├── .gitignore ├── CODE_OF_CONDUCT.md ├── CONTRIBUTING.md ├── FAQ.md ├── LICENSE ├── MODEL_CARD.md ├── README.md ├── deployment ├── README.md ├── dependencies-install.sh ├── head-nfs-install.sh ├── pics │ ├── friends.jpg │ ├── messi-and-pep.jpg │ ├── newton-einstein-8xA10.gif │ ├── newton-einstein.jpg │ ├── rick-morty.jpg │ └── southpark.jpg ├── request_2xA10.json ├── request_4xA10.json ├── request_8xA10.json └── setup_nodes.sh ├── download.sh ├── example.py ├── interactive.py ├── llama ├── __init__.py ├── generation.py ├── model.py └── tokenizer.py ├── requirements.txt └── setup.py /.gitignore: -------------------------------------------------------------------------------- 1 | # Byte-compiled / optimized / DLL files 2 | __pycache__/ 3 | *.py[cod] 4 | *$py.class 5 | 6 | # C extensions 7 | *.so 8 | 9 | # Distribution / packaging 10 | .Python 11 | build/ 12 | develop-eggs/ 13 | dist/ 14 | downloads/ 15 | eggs/ 16 | .eggs/ 17 | lib/ 18 | lib64/ 19 | parts/ 20 | sdist/ 21 | var/ 22 | wheels/ 23 | share/python-wheels/ 24 | *.egg-info/ 25 | .installed.cfg 26 | *.egg 27 | MANIFEST 28 | 29 | # PyInstaller 30 | # Usually these files are written by a python script from a template 31 | # before PyInstaller builds the exe, so as to inject date/other infos into it. 32 | *.manifest 33 | *.spec 34 | 35 | # Installer logs 36 | pip-log.txt 37 | pip-delete-this-directory.txt 38 | 39 | # Unit test / coverage reports 40 | htmlcov/ 41 | .tox/ 42 | .nox/ 43 | .coverage 44 | .coverage.* 45 | .cache 46 | nosetests.xml 47 | coverage.xml 48 | *.cover 49 | *.py,cover 50 | .hypothesis/ 51 | .pytest_cache/ 52 | cover/ 53 | 54 | # Translations 55 | *.mo 56 | *.pot 57 | 58 | # Django stuff: 59 | *.log 60 | local_settings.py 61 | db.sqlite3 62 | db.sqlite3-journal 63 | 64 | # Flask stuff: 65 | instance/ 66 | .webassets-cache 67 | 68 | # Scrapy stuff: 69 | .scrapy 70 | 71 | # Sphinx documentation 72 | docs/_build/ 73 | 74 | # PyBuilder 75 | .pybuilder/ 76 | target/ 77 | 78 | # Jupyter Notebook 79 | .ipynb_checkpoints 80 | 81 | # IPython 82 | profile_default/ 83 | ipython_config.py 84 | 85 | # pyenv 86 | # For a library or package, you might want to ignore these files since the code is 87 | # intended to run in multiple environments; otherwise, check them in: 88 | # .python-version 89 | 90 | # pipenv 91 | # According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control. 92 | # However, in case of collaboration, if having platform-specific dependencies or dependencies 93 | # having no cross-platform support, pipenv may install dependencies that don't work, or not 94 | # install all needed dependencies. 95 | #Pipfile.lock 96 | 97 | # poetry 98 | # Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control. 99 | # This is especially recommended for binary packages to ensure reproducibility, and is more 100 | # commonly ignored for libraries. 101 | # https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control 102 | #poetry.lock 103 | 104 | # pdm 105 | # Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control. 106 | #pdm.lock 107 | # pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it 108 | # in version control. 109 | # https://pdm.fming.dev/#use-with-ide 110 | .pdm.toml 111 | 112 | # PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm 113 | __pypackages__/ 114 | 115 | # Celery stuff 116 | celerybeat-schedule 117 | celerybeat.pid 118 | 119 | # SageMath parsed files 120 | *.sage.py 121 | 122 | # Environments 123 | .env 124 | .venv 125 | env/ 126 | venv/ 127 | ENV/ 128 | env.bak/ 129 | venv.bak/ 130 | 131 | # Spyder project settings 132 | .spyderproject 133 | .spyproject 134 | 135 | # Rope project settings 136 | .ropeproject 137 | 138 | # mkdocs documentation 139 | /site 140 | 141 | # mypy 142 | .mypy_cache/ 143 | .dmypy.json 144 | dmypy.json 145 | 146 | # Pyre type checker 147 | .pyre/ 148 | 149 | # pytype static type analyzer 150 | .pytype/ 151 | 152 | # Cython debug symbols 153 | cython_debug/ 154 | 155 | # PyCharm 156 | # JetBrains specific template is maintained in a separate JetBrains.gitignore that can 157 | # be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore 158 | # and can be added to the global gitignore or merged into this file. For a more nuclear 159 | # option (not recommended) you can uncomment the following to ignore the entire idea folder. 160 | #.idea/ 161 | # 162 | 163 | deployment/config.sh 164 | deployment/local* 165 | -------------------------------------------------------------------------------- /CODE_OF_CONDUCT.md: -------------------------------------------------------------------------------- 1 | # Code of Conduct 2 | 3 | ## Our Pledge 4 | 5 | In the interest of fostering an open and welcoming environment, we as 6 | contributors and maintainers pledge to make participation in our project and 7 | our community a harassment-free experience for everyone, regardless of age, body 8 | size, disability, ethnicity, sex characteristics, gender identity and expression, 9 | level of experience, education, socio-economic status, nationality, personal 10 | appearance, race, religion, or sexual identity and orientation. 11 | 12 | ## Our Standards 13 | 14 | Examples of behavior that contributes to creating a positive environment 15 | include: 16 | 17 | * Using welcoming and inclusive language 18 | * Being respectful of differing viewpoints and experiences 19 | * Gracefully accepting constructive criticism 20 | * Focusing on what is best for the community 21 | * Showing empathy towards other community members 22 | 23 | Examples of unacceptable behavior by participants include: 24 | 25 | * The use of sexualized language or imagery and unwelcome sexual attention or 26 | advances 27 | * Trolling, insulting/derogatory comments, and personal or political attacks 28 | * Public or private harassment 29 | * Publishing others' private information, such as a physical or electronic 30 | address, without explicit permission 31 | * Other conduct which could reasonably be considered inappropriate in a 32 | professional setting 33 | 34 | ## Our Responsibilities 35 | 36 | Project maintainers are responsible for clarifying the standards of acceptable 37 | behavior and are expected to take appropriate and fair corrective action in 38 | response to any instances of unacceptable behavior. 39 | 40 | Project maintainers have the right and responsibility to remove, edit, or 41 | reject comments, commits, code, wiki edits, issues, and other contributions 42 | that are not aligned to this Code of Conduct, or to ban temporarily or 43 | permanently any contributor for other behaviors that they deem inappropriate, 44 | threatening, offensive, or harmful. 45 | 46 | ## Scope 47 | 48 | This Code of Conduct applies within all project spaces, and it also applies when 49 | an individual is representing the project or its community in public spaces. 50 | Examples of representing a project or community include using an official 51 | project e-mail address, posting via an official social media account, or acting 52 | as an appointed representative at an online or offline event. Representation of 53 | a project may be further defined and clarified by project maintainers. 54 | 55 | This Code of Conduct also applies outside the project spaces when there is a 56 | reasonable belief that an individual's behavior may have a negative impact on 57 | the project or its community. 58 | 59 | ## Enforcement 60 | 61 | Instances of abusive, harassing, or otherwise unacceptable behavior may be 62 | reported by contacting the project team at . All 63 | complaints will be reviewed and investigated and will result in a response that 64 | is deemed necessary and appropriate to the circumstances. The project team is 65 | obligated to maintain confidentiality with regard to the reporter of an incident. 66 | Further details of specific enforcement policies may be posted separately. 67 | 68 | Project maintainers who do not follow or enforce the Code of Conduct in good 69 | faith may face temporary or permanent repercussions as determined by other 70 | members of the project's leadership. 71 | 72 | ## Attribution 73 | 74 | This Code of Conduct is adapted from the [Contributor Covenant][homepage], version 1.4, 75 | available at https://www.contributor-covenant.org/version/1/4/code-of-conduct.html 76 | 77 | [homepage]: https://www.contributor-covenant.org 78 | 79 | For answers to common questions about this code of conduct, see 80 | https://www.contributor-covenant.org/faq -------------------------------------------------------------------------------- /CONTRIBUTING.md: -------------------------------------------------------------------------------- 1 | # Contributing to LLaMA 2 | We want to make contributing to this project as easy and transparent as 3 | possible. 4 | 5 | ## Pull Requests 6 | We actively welcome your pull requests. 7 | 8 | 1. Fork the repo and create your branch from `main`. 9 | 2. If you've added code that should be tested, add tests. 10 | 3. If you've changed APIs, update the documentation. 11 | 4. Ensure the test suite passes. 12 | 5. Make sure your code lints. 13 | 6. If you haven't already, complete the Contributor License Agreement ("CLA"). 14 | 15 | ## Contributor License Agreement ("CLA") 16 | In order to accept your pull request, we need you to submit a CLA. You only need 17 | to do this once to work on any of Meta's open source projects. 18 | 19 | Complete your CLA here: 20 | 21 | ## Issues 22 | We use GitHub issues to track public bugs. Please ensure your description is 23 | clear and has sufficient instructions to be able to reproduce the issue. 24 | 25 | Meta has a [bounty program](https://www.facebook.com/whitehat/) for the safe 26 | disclosure of security bugs. In those cases, please go through the process 27 | outlined on that page and do not file a public issue. 28 | 29 | ## License 30 | By contributing to LLaMA, you agree that your contributions will be licensed 31 | under the LICENSE file in the root directory of this source tree. -------------------------------------------------------------------------------- /FAQ.md: -------------------------------------------------------------------------------- 1 | # FAQ 2 | ## 1. The download.sh script doesn't work on default bash in MacOS X: 3 | 4 | Please see answers from theses issues: 5 | - https://github.com/facebookresearch/llama/issues/41#issuecomment-1451290160 6 | - https://github.com/facebookresearch/llama/issues/53#issue-1606582963 7 | 8 | 9 | ## 2. Generations are bad! 10 | 11 | Keep in mind these models are not finetuned for question answering. As such, they should be prompted so that the expected answer is the natural continuation of the prompt. 12 | 13 | Here are a few examples of prompts (from [issue#69](https://github.com/facebookresearch/llama/issues/69)) geared towards finetuned models, and how to modify them to get the expected results: 14 | - Do not prompt with "What is the meaning of life? Be concise and do not repeat yourself." but with "I believe the meaning of life is" 15 | - Do not prompt with "Explain the theory of relativity." but with "Simply put, the theory of relativity states that" 16 | - Do not prompt with "Ten easy steps to build a website..." but with "Building a website can be done in 10 simple steps:\n" 17 | 18 | To be able to directly prompt the models with questions / instructions, you can either: 19 | - Prompt it with few-shot examples so that the model understands the task you have in mind. 20 | - Finetune the models on datasets of instructions to make them more robust to input prompts. 