diff --git a/README.md b/README.md index 2610d4256..c7b87ff79 100644 --- a/README.md +++ b/README.md @@ -1,29 +1,27 @@ # koboldcpp -KoboldCpp is an easy-to-use AI text-generation software for GGML and GGUF models. It's a single self contained distributable from Concedo, that builds off llama.cpp, and adds a versatile Kobold API endpoint, additional format support, Stable Diffusion image generation, backward compatibility, as well as a fancy UI with persistent stories, editing tools, save formats, memory, world info, author's note, characters, scenarios and everything KoboldAI and KoboldAI Lite have to offer. +KoboldCpp is an easy-to-use AI text-generation software for GGML and GGUF models, inspired by the original KoboldAI. It's a single self-contained distributable from Concedo, that builds off llama.cpp, and adds a versatile **KoboldAI API endpoint**, additional format support, Stable Diffusion image generation, speech-to-text, backward compatibility, as well as a fancy UI with persistent stories, editing tools, save formats, memory, world info, author's note, characters, scenarios and everything KoboldAI and KoboldAI Lite have to offer. ![Preview](media/preview.png) ![Preview](media/preview2.png) ![Preview](media/preview3.png) ![Preview](media/preview4.png) -## Windows Usage -- **[Download the latest .exe release here](https://github.com/LostRuins/koboldcpp/releases/latest)** or clone the git repo. -- Windows binaries are provided in the form of **koboldcpp.exe**, which is a pyinstaller wrapper for a few **.dll** files and **koboldcpp.py**. You can also rebuild it yourself with the provided makefiles and scripts. -- Weights are not included, you can use the official llama.cpp `quantize.exe` to generate them from your official weight files (or download them from other places such as [TheBloke's Huggingface](https://huggingface.co/TheBloke). +## Windows Usage (Precompiled Binary, Recommended) +- Windows binaries are provided in the form of **koboldcpp.exe**, which is a pyinstaller wrapper containing all necessary files. **[Download the latest koboldcpp.exe release here](https://github.com/LostRuins/koboldcpp/releases/latest)** - To run, simply execute **koboldcpp.exe**. - Launching with no command line arguments displays a GUI containing a subset of configurable settings. Generally you dont have to change much besides the `Presets` and `GPU Layers`. Read the `--help` for more info about each settings. - By default, you can connect to http://localhost:5001 - You can also run it using the command line. For info, please check `koboldcpp.exe --help` -### Improving Performance -- **(Nvidia Only) GPU Acceleration**: If you're on Windows with an Nvidia GPU you can get CUDA support out of the box using the `--usecublas` flag, make sure you select the correct .exe with CUDA support. -- **Any GPU Acceleration**: As a slightly slower alternative, try CLBlast with `--useclblast` flags for a slightly slower but more GPU compatible speedup. -- **GPU Layer Offloading**: Want even more speedup? Combine one of the above GPU flags with `--gpulayers` to offload entire layers to the GPU! **Much faster, but uses more VRAM**. Experiment to determine number of layers to offload, and reduce by a few if you run out of memory. -- **Increasing Context Size**: Try `--contextsize 4096` to 2x your context size! without much perplexity gain. Note that you'll have to increase the max context in the KoboldAI Lite UI as well (click and edit the number text field). -- If you are having crashes or issues, you can try turning off BLAS with the `--noblas` flag. You can also try running in a non-avx2 compatibility mode with `--noavx2`. Lastly, you can try turning off mmap with `--nommap`. +## Linux Usage (Precompiled Binary, Recommended) +On modern Linux systems, you should download the `koboldcpp-linux-x64-cuda1150` prebuilt PyInstaller binary on the **[releases page](https://github.com/LostRuins/koboldcpp/releases/latest)**. Simply download and run the binary. -For more information, be sure to run the program with the `--help` flag, or [check the wiki](https://github.com/LostRuins/koboldcpp/wiki). +Alternatively, you can also install koboldcpp to the current directory by running the following terminal command: +``` +curl -fLo koboldcpp https://github.com/LostRuins/koboldcpp/releases/latest/download/koboldcpp-linux-x64-cuda1150 && chmod +x koboldcpp +``` +After running this command you can launch Koboldcpp from the current directory using `./koboldcpp` in the terminal (for CLI usage, run with `--help`). ## Run on Colab - KoboldCpp now has an **official Colab GPU Notebook**! This is an easy way to get started without installing anything in a minute or two. [Try it here!](https://colab.research.google.com/github/LostRuins/koboldcpp/blob/concedo/colab.ipynb). @@ -32,21 +30,32 @@ For more information, be sure to run the program with the `--help` flag, or [che ## Run on RunPod - KoboldCpp can now be used on RunPod cloud GPUs! This is an easy way to get started without installing anything in a minute or two, and is very scalable, capable of running 70B+ models at afforable cost. [Try our RunPod image here!](https://koboldai.org/runpodcpp). -## OSX and Linux +## Docker +- The official docker can be found at https://hub.docker.com/r/koboldai/koboldcpp +- If you're building your own docker, remember to set CUDA_DOCKER_ARCH or enable LLAMA_PORTABLE -### Linux Usage (Precompiled Binary, Recommended) -On Linux, we provide a `koboldcpp-linux-x64-cuda1150` PyInstaller prebuilt binary on the **[releases](https://github.com/LostRuins/koboldcpp/releases/latest)** page for modern systems. Simply download and run the binary. +## MacOS +- You will need to clone the repo and compile from source code, see Compiling for MacOS below. -Alternatively, you can also install koboldcpp to the current directory by running the following terminal command: -``` -curl -fLo koboldcpp https://github.com/LostRuins/koboldcpp/releases/latest/download/koboldcpp-linux-x64-cuda1150 && chmod +x koboldcpp -``` -After running this command you can launch Koboldcpp from the current directory using `./koboldcpp` in the terminal (for CLI usage, run with `--help`). +## Obtaining a GGUF model +- KoboldCpp uses GGUF models. They are not included here, but you can download GGUF files from other places such as [TheBloke's Huggingface](https://huggingface.co/TheBloke). Search for "GGUF" on huggingface.co for plenty of compatible models in the `.gguf` format. +- For beginners, we recommend the models [MistRP Airoboros](https://huggingface.co/TheBloke/MistRP-Airoboros-7B-GGUF/resolve/main/mistrp-airoboros-7b.Q4_K_S.gguf) or [Tiefighter 13B](https://huggingface.co/KoboldAI/LLaMA2-13B-Tiefighter-GGUF/resolve/main/LLaMA2-13B-Tiefighter.Q4_K_S.gguf) (larger model). +- [Alternatively, you can download the tools to convert models to the GGUF format yourself here](https://github.com/LostRuins/koboldcpp/releases/download/v1.69.1/koboldcpp_tools_6jul.zip). Run `convert-hf-to-gguf.py` to convert them, then `quantize_gguf.exe` to quantize the result. +## Improving Performance +- **GPU Acceleration**: If you're on Windows with an Nvidia GPU you can get CUDA support out of the box using the `--usecublas` flag (Nvidia Only), or `--usevulkan` (Any GPU), make sure you select the correct .exe with CUDA support. +- **GPU Layer Offloading**: Add `--gpulayers` to offload model layers to the GPU. The more layers you offload to VRAM, the faster generation speed will become. Experiment to