t)
else:
return TypeError(f"{type(t)!r} is not compatible with {cls._tensor_type!r}")
class LazyNumpyTensor(LazyBase):
_tensor_type = np.ndarray
shape: tuple[int, ...] # Makes the type checker happy in quants.py
@classmethod
def meta_with_dtype_and_shape(cls, dtype: DTypeLike, shape: tuple[int, ...]) -> np.ndarray[Any, Any]:
# The initial idea was to use np.nan as the fill value,
# but non-float types like np.int16 can't use that.
# So zero it is.
cheat = np.zeros(1, dtype)
return np.lib.stride_tricks.as_strided(cheat, shape, (0 for _ in shape))
def astype(self, dtype, *args, **kwargs):
meta = type(self).meta_with_dtype_and_shape(dtype, self._meta.shape)
full_args = (self, dtype,) + args
return type(self)(meta=meta, args=full_args, kwargs=kwargs, func=(lambda a, *args, **kwargs: a.astype(*args, **kwargs)))
def tofile(self, *args, **kwargs):
eager = LazyNumpyTensor.to_eager(self)
return eager.tofile(*args, **kwargs)
# TODO: __array_function__

510
packages_3rdparty/gguf/metadata.py vendored Executable file
View File

@@ -0,0 +1,510 @@
from __future__ import annotations
import re
import json
import yaml
import logging
from pathlib import Path
from typing import Any, Literal, Optional
from dataclasses import dataclass
from .constants import Keys
import gguf
logger = logging.getLogger("metadata")
@dataclass
class Metadata:
# Authorship Metadata to be written to GGUF KV Store
name: Optional[str] = None
author: Optional[str] = None
version: Optional[str] = None
organization: Optional[str] = None
finetune: Optional[str] = None
basename: Optional[str] = None
description: Optional[str] = None
quantized_by: Optional[str] = None
size_label: Optional[str] = None
url: Optional[str] = None
doi: Optional[str] = None
uuid: Optional[str] = None
repo_url: Optional[str] = None
source_url: Optional[str] = None
source_doi: Optional[str] = None
source_uuid: Optional[str] = None
source_repo_url: Optional[str] = None
license: Optional[str] = None
license_name: Optional[str] = None
license_link: Optional[str] = None
base_models: Optional[list[dict]] = None
tags: Optional[list[str]] = None
languages: Optional[list[str]] = None
datasets: Optional[list[str]] = None
@staticmethod
def load(metadata_override_path: Optional[Path] = None, model_path: Optional[Path] = None, model_name: Optional[str] = None, total_params: int = 0) -> Metadata:
# This grabs as many contextual authorship metadata as possible from the model repository
# making any conversion as required to match the gguf kv store metadata format
# as well as giving users the ability to override any authorship metadata that may be incorrect
# Create a new Metadata instance
metadata = Metadata()
model_card = Metadata.load_model_card(model_path)
hf_params = Metadata.load_hf_parameters(model_path)
# TODO: load adapter_config.json when possible, it usually contains the base model of the LoRA adapter
# heuristics
metadata = Metadata.apply_metadata_heuristic(metadata, model_card, hf_params, model_path, total_params)
# Metadata Override File Provided
# This is based on LLM_KV_NAMES mapping in llama.cpp
metadata_override = Metadata.load_metadata_override(metadata_override_path)
metadata.name = metadata_override.get(Keys.General.NAME, metadata.name)
metadata.author = metadata_override.get(Keys.General.AUTHOR, metadata.author)
metadata.version = metadata_override.get(Keys.General.VERSION, metadata.version)
metadata.organization = metadata_override.get(Keys.General.ORGANIZATION, metadata.organization)
metadata.finetune = metadata_override.get(Keys.General.FINETUNE, metadata.finetune)
metadata.basename = metadata_override.get(Keys.General.BASENAME, metadata.basename)
metadata.description = metadata_override.get(Keys.General.DESCRIPTION, metadata.description)
metadata.quantized_by = metadata_override.get(Keys.General.QUANTIZED_BY, metadata.quantized_by)
metadata.size_label = metadata_override.get(Keys.General.SIZE_LABEL, metadata.size_label)
metadata.license_name = metadata_override.get(Keys.General.LICENSE_NAME, metadata.license_name)
metadata.license_link = metadata_override.get(Keys.General.LICENSE_LINK, metadata.license_link)
metadata.url = metadata_override.get(Keys.General.URL, metadata.url)
metadata.doi = metadata_override.get(Keys.General.DOI, metadata.doi)
metadata.uuid = metadata_override.get(Keys.General.UUID, metadata.uuid)
metadata.repo_url = metadata_override.get(Keys.General.REPO_URL, metadata.repo_url)
metadata.source_url = metadata_override.get(Keys.General.SOURCE_URL, metadata.source_url)
metadata.source_doi = metadata_override.get(Keys.General.SOURCE_DOI, metadata.source_doi)
metadata.source_uuid = metadata_override.get(Keys.General.SOURCE_UUID, metadata.source_uuid)
metadata.source_repo_url = metadata_override.get(Keys.General.SOURCE_REPO_URL, metadata.source_repo_url)
# Base Models is received here as an array of models
metadata.base_models = metadata_override.get("general.base_models", metadata.base_models)
metadata.tags = metadata_override.get(Keys.General.TAGS, metadata.tags)
metadata.languages = metadata_override.get(Keys.General.LANGUAGES, metadata.languages)
metadata.datasets = metadata_override.get(Keys.General.DATASETS, metadata.datasets)
# Direct Metadata Override (via direct cli argument)
if model_name is not None:
metadata.name = model_name
return metadata
@staticmethod
def load_metadata_override(metadata_override_path: Optional[Path] = None) -> dict[str, Any]:
if metadata_override_path is None or not metadata_override_path.is_file():
return {}
with open(metadata_override_path, "r", encoding="utf-8") as f:
return json.load(f)
@staticmethod
def load_model_card(model_path: Optional[Path] = None) -> dict[str, Any]:
if model_path is None or not model_path.is_dir():
return {}
model_card_path = model_path / "README.md"
if not model_card_path.is_file():
return {}
# The model card metadata is assumed to always be in YAML
# ref: https://github.com/huggingface/transformers/blob/a5c642fe7a1f25d3bdcd76991443ba6ff7ee34b2/src/transformers/modelcard.py#L468-L473
with open(model_card_path, "r", encoding="utf-8") as f:
if f.readline() == "---\n":
raw = f.read().partition("---\n")[0]
data = yaml.safe_load(raw)
if isinstance(data, dict):
return data
else:
logger.error(f"while reading YAML model card frontmatter, data is {type(data)} instead of dict")
return {}
else:
return {}
@staticmethod
def load_hf_parameters(model_path: Optional[Path] = None) -> dict[str, Any]:
if model_path is None or not model_path.is_dir():
return {}
config_path = model_path / "config.json"
if not config_path.is_file():
return {}
with open(config_path, "r", encoding="utf-8") as f:
return json.load(f)
@staticmethod
def id_to_title(string):
# Convert capitalization into title form unless acronym or version number
return ' '.join([w.title() if w.islower() and not re.match(r'^(v\d+(?:\.\d+)*|\d.*)$', w) else w for w in string.strip().replace('-', ' ').split()])
@staticmethod
def get_model_id_components(model_id: Optional[str] = None, total_params: int = 0) -> tuple[str | None, str | None, str | None, str | None, str | None, str | None]:
# Huggingface often store model id as '<org>/<model name>'
# so let's parse it and apply some heuristics if possible for model name components
if model_id is None:
# model ID missing
return None, None, None, None, None, None
if ' ' in model_id:
# model ID is actually a normal human sentence
# which means its most likely a normal model name only
# not part of the hugging face naming standard, but whatever
return model_id, None, None, None, None, None
if '/' in model_id:
# model ID (huggingface style)
org_component, model_full_name_component = model_id.split('/', 1)
else:
# model ID but missing org components
org_component, model_full_name_component = None, model_id
# Check if we erroneously matched against './' or '../' etc...
if org_component is not None and len(org_component) > 0 and org_component[0] == '.':
org_component = None
name_parts: list[str] = model_full_name_component.split('-')
# Remove empty parts
for i in reversed(range(len(name_parts))):
if len(name_parts[i]) == 0:
del name_parts[i]
name_types: list[
set[Literal["basename", "size_label", "finetune", "version", "type"]]
] = [set() for _ in name_parts]
# Annotate the name
for i, part in enumerate(name_parts):
# Version
if re.fullmatch(r'(v|iter)?\d+([.]\d+)*', part, re.IGNORECASE):
name_types[i].add("version")
# Quant type (should not be there for base models, but still annotated)
elif re.fullmatch(r'i?q\d(_\w)*|b?fp?(16|32)', part, re.IGNORECASE):
name_types[i].add("type")
name_parts[i] = part.upper()
# Model size
elif i > 0 and re.fullmatch(r'(([A]|\d+[x])?\d+([._]\d+)?[KMBT][\d]?|small|mini|medium|large|x?xl)', part, re.IGNORECASE):
part = part.replace("_", ".")
# Handle weird bloom-7b1 notation
if part[-1].isdecimal():
part = part[:-2] + "." + part[-1] + part[-2]
# Normalize the size suffixes
if len(part) > 1 and part[-2].isdecimal():
if part[-1] in "kmbt":
part = part[:-1] + part[-1].upper()
if total_params != 0:
try:
label_params = float(part[:-1]) * pow(1000, " KMBT".find(part[-1]))
# Only use it as a size label if it's close or bigger than the model size
# Note that LoRA adapters don't necessarily include all layers,
# so this is why bigger label sizes are accepted.
# Do not use the size label when it's smaller than 1/8 of the model size
if (total_params < 0 and label_params < abs(total_params) // 8) or (
# Check both directions when the current model isn't a LoRA adapter
total_params > 0 and abs(label_params - total_params) > 7 * total_params // 8
):
# Likely a context length
name_types[i].add("finetune")
# Lowercase the size when it's a context length
part = part[:-1] + part[-1].lower()
except ValueError:
# Failed to convert the size label to float, use it anyway
pass
if len(name_types[i]) == 0:
name_types[i].add("size_label")
name_parts[i] = part
# Some easy to recognize finetune names
elif i > 0 and re.fullmatch(r'chat|instruct|vision|lora', part, re.IGNORECASE):
if total_params < 0 and part.lower() == "lora":
# ignore redundant "lora" in the finetune part when the output is a lora adapter
name_types[i].add("type")
else:
name_types[i].add("finetune")
# Ignore word-based size labels when there is at least a number-based one present
# TODO: should word-based size labels always be removed instead?
if any(c.isdecimal() for n, t in zip(name_parts, name_types) if "size_label" in t for c in n):
for n, t in zip(name_parts, name_types):
if "size_label" in t:
if all(c.isalpha() for c in n):
t.remove("size_label")
at_start = True
# Find the basename through the annotated name
for part, t in zip(name_parts, name_types):
if at_start and ((len(t) == 0 and part[0].isalpha()) or "version" in t):
t.add("basename")
else:
if at_start:
at_start = False
if len(t) == 0:
t.add("finetune")
# Remove the basename annotation from trailing version
for part, t in zip(reversed(name_parts), reversed(name_types)):
if "basename" in t and len(t) > 1:
t.remove("basename")
else:
break
basename = "-".join(n for n, t in zip(name_parts, name_types) if "basename" in t) or None
# Deduplicate size labels using order-preserving 'dict' ('set' seems to sort the keys)
size_label = "-".join(dict.fromkeys(s for s, t in zip(name_parts, name_types) if "size_label" in t).keys()) or None
finetune = "-".join(f for f, t in zip(name_parts, name_types) if "finetune" in t) or None
# TODO: should the basename version always be excluded?
# NOTE: multiple finetune versions are joined together
version = "-".join(v for v, t, in zip(name_parts, name_types) if "version" in t and "basename" not in t) or None
if size_label is None and finetune is None and version is None:
# Too ambiguous, output nothing
basename = None
return model_full_name_component, org_component, basename, finetune, version, size_label
@staticmethod
def apply_metadata_heuristic(metadata: Metadata, model_card: Optional[dict] = None, hf_params: Optional[dict] = None, model_path: Optional[Path] = None, total_params: int = 0) -> Metadata:
# Reference Model Card Metadata: https://github.com/huggingface/hub-docs/blob/main/modelcard.md?plain=1
# Model Card Heuristics
########################
if model_card is not None:
def use_model_card_metadata(metadata_key: str, model_card_key: str):
if model_card_key in model_card and getattr(metadata, metadata_key, None) is None:
setattr(metadata, metadata_key, model_card.get(model_card_key))
def use_array_model_card_metadata(metadata_key: str, model_card_key: str):
# Note: Will append rather than replace if already exist
tags_value = model_card.get(model_card_key, None)
if tags_value is None:
return
current_value = getattr(metadata, metadata_key, None)
if current_value is None:
current_value = []
if isinstance(tags_value, str):
current_value.append(tags_value)
elif isinstance(tags_value, list):
current_value.extend(tags_value)
setattr(metadata, metadata_key, current_value)
# LLAMA.cpp's direct internal convention
# (Definitely not part of hugging face formal/informal standard)
#########################################
use_model_card_metadata("name", "name")
use_model_card_metadata("author", "author")
use_model_card_metadata("version", "version")
use_model_card_metadata("organization", "organization")
use_model_card_metadata("description", "description")
use_model_card_metadata("finetune", "finetune")
use_model_card_metadata("basename", "basename")
use_model_card_metadata("size_label", "size_label")
use_model_card_metadata("source_url", "url")
use_model_card_metadata("source_doi", "doi")
use_model_card_metadata("source_uuid", "uuid")
use_model_card_metadata("source_repo_url", "repo_url")
# LLAMA.cpp's huggingface style convention
# (Definitely not part of hugging face formal/informal standard... but with model_ appended to match their style)
###########################################
use_model_card_metadata("name", "model_name")
use_model_card_metadata("author", "model_author")
use_model_card_metadata("version", "model_version")
use_model_card_metadata("organization", "model_organization")
use_model_card_metadata("description", "model_description")
use_model_card_metadata("finetune", "model_finetune")
use_model_card_metadata("basename", "model_basename")
use_model_card_metadata("size_label", "model_size_label")
use_model_card_metadata("source_url", "model_url")
use_model_card_metadata("source_doi", "model_doi")
use_model_card_metadata("source_uuid", "model_uuid")
use_model_card_metadata("source_repo_url", "model_repo_url")
# Hugging Face Direct Convention
#################################
# Not part of huggingface model card standard but notice some model creator using it
# such as TheBloke in 'TheBloke/Mistral-7B-Instruct-v0.2-GGUF'
use_model_card_metadata("name", "model_name")
use_model_card_metadata("author", "model_creator")
use_model_card_metadata("basename", "model_type")
if "base_model" in model_card:
# This represents the parent models that this is based on
# Example: stabilityai/stable-diffusion-xl-base-1.0. Can also be a list (for merges)
# Example of merges: https://huggingface.co/EmbeddedLLM/Mistral-7B-Merge-14-v0.1/blob/main/README.md
metadata_base_models = []
base_model_value = model_card.get("base_model", None)
if base_model_value is not None:
if isinstance(base_model_value, str):
metadata_base_models.append(base_model_value)
elif isinstance(base_model_value, list):
metadata_base_models.extend(base_model_value)
if metadata.base_models is None:
metadata.base_models = []
for model_id in metadata_base_models:
# NOTE: model size of base model is assumed to be similar to the size of the current model
model_full_name_component, org_component, basename, finetune, version, size_label = Metadata.get_model_id_components(model_id, total_params)
base_model = {}
if model_full_name_component is not None:
base_model["name"] = Metadata.id_to_title(model_full_name_component)
if org_component is not None:
base_model["organization"] = Metadata.id_to_title(org_component)
if version is not None:
base_model["version"] = version
if org_component is not None and model_full_name_component is not None:
base_model["repo_url"] = f"https://huggingface.co/{org_component}/{model_full_name_component}"
metadata.base_models.append(base_model)
use_model_card_metadata("license", "license")
use_model_card_metadata("license_name", "license_name")
use_model_card_metadata("license_link", "license_link")
use_array_model_card_metadata("tags", "tags")
use_array_model_card_metadata("tags", "pipeline_tag")
use_array_model_card_metadata("languages", "languages")
use_array_model_card_metadata("languages", "language")
use_array_model_card_metadata("datasets", "datasets")
use_array_model_card_metadata("datasets", "dataset")
# Hugging Face Parameter Heuristics
####################################
if hf_params is not None:
hf_name_or_path = hf_params.get("_name_or_path")
if hf_name_or_path is not None and hf_name_or_path.count('/') <= 1:
# Use _name_or_path only if its actually a model name and not some computer path
# e.g. 'meta-llama/Llama-2-7b-hf'
model_id = hf_name_or_path
model_full_name_component, org_component, basename, finetune, version, size_label = Metadata.get_model_id_components(model_id, total_params)
if metadata.name is None and model_full_name_component is not None:
metadata.name = Metadata.id_to_title(model_full_name_component)
if metadata.organization is None and org_component is not None:
metadata.organization = Metadata.id_to_title(org_component)
if metadata.basename is None and basename is not None:
metadata.basename = basename
if metadata.finetune is None and finetune is not None:
metadata.finetune = finetune
if metadata.version is None and version is not None:
metadata.version = version
if metadata.size_label is None and size_label is not None:
metadata.size_label = size_label
# Directory Folder Name Fallback Heuristics
############################################
if model_path is not None:
model_id = model_path.name
model_full_name_component, org_component, basename, finetune, version, size_label = Metadata.get_model_id_components(model_id, total_params)
if metadata.name is None and model_full_name_component is not None:
metadata.name = Metadata.id_to_title(model_full_name_component)
if metadata.organization is None and org_component is not None:
metadata.organization = Metadata.id_to_title(org_component)
if metadata.basename is None and basename is not None:
metadata.basename = basename
if metadata.finetune is None and finetune is not None:
metadata.finetune = finetune
if metadata.version is None and version is not None:
metadata.version = version
if metadata.size_label is None and size_label is not None:
metadata.size_label = size_label
return metadata
def set_gguf_meta_model(self, gguf_writer: gguf.GGUFWriter):
assert self.name is not None
gguf_writer.add_name(self.name)
if self.author is not None:
gguf_writer.add_author(self.author)
if self.version is not None:
gguf_writer.add_version(self.version)
if self.organization is not None:
gguf_writer.add_organization(self.organization)
if self.finetune is not None:
gguf_writer.add_finetune(self.finetune)
if self.basename is not None:
gguf_writer.add_basename(self.basename)
if self.description is not None:
gguf_writer.add_description(self.description)
if self.quantized_by is not None:
gguf_writer.add_quantized_by(self.quantized_by)
if self.size_label is not None:
gguf_writer.add_size_label(self.size_label)
if self.license is not None:
gguf_writer.add_license(self.license)
if self.license_name is not None:
gguf_writer.add_license_name(self.license_name)
if self.license_link is not None:
gguf_writer.add_license_link(self.license_link)
if self.url is not None:
gguf_writer.add_url(self.url)
if self.doi is not None:
gguf_writer.add_doi(self.doi)
if self.uuid is not None:
gguf_writer.add_uuid(self.uuid)
if self.repo_url is not None:
gguf_writer.add_repo_url(self.repo_url)
if self.source_url is not None:
gguf_writer.add_source_url(self.source_url)
if self.source_doi is not None:
gguf_writer.add_source_doi(self.source_doi)
if self.source_uuid is not None:
gguf_writer.add_source_uuid(self.source_uuid)
if self.source_repo_url is not None:
gguf_writer.add_source_repo_url(self.source_repo_url)
if self.base_models is not None:
gguf_writer.add_base_model_count(len(self.base_models))
for key, base_model_entry in enumerate(self.base_models):
if "name" in base_model_entry:
gguf_writer.add_base_model_name(key, base_model_entry["name"])
if "author" in base_model_entry:
gguf_writer.add_base_model_author(key, base_model_entry["author"])
if "version" in base_model_entry:
gguf_writer.add_base_model_version(key, base_model_entry["version"])
if "organization" in base_model_entry:
gguf_writer.add_base_model_organization(key, base_model_entry["organization"])
if "url" in base_model_entry:
gguf_writer.add_base_model_url(key, base_model_entry["url"])
if "doi" in base_model_entry:
gguf_writer.add_base_model_doi(key, base_model_entry["doi"])
if "uuid" in base_model_entry:
gguf_writer.add_base_model_uuid(key, base_model_entry["uuid"])
if "repo_url" in base_model_entry:
gguf_writer.add_base_model_repo_url(key, base_model_entry["repo_url"])
if self.tags is not None:
gguf_writer.add_tags(self.tags)
if self.languages is not None:
gguf_writer.add_languages(self.languages)
if self.datasets is not None:
gguf_writer.add_datasets(self.datasets)