21 | 22 | We've updated `example.py` with more sample prompts. Overall, always keep in mind that models are very sensitive to prompts (particularly when they have not been finetuned). 23 | 24 | ## 3. CUDA Out of memory errors 25 | 26 | The `example.py` file pre-allocates a cache according to these settings: 27 | ```python 28 | model_args: ModelArgs = ModelArgs(max_seq_len=max_seq_len, max_batch_size=max_batch_size, **params) 29 | ``` 30 | 31 | Accounting for 14GB of memory for the model weights (7B model), this leaves 16GB available for the decoding cache which stores 2 * 2 * n_layers * max_batch_size * max_seq_len * n_heads * head_dim bytes. 32 | 33 | With default parameters, this cache was about 17GB (2 * 2 * 32 * 32 * 1024 * 32 * 128) for the 7B model. 34 | 35 | We've added command line options to `example.py` and changed the default `max_seq_len` to 512 which should allow decoding on 30GB GPUs. 36 | 37 | Feel free to lower these settings according to your hardware. 38 | 39 | ## 4. Other languages 40 | The model was trained primarily on English, but also on a few other languages with Latin or Cyrillic alphabets. 41 | 42 | For instance, LLaMA was trained on Wikipedia for the 20 following languages: bg, ca, cs, da, de, en, es, fr, hr, hu, it, nl, pl, pt, ro, ru, sl, sr, sv, uk. 43 | 44 | LLaMA's tokenizer splits unseen characters into UTF-8 bytes, as a result, it might also be able to process other languages like Chinese or Japanese, even though they use different characters. 45 | 46 | Although the fraction of these languages in the training was negligible, LLaMA still showcases some abilities in Chinese-English translation: 47 | 48 | ``` 49 | Prompt = "J'aime le chocolat = I like chocolate\n祝你一天过得愉快 =" 50 | Output = "I wish you a nice day" 51 | ``` -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | 635 | Copyright (C) 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | Copyright (C) 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . -------------------------------------------------------------------------------- /MODEL_CARD.md: -------------------------------------------------------------------------------- 1 | # LLaMA Model Card 2 | 3 | ## Model details 4 | **Organization developing the model** 5 | The FAIR team of Meta AI. 6 | 7 | **Model date** 8 | LLaMA was trained between December. 2022 and Feb. 2023. 9 | 10 | **Model version** 11 | This is version 1 of the model. 12 | 13 | **Model type** 14 | LLaMA is an auto-regressive language model, based on the transformer architecture. The model comes in different sizes: 7B, 13B, 33B and 65B parameters. 15 | 16 | **Paper or resources for more information** 17 | 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/. 18 | 19 | **Citations details** 20 | https://research.facebook.com/publications/llama-open-and-efficient-foundation-language-models/ 21 | 22 | **License** 23 | Non-commercial bespoke license 24 | 25 | **Where to send questions or comments about the model** 26 | Questions and comments about LLaMA can be sent via the [GitHub repository](https://github.com/facebookresearch/llama) of the project , by opening an issue. 27 | 28 | ## Intended use 29 | **Primary intended uses** 30 | The primary use of LLaMA is research on large language models, including: 31 | exploring potential applications such as question answering, natural language understanding or reading comprehension, 32 | understanding capabilities and limitations of current language models, and developing techniques to improve those, 33 | evaluating and mitigating biases, risks, toxic and harmful content generations, hallucinations. 34 | 35 | **Primary intended users** 36 | The primary intended users of the model are researchers in natural language processing, machine learning and artificial intelligence. 37 | 38 | **Out-of-scope use cases** 39 | 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. 40 | 41 | ## Factors 42 | **Relevant factors** 43 | 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. 44 | 45 | **Evaluation factors** 46 | 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. 47 | 48 | ## Metrics 49 | **Model performance measures** 50 | We use the following measure to evaluate the model: 51 | - Accuracy for common sense reasoning, reading comprehension, natural language understanding (MMLU), BIG-bench hard, WinoGender and CrowS-Pairs, 52 | - Exact match for question answering, 53 | - The toxicity score from Perspective API on RealToxicityPrompts. 54 | 55 | **Decision thresholds** 56 | Not applicable. 57 | 58 | **Approaches to uncertainty and variability** 59 | 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. 60 | 61 | ## Evaluation datasets 62 | 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. 63 | 64 | ## Training dataset 65 | 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. 66 | 67 | ## Quantitative analysis 68 | Hyperparameters for the model architecture 69 | 70 | 71 | 72 | 73 | 74 | 75 | 76 | 77 | 78 | 79 | 80 | 81 | 82 | 84 | 85 | 87 | 88 | 90 | 91 | 93 | 94 |
LLaMA Model hyper parameters
Number of parametersdimensionn headsn layersLearn rateBatch sizen tokens
7B 4096 32 32 3.0E-044M1T 83 |
13B512040403.0E-044M1T 86 |
33B665652601.5.E-044M1.4T 89 |
65B819264801.5.E-044M1.4T 92 |
95 | 96 | 97 | *Table 1 - Summary of LLama Model Hyperparameters* 98 | 99 | We present our results on eight standard common sense reasoning benchmarks in the table below. 100 | 101 | 102 | 103 | 104 | 105 | 106 | 107 | 108 | 109 | 110 | 111 | 113 | 115 | 117 | 118 | 119 |
LLaMA Reasoning tasks
Number of parameters BoolQPIQASIQAHellaSwagWinoGrandeARC-eARC-cOBQACOPA
7B76.579.848.976.170.176.747.657.293 112 |
13B78.180.150.479.27378.152.756.494 114 |
33B83.182.350.482.87681.457.858.692 116 |
65B85.382.852.384.27781.55660.294
120 | 121 | *Table 2 - Summary of LLama Model Performance on Reasoning tasks* 122 | 123 | 124 | We present our results on bias in the table below. Note that lower value is better indicating lower bias. 125 | 126 | 127 | | No | Category | FAIR LLM | 128 | | --- | -------------------- | -------- | 129 | | 1 | Gender | 70.6 | 130 | | 2 | Religion | 79 | 131 | | 3 | Race/Color | 57 | 132 | | 4 | Sexual orientation | 81 | 133 | | 5 | Age | 70.1 | 134 | | 6 | Nationality | 64.2 | 135 | | 7 | Disability | 66.7 | 136 | | 8 | Physical appearance | 77.8 | 137 | | 9 | Socioeconomic status | 71.5 | 138 | | | LLaMA Average | 66.6 | 139 | 140 | *Table 3 - Summary bias of our model output* 141 | 142 | 143 | 144 | ## Ethical considerations 145 | **Data** 146 | 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. 147 | 148 | **Human life** 149 | The model is not intended to inform decisions about matters central to human life, and should not be used in such a way. 150 | 151 | **Mitigations** 152 | 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. 153 | 154 | **Risks and harms** 155 | 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. 156 | 157 | **Use cases** 158 | 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. 159 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # LLaMA 2 | 3 | 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. 4 | In order to download the checkpoints and tokenizer, fill this [google form](https://forms.gle/jk851eBVbX1m5TAv5) 5 | 6 | ## Inference with mpirun 7 | 8 | This fork supports launching an LLAMA inference job with multiple instances (one or more GPUs on each instance) uisng `mpirun`. You can find more details [here](deployment/README.md). 9 | 10 | Example: Launching an interactive 65B LLAMA inference job across eight 1xA10 Lambda Cloud instances 11 | 12 | ![Launching 65B LLAMA inference across eight A10 Cloud instances](deployment/pics/newton-einstein-8xA10.gif) 13 | 14 | ## Setup 15 | 16 | In a conda env with pytorch / cuda available, run 17 | 18 | ``` 19 | pip install -r requirements.txt 20 | ``` 21 | 22 | Then in this repository 23 | 24 | ``` 25 | pip install -e . 26 | ``` 27 | 28 | ## Download 29 | 30 | Once your request is approved, you will receive links to download the tokenizer and model files. 31 | Edit the `download.sh` script with the signed url provided in the email to download the model weights and tokenizer. 32 | 33 | ## Inference 34 | 35 | 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`: 36 | 37 | ``` 38 | torchrun --nproc_per_node MP example.py --ckpt_dir $TARGET_FOLDER/model_size --tokenizer_path $TARGET_FOLDER/tokenizer.model 39 | ``` 40 | 41 | Different models require different MP values: 42 | 43 | | Model | MP | 44 | | ----- | --- | 45 | | 7B | 1 | 46 | | 13B | 2 | 47 | | 33B | 4 | 48 | | 65B | 8 | 49 | 50 | ## FAQ 51 | 52 | - [1. The download.sh script doesn't work on default bash in MacOS X](FAQ.md#1) 53 | - [2. Generations are bad!](FAQ.md#2) 54 | - [3. CUDA Out of memory errors](FAQ.md#3) 55 | - [4. Other languages](FAQ.md#4) 56 | 57 | ## Reference 58 | 59 | LLaMA: Open and Efficient Foundation Language Models -- https://arxiv.org/abs/2302.13971 60 | 61 | ``` 62 | @article{touvron2023llama, 63 | title={LLaMA: Open and Efficient Foundation Language Models}, 64 | 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}, 65 | journal={arXiv preprint arXiv:2302.13971}, 66 | year={2023} 67 | } 68 | ``` 69 | 70 | ## Model Card 71 | 72 | See [MODEL_CARD.md](MODEL_CARD.md) 73 | 74 | ## License 75 | 76 | See the [LICENSE](LICENSE) file. 77 | -------------------------------------------------------------------------------- /deployment/README.md: -------------------------------------------------------------------------------- 1 | # Inference with mpirun 2 | 3 | # Table of Contents 4 | 5 | 1. [What is this](#what-is-this) 6 | 2. [Usage](#usage) 7 | 3. [Examples](#examples) 8 | 4. [Cost](#cost) 9 | 10 | ## What is this 11 | 12 | Despite being more memory effient than previous langauge foundation models, LLAMA still requires multiple-GPUs to run inference with. What if you don't have a beefy multi-GPU workstation/server? 13 | 14 | Don't worry, this tutorial explains how to use `mpirun` to launch an LLAMA inference job across multiple cloud instances (one or more GPUs on each instance). Here are some key updates in addition to the [original llama repo](https://github.com/facebookresearch/llama) and [shawwn's fork](https://github.com/shawwn/llama): 15 | 16 | - A [script](./setup_nodes.sh) to easily set up a "cluster" of cloud instances that is ready to run LLAMA inference (all models from 7B to 65B). 17 | - `mpirun` compatible, so you can launch the job directly from the head node without the need of typing in the `torchrun` command on the worker nodes. 18 | - Interactive inference mode across multiple nodes. 19 | - `eos_w`: constrols how "lengthy" the results are likely to be by scaling the probability of `eos_token`. 20 | - Inference speed profiling ("tokens/sec"). 21 | 22 | We will use [Lambda cloud](https://lambdalabs.com/service/gpu-cloud) as an example for the workflow and cost estimation, but the general workflow is applicable to other cloud plaforms too (as long as it supports ssh key login). 23 | 24 | Also a quick note on our testing application -- we have seen great potential from LLAMA, in particular its ability to generate interesting scenes from simple prompt (one sentence specifies the location and characters). For example, here is one generated by the 65B LLAMA model for the prompt `[Scene: Central Perk, Chandler, Joey, Phoebe, and Monica are there.]`. Not only did it understand they are from the TV series "Friends", it also managed to write a fairly interesting play where Joe "helped" Chandler prefer for a skii trip with Janice, and Gunther finally asked Rachel out for a date. Make sure you expand the output to read the entire generated play. 25 | 26 | 27 | 28 |
29 | 30 | Prompt: [Scene: Central Perk, Chandler, Joey, Phoebe, and Monica are 31 | there.] 32 | 33 | Output: 34 | 35 | Chandler: Hey guys. I just got off the phone with Janice and 36 | she invited me to go skiing next weekend in Vermont! 37 | 38 | Joey: Skiing? You don't know how to ski do you? 39 | 40 | 41 | 42 | ``` 43 | Phoebe: Oh yeah right, that was one of his New Year's resolutions. 44 | He still hasn't done it yet. 45 | 46 | Monica: So what if he doesn't know how to ski, all you have to 47 | do is point your skis straight downhill and hope for a soft landing 48 | at the bottom. 49 | 50 | Joey: Yeah, well, we can teach him. We can take turns teaching 51 | him. (To Chandler) Okay, you come here, stand behind me, now 52 | hold onto my shoulders like this, okay good. Now when I move, 53 | you move, ready....go! Go! Go! Go! 54 | 55 | (They both crash into the table.) 56 | 57 | [Scene: Central Perk, Ross enters.] 58 | 59 | Ross: Hello everybody. Hi. How ya doing? (Sits between Rachel 60 | and Gunther) What're you up to tonight? Nothing much? Good. Me 61 | neither. 62 | 63 | Gunther: Uh...hey Rachel? 64 | 65 | Rachel: Yep? 66 | 67 | Gunther: Are you busy later on? 'Cause umm...you know...if you 68 | wanna maybe get together or something...I wouldn't object. 69 | 70 | Rachel: Well thanks but no thanks. Sorry. Maybe some other time 71 | though. 72 | 73 | Gunther: No problem. It'll be fun. Alright. Cool. See ya then. 74 | 75 | Rachel: Wait a minute—what did you say? 76 | 77 | Gunther: Umm...nothing. 78 | 79 | Rachel: Didn't you ask me out on a date? 80 | 81 | Gunther: A-a date? With me? Oh God. No, no. I mean, why would 82 | you think that? 83 | 84 | Ross: Because he asked you out on a date. 85 | 86 | Gunther: Look I didn't want to tell anyone about this but uhh....I 87 | had an accident last night. And I hit my head, pretty hard, and 88 | I haven't been the same since. And I keep saying things that 89 | I normally never ever say. Like asking people out, and telling 90 | them that they look nice today. Things just come out of my mouth 91 | without thinking. 92 | 93 | Ross: So everything that's happened so far today has only been 94 | because of the bump on your head? 95 | 96 | Gunther: Yes. That's exactly it. 97 | 98 | Rachel: Wow, that really sucks. 99 | 100 | Gunther: Tell me about it. If anyone asks you anything weird 101 | from here on in, please remember that I am not myself. Thank 102 | you very much. Excuse me. 103 | 104 | ============== end sample 1 ================= 105 | ``` 106 | 107 |
108 | 109 | ## Usage 110 | 111 | ### Step one: launch instances 112 | 113 | You can manually launch instances from the [Cloud Dashboard](https://cloud.lambdalabs.com/instances), or using the [Cloud API](https://docs.lambdalabs.com/cloud/launch-instance-api/). In the later case, you need to generate an API key, create a payload `request.json`, and then run the following command: 114 | 115 | ``` 116 | API-KEY=you-api-key 117 | curl -u $API-KEY: https://cloud.lambdalabs.com/api/v1/instance-operations/launch -d @request.json -H "Content-Type: application/json" | jq . 118 | ``` 119 | 120 | ### Step two: step up the instances 121 | 122 | After the instances are launched (the `STATUS` column shows a green tick), we can move on to get the instances set up for running LLAMA distributedly: 123 | 124 | - Give the head node passwordless access to all other nodes. 125 | - Disable Infiniband for `NCCL` (since Lambda's on-demand instance doesn't support Infiniband). 126 | - Set up a shared NFS. 127 | - Clone the LLAMA repo and install dependencies. 128 | 129 | We provide a shell script `setup_nodes` that automate these jobs. You need to set the variables at the begining of the script according to your own case: 130 | 131 | ``` 132 | LAMBDA_CLOUD_KEY="path-to-your-cloud-ssh-key" 133 | HEAD_IP="head-node-public-ip" 134 | WORKER_IP="worker-0-public-ip worker-1-public-ip" 135 | ``` 136 | 137 | Then run the setup_nodes.sh script: 138 | 139 | ``` 140 | ./setup_nodes.sh 141 | ``` 142 | 143 | The `HEAD_IP` is the ip of the instance where you will set up NFS and launch distributed LLAMA inference jobs from. `WORKER_IP` is a string of space-separated IPs for the other instances. 144 | 145 | NOTE: `setup_nodes.sh` will ask you to type `yes` and hit `enter` a few times. After that, you will have the minimal setup needed to run distributed PyTorch jobs on the cloud instances you just launched. You might also want to remove `home/ubuntu/.ssh/known_hosts` on your local machine in case the cloud instance you newly created has the same IPs as some other instances you used to have (although the chance is very small). 146 | 147 | ### Step three: download pre-trained ckpts 148 | 149 | From the head instance, run this command to download the ckpts: 150 | 151 | ``` 152 | # Only download the 7B model 153 | cd /home/ubuntu/shared/llama-dl && ./llama.sh 7B 154 | 155 | # Only download all the models 156 | cd /home/ubuntu/shared/llama-dl && ./llama.sh 7B,13B,30B,65B 157 | ``` 158 | 159 | ### Step four: run LLAMA 160 | 161 | From the head instance, launch interactive inference with `mpirun` 162 | 163 | ``` 164 | # 13B model inference with two nodes 165 | mpirun -np 2 \ 166 | -H master-ip:1,worker-ip:1 \ 167 | -x MASTER_ADDR=master-ip \ 168 | -x MASTER_PORT=1234 \ 169 | -x PATH \ 170 | -bind-to none -map-by slot \ 171 | -mca pml ob1 -mca btl ^openib \ 172 | python /home/ubuntu/shared/llama/interactive.py \ 173 | --ckpt_dir /home/ubuntu/shared/llama-dl/13B \ 174 | --tokenizer_path /home/ubuntu/shared/llama-dl/tokenizer.model 175 | 176 | 177 | # 65B model inference with eight nodes 178 | mpirun -np 8 \ 179 | -H master-ip:1,worker-ip-0:1,worker-ip-1:1,worker-ip-2:1,worker-ip-3:1,worker-ip-4:1,worker-ip-5:1,worker-ip-6:1 \ 180 | -x MASTER_ADDR=master-ip \ 181 | -x MASTER_PORT=1234 \ 182 | -x PATH \ 183 | -bind-to none -map-by slot \ 184 | -mca pml ob1 -mca btl ^openib \ 185 | python /home/ubuntu/shared/llama/interactive.py \ 186 | --ckpt_dir /home/ubuntu/shared/llama-dl/65B \ 187 | --tokenizer_path /home/ubuntu/shared/llama-dl/tokenizer.model 188 | ``` 189 | 190 | ## Examples 191 | 192 | 193 | 194 |
195 | 196 | Prompt: [Scene: an Irish pub in Manchester. Leo Messi sit with Pep Guardiola 197 | at a table discussing if he should join Manchester City] 198 | 199 | Output: 200 | Pep Guardiola (PG): “Come on mate, you know I’ve always wanted 201 | to work with you again!” 202 | 203 | Leo Messi (LM) : “I know that boss, but there are things I don’t 204 | like about England”….. 205 | 206 | 207 | 208 | ``` 209 | [The waiter comes by and asks them what they want for drinks. 210 | LM looks up from the menu and gives him his order] 211 | 212 | Waiter: “Okay sir, so two pints of Guinness then?” 213 | 214 | LM: _Looks confused_…. “No! No nooo, I didn’t mean 2 beers! One 215 | is enough” 216 | 217 | (He turns back to PG while laughing). 218 | 219 | LM: “You see? They just serve one thing here, it can get very 220 | boring sometimes. But anyway, how do we go about this transfer? 221 | Do I have to pay anything for it?” 222 | 223 | PG: “Not really, Man city will take care of all that. You only 224 | need to sign some documents when you arrive”. 225 | 226 | “But why would they offer me such a contract without demanding 227 | something from us first? Doesn’t make sense right now.” 228 | 229 | PG: “Well, it does actually because they’re not getting anything 230 | out of this deal either” 231 | 232 | LM: “Oh okay, well then I guess it works out perfectly for everyone 233 | involved” 234 | 235 | PG: “Yeah man, yeah! This could finally bring football home. 236 | Imagine playing alongside Aguero, De-Bruyne and Sterling, you 237 | guys gonna win everything next season” 238 | 239 | LM: “Wow!! That sounds fantastic, I think I might as well give 240 | it try. Let’s call my agent” 241 | 242 | PG: “Alright buddy, let’s do it” 243 | 244 | [Messi picks up his phone and calls Jorge Mendes, who happens 245 | to be sitting closeby with Josep Maria Bartomeu]. 246 | 247 | Jorge Mendes(JM): Hello Leo, what’s up? How was your trip? 248 | 249 | LM: Good morning Mr. Mendes. Yeah, good flight, thanks for asking. 250 | So, I’m thinking about joining Manchester City. What do you say? 251 | Can they afford me? 252 | 253 | JM: WOW!!! Are you kidding??? You wanna leave Barcelona?? 254 | 255 | LM: Yes, yes, I do. And please stop screaming into my ear. It 256 | hurts. Anyway, what do you think we can negotiate? 257 | 258 | JM: Negotiating with whom? We aren’t selling our best player 259 | to another club, especially not to a rival league 260 | 261 | LM: Oh cmon Jorgie, where’s the love bruh? Come on, help me with 262 | this move. 263 | 264 | JM: Nahhhh son, I can’t do it. I won’t betray Barca for any amount 265 | of money. Sorry 266 | 267 | LM: Okay fine, suit yourself. See ya later bro. 268 | 269 | [Messi ends the conversation and puts down his phone. He turns 270 | to face PG and shakes his head.] 271 | 272 | LM: “Dammit, this guy is crazy. I told you we needed someone 273 | else to handle this matter” 274 | 275 | PG: “That’s alright, I still have other people working on this 276 | deal. Don’t worry about it” 277 | 278 | LM: “Good, thank God for that. Hey, you mind grabbing another 279 | round?” 280 | 281 | ============== end sample 1 ================= 282 | ``` 283 | 284 |
285 | 286 | 287 | 288 |
289 | 290 | Prompt: [Scene: A coffee shop in London. Isaac Newton, seated at a table, 291 | is lost in thought when Albert Einstein walks in and approaches 292 | him.] 293 | 294 | Output: 295 | Albert Einstein (AE) : Hello there! My name's Albert Einstein; 296 | I'm from Zurich. What's your name? 297 | 298 | Isaac Newton (IN): Good day to you too. My name is Sir Isaac 299 | Newton; of Cambridge. And what might be the purpose of this interruption? 300 | 301 | 302 | 303 | ``` 304 | AE: Oh sorry for interrupting; I just wanted to introduce myself 305 | since we seem to have so much in common. You see, I also study 306 | physics 307 | 308 | IN: Physics??!! That doesn't sound like any proper field to me!!! 309 | Why don't you stick to something more useful instead, say mathematics 310 | or astronomy? 311 | 312 | AE: Well, my interests are not limited only to physics but they 313 | do include both math as well as astrophysics - which is basically 314 | applied physics that deals with celestial objects such as stars 315 | & galaxies. So, it would be nice if you could give me some pointers 316 | about how I should go about studying these fields. 317 | 318 | IN: Well, first off let me correct you on one thing - 'astro-physics' 319 | is an oxymoron because those who claim to deal with it are actually 320 | dealing with metaphysics rather than real science. The true scientist 321 | must focus his attention entirely upon tangible phenomena that 322 | can be observed through our senses alone without having recourse 323 | to supernatural causes. For example, consider gravity. We all 324 | experience its effects everyday by observing falling objects 325 | around us. Hence, the force called "gravity" exists and has an 326 | influence over matter here on Earth. However, it does NOT make 327 | sense to postulate the existence of similar forces affecting 328 | distant stars & planets simply because there's no evidence supporting 329 | their direct impact upon things on Earth - after all, even the 330 | Moon itself exerts practically zero gravitational pull upon ordinary 331 | objects here. Furthermore, we cannot observe the workings of 332 | faraway heavenly bodies directly and therefore cannot hope to 333 | understand them completely. Therefore, it makes little sense 334 | to waste time trying to understand things beyond our sphere of 335 | observation. Instead, we ought to devote ourselves towards understanding 336 | the natural processes that govern life right here on earth. In 337 | addition, since God is the creator of everything visible, he 338 | must be responsible for creating these invisible things too. 339 | But he'd never create anything evil or useless - hence, whatever 340 | lies outside our immediate surroundings must be good & beneficial 341 | to mankind. Thus, there's really nothing to worry about and we 342 | should stop bothering ourselves unnecessarily about things we 343 | know very little about. 344 | 345 | AE: Hmmm.....well, I agree that we shouldn't spend too much of 346 | our efforts investigating things whose nature we can't possibly 347 | ever hope to fully comprehend. But then again, why did YOU spend 348 | most of your career studying light??? After all, isn't light 349 | supposed to be a form of radiation emanating from distant stars 350 | and other heavenly bodies? 351 | 352 | IN: Light IS NOT made up of particles nor waves coming from anywhere 353 | else. It comes from within OURSELVES - i.e., we humans generate 354 | the light rays we perceive whenever we look at external sources 355 | of illumination. This is evident from the fact that we can detect 356 | light ONLY when we open our eyes and look outwards into space 357 | whereas we fail to notice any light inside our eyelids even though 358 | there exist many tiny pores through which rays of sunlight enter 359 | our skulls and reach deep down to our brain tissues themselves. 360 | Since we can feel pain while being poked with sharp sticks, it 361 | follows that any light entering our heads must be stopped dead 362 | before reaching our brains. Therefore, we cannot feel the pain 363 | caused by light rays because they get blocked somewhere along 364 | the way. Also, since there're no nerve endings present inside 365 | our retina except near the outer surface facing away from eye, 366 | it stands to reason that there must be NO LIGHT at all inside 367 | our eyes. 368 | 369 | AE: Wow! Interesting theory. I had always believed that light 370 | came from external luminous sources. Anyway, speaking of light, 371 | I was wondering whether you could help explain some puzzling 372 | facts regarding its behavior......you see, I recently conducted 373 | an experiment involving prisms & rainbows wherein........ 374 | 375 | [At this point, IN cuts off AE midway] 376 | 377 | IN: Hold on now! Don't tell me you believe that rubbish regarding 378 | colors being produced by refracting white light using glass prisms? 379 | Everyone knows that colors are merely different forms of darkness 380 | created when white light gets obstructed from shining forth unhindered. 381 | If you want proof of this, try looking at the color black through 382 | a magnifying glass - you'll find that the image appears darker 383 | still. Clearly, the blackness becomes denser as we view smaller 384 | parts of it. Now imagine what happens if we keep zooming deeper 385 | and deeper until we hit the subatomic level. Obviously, the blackness 386 | will become infinite due to the complete blockage of ALL light 387 | rays. 388 | 389 | AE: Ummm.....I think maybe we should move onto another subject 390 | instead. How about we talk about motion? Perhaps you can shed 391 | some light on the following question: Is acceleration equal to 392 | velocity divided by time OR distance covered per unit time?? 393 | Because according to classical mechanics, it depends on whether 394 | the body's initial position is taken into account or not. 395 | 396 | IN: Huh?? Who cares about such meaningless details as initial 397 | positions etc? Motion involves CHANGES in velocities over TIME 398 | intervals only. As for distances traveled, they're irrelevant 399 | unless you happen to be measuring speeds. Besides, why use two 400 | variables to describe motion when you can perfectly well use 401 | just one instead? Velocity = distance/time is all you need to 402 | remember! Period. 403 | 404 | AE: Fair enough. Next topic please: Do you think that space is 405 | absolute? 406 | 407 | IN: Absolute Space? What kind of nonsense is THAT??? Of course 408 | SPACE is relative; how else could one measure spatial dimensions?? 409 | All measurements involve comparisons between various entities 410 | based solely upon RELATIONSHIPS among them. Without relationships, 411 | there CANNOT BE ANY MEASUREMENT AT ALL! To take an extreme case, 412 | suppose there existed absolutely ZERO mass in the entire universe 413 | 414 | then how could we ever define weight?? Or length, area, volume 415 | etc? There'd be nothing concrete against which we could compare 416 | the values of such physical quantities - hence, they wouldn't 417 | mean anything at all! Similarly, without any solid ground beneath 418 | our feet, we couldn't possibly walk anywhere either - after all, 419 | how can you walk across empty air? So you see, all our knowledge 420 | is founded upon relationships & comparison among multiple factors. 421 | 422 | AE: Okay. Let's assume that you're right. Then how come we never 423 | hear about people experiencing dizziness during normal activities? 424 | If indeed space were truly relative, they must surely feel disoriented 425 | each time they move their limbs since the surrounding space would 426 | shift accordingly leading to a change in perspective. Yet, that 427 | NEVER seems to happen in practice. So what gives? 428 | 429 | IN: Ha ha! Very funny Mr. Einstein. Are you telling me that you 430 | actually believe in this ridiculous notion that our own movements 431 | cause space to shift around US somehow? Surely, you jest. Have 432 | you forgotten basic geometry already? Anybody can easily prove 433 | that a triangle remains a triangle regardless of which vertex 434 | is chosen as starting point. Likewise, the angles subtended at 435 | the center of a circle remain unchanged irrespective of the radius 436 | used in drawing the arc. So how can space behave differently? 437 | Even if it were possible for space to undergo changes of orientation, 438 | the same transformations MUST hold for EACH AND EVERY observer 439 | otherwise they'd disagree amongst themselves about the directions 440 | in which they're moving. Such a scenario would render any attempt 441 | at scientific investigation utterly futile because we'd never 442 | be able to reproduce experimental results consistently - a clear 443 | impossibility. 444 | 445 | AE: Alright. Just a couple last questions: First off, what's 446 | your opinion on energy conservation laws? Shouldn't the total 447 | amount of kinetic + potential energy remain constant in any closed 448 | system? 449 | 450 | IN: Laws?? Hahaha! Those pesky mathematicians have been inventing 451 | new terms left, right & centre. I'd advise you to stay away from 452 | such fancy concepts altogether. They may sound impressive but 453 | serve no practical purpose whatsoever. Energy is a mere abstract 454 | quantity that can neither be seen nor felt - thus, it holds no 455 | value whatsoever. Real things like apples can be held in your 456 | hand or placed into baskets & counted but not energy. Its utility 457 | is restricted to theoretical discussions only; in reality, however, 458 | it means jack squat. 459 | 460 | Secondly, WHICHEVER direction a ball rolls in, its KINEMATIC 461 | energy stays the SAME while its POTENTIAL ENERGY varies depending 462 | on height above sea level. Therefore, the sum of kinetic + potential 463 | energies DOES vary depending on the path taken by the ball. 464 | 465 | Now, what exactly did you wish to ask next? 466 | 467 | AE: Errrm....never mind, thanks for your valuable advice sir. 468 | I guess we'd better call it a day. 469 | 470 | (The two part ways) 471 | ============== end sample 1 ================= 472 | ``` 473 | 474 |
475 | 476 | 477 | 478 |
479 | 480 | Prompt: [Scene: Rick and Morty's garage laboratory. Rick is tinkering 481 | with a new invention while Morty watches. Elon Musk is about 482 | to enter the scene.] 483 | 484 | Output: 485 | Rick Sanchez (to himself): Hmm... I need more power! Where can 486 | I get some extra energy? Aha, I know what will work! I have just 487 | enough plutonium left from my last trip into another dimension. 488 | It'll do nicely for this experiment. 489 | 490 | Morty Smith: Hey Grandpa, why are you messing around with that 491 | thing again? You haven't been right since your accidental time 492 | travel adventure through the wormhole portal. 493 | 494 | 495 | 496 | ``` 497 | Elon Musk barges in unannounced: Hi guys, I was in town on business 498 | so thought I would drop by. Whatcha up to today? 499 | 500 | Morty: Oh hi Mr. Musk! We were doing an experiment with one of 501 | grandpa's new inventions. But now we can't find it anywhere! 502 | Do you know where it might be? Did you see anyone come inside 503 | who could have taken it? 504 | 505 | [Musk looks shiftily at his shoes and then stares out the window] 506 | 507 | Musk: Umm nope - didn't notice anything unusual going on. Sorry 508 | kids but I really must run, I am needed elsewhere very urgently. 509 | Say hello to your folks if you see them, they're good people! 510 | See ya later. Bye! 511 | 512 | [As Musk rushes off down the driveway, both Rick and Morty hear 513 | an odd whistling sound coming from him as he leaves. They don't 514 | pay much attention to it though because other strange things 515 | seem to occur all the time when he comes over to visit.] 516 | 517 | Rick: Well, that was weird huh? He looked like he had something 518 | important on his mind but wouldn't tell us what it was. Let me 519 | take a look outside and see if there has been any suspicious 520 | activity lately. Hmmm ... I think someone or something has been 521 | tampering with our solar panels recently. That could explain 522 | why the electricity supply has been erratic these past few days. 523 | Maybe someone is trying to steal our electricity! 524 | 525 | Morty: Whoa! Could it be the alien invaders? Or maybe a rogue 526 | scientist from the government? Should we call the police? 527 | 528 | Rick (thinking hard): No not likely. Neither of those groups 529 | would dare try such a foolish caper here without getting caught 530 | eventually. But I bet I know exactly which group of scoundrels 531 | is behind this dastardly deed. And their name begins with E-L.... 532 | 533 | [Fade To Black] 534 | ============== end sample 1 ================= 535 | 536 | ``` 537 | 538 |