determine number of layers to offload, and reduce by a few if you run out of memory. +- **Increasing Context Size**: Use `--contextsize (number)` to increase context size, allowing the model to read more text. Note that you may also need to increase the max context in the KoboldAI Lite UI as well (click and edit the number text field). +- **Old CPU Compatibility**: If you are having crashes or issues, you can try turning off BLAS with the `--noblas` flag. You can also try running in a non-avx2 compatibility mode with `--noavx2`. Lastly, you can try turning off mmap with `--nommap`. -### Linux Usage (koboldcpp.sh automated compiler script) -when you can't use the precompiled binary directly, we provide an automated build script which uses conda to obtain all dependencies, and generates (from source) a ready-to-use a pyinstaller binary for linux users. Simply execute the build script with `./koboldcpp.sh dist` and run the generated binary. (Not recomended for systems that already have an existing installation of conda. Dependencies: curl, bzip2) +For more information, be sure to run the program with the `--help` flag, or **[check the wiki](https://github.com/LostRuins/koboldcpp/wiki).** +## Compiling KoboldCpp From Source Code + +### Compiling on Linux (Using koboldcpp.sh automated compiler script) +when you can't use the precompiled binary directly, we provide an automated build script which uses conda to obtain all dependencies, and generates (from source) a ready-to-use a pyinstaller binary for linux users. +- Clone the repo with `git clone https://github.com/LostRuins/koboldcpp.git` +- Simply execute the build script with `./koboldcpp.sh dist` and run the generated binary. (Not recomended for systems that already have an existing installation of conda. Dependencies: curl, bzip2) ``` ./koboldcpp.sh # This launches the GUI for easy configuration and launching (X11 required). ./koboldcpp.sh --help # List all available terminal commands for using Koboldcpp, you can use koboldcpp.sh the same way as our python script and binaries. @@ -54,38 +63,41 @@ when you can't use the precompiled binary directly, we provide an automated buil ./koboldcpp.sh dist # Generate your own precompiled binary (Due to the nature of Linux compiling these will only work on distributions equal or newer than your own.) ``` -## OSX and Linux Manual Compiling -- Otherwise, you will have to compile your binaries from source. A makefile is provided, simply run `make`. -- If you want you can also link your own install of OpenBLAS manually with `make LLAMA_OPENBLAS=1` -- If you want you can also link your own install of CLBlast manually with `make LLAMA_CLBLAST=1`, for this you will need to obtain and link OpenCL and CLBlast libraries. +### Compiling on Linux (Manual Method) +- To compile your binaries from source, clone the repo with `git clone https://github.com/LostRuins/koboldcpp.git` +- A makefile is provided, simply run `make`. +- Optional OpenBLAS: Link your own install of OpenBLAS manually with `make LLAMA_OPENBLAS=1` +- Optional CLBlast: Link your own install of CLBlast manually with `make LLAMA_CLBLAST=1` +- Note: for these you will need to obtain and link OpenCL and CLBlast libraries. - For Arch Linux: Install `cblas` `openblas` and `clblast`. - For Debian: Install `libclblast-dev` and `libopenblas-dev`. - You can attempt a CuBLAS build with `LLAMA_CUBLAS=1`. You will need CUDA Toolkit installed. Some have also reported success with the CMake file, though that is more for windows. - For a full featured build (all backends), do `make LLAMA_OPENBLAS=1 LLAMA_CLBLAST=1 LLAMA_CUBLAS=1 LLAMA_VULKAN=1`. (Note that `LLAMA_CUBLAS=1` will not work on windows, you need visual studio) -- After all binaries are built, you can run the python script with the command `koboldcpp.py [ggml_model.bin] [port]` +- After all binaries are built, you can run the python script with the command `koboldcpp.py [ggml_model.gguf] [port]` -- Note: Many OSX users have found that the using Accelerate is actually faster than OpenBLAS. To try, you may wish to run with `--noblas` and compare speeds. - -### Arch Linux Packages -There are some community made AUR packages available: [CUBLAS](https://aur.archlinux.org/packages/koboldcpp-cuda), and [HIPBLAS](https://aur.archlinux.org/packages/koboldcpp-hipblas). They are intended for users with NVIDIA GPUs, and users with a supported AMD GPU. Note that these packages may be outdated, and it's probably better to use official KoboldCpp binaries. - -## Compiling on Windows +### Compiling on Windows - You're encouraged to use the .exe released, but if you want to compile your binaries from source at Windows, the easiest way is: - - Use the latest release of w64devkit (https://github.com/skeeto/w64devkit). Be sure to use the "vanilla one", not i686 or other different stuff. If you try they will conflit with the precompiled libs! - - Make sure you are using the w64devkit integrated terminal, then run 'make' at the KoboldCpp source folder. This will create the .dll files. - - If you want to generate the .exe file, make sure you have the python module PyInstaller installed with pip ('pip install PyInstaller'). - - Run the script make_pyinstaller.bat at a regular terminal (or Windows Explorer). + - Get the latest release of w64devkit (https://github.com/skeeto/w64devkit). Be sure to use the "vanilla one", not i686 or other different stuff. If you try they will conflit with the precompiled libs! + - Clone the repo with `git clone https://github.com/LostRuins/koboldcpp.git` + - Make sure you are using the w64devkit integrated terminal, then run `make` at the KoboldCpp source folder. This will create the .dll files. + - If you want to generate the .exe file, make sure you have the python module PyInstaller installed with pip (`pip install PyInstaller`). Then run the script `make_pyinstaller.bat` - The koboldcpp.exe file will be at your dist folder. -- If you wish to use your own version of the additional Windows libraries (OpenCL, CLBlast and OpenBLAS), you can do it with: +- **Building with CUDA**: Visual Studio, CMake and CUDA Toolkit is required. Clone the repo, then open the CMake file and compile it in Visual Studio. Copy the `koboldcpp_cublas.dll` generated into the same directory as the `koboldcpp.py` file. If you are bundling executables, you may need to include CUDA dynamic libraries (such as `cublasLt64_11.dll` and `cublas64_11.dll`) in order for the executable to work correctly on a different PC. +- **Replacing Libraries (Not Recommended)**: If you wish to use your own version of the additional Windows libraries (OpenCL, CLBlast and OpenBLAS), you can do it with: - OpenCL - tested with https://github.com/KhronosGroup/OpenCL-SDK . If you wish to compile it, follow the repository instructions. You will need vcpkg. - CLBlast - tested with https://github.com/CNugteren/CLBlast . If you wish to compile it you will need to reference the OpenCL files. It will only generate the ".lib" file if you compile using MSVC. - OpenBLAS - tested with https://github.com/xianyi/OpenBLAS . - Move the respectives .lib files to the /lib folder of your project, overwriting the older files. - Also, replace the existing versions of the corresponding .dll files located in the project directory root (e.g. libopenblas.dll). - - You can attempt a CuBLAS build with using the provided CMake file with visual studio. If you use the CMake file to build, copy the `koboldcpp_cublas.dll` generated into the same directory as the `koboldcpp.py` file. If you are bundling executables, you may need to include CUDA dynamic libraries (such as `cublasLt64_11.dll` and `cublas64_11.dll`) in order for the executable to work correctly on a different PC. - - Make the KoboldCPP project using the instructions above. + - Make the KoboldCpp project using the instructions above. -## Compiling on Android (Termux Installation) +### Compiling on MacOS +- You can compile your binaries from source. You can clone the repo with `git clone https://github.com/LostRuins/koboldcpp.git` +- A makefile is provided, simply run `make`. +- If you want Metal GPU support, instead run `make LLAMA_METAL=1`, note that MacOS metal libraries need to be installed. +- After all binaries are built, you can run the python script with the command `koboldcpp.py --model [ggml_model.gguf]` (and add `--gpulayers (number of layer)` if you wish to offload layers to GPU). + +### Compiling on Android (Termux Installation) - [Install and run Termux from F-Droid](https://f-droid.org/en/packages/com.termux/) - Enter the command `termux-change-repo` and choose `Mirror by BFSU` - Install dependencies with `pkg install wget git python` (plus any other missing packages) @@ -99,16 +111,9 @@ There are some community made AUR packages available: [CUBLAS](https://aur.archl - If you encounter any errors, make sure your packages are up-to-date with `pkg up` - GPU acceleration for Termux may be possible but I have not explored it. If you find a good cross-device solution, do share or PR it. -## AMD +## AMD Users - Please check out https://github.com/YellowRoseCx/koboldcpp-rocm -## Docker -- The official docker can be found at https://hub.docker.com/r/koboldai/koboldcpp -- KoboldCpp also has a few unofficial third-party community created docker images. Feel free to try them out, but do not expect up-to-date support: - - https://github.com/korewaChino/koboldCppDocker - - https://github.com/noneabove1182/koboldcpp-docker -- If you're building your own docker, remember to set CUDA_DOCKER_ARCH or enable LLAMA_PORTABLE - ## Questions and Help - **First, please check out [The KoboldCpp FAQ and Knowledgebase](https://github.com/LostRuins/koboldcpp/wiki) which may already have answers to your questions! Also please search through past issues and discussions.** - If you cannot find an answer, open an issue on this github, or find us on the [KoboldAI Discord](https://koboldai.org/discord). @@ -117,11 +122,11 @@ There are some community made AUR packages available: [CUBLAS](https://aur.archl - For Windows: No installation, single file executable, (It Just Works) - Since v1.0.6, requires libopenblas, the prebuilt windows binaries are included in this repo. If not found, it will fall back to a mode without BLAS. - Since v1.15, requires CLBlast if enabled, the prebuilt windows binaries are included in this repo. If not found, it will fall back to a mode without CLBlast. -- Since v1.33, you can set the context size to be above what the model supports officially. It does increases perplexity but should still work well below 4096 even on untuned models. (For GPT-NeoX, GPT-J, and LLAMA models) Customize this with `--ropeconfig`. +- Since v1.33, you can set the context size to be above what the model supports officially. It does increases perplexity but should still work well below 4096 even on untuned models. (For GPT-NeoX, GPT-J, and Llama models) Customize this with `--ropeconfig`. - Since v1.42, supports GGUF models for LLAMA and Falcon - Since v1.55, lcuda paths on Linux are hardcoded and may require manual changes to the makefile if you do not use koboldcpp.sh for the compilation. - Since v1.60, provides native image generation with StableDiffusion.cpp, you can load any SD1.5 or SDXL .safetensors model and it will provide an A1111 compatible API to use. -- **I plan to keep backwards compatibility with ALL past llama.cpp AND alpaca.cpp models**. But you are also encouraged to reconvert/update your models if possible for best results. +- **I try to keep backwards compatibility with ALL past llama.cpp models**. But you are also encouraged to reconvert/update your models if possible for best results. ## License - The original GGML library and llama.cpp by ggerganov are licensed under the MIT License @@ -129,17 +134,18 @@ There are some community made AUR packages available: [CUBLAS](https://aur.archl - The other files are also under the AGPL v3.0 License unless otherwise stated ## Notes -- Generation delay scales linearly with original prompt length. If OpenBLAS is enabled then prompt ingestion becomes about 2-3x faster. This is automatic on windows, but will require linking on OSX and Linux. CLBlast speeds this up even further, and `--gpulayers` + `--useclblast` or `--usecublas` more so. -- I have heard of someone claiming a false AV positive report. The exe is a simple pyinstaller bundle that includes the necessary python scripts and dlls to run. If this still concerns you, you might wish to rebuild everything from source code using the makefile, and you can rebuild the exe yourself with pyinstaller by using `make_pyinstaller.bat` -- API documentation available at `/api` and https://lite.koboldai.net/koboldcpp_api -- Supported GGML models (Includes backward compatibility for older versions/legacy GGML models, though some newer features might be unavailable): - - All up-to-date GGUF models are supported (Mistral/Mixtral/QWEN/Gemma and more) - - LLAMA and LLAMA2 (LLaMA / Alpaca / GPT4All / Vicuna / Koala / Pygmalion 7B / Metharme 7B / WizardLM and many more) +- If you wish, after building the koboldcpp libraries with `make`, you can rebuild the exe yourself with pyinstaller by using `make_pyinstaller.bat` +- API documentation available at `/api` (e.g. `http://localhost:5001/api`) and https://lite.koboldai.net/koboldcpp_api. An OpenAI compatible API is also provided at `/v1` route (e.g. `http://localhost:5001/v1`). +- **All up-to-date GGUF models are supported**, and KoboldCpp also includes backward compatibility for older versions/legacy GGML `.bin` models, though some newer features might be unavailable. +- An incomplete list of models and architectures is listed, but there are *many hundreds of other GGUF models*. In general, if it's GGUF, it should work. + - Llama / Llama2 / Llama3 / Alpaca / GPT4All / Vicuna / Koala / Pygmalion / Metharme / WizardLM + - Mistral / Mixtral / Miqu + - Qwen / Qwen2 / Yi + - Gemma / Gemma2 - GPT-2 / Cerebras - - GPT-J - - RWKV + - Phi-2 / Phi-3 - GPT-NeoX / Pythia / StableLM / Dolly / RedPajama - - MPT models - - Falcon (GGUF only) - - [Stable Diffusion and SDXL models](https://github.com/LostRuins/koboldcpp/wiki#can-i-generate-images-with-koboldcpp) - + - GPT-J / RWKV4 / MPT / Falcon / Starcoder / Deepseek and many more + - [Stable Diffusion 1.5 and SDXL safetensor models](https://github.com/LostRuins/koboldcpp/wiki#can-i-generate-images-with-koboldcpp) + - [LLaVA based Vision models and multimodal projectors (mmproj)](https://github.com/LostRuins/koboldcpp/wiki#what-is-llava-and-mmproj) + - [Whisper models for Speech-To-Text](https://huggingface.co/koboldcpp/whisper/tree/main) diff --git a/colab.ipynb b/colab.ipynb index b87209aa2..a806fa284 100644 --- a/colab.ipynb +++ b/colab.ipynb @@ -48,7 +48,7 @@ "source": [ "#@title v-- Enter your model below and then click this to start Koboldcpp\r\n", "\r\n", - "Model = \"https://huggingface.co/KoboldAI/LLaMA2-13B-Tiefighter-GGUF/resolve/main/LLaMA2-13B-Tiefighter.Q4_K_S.gguf\" #@param [\"https://huggingface.co/KoboldAI/LLaMA2-13B-Tiefighter-GGUF/resolve/main/LLaMA2-13B-Tiefighter.Q4_K_S.gguf\",\"https://huggingface.co/KoboldAI/LLaMA2-13B-Estopia-GGUF/resolve/main/LLaMA2-13B-Estopia.Q4_K_S.gguf\",\"https://huggingface.co/Sao10K/Fimbulvetr-11B-v2-GGUF/resolve/main/Fimbulvetr-11B-v2-Test-14.q4_K_M.gguf\",\"https://huggingface.co/TheBloke/MythoMax-L2-13B-GGUF/resolve/main/mythomax-l2-13b.Q4_K_M.gguf\",\"https://huggingface.co/TheBloke/ReMM-SLERP-L2-13B-GGUF/resolve/main/remm-slerp-l2-13b.Q4_K_M.gguf\",\"https://huggingface.co/TheBloke/Xwin-LM-13B-v0.2-GGUF/resolve/main/xwin-lm-13b-v0.2.Q4_K_M.gguf\",\"https://huggingface.co/TheBloke/Stheno-L2-13B-GGUF/resolve/main/stheno-l2-13b.Q4_K_M.gguf\",\"https://huggingface.co/TheBloke/MythoMax-L2-Kimiko-v2-13B-GGUF/resolve/main/mythomax-l2-kimiko-v2-13b.Q4_K_M.gguf\",\"https://huggingface.co/TheBloke/airoboros-mistral2.2-7B-GGUF/resolve/main/airoboros-mistral2.2-7b.Q4_K_S.gguf\",\"https://huggingface.co/concedo/KobbleTinyV2-1.1B-GGUF/resolve/main/KobbleTiny-Q4_K.gguf\",\"https://huggingface.co/grimjim/kukulemon-7B-GGUF/resolve/main/kukulemon-7B.Q8_0.gguf\",\"https://huggingface.co/mradermacher/LemonKunoichiWizardV3-GGUF/resolve/main/LemonKunoichiWizardV3.Q4_K_M.gguf\",\"https://huggingface.co/Lewdiculous/Kunoichi-DPO-v2-7B-GGUF-Imatrix/resolve/main/Kunoichi-DPO-v2-7B-Q4_K_M-imatrix.gguf\",\"https://huggingface.co/mradermacher/L3-8B-Stheno-v3.2-i1-GGUF/resolve/main/L3-8B-Stheno-v3.2.i1-Q4_K_M.gguf\",\"https://huggingface.co/Lewdiculous/Llama-3-Lumimaid-8B-v0.1-OAS-GGUF-IQ-Imatrix/resolve/main/v2-Llama-3-Lumimaid-8B-v0.1-OAS-Q4_K_M-imat.gguf\",\"https://huggingface.co/bartowski/NeuralDaredevil-8B-abliterated-GGUF/resolve/main/NeuralDaredevil-8B-abliterated-Q4_K_M.gguf\",\"https://huggingface.co/bartowski/L3-8B-Lunaris-v1-GGUF/resolve/main/L3-8B-Lunaris-v1-Q4_K_M.gguf\",\"https://huggingface.co/mradermacher/L3-Umbral-Mind-RP-v2.0-8B-GGUF/resolve/main/L3-Umbral-Mind-RP-v2.0-8B.Q4_K_M.gguf\"]{allow-input: true}\r\n", + "Model = \"https://huggingface.co/KoboldAI/LLaMA2-13B-Tiefighter-GGUF/resolve/main/LLaMA2-13B-Tiefighter.Q4_K_S.gguf\" #@param [\"https://huggingface.co/KoboldAI/LLaMA2-13B-Tiefighter-GGUF/resolve/main/LLaMA2-13B-Tiefighter.Q4_K_S.gguf\",\"https://huggingface.co/KoboldAI/LLaMA2-13B-Estopia-GGUF/resolve/main/LLaMA2-13B-Estopia.Q4_K_S.gguf\",\"https://huggingface.co/Sao10K/Fimbulvetr-11B-v2-GGUF/resolve/main/Fimbulvetr-11B-v2-Test-14.q4_K_M.gguf\",\"https://huggingface.co/TheBloke/MythoMax-L2-13B-GGUF/resolve/main/mythomax-l2-13b.Q4_K_M.gguf\",\"https://huggingface.co/TheBloke/ReMM-SLERP-L2-13B-GGUF/resolve/main/remm-slerp-l2-13b.Q4_K_M.gguf\",\"https://huggingface.co/TheBloke/Xwin-LM-13B-v0.2-GGUF/resolve/main/xwin-lm-13b-v0.2.Q4_K_M.gguf\",\"https://huggingface.co/TheBloke/Stheno-L2-13B-GGUF/resolve/main/stheno-l2-13b.Q4_K_M.gguf\",\"https://huggingface.co/TheBloke/MythoMax-L2-Kimiko-v2-13B-GGUF/resolve/main/mythomax-l2-kimiko-v2-13b.Q4_K_M.gguf\",\"https://huggingface.co/TheBloke/MistRP-Airoboros-7B-GGUF/resolve/main/mistrp-airoboros-7b.Q4_K_S.gguf\",\"https://huggingface.co/TheBloke/airoboros-mistral2.2-7B-GGUF/resolve/main/airoboros-mistral2.2-7b.Q4_K_S.gguf\",\"https://huggingface.co/concedo/KobbleTinyV2-1.1B-GGUF/resolve/main/KobbleTiny-Q4_K.gguf\",\"https://huggingface.co/grimjim/kukulemon-7B-GGUF/resolve/main/kukulemon-7B.Q8_0.gguf\",\"https://huggingface.co/mradermacher/LemonKunoichiWizardV3-GGUF/resolve/main/LemonKunoichiWizardV3.Q4_K_M.gguf\",\"https://huggingface.co/Lewdiculous/Kunoichi-DPO-v2-7B-GGUF-Imatrix/resolve/main/Kunoichi-DPO-v2-7B-Q4_K_M-imatrix.gguf\",\"https://huggingface.co/mradermacher/L3-8B-Stheno-v3.2-i1-GGUF/resolve/main/L3-8B-Stheno-v3.2.i1-Q4_K_M.gguf\",\"https://huggingface.co/Lewdiculous/Llama-3-Lumimaid-8B-v0.1-OAS-GGUF-IQ-Imatrix/resolve/main/v2-Llama-3-Lumimaid-8B-v0.1-OAS-Q4_K_M-imat.gguf\",\"https://huggingface.co/bartowski/NeuralDaredevil-8B-abliterated-GGUF/resolve/main/NeuralDaredevil-8B-abliterated-Q4_K_M.gguf\",\"https://huggingface.co/bartowski/L3-8B-Lunaris-v1-GGUF/resolve/main/L3-8B-Lunaris-v1-Q4_K_M.gguf\",\"https://huggingface.co/mradermacher/L3-Umbral-Mind-RP-v2.0-8B-GGUF/resolve/main/L3-Umbral-Mind-RP-v2.0-8B.Q4_K_M.gguf\"]{allow-input: true}\r\n", "Layers = 99 #@param [99]{allow-input: true}\r\n", "ContextSize = 4096 #@param [4096] {allow-input: true}\r\n", "#@markdown
\r\n", diff --git a/common/common.cpp b/common/common.cpp index 3352a29bd..542ad6c6d 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -473,6 +473,14 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa else { invalid_param = true; } return true; } + if (arg == "--attention") { + CHECK_ARG + std::string value(argv[i]); + /**/ if (value == "causal") { params.attention_type = LLAMA_ATTENTION_TYPE_CAUSAL; } + else if (value == "non-causal") { params.attention_type = LLAMA_ATTENTION_TYPE_NON_CAUSAL; } + else { invalid_param = true; } + return true; + } if (arg == "--defrag-thold" || arg == "-dt") { CHECK_ARG params.defrag_thold = std::stof(argv[i]); @@ -758,7 +766,7 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa params.cache_type_v = argv[++i]; return true; } - if (arg == "--multiline-input") { + if (arg == "-mli" || arg == "--multiline-input") { params.multiline_input = true; return true; } @@ -1395,7 +1403,9 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param options.push_back({ "*", " --keep N", "number of tokens to keep from the initial prompt (default: %d, -1 = all)", params.n_keep }); options.push_back({ "*", " --chunks N", "max number of chunks to process (default: %d, -1 = all)", params.n_chunks }); options.push_back({ "*", "-fa, --flash-attn", "enable Flash Attention (default: %s)", params.flash_attn ? "enabled" : "disabled" }); - options.push_back({ "*", "-p, --prompt PROMPT", "prompt to start generation with (default: '%s')", params.prompt.c_str() }); + options.push_back({ "*", "-p, --prompt PROMPT", "prompt to start generation with\n" + "in conversation mode, this will be used as system prompt\n" + "(default: '%s')", params.prompt.c_str() }); options.push_back({ "*", "-f, --file FNAME", "a file containing the prompt (default: none)" }); options.push_back({ "*", " --in-file FNAME", "an input file (repeat to specify multiple files)" }); options.push_back({ "*", "-bf, --binary-file FNAME", "binary file containing the prompt (default: none)" }); @@ -1410,7 +1420,9 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param "halt generation at PROMPT, return control in interactive mode\n" "can be specified more than once for multiple prompts" }); options.push_back({ "main", "-sp, --special", "special tokens output enabled (default: %s)", params.special ? "true" : "false" }); - options.push_back({ "main", "-cnv, --conversation", "run in conversation mode (does not print special tokens and suffix/prefix, use default chat template) (default: %s)", params.conversation ? "true" : "false" }); + options.push_back({ "main", "-cnv, --conversation", "run in conversation mode, does not print special tokens and suffix/prefix\n" + "if suffix/prefix are not specified, default chat template will be used\n" + "(default: %s)", params.conversation ? "true" : "false" }); options.push_back({ "main infill", "-i, --interactive", "run in interactive mode (default: %s)", params.interactive ? "true" : "false" }); options.push_back({ "main infill", "-if, --interactive-first", "run in interactive mode and wait for input right away (default: %s)", params.interactive_first ? "true" : "false" }); options.push_back({ "main