1524
packages_3rdparty/gguf/quants.py vendored Executable file
View File

File diff suppressed because it is too large Load Diff

76
packages_3rdparty/gguf/quick_4bits_ops.py vendored Executable file
View File

@@ -0,0 +1,76 @@
# By Forge
import torch
def native_unpack_4x4bits_in_1x16bits_to_4x8bits_in_1x32bits(x):
x = x.view(torch.uint8).view(x.size(0), -1)
unpacked = torch.stack([x & 15, x >> 4], dim=-1)
reshaped = unpacked.view(x.size(0), -1)
reshaped = reshaped.view(torch.int8) - 8
return reshaped.view(torch.int32)
def native_unpack_4x4bits_in_1x16bits_to_4x8bits_in_1x32bits_u(x):
x = x.view(torch.uint8).view(x.size(0), -1)
unpacked = torch.stack([x & 15, x >> 4], dim=-1)
reshaped = unpacked.view(x.size(0), -1)
return reshaped.view(torch.int32)
disable_all_optimizations = False
if not hasattr(torch, 'uint16'):
disable_all_optimizations = True
if disable_all_optimizations:
print('You are using PyTorch below version 2.3. Some optimizations will be disabled.')
if not disable_all_optimizations:
native_4bits_lookup_table = native_unpack_4x4bits_in_1x16bits_to_4x8bits_in_1x32bits(torch.arange(start=0, end=256*256, dtype=torch.long).to(torch.uint16))[:, 0]
native_4bits_lookup_table_u = native_unpack_4x4bits_in_1x16bits_to_4x8bits_in_1x32bits_u(torch.arange(start=0, end=256*256, dtype=torch.long).to(torch.uint16))[:, 0]
def quick_unpack_4bits(x):
if disable_all_optimizations:
return torch.stack([x & 15, x >> 4], dim=-1).view(x.size(0), -1).view(torch.int8) - 8
global native_4bits_lookup_table
s0 = x.size(0)
x = x.view(torch.uint16)
if native_4bits_lookup_table.device != x.device:
native_4bits_lookup_table = native_4bits_lookup_table.to(device=x.device)
y = torch.index_select(input=native_4bits_lookup_table, dim=0, index=x.to(dtype=torch.int32).flatten())
y = y.view(torch.int8)
y = y.view(s0, -1)
return y
def quick_unpack_4bits_u(x):
if disable_all_optimizations:
return torch.stack([x & 15, x >> 4], dim=-1).view(x.size(0), -1)
global native_4bits_lookup_table_u
s0 = x.size(0)
x = x.view(torch.uint16)
if native_4bits_lookup_table_u.device != x.device:
native_4bits_lookup_table_u = native_4bits_lookup_table_u.to(device=x.device)
y = torch.index_select(input=native_4bits_lookup_table_u, dim=0, index=x.to(dtype=torch.int32).flatten())
y = y.view(torch.uint8)
y = y.view(s0, -1)
return y
def change_4bits_order(x):
y = torch.stack([x & 15, x >> 4], dim=-2).view(x.size(0), -1)
z = y[:, ::2] | (y[:, 1::2] << 4)
return z

649
packages_3rdparty/gguf/tensor_mapping.py vendored Executable file
View File

@@ -0,0 +1,649 @@
from __future__ import annotations
from typing import Sequence
from .constants import MODEL_ARCH, MODEL_TENSOR, MODEL_TENSORS, TENSOR_NAMES
class TensorNameMap:
mappings_cfg: dict[MODEL_TENSOR, tuple[str, ...]] = {
# Token embeddings
MODEL_TENSOR.TOKEN_EMBD: (
"gpt_neox.embed_in", # gptneox
"transformer.wte", # gpt2 gpt-j mpt refact qwen dbrx jais
"transformer.word_embeddings", # falcon
"word_embeddings", # bloom
"model.embed_tokens", # llama-hf
"tok_embeddings", # llama-pth
"embeddings.word_embeddings", # bert nomic-bert
"language_model.embedding.word_embeddings", # persimmon
"wte", # gpt2
"transformer.embd.wte", # phi2
"model.tok_embeddings", # internlm2
"model.embedding", # mamba-qbert
"backbone.embedding", # mamba
"backbone.embeddings", # mamba-hf
"transformer.in_out_embed", # Grok
"embedding.word_embeddings", # chatglm
"transformer.token_embeddings", # openelm
"shared", # t5
),
# Token type embeddings
MODEL_TENSOR.TOKEN_TYPES: (
"embeddings.token_type_embeddings", # bert nomic-bert
),
# Normalization of token embeddings
MODEL_TENSOR.TOKEN_EMBD_NORM: (
"word_embeddings_layernorm", # bloom
"embeddings.LayerNorm", # bert
"emb_ln", # nomic-bert
"transformer.norm", # openelm
),
# Position embeddings
MODEL_TENSOR.POS_EMBD: (
"transformer.wpe", # gpt2
"embeddings.position_embeddings", # bert
"wpe", # gpt2
),
# Output
MODEL_TENSOR.OUTPUT: (
"embed_out", # gptneox
"lm_head", # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx jais
"output", # llama-pth bloom internlm2
"word_embeddings_for_head", # persimmon
"lm_head.linear", # phi2
"output_layer", # chatglm
),
# Output norm
MODEL_TENSOR.OUTPUT_NORM: (
"gpt_neox.final_layer_norm", # gptneox
"transformer.ln_f", # gpt2 gpt-j falcon jais
"model.norm", # llama-hf baichuan internlm2
"norm", # llama-pth
"transformer.norm_f", # mpt dbrx
"ln_f", # refact bloom qwen gpt2
"language_model.encoder.final_layernorm", # persimmon
"model.final_layernorm", # persimmon
"lm_head.ln", # phi2
"model.norm_f", # mamba-qbert
"backbone.norm_f", # mamba
"transformer.rms_norm", # Grok
"encoder.final_layernorm", # chatglm
"transformer.norm", # openelm
),
# Rope frequencies
MODEL_TENSOR.ROPE_FREQS: (
"rope.freqs", # llama-pth
"rotary_pos_emb.inv_freq", # chatglm
),
}
block_mappings_cfg: dict[MODEL_TENSOR, tuple[str, ...]] = {
# Attention norm
MODEL_TENSOR.ATTN_NORM: (
"gpt_neox.layers.{bid}.input_layernorm", # gptneox
"transformer.h.{bid}.ln_1", # gpt2 gpt-j refact qwen jais
"transformer.blocks.{bid}.norm_1", # mpt
"transformer.h.{bid}.input_layernorm", # falcon7b
"h.{bid}.input_layernorm", # bloom
"transformer.h.{bid}.ln_mlp", # falcon40b
"model.layers.{bid}.input_layernorm", # llama-hf
"layers.{bid}.attention_norm", # llama-pth
"language_model.encoder.layers.{bid}.input_layernorm", # persimmon
"model.layers.{bid}.ln1", # yi
"h.{bid}.ln_1", # gpt2
"transformer.h.{bid}.ln", # phi2
"model.layers.layers.{bid}.norm", # plamo
"model.layers.{bid}.attention_norm", # internlm2
"model.layers.{bid}.norm", # mamba-qbert
"backbone.layers.{bid}.norm", # mamba
"transformer.decoder_layer.{bid}.rms_norm", # Grok
"transformer.blocks.{bid}.norm_attn_norm.norm_1", # dbrx
"encoder.layers.{bid}.input_layernorm", # chatglm
"transformer.layers.{bid}.attn_norm", # openelm
),
# Attention norm 2
MODEL_TENSOR.ATTN_NORM_2: (
"transformer.h.{bid}.ln_attn", # falcon40b
"encoder.layer.{bid}.layer_norm_1", # jina-v2-code
),
# Attention query-key-value
MODEL_TENSOR.ATTN_QKV: (
"gpt_neox.layers.{bid}.attention.query_key_value", # gptneox
"transformer.h.{bid}.attn.c_attn", # gpt2 qwen jais
"transformer.blocks.{bid}.attn.Wqkv", # mpt
"transformer.blocks.{bid}.norm_attn_norm.attn.Wqkv", # dbrx
"transformer.h.{bid}.self_attention.query_key_value", # falcon
"h.{bid}.self_attention.query_key_value", # bloom
"language_model.encoder.layers.{bid}.self_attention.query_key_value", # persimmon
"model.layers.{bid}.self_attn.query_key_value", # persimmon
"h.{bid}.attn.c_attn", # gpt2
"transformer.h.{bid}.mixer.Wqkv", # phi2
"encoder.layers.{bid}.attn.Wqkv", # nomic-bert
"model.layers.{bid}.self_attn.qkv_proj", # phi3
"encoder.layers.{bid}.self_attention.query_key_value", # chatglm
"transformer.layers.{bid}.attn.qkv_proj", # openelm
),
# Attention query
MODEL_TENSOR.ATTN_Q: (
"model.layers.{bid}.self_attn.q_proj", # llama-hf
"layers.{bid}.attention.wq", # llama-pth
"encoder.layer.{bid}.attention.self.query", # bert
"transformer.h.{bid}.attn.q_proj", # gpt-j
"model.layers.layers.{bid}.self_attn.q_proj", # plamo
"model.layers.{bid}.attention.wq", # internlm2
"transformer.decoder_layer.{bid}.multi_head_attention.query",# Grok
),
# Attention key
MODEL_TENSOR.ATTN_K: (
"model.layers.{bid}.self_attn.k_proj", # llama-hf
"layers.{bid}.attention.wk", # llama-pth
"encoder.layer.{bid}.attention.self.key", # bert
"transformer.h.{bid}.attn.k_proj", # gpt-j
"transformer.h.{bid}.attn.k", # refact
"model.layers.layers.{bid}.self_attn.k_proj", # plamo
"model.layers.{bid}.attention.wk", # internlm2
"transformer.decoder_layer.{bid}.multi_head_attention.key",# Grok
),
# Attention value
MODEL_TENSOR.ATTN_V: (
"model.layers.{bid}.self_attn.v_proj", # llama-hf
"layers.{bid}.attention.wv", # llama-pth
"encoder.layer.{bid}.attention.self.value", # bert
"transformer.h.{bid}.attn.v_proj", # gpt-j
"transformer.h.{bid}.attn.v", # refact
"model.layers.layers.{bid}.self_attn.v_proj", # plamo
"model.layers.{bid}.attention.wv", # internlm2
"transformer.decoder_layer.{bid}.multi_head_attention.value" # Grok
),
# Attention output
MODEL_TENSOR.ATTN_OUT: (
"gpt_neox.layers.{bid}.attention.dense", # gptneox
"transformer.h.{bid}.attn.c_proj", # gpt2 refact qwen jais
"transformer.blocks.{bid}.attn.out_proj", # mpt
"transformer.h.{bid}.self_attention.dense", # falcon
"h.{bid}.self_attention.dense", # bloom
"model.layers.{bid}.self_attn.o_proj", # llama-hf
"layers.{bid}.attention.wo", # llama-pth
"encoder.layer.{bid}.attention.output.dense", # bert
"transformer.h.{bid}.attn.out_proj", # gpt-j
"language_model.encoder.layers.{bid}.self_attention.dense", # persimmon
"model.layers.{bid}.self_attn.dense", # persimmon
"h.{bid}.attn.c_proj", # gpt2
"transformer.h.{bid}.mixer.out_proj", # phi2
"model.layers.layers.{bid}.self_attn.o_proj", # plamo
"model.layers.{bid}.attention.wo", # internlm2
"encoder.layers.{bid}.attn.out_proj", # nomic-bert
"transformer.decoder_layer.{bid}.multi_head_attention.linear", # Grok
"transformer.blocks.{bid}.norm_attn_norm.attn.out_proj", # dbrx
"encoder.layers.{bid}.self_attention.dense", # chatglm
"transformer.layers.{bid}.attn.out_proj", # openelm
),
# Attention output norm
MODEL_TENSOR.ATTN_OUT_NORM: (
"encoder.layer.{bid}.attention.output.LayerNorm", # bert
"encoder.layers.{bid}.norm1", # nomic-bert
"transformer.decoder_layer.{bid}.rms_norm_1", # Grok
"transformer.blocks.{bid}.norm_attn_norm.norm_2", # dbrx
),
MODEL_TENSOR.ATTN_POST_NORM: (
"model.layers.{bid}.post_attention_layernorm", # gemma2
),
# Rotary embeddings
MODEL_TENSOR.ATTN_ROT_EMBD: (
"model.layers.{bid}.self_attn.rotary_emb.inv_freq", # llama-hf
"layers.{bid}.attention.inner_attention.rope.freqs", # llama-pth
"model.layers.layers.{bid}.self_attn.rotary_emb.inv_freq", # plamo
"transformer.h.{bid}.attn.rotary_emb.inv_freq", # codeshell
),
# Feed-forward norm
MODEL_TENSOR.FFN_NORM: (
"gpt_neox.layers.{bid}.post_attention_layernorm", # gptneox
"transformer.h.{bid}.ln_2", # gpt2 refact qwen jais
"h.{bid}.post_attention_layernorm", # bloom
"transformer.blocks.{bid}.norm_2", # mpt
"model.layers.{bid}.post_attention_layernorm", # llama-hf
"layers.{bid}.ffn_norm", # llama-pth
"language_model.encoder.layers.{bid}.post_attention_layernorm", # persimmon
"model.layers.{bid}.ln2", # yi
"h.{bid}.ln_2", # gpt2
"model.layers.{bid}.ffn_norm", # internlm2
"transformer.decoder_layer.{bid}.rms_norm_2", # Grok
"encoder.layers.{bid}.post_attention_layernorm", # chatglm
"transformer.layers.{bid}.ffn_norm", # openelm
),
# Post feed-forward norm
MODEL_TENSOR.FFN_PRE_NORM: (
"model.layers.{bid}.pre_feedforward_layernorm", # gemma2
),
# Post feed-forward norm
MODEL_TENSOR.FFN_POST_NORM: (
"model.layers.{bid}.post_feedforward_layernorm", # gemma2
),
MODEL_TENSOR.FFN_GATE_INP: (
"layers.{bid}.feed_forward.gate", # mixtral
"model.layers.{bid}.block_sparse_moe.gate", # mixtral
"model.layers.{bid}.mlp.gate", # qwen2moe
"transformer.decoder_layer.{bid}.router", # Grok
"transformer.blocks.{bid}.ffn.router.layer", # dbrx
),
MODEL_TENSOR.FFN_GATE_INP_SHEXP: (
"model.layers.{bid}.mlp.shared_expert_gate", # qwen2moe
),
# Feed-forward up
MODEL_TENSOR.FFN_UP: (
"gpt_neox.layers.{bid}.mlp.dense_h_to_4h", # gptneox
"transformer.h.{bid}.mlp.c_fc", # gpt2 jais
"transformer.blocks.{bid}.ffn.up_proj", # mpt
"transformer.h.{bid}.mlp.dense_h_to_4h", # falcon
"h.{bid}.mlp.dense_h_to_4h", # bloom
"model.layers.{bid}.mlp.up_proj", # llama-hf refact
"layers.{bid}.feed_forward.w3", # llama-pth
"encoder.layer.{bid}.intermediate.dense", # bert
"transformer.h.{bid}.mlp.fc_in", # gpt-j
"transformer.h.{bid}.mlp.linear_3", # refact
"language_model.encoder.layers.{bid}.mlp.dense_h_to_4h", # persimmon
"model.layers.{bid}.mlp.dense_h_to_4h", # persimmon
"transformer.h.{bid}.mlp.w1", # qwen
"h.{bid}.mlp.c_fc", # gpt2
"transformer.h.{bid}.mlp.fc1", # phi2
"model.layers.{bid}.mlp.fc1", # phi2
"model.layers.{bid}.mlp.gate_up_proj", # phi3
"model.layers.layers.{bid}.mlp.up_proj", # plamo
"model.layers.{bid}.feed_forward.w3", # internlm2
"encoder.layers.{bid}.mlp.fc11", # nomic-bert
"model.layers.{bid}.mlp.c_fc", # starcoder2
"encoder.layer.{bid}.mlp.gated_layers_v", # jina-bert-v2
"model.layers.{bid}.residual_mlp.w3", # arctic
"encoder.layers.{bid}.mlp.dense_h_to_4h", # chatglm
),
MODEL_TENSOR.FFN_UP_EXP: (
"layers.{bid}.feed_forward.experts.w3", # mixtral (merged)
"transformer.decoder_layer.{bid}.moe.linear_v", # Grok (merged)
"transformer.blocks.{bid}.ffn.experts.mlp.v1", # dbrx
"model.layers.{bid}.mlp.experts.up_proj", # qwen2moe (merged)
),
MODEL_TENSOR.FFN_UP_SHEXP: (
"model.layers.{bid}.mlp.shared_expert.up_proj", # qwen2moe
"model.layers.{bid}.mlp.shared_experts.up_proj", # deepseek2
),
# AWQ-activation gate
MODEL_TENSOR.FFN_ACT: (
"transformer.blocks.{bid}.ffn.act", # mpt
),
# Feed-forward gate
MODEL_TENSOR.FFN_GATE: (
"model.layers.{bid}.mlp.gate_proj", # llama-hf refact
"layers.{bid}.feed_forward.w1", # llama-pth
"transformer.h.{bid}.mlp.w2", # qwen
"transformer.h.{bid}.mlp.c_fc2", # jais
"model.layers.layers.{bid}.mlp.gate_proj", # plamo
"model.layers.{bid}.feed_forward.w1", # internlm2
"encoder.layers.{bid}.mlp.fc12", # nomic-bert
"encoder.layer.{bid}.mlp.gated_layers_w", # jina-bert-v2
"transformer.h.{bid}.mlp.linear_1", # refact
"model.layers.{bid}.residual_mlp.w1", # arctic
),
MODEL_TENSOR.FFN_GATE_EXP: (
"layers.{bid}.feed_forward.experts.w1", # mixtral (merged)
"transformer.decoder_layer.{bid}.moe.linear", # Grok (merged)
"transformer.blocks.{bid}.ffn.experts.mlp.w1", # dbrx
"model.layers.{bid}.mlp.experts.gate_proj", # qwen2moe (merged)
),
MODEL_TENSOR.FFN_GATE_SHEXP: (
"model.layers.{bid}.mlp.shared_expert.gate_proj", # qwen2moe
"model.layers.{bid}.mlp.shared_experts.gate_proj", # deepseek2
),
# Feed-forward down
MODEL_TENSOR.FFN_DOWN: (
"gpt_neox.layers.{bid}.mlp.dense_4h_to_h", # gptneox
"transformer.h.{bid}.mlp.c_proj", # gpt2 refact qwen jais
"transformer.blocks.{bid}.ffn.down_proj", # mpt
"transformer.h.{bid}.mlp.dense_4h_to_h", # falcon
"h.{bid}.mlp.dense_4h_to_h", # bloom
"model.layers.{bid}.mlp.down_proj", # llama-hf
"layers.{bid}.feed_forward.w2", # llama-pth
"encoder.layer.{bid}.output.dense", # bert
"transformer.h.{bid}.mlp.fc_out", # gpt-j
"language_model.encoder.layers.{bid}.mlp.dense_4h_to_h", # persimmon
"model.layers.{bid}.mlp.dense_4h_to_h", # persimmon
"h.{bid}.mlp.c_proj", # gpt2
"transformer.h.{bid}.mlp.fc2", # phi2
"model.layers.{bid}.mlp.fc2", # phi2
"model.layers.layers.{bid}.mlp.down_proj", # plamo
"model.layers.{bid}.feed_forward.w2", # internlm2
"encoder.layers.{bid}.mlp.fc2", # nomic-bert
"model.layers.{bid}.mlp.c_proj", # starcoder2
"encoder.layer.{bid}.mlp.wo", # jina-bert-v2
"transformer.layers.{bid}.ffn.proj_2", # openelm
"model.layers.{bid}.residual_mlp.w2", # arctic
"encoder.layer.{bid}.mlp.down_layer", # jina-bert-v2
"encoder.layers.{bid}.mlp.dense_4h_to_h", # chatglm
),
MODEL_TENSOR.FFN_DOWN_EXP: (
"layers.{bid}.feed_forward.experts.w2", # mixtral (merged)
"transformer.decoder_layer.{bid}.moe.linear_1", # Grok (merged)
"transformer.blocks.{bid}.ffn.experts.mlp.w2", # dbrx
"model.layers.{bid}.mlp.experts.down_proj", # qwen2moe (merged)
),
MODEL_TENSOR.FFN_DOWN_SHEXP: (
"model.layers.{bid}.mlp.shared_expert.down_proj", # qwen2moe
"model.layers.{bid}.mlp.shared_experts.down_proj", # deepseek2
),
MODEL_TENSOR.ATTN_Q_NORM: (
"language_model.encoder.layers.{bid}.self_attention.q_layernorm",
"model.layers.{bid}.self_attn.q_layernorm", # persimmon
"model.layers.{bid}.self_attn.q_norm", # cohere
"transformer.blocks.{bid}.attn.q_ln", # sea-lion
"encoder.layer.{bid}.attention.self.layer_norm_q", # jina-bert-v2
"transformer.layers.{bid}.attn.q_norm", # openelm
),
MODEL_TENSOR.ATTN_K_NORM: (
"language_model.encoder.layers.{bid}.self_attention.k_layernorm",
"model.layers.{bid}.self_attn.k_layernorm", # persimmon
"model.layers.{bid}.self_attn.k_norm", # cohere
"transformer.blocks.{bid}.attn.k_ln", # sea-lion
"encoder.layer.{bid}.attention.self.layer_norm_k", # jina-bert-v2
"transformer.layers.{bid}.attn.k_norm", # openelm
),
MODEL_TENSOR.ROPE_FREQS: (
"language_model.encoder.layers.{bid}.self_attention.rotary_emb.inv_freq", # persimmon
),
MODEL_TENSOR.LAYER_OUT_NORM: (
"encoder.layer.{bid}.output.LayerNorm", # bert
"encoder.layers.{bid}.norm2", # nomic-bert
"transformer.decoder_layer.{bid}.rms_norm_3", # Grok
"encoder.layer.{bid}.mlp.layernorm", # jina-bert-v2
"encoder.layer.{bid}.layer_norm_2" # jina-v2-code
),
MODEL_TENSOR.SSM_IN: (
"model.layers.{bid}.in_proj",
"backbone.layers.{bid}.mixer.in_proj",
),
MODEL_TENSOR.SSM_CONV1D: (
"model.layers.{bid}.conv1d",
"backbone.layers.{bid}.mixer.conv1d",
),
MODEL_TENSOR.SSM_X: (
"model.layers.{bid}.x_proj",
"backbone.layers.{bid}.mixer.x_proj",
),
MODEL_TENSOR.SSM_DT: (
"model.layers.{bid}.dt_proj",
"backbone.layers.{bid}.mixer.dt_proj",
),
MODEL_TENSOR.SSM_A: (
"model.layers.{bid}.A_log",
"backbone.layers.{bid}.mixer.A_log",
),
MODEL_TENSOR.SSM_D: (
"model.layers.{bid}.D",
"backbone.layers.{bid}.mixer.D",
),
MODEL_TENSOR.SSM_OUT: (
"model.layers.{bid}.out_proj",
"backbone.layers.{bid}.mixer.out_proj",
),
MODEL_TENSOR.ATTN_Q_A: (
"model.layers.{bid}.self_attn.q_a_proj", # deepseek2
),
MODEL_TENSOR.ATTN_Q_B: (
"model.layers.{bid}.self_attn.q_b_proj", # deepseek2
),
MODEL_TENSOR.ATTN_KV_A_MQA: (
"model.layers.{bid}.self_attn.kv_a_proj_with_mqa", # deepseek2
),
MODEL_TENSOR.ATTN_KV_B: (
"model.layers.{bid}.self_attn.kv_b_proj", # deepseek2
),
MODEL_TENSOR.ATTN_Q_A_NORM: (
"model.layers.{bid}.self_attn.q_a_layernorm", # deepseek2
),
MODEL_TENSOR.ATTN_KV_A_NORM: (
"model.layers.{bid}.self_attn.kv_a_layernorm", # deepseek2
),
MODEL_TENSOR.ATTN_SUB_NORM: (
"model.layers.{bid}.self_attn.inner_attn_ln", # bitnet
),
MODEL_TENSOR.FFN_SUB_NORM: (
"model.layers.{bid}.mlp.ffn_layernorm", # bitnet
),
MODEL_TENSOR.DEC_ATTN_NORM: (
"decoder.block.{bid}.layer.0.layer_norm", # t5
),
MODEL_TENSOR.DEC_ATTN_Q: (
"decoder.block.{bid}.layer.0.SelfAttention.q", # t5
),
MODEL_TENSOR.DEC_ATTN_K: (
"decoder.block.{bid}.layer.0.SelfAttention.k", # t5
),
MODEL_TENSOR.DEC_ATTN_V: (
"decoder.block.{bid}.layer.0.SelfAttention.v", # t5
),
MODEL_TENSOR.DEC_ATTN_OUT: (
"decoder.block.{bid}.layer.0.SelfAttention.o", # t5
),
MODEL_TENSOR.DEC_ATTN_REL_B: (
"decoder.block.{bid}.layer.0.SelfAttention.relative_attention_bias", # t5
),
MODEL_TENSOR.DEC_CROSS_ATTN_NORM: (
"decoder.block.{bid}.layer.1.layer_norm", # t5
),
MODEL_TENSOR.DEC_CROSS_ATTN_Q: (
"decoder.block.{bid}.layer.1.EncDecAttention.q", # t5
),
MODEL_TENSOR.DEC_CROSS_ATTN_K: (
"decoder.block.{bid}.layer.1.EncDecAttention.k", # t5
),
MODEL_TENSOR.DEC_CROSS_ATTN_V: (
"decoder.block.{bid}.layer.1.EncDecAttention.v", # t5
),
MODEL_TENSOR.DEC_CROSS_ATTN_OUT: (
"decoder.block.{bid}.layer.1.EncDecAttention.o", # t5
),
MODEL_TENSOR.DEC_CROSS_ATTN_REL_B: (
"decoder.block.{bid}.layer.1.EncDecAttention.relative_attention_bias", # t5
),
MODEL_TENSOR.DEC_FFN_NORM: (
"decoder.block.{bid}.layer.2.layer_norm", # t5
),
MODEL_TENSOR.DEC_FFN_GATE: (
"decoder.block.{bid}.layer.2.DenseReluDense.wi_0", # flan-t5
),
MODEL_TENSOR.DEC_FFN_UP: (
"decoder.block.{bid}.layer.2.DenseReluDense.wi", # t5
"decoder.block.{bid}.layer.2.DenseReluDense.wi_1", # flan-t5
),
MODEL_TENSOR.DEC_FFN_DOWN: (
"decoder.block.{bid}.layer.2.DenseReluDense.wo", # t5
),
MODEL_TENSOR.DEC_OUTPUT_NORM: (
"decoder.final_layer_norm", # t5
),
MODEL_TENSOR.ENC_ATTN_NORM: (
"encoder.block.{bid}.layer.0.layer_norm", # t5
),
MODEL_TENSOR.ENC_ATTN_Q: (
"encoder.block.{bid}.layer.0.SelfAttention.q", # t5
),
MODEL_TENSOR.ENC_ATTN_K: (
"encoder.block.{bid}.layer.0.SelfAttention.k", # t5
),
MODEL_TENSOR.ENC_ATTN_V: (
"encoder.block.{bid}.layer.0.SelfAttention.v", # t5
),
MODEL_TENSOR.ENC_ATTN_OUT: (
"encoder.block.{bid}.layer.0.SelfAttention.o", # t5
),
MODEL_TENSOR.ENC_ATTN_REL_B: (
"encoder.block.{bid}.layer.0.SelfAttention.relative_attention_bias", # t5
),
MODEL_TENSOR.ENC_FFN_NORM: (
"encoder.block.{bid}.layer.1.layer_norm", # t5
),
MODEL_TENSOR.ENC_FFN_GATE: (
"encoder.block.{bid}.layer.1.DenseReluDense.wi_0", # flan-t5
),
MODEL_TENSOR.ENC_FFN_UP: (
"encoder.block.{bid}.layer.1.DenseReluDense.wi", # t5
"encoder.block.{bid}.layer.1.DenseReluDense.wi_1", # flan-t5
),
MODEL_TENSOR.ENC_FFN_DOWN: (
"encoder.block.{bid}.layer.1.DenseReluDense.wo", # t5
),
MODEL_TENSOR.ENC_OUTPUT_NORM: (
"encoder.final_layer_norm", # t5
),
}
# architecture-specific block mappings
arch_block_mappings_cfg: dict[MODEL_ARCH, dict[MODEL_TENSOR, tuple[str, ...]]] = {
MODEL_ARCH.ARCTIC: {
MODEL_TENSOR.FFN_NORM: (
"model.layers.{bid}.residual_layernorm",
),
MODEL_TENSOR.FFN_NORM_EXP: (
"model.layers.{bid}.post_attention_layernorm",
),
},
}
mapping: dict[str, tuple[MODEL_TENSOR, str]]
def __init__(self, arch: MODEL_ARCH, n_blocks: int):
self.mapping = {}
for tensor, keys in self.mappings_cfg.items():
if tensor not in MODEL_TENSORS[arch]:
continue
tensor_name = TENSOR_NAMES[tensor]
self.mapping[tensor_name] = (tensor, tensor_name)
for key in keys:
self.mapping[key] = (tensor, tensor_name)
if arch in self.arch_block_mappings_cfg:
self.block_mappings_cfg.update(self.arch_block_mappings_cfg[arch])
for bid in range(n_blocks):
for tensor, keys in self.block_mappings_cfg.items():
if tensor not in MODEL_TENSORS[arch]:
continue
tensor_name = TENSOR_NAMES[tensor].format(bid = bid)
self.mapping[tensor_name] = (tensor, tensor_name)
for key in keys:
key = key.format(bid = bid)
self.mapping[key] = (tensor, tensor_name)
def get_type_and_name(self, key: str, try_suffixes: Sequence[str] = ()) -> tuple[MODEL_TENSOR, str] | None:
result = self.mapping.get(key)
if result is not None:
return result
for suffix in try_suffixes:
if key.endswith(suffix):
result = self.mapping.get(key[:-len(suffix)])
if result is not None:
return result[0], result[1] + suffix
return None
def get_name(self, key: str, try_suffixes: Sequence[str] = ()) -> str | None:
result = self.get_type_and_name(key, try_suffixes = try_suffixes)
if result is None:
return None
return result[1]
def get_type(self, key: str, try_suffixes: Sequence[str] = ()) -> MODEL_TENSOR | None:
result = self.get_type_and_name(key, try_suffixes = try_suffixes)
if result is None:
return None
return result[0]
def __getitem__(self, key: str) -> str:
try:
return self.mapping[key][1]
except KeyError:
raise KeyError(key)
def __contains__(self, key: str) -> bool:
return key in self.mapping
def __repr__(self) -> str:
return repr(self.mapping)
def get_tensor_name_map(arch: MODEL_ARCH, n_blocks: int) -> TensorNameMap:
return TensorNameMap(arch, n_blocks)

69
packages_3rdparty/gguf/utility.py vendored Executable file
View File

@@ -0,0 +1,69 @@
from __future__ import annotations
from typing import Literal
def fill_templated_filename(filename: str, output_type: str | None) -> str:
# Given a file name fill in any type templates e.g. 'some-model-name.{ftype}.gguf'
ftype_lowercase: str = output_type.lower() if output_type is not None else ""
ftype_uppercase: str = output_type.upper() if output_type is not None else ""
return filename.format(ftype_lowercase,
outtype=ftype_lowercase, ftype=ftype_lowercase,
OUTTYPE=ftype_uppercase, FTYPE=ftype_uppercase)
def model_weight_count_rounded_notation(model_params_count: int, min_digits: int = 2) -> str:
if model_params_count > 1e12 :
# Trillions Of Parameters
scaled_model_params = model_params_count * 1e-12
scale_suffix = "T"
elif model_params_count > 1e9 :
# Billions Of Parameters
scaled_model_params = model_params_count * 1e-9
scale_suffix = "B"
elif model_params_count > 1e6 :
# Millions Of Parameters
scaled_model_params = model_params_count * 1e-6
scale_suffix = "M"
else:
# Thousands Of Parameters
scaled_model_params = model_params_count * 1e-3
scale_suffix = "K"
fix = max(min_digits - len(str(round(scaled_model_params)).lstrip('0')), 0)
return f"{scaled_model_params:.{fix}f}{scale_suffix}"
def size_label(total_params: int, shared_params: int, expert_params: int, expert_count: int) -> str:
if expert_count > 0:
pretty_size = model_weight_count_rounded_notation(abs(shared_params) + abs(expert_params), min_digits=2)
size_class = f"{expert_count}x{pretty_size}"
else:
size_class = model_weight_count_rounded_notation(abs(total_params), min_digits=2)
return size_class
def naming_convention(model_name: str | None, base_name: str | None, finetune_string: str | None, version_string: str | None, size_label: str | None, output_type: str | None, model_type: Literal['vocab', 'LoRA'] | None = None) -> str:
# Reference: https://github.com/ggerganov/ggml/blob/master/docs/gguf.md#gguf-naming-convention
if base_name is not None:
name = base_name.strip().replace(' ', '-').replace('/', '-')
elif model_name is not None:
name = model_name.strip().replace(' ', '-').replace('/', '-')
else:
name = "ggml-model"
parameters = f"-{size_label}" if size_label is not None else ""
finetune = f"-{finetune_string.strip().replace(' ', '-')}" if finetune_string is not None else ""
version = f"-{version_string.strip().replace(' ', '-')}" if version_string is not None else ""
encoding = f"-{output_type.strip().replace(' ', '-').upper()}" if output_type is not None else ""
kind = f"-{model_type.strip().replace(' ', '-')}" if model_type is not None else ""
return f"{name}{parameters}{finetune}{version}{encoding}{kind}"

465
packages_3rdparty/gguf/vocab.py vendored Executable file
View File

@@ -0,0 +1,465 @@
from __future__ import annotations
import re
import logging
import json
import os
from pathlib import Path
from typing import Any, Callable, Sequence, Mapping, Iterable, Protocol, ClassVar, runtime_checkable
from sentencepiece import SentencePieceProcessor
import gguf
from .gguf_writer import GGUFWriter
logger = logging.getLogger(__name__)
class SpecialVocab:
merges: list[str]
add_special_token: dict[str, bool]
special_token_ids: dict[str, int]
chat_template: str | Sequence[Mapping[str, str]] | None
def __init__(
self, path: str | os.PathLike[str], load_merges: bool = False,
special_token_types: Iterable[str] | None = None,
n_vocab: int | None = None,
):
self.special_token_ids = {}
self.add_special_token = {}
self.n_vocab = n_vocab
self.load_merges = load_merges
self.merges = []
self.chat_template = None
if special_token_types is not None:
self.special_token_types = special_token_types
else:
self.special_token_types = ('bos', 'eos', 'unk', 'sep', 'pad', 'cls', 'mask')
self._load(Path(path))
def __repr__(self) -> str:
return '<SpecialVocab with {} merges, special tokens {}, add special tokens {}>'.format(
len(self.merges), self.special_token_ids or "unset", self.add_special_token or "unset",
)
def add_to_gguf(self, gw: GGUFWriter, quiet: bool = False) -> None:
if self.merges:
if not quiet:
logger.info(f'Adding {len(self.merges)} merge(s).')
gw.add_token_merges(self.merges)
elif self.load_merges:
logger.warning('Adding merges requested but no merges found, output may be non-functional.')
for typ, tokid in self.special_token_ids.items():
id_handler: Callable[[int], None] | None = getattr(gw, f'add_{typ}_token_id', None)
if id_handler is None:
logger.warning(f'No handler for special token type {typ} with id {tokid} - skipping')
continue
if not quiet:
logger.info(f'Setting special token type {typ} to {tokid}')
id_handler(tokid)
for typ, value in self.add_special_token.items():
add_handler: Callable[[bool], None] | None = getattr(gw, f'add_add_{typ}_token', None)
if add_handler is None:
logger.warning(f'No handler for add_{typ}_token with value {value} - skipping')
continue
if not quiet:
logger.info(f'Setting add_{typ}_token to {value}')
add_handler(value)
if self.chat_template is not None:
if not quiet:
logger.info(f'Setting chat_template to {self.chat_template}')
gw.add_chat_template(self.chat_template)
def _load(self, path: Path) -> None:
self._try_load_from_tokenizer_json(path)
self._try_load_from_config_json(path)
if self.load_merges and not self.merges:
self._try_load_merges_txt(path)
def _try_load_merges_txt(self, path: Path) -> bool:
merges_file = path / 'merges.txt'
if not merges_file.is_file():
return False
with open(merges_file, 'r', encoding = 'utf-8') as fp:
first_line = next(fp, '').strip()
if not first_line.startswith('#'):
fp.seek(0)
line_num = 0
else:
line_num = 1
merges = []
for line in fp:
line_num += 1
line = line.strip()
if not line:
continue
parts = line.split(None, 3)
if len(parts) != 2:
logger.warning(f'{merges_file.name}: Line {line_num}: Entry malformed, ignoring')
continue
merges.append(f'{parts[0]} {parts[1]}')
self.merges = merges
return True
def _set_special_token(self, typ: str, tid: Any) -> None:
if not isinstance(tid, int):
return
if tid < 0:
raise ValueError(f'invalid value for special token type {typ}: {tid}')
if self.n_vocab is None or tid < self.n_vocab:
if typ in self.special_token_ids:
return
self.special_token_ids[typ] = tid
return
logger.warning(f'Special token type {typ}, id {tid} out of range, must be under {self.n_vocab} - skipping')
def _try_load_from_tokenizer_json(self, path: Path) -> bool:
tokenizer_file = path / 'tokenizer.json'
if tokenizer_file.is_file():
with open(tokenizer_file, encoding = 'utf-8') as f:
tokenizer = json.load(f)
if self.load_merges:
merges = tokenizer.get('model', {}).get('merges')
if isinstance(merges, list) and merges and isinstance(merges[0], str):
self.merges = merges
added_tokens = tokenizer.get('added_tokens', {})
else:
added_tokens = {}
tokenizer_config_file = path / 'tokenizer_config.json'
if not tokenizer_config_file.is_file():
return True
with open(tokenizer_config_file, encoding = 'utf-8') as f:
tokenizer_config = json.load(f)
chat_template = tokenizer_config.get('chat_template')
if chat_template is None or isinstance(chat_template, (str, list)):
self.chat_template = chat_template
else:
logger.warning(f'Bad type for chat_template field in {tokenizer_config_file!r} - ignoring')
for typ in self.special_token_types:
add_entry = tokenizer_config.get(f'add_{typ}_token')
if isinstance(add_entry, bool):
self.add_special_token[typ] = add_entry
entry = tokenizer_config.get(f'{typ}_token')
if isinstance(entry, str):
tc_content = entry
elif isinstance(entry, dict):
entry_content = entry.get('content')
if not isinstance(entry_content, str):
continue
tc_content = entry_content
else:
continue
# We only need the first match here.
maybe_token_id = next(
(atok.get('id') for atok in added_tokens if atok.get('content') == tc_content),
None,
)
self._set_special_token(typ, maybe_token_id)
return True
def _try_load_from_config_json(self, path: Path) -> bool:
config_file = path / 'config.json'
if not config_file.is_file():
return False
with open(config_file, encoding = 'utf-8') as f:
config = json.load(f)
for typ in self.special_token_types:
self._set_special_token(typ, config.get(f'{typ}_token_id'))
return True
@runtime_checkable
class BaseVocab(Protocol):
tokenizer_model: ClassVar[str]
name: ClassVar[str]
@runtime_checkable
class Vocab(BaseVocab, Protocol):
vocab_size: int
added_tokens_dict: dict[str, int]
added_tokens_list: list[str]
fname_tokenizer: Path
def __init__(self, base_path: Path): ...
def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: ...
class NoVocab(BaseVocab):
tokenizer_model = "no_vocab"
name = "no_vocab"
def __repr__(self) -> str:
return "<NoVocab for a model without integrated vocabulary>"
class BpeVocab(Vocab):
tokenizer_model = "gpt2"
name = "bpe"
def __init__(self, base_path: Path):
added_tokens: dict[str, int] = {}
if (fname_tokenizer := base_path / 'vocab.json').exists():
# "slow" tokenizer
with open(fname_tokenizer, encoding="utf-8") as f:
self.vocab = json.load(f)
try:
# FIXME: Verify that added tokens here _cannot_ overlap with the main vocab.
with open(base_path / 'added_tokens.json', encoding="utf-8") as f:
added_tokens = json.load(f)
except FileNotFoundError:
pass
else:
# "fast" tokenizer
fname_tokenizer = base_path / 'tokenizer.json'
# if this fails, FileNotFoundError propagates to caller
with open(fname_tokenizer, encoding="utf-8") as f:
tokenizer_json = json.load(f)
tokenizer_model: dict[str, Any] = tokenizer_json['model']
if (
tokenizer_model['type'] != 'BPE' or tokenizer_model.get('byte_fallback', False)
or tokenizer_json['decoder']['type'] != 'ByteLevel'
):
raise FileNotFoundError('Cannot find GPT-2 BPE tokenizer')
self.vocab = tokenizer_model["vocab"]
if (added := tokenizer_json.get('added_tokens')) is not None:
# Added tokens here can be duplicates of the main vocabulary.
added_tokens = {item['content']: item['id']
for item in added
if item['content'] not in self.vocab}
vocab_size = len(self.vocab)
expected_ids = list(range(vocab_size, vocab_size + len(added_tokens)))
actual_ids = sorted(added_tokens.values())
if expected_ids != actual_ids:
expected_end_id = vocab_size + len(actual_ids) - 1
raise ValueError(f"Expected the {len(actual_ids)} added token ID(s) to be sequential in the range "
f"{vocab_size} - {expected_end_id}; got {actual_ids}")
items = sorted(added_tokens.items(), key=lambda text_idx: text_idx[1])
self.added_tokens_dict = added_tokens
self.added_tokens_list = [text for (text, idx) in items]
self.vocab_size_base = vocab_size
self.vocab_size = self.vocab_size_base + len(self.added_tokens_list)
self.fname_tokenizer = fname_tokenizer
def bpe_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
reverse_vocab = {id: encoded_tok for encoded_tok, id in self.vocab.items()}
for i, _ in enumerate(self.vocab):
yield reverse_vocab[i], 0.0, gguf.TokenType.NORMAL
def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
for text in self.added_tokens_list:
score = -1000.0
yield text.encode("utf-8"), score, gguf.TokenType.CONTROL
def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
yield from self.bpe_tokens()
yield from self.added_tokens()
def __repr__(self) -> str:
return f"<BpeVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
class SentencePieceVocab(Vocab):
tokenizer_model = "llama"
name = "spm"
def __init__(self, base_path: Path):
added_tokens: dict[str, int] = {}
if (fname_tokenizer := base_path / 'tokenizer.model').exists():
# normal location
try:
with open(base_path / 'added_tokens.json', encoding="utf-8") as f:
added_tokens = json.load(f)
except FileNotFoundError:
pass
elif not (fname_tokenizer := base_path.parent / 'tokenizer.model').exists():
# not found in alternate location either
raise FileNotFoundError('Cannot find tokenizer.model')
self.sentencepiece_tokenizer = SentencePieceProcessor()
self.sentencepiece_tokenizer.LoadFromFile(str(fname_tokenizer))
vocab_size = self.sentencepiece_tokenizer.vocab_size()
new_tokens = {id: piece for piece, id in added_tokens.items() if id >= vocab_size}
expected_new_ids = list(range(vocab_size, vocab_size + len(new_tokens)))
actual_new_ids = sorted(new_tokens.keys())
if expected_new_ids != actual_new_ids:
raise ValueError(f"Expected new token IDs {expected_new_ids} to be sequential; got {actual_new_ids}")
# Token pieces that were added to the base vocabulary.
self.added_tokens_dict = added_tokens
self.added_tokens_list = [new_tokens[id] for id in actual_new_ids]
self.vocab_size_base = vocab_size
self.vocab_size = self.vocab_size_base + len(self.added_tokens_list)
self.fname_tokenizer = fname_tokenizer
def sentencepiece_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
tokenizer = self.sentencepiece_tokenizer
for i in range(tokenizer.vocab_size()):
piece = tokenizer.IdToPiece(i)
text = piece.encode("utf-8")
score: float = tokenizer.GetScore(i)
toktype = gguf.TokenType.NORMAL
if tokenizer.IsUnknown(i):
toktype = gguf.TokenType.UNKNOWN
if tokenizer.IsControl(i):
toktype = gguf.TokenType.CONTROL
# NOTE: I think added_tokens are user defined.
# ref: https://github.com/google/sentencepiece/blob/master/src/sentencepiece_model.proto
# if tokenizer.is_user_defined(i): toktype = gguf.TokenType.USER_DEFINED
if tokenizer.IsUnused(i):
toktype = gguf.TokenType.UNUSED
if tokenizer.IsByte(i):
toktype = gguf.TokenType.BYTE
yield text, score, toktype
def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
for text in self.added_tokens_list:
score = -1000.0
yield text.encode("utf-8"), score, gguf.TokenType.USER_DEFINED
def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
yield from self.sentencepiece_tokens()
yield from self.added_tokens()
def __repr__(self) -> str:
return f"<SentencePieceVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
class LlamaHfVocab(Vocab):
tokenizer_model = "llama"
name = "hfft"
def __init__(self, base_path: Path):
fname_tokenizer = base_path / 'tokenizer.json'
# if this fails, FileNotFoundError propagates to caller
with open(fname_tokenizer, encoding='utf-8') as f:
tokenizer_json = json.load(f)
# pre-check so we know if we need transformers
tokenizer_model: dict[str, Any] = tokenizer_json['model']
is_llama3 = (
tokenizer_model['type'] == 'BPE' and tokenizer_model.get('ignore_merges', False)
and not tokenizer_model.get('byte_fallback', True)
)
if is_llama3:
raise TypeError('Llama 3 must be converted with BpeVocab')
if not is_llama3 and (
tokenizer_model['type'] != 'BPE' or not tokenizer_model.get('byte_fallback', False)
or tokenizer_json['decoder']['type'] != 'Sequence'
):
raise FileNotFoundError('Cannot find Llama BPE tokenizer')
try:
from transformers import AutoTokenizer
except ImportError as e:
raise ImportError(
"To use LlamaHfVocab, please install the `transformers` package. "
"You can install it with `pip install transformers`."
) from e
# Allow the tokenizer to default to slow or fast versions.
# Explicitly set tokenizer to use local paths.
self.tokenizer = AutoTokenizer.from_pretrained(
base_path,
cache_dir=base_path,
local_files_only=True,
)
assert self.tokenizer.is_fast # assume tokenizer.json is used
# Initialize lists and dictionaries for added tokens
self.added_tokens_list = []
self.added_tokens_dict = dict()
self.added_tokens_ids = set()
# Process added tokens
for tok, tokidx in sorted(
self.tokenizer.get_added_vocab().items(), key=lambda x: x[1]
):
# Only consider added tokens that are not in the base vocabulary
if tokidx >= self.tokenizer.vocab_size:
self.added_tokens_list.append(tok)
self.added_tokens_dict[tok] = tokidx
self.added_tokens_ids.add(tokidx)
# Store special tokens and their IDs
self.specials = {
tok: self.tokenizer.get_vocab()[tok]
for tok in self.tokenizer.all_special_tokens
}
self.special_ids = set(self.tokenizer.all_special_ids)
# Set vocabulary sizes
self.vocab_size_base = self.tokenizer.vocab_size
self.vocab_size = self.vocab_size_base + len(self.added_tokens_list)
self.fname_tokenizer = fname_tokenizer
def hf_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
reverse_vocab = {
id: encoded_tok for encoded_tok, id in self.tokenizer.get_vocab().items()
}
for token_id in range(self.vocab_size_base):
# Skip processing added tokens here
if token_id in self.added_tokens_ids:
continue
# Convert token text to bytes
token_text = reverse_vocab[token_id].encode("utf-8")
# Yield token text, score, and type
yield token_text, self.get_token_score(token_id), self.get_token_type(
token_id, token_text, self.special_ids # Reuse already stored special IDs
)
def get_token_type(self, token_id: int, token_text: bytes, special_ids: set[int]) -> gguf.TokenType:
# Special case for byte tokens
if re.fullmatch(br"<0x[0-9A-Fa-f]{2}>", token_text):
return gguf.TokenType.BYTE
# Determine token type based on whether it's a special token
return gguf.TokenType.CONTROL if token_id in special_ids else gguf.TokenType.NORMAL
def get_token_score(self, token_id: int) -> float:
# Placeholder for actual logic to determine the token's score
# This needs to be implemented based on specific requirements
return -1000.0 # Default score
def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
for text in self.added_tokens_list:
if text in self.specials:
toktype = self.get_token_type(self.specials[text], b'', self.special_ids)
score = self.get_token_score(self.specials[text])
else:
toktype = gguf.TokenType.USER_DEFINED
score = -1000.0
yield text.encode("utf-8"), score, toktype
def has_newline_token(self):
return "<0x0A>" in self.tokenizer.vocab or "\n" in self.tokenizer.vocab
def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
yield from self.hf_tokens()
yield from self.added_tokens()
def __repr__(self) -> str:
return f"<LlamaHfVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"

688
packages_3rdparty/webui_lora_collection/LICENSE.txt vendored Executable file
View File

@@ -0,0 +1,688 @@
GNU AFFERO GENERAL PUBLIC LICENSE
Version 3, 19 November 2007
Copyright (c) 2023 AUTOMATIC1111
Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
Everyone is permitted to copy and distribute verbatim copies
of this license document, but changing it is not allowed.
Preamble
The GNU Affero General Public License is a free, copyleft license for
software and other kinds of works, specifically designed to ensure
cooperation with the community in the case of network server software.
The licenses for most software and other practical works are designed
to take away your freedom to share and change the works. By contrast,
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
software for all its users.
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.
Developers that use our General Public Licenses protect your rights
with two steps: (1) assert copyright on the software, and (2) offer
you this License which gives you legal permission to copy, distribute
and/or modify the software.
A secondary benefit of defending all users' freedom is that
improvements made in alternate versions of the program, if they
receive widespread use, become available for other developers to
incorporate. Many developers of free software are heartened and
encouraged by the resulting cooperation. However, in the case of
software used on network servers, this result may fail to come about.
The GNU General Public License permits making a modified version and
letting the public access it on a server without ever releasing its
source code to the public.
The GNU Affero General Public License is designed specifically to
ensure that, in such cases, the modified source code becomes available
to the community. It requires the operator of a network server to
provide the source code of the modified version running there to the
users of that server. Therefore, public use of a modified version, on
a publicly accessible server, gives the public access to the source
code of the modified version.
An older license, called the Affero General Public License and
published by Affero, was designed to accomplish similar goals. This is
a different license, not a version of the Affero GPL, but Affero has
released a new version of the Affero GPL which permits relicensing under
this license.
The precise terms and conditions for copying, distribution and
modification follow.
TERMS AND CONDITIONS
0. Definitions.
"This License" refers to version 3 of the GNU Affero General Public License.
"Copyright" also means copyright-like laws that apply to other kinds of
works, such as semiconductor masks.
"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.
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.
A "covered work" means either the unmodified Program or a work based
on the Program.
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 "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.
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.
1. Source Code.
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.
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.
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 "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.
The Corresponding Source need not include anything that users
can regenerate automatically from other parts of the Corresponding
Source.
The Corresponding Source for a work in source code form is that
same work.
2. Basic Permissions.
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.
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.
Conveying under any other circumstances is permitted solely under
the conditions stated below. Sublicensing is not allowed; section 10
makes it unnecessary.
3. Protecting Users' Legal Rights From Anti-Circumvention Law.
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.
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.
4. Conveying Verbatim Copies.
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.
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.
5. Conveying Modified Source Versions.
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:
a) The work must carry prominent notices stating that you modified
it, and giving a relevant date.
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".
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.
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.
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.
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 AGPL, see
<https://www.gnu.org/licenses/>.
---------------------------------Facebook BNB-------------------------------
MIT License
Copyright (c) Facebook, Inc. and its affiliates.
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.

2
packages_3rdparty/webui_lora_collection/lora.py vendored Executable file
View File

@@ -0,0 +1,2 @@
# TODO: Implement API

68
packages_3rdparty/webui_lora_collection/lyco_helpers.py vendored Executable file
View File

@@ -0,0 +1,68 @@
import torch
def make_weight_cp(t, wa, wb):
temp = torch.einsum('i j k l, j r -> i r k l', t, wb)
return torch.einsum('i j k l, i r -> r j k l', temp, wa)
def rebuild_conventional(up, down, shape, dyn_dim=None):
up = up.reshape(up.size(0), -1)
down = down.reshape(down.size(0), -1)
if dyn_dim is not None:
up = up[:, :dyn_dim]
down = down[:dyn_dim, :]
return (up @ down).reshape(shape)
def rebuild_cp_decomposition(up, down, mid):
up = up.reshape(up.size(0), -1)
down = down.reshape(down.size(0), -1)
return torch.einsum('n m k l, i n, m j -> i j k l', mid, up, down)
# copied from https://github.com/KohakuBlueleaf/LyCORIS/blob/dev/lycoris/modules/lokr.py
def factorization(dimension: int, factor:int=-1) -> tuple[int, int]:
'''
return a tuple of two value of input dimension decomposed by the number closest to factor
second value is higher or equal than first value.
In LoRA with Kroneckor Product, first value is a value for weight scale.
secon value is a value for weight.
Because of non-commutative property, A⊗B ≠ B⊗A. Meaning of two matrices is slightly different.
examples)
factor
-1 2 4 8 16 ...
127 -> 1, 127 127 -> 1, 127 127 -> 1, 127 127 -> 1, 127 127 -> 1, 127
128 -> 8, 16 128 -> 2, 64 128 -> 4, 32 128 -> 8, 16 128 -> 8, 16
250 -> 10, 25 250 -> 2, 125 250 -> 2, 125 250 -> 5, 50 250 -> 10, 25
360 -> 8, 45 360 -> 2, 180 360 -> 4, 90 360 -> 8, 45 360 -> 12, 30
512 -> 16, 32 512 -> 2, 256 512 -> 4, 128 512 -> 8, 64 512 -> 16, 32
1024 -> 32, 32 1024 -> 2, 512 1024 -> 4, 256 1024 -> 8, 128 1024 -> 16, 64
'''
if factor > 0 and (dimension % factor) == 0:
m = factor
n = dimension // factor
if m > n:
n, m = m, n
return m, n
if factor < 0:
factor = dimension
m, n = 1, dimension
length = m + n
while m<n:
new_m = m + 1
while dimension%new_m != 0:
new_m += 1
new_n = dimension // new_m
if new_m + new_n > length or new_m>factor:
break
else:
m, n = new_m, new_n
if m > n:
n, m = m, n
return m, n

228
packages_3rdparty/webui_lora_collection/network.py vendored Executable file
View File

@@ -0,0 +1,228 @@
from __future__ import annotations
import os
from collections import namedtuple
import enum
import torch.nn as nn
import torch.nn.functional as F
from modules import sd_models, cache, errors, hashes, shared
import modules.models.sd3.mmdit
NetworkWeights = namedtuple('NetworkWeights', ['network_key', 'sd_key', 'w', 'sd_module'])
metadata_tags_order = {"ss_sd_model_name": 1, "ss_resolution": 2, "ss_clip_skip": 3, "ss_num_train_images": 10, "ss_tag_frequency": 20}
class SdVersion(enum.Enum):
Unknown = 1
SD1 = 2
SD2 = 3
SDXL = 4
class NetworkOnDisk:
def __init__(self, name, filename):
self.name = name
self.filename = filename
self.metadata = {}
self.is_safetensors = os.path.splitext(filename)[1].lower() == ".safetensors"
def read_metadata():
metadata = sd_models.read_metadata_from_safetensors(filename)
return metadata
if self.is_safetensors:
try:
self.metadata = cache.cached_data_for_file('safetensors-metadata', "lora/" + self.name, filename, read_metadata)
except Exception as e:
errors.display(e, f"reading lora {filename}")
if self.metadata:
m = {}
for k, v in sorted(self.metadata.items(), key=lambda x: metadata_tags_order.get(x[0], 999)):
m[k] = v
self.metadata = m
self.alias = self.metadata.get('ss_output_name', self.name)
self.hash = None
self.shorthash = None
self.set_hash(
self.metadata.get('sshs_model_hash') or
hashes.sha256_from_cache(self.filename, "lora/" + self.name, use_addnet_hash=self.is_safetensors) or
''
)
self.sd_version = self.detect_version()
def detect_version(self):
if str(self.metadata.get('ss_base_model_version', "")).startswith("sdxl_"):
return SdVersion.SDXL
elif str(self.metadata.get('ss_v2', "")) == "True":
return SdVersion.SD2
elif len(self.metadata):
return SdVersion.SD1
return SdVersion.Unknown
def set_hash(self, v):
self.hash = v
self.shorthash = self.hash[0:12]
if self.shorthash:
import networks
networks.available_network_hash_lookup[self.shorthash] = self
def read_hash(self):
if not self.hash:
self.set_hash(hashes.sha256(self.filename, "lora/" + self.name, use_addnet_hash=self.is_safetensors) or '')
def get_alias(self):
import networks
if shared.opts.lora_preferred_name == "Filename" or self.alias.lower() in networks.forbidden_network_aliases:
return self.name
else:
return self.alias
class Network: # LoraModule
def __init__(self, name, network_on_disk: NetworkOnDisk):
self.name = name
self.network_on_disk = network_on_disk
self.te_multiplier = 1.0
self.unet_multiplier = 1.0
self.dyn_dim = None
self.modules = {}
self.bundle_embeddings = {}
self.mtime = None
self.mentioned_name = None
"""the text that was used to add the network to prompt - can be either name or an alias"""
class ModuleType:
def create_module(self, net: Network, weights: NetworkWeights) -> Network | None:
return None
class NetworkModule:
def __init__(self, net: Network, weights: NetworkWeights):
self.network = net
self.network_key = weights.network_key
self.sd_key = weights.sd_key
self.sd_module = weights.sd_module
if isinstance(self.sd_module, modules.models.sd3.mmdit.QkvLinear):
s = self.sd_module.weight.shape
self.shape = (s[0] // 3, s[1])
elif hasattr(self.sd_module, 'weight'):
self.shape = self.sd_module.weight.shape
elif isinstance(self.sd_module, nn.MultiheadAttention):
# For now, only self-attn use Pytorch's MHA
# So assume all qkvo proj have same shape
self.shape = self.sd_module.out_proj.weight.shape
else:
self.shape = None
self.ops = None
self.extra_kwargs = {}
if isinstance(self.sd_module, nn.Conv2d):
self.ops = F.conv2d
self.extra_kwargs = {
'stride': self.sd_module.stride,
'padding': self.sd_module.padding
}
elif isinstance(self.sd_module, nn.Linear):
self.ops = F.linear
elif isinstance(self.sd_module, nn.LayerNorm):
self.ops = F.layer_norm
self.extra_kwargs = {
'normalized_shape': self.sd_module.normalized_shape,
'eps': self.sd_module.eps
}
elif isinstance(self.sd_module, nn.GroupNorm):
self.ops = F.group_norm
self.extra_kwargs = {
'num_groups': self.sd_module.num_groups,
'eps': self.sd_module.eps
}
self.dim = None
self.bias = weights.w.get("bias")
self.alpha = weights.w["alpha"].item() if "alpha" in weights.w else None
self.scale = weights.w["scale"].item() if "scale" in weights.w else None
self.dora_scale = weights.w.get("dora_scale", None)
self.dora_norm_dims = len(self.shape) - 1
def multiplier(self):
if 'transformer' in self.sd_key[:20]:
return self.network.te_multiplier
else:
return self.network.unet_multiplier
def calc_scale(self):
if self.scale is not None:
return self.scale
if self.dim is not None and self.alpha is not None:
return self.alpha / self.dim
return 1.0
def apply_weight_decompose(self, updown, orig_weight):
# Match the device/dtype
orig_weight = orig_weight.to(updown.dtype)
dora_scale = self.dora_scale.to(device=orig_weight.device, dtype=updown.dtype)
updown = updown.to(orig_weight.device)
merged_scale1 = updown + orig_weight
merged_scale1_norm = (
merged_scale1.transpose(0, 1)
.reshape(merged_scale1.shape[1], -1)
.norm(dim=1, keepdim=True)
.reshape(merged_scale1.shape[1], *[1] * self.dora_norm_dims)
.transpose(0, 1)
)
dora_merged = (
merged_scale1 * (dora_scale / merged_scale1_norm)
)
final_updown = dora_merged - orig_weight
return final_updown
def finalize_updown(self, updown, orig_weight, output_shape, ex_bias=None):
if self.bias is not None:
updown = updown.reshape(self.bias.shape)
updown += self.bias.to(orig_weight.device, dtype=updown.dtype)
updown = updown.reshape(output_shape)
if len(output_shape) == 4:
updown = updown.reshape(output_shape)
if orig_weight.size().numel() == updown.size().numel():
updown = updown.reshape(orig_weight.shape)
if ex_bias is not None:
ex_bias = ex_bias * self.multiplier()
updown = updown * self.calc_scale()
if self.dora_scale is not None:
updown = self.apply_weight_decompose(updown, orig_weight)
return updown * self.multiplier(), ex_bias
def calc_updown(self, target):
raise NotImplementedError()
def forward(self, x, y):
"""A general forward implementation for all modules"""
if self.ops is None:
raise NotImplementedError()
else:
updown, ex_bias = self.calc_updown(self.sd_module.weight)
return y + self.ops(x, weight=updown, bias=ex_bias, **self.extra_kwargs)

27
packages_3rdparty/webui_lora_collection/network_full.py vendored Executable file
View File

@@ -0,0 +1,27 @@
import network
class ModuleTypeFull(network.ModuleType):
def create_module(self, net: network.Network, weights: network.NetworkWeights):
if all(x in weights.w for x in ["diff"]):
return NetworkModuleFull(net, weights)
return None
class NetworkModuleFull(network.NetworkModule):
def __init__(self, net: network.Network, weights: network.NetworkWeights):
super().__init__(net, weights)
self.weight = weights.w.get("diff")
self.ex_bias = weights.w.get("diff_b")
def calc_updown(self, orig_weight):
output_shape = self.weight.shape
updown = self.weight.to(orig_weight.device)
if self.ex_bias is not None:
ex_bias = self.ex_bias.to(orig_weight.device)
else:
ex_bias = None
return self.finalize_updown(updown, orig_weight, output_shape, ex_bias)

33
packages_3rdparty/webui_lora_collection/network_glora.py vendored Executable file
View File

@@ -0,0 +1,33 @@
import network
class ModuleTypeGLora(network.ModuleType):
def create_module(self, net: network.Network, weights: network.NetworkWeights):
if all(x in weights.w for x in ["a1.weight", "a2.weight", "alpha", "b1.weight", "b2.weight"]):
return NetworkModuleGLora(net, weights)
return None
# adapted from https://github.com/KohakuBlueleaf/LyCORIS
class NetworkModuleGLora(network.NetworkModule):
def __init__(self, net: network.Network, weights: network.NetworkWeights):
super().__init__(net, weights)
if hasattr(self.sd_module, 'weight'):
self.shape = self.sd_module.weight.shape
self.w1a = weights.w["a1.weight"]
self.w1b = weights.w["b1.weight"]
self.w2a = weights.w["a2.weight"]
self.w2b = weights.w["b2.weight"]
def calc_updown(self, orig_weight):
w1a = self.w1a.to(orig_weight.device)
w1b = self.w1b.to(orig_weight.device)
w2a = self.w2a.to(orig_weight.device)
w2b = self.w2b.to(orig_weight.device)
output_shape = [w1a.size(0), w1b.size(1)]
updown = ((w2b @ w1b) + ((orig_weight.to(dtype = w1a.dtype) @ w2a) @ w1a))
return self.finalize_updown(updown, orig_weight, output_shape)

55
packages_3rdparty/webui_lora_collection/network_hada.py vendored Executable file
View File

@@ -0,0 +1,55 @@
import lyco_helpers
import network
class ModuleTypeHada(network.ModuleType):
def create_module(self, net: network.Network, weights: network.NetworkWeights):
if all(x in weights.w for x in ["hada_w1_a", "hada_w1_b", "hada_w2_a", "hada_w2_b"]):
return NetworkModuleHada(net, weights)
return None
class NetworkModuleHada(network.NetworkModule):
def __init__(self, net: network.Network, weights: network.NetworkWeights):
super().__init__(net, weights)
if hasattr(self.sd_module, 'weight'):
self.shape = self.sd_module.weight.shape
self.w1a = weights.w["hada_w1_a"]
self.w1b = weights.w["hada_w1_b"]
self.dim = self.w1b.shape[0]
self.w2a = weights.w["hada_w2_a"]
self.w2b = weights.w["hada_w2_b"]
self.t1 = weights.w.get("hada_t1")
self.t2 = weights.w.get("hada_t2")
def calc_updown(self, orig_weight):
w1a = self.w1a.to(orig_weight.device)
w1b = self.w1b.to(orig_weight.device)
w2a = self.w2a.to(orig_weight.device)
w2b = self.w2b.to(orig_weight.device)
output_shape = [w1a.size(0), w1b.size(1)]
if self.t1 is not None:
output_shape = [w1a.size(1), w1b.size(1)]
t1 = self.t1.to(orig_weight.device)
updown1 = lyco_helpers.make_weight_cp(t1, w1a, w1b)
output_shape += t1.shape[2:]
else:
if len(w1b.shape) == 4:
output_shape += w1b.shape[2:]
updown1 = lyco_helpers.rebuild_conventional(w1a, w1b, output_shape)
if self.t2 is not None:
t2 = self.t2.to(orig_weight.device)
updown2 = lyco_helpers.make_weight_cp(t2, w2a, w2b)
else:
updown2 = lyco_helpers.rebuild_conventional(w2a, w2b, output_shape)
updown = updown1 * updown2
return self.finalize_updown(updown, orig_weight, output_shape)

30
packages_3rdparty/webui_lora_collection/network_ia3.py vendored Executable file
View File

@@ -0,0 +1,30 @@
import network
class ModuleTypeIa3(network.ModuleType):
def create_module(self, net: network.Network, weights: network.NetworkWeights):
if all(x in weights.w for x in ["weight"]):
return NetworkModuleIa3(net, weights)
return None
class NetworkModuleIa3(network.NetworkModule):
def __init__(self, net: network.Network, weights: network.NetworkWeights):
super().__init__(net, weights)
self.w = weights.w["weight"]
self.on_input = weights.w["on_input"].item()
def calc_updown(self, orig_weight):
w = self.w.to(orig_weight.device)
output_shape = [w.size(0), orig_weight.size(1)]
if self.on_input:
output_shape.reverse()
else:
w = w.reshape(-1, 1)
updown = orig_weight * w
return self.finalize_updown(updown, orig_weight, output_shape)

64
packages_3rdparty/webui_lora_collection/network_lokr.py vendored Executable file
View File

@@ -0,0 +1,64 @@
import torch
import lyco_helpers
import network
class ModuleTypeLokr(network.ModuleType):
def create_module(self, net: network.Network, weights: network.NetworkWeights):
has_1 = "lokr_w1" in weights.w or ("lokr_w1_a" in weights.w and "lokr_w1_b" in weights.w)
has_2 = "lokr_w2" in weights.w or ("lokr_w2_a" in weights.w and "lokr_w2_b" in weights.w)
if has_1 and has_2:
return NetworkModuleLokr(net, weights)
return None
def make_kron(orig_shape, w1, w2):
if len(w2.shape) == 4:
w1 = w1.unsqueeze(2).unsqueeze(2)
w2 = w2.contiguous()
return torch.kron(w1, w2).reshape(orig_shape)
class NetworkModuleLokr(network.NetworkModule):
def __init__(self, net: network.Network, weights: network.NetworkWeights):
super().__init__(net, weights)
self.w1 = weights.w.get("lokr_w1")
self.w1a = weights.w.get("lokr_w1_a")
self.w1b = weights.w.get("lokr_w1_b")
self.dim = self.w1b.shape[0] if self.w1b is not None else self.dim
self.w2 = weights.w.get("lokr_w2")
self.w2a = weights.w.get("lokr_w2_a")
self.w2b = weights.w.get("lokr_w2_b")
self.dim = self.w2b.shape[0] if self.w2b is not None else self.dim
self.t2 = weights.w.get("lokr_t2")
def calc_updown(self, orig_weight):
if self.w1 is not None:
w1 = self.w1.to(orig_weight.device)
else:
w1a = self.w1a.to(orig_weight.device)
w1b = self.w1b.to(orig_weight.device)
w1 = w1a @ w1b
if self.w2 is not None:
w2 = self.w2.to(orig_weight.device)
elif self.t2 is None:
w2a = self.w2a.to(orig_weight.device)
w2b = self.w2b.to(orig_weight.device)
w2 = w2a @ w2b
else:
t2 = self.t2.to(orig_weight.device)
w2a = self.w2a.to(orig_weight.device)
w2b = self.w2b.to(orig_weight.device)
w2 = lyco_helpers.make_weight_cp(t2, w2a, w2b)
output_shape = [w1.size(0) * w2.size(0), w1.size(1) * w2.size(1)]
if len(orig_weight.shape) == 4:
output_shape = orig_weight.shape
updown = make_kron(output_shape, w1, w2)
return self.finalize_updown(updown, orig_weight, output_shape)

94
packages_3rdparty/webui_lora_collection/network_lora.py vendored Executable file
View File

@@ -0,0 +1,94 @@
import torch
import lyco_helpers
import modules.models.sd3.mmdit
import network
from modules import devices
class ModuleTypeLora(network.ModuleType):
def create_module(self, net: network.Network, weights: network.NetworkWeights):
if all(x in weights.w for x in ["lora_up.weight", "lora_down.weight"]):
return NetworkModuleLora(net, weights)
if all(x in weights.w for x in ["lora_A.weight", "lora_B.weight"]):
w = weights.w.copy()
weights.w.clear()
weights.w.update({"lora_up.weight": w["lora_B.weight"], "lora_down.weight": w["lora_A.weight"]})
return NetworkModuleLora(net, weights)
return None
class NetworkModuleLora(network.NetworkModule):
def __init__(self, net: network.Network, weights: network.NetworkWeights):
super().__init__(net, weights)
self.up_model = self.create_module(weights.w, "lora_up.weight")
self.down_model = self.create_module(weights.w, "lora_down.weight")
self.mid_model = self.create_module(weights.w, "lora_mid.weight", none_ok=True)
self.dim = weights.w["lora_down.weight"].shape[0]
def create_module(self, weights, key, none_ok=False):
weight = weights.get(key)
if weight is None and none_ok:
return None
is_linear = type(self.sd_module) in [torch.nn.Linear, torch.nn.modules.linear.NonDynamicallyQuantizableLinear, torch.nn.MultiheadAttention, modules.models.sd3.mmdit.QkvLinear]
is_conv = type(self.sd_module) in [torch.nn.Conv2d]
if is_linear:
weight = weight.reshape(weight.shape[0], -1)
module = torch.nn.Linear(weight.shape[1], weight.shape[0], bias=False)
elif is_conv and key == "lora_down.weight" or key == "dyn_up":
if len(weight.shape) == 2:
weight = weight.reshape(weight.shape[0], -1, 1, 1)
if weight.shape[2] != 1 or weight.shape[3] != 1:
module = torch.nn.Conv2d(weight.shape[1], weight.shape[0], self.sd_module.kernel_size, self.sd_module.stride, self.sd_module.padding, bias=False)
else:
module = torch.nn.Conv2d(weight.shape[1], weight.shape[0], (1, 1), bias=False)
elif is_conv and key == "lora_mid.weight":
module = torch.nn.Conv2d(weight.shape[1], weight.shape[0], self.sd_module.kernel_size, self.sd_module.stride, self.sd_module.padding, bias=False)
elif is_conv and key == "lora_up.weight" or key == "dyn_down":
module = torch.nn.Conv2d(weight.shape[1], weight.shape[0], (1, 1), bias=False)
else:
raise AssertionError(f'Lora layer {self.network_key} matched a layer with unsupported type: {type(self.sd_module).__name__}')
with torch.no_grad():
if weight.shape != module.weight.shape:
weight = weight.reshape(module.weight.shape)
module.weight.copy_(weight)
module.to(device=devices.cpu, dtype=devices.dtype)
module.weight.requires_grad_(False)
return module
def calc_updown(self, orig_weight):
up = self.up_model.weight.to(orig_weight.device)
down = self.down_model.weight.to(orig_weight.device)
output_shape = [up.size(0), down.size(1)]
if self.mid_model is not None:
# cp-decomposition
mid = self.mid_model.weight.to(orig_weight.device)
updown = lyco_helpers.rebuild_cp_decomposition(up, down, mid)
output_shape += mid.shape[2:]
else:
if len(down.shape) == 4:
output_shape += down.shape[2:]
updown = lyco_helpers.rebuild_conventional(up, down, output_shape, self.network.dyn_dim)
return self.finalize_updown(updown, orig_weight, output_shape)
def forward(self, x, y):
self.up_model.to(device=devices.device)
self.down_model.to(device=devices.device)
return y + self.up_model(self.down_model(x)) * self.multiplier() * self.calc_scale()

28
packages_3rdparty/webui_lora_collection/network_norm.py vendored Executable file
View File

@@ -0,0 +1,28 @@
import network
class ModuleTypeNorm(network.ModuleType):
def create_module(self, net: network.Network, weights: network.NetworkWeights):
if all(x in weights.w for x in ["w_norm", "b_norm"]):
return NetworkModuleNorm(net, weights)
return None
class NetworkModuleNorm(network.NetworkModule):
def __init__(self, net: network.Network, weights: network.NetworkWeights):
super().__init__(net, weights)
self.w_norm = weights.w.get("w_norm")
self.b_norm = weights.w.get("b_norm")
def calc_updown(self, orig_weight):
output_shape = self.w_norm.shape
updown = self.w_norm.to(orig_weight.device)
if self.b_norm is not None:
ex_bias = self.b_norm.to(orig_weight.device)
else:
ex_bias = None
return self.finalize_updown(updown, orig_weight, output_shape, ex_bias)

119
packages_3rdparty/webui_lora_collection/network_oft.py vendored Executable file
View File

@@ -0,0 +1,119 @@
import torch
import network
from einops import rearrange
class ModuleTypeOFT(network.ModuleType):
def create_module(self, net: network.Network, weights: network.NetworkWeights):
if all(x in weights.w for x in ["oft_blocks"]) or all(x in weights.w for x in ["oft_diag"]):
return NetworkModuleOFT(net, weights)
return None
# TODO: Convert to forge patcher
# Supports both kohya-ss' implementation of COFT https://github.com/kohya-ss/sd-scripts/blob/main/networks/oft.py
# and KohakuBlueleaf's implementation of OFT/COFT https://github.com/KohakuBlueleaf/LyCORIS/blob/dev/lycoris/modules/diag_oft.py
class NetworkModuleOFT(network.NetworkModule):
def __init__(self, net: network.Network, weights: network.NetworkWeights):
super().__init__(net, weights)
self.lin_module = None
self.org_module: list[torch.Module] = [self.sd_module]
self.scale = 1.0
self.is_R = False
self.is_boft = False
# kohya-ss/New LyCORIS OFT/BOFT
if "oft_blocks" in weights.w.keys():
self.oft_blocks = weights.w["oft_blocks"] # (num_blocks, block_size, block_size)
self.alpha = weights.w.get("alpha", None) # alpha is constraint
self.dim = self.oft_blocks.shape[0] # lora dim
# Old LyCORIS OFT
elif "oft_diag" in weights.w.keys():
self.is_R = True
self.oft_blocks = weights.w["oft_diag"]
# self.alpha is unused
self.dim = self.oft_blocks.shape[1] # (num_blocks, block_size, block_size)
is_linear = type(self.sd_module) in [torch.nn.Linear, torch.nn.modules.linear.NonDynamicallyQuantizableLinear]
is_conv = type(self.sd_module) in [torch.nn.Conv2d]
is_other_linear = type(self.sd_module) in [torch.nn.MultiheadAttention] # unsupported
if is_linear:
self.out_dim = self.sd_module.out_features
elif is_conv:
self.out_dim = self.sd_module.out_channels
elif is_other_linear:
self.out_dim = self.sd_module.embed_dim
# LyCORIS BOFT
if self.oft_blocks.dim() == 4:
self.is_boft = True
self.rescale = weights.w.get('rescale', None)
if self.rescale is not None and not is_other_linear:
self.rescale = self.rescale.reshape(-1, *[1]*(self.org_module[0].weight.dim() - 1))
self.num_blocks = self.dim
self.block_size = self.out_dim // self.dim
self.constraint = (0 if self.alpha is None else self.alpha) * self.out_dim
if self.is_R:
self.constraint = None
self.block_size = self.dim
self.num_blocks = self.out_dim // self.dim
elif self.is_boft:
self.boft_m = self.oft_blocks.shape[0]
self.num_blocks = self.oft_blocks.shape[1]
self.block_size = self.oft_blocks.shape[2]
self.boft_b = self.block_size
def calc_updown(self, orig_weight):
oft_blocks = self.oft_blocks.to(orig_weight.device)
eye = torch.eye(self.block_size, device=oft_blocks.device)
if not self.is_R:
block_Q = oft_blocks - oft_blocks.transpose(-1, -2) # ensure skew-symmetric orthogonal matrix
if self.constraint != 0:
norm_Q = torch.norm(block_Q.flatten())
new_norm_Q = torch.clamp(norm_Q, max=self.constraint.to(oft_blocks.device))
block_Q = block_Q * ((new_norm_Q + 1e-8) / (norm_Q + 1e-8))
oft_blocks = torch.matmul(eye + block_Q, (eye - block_Q).float().inverse())
R = oft_blocks.to(orig_weight.device)
if not self.is_boft:
# This errors out for MultiheadAttention, might need to be handled up-stream
merged_weight = rearrange(orig_weight, '(k n) ... -> k n ...', k=self.num_blocks, n=self.block_size)
merged_weight = torch.einsum(
'k n m, k n ... -> k m ...',
R,
merged_weight
)
merged_weight = rearrange(merged_weight, 'k m ... -> (k m) ...')
else:
# TODO: determine correct value for scale
scale = 1.0
m = self.boft_m
b = self.boft_b
r_b = b // 2
inp = orig_weight
for i in range(m):
bi = R[i] # b_num, b_size, b_size
if i == 0:
# Apply multiplier/scale and rescale into first weight
bi = bi * scale + (1 - scale) * eye
inp = rearrange(inp, "(c g k) ... -> (c k g) ...", g=2, k=2**i * r_b)
inp = rearrange(inp, "(d b) ... -> d b ...", b=b)
inp = torch.einsum("b i j, b j ... -> b i ...", bi, inp)
inp = rearrange(inp, "d b ... -> (d b) ...")
inp = rearrange(inp, "(c k g) ... -> (c g k) ...", g=2, k=2**i * r_b)
merged_weight = inp
# Rescale mechanism
if self.rescale is not None:
merged_weight = self.rescale.to(merged_weight) * merged_weight
updown = merged_weight.to(orig_weight.device) - orig_weight.to(merged_weight.dtype)
output_shape = orig_weight.shape
return self.finalize_updown(updown, orig_weight, output_shape)