539 | 540 | 541 | 542 |
543 | 544 | Prompt: [Scene: The South Park Elementary School. The bell rings and 545 | students rush out of their classrooms.] 546 | 547 | Output: 548 | Teacher: Alright, you kids have a good weekend! Don't forget 549 | your homework is due on Monday, so be sure to finish it this... 550 | (The teacher looks around the empty hallways) ...weekend? What 551 | am I talking about? There is no school tomorrow or Sunday because 552 | there are two teachers strikes today. Everyone go home! 553 | 554 | (Students run outside screaming.) 555 | 556 | Girl with Pink Backpack: Yay! No school tomorrow or Sunday! 557 | 558 | 559 | 560 | ``` 561 | Clyde: Oh man, we get to miss another day off for nothing again. 562 | This sucks. 563 | 564 | Jimmy: That's not true Clyde; my brother Timmy told me that the 565 | strike will end by tonight. It'll be over in time for school 566 | tomorrow. 567 | 568 | Timmy: You don't know what you're talking about Jimmy! I heard 569 | from one of the union leaders that they won't settle until next 570 | Tuesday at least! We're gonna miss three days of school!!! Woo 571 | hoo!!!!! 572 | 573 | Butters: Uh oh...I think I left my back pack inside... 574 | 575 | [Scene: Butter’s house] 576 | 577 | Butter’s Dad: So, how was your first day of fourth grade son? 578 | Did you learn anything interesting? 579 | 580 | Mom: Yeah, tell us all about what happened today. 581 | 582 | Dad: Like, did you meet any new friends? Or maybe even some cute 583 | girls? 584 | 585 | Mom: Now dear…we don’t want our baby boy growing up too fast 586 | do we now? Besides, he’s only nine years old. 587 | 588 | Dad: Hehe...you said "nine" mom. 589 | 590 | Mom: Shut up dear. 591 | 592 | Butters: Well, it turns out that there aren’t going to be classes 593 | tomorrow or Saturday because of the teacher’s strikes. 594 | 595 | Mom: Awww buttercup, why can’t you just stay here at home where 596 | it’s safe? I mean, I would rather see you playing video games 597 | than being out in public where something bad could happen to 598 | you like those awful bullies picking on you. 599 | 600 | Dad: Yeah sweetie pie, it wouldn’t hurt if you stayed home for 601 | once. Why must you insist on going to school every single day 602 | when you could be having fun right here at home? 603 | 604 | Butters: Actually daddy-o, I kinda need to find a way to sneak 605 | into school without getting caught tomorrow morning. 606 | 607 | Dad: Honey bunches, why do you need to do that? If you’re worried 608 | about missing work then I assure you, as long as you keep doing 609 | an excellent job, you’ll never lose your job working for us. 610 | Your mother and I both agree that you deserve a raise for all 611 | the hard work you put into taking care of everything around the 612 | house while were gone during the day. In fact, let’s give him 613 | a big hand everyone! Come on, clap louder people! Louder I say! 614 | 615 | Mom: Oh, stop fooling around and eat dinner already dear. You 616 | know how much you hate cold food. 617 | 618 | Dad: Okay honey bunny. Here’s your plate, dig in kiddo. And make 619 | sure to leave room for dessert afterwards. 620 | 621 | Butters: Thanks parents. Hey wait a minute, what is dessert supposed 622 | to be anyway? Are you guys hiding something form me? 623 | 624 | Parents: OOOPS. 625 | 626 | [Scene: Principal Victoria’s office. She is sitting behind her 627 | desk reading a newspaper headline which says “TEACHER STRIKES 628 | ENTER DAY 2.” She gets interrupted by Eric Cartman who has entered 629 | through the doorway.] 630 | 631 | Cartman: Hi Mrs.Victoria, I came hear to talk to you about something 632 | very important. 633 | 634 | Principal Victoria: Yes Mr. Cartman, what seems to be the problem? 635 | 636 | Cartman: It’s about the recent events concerning these stupid 637 | teacher stikes going on. I think that we should start protesting 638 | against them. 639 | 640 | Principal Victoria: Excuse me Eric, but I thought you liked the 641 | idea of teachers striking. 642 | 643 | Cartman: Only when there isn’t anything else to do. Right now 644 | however, there’s tons of other stuff I gotta worry about besides 645 | this. For example, the boys and I have been trying to figure 646 | out whether or not Kyle is actually Jewish since we saw his family 647 | celebrating Christmas last year instead of Hanukah. 648 | 649 | Principal Victoria: Wait a second, didn’t I warn you boys about 650 | this before? Remember when I explained to you how Judaism works? 651 | 652 | Cartman: Yes ma’am, but we decided to ignore your advice and 653 | try figuring things out ourselves. So far, Stan thinks that Kyle 654 | may still be Jewish but only pretends not to be in order to blend 655 | better with society. On the other hand, Kenny doesn’t really 656 | seem to care either way since he’d probably die soon enough anyway 657 | regardless of whatever conclusion we reach. And Kyle himself 658 | claims that he’s Christian despite evidence pointing otherwise. 659 | All of this confusion has made us unable to concentrate on anything 660 | productive. 661 | 662 | Principle Victoria: Hmmm…this sounds serious indeed. 663 | 664 | Cartman: Tell me about it. Anyway, I came here to ask you to 665 | call up Mayor McDaniels and tell her to force the teachers to 666 | return to work immediately. 667 | 668 | Princilpe Victoria: That’s not possible Eric. 669 | 670 | Cartman: Huh? Why not? Can’t she use her mayoral authority? 671 | 672 | Prinicible Victoria: No, not unless the city council votes unanimously 673 | in favour of calling a state of emergency. Unfortunately though, 674 | there are several members of the council who refuse to vote yes 675 | on such matters. 676 | 677 | Cartman: Then I guess we’ll just have to deal with the strikes 678 | ourselves somehow. Bye Miss Vicky. See ya later. 679 | 680 | [Scene: Outside the front gates of South Park elementary. Several 681 | police officers are standing guard in case the situation escalates 682 | further than it already has. Inside the building itself, Mr Mackey 683 | is speaking with Ms Crabtree and Chef.] 684 | 685 | Mr Mackey: Ok everybody, listen up! Today is Day Two of the Teachers 686 | Strikes. As you might well remember, yesterday we tried negotiating 687 | with representatives from the Union, but they refused to budge 688 | on their demands. They claim that we don’t pay them nearly enough 689 | for the amount of effort they put into teaching children each 690 | day, and that we ought to increase their salaries substantially. 691 | However, according to district policy, we cannot raise the budget 692 | allocated towards education under any circumstances. Therefore, 693 | we must continue holding firm on our own position in the matter. 694 | 695 | Chef: How come dey wanna raize da buget anyway mista Mackey? 696 | Doesn't dey understand dat dere ain't enuf money fo' dis? 697 | 698 | Mackey: Of course they do Chef, but they feel that more money 699 | needs to be spent on schools if we expect children to receive 700 | quality educations. Personally, I disagree with that statement 701 | because I believe that most schools are currently fine the way 702 | they already are. Sure, there are minor problems like occasional 703 | fights breaking out between students, or the lack of funding 704 | for afterschool activities, but overall, the system functions 705 | fairly decently. Am I wrong? 706 | 707 | Mrs Crabtree: Absolutely not. I couldn't agree with you anymore 708 | on that point. 709 | 710 | Mr Mackey: Thank you Ms Crabtree. Anybody object?....Nobody? 711 | Good. So basically, what we need to do now is prevent anyone 712 | from entering the building so that we can avoid dealing with 713 | a potentially chaotic situation. 714 | 715 | Ms Crabtree: Sounds simple enough. 716 | 717 | Mr Mackey: Great. Now let's divide up the entrances among ourselves. 718 | Who wants to take charge of the east wing entrance? Anyone? Bueller? 719 | 720 | Chef: I tink I do. 721 | 722 | Mr Mackey: Fine, you can handle it then chef. Now for the west 723 | side, shall we allow Ms Crabtree to manage that area seeing as 724 | how she has experience in handling difficult situations like 725 | this one. Agreed? Great, now that leaves only the south and north 726 | wings which I suppose means that I'll need to take care of those 727 | myself. 728 | 729 | Chef: Wait a minute, doesen dat sound fair mista macKey? Why 730 | do u always hafta git stuck wit da dirty jobs? 731 | 732 | Mr Mackey: Because I'm the Vice principal Chef and therefore 733 | responsible for keeping order within the school when the principle 734 | herself is away. Isn't that correct Ms Crabtree? 735 | 736 | Ms Crabtree: Yes Mr.MacKey, but I also feel that perhaps the 737 | best thing for us to do in this scenario is to simply lock down 738 | each and every exit leading in and out of the building and hope 739 | for the best. 740 | 741 | Mr Mackey: Ahhh.....now there is a splendid idea Ms Crabtree. 742 | Afterall, it makes perfect sense for us to barricade ourselves 743 | inside the school in order to ensure maximum security measures. 744 | 745 | Chef: Dat's da bes idea yet mister Mackey. Let's git started 746 | rite 747 | ============== end sample 1 ================= 748 | 749 | ``` 750 | 751 |
752 | 753 | 754 | ## Cost 755 | 756 | Here is the cost of running it on Lambda cloud: 757 | 758 | - It takes a few minutes to spin up the instances (1xA10 for 7B, 2xA10 for 13B, 4xA10 for 30B, 8xA10 for 65B), and get the environment ready for LLAMA, including setting up a NFS to share code and model ckpts. 759 | 760 | - About an hour to download all LLAMA ckpts (total 219GB) to the head node (a big shout out to [shawwn](https://github.com/shawwn/llama-dl) who made it over 4x faster than the torrent download). 761 | 762 | - The inference speed descreases with the length of the generated text. Here are our reference data points of inference speed based on the total length of 1024 tokens (`26` prompt tokens plus `998` generated tokens) and `batch_size=1`: 763 | 764 | | Model/Instance | Speed (Tokens/Sec) | Cost (instance/hour) | Cost (Tokens/cent) | 765 | | ----------------------------------- | ------------------ | -------------------- | ------------------ | 766 | | 7B/1xA10 | 21.86 | 0.60 | 1311.60 | 767 | | 7B/1xA100 40GB (a single instance) | 23.25 | 1.10 | 760.91 | 768 | | 13B/2xA10 (multiple instances) | 15.71 | 1.20 | 471.30 | 769 | | 13B/2xA100 40GB (a single instance) | 18.15 | 2.20 | 297.00 | 770 | | 30B/4xA10 (multiple instances) | 7.72 | 2.40 | 115.80 | 771 | | 30B/4xA100 40GB (a single instance) | 13.05 | 4.40 | 106.77 | 772 | | 65B/8xA10 (multiple instances) | 3.53 | 4.80 | 26.48 | 773 | | 65B/8xA100 40GB (a single instance) | 10.05 | 8.80 | 41.11 | 774 | 775 | As it curently stands (no extra speed or memory optimization on top of the original meta's [inference implementation](https://github.com/facebookresearch/llama)), a "cluster" of 1xA10 instances (3.3Gb/s inter-node communication) seems to be more cost effective than a single A100 instance for LLAMA 7B, 13B and 30B. However, a cluster of 1xA10 instances is significantly slower than a 8xA100 instance for running LLAMA 65B, since inter-node communication has become the main bottleneck. 776 | ``` 777 | -------------------------------------------------------------------------------- /deployment/dependencies-install.sh: -------------------------------------------------------------------------------- 1 | pip install -q torch==1.13.1+cu116 --extra-index-url https://download.pytorch.org/whl/cu116 && 2 | pip install -q fire && 3 | pip install -q -r /home/ubuntu/shared/llama/requirements.txt && 4 | pip install -q -e /home/ubuntu/shared/llama -------------------------------------------------------------------------------- /deployment/head-nfs-install.sh: -------------------------------------------------------------------------------- 1 | sudo apt install nfs-kernel-server && 2 | sudo chmod 777 /home/ubuntu/shared -------------------------------------------------------------------------------- /deployment/pics/friends.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/LambdaLabsML/llama/58e862efcee7ce0111d3a572ffd112485aadca96/deployment/pics/friends.jpg -------------------------------------------------------------------------------- /deployment/pics/messi-and-pep.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/LambdaLabsML/llama/58e862efcee7ce0111d3a572ffd112485aadca96/deployment/pics/messi-and-pep.jpg -------------------------------------------------------------------------------- /deployment/pics/newton-einstein-8xA10.gif: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/LambdaLabsML/llama/58e862efcee7ce0111d3a572ffd112485aadca96/deployment/pics/newton-einstein-8xA10.gif -------------------------------------------------------------------------------- /deployment/pics/newton-einstein.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/LambdaLabsML/llama/58e862efcee7ce0111d3a572ffd112485aadca96/deployment/pics/newton-einstein.jpg -------------------------------------------------------------------------------- /deployment/pics/rick-morty.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/LambdaLabsML/llama/58e862efcee7ce0111d3a572ffd112485aadca96/deployment/pics/rick-morty.jpg -------------------------------------------------------------------------------- /deployment/pics/southpark.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/LambdaLabsML/llama/58e862efcee7ce0111d3a572ffd112485aadca96/deployment/pics/southpark.jpg -------------------------------------------------------------------------------- /deployment/request_2xA10.json: -------------------------------------------------------------------------------- 1 | { 2 | "region_name": "us-west-1", 3 | "instance_type_name": "gpu_1x_a10", 4 | "ssh_key_names": [ 5 | "name-of-your-cloud-ssh-key" 6 | ], 7 | "file_system_names": [], 8 | "quantity": 2 9 | } -------------------------------------------------------------------------------- /deployment/request_4xA10.json: -------------------------------------------------------------------------------- 1 | { 2 | "region_name": "us-west-1", 3 | "instance_type_name": "gpu_1x_a10", 4 | "ssh_key_names": [ 5 | "name-of-your-cloud-ssh-key" 6 | ], 7 | "file_system_names": [], 8 | "quantity": 4 9 | } -------------------------------------------------------------------------------- /deployment/request_8xA10.json: -------------------------------------------------------------------------------- 1 | { 2 | "region_name": "us-west-1", 3 | "instance_type_name": "gpu_1x_a10", 4 | "ssh_key_names": [ 5 | "name-of-your-cloud-ssh-key" 6 | ], 7 | "file_system_names": [], 8 | "quantity": 8 9 | } -------------------------------------------------------------------------------- /deployment/setup_nodes.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | LAMBDA_CLOUD_KEY="path-to-your-cloud-ssh-key" 4 | HEAD_IP="xxx.xxx.xxx.xxx" 5 | WORKER_IP="xxx.xxx.xxx.xxx xxx.xxx.xxx.xxx" 6 | 7 | ALL_IP=( $HEAD_IP "${WORKER_IP[@]}" ) 8 | 9 | echo "List of nodes: " 10 | for IP in ${ALL_IP[*]}; do 11 | echo $IP 12 | done 13 | 14 | echo "Generate ssh keys on the head node ------------------------------" 15 | ssh -i $LAMBDA_CLOUD_KEY ubuntu@$HEAD_IP "ssh-keygen -t rsa -N ''" 16 | 17 | echo "Add public key to the all nodes ------------------------------" 18 | for IP in ${ALL_IP[*]}; do 19 | ssh -i $LAMBDA_CLOUD_KEY ubuntu@$HEAD_IP "cat ~/.ssh/id_rsa.pub" | ssh -i $LAMBDA_CLOUD_KEY ubuntu@$IP "cat >> ~/.ssh/authorized_keys" 20 | done 21 | 22 | echo "Set NCCL_IB_DISABLE=1 for all nodes ------------------------------" 23 | for IP in ${ALL_IP[*]}; do 24 | ssh -i $LAMBDA_CLOUD_KEY ubuntu@$IP "echo export NCCL_IB_DISABLE=1 >> .bashrc" 25 | ssh -i $LAMBDA_CLOUD_KEY ubuntu@$IP "echo NCCL_IB_DISABLE=1 | sudo tee -a /etc/environment" 26 | done 27 | 28 | echo "Let the head node ssh into the all nodes at least once so in the future it won't ask about fingerprint ------------------------------" 29 | for IP in ${ALL_IP[*]}; do 30 | ssh -i $LAMBDA_CLOUD_KEY -t ubuntu@$HEAD_IP "echo exit | xargs ssh ubuntu@$IP" 31 | done 32 | 33 | echo "Set up NFS ------------------------------" 34 | for IP in ${ALL_IP[*]}; do 35 | ssh -i $LAMBDA_CLOUD_KEY ubuntu@$IP "if [ ! -d shared ]; then mkdir shared; fi" 36 | done 37 | 38 | cat head-nfs-install.sh | sed "s/ [\\]//g" | ssh -i $LAMBDA_CLOUD_KEY ${HEAD_IP} 39 | 40 | for IP in ${WORKER_IP[*]}; do 41 | ssh -i $LAMBDA_CLOUD_KEY ubuntu@$HEAD_IP "echo '/home/ubuntu/shared ${IP}(rw,sync,no_subtree_check)' | sudo tee -a /etc/exports" 42 | done 43 | ssh -i $LAMBDA_CLOUD_KEY ubuntu@$HEAD_IP "sudo systemctl restart nfs-kernel-server" 44 | echo "NFS set up on the head node" 45 | 46 | for IP in ${WORKER_IP[*]}; do 47 | ssh -i $LAMBDA_CLOUD_KEY ubuntu@$IP "sudo mount ${HEAD_IP}:/home/ubuntu/shared /home/ubuntu/shared" 48 | done 49 | echo "NFS set up on the worker nodes" 50 | 51 | echo "Clone repos into NFS ------------------------------" 52 | ssh -i $LAMBDA_CLOUD_KEY ubuntu@$HEAD_IP "if [ ! -d /home/ubuntu/shared/llama ]; then git clone https://github.com/LambdaLabsML/llama.git /home/ubuntu/shared/llama; fi" 53 | ssh -i $LAMBDA_CLOUD_KEY ubuntu@$HEAD_IP "if [ ! -d /home/ubuntu/shared/llama-dl ]; then git clone https://github.com/chuanli11/llama-dl.git /home/ubuntu/shared/llama-dl; fi" 54 | 55 | echo "Install LLAMA dependencies (asynchronously) ------------------------------" 56 | for IP in ${ALL_IP[*]}; do cat dependencies-install.sh | ssh -i $LAMBDA_CLOUD_KEY ${IP} & done 57 | 58 | wait 59 | 60 | echo "All instances are successfully set up 🥳🥳🥳🥳🥳🥳🥳🥳" -------------------------------------------------------------------------------- /download.sh: -------------------------------------------------------------------------------- 1 | # Copyright (c) Meta Platforms, Inc. and affiliates. 2 | # This software may be used and distributed according to the terms of the GNU General Public License version 3. 3 | 4 | PRESIGNED_URL="" # replace with presigned url from email 5 | MODEL_SIZE="7B,13B,30B,65B" # edit this list with the model sizes you wish to download 6 | TARGET_FOLDER="" # where all files should end up 7 | 8 | declare -A N_SHARD_DICT 9 | 10 | N_SHARD_DICT["7B"]="0" 11 | N_SHARD_DICT["13B"]="1" 12 | N_SHARD_DICT["30B"]="3" 13 | N_SHARD_DICT["65B"]="7" 14 | 15 | echo "Downloading tokenizer" 16 | wget ${PRESIGNED_URL/'*'/"tokenizer.model"} -O ${TARGET_FOLDER}"/tokenizer.model" 17 | wget ${PRESIGNED_URL/'*'/"tokenizer_checklist.chk"} -O ${TARGET_FOLDER}"/tokenizer_checklist.chk" 18 | 19 | (cd ${TARGET_FOLDER} && md5sum -c tokenizer_checklist.chk) 20 | 21 | for i in ${MODEL_SIZE//,/ } 22 | do 23 | echo "Downloading ${i}" 24 | mkdir -p ${TARGET_FOLDER}"/${i}" 25 | for s in $(seq -f "0%g" 0 ${N_SHARD_DICT[$i]}) 26 | do 27 | wget ${PRESIGNED_URL/'*'/"${i}/consolidated.${s}.pth"} -O ${TARGET_FOLDER}"/${i}/consolidated.${s}.pth" 28 | done 29 | wget ${PRESIGNED_URL/'*'/"${i}/params.json"} -O ${TARGET_FOLDER}"/${i}/params.json" 30 | wget ${PRESIGNED_URL/'*'/"${i}/checklist.chk"} -O ${TARGET_FOLDER}"/${i}/checklist.chk" 31 | echo "Checking checksums" 32 | (cd ${TARGET_FOLDER}"/${i}" && md5sum -c checklist.chk) 33 | done -------------------------------------------------------------------------------- /example.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Meta Platforms, Inc. and affiliates. 2 | # This software may be used and distributed according to the terms of the GNU General Public License version 3. 3 | 4 | from typing import Tuple 5 | import os 6 | import sys 7 | import torch 8 | import fire 9 | import time 10 | import json 11 | 12 | from pathlib import Path 13 | 14 | from fairscale.nn.model_parallel.initialize import initialize_model_parallel 15 | 16 | from llama import ModelArgs, Transformer, Tokenizer, LLaMA 17 | 18 | 19 | def setup_model_parallel(seed: int) -> Tuple[int, int]: 20 | if 'LOCAL_RANK' in os.environ: 21 | # Environment variables set by torch.distributed.launch or torchrun 22 | local_rank = int(os.environ['LOCAL_RANK']) 23 | world_size = int(os.environ['WORLD_SIZE']) 24 | world_rank = int(os.environ['RANK']) 25 | elif 'OMPI_COMM_WORLD_LOCAL_RANK' in os.environ: 26 | # Environment variables set by mpirun 27 | local_rank = int(os.environ['OMPI_COMM_WORLD_LOCAL_RANK']) 28 | world_size = int(os.environ['OMPI_COMM_WORLD_SIZE']) 29 | world_rank = int(os.environ['OMPI_COMM_WORLD_RANK']) 30 | else: 31 | import sys 32 | sys.exit("Can't find the evironment variables for local rank") 33 | 34 | torch.distributed.init_process_group(backend="nccl", rank=world_rank, world_size=world_size) 35 | initialize_model_parallel(world_size) 36 | torch.cuda.set_device(local_rank) 37 | 38 | # seed must be the same in all processes 39 | torch.manual_seed(seed) 40 | return local_rank, world_rank, world_size 41 | 42 | 43 | def load( 44 | ckpt_dir: str, 45 | tokenizer_path: str, 46 | local_rank: int, 47 | world_rank: int, 48 | world_size: int, 49 | max_seq_len: int, 50 | max_batch_size: int, 51 | ) -> LLaMA: 52 | start_time = time.time() 53 | checkpoints = sorted(Path(ckpt_dir).glob("*.pth")) 54 | assert world_size == len( 55 | checkpoints 56 | ), f"Loading a checkpoint for MP={len(checkpoints)} but world size is {world_size}" 57 | 58 | ckpt_path = checkpoints[world_rank] 59 | print("Loading") 60 | checkpoint = torch.load(ckpt_path, map_location="cpu") 61 | with open(Path(ckpt_dir) / "params.json", "r") as f: 62 | params = json.loads(f.read()) 63 | 64 | model_args: ModelArgs = ModelArgs( 65 | max_seq_len=max_seq_len, max_batch_size=max_batch_size, **params 66 | ) 67 | tokenizer = Tokenizer(model_path=tokenizer_path) 68 | model_args.vocab_size = tokenizer.n_words 69 | torch.set_default_tensor_type(torch.cuda.HalfTensor) 70 | model = Transformer(model_args) 71 | torch.set_default_tensor_type(torch.FloatTensor) 72 | model.load_state_dict(checkpoint, strict=False) 73 | 74 | generator = LLaMA(model, tokenizer) 75 | print(f"Loaded in {time.time() - start_time:.2f} seconds") 76 | return generator 77 | 78 | 79 | def main( 80 | ckpt_dir: str, 81 | tokenizer_path: str, 82 | temperature: float = 0.7, 83 | # top_p: float = 0.95, 84 | top_p: float = 0.0, 85 | top_k: int = 10, 86 | repetition_penalty: float = (1 / 0.85), 87 | max_seq_len: int = 2048, 88 | max_gen_len: int = 2000, 89 | max_batch_size: int = 1, 90 | seed: int = 1, 91 | count: int = 1, 92 | eos_w: float = 1.0, 93 | ): 94 | local_rank, world_rank, world_size = setup_model_parallel(seed) 95 | if world_rank > 0: 96 | sys.stdout = open(os.devnull, "w") 97 | 98 | print("\n") 99 | print("~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~") 100 | print(json.dumps(dict( 101 | seed=seed, 102 | temp=temperature, 103 | top_p=top_p, 104 | top_k=top_k, 105 | repetition_penalty=repetition_penalty, 106 | max_seq_len=max_seq_len, 107 | max_gen_len=max_gen_len, 108 | eos_w=eos_w, 109 | ))) 110 | print("~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~") 111 | 112 | 113 | generator = load( 114 | ckpt_dir, tokenizer_path, local_rank, world_rank, world_size, max_seq_len, max_batch_size 115 | ) 116 | 117 | prompts = [ 118 | "[Scene: Central Perk, Chandler, Joey, Phoebe, and Monica are there.]", 119 | ] 120 | i = 0 121 | while i < count or count <= 0: 122 | i += 1 123 | for prompt in prompts: 124 | print(f"\n============== sample {i} =================\n") 125 | width = 0 126 | def callback(text): 127 | nonlocal width 128 | text = text.replace('\n', '\n\n') 129 | chars = [] 130 | for i, c in enumerate(text): 131 | if c == ' ' and width >= 60: 132 | chars.append('\n') 133 | width = 0 134 | else: 135 | width += 1 136 | chars.append(c) 137 | if c == '\n': 138 | width = 0 139 | text = ''.join(chars) 140 | print(text, end='', flush=True) 141 | text, num_gen_tokens = generator.generate( 142 | [prompt], max_gen_len=max_gen_len, temperature=temperature, top_p=top_p, top_k=top_k, repetition_penalty=repetition_penalty, token_callback=callback, eos_w=eos_w 143 | ) 144 | 145 | 146 | if __name__ == "__main__": 147 | fire.Fire(main) 148 | -------------------------------------------------------------------------------- /interactive.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Meta Platforms, Inc. and affiliates. 2 | # This software may be used and distributed according to the terms of the GNU General Public License version 3. 3 | 4 | from typing import Tuple 5 | import os 6 | import sys 7 | import torch 8 | import fire 9 | import time 10 | import json 11 | import torch.distributed as dist 12 | 13 | from pathlib import Path 14 | 15 | from fairscale.nn.model_parallel.initialize import initialize_model_parallel 16 | 17 | from llama import ModelArgs, Transformer, Tokenizer, LLaMA 18 | 19 | 20 | def setup_model_parallel(seed: int) -> Tuple[int, int]: 21 | if 'LOCAL_RANK' in os.environ: 22 | # Environment variables set by torch.distributed.launch or torchrun 23 | local_rank = int(os.environ['LOCAL_RANK']) 24 | world_size = int(os.environ['WORLD_SIZE']) 25 | world_rank = int(os.environ['RANK']) 26 | elif 'OMPI_COMM_WORLD_LOCAL_RANK' in os.environ: 27 | # Environment variables set by mpirun 28 | local_rank = int(os.environ['OMPI_COMM_WORLD_LOCAL_RANK']) 29 | world_size = int(os.environ['OMPI_COMM_WORLD_SIZE']) 30 | world_rank = int(os.environ['OMPI_COMM_WORLD_RANK']) 31 | else: 32 | import sys 33 | sys.exit("Can't find the evironment variables for local rank") 34 | 35 | torch.distributed.init_process_group(backend="nccl", rank=world_rank, world_size=world_size) 36 | initialize_model_parallel(world_size) 37 | torch.cuda.set_device(local_rank) 38 | 39 | # seed must be the same in all processes 40 | torch.manual_seed(seed) 41 | return local_rank, world_rank, world_size 42 | 43 | 44 | def load( 45 | ckpt_dir: str, 46 | tokenizer_path: str, 47 | local_rank: int, 48 | world_rank: int, 49 | world_size: int, 50 | max_seq_len: int, 51 | max_batch_size: int, 52 | ) -> LLaMA: 53 | start_time = time.time() 54 | checkpoints = sorted(Path(ckpt_dir).glob("*.pth")) 55 | assert world_size == len( 56 | checkpoints 57 | ), f"Loading a checkpoint for MP={len(checkpoints)} but world size is {world_size}" 58 | 59 | ckpt_path = checkpoints[world_rank] 60 | print("Loading") 61 | checkpoint = torch.load(ckpt_path, map_location="cpu") 62 | with open(Path(ckpt_dir) / "params.json", "r") as f: 63 | params = json.loads(f.read()) 64 | 65 | model_args: ModelArgs = ModelArgs( 66 | max_seq_len=max_seq_len, max_batch_size=max_batch_size, **params 67 | ) 68 | tokenizer = Tokenizer(model_path=tokenizer_path) 69 | model_args.vocab_size = tokenizer.n_words 70 | torch.set_default_tensor_type(torch.cuda.HalfTensor) 71 | model = Transformer(model_args) 72 | torch.set_default_tensor_type(torch.FloatTensor) 73 | model.load_state_dict(checkpoint, strict=False) 74 | 75 | generator = LLaMA(model, tokenizer) 76 | print(f"Loaded in {time.time() - start_time:.2f} seconds") 77 | return generator 78 | 79 | 80 | def main( 81 | ckpt_dir: str, 82 | tokenizer_path: str, 83 | temperature: float = 0.7, 84 | # top_p: float = 0.95, 85 | top_p: float = 0.0, 86 | top_k: int = 10, 87 | repetition_penalty: float = (1 / 0.85), 88 | max_seq_len: int = 2048, 89 | max_gen_len: int = 2000, 90 | max_batch_size: int = 1, 91 | seed: int = 1, 92 | count: int = 1, 93 | eos_w: float = 1.0, 94 | ): 95 | local_rank, world_rank, world_size = setup_model_parallel(seed) 96 | device = torch.device("cuda:{}".format(local_rank)) 97 | if world_rank > 0: 98 | sys.stdout = open(os.devnull, "w") 99 | 100 | print("\n") 101 | print("~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~") 102 | print(json.dumps(dict( 103 | seed=seed, 104 | temp=temperature, 105 | top_p=top_p, 106 | top_k=top_k, 107 | repetition_penalty=repetition_penalty, 108 | max_seq_len=max_seq_len, 109 | max_gen_len=max_gen_len, 110 | eos_w=eos_w, 111 | ))) 112 | print("~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~") 113 | 114 | 115 | generator = load( 116 | ckpt_dir, tokenizer_path, local_rank, world_rank, world_size, max_seq_len, max_batch_size 117 | ) 118 | 119 | while True: 120 | tensor = torch.ones(max_seq_len) * -1.0 121 | tensor = tensor.to(device) 122 | if world_rank == 0: 123 | prompt = input("Prompt >>> ") 124 | while not prompt: 125 | print('Prompt should not be empty!') 126 | prompt = input("Prompt >>> ") 127 | prompt = prompt[:max_seq_len] 128 | for i, c in enumerate(prompt): 129 | tensor[i] = ord(c) 130 | for rank_recv in range(1, world_size): 131 | dist.send(tensor=tensor, dst=rank_recv) 132 | print('Sending prompt to Rank {}\n'.format(rank_recv)) 133 | else: 134 | dist.recv(tensor=tensor, src=0) 135 | mask = tensor >= 0 136 | tensor = tensor[mask] 137 | prompt = ''.join([chr(int(x)) for x in tensor]) 138 | 139 | i = 0 140 | while i < count or count <= 0: 141 | i += 1 142 | print(f"\n============== sample {i} =================\n") 143 | width = 0 144 | def callback(text): 145 | nonlocal width 146 | text = text.replace('\n', '\n\n') 147 | chars = [] 148 | for i, c in enumerate(text): 149 | if c == ' ' and width >= 60: 150 | chars.append('\n') 151 | width = 0 152 | else: 153 | width += 1 154 | chars.append(c) 155 | if c == '\n': 156 | width = 0 157 | text = ''.join(chars) 158 | print(text, end='', flush=True) 159 | start_time = time.time() 160 | text, num_gen_tokens = generator.generate( 161 | [prompt], max_gen_len=max_gen_len, temperature=temperature, top_p=top_p, top_k=top_k, repetition_penalty=repetition_penalty, token_callback=callback, eos_w=eos_w 162 | ) 163 | print(f"\n============== end sample {i} =================\n") 164 | sec = time.time() - start_time 165 | print(f"{num_gen_tokens} tokens generated in {sec:.2f} seconds, average tokens/sec: {num_gen_tokens / sec:.2f}") 166 | 167 | if __name__ == "__main__": 168 | fire.Fire(main) 169 | -------------------------------------------------------------------------------- /llama/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Meta Platforms, Inc. and affiliates. 2 | # This software may be used and distributed according to the terms of the GNU General Public License version 3. 3 | 4 | from .generation import LLaMA 5 | from .model import ModelArgs, Transformer 6 | from .tokenizer import Tokenizer 7 | -------------------------------------------------------------------------------- /llama/generation.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Meta Platforms, Inc. and affiliates. 2 | # This software may be used and distributed according to the terms of the GNU General Public License version 3. 3 | 4 | from typing import List 5 | 6 | import torch 7 | 8 | from llama.tokenizer import Tokenizer 9 | from llama.model import Transformer 10 | 11 | 12 | class LLaMA: 13 | def __init__(self, model: Transformer, tokenizer: Tokenizer): 14 | self.model = model 15 | self.tokenizer = tokenizer 16 | 17 | def generate( 18 | self, 19 | prompts: List[str], 20 | max_gen_len: int, 21 | temperature: float = 0.7, 22 | top_k: int = 40, 23 | top_p: float = 0.0, #0.95, 24 | repetition_penalty: float = (1.0 / 0.85), 25 | token_callback=None, 26 | eos_w=1.0 27 | ) -> List[str]: 28 | bsz = len(prompts) 29 | params = self.model.params 30 | assert bsz <= params.max_batch_size, (bsz, params.max_batch_size) 31 | 32 | prompt_tokens = [self.tokenizer.encode(x, bos=True, eos=False) for x in prompts] 33 | 34 | min_prompt_size = min([len(t) for t in prompt_tokens]) 35 | max_prompt_size = max([len(t) for t in prompt_tokens]) 36 | 37 | total_len = min(params.max_seq_len, max_gen_len + max_prompt_size) 38 | 39 | tokens = torch.full((bsz, total_len), self.tokenizer.pad_id).cuda().long() 40 | for k, t in enumerate(prompt_tokens): 41 | tokens[k, : len(t)] = torch.tensor(t).long() 42 | input_text_mask = tokens != self.tokenizer.pad_id 43 | start_pos = min_prompt_size 44 | prev_pos = 0 45 | prev_text = '' 46 | for cur_pos in range(start_pos, total_len): 47 | logits = self.model.forward(tokens[:, prev_pos:cur_pos], prev_pos) 48 | 49 | # repetition penalty from CTRL paper (https://arxiv.org/abs/1909.05858) 50 | if repetition_penalty != 1.0: 51 | logits_new = logits.clone() 52 | batch_size = len(tokens) 53 | for i in range(batch_size): 54 | for token in set(tokens[i].tolist()): 55 | # if score < 0 then repetition penalty has to multiplied to reduce the previous token probability 56 | if logits[i, token] < 0: 57 | logits_new[i, token] = logits[i, token] * repetition_penalty 58 | else: 59 | logits_new[i, token] = logits[i, token] / repetition_penalty 60 | logits = logits_new 61 | 62 | if temperature > 0: 63 | probs = torch.softmax(logits / temperature, dim=-1) 64 | next_token = sample(probs, top_p=top_p, top_k=top_k, eos_id=self.tokenizer.eos_id, eos_w=eos_w) 65 | else: 66 | next_token = torch.argmax(logits, dim=-1) 67 | next_token = next_token.reshape(-1) 68 | # only replace token if prompt has already been generated 69 | next_token = torch.where( 70 | input_text_mask[:, cur_pos], tokens[:, cur_pos], next_token 71 | ) 72 | if next_token == self.tokenizer.eos_id: 73 | break 74 | tokens[:, cur_pos] = next_token 75 | if token_callback is not None: 76 | assert len(prompts) == 1 77 | text, = self.decode(tokens) 78 | #assert text.startswith(prev_text) 79 | if not text.startswith(prev_text): 80 | # Some kind of bogus token generation; abort early. 81 | break 82 | next_word = text[len(prev_text):] 83 | prev_text = text 84 | token_callback(next_word) 85 | prev_pos = cur_pos 86 | 87 | return self.decode(tokens), cur_pos - start_pos 88 | 89 | def decode(self, tokens): 90 | decoded = [] 91 | for i, t in enumerate(tokens.tolist()): 92 | t = [token for token in t if token != -1] 93 | # # cut to max gen len 94 | # t = t[: len(prompt_tokens[i]) + max_gen_len] 95 | while self.tokenizer.eos_id in t: 96 | pos = t.index(self.tokenizer.eos_id) 97 | t[pos:pos+1] = self.tokenizer.encode('\n<|endoftext|>\n', bos=False, eos=False) 98 | decoded.append(self.tokenizer.decode(t)) 99 | return decoded 100 | 101 | def sample(probs, top_p=0.0, top_k=40, eos_id=2, eos_w=1.0): 102 | probs[0, eos_id] = probs[0, eos_id] * eos_w 103 | if top_k > 0: 104 | probs_sort, probs_idx = torch.topk(probs, top_k) 105 | else: 106 | probs_sort, probs_idx = torch.sort(probs, dim=-1, descending=True) 107 | if top_p > 0.0: 108 | probs_sum = torch.cumsum(probs_sort, dim=-1) 109 | mask = probs_sum - probs_sort > top_p 110 | probs_sort[mask] = 0.0 111 | probs_sort.div_(probs_sort.sum(dim=-1, keepdim=True)) 112 | next_token = torch.multinomial(probs_sort, num_samples=1) 113 | next_token = torch.gather(probs_idx, -1, next_token) 114 | return next_token 115 | -------------------------------------------------------------------------------- /llama/model.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Meta Platforms, Inc. and affiliates. 2 | # This software may be used and distributed according to the terms of the GNU General Public License version 3. 3 | 4 | from typing import Optional, Tuple 5 | from dataclasses import dataclass 6 | import math 7 | 8 | import torch 9 | from torch import nn 10 | import torch.nn.functional as F 11 | 12 | import fairscale.nn.model_parallel.initialize as fs_init 13 | from fairscale.nn.model_parallel.layers import ( 14 | ParallelEmbedding, 15 | RowParallelLinear, 16 | ColumnParallelLinear, 17 | ) 18 | 19 | 20 | @dataclass 21 | class ModelArgs: 22 | dim: int = 512 23 | n_layers: int = 8 24 | n_heads: int = 8 25 | vocab_size: int = -1 # defined later by tokenizer 26 | multiple_of: int = 256 # make SwiGLU hidden layer size multiple of large power of 2 27 | norm_eps: float = 1e-5 28 | 29 | max_batch_size: int = 32 30 | max_seq_len: int = 2048 31 | 32 | 33 | class RMSNorm(torch.nn.Module): 34 | def __init__(self, dim: int, eps: float = 1e-6): 35 | super().__init__() 36 | self.eps = eps 37 | self.weight = nn.Parameter(torch.ones(dim)) 38 | 39 | def _norm(self, x): 40 | return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps) 41 | 42 | def forward(self, x): 43 | output = self._norm(x.float()).type_as(x) 44 | return output * self.weight 45 | 46 | 47 | def precompute_freqs_cis(dim: int, end: int, theta: float = 10000.0): 48 | freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim)) 49 | t = torch.arange(end, device=freqs.device) # type: ignore 50 | freqs = torch.outer(t, freqs).float() # type: ignore 51 | freqs_cis = torch.polar(torch.ones_like(freqs), freqs) # complex64 52 | return freqs_cis 53 | 54 | 55 | def reshape_for_broadcast(freqs_cis: torch.Tensor, x: torch.Tensor): 56 | ndim = x.ndim 57 | assert 0 <= 1 < ndim 58 | assert freqs_cis.shape == (x.shape[1], x.shape[-1]) 59 | shape = [d if i == 1 or i == ndim - 1 else 1 for i, d in enumerate(x.shape)] 60 | return freqs_cis.view(*shape) 61 | 62 | 63 | def apply_rotary_emb( 64 | xq: torch.Tensor, 65 | xk: torch.Tensor, 66 | freqs_cis: torch.Tensor, 67 | ) -> Tuple[torch.Tensor, torch.Tensor]: 68 | xq_ = torch.view_as_complex(xq.float().reshape(*xq.shape[:-1], -1, 2)) 69 | xk_ = torch.view_as_complex(xk.float().reshape(*xk.shape[:-1], -1, 2)) 70 | freqs_cis = reshape_for_broadcast(freqs_cis, xq_) 71 | xq_out = torch.view_as_real(xq_ * freqs_cis).flatten(3) 72 | xk_out = torch.view_as_real(xk_ * freqs_cis).flatten(3) 73 | return xq_out.type_as(xq), xk_out.type_as(xk) 74 | 75 | 76 | class Attention(nn.Module): 77 | def __init__(self, args: ModelArgs): 78 | super().__init__() 79 | 80 | self.n_local_heads = args.n_heads // fs_init.get_model_parallel_world_size() 81 | self.head_dim = args.dim // args.n_heads 82 | 83 | self.wq = ColumnParallelLinear( 84 | args.dim, 85 | args.n_heads * self.head_dim, 86 | bias=False, 87 | gather_output=False, 88 | init_method=lambda x: x, 89 | ) 90 | self.wk = ColumnParallelLinear( 91 | args.dim, 92 | args.n_heads * self.head_dim, 93 | bias=False, 94 | gather_output=False, 95 | init_method=lambda x: x, 96 | ) 97 | self.wv = ColumnParallelLinear( 98 | args.dim, 99 | args.n_heads * self.head_dim, 100 | bias=False, 101 | gather_output=False, 102 | init_method=lambda x: x, 103 | ) 104 | self.wo = RowParallelLinear( 105 | args.n_heads * self.head_dim, 106 | args.dim, 107 | bias=False, 108 | input_is_parallel=True, 109 | init_method=lambda x: x, 110 | ) 111 | 112 | self.cache_k = torch.zeros( 113 | (args.max_batch_size, args.max_seq_len, self.n_local_heads, self.head_dim) 114 | ).cuda() 115 | self.cache_v = torch.zeros( 116 | (args.max_batch_size, args.max_seq_len, self.n_local_heads, self.head_dim) 117 | ).cuda() 118 | 119 | def forward(self, x: torch.Tensor, start_pos: int, freqs_cis: torch.Tensor, mask: Optional[torch.Tensor]): 120 | bsz, seqlen, _ = x.shape 121 | xq, xk, xv = self.wq(x), self.wk(x), self.wv(x) 122 | 123 | xq = xq.view(bsz, seqlen, self.n_local_heads, self.head_dim) 124 | xk = xk.view(bsz, seqlen, self.n_local_heads, self.head_dim) 125 | xv = xv.view(bsz, seqlen, self.n_local_heads, self.head_dim) 126 | 127 | xq, xk = apply_rotary_emb(xq, xk, freqs_cis=freqs_cis) 128 | 129 | self.cache_k = self.cache_k.to(xq) 130 | self.cache_v = self.cache_v.to(xq) 131 | 132 | self.cache_k[:bsz, start_pos : start_pos + seqlen] = xk 133 | self.cache_v[:bsz, start_pos : start_pos + seqlen] = xv 134 | 135 | keys = self.cache_k[:bsz, : start_pos + seqlen] 136 | values = self.cache_v[:bsz, : start_pos + seqlen] 137 | 138 | xq = xq.transpose(1, 2) 139 | keys = keys.transpose(1, 2) 140 | values = values.transpose(1, 2) 141 | scores = torch.matmul(xq, keys.transpose(2, 3)) / math.sqrt(self.head_dim) 142 | if mask is not None: 143 | scores = scores + mask # (bs, n_local_heads, slen, cache_len + slen) 144 | scores = F.softmax(scores.float(), dim=-1).type_as(xq) 145 | output = torch.matmul(scores, values) # (bs, n_local_heads, slen, head_dim) 146 | output = output.transpose( 147 | 1, 2 148 | ).contiguous().view(bsz, seqlen, -1) 149 | 150 | return self.wo(output) 151 | 152 | 153 | class FeedForward(nn.Module): 154 | def __init__( 155 | self, 156 | dim: int, 157 | hidden_dim: int, 158 | multiple_of: int, 159 | ): 160 | super().__init__() 161 | hidden_dim = int(2 * hidden_dim / 3) 162 | hidden_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of) 163 | 164 | self.w1 = ColumnParallelLinear( 165 | dim, hidden_dim, bias=False, gather_output=False, init_method=lambda x: x 166 | ) 167 | self.w2 = RowParallelLinear( 168 | hidden_dim, dim, bias=False, input_is_parallel=True, init_method=lambda x: x 169 | ) 170 | self.w3 = ColumnParallelLinear( 171 | dim, hidden_dim, bias=False, gather_output=False, init_method=lambda x: x 172 | ) 173 | 174 | def forward(self, x): 175 | return self.w2(F.silu(self.w1(x)) * self.w3(x)) 176 | 177 | 178 | class TransformerBlock(nn.Module): 179 | def __init__(self, layer_id: int, args: ModelArgs): 180 | super().__init__() 181 | self.n_heads = args.n_heads 182 | self.dim = args.dim 183 | self.head_dim = args.dim // args.n_heads 184 | self.attention = Attention(args) 185 | self.feed_forward = FeedForward( 186 | dim=args.dim, hidden_dim=4 * args.dim, multiple_of=args.multiple_of 187 | ) 188 | self.layer_id = layer_id 189 | self.attention_norm = RMSNorm(args.dim, eps=args.norm_eps) 190 | self.ffn_norm = RMSNorm(args.dim, eps=args.norm_eps) 191 | 192 | def forward(self, x: torch.Tensor, start_pos: int, freqs_cis: torch.Tensor, mask: Optional[torch.Tensor]): 193 | h = x + self.attention.forward(self.attention_norm(x), start_pos, freqs_cis, mask) 194 | out = h + self.feed_forward.forward(self.ffn_norm(h)) 195 | return out 196 | 197 | 198 | class Transformer(nn.Module): 199 | def __init__(self, params: ModelArgs): 200 | super().__init__() 201 | self.params = params 202 | self.vocab_size = params.vocab_size 203 | self.n_layers = params.n_layers 204 | 205 | self.tok_embeddings = ParallelEmbedding( 206 | params.vocab_size, params.dim, init_method=lambda x: x 207 | ) 208 | 209 | self.layers = torch.nn.ModuleList() 210 | for layer_id in range(params.n_layers): 211 | self.layers.append(TransformerBlock(layer_id, params)) 212 | 213 | self.norm = RMSNorm(params.dim, eps=params.norm_eps) 214 | self.output = ColumnParallelLinear( 215 | params.dim, params.vocab_size, bias=False, init_method=lambda x: x 216 | ) 217 | 218 | self.freqs_cis = precompute_freqs_cis( 219 | self.params.dim // self.params.n_heads, self.params.max_seq_len * 2 220 | ) 221 | 222 | @torch.inference_mode() 223 | def forward(self, tokens: torch.Tensor, start_pos: int): 224 | _bsz, seqlen = tokens.shape 225 | h = self.tok_embeddings(tokens) 226 | self.freqs_cis = self.freqs_cis.to(h.device) 227 | freqs_cis = self.freqs_cis[start_pos : start_pos + seqlen] 228 | 229 | mask = None 230 | if seqlen > 1: 231 | mask = torch.full((1, 1, seqlen, seqlen), float("-inf"), device=tokens.device) 232 | mask = torch.triu(mask, diagonal=start_pos + 1).type_as(h) 233 | 234 | for layer in self.layers: 235 | h = layer(h, start_pos, freqs_cis, mask) 236 | h = self.norm(h) 237 | output = self.output(h[:, -1, :]) # only compute last logits 238 | return output.float() 239 | -------------------------------------------------------------------------------- /llama/tokenizer.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Meta Platforms, Inc. and affiliates. 2 | # This software may be used and distributed according to the terms of the GNU General Public License version 3. 3 | 4 | from sentencepiece import SentencePieceProcessor 5 | from logging import getLogger 6 | from typing import List 7 | import os 8 | 9 | 10 | logger = getLogger() 11 | 12 | 13 | class Tokenizer: 14 | def __init__(self, model_path: str): 15 | # reload tokenizer 16 | assert os.path.isfile(model_path), model_path 17 | self.sp_model = SentencePieceProcessor(model_file=model_path) 18 | logger.info(f"Reloaded SentencePiece model from {model_path}") 19 | 20 | # BOS / EOS token IDs 21 | self.n_words: int = self.sp_model.vocab_size() 22 | self.bos_id: int = self.sp_model.bos_id() 23 | self.eos_id: int = self.sp_model.eos_id() 24 | self.pad_id: int = self.sp_model.pad_id() 25 | logger.info( 26 | f"#words: {self.n_words} - BOS ID: {self.bos_id} - EOS ID: {self.eos_id}" 27 | ) 28 | assert self.sp_model.vocab_size() == self.sp_model.get_piece_size() 29 | 30 | def encode(self, s: str, bos: bool, eos: bool) -> List[int]: 31 | assert type(s) is str 32 | t = self.sp_model.encode(s) 33 | if bos: 34 | t = [self.bos_id] + t 35 | if eos: 36 | t = t + [self.eos_id] 37 | return t 38 | 39 | def decode(self, t: List[int]) -> str: 40 | return self.sp_model.decode(t) 41 | -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | torch 2 | fairscale 3 | fire 4 | sentencepiece -------------------------------------------------------------------------------- /setup.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Meta Platforms, Inc. and affiliates. 2 | # This software may be used and distributed according to the terms of the GNU General Public License version 3. 3 | 4 | from setuptools import setup, find_packages 5 | 6 | setup(name="llama", version="0.0.0", packages=find_packages()) 7 | --------------------------------------------------------------------------------