infill", "-mli, --multiline-input", "allows you to write or paste multiple lines without ending each in '\\'" }); @@ -1454,6 +1466,7 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param options.push_back({ "main", " --cfg-scale N", "strength of guidance (default: %.1f, 1.0 = disable)", (double)sparams.cfg_scale }); options.push_back({ "main", " --chat-template JINJA_TEMPLATE", "set custom jinja chat template (default: template taken from model's metadata)\n" + "if suffix/prefix are specified, template will be disabled\n" "only commonly used templates are accepted:\n" "https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template" }); options.push_back({ "grammar" }); @@ -1464,8 +1477,10 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param "For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead" }); options.push_back({ "embedding" }); - options.push_back({ "embedding", " --pooling {none,mean,cls}", + options.push_back({ "embedding", " --pooling {none,mean,cls,last}", "pooling type for embeddings, use model default if unspecified" }); + options.push_back({ "embedding", " --attention {causal,non-causal}", + "attention type for embeddings, use model default if unspecified" }); options.push_back({ "context hacking" }); options.push_back({ "*", " --rope-scaling {none,linear,yarn}", @@ -2071,7 +2086,24 @@ std::tuple llama_init_from_gpt_par if (params.warmup) { LOG("warming up the model with an empty run\n"); - std::vector tmp = { llama_token_bos(model), llama_token_eos(model), }; + std::vector tmp; + llama_token bos = llama_token_bos(model); + llama_token eos = llama_token_eos(model); + // some models (e.g. T5) don't have a BOS token + if (bos != -1) { + tmp.push_back(bos); + } + tmp.push_back(eos); + + if (llama_model_has_encoder(model)) { + llama_encode(lctx, llama_batch_get_one(tmp.data(), tmp.size(), 0, 0)); + llama_token decoder_start_token_id = llama_model_decoder_start_token(model); + if (decoder_start_token_id == -1) { + decoder_start_token_id = bos; + } + tmp.clear(); + tmp.push_back(decoder_start_token_id); + } llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0)); llama_kv_cache_clear(lctx); llama_synchronize(lctx); @@ -2154,6 +2186,7 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param cparams.yarn_beta_slow = params.yarn_beta_slow; cparams.yarn_orig_ctx = params.yarn_orig_ctx; cparams.pooling_type = params.pooling_type; + cparams.attention_type = params.attention_type; cparams.defrag_thold = params.defrag_thold; cparams.cb_eval = params.cb_eval; cparams.cb_eval_user_data = params.cb_eval_user_data; diff --git a/common/common.h b/common/common.h index a811713f4..59e10f9bd 100644 --- a/common/common.h +++ b/common/common.h @@ -95,6 +95,7 @@ struct gpt_params { enum llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs enum llama_rope_scaling_type rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED; enum llama_pooling_type pooling_type = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings + enum llama_attention_type attention_type = LLAMA_ATTENTION_TYPE_UNSPECIFIED; // attention type for embeddings // sampling parameters int32_t top_k = 40; // <= 0 to use vocab size @@ -475,5 +476,3 @@ void yaml_dump_string_multiline(FILE * stream, const char * prop_name, const cha void yaml_dump_non_result_info( FILE * stream, const gpt_params & params, const llama_context * lctx, const std::string & timestamp, const std::vector & prompt_tokens, const char * model_desc); - - diff --git a/convert-hf-to-gguf.py b/convert_hf_to_gguf.py similarity index 91% rename from convert-hf-to-gguf.py rename to convert_hf_to_gguf.py index 05fd70171..455eea883 100755 --- a/convert-hf-to-gguf.py +++ b/convert_hf_to_gguf.py @@ -13,7 +13,7 @@ import sys from enum import IntEnum from pathlib import Path from hashlib import sha256 -from typing import TYPE_CHECKING, Any, Callable, ContextManager, Iterable, Iterator, Sequence, TypeVar, cast +from typing import TYPE_CHECKING, Any, Callable, ContextManager, Iterable, Iterator, Literal, Sequence, TypeVar, cast import math import numpy as np @@ -404,7 +404,7 @@ class Model: return tokens, toktypes, tokpre - # NOTE: this function is generated by convert-hf-to-gguf-update.py + # NOTE: this function is generated by convert_hf_to_gguf_update.py # do not modify it manually! # ref: https://github.com/ggerganov/llama.cpp/pull/6920 # Marker: Start get_vocab_base_pre @@ -424,7 +424,7 @@ class Model: res = None - # NOTE: if you get an error here, you need to update the convert-hf-to-gguf-update.py script + # NOTE: if you get an error here, you need to update the convert_hf_to_gguf_update.py script # or pull the latest version of the model from Huggingface # don't edit the hashes manually! if chkhsh == "0ef9807a4087ebef797fc749390439009c3b9eda9ad1a097abbe738f486c01e5": @@ -490,15 +490,18 @@ class Model: if chkhsh == "7fc505bd3104ca1083b150b17d088b59534ede9bde81f0dd2090967d7fe52cee": # ref: https://huggingface.co/LumiOpen/Viking-7B res = "viking" + if chkhsh == "b53802fb28e26d645c3a310b34bfe07da813026ec7c7716883404d5e0f8b1901": + # ref: https://huggingface.co/core42/jais-13b + res = "jais" if res is None: logger.warning("\n") logger.warning("**************************************************************************************") logger.warning("** WARNING: The BPE pre-tokenizer was not recognized!") logger.warning("** There are 2 possible reasons for this:") - logger.warning("** - the model has not been added to convert-hf-to-gguf-update.py yet") + logger.warning("** - the model has not been added to convert_hf_to_gguf_update.py yet") logger.warning("** - the pre-tokenization config has changed upstream") - logger.warning("** Check your model files and convert-hf-to-gguf-update.py and update them accordingly.") + logger.warning("** Check your model files and convert_hf_to_gguf_update.py and update them accordingly.") logger.warning("** ref: https://github.com/ggerganov/llama.cpp/pull/6920") logger.warning("**") logger.warning(f"** chkhsh: {chkhsh}") @@ -674,6 +677,51 @@ class Model: special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens)) special_vocab.add_to_gguf(self.gguf_writer) + def _set_vocab_builtin(self, model_name: Literal["gpt-neox", "llama-spm"], vocab_size: int): + tokenizer_path = Path(sys.path[0]) / "models" / f"ggml-vocab-{model_name}.gguf" + logger.warning(f"Using tokenizer from '{os.path.relpath(tokenizer_path, os.getcwd())}'") + vocab_reader = gguf.GGUFReader(tokenizer_path, "r") + + default_pre = "mpt" if model_name == "gpt-neox" else "default" + + field = vocab_reader.get_field(gguf.Keys.Tokenizer.MODEL) + assert field # tokenizer model + self.gguf_writer.add_tokenizer_model(bytes(field.parts[-1]).decode("utf-8")) + + field = vocab_reader.get_field(gguf.Keys.Tokenizer.PRE) + self.gguf_writer.add_tokenizer_pre(bytes(field.parts[-1]).decode("utf-8") if field else default_pre) + + field = vocab_reader.get_field(gguf.Keys.Tokenizer.LIST) + assert field # token list + self.gguf_writer.add_token_list([bytes(field.parts[i]) for i in field.data][:vocab_size]) + + if model_name == "llama-spm": + field = vocab_reader.get_field(gguf.Keys.Tokenizer.SCORES) + assert field # token scores + self.gguf_writer.add_token_scores([field.parts[i].tolist()[0] for i in field.data][:vocab_size]) + + field = vocab_reader.get_field(gguf.Keys.Tokenizer.TOKEN_TYPE) + assert field # token types + self.gguf_writer.add_token_types([field.parts[i].tolist()[0] for i in field.data][:vocab_size]) + + if model_name != "llama-spm": + field = vocab_reader.get_field(gguf.Keys.Tokenizer.MERGES) + assert field # token merges + self.gguf_writer.add_token_merges([bytes(field.parts[i]) for i in field.data]) + + if (field := vocab_reader.get_field(gguf.Keys.Tokenizer.BOS_ID)) is not None: + self.gguf_writer.add_bos_token_id(field.parts[-1].tolist()[0]) + if (field := vocab_reader.get_field(gguf.Keys.Tokenizer.EOS_ID)) is not None: + self.gguf_writer.add_eos_token_id(field.parts[-1].tolist()[0]) + if (field := vocab_reader.get_field(gguf.Keys.Tokenizer.UNK_ID)) is not None: + self.gguf_writer.add_unk_token_id(field.parts[-1].tolist()[0]) + if (field := vocab_reader.get_field(gguf.Keys.Tokenizer.PAD_ID)) is not None: + self.gguf_writer.add_pad_token_id(field.parts[-1].tolist()[0]) + if (field := vocab_reader.get_field(gguf.Keys.Tokenizer.ADD_BOS)) is not None: + self.gguf_writer.add_add_bos_token(field.parts[-1].tolist()[0]) + if (field := vocab_reader.get_field(gguf.Keys.Tokenizer.ADD_EOS)) is not None: + self.gguf_writer.add_add_eos_token(field.parts[-1].tolist()[0]) + @Model.register("GPTNeoXForCausalLM") class GPTNeoXModel(Model): @@ -1939,7 +1987,7 @@ class Phi3MiniModel(Model): if len(rope_scaling_type) == 0: raise KeyError('Missing the required key rope_scaling.type') - if rope_scaling_type == 'su': + if rope_scaling_type == 'su' or rope_scaling_type == 'longrope': attn_factor = math.sqrt(1 + math.log(scale) / math.log(orig_max_pos_embds)) if scale > 1.0 else 1.0 elif rope_scaling_type == 'yarn': attn_factor = 0.1 * math.log(scale) + 1.0 if scale > 1.0 else 1.0 @@ -2313,6 +2361,8 @@ class GemmaModel(Model): special_vocab._set_special_token("eot", 107) special_vocab.add_to_gguf(self.gguf_writer) + self.gguf_writer.add_add_space_prefix(False) + def set_gguf_parameters(self): hparams = self.hparams block_count = hparams["num_hidden_layers"] @@ -2363,6 +2413,7 @@ class Gemma2Model(Model): special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens)) special_vocab.add_to_gguf(self.gguf_writer) + self.gguf_writer.add_add_space_prefix(False) def set_gguf_parameters(self): @@ -2394,7 +2445,7 @@ class Gemma2Model(Model): raise ValueError("query_pre_attn_scalar must be equal to n_embd / n_head") def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: - del bid # unusem + del bid # unused # lm_head is not used in llama.cpp, while autoawq will include this tensor in model # To prevent errors, skip loading lm_head.weight. @@ -2433,39 +2484,7 @@ class MambaModel(Model): self._set_vocab_sentencepiece() else: # Use the GPT-NeoX tokenizer when no tokenizer files are present - tokenizer_path = Path(sys.path[0]) / "models" / "ggml-vocab-gpt-neox.gguf" - logger.warning(f"Using tokenizer from '{os.path.relpath(tokenizer_path, os.getcwd())}'") - neox_reader = gguf.GGUFReader(tokenizer_path, "r") - - field = neox_reader.get_field(gguf.Keys.Tokenizer.MODEL) - self.gguf_writer.add_tokenizer_model(bytes(field.parts[-1]).decode("utf-8") if field else "gpt2") - - field = neox_reader.get_field(gguf.Keys.Tokenizer.PRE) - self.gguf_writer.add_tokenizer_pre(bytes(field.parts[-1]).decode("utf-8") if field else "mpt") - - field = neox_reader.get_field(gguf.Keys.Tokenizer.LIST) - assert field - self.gguf_writer.add_token_list([bytes(field.parts[i]) for i in field.data][:vocab_size]) - - field = neox_reader.get_field(gguf.Keys.Tokenizer.TOKEN_TYPE) - assert field - self.gguf_writer.add_token_types([field.parts[i].tolist()[0] for i in field.data][:vocab_size]) - - field = neox_reader.get_field(gguf.Keys.Tokenizer.MERGES) - assert field - self.gguf_writer.add_token_merges([bytes(field.parts[i]) for i in field.data]) - - field = neox_reader.get_field(gguf.Keys.Tokenizer.BOS_ID) - self.gguf_writer.add_bos_token_id(field.parts[-1].tolist()[0] if field else 1) - - field = neox_reader.get_field(gguf.Keys.Tokenizer.EOS_ID) - self.gguf_writer.add_eos_token_id(field.parts[-1].tolist()[0] if field else 0) - - field = neox_reader.get_field(gguf.Keys.Tokenizer.UNK_ID) - self.gguf_writer.add_unk_token_id(field.parts[-1].tolist()[0] if field else 0) - - field = neox_reader.get_field(gguf.Keys.Tokenizer.PAD_ID) - self.gguf_writer.add_pad_token_id(field.parts[-1].tolist()[0] if field else 0) + self._set_vocab_builtin("gpt-neox", vocab_size) def set_gguf_parameters(self): d_model = self.find_hparam(["hidden_size", "d_model"]) @@ -2617,6 +2636,82 @@ class JinaBertV2Model(BertModel): self.gguf_writer.add_add_eos_token(True) +@Model.register("OpenELMForCausalLM") +class OpenELMModel(Model): + model_arch = gguf.MODEL_ARCH.OPENELM + + @staticmethod + def _make_divisible(v: float | int, divisor: int) -> int: + # ref: https://huggingface.co/apple/OpenELM-270M-Instruct/blob/eb111ff2e6724348e5b905984063d4064d4bc579/configuration_openelm.py#L34-L38 + new_v = max(divisor, int(v + divisor / 2) // divisor * divisor) + # Make sure that round down does not go down by more than 10%. + if new_v < 0.9 * v: + new_v += divisor + return new_v + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + ffn_multipliers: list[float] = self.hparams["ffn_multipliers"] + ffn_dim_divisor: int = self.hparams["ffn_dim_divisor"] + self._n_embd: int = self.hparams["model_dim"] + self._num_kv_heads: list[int] = self.hparams["num_kv_heads"] + self._num_query_heads: list[int] = self.hparams["num_query_heads"] + self._ffn_dims: list[int] = [ + OpenELMModel._make_divisible(multiplier * self._n_embd, ffn_dim_divisor) + for multiplier in ffn_multipliers + ] + assert isinstance(self._num_kv_heads, list) and isinstance(self._num_kv_heads[0], int) + assert isinstance(self._num_query_heads, list) and isinstance(self._num_query_heads[0], int) + + # Uses the tokenizer from meta-llama/Llama-2-7b-hf + def set_vocab(self): + try: + self._set_vocab_sentencepiece() + except FileNotFoundError: + self._set_vocab_builtin("llama-spm", self.hparams["vocab_size"]) + + def set_gguf_parameters(self): + n_embd = self._n_embd + head_dim = self.hparams["head_dim"] + rot_pct = 1.0 + assert self.block_count == len(self._num_kv_heads) + assert self.block_count == len(self._num_query_heads) + assert self.block_count == len(self._ffn_dims) + + self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name) + self.gguf_writer.add_block_count(self.block_count) + self.gguf_writer.add_context_length(self.hparams["max_context_length"]) + self.gguf_writer.add_embedding_length(n_embd) + self.gguf_writer.add_feed_forward_length(self._ffn_dims) + self.gguf_writer.add_head_count(self._num_query_heads) + self.gguf_writer.add_head_count_kv(self._num_kv_heads) + self.gguf_writer.add_rope_freq_base(self.hparams["rope_freq_constant"]) + # https://huggingface.co/apple/OpenELM-270M-Instruct/blob/c401df2/modeling_openelm.py#L30 + self.gguf_writer.add_layer_norm_rms_eps(1e-6) + self.gguf_writer.add_rope_dimension_count(int(rot_pct * head_dim)) + self.gguf_writer.add_key_length(head_dim) + self.gguf_writer.add_value_length(head_dim) + self.gguf_writer.add_file_type(self.ftype) + + def find_hparam(self, keys: Iterable[str], optional: bool = False) -> Any: + if "n_layers" in keys: + return self.hparams["num_transformer_layers"] + + return super().find_hparam(keys, optional) + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + + # split ff + if bid is not None and name == f"transformer.layers.{bid}.ffn.proj_1.weight": + ff_dim = self._ffn_dims[bid] + yield (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE, bid), data_torch[:ff_dim]) + yield (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP, bid), data_torch[ff_dim:]) + return + + yield (self.map_tensor_name(name), data_torch) + + @Model.register("ArcticForCausalLM") class ArcticModel(Model): model_arch = gguf.MODEL_ARCH.ARCTIC @@ -2847,11 +2942,17 @@ class DeepseekV2Model(Model): raise ValueError(f"Unprocessed experts: {experts}") -@Model.register("T5ForConditionalGeneration") @Model.register("T5WithLMHeadModel") +@Model.register("T5ForConditionalGeneration") +@Model.register("MT5ForConditionalGeneration") +@Model.register("UMT5ForConditionalGeneration") class T5Model(Model): model_arch = gguf.MODEL_ARCH.T5 + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + self.shared_token_embeddings_found = False + def set_vocab(self): # to avoid TypeError: Descriptors cannot be created directly # exception when importing sentencepiece_model_pb2 @@ -2859,17 +2960,29 @@ class T5Model(Model): from sentencepiece import SentencePieceProcessor from sentencepiece import sentencepiece_model_pb2 as model - tokenizer_path = self.dir_model / 'spiece.model' + tokenizer_path = self.dir_model / 'tokenizer.model' + + # many older models use spiece.model tokenizer model filename + if not tokenizer_path.is_file(): + tokenizer_path = self.dir_model / 'spiece.model' if not tokenizer_path.is_file(): raise FileNotFoundError(f"File not found: {tokenizer_path}") sentencepiece_model = model.ModelProto() sentencepiece_model.ParseFromString(open(tokenizer_path, "rb").read()) + + # some models like Pile-T5 family use BPE tokenizer instead of Unigram + if sentencepiece_model.trainer_spec.model_type == 2: # BPE + # assure the tokenizer model file name is correct + assert tokenizer_path.name == 'tokenizer.model' + return self._set_vocab_sentencepiece() + else: + assert sentencepiece_model.trainer_spec.model_type == 1 # UNIGRAM + add_prefix = sentencepiece_model.normalizer_spec.add_dummy_prefix remove_whitespaces = sentencepiece_model.normalizer_spec.remove_extra_whitespaces precompiled_charsmap = sentencepiece_model.normalizer_spec.precompiled_charsmap - assert sentencepiece_model.trainer_spec.model_type == 1 # UNIGRAM tokenizer = SentencePieceProcessor() tokenizer.LoadFromFile(str(tokenizer_path)) @@ -2939,7 +3052,10 @@ class T5Model(Model): def set_gguf_parameters(self): self.gguf_writer.add_name("T5") - self.gguf_writer.add_context_length(self.hparams["n_positions"]) + if (n_ctx := self.find_hparam(["n_positions"], optional=True)) is None: + logger.warning("Couldn't find context length in config.json, assuming default value of 512") + n_ctx = 512 + self.gguf_writer.add_context_length(n_ctx) self.gguf_writer.add_embedding_length(self.hparams["d_model"]) self.gguf_writer.add_feed_forward_length(self.hparams["d_ff"]) self.gguf_writer.add_block_count(self.hparams["num_layers"]) @@ -2955,16 +3071,111 @@ class T5Model(Model): def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: del bid # unused - # Sometimes T5 and Flan-T5 based models contain "encoder.embed_tokens.weight" tensor or - # "decoder.embed_tokens.weight" tensors that are duplicates of "shared.weight" tensor - # To prevent errors caused by an unnecessary unmapped tensor, skip both of them and use only "shared.weight". - if name == "decoder.embed_tokens.weight" or name == "encoder.embed_tokens.weight": - logger.debug(f"Skipping tensor {name!r} in safetensors so that convert can end normally.") - return [] + # T5 based models contain shared token embeddings tensors saved randomly as either "encoder.embed_tokens.weight", + # "decoder.embed_tokens.weight" or "shared.weight" tensor. In some models there are even multiple of them stored + # in the safetensors files. We use the first tensor from these three as the token embeddings for both encoder + # and decoder and ignore the remaining ones. + if name in ["decoder.embed_tokens.weight", "encoder.embed_tokens.weight", "shared.weight"]: + if not self.shared_token_embeddings_found: + name = "shared.weight" + self.shared_token_embeddings_found = True + else: + logger.debug(f"Skipping shared tensor {name!r} in safetensors so that convert can end normally.") + return [] return [(self.map_tensor_name(name), data_torch)] +@Model.register("JAISLMHeadModel") +class JaisModel(Model): + model_arch = gguf.MODEL_ARCH.JAIS + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + # SwigLU activation + assert self.hparams["activation_function"] == "swiglu" + # ALiBi position embedding + assert self.hparams["position_embedding_type"] == "alibi" + + # Embeddings scale + self.embeddings_scale = 1.0 + # note: For some JAIS flavors, output is tied to (same as) wte in original model + self.output_is_wte = False + if 'mup_embeddings_scale' in self.hparams: + self.output_is_wte = True # Hack (?) + self.embeddings_scale = self.hparams['mup_embeddings_scale'] + elif 'embeddings_scale' in self.hparams: + self.embeddings_scale = self.hparams['embeddings_scale'] + else: + assert False + + self.width_scale = 1.0 + if 'mup_output_alpha' in self.hparams: + assert 'mup_width_scale' in self.hparams + self.width_scale = self.hparams['mup_output_alpha'] * self.hparams['mup_width_scale'] + elif 'width_scale' in self.hparams: + self.width_scale = self.hparams['width_scale'] + else: + assert False + + self.max_alibi_bias = 8.0 + + def set_vocab(self): + self._set_vocab_gpt2() + + def set_gguf_parameters(self): + self.gguf_writer.add_name(self.dir_model.name) + self.gguf_writer.add_block_count(self.hparams["n_layer"]) + self.gguf_writer.add_context_length(self.hparams["n_positions"]) + self.gguf_writer.add_embedding_length(self.hparams["n_embd"]) + self.gguf_writer.add_feed_forward_length(self.hparams["n_inner"]) + self.gguf_writer.add_head_count(self.hparams["n_head"]) + self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"]) + self.gguf_writer.add_file_type(self.ftype) + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + del bid # unused + + tensors: list[tuple[str, Tensor]] = [] + + # we don't need these + if name.endswith((".attn.bias")): + return tensors + + if name.endswith(("relative_pe.slopes")): + # Calculate max ALiBi bias (this is the inverse of the ALiBi calculation) + # Some other models has max_alibi_bias spelled out explicitly in the hyperparams, + # but Jais's PyTorch model simply precalculates the slope values and places them + # in relative_pes.slopes + n_head_closest_log2 = 2 ** math.floor(math.log2(self.hparams["n_head"])) + first_val = float(data_torch._data[0]) + self.max_alibi_bias = -round(math.log2(first_val) * n_head_closest_log2) + + return tensors + + if name.endswith((".c_attn.weight", ".c_proj.weight", ".c_fc.weight", ".c_fc2.weight")): + data_torch = data_torch.transpose(1, 0) + + new_name = self.map_tensor_name(name) + + if new_name == self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD): + tensors.append((new_name, data_torch * self.embeddings_scale)) + if self.output_is_wte: + tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch * self.width_scale)) + elif new_name == self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT): + assert not self.output_is_wte + tensors.append((new_name, data_torch * self.width_scale)) + else: + tensors.append((new_name, data_torch)) + + return tensors + + def write_tensors(self): + super().write_tensors() + self.gguf_writer.add_max_alibi_bias(self.max_alibi_bias) + + ###### CONVERSION LOGIC ###### @@ -3014,10 +3225,6 @@ def parse_args() -> argparse.Namespace: "--vocab-only", action="store_true", help="extract only the vocab", ) - parser.add_argument( - "--awq-path", type=Path, default=None, - help="Path to scale awq cache file", - ) parser.add_argument( "--outfile", type=Path, help="path to write to; default: based on input. {ftype} will be replaced by the outtype.", @@ -3095,19 +3302,6 @@ def main() -> None: dir_model = args.model - if args.awq_path: - sys.path.insert(1, str(Path(__file__).parent / 'awq-py')) - from awq.apply_awq import add_scale_weights # type: ignore[import-not-found] - tmp_model_path = args.model / "weighted_model" - dir_model = tmp_model_path - if tmp_model_path.is_dir(): - logger.info(f"{tmp_model_path} exists as a weighted model.") - else: - tmp_model_path.mkdir(parents=True, exist_ok=True) - logger.info("Saving new weighted model ...") - add_scale_weights(str(args.model), str(args.awq_path), str(tmp_model_path)) - logger.info(f"Saved weighted model at {tmp_model_path}.") - if not dir_model.is_dir(): logger.error(f'Error: {args.model} is not a directory') sys.exit(1) diff --git a/convert-hf-to-gguf-update.py b/convert_hf_to_gguf_update.py similarity index 87% rename from convert-hf-to-gguf-update.py rename to convert_hf_to_gguf_update.py index 2758214fa..e4165ae2d 100755 --- a/convert-hf-to-gguf-update.py +++ b/convert_hf_to_gguf_update.py @@ -2,7 +2,7 @@ # -*- coding: utf-8 -*- # This script downloads the tokenizer models of the specified models from Huggingface and -# generates the get_vocab_base_pre() function for convert-hf-to-gguf.py +# generates the get_vocab_base_pre() function for convert_hf_to_gguf.py # # This is necessary in order to analyze the type of pre-tokenizer used by the model and # provide the necessary information to llama.cpp via the GGUF header in order to implement @@ -15,9 +15,9 @@ # - Add a new model to the "models" list # - Run the script with your huggingface token: # -# python3 convert-hf-to-gguf-update.py +# python3 convert_hf_to_gguf_update.py # -# - Copy-paste the generated get_vocab_base_pre() function into convert-hf-to-gguf.py +# - Copy-paste the generated get_vocab_base_pre() function into convert_hf_to_gguf.py # - Update llama.cpp with the new pre-tokenizer if necessary # # TODO: generate tokenizer tests for llama.cpp @@ -37,7 +37,7 @@ from enum import IntEnum, auto from transformers import AutoTokenizer logging.basicConfig(level=logging.DEBUG) -logger = logging.getLogger("convert-hf-to-gguf-update") +logger = logging.getLogger("convert_hf_to_gguf_update") sess = requests.Session() @@ -45,6 +45,7 @@ class TOKENIZER_TYPE(IntEnum): SPM = auto() BPE = auto() WPM = auto() + UGM = auto() # TODO: this string has to exercise as much pre-tokenizer functionality as possible @@ -55,10 +56,10 @@ if len(sys.argv) == 2: token = sys.argv[1] if not token.startswith("hf_"): logger.info("Huggingface token seems invalid") - logger.info("Usage: python convert-hf-to-gguf-update.py ") + logger.info("Usage: python convert_hf_to_gguf_update.py ") sys.exit(1) else: - logger.info("Usage: python convert-hf-to-gguf-update.py ") + logger.info("Usage: python convert_hf_to_gguf_update.py ") sys.exit(1) # TODO: add models here, base models preferred @@ -86,6 +87,10 @@ models = [ {"name": "poro-chat", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LumiOpen/Poro-34B-chat", }, {"name": "jina-v2-code", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-code", }, {"name": "viking", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LumiOpen/Viking-7B", }, # Also used for Viking 13B and 33B + {"name": "gemma", "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/google/gemma-2b", }, + {"name": "gemma-2", "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/google/gemma-2-9b", }, + {"name": "jais", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/core42/jais-13b", }, + {"name": "t5", "tokt": TOKENIZER_TYPE.UGM, "repo": "https://huggingface.co/google-t5/t5-small", }, ] @@ -107,9 +112,13 @@ def download_model(model): os.makedirs(f"models/tokenizers/{name}", exist_ok=True) files = ["config.json", "tokenizer.json", "tokenizer_config.json"] + if tokt == TOKENIZER_TYPE.SPM: files.append("tokenizer.model") + if tokt == TOKENIZER_TYPE.UGM: + files.append("spiece.model") + for file in files: save_path = f"models/tokenizers/{name}/{file}" if os.path.isfile(save_path): @@ -125,14 +134,14 @@ for model in models: logger.error(f"Failed to download model {model['name']}. Error: {e}") -# generate the source code for the convert-hf-to-gguf.py:get_vocab_base_pre() function: +# generate the source code for the convert_hf_to_gguf.py:get_vocab_base_pre() function: src_ifs = "" for model in models: name = model["name"] tokt = model["tokt"] - if tokt == TOKENIZER_TYPE.SPM: + if tokt == TOKENIZER_TYPE.SPM or tokt == TOKENIZER_TYPE.UGM: continue # Skip if the tokenizer folder does not exist or there are other download issues previously @@ -142,7 +151,10 @@ for model in models: # create the tokenizer try: - tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}") + if name == "t5": + tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}", use_fast=False) + else: + tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}") except OSError as e: logger.error(f"Error loading tokenizer for model {name}. The model may not exist or is not accessible with the provided token. Error: {e}") continue # Skip to the next model if the tokenizer can't be loaded @@ -189,7 +201,7 @@ src_func = f""" res = None - # NOTE: if you get an error here, you need to update the convert-hf-to-gguf-update.py script + # NOTE: if you get an error here, you need to update the convert_hf_to_gguf_update.py script # or pull the latest version of the model from Huggingface # don't edit the hashes manually! {src_ifs} @@ -198,9 +210,9 @@ src_func = f""" logger.warning("**************************************************************************************") logger.warning("** WARNING: The BPE pre-tokenizer was not recognized!") logger.warning("** There are 2 possible reasons for this:") - logger.warning("** - the model has not been added to convert-hf-to-gguf-update.py yet") + logger.warning("** - the model has not been added to convert_hf_to_gguf_update.py yet") logger.warning("** - the pre-tokenization config has changed upstream") - logger.warning("** Check your model files and convert-hf-to-gguf-update.py and update them accordingly.") + logger.warning("** Check your model files and convert_hf_to_gguf_update.py and update them accordingly.") logger.warning("** ref: https://github.com/ggerganov/llama.cpp/pull/6920") logger.warning("**") logger.warning(f"** chkhsh: {{chkhsh}}") @@ -214,7 +226,7 @@ src_func = f""" return res """ -convert_py_pth = pathlib.Path("convert-hf-to-gguf.py") +convert_py_pth = pathlib.Path("convert_hf_to_gguf.py") convert_py = convert_py_pth.read_text(encoding="utf-8") convert_py = re.sub( r"(# Marker: Start get_vocab_base_pre)(.+?)( +# Marker: End get_vocab_base_pre)", @@ -225,7 +237,7 @@ convert_py = re.sub( convert_py_pth.write_text(convert_py, encoding="utf-8") -logger.info("+++ convert-hf-to-gguf.py was updated") +logger.info("+++ convert_hf_to_gguf.py was updated") # generate tests for each tokenizer model @@ -263,6 +275,7 @@ tests = [ "\n =", "' era", "Hello, y'all! How are you 😁 ?我想在apple工作1314151天～", + "!!!!!!", "3", "33", "333", @@ -272,7 +285,8 @@ tests = [ "3333333", "33333333", "333333333", - # "Cửa Việt", # llama-bpe fails on this + "Cửa Việt", # llama-bpe fails on this + " discards", chktxt, ] @@ -300,7 +314,10 @@ for model in models: # create the tokenizer try: - tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}") + if name == "t5": + tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}", