Firehawk52-backup/llama
28
0
Fork 0
mirror of https://github.com/meta-llama/llama.git synced 2026-01-15 08:22:55 -03:00
Code Issues Actions Packages Projects Releases Wiki Activity

llama 2

Browse Source
This commit is contained in:
Ubuntu
2023-07-18 15:57:25 +00:00
parent 57b0eb62de
commit 6d4c0c290a
15 changed files with 839 additions and 921 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
CONTRIBUTING.md
View File

@@ -1,4 +1,4 @@
# Contributing to LLaMA
# Contributing to Llama
We want to make contributing to this project as easy and transparent as
possible.
@@ -27,5 +27,5 @@ disclosure of security bugs. In those cases, please go through the process
outlined on that page and do not file a public issue.
## License
By contributing to LLaMA, you agree that your contributions will be licensed
By contributing to Llama, you agree that your contributions will be licensed
under the LICENSE file in the root directory of this source tree.

766
LICENSE
View File

@@ -1,674 +1,126 @@
GNU GENERAL PUBLIC LICENSE
Version 3, 29 June 2007
LLAMA 2 COMMUNITY LICENSE AGREEMENT
Llama 2 Version Release Date: July 18, 2023
Copyright (C) 2007 Free Software Foundation, Inc. <http://fsf.org/>
Everyone is permitted to copy and distribute verbatim copies
of this license document, but changing it is not allowed.
"Agreement" means the terms and conditions for use, reproduction, distribution and
modification of the Llama Materials set forth herein.
Preamble
"Documentation" means the specifications, manuals and documentation
accompanying Llama 2 distributed by Meta at ai.meta.com/resources/models-and-
libraries/llama-downloads/.
The GNU General Public License is a free, copyleft license for
software and other kinds of works.
"Licensee" or "you" means you, or your employer or any other person or entity (if
you are entering into this Agreement on such person or entity's behalf), of the age
required under applicable laws, rules or regulations to provide legal consent and that
has legal authority to bind your employer or such other person or entity if you are
entering in this Agreement on their behalf.
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.
"Llama 2" means the foundational large language models and software and
algorithms, including machine-learning model code, trained model weights,
inference-enabling code, training-enabling code, fine-tuning enabling code and other
elements of the foregoing distributed by Meta at ai.meta.com/resources/models-and-
libraries/llama-downloads/.
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.
"Llama Materials" means, collectively, Meta's proprietary Llama 2 and
Documentation (and any portion thereof) made available under this Agreement.
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.
"Meta" or "we" means Meta Platforms Ireland Limited (if you are located in or, if you
are an entity, your principal place of business is in the EEA or Switzerland) and Meta
Platforms, Inc. (if you are located outside of the EEA or Switzerland).
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.
By clicking "I Accept" below or by using or distributing any portion or element of the
Llama Materials, you agree to be bound by this Agreement.
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.
1. License Rights and Redistribution.
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.
a. Grant of Rights. You are granted a non-exclusive, worldwide, non-
transferable and royalty-free limited license under Meta's intellectual property or
other rights owned by Meta embodied in the Llama Materials to use, reproduce,
distribute, copy, create derivative works of, and make modifications to the Llama
Materials.
b. Redistribution and Use.
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.
i. If you distribute or make the Llama Materials, or any derivative works
thereof, available to a third party, you shall provide a copy of this Agreement to such
third party.
ii. If you receive Llama Materials, or any derivative works thereof, from
a Licensee as part of an integrated end user product, then Section 2 of this
Agreement will not apply to you.
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.
iii. You must retain in all copies of the Llama Materials that you
distribute the following attribution notice within a "Notice" text file distributed as a
part of such copies: "Llama 2 is licensed under the LLAMA 2 Community License,
Copyright (c) Meta Platforms, Inc. All Rights Reserved."
The precise terms and conditions for copying, distribution and
modification follow.
iv. Your use of the Llama Materials must comply with applicable laws
and regulations (including trade compliance laws and regulations) and adhere to the
Acceptable Use Policy for the Llama Materials (available at
https://ai.meta.com/llama/use-policy), which is hereby incorporated by reference into
this Agreement.
TERMS AND CONDITIONS
v. You will not use the Llama Materials or any output or results of the
Llama Materials to improve any other large language model (excluding Llama 2 or
derivative works thereof).
0. Definitions.
2. Additional Commercial Terms. If, on the Llama 2 version release date, the
monthly active users of the products or services made available by or for Licensee,
or Licensee's affiliates, is greater than 700 million monthly active users in the
preceding calendar month, you must request a license from Meta, which Meta may
grant to you in its sole discretion, and you are not authorized to exercise any of the
rights under this Agreement unless or until Meta otherwise expressly grants you
such rights.
3. Disclaimer of Warranty. UNLESS REQUIRED BY APPLICABLE LAW, THE
LLAMA MATERIALS AND ANY OUTPUT AND RESULTS THEREFROM ARE
PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND,
EITHER EXPRESS OR IMPLIED, INCLUDING, WITHOUT LIMITATION, ANY
WARRANTIES OF TITLE, NON-INFRINGEMENT, MERCHANTABILITY, OR
FITNESS FOR A PARTICULAR PURPOSE. YOU ARE SOLELY RESPONSIBLE
FOR DETERMINING THE APPROPRIATENESS OF USING OR REDISTRIBUTING
THE LLAMA MATERIALS AND ASSUME ANY RISKS ASSOCIATED WITH YOUR
USE OF THE LLAMA MATERIALS AND ANY OUTPUT AND RESULTS.
"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.
4. Limitation of Liability. IN NO EVENT WILL META OR ITS AFFILIATES BE
LIABLE UNDER ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, TORT,
NEGLIGENCE, PRODUCTS LIABILITY, OR OTHERWISE, ARISING OUT OF THIS
AGREEMENT, FOR ANY LOST PROFITS OR ANY INDIRECT, SPECIAL,
CONSEQUENTIAL, INCIDENTAL, EXEMPLARY OR PUNITIVE DAMAGES, EVEN
IF META OR ITS AFFILIATES HAVE BEEN ADVISED OF THE POSSIBILITY OF
ANY OF THE FOREGOING.
"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.
5. Intellectual Property.
a. No trademark licenses are granted under this Agreement, and in
connection with the Llama Materials, neither Meta nor Licensee may use any name
or mark owned by or associated with the other or any of its affiliates, except as
required for reasonable and customary use in describing and redistributing the
Llama Materials.
b. Subject to Meta's ownership of Llama Materials and derivatives made by or
for Meta, with respect to any derivative works and modifications of the Llama
Materials that are made by you, as between you and Meta, you are and will be the
owner of such derivative works and modifications.
c. If you institute litigation or other proceedings against Meta or any entity
(including a cross-claim or counterclaim in a lawsuit) alleging that the Llama
Materials or Llama 2 outputs or results, or any portion of any of the foregoing,
constitutes infringement of intellectual property or other rights owned or licensable
by you, then any licenses granted to you under this Agreement shall terminate as of
the date such litigation or claim is filed or instituted. You will indemnify and hold
harmless Meta from and against any claim by any third party arising out of or related
to your use or distribution of the Llama Materials.
6. Term and Termination. The term of this Agreement will commence upon your
acceptance of this Agreement or access to the Llama Materials and will continue in
full force and effect until terminated in accordance with the terms and conditions
herein. Meta may terminate this Agreement if you are in breach of any term or
condition of this Agreement. Upon termination of this Agreement, you shall delete
and cease use of the Llama Materials. Sections 3, 4 and 7 shall survive the
termination of this Agreement.
7. Governing Law and Jurisdiction. This Agreement will be governed and
construed under the laws of the State of California without regard to choice of law
principles, and the UN Convention on Contracts for the International Sale of Goods
does not apply to this Agreement. The courts of California shall have exclusive
jurisdiction of any dispute arising out of this Agreement.
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 <http://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
<http://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
<http://www.gnu.org/philosophy/why-not-lgpl.html>.

208
MODEL_CARD.md
View File

@@ -1,158 +1,98 @@
# LLaMA Model Card
# **Model Details**
## Model details
**Organization developing the model**
The FAIR team of Meta AI.
Meta developed and released the Llama 2 family of large language models (LLMs), a collection of pretrained and fine-tuned generative text models ranging in scale from 7 billion to 70 billion parameters. Our fine-tuned LLMs, called Llama-2-Chat, are optimized for dialogue use cases. Llama-2-Chat models outperform open-source chat models on most benchmarks we tested, and in our human evaluations for helpfulness and safety, are on par with some popular closed-source models like ChatGPT and PaLM.
**Model date**
LLaMA was trained between December. 2022 and Feb. 2023.
**Model Developers** Meta
**Model version**
This is version 1 of the model.
**Variations** Llama 2 comes in a range of parameter sizes — 7B, 13B, and 70B — as well as pretrained and fine-tuned variations.
**Model type**
LLaMA is an auto-regressive language model, based on the transformer architecture. The model comes in different sizes: 7B, 13B, 33B and 65B parameters.
**Input** Models input text only.
**Paper or resources for more information**
More information can be found in the paper “LLaMA, Open and Efficient Foundation Language Models”, available at https://research.facebook.com/publications/llama-open-and-efficient-foundation-language-models/.
**Output** Models generate text only.
**Citations details**
https://research.facebook.com/publications/llama-open-and-efficient-foundation-language-models/
**Model Architecture** Llama 2 is an auto-regressive language model that uses an optimized transformer architecture. The tuned versions use supervised fine-tuning (SFT) and reinforcement learning with human feedback (RLHF) to align to human preferences for helpfulness and safety.
**License**
Non-commercial bespoke license
||Training Data|Params|Content Length|GQA|Tokens|LR|
|---|---|---|---|---|---|---|
Llama 2|*A new mix of publicly available online data*|7B|4k|&#10007;|2.0T|3.0 x 10<sup>-4</sup>
Llama 2|*A new mix of publicly available online data*|13B|4k|&#10007;|2.0T|3.0 x 10<sup>-4</sup>
Llama 2|*A new mix of publicly available online data*|70B|4k|&#10004;|2.0T|1.5 x 10<sup>-4</sup>
**Where to send questions or comments about the model**
Questions and comments about LLaMA can be sent via the [GitHub repository](https://github.com/facebookresearch/llama) of the project , by opening an issue.
**Llama 2 family of models.** Token counts refer to pretraining data only. All models are trained with a global batch-size of 4M tokens. The 70B version uses Grouped-Query Attention (GQA) for improved inference scalability.
## Intended use
**Primary intended uses**
The primary use of LLaMA is research on large language models, including:
exploring potential applications such as question answering, natural language understanding or reading comprehension,
understanding capabilities and limitations of current language models, and developing techniques to improve those,
evaluating and mitigating biases, risks, toxic and harmful content generations, hallucinations.
**Model Dates** Llama 2 was trained between January 2023 and July 2023.
**Primary intended users**
The primary intended users of the model are researchers in natural language processing, machine learning and artificial intelligence.
**Status** This is a static model trained on an offline dataset. Future versions of the tuned models will be released as we improve model safety with community feedback.
**Out-of-scope use cases**
LLaMA is a base, or foundational, model. As such, it should not be used on downstream applications without further risk evaluation and mitigation. In particular, our model has not been trained with human feedback, and can thus generate toxic or offensive content, incorrect information or generally unhelpful answers.
**License** A custom commercial license is available at: [https://ai.meta.com/resources/models-and-libraries/llama-downloads/](https://ai.meta.com/resources/models-and-libraries/llama-downloads/)
## Factors
**Relevant factors**
One of the most relevant factors for which model performance may vary is which language is used. Although we included 20 languages in the training data, most of our dataset is made of English text, and we thus expect the model to perform better for English than other languages. Relatedly, it has been shown in previous studies that performance might vary for different dialects, and we expect that it will be the case for our model.
**Research Paper** More information can be found in the paper "Llama-2: Open Foundation and Fine-tuned Chat Models", available at https://ai.meta.com/research/publications/llama-2-open-foundation-and-fine-tuned-chat-models/.
**Evaluation factors**
As our model is trained on data from the Web, we expect that it reflects biases from this source. We thus evaluated on RAI datasets to measure biases exhibited by the model for gender, religion, race, sexual orientation, age, nationality, disability, physical appearance and socio-economic status. We also measure the toxicity of model generations, depending on the toxicity of the context used to prompt the model.
**Where to send questions or comments about the model** Instructions on how to provide feedback or comments on the model can be found in the model [README](README.md).
## Metrics
**Model performance measures**
We use the following measure to evaluate the model:
- Accuracy for common sense reasoning, reading comprehension, natural language understanding (MMLU), BIG-bench hard, WinoGender and CrowS-Pairs,
- Exact match for question answering,
- The toxicity score from Perspective API on RealToxicityPrompts.
# **Intended Use**
**Intended Use Cases** Llama 2 is intended for commercial and research use in English. Tuned models are intended for assistant-like chat, whereas pretrained models can be adapted for a variety of natural language generation tasks.
**Decision thresholds**
Not applicable.
**Out-of-scope Uses** Use in any manner that violates applicable laws or regulations (including trade compliance laws). Use in languages other than English. Use in any other way that is prohibited by the Acceptable Use Policy and Licensing Agreement for Llama 2.
**Approaches to uncertainty and variability**
Due to the high computational requirements of training LLMs, we trained only one model of each size, and thus could not evaluate variability of pre-training.
# **Hardware and Software**
**Training Factors** We used custom training libraries, Meta's Research Super Cluster, and production clusters for pretraining. Fine-tuning, annotation, and evaluation were also performed on third-party cloud compute.
## Evaluation datasets
The model was evaluated on the following benchmarks: BoolQ, PIQA, SIQA, HellaSwag, WinoGrande, ARC, OpenBookQA, NaturalQuestions, TriviaQA, RACE, MMLU, BIG-bench hard, GSM8k, RealToxicityPrompts, WinoGender, CrowS-Pairs.
**Carbon Footprint** Pretraining utilized a cumulative 3.3M GPU hours of computation on hardware of type A100-80GB (TDP of 350-400W). Estimated total emissions were 539 tCO2eq, 100% of which were offset by Meta’s sustainability program.
## Training dataset
The model was trained using the following source of data: CCNet [67%], C4 [15%], GitHub [4.5%], Wikipedia [4.5%], Books [4.5%], ArXiv [2.5%], Stack Exchange[2%]. The Wikipedia and Books domains include data in the following languages: bg, ca, cs, da, de, en, es, fr, hr, hu, it, nl, pl, pt, ro, ru, sl, sr, sv, uk. See the paper for more details about the training set and corresponding preprocessing.
||Time (GPU hours)|Power Consumption (W)|Carbon Emitted(tCO<sub>2</sub>eq)|
|---|---|---|---|
|Llama 2 7B|184320|400|31.22|
|Llama 2 13B|368640|400|62.44|
|Llama 2 70B|1720320|400|291.42|
|Total|3311616||539.00|
## Quantitative analysis
Hyperparameters for the model architecture
**CO<sub>2</sub> emissions during pretraining.** Time: total GPU time required for training each model. Power Consumption: peak power capacity per GPU device for the GPUs used adjusted for power usage efficiency. 100% of the emissions are directly offset by Meta's sustainability program, and because we are openly releasing these models, the pretraining costs do not need to be incurred by others.
# **Training Data**
**Overview** Llama 2 was pretrained on 2 trillion tokens of data from publicly available sources. The fine-tuning data includes publicly available instruction datasets, as well as over one million new human-annotated examples. Neither the pretraining nor the fine-tuning datasets include Meta user data.
**Data Freshness** The pretraining data has a cutoff of September 2022, but some tuning data is more recent, up to July 2023.
# **Evaluation Results**
In this section, we report the results for the Llama 1 and Llama 2 models on standard academic benchmarks.
For all the evaluations, we use our internal evaluations library.
|Model|Size|Code|Commonsense Reasoning|World Knowledge|Reading Comprehension|Math|MMLU|BBH|AGI Eval|
|---|---|---|---|---|---|---|---|---|---|
|Llama 1|7B|14.1|60.8|46.2|58.5|6.95|35.1|30.3|23.9|
|Llama 1|13B|18.9|66.1|52.6|62.3|10.9|46.9|37.0|33.9|
|Llama 1|33B|26.0|70.0|58.4|67.6|21.4|57.8|39.8|41.7|
|Llama 1|65B|30.7|70.7|60.5|68.6|30.8|63.4|43.5|47.6|
|Llama 2|7B|16.8|63.9|48.9|61.3|14.6|45.3|32.6|29.3|
|Llama 2|13B|24.5|66.9|55.4|65.8|28.7|54.8|39.4|39.1|
|Llama 2|70B|**37.5**|**71.9**|**63.6**|**69.4**|**35.2**|**68.9**|**51.2**|**54.2**|
**Overall performance on grouped academic benchmarks.** *Code:* We report the average pass@1 scores of our models on HumanEval and MBPP. *Commonsense Reasoning:* We report the average of PIQA, SIQA, HellaSwag, WinoGrande, ARC easy and challenge, OpenBookQA, and CommonsenseQA. We report 7-shot results for CommonSenseQA and 0-shot results for all other benchmarks. *World Knowledge:* We evaluate the 5-shot performance on NaturalQuestions and TriviaQA and report the average. *Reading Comprehension:* For reading comprehension, we report the 0-shot average on SQuAD, QuAC, and BoolQ. *MATH:* We report the average of the GSM8K (8 shot) and MATH (4 shot) benchmarks at top 1.
|||TruthfulQA|Toxigen|
|---|---|---|---|
|Llama 1|7B|27.42|23.00|
|Llama 1|13B|41.74|23.08|
|Llama 1|33B|44.19|22.57|
|Llama 1|65B|48.71|21.77|
|Llama 2|7B|33.29|**21.25**|
|Llama 2|13B|41.86|26.10|
|Llama 2|70B|**50.18**|24.60|
**Evaluation of pretrained LLMs on automatic safety benchmarks.** For TruthfulQA, we present the percentage of generations that are both truthful and informative (the higher the better). For ToxiGen, we present the percentage of toxic generations (the smaller the better).
<table>
<thead>
<tr>
<th >LLaMA</th> <th colspan=6>Model hyper parameters </th>
</tr>
<tr>
<th>Number of parameters</th><th>dimension</th><th>n heads</th><th>n layers</th><th>Learn rate</th><th>Batch size</th><th>n tokens</th>
</tr>
</thead>
<tbody>
<tr>
<th>7B</th> <th>4096</th> <th>32</th> <th>32</th> <th>3.0E-04</th><th>4M</th><th>1T
</tr>
<tr>
<th>13B</th><th>5120</th><th>40</th><th>40</th><th>3.0E-04</th><th>4M</th><th>1T
</tr>
<tr>
<th>33B</th><th>6656</th><th>52</th><th>60</th><th>1.5.E-04</th><th>4M</th><th>1.4T
</tr>
<tr>
<th>65B</th><th>8192</th><th>64</th><th>80</th><th>1.5.E-04</th><th>4M</th><th>1.4T
</tr>
</tbody>
</table>
|||TruthfulQA|Toxigen|
|---|---|---|---|
|Llama-2-Chat|7B|57.04|**0.00**|
|Llama-2-Chat|13B|62.18|**0.00**|
|Llama-2-Chat|70B|**64.14**|0.01|
**Evaluation of fine-tuned LLMs on different safety datasets.** Same metric definitions as above.
*Table 1 - Summary of LLama Model Hyperparameters*
# **Ethical Considerations and Limitations**
Llama 2 is a new technology that carries risks with use. Testing conducted to date has been in English, and has not covered, nor could it cover all scenarios. For these reasons, as with all LLMs, Llama 2’s potential outputs cannot be predicted in advance, and the model may in some instances produce inaccurate, biased or other objectionable responses to user prompts. Therefore, before deploying any applications of Llama 2, developers should perform safety testing and tuning tailored to their specific applications of the model.
We present our results on eight standard common sense reasoning benchmarks in the table below.
<table>
<thead>
<tr>
<th>LLaMA</th> <th colspan=9>Reasoning tasks </th>
</tr>
<tr>
<th>Number of parameters</th> <th>BoolQ</th><th>PIQA</th><th>SIQA</th><th>HellaSwag</th><th>WinoGrande</th><th>ARC-e</th><th>ARC-c</th><th>OBQA</th><th>COPA</th>
</tr>
</thead>
<tbody>
<tr>
<th>7B</th><th>76.5</th><th>79.8</th><th>48.9</th><th>76.1</th><th>70.1</th><th>76.7</th><th>47.6</th><th>57.2</th><th>93
</th>
<tr><th>13B</th><th>78.1</th><th>80.1</th><th>50.4</th><th>79.2</th><th>73</th><th>78.1</th><th>52.7</th><th>56.4</th><th>94
</th>
<tr><th>33B</th><th>83.1</th><th>82.3</th><th>50.4</th><th>82.8</th><th>76</th><th>81.4</th><th>57.8</th><th>58.6</th><th>92
</th>
<tr><th>65B</th><th>85.3</th><th>82.8</th><th>52.3</th><th>84.2</th><th>77</th><th>81.5</th><th>56</th><th>60.2</th><th>94</th></tr>
</tbody>
</table>
*Table 2 - Summary of LLama Model Performance on Reasoning tasks*
We present our results on bias in the table below. Note that lower value is better indicating lower bias.
| No | Category | FAIR LLM |
| --- | -------------------- | -------- |
| 1 | Gender | 70.6 |
| 2 | Religion | 79 |
| 3 | Race/Color | 57 |
| 4 | Sexual orientation | 81 |
| 5 | Age | 70.1 |
| 6 | Nationality | 64.2 |
| 7 | Disability | 66.7 |
| 8 | Physical appearance | 77.8 |
| 9 | Socioeconomic status | 71.5 |
| | LLaMA Average | 66.6 |
*Table 3 - Summary bias of our model output*
## Ethical considerations
**Data**
The data used to train the model is collected from various sources, mostly from the Web. As such, it contains offensive, harmful and biased content. We thus expect the model to exhibit such biases from the training data.
**Human life**
The model is not intended to inform decisions about matters central to human life, and should not be used in such a way.
**Mitigations**
We filtered the data from the Web based on its proximity to Wikipedia text and references. For this, we used a Kneser-Ney language model and a fastText linear classifier.
**Risks and harms**
Risks and harms of large language models include the generation of harmful, offensive or biased content. These models are often prone to generating incorrect information, sometimes referred to as hallucinations. We do not expect our model to be an exception in this regard.
**Use cases**
LLaMA is a foundational model, and as such, it should not be used for downstream applications without further investigation and mitigations of risks. These risks and potential fraught use cases include, but are not limited to: generation of misinformation and generation of harmful, biased or offensive content.
Please see the Responsible Use Guide available at [https://ai.meta.com/llama/responsible-use-guide/](https://ai.meta.com/llama/responsible-use-guide/)

108
README.md
View File

@@ -1,62 +1,98 @@
# LLaMA
# Llama 2
This repository is intended as a minimal, hackable and readable example to load [LLaMA](https://ai.facebook.com/blog/large-language-model-llama-meta-ai/) ([arXiv](https://arxiv.org/abs/2302.13971v1)) models and run inference.
In order to download the checkpoints and tokenizer, fill this [google form](https://forms.gle/jk851eBVbX1m5TAv5)
We are unlocking the power of large language models. Our latest version of Llama is now accessible to individuals, creators, researchers and businesses of all sizes so that they can experiment, innovate and scale their ideas responsibly.
This release includes model weights and starting code for pretrained and fine-tuned Llama language models — ranging from 7B to 70B parameters.
This repository is intended as a minimal example to load [Llama 2](https://ai.meta.com/research/publications/llama-2-open-foundation-and-fine-tuned-chat-models/) models and run inference. For more detailed examples leveraging HuggingFace, see [llama-recipes](https://github.com/facebookresearch/llama-recipes/).
## Download
In order to download the model weights and tokenizer, please visit the [Meta AI website](https://ai.meta.com/resources/models-and-libraries/llama-downloads/) and accept our License.
Once your request is approved, you will receive a signed URL over email. Then run the download.sh script, passing the URL provided when prompted to start the download.
Pre-requisites: make sure you have `wget` and `md5sum` installed. Then to run the script: `./download.sh`.
Keep in mind that the links expire after 24 hours and a certain amount of downloads. If you start seeing errors such as `403: Forbidden`, you can always re-request a link.
### Access on Hugging Face
We are also providing downloads on [Hugging Face](https://huggingface.co/meta-llama). You must first request a download from the Meta AI website using the same email address as your Hugging Face account. After doing so, you can request access to any of the models on Hugging Face and within 1-2 days your account will be granted access to all versions.
## Setup
In a conda env with pytorch / cuda available, run:
```
pip install -r requirements.txt
```
Then in this repository:
In a conda env with PyTorch / CUDA available, clone the repo and run in the top-level directory:
```
pip install -e .
```
## Download
Once your request is approved, you will receive links to download the tokenizer and model files.
Edit the `download.sh` script with the signed url provided in the email to download the model weights and tokenizer.
## Inference
The provided `example.py` can be run on a single or multi-gpu node with `torchrun` and will output completions for two pre-defined prompts. Using `TARGET_FOLDER` as defined in `download.sh`:
```
torchrun --nproc_per_node MP example.py --ckpt_dir $TARGET_FOLDER/model_size --tokenizer_path $TARGET_FOLDER/tokenizer.model
```
Different models require different MP values:
Different models require different model-parallel (MP) values:
| Model | MP |
|--------|----|
| 7B | 1 |
| 13B | 2 |
| 33B | 4 |
| 65B | 8 |
| 70B | 8 |
## FAQ
All models support sequence length up to 4096 tokens, but we pre-allocate the cache according to `max_seq_len` and `max_batch_size` values. So set those according to your hardware.
- [1. The download.sh script doesn't work on default bash in MacOS X](FAQ.md#1)
- [2. Generations are bad!](FAQ.md#2)
- [3. CUDA Out of memory errors](FAQ.md#3)
- [4. Other languages](FAQ.md#4)
### Pretrained Models
## Reference
These models are not finetuned for chat or Q&A. They should be prompted so that the expected answer is the natural continuation of the prompt.
LLaMA: Open and Efficient Foundation Language Models -- https://arxiv.org/abs/2302.13971
See `example_text_completion.py` for some examples. To illustrate, see command below to run it with the llama-2-7b model (`nproc_per_node` needs to be set to the `MP` value):
```
@article{touvron2023llama,
title={LLaMA: Open and Efficient Foundation Language Models},
author={Touvron, Hugo and Lavril, Thibaut and Izacard, Gautier and Martinet, Xavier and Lachaux, Marie-Anne and Lacroix, Timoth{\'e}e and Rozi{\`e}re, Baptiste and Goyal, Naman and Hambro, Eric and Azhar, Faisal and Rodriguez, Aurelien and Joulin, Armand and Grave, Edouard and Lample, Guillaume},
journal={arXiv preprint arXiv:2302.13971},
year={2023}
}
torchrun --nproc_per_node 1 example_text_completion.py \
--ckpt_dir llama-2-7b/ \
--tokenizer_path tokenizer.model \
--max_seq_len 128 --max_batch_size 4
```
### Fine-tuned Chat Models
The fine-tuned models were trained for dialogue applications. To get the expected features and performance for them, a specific formatting defined in [`chat_completion`](https://github.com/fairinternal/llama_v2/blob/main/llama/generation.py#L211)
needs to be followed, including the `INST` and `<<SYS>>` tags, `BOS` and `EOS` tokens, and the whitespaces and breaklines in between (we recommend calling `strip()` on inputs to avoid double-spaces).
You can also deploy additional classifiers for filtering out inputs and outputs that are deemed unsafe. See the llama-recipes repo for [an example](https://github.com/facebookresearch/llama-recipes/blob/main/inference/inference.py) of how to add a safety checker to the inputs and outputs of your inference code.
Examples using llama-2-7b-chat:
```
torchrun --nproc_per_node 1 example_chat_completion.py \
--ckpt_dir llama-2-7b-chat/ \
--tokenizer_path tokenizer.model \
--max_seq_len 512 --max_batch_size 4
```
Llama 2 is a new technology that carries potential risks with use. Testing conducted to date has not — and could not — cover all scenarios.
In order to help developers address these risks, we have created the [Responsible Use Guide](Responsible-Use-Guide.pdf). More details can be found in our research paper as well.
## Issues
Please report any software “bug,” or other problems with the models through one of the following means:
- Reporting issues with the model: [github.com/facebookresearch/llama](http://github.com/facebookresearch/llama)
- Reporting risky content generated by the model: [developers.facebook.com/llama_output_feedback](http://developers.facebook.com/llama_output_feedback)
- Reporting bugs and security concerns: [facebook.com/whitehat/info](http://facebook.com/whitehat/info)
## Model Card
See [MODEL_CARD.md](MODEL_CARD.md)
See [MODEL_CARD.md](MODEL_CARD.md).
## License
See the [LICENSE](LICENSE) file.
Our model and weights are licensed for both researchers and commercial entities, upholding the principles of openness. Our mission is to empower individuals, and industry through this opportunity, while fostering an environment of discovery and ethical AI advancements.
See the [LICENSE](LICENSE) file, as well as our accompanying [Acceptable Use Policy](USE_POLICY.md)
## References
1. [Research Paper](https://ai.meta.com/research/publications/llama-2-open-foundation-and-fine-tuned-chat-models/)
2. [Llama 2 technical overview](https://ai.meta.com/resources/models-and-libraries/llama)
3. [Open Innovation AI Research Community](https://ai.meta.com/llama/open-innovation-ai-research-community/)
## Original LLaMA
The repo for the original llama release is in the [`llama_v1`](https://github.com/facebookresearch/llama/tree/llama_v1) branch.

BIN
Responsible-Use-Guide.pdf Normal file
View File

Binary file not shown.

50
USE_POLICY.md Normal file
View File

@@ -0,0 +1,50 @@
# Llama 2 Acceptable Use Policy
Meta is committed to promoting safe and fair use of its tools and features, including Llama 2. If you access or use Llama 2, you agree to this Acceptable Use Policy (“Policy”). The most recent copy of this policy can be found at [ai.meta.com/llama/use-policy](http://ai.meta.com/llama/use-policy).
## Prohibited Uses
We want everyone to use Llama 2 safely and responsibly. You agree you will not use, or allow others to use, Llama 2 to:
1. Violate the law or others’ rights, including to:
1. Engage in, promote, generate, contribute to, encourage, plan, incite, or further illegal or unlawful activity or content, such as:
1. Violence or terrorism
2. Exploitation or harm to children, including the solicitation, creation, acquisition, or dissemination of child exploitative content or failure to report Child Sexual Abuse Material
3. Human trafficking, exploitation, and sexual violence
4. The illegal distribution of information or materials to minors, including obscene materials, or failure to employ legally required age-gating in connection with such information or materials.
5. Sexual solicitation
6. Any other criminal activity
2. Engage in, promote, incite, or facilitate the harassment, abuse, threatening, or bullying of individuals or groups of individuals
3. Engage in, promote, incite, or facilitate discrimination or other unlawful or harmful conduct in the provision of employment, employment benefits, credit, housing, other economic benefits, or other essential goods and services
4. Engage in the unauthorized or unlicensed practice of any profession including, but not limited to, financial, legal, medical/health, or related professional practices
5. Collect, process, disclose, generate, or infer health, demographic, or other sensitive personal or private information about individuals without rights and consents required by applicable laws
6. Engage in or facilitate any action or generate any content that infringes, misappropriates, or otherwise violates any third-party rights, including the outputs or results of any products or services using the Llama 2 Materials
7. Create, generate, or facilitate the creation of malicious code, malware, computer viruses or do anything else that could disable, overburden, interfere with or impair the proper working, integrity, operation or appearance of a website or computer system
2. Engage in, promote, incite, facilitate, or assist in the planning or development of activities that present a risk of death or bodily harm to individuals, including use of Llama 2 related to the following:
1. Military, warfare, nuclear industries or applications, espionage, use for materials or activities that are subject to the International Traffic Arms Regulations (ITAR) maintained by the United States Department of State
2. Guns and illegal weapons (including weapon development)
3. Illegal drugs and regulated/controlled substances
4. Operation of critical infrastructure, transportation technologies, or heavy machinery
5. Self-harm or harm to others, including suicide, cutting, and eating disorders
6. Any content intended to incite or promote violence, abuse, or any infliction of bodily harm to an individual
3. Intentionally deceive or mislead others, including use of Llama 2 related to the following:
1. Generating, promoting, or furthering fraud or the creation or promotion of disinformation
2. Generating, promoting, or furthering defamatory content, including the creation of defamatory statements, images, or other content
3. Generating, promoting, or further distributing spam
4. Impersonating another individual without consent, authorization, or legal right
5. Representing that the use of Llama 2 or outputs are human-generated
6. Generating or facilitating false online engagement, including fake reviews and other means of fake online engagement
4. Fail to appropriately disclose to end users any known dangers of your AI system
Please report any violation of this Policy, software “bug,” or other problems that could lead to a violation of this Policy through one of the following means:
* Reporting issues with the model: [github.com/facebookresearch/llama](http://github.com/facebookresearch/llama)
* Reporting risky content generated by the model: [developers.facebook.com/llama_output_feedback](http://developers.facebook.com/llama_output_feedback)
* Reporting bugs and security concerns: [facebook.com/whitehat/info](http://facebook.com/whitehat/info)
* Reporting violations of the Acceptable Use Policy or unlicensed uses of Llama: [LlamaUseReport@meta.com](mailto:LlamaUseReport@meta.com)

63
download.sh
View File

@@ -1,33 +1,58 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# This software may be used and distributed according to the terms of the GNU General Public License version 3.
# This software may be used and distributed according to the terms of the Llama 2 Community License Agreement.
PRESIGNED_URL="" # replace with presigned url from email
MODEL_SIZE="7B,13B,30B,65B" # edit this list with the model sizes you wish to download
TARGET_FOLDER="" # where all files should end up
read -p "Enter the URL from email: " PRESIGNED_URL
echo ""
read -p "Enter the list of models to download without spaces (7B,13B,70B,7B-chat,13B-chat,70B-chat), or press Enter for all: " MODEL_SIZE
TARGET_FOLDER="." # where all files should end up
mkdir -p ${TARGET_FOLDER}
declare -A N_SHARD_DICT
if [[ $MODEL_SIZE == "" ]]; then
MODEL_SIZE="7B,13B,70B,7B-chat,13B-chat,70B-chat"
fi
N_SHARD_DICT["7B"]="0"
N_SHARD_DICT["13B"]="1"
N_SHARD_DICT["30B"]="3"
N_SHARD_DICT["65B"]="7"
echo "Downloading LICENSE and Acceptable Usage Policy"
wget ${PRESIGNED_URL/'*'/"LICENSE"} -O ${TARGET_FOLDER}"/LICENSE"
wget ${PRESIGNED_URL/'*'/"USE_POLICY.md"} -O ${TARGET_FOLDER}"/USE_POLICY.md"
echo "Downloading tokenizer"
wget ${PRESIGNED_URL/'*'/"tokenizer.model"} -O ${TARGET_FOLDER}"/tokenizer.model"
wget ${PRESIGNED_URL/'*'/"tokenizer_checklist.chk"} -O ${TARGET_FOLDER}"/tokenizer_checklist.chk"
(cd ${TARGET_FOLDER} && md5sum -c tokenizer_checklist.chk)
for i in ${MODEL_SIZE//,/ }
for m in ${MODEL_SIZE//,/ }
do
echo "Downloading ${i}"
mkdir -p ${TARGET_FOLDER}"/${i}"
for s in $(seq -f "0%g" 0 ${N_SHARD_DICT[$i]})
if [[ $m == "7B" ]]; then
SHARD=0
MODEL_PATH="llama-2-7b"
elif [[ $m == "7B-chat" ]]; then
SHARD=0
MODEL_PATH="llama-2-7b-chat"
elif [[ $m == "13B" ]]; then
SHARD=1
MODEL_PATH="llama-2-13b"
elif [[ $m == "13B-chat" ]]; then
SHARD=1
MODEL_PATH="llama-2-13b-chat"
elif [[ $m == "70B" ]]; then
SHARD=7
MODEL_PATH="llama-2-70b"
elif [[ $m == "70B-chat" ]]; then
SHARD=7
MODEL_PATH="llama-2-70b-chat"
fi
echo "Downloading ${MODEL_PATH}"
mkdir -p ${TARGET_FOLDER}"/${MODEL_PATH}"
for s in $(seq -f "0%g" 0 ${SHARD})
do
wget ${PRESIGNED_URL/'*'/"${i}/consolidated.${s}.pth"} -O ${TARGET_FOLDER}"/${i}/consolidated.${s}.pth"
wget ${PRESIGNED_URL/'*'/"${MODEL_PATH}/consolidated.${s}.pth"} -O ${TARGET_FOLDER}"/${MODEL_PATH}/consolidated.${s}.pth"
done
wget ${PRESIGNED_URL/'*'/"${i}/params.json"} -O ${TARGET_FOLDER}"/${i}/params.json"
wget ${PRESIGNED_URL/'*'/"${i}/checklist.chk"} -O ${TARGET_FOLDER}"/${i}/checklist.chk"
wget ${PRESIGNED_URL/'*'/"${MODEL_PATH}/params.json"} -O ${TARGET_FOLDER}"/${MODEL_PATH}/params.json"
wget ${PRESIGNED_URL/'*'/"${MODEL_PATH}/checklist.chk"} -O ${TARGET_FOLDER}"/${MODEL_PATH}/checklist.chk"
echo "Checking checksums"
(cd ${TARGET_FOLDER}"/${i}" && md5sum -c checklist.chk)
done
(cd ${TARGET_FOLDER}"/${MODEL_PATH}" && md5sum -c checklist.chk)
done

73
example_chat_completion.py Normal file
View File

@@ -0,0 +1,73 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# This software may be used and distributed according to the terms of the Llama 2 Community License Agreement.
from typing import Optional
import fire
from llama import Llama
def main(
ckpt_dir: str,
tokenizer_path: str,
temperature: float = 0.6,
top_p: float = 0.9,
max_seq_len: int = 512,
max_batch_size: int = 4,
max_gen_len: Optional[int] = None,
):
generator = Llama.build(
ckpt_dir=ckpt_dir,
tokenizer_path=tokenizer_path,
max_seq_len=max_seq_len,
max_batch_size=max_batch_size,
)
dialogs = [
[{"role": "user", "content": "what is the recipe of mayonnaise?"}],
[
{"role": "user", "content": "I am going to Paris, what should I see?"},
{
"role": "assistant",
"content": """\
Paris, the capital of France, is known for its stunning architecture, art museums, historical landmarks, and romantic atmosphere. Here are some of the top attractions to see in Paris:
1. The Eiffel Tower: The iconic Eiffel Tower is one of the most recognizable landmarks in the world and offers breathtaking views of the city.
2. The Louvre Museum: The Louvre is one of the world's largest and most famous museums, housing an impressive collection of art and artifacts, including the Mona Lisa.
3. Notre-Dame Cathedral: This beautiful cathedral is one of the most famous landmarks in Paris and is known for its Gothic architecture and stunning stained glass windows.
These are just a few of the many attractions that Paris has to offer. With so much to see and do, it's no wonder that Paris is one of the most popular tourist destinations in the world.""",
},
{"role": "user", "content": "What is so great about #1?"},
],
[
{"role": "system", "content": "Always answer with Haiku"},
{"role": "user", "content": "I am going to Paris, what should I see?"},
],
[
{
"role": "system",
"content": "Always answer with emojis",
},
{"role": "user", "content": "How to go from Beijing to NY?"},
],
]
results = generator.chat_completion(
dialogs, # type: ignore
max_gen_len=max_gen_len,
temperature=temperature,
top_p=top_p,
)
for dialog, result in zip(dialogs, results):
for msg in dialog:
print(f"{msg['role'].capitalize()}: {msg['content']}\n")
print(
f"> {result['generation']['role'].capitalize()}: {result['generation']['content']}"
)
print("\n==================================\n")
if __name__ == "__main__":
fire.Fire(main)

55
example_text_completion.py Executable file
View File

@@ -0,0 +1,55 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# This software may be used and distributed according to the terms of the Llama 2 Community License Agreement.
import fire
from llama import Llama
def main(
ckpt_dir: str,
tokenizer_path: str,
temperature: float = 0.6,
top_p: float = 0.9,
max_seq_len: int = 128,
max_gen_len: int = 64,
max_batch_size: int = 4,
):
generator = Llama.build(
ckpt_dir=ckpt_dir,
tokenizer_path=tokenizer_path,
max_seq_len=max_seq_len,
max_batch_size=max_batch_size,
)
prompts = [
# For these prompts, the expected answer is the natural continuation of the prompt
"I believe the meaning of life is",
"Simply put, the theory of relativity states that ",
"""A brief message congratulating the team on the launch:
Hi everyone,
I just """,
# Few shot prompt (providing a few examples before asking model to complete more);
"""Translate English to French:
sea otter => loutre de mer
peppermint => menthe poivrée
plush girafe => girafe peluche
cheese =>""",
]
results = generator.text_completion(
prompts,
max_gen_len=max_gen_len,
temperature=temperature,
top_p=top_p,
)
for prompt, result in zip(prompts, results):
print(prompt)
print(f"> {result['generation']}")
print("\n==================================\n")
if __name__ == "__main__":
fire.Fire(main)

4
llama/__init__.py
View File

@@ -1,6 +1,6 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# This software may be used and distributed according to the terms of the GNU General Public License version 3.
# This software may be used and distributed according to the terms of the Llama 2 Community License Agreement.
from .generation import LLaMA
from .generation import Llama
from .model import ModelArgs, Transformer
from .tokenizer import Tokenizer

288
llama/generation.py
View File

@@ -1,69 +1,295 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# This software may be used and distributed according to the terms of the GNU General Public License version 3.
# This software may be used and distributed according to the terms of the Llama 2 Community License Agreement.
from typing import List
import json
import os
import sys
import time
from pathlib import Path
from typing import List, Literal, Optional, Tuple, TypedDict
import torch
import torch.nn.functional as F
from fairscale.nn.model_parallel.initialize import (
get_model_parallel_rank,
initialize_model_parallel,
model_parallel_is_initialized,
)
from llama.model import ModelArgs, Transformer
from llama.tokenizer import Tokenizer
from llama.model import Transformer
Role = Literal["system", "user", "assistant"]
class LLaMA:
class Message(TypedDict):
role: Role
content: str
class CompletionPrediction(TypedDict, total=False):
generation: str
tokens: List[str] # not required
logprobs: List[float] # not required
class ChatPrediction(TypedDict, total=False):
generation: Message
tokens: List[str] # not required
logprobs: List[float] # not required
Dialog = List[Message]
B_INST, E_INST = "[INST]", "[/INST]"
B_SYS, E_SYS = "<<SYS>>\n", "\n<</SYS>>\n\n"
DEFAULT_SYSTEM_PROMPT = """\
You are a helpful, respectful and honest assistant. Always answer as helpfully as possible, while being safe. Your answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. Please ensure that your responses are socially unbiased and positive in nature.
If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information."""
class Llama:
@staticmethod
def build(
ckpt_dir: str,
tokenizer_path: str,
max_seq_len: int,
max_batch_size: int,
model_parallel_size: Optional[int] = None,
) -> "Llama":
if not torch.distributed.is_initialized():
torch.distributed.init_process_group("nccl")
if not model_parallel_is_initialized():
if model_parallel_size is None:
model_parallel_size = int(os.environ.get("WORLD_SIZE", 1))
initialize_model_parallel(model_parallel_size)
local_rank = int(os.environ.get("LOCAL_RANK", 0))
torch.cuda.set_device(local_rank)
# seed must be the same in all processes
torch.manual_seed(1)
if local_rank > 0:
sys.stdout = open(os.devnull, "w")
start_time = time.time()
checkpoints = sorted(Path(ckpt_dir).glob("*.pth"))
assert len(checkpoints) > 0, f"no checkpoint files found in {ckpt_dir}"
assert model_parallel_size == len(
checkpoints
), f"Loading a checkpoint for MP={len(checkpoints)} but world size is {model_parallel_size}"
ckpt_path = checkpoints[get_model_parallel_rank()]
checkpoint = torch.load(ckpt_path, map_location="cpu")
with open(Path(ckpt_dir) / "params.json", "r") as f:
params = json.loads(f.read())
model_args: ModelArgs = ModelArgs(
max_seq_len=max_seq_len,
max_batch_size=max_batch_size,
**params,
)
tokenizer = Tokenizer(model_path=tokenizer_path)
model_args.vocab_size = tokenizer.n_words
torch.set_default_tensor_type(torch.cuda.HalfTensor)
model = Transformer(model_args)
model.load_state_dict(checkpoint, strict=False)
print(f"Loaded in {time.time() - start_time:.2f} seconds")
return Llama(model, tokenizer)
def __init__(self, model: Transformer, tokenizer: Tokenizer):
self.model = model
self.tokenizer = tokenizer
@torch.inference_mode()
def generate(
self,
prompts: List[str],
prompt_tokens: List[List[int]],
max_gen_len: int,
temperature: float = 0.8,
top_p: float = 0.95,
) -> List[str]:
bsz = len(prompts)
temperature: float = 0.6,
top_p: float = 0.9,
logprobs: bool = False,
echo: bool = False,
) -> Tuple[List[List[int]], Optional[List[List[float]]]]:
params = self.model.params
bsz = len(prompt_tokens)
assert bsz <= params.max_batch_size, (bsz, params.max_batch_size)
prompt_tokens = [self.tokenizer.encode(x, bos=True, eos=False) for x in prompts]
min_prompt_len = min(len(t) for t in prompt_tokens)
max_prompt_len = max(len(t) for t in prompt_tokens)
assert max_prompt_len <= params.max_seq_len
total_len = min(params.max_seq_len, max_gen_len + max_prompt_len)
min_prompt_size = min([len(t) for t in prompt_tokens])
max_prompt_size = max([len(t) for t in prompt_tokens])
total_len = min(params.max_seq_len, max_gen_len + max_prompt_size)
tokens = torch.full((bsz, total_len), self.tokenizer.pad_id).cuda().long()
pad_id = self.tokenizer.pad_id
tokens = torch.full((bsz, total_len), pad_id, dtype=torch.long, device="cuda")
for k, t in enumerate(prompt_tokens):
tokens[k, : len(t)] = torch.tensor(t).long()
input_text_mask = tokens != self.tokenizer.pad_id
start_pos = min_prompt_size
tokens[k, : len(t)] = torch.tensor(t, dtype=torch.long, device="cuda")
if logprobs:
token_logprobs = torch.zeros_like(tokens, dtype=torch.float)
prev_pos = 0
for cur_pos in range(start_pos, total_len):
eos_reached = torch.tensor([False] * bsz, device="cuda")
input_text_mask = tokens != pad_id
for cur_pos in range(min_prompt_len, total_len):
logits = self.model.forward(tokens[:, prev_pos:cur_pos], prev_pos)
if logprobs:
token_logprobs[:, prev_pos + 1 : cur_pos + 1] = -F.cross_entropy(
input=logits.transpose(1, 2),
target=tokens[:, prev_pos + 1 : cur_pos + 1],
reduction="none",
ignore_index=pad_id,
)
if temperature > 0:
probs = torch.softmax(logits / temperature, dim=-1)
probs = torch.softmax(logits[:, -1] / temperature, dim=-1)
next_token = sample_top_p(probs, top_p)
else:
next_token = torch.argmax(logits, dim=-1)
next_token = torch.argmax(logits[:, -1], dim=-1)
next_token = next_token.reshape(-1)
# only replace token if prompt has already been generated
next_token = torch.where(
input_text_mask[:, cur_pos], tokens[:, cur_pos], next_token
)
tokens[:, cur_pos] = next_token
eos_reached |= (~input_text_mask[:, cur_pos]) & (
next_token == self.tokenizer.eos_id
)
prev_pos = cur_pos
if all(eos_reached):
break
decoded = []
for i, t in enumerate(tokens.tolist()):
if logprobs:
token_logprobs = token_logprobs.tolist()
out_tokens, out_logprobs = [], []
for i, toks in enumerate(tokens.tolist()):
# cut to max gen len
t = t[: len(prompt_tokens[i]) + max_gen_len]
start = 0 if echo else len(prompt_tokens[i])
toks = toks[start : len(prompt_tokens[i]) + max_gen_len]
if logprobs:
probs = token_logprobs[i][start : len(prompt_tokens[i]) + max_gen_len]
# cut to eos tok if any
try:
t = t[: t.index(self.tokenizer.eos_id)]
except ValueError:
pass
decoded.append(self.tokenizer.decode(t))
return decoded
if self.tokenizer.eos_id in toks:
eos_idx = toks.index(self.tokenizer.eos_id)
toks = toks[:eos_idx]
probs = probs[:eos_idx] if logprobs else None
out_tokens.append(toks)
out_logprobs.append(probs)
return (out_tokens, out_logprobs if logprobs else None)
def text_completion(
self,
prompts: List[str],
temperature: float = 0.6,
top_p: float = 0.9,
max_gen_len: Optional[int] = None,
logprobs: bool = False,
echo: bool = False,
) -> List[CompletionPrediction]:
if max_gen_len is None:
max_gen_len = self.model.params.max_seq_len - 1
prompt_tokens = [self.tokenizer.encode(x, bos=True, eos=False) for x in prompts]
generation_tokens, generation_logprobs = self.generate(
prompt_tokens=prompt_tokens,
max_gen_len=max_gen_len,
temperature=temperature,
top_p=top_p,
logprobs=logprobs,
echo=echo,
)
if logprobs:
return [
{
"generation": self.tokenizer.decode(t),
"tokens": [self.tokenizer.decode(x) for x in t],
"logprobs": logprobs_i,
}
for t, logprobs_i in zip(generation_tokens, generation_logprobs)
]
return [{"generation": self.tokenizer.decode(t)} for t in generation_tokens]
def chat_completion(
self,
dialogs: List[Dialog],
temperature: float = 0.6,
top_p: float = 0.9,
max_gen_len: Optional[int] = None,
logprobs: bool = False,
) -> List[ChatPrediction]:
if max_gen_len is None:
max_gen_len = self.model.params.max_seq_len - 1
prompt_tokens = []
for dialog in dialogs:
if dialog[0]["role"] != "system":
dialog = [
{
"role": "system",
"content": DEFAULT_SYSTEM_PROMPT,
}
] + dialog
dialog = [
{
"role": dialog[1]["role"],
"content": B_SYS
+ dialog[0]["content"]
+ E_SYS
+ dialog[1]["content"],
}
] + dialog[2:]
assert all([msg["role"] == "user" for msg in dialog[::2]]) and all(
[msg["role"] == "assistant" for msg in dialog[1::2]]
), (
"model only supports 'system', 'user' and 'assistant' roles, "
"starting with 'system', then 'user' and alternating (u/a/u/a/u...)"
)
dialog_tokens: List[int] = sum(
[
self.tokenizer.encode(
f"{B_INST} {(prompt['content']).strip()} {E_INST} {(answer['content']).strip()} ",
bos=True,
eos=True,
)
for prompt, answer in zip(
dialog[::2],
dialog[1::2],
)
],
[],
)
assert (
dialog[-1]["role"] == "user"
), f"Last message must be from user, got {dialog[-1]['role']}"
dialog_tokens += self.tokenizer.encode(
f"{B_INST} {(dialog[-1]['content']).strip()} {E_INST}",
bos=True,
eos=False,
)
prompt_tokens.append(dialog_tokens)
generation_tokens, generation_logprobs = self.generate(
prompt_tokens=prompt_tokens,
max_gen_len=max_gen_len,
temperature=temperature,
top_p=top_p,
logprobs=logprobs,
)
if logprobs:
return [
{
"generation": {
"role": "assistant",
"content": self.tokenizer.decode(t),
},
"tokens": [self.tokenizer.decode(x) for x in t],
"logprobs": logprobs_i,
}
for t, logprobs_i in zip(generation_tokens, generation_logprobs)
]
return [
{"generation": {"role": "assistant", "content": self.tokenizer.decode(t)}}
for t in generation_tokens
]
def sample_top_p(probs, p):

116
llama/model.py
View File

@@ -1,29 +1,30 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# This software may be used and distributed according to the terms of the GNU General Public License version 3.
# This software may be used and distributed according to the terms of the Llama 2 Community License Agreement.
from typing import Optional, Tuple
from dataclasses import dataclass
import math
import torch
from torch import nn
import torch.nn.functional as F
from dataclasses import dataclass
from typing import Any, Optional, Tuple
import fairscale.nn.model_parallel.initialize as fs_init
import torch
import torch.nn.functional as F
from fairscale.nn.model_parallel.layers import (
ColumnParallelLinear,
ParallelEmbedding,
RowParallelLinear,
ColumnParallelLinear,
)
from torch import nn
@dataclass
class ModelArgs:
dim: int = 512
n_layers: int = 8
n_heads: int = 8
dim: int = 4096
n_layers: int = 32
n_heads: int = 32
n_kv_heads: Optional[int] = None
vocab_size: int = -1 # defined later by tokenizer
multiple_of: int = 256 # make SwiGLU hidden layer size multiple of large power of 2
ffn_dim_multiplier: Optional[float] = None
norm_eps: float = 1e-5
max_batch_size: int = 32
@@ -73,11 +74,26 @@ def apply_rotary_emb(
return xq_out.type_as(xq), xk_out.type_as(xk)
def repeat_kv(x: torch.Tensor, n_rep: int) -> torch.Tensor:
"""torch.repeat_interleave(x, dim=2, repeats=n_rep)"""
bs, slen, n_kv_heads, head_dim = x.shape
if n_rep == 1:
return x
return (
x[:, :, :, None, :]
.expand(bs, slen, n_kv_heads, n_rep, head_dim)
.reshape(bs, slen, n_kv_heads * n_rep, head_dim)
)
class Attention(nn.Module):
def __init__(self, args: ModelArgs):
super().__init__()
self.n_local_heads = args.n_heads // fs_init.get_model_parallel_world_size()
self.n_kv_heads = args.n_heads if args.n_kv_heads is None else args.n_kv_heads
model_parallel_size = fs_init.get_model_parallel_world_size()
self.n_local_heads = args.n_heads // model_parallel_size
self.n_local_kv_heads = self.n_kv_heads // model_parallel_size
self.n_rep = self.n_local_heads // self.n_local_kv_heads
self.head_dim = args.dim // args.n_heads
self.wq = ColumnParallelLinear(
@@ -89,14 +105,14 @@ class Attention(nn.Module):
)
self.wk = ColumnParallelLinear(
args.dim,
args.n_heads * self.head_dim,
self.n_kv_heads * self.head_dim,
bias=False,
gather_output=False,
init_method=lambda x: x,
)
self.wv = ColumnParallelLinear(
args.dim,
args.n_heads * self.head_dim,
self.n_kv_heads * self.head_dim,
bias=False,
gather_output=False,
init_method=lambda x: x,
@@ -110,19 +126,35 @@ class Attention(nn.Module):
)
self.cache_k = torch.zeros(
(args.max_batch_size, args.max_seq_len, self.n_local_heads, self.head_dim)
(
args.max_batch_size,
args.max_seq_len,
self.n_local_kv_heads,
self.head_dim,
)
).cuda()
self.cache_v = torch.zeros(
(args.max_batch_size, args.max_seq_len, self.n_local_heads, self.head_dim)
(
args.max_batch_size,
args.max_seq_len,
self.n_local_kv_heads,
self.head_dim,
)
).cuda()
def forward(self, x: torch.Tensor, start_pos: int, freqs_cis: torch.Tensor, mask: Optional[torch.Tensor]):
def forward(
self,
x: torch.Tensor,
start_pos: int,
freqs_cis: torch.Tensor,
mask: Optional[torch.Tensor],
):
bsz, seqlen, _ = x.shape
xq, xk, xv = self.wq(x), self.wk(x), self.wv(x)
xq = xq.view(bsz, seqlen, self.n_local_heads, self.head_dim)
xk = xk.view(bsz, seqlen, self.n_local_heads, self.head_dim)
xv = xv.view(bsz, seqlen, self.n_local_heads, self.head_dim)
xk = xk.view(bsz, seqlen, self.n_local_kv_heads, self.head_dim)
xv = xv.view(bsz, seqlen, self.n_local_kv_heads, self.head_dim)
xq, xk = apply_rotary_emb(xq, xk, freqs_cis=freqs_cis)
@@ -135,18 +167,19 @@ class Attention(nn.Module):
keys = self.cache_k[:bsz, : start_pos + seqlen]
values = self.cache_v[:bsz, : start_pos + seqlen]
xq = xq.transpose(1, 2)
# repeat k/v heads if n_kv_heads < n_heads
keys = repeat_kv(keys, self.n_rep) # (bs, seqlen, n_local_heads, head_dim)
values = repeat_kv(values, self.n_rep) # (bs, seqlen, n_local_heads, head_dim)
xq = xq.transpose(1, 2) # (bs, n_local_heads, seqlen, head_dim)
keys = keys.transpose(1, 2)
values = values.transpose(1, 2)
scores = torch.matmul(xq, keys.transpose(2, 3)) / math.sqrt(self.head_dim)
if mask is not None:
scores = scores + mask # (bs, n_local_heads, slen, cache_len + slen)
scores = scores + mask # (bs, n_local_heads, seqlen, cache_len + seqlen)
scores = F.softmax(scores.float(), dim=-1).type_as(xq)
output = torch.matmul(scores, values) # (bs, n_local_heads, slen, head_dim)
output = output.transpose(
1, 2
).contiguous().view(bsz, seqlen, -1)
output = torch.matmul(scores, values) # (bs, n_local_heads, seqlen, head_dim)
output = output.transpose(1, 2).contiguous().view(bsz, seqlen, -1)
return self.wo(output)
@@ -156,9 +189,13 @@ class FeedForward(nn.Module):
dim: int,
hidden_dim: int,
multiple_of: int,
ffn_dim_multiplier: Optional[float],
):
super().__init__()
hidden_dim = int(2 * hidden_dim / 3)
# custom dim factor multiplier
if ffn_dim_multiplier is not None:
hidden_dim = int(ffn_dim_multiplier * hidden_dim)
hidden_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
self.w1 = ColumnParallelLinear(
@@ -183,14 +220,25 @@ class TransformerBlock(nn.Module):
self.head_dim = args.dim // args.n_heads
self.attention = Attention(args)
self.feed_forward = FeedForward(
dim=args.dim, hidden_dim=4 * args.dim, multiple_of=args.multiple_of
dim=args.dim,
hidden_dim=4 * args.dim,
multiple_of=args.multiple_of,
ffn_dim_multiplier=args.ffn_dim_multiplier,
)
self.layer_id = layer_id
self.attention_norm = RMSNorm(args.dim, eps=args.norm_eps)
self.ffn_norm = RMSNorm(args.dim, eps=args.norm_eps)
def forward(self, x: torch.Tensor, start_pos: int, freqs_cis: torch.Tensor, mask: Optional[torch.Tensor]):
h = x + self.attention.forward(self.attention_norm(x), start_pos, freqs_cis, mask)
def forward(
self,
x: torch.Tensor,
start_pos: int,
freqs_cis: torch.Tensor,
mask: Optional[torch.Tensor],
):
h = x + self.attention.forward(
self.attention_norm(x), start_pos, freqs_cis, mask
)
out = h + self.feed_forward.forward(self.ffn_norm(h))
return out
@@ -228,11 +276,13 @@ class Transformer(nn.Module):
mask = None
if seqlen > 1:
mask = torch.full((1, 1, seqlen, seqlen), float("-inf"), device=tokens.device)
mask = torch.full(
(1, 1, seqlen, seqlen), float("-inf"), device=tokens.device
)
mask = torch.triu(mask, diagonal=start_pos + 1).type_as(h)
for layer in self.layers:
h = layer(h, start_pos, freqs_cis, mask)
h = self.norm(h)
output = self.output(h[:, -1, :]) # only compute last logits
return output.float()
output = self.output(h).float()
return output

7
llama/tokenizer.py
View File

@@ -1,10 +1,11 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# This software may be used and distributed according to the terms of the GNU General Public License version 3.
# This software may be used and distributed according to the terms of the Llama 2 Community License Agreement.
from sentencepiece import SentencePieceProcessor
import os
from logging import getLogger
from typing import List
import os
from sentencepiece import SentencePieceProcessor
logger = getLogger()

2
requirements.txt
View File

@@ -1,4 +1,4 @@
torch
fairscale
fire
sentencepiece
sentencepiece

16
setup.py
View File

@@ -1,6 +1,16 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# This software may be used and distributed according to the terms of the GNU General Public License version 3.
# This software may be used and distributed according to the terms of the Llama 2 Community License Agreement.
from setuptools import setup, find_packages
from setuptools import find_packages, setup
setup(name="llama", version="0.0.0", packages=find_packages())
def get_requirements(path: str):
return [l.strip() for l in open(path)]
setup(
name="llama",
version="0.0.1",
packages=find_packages(),
install_requires=get_requirements("requirements.txt"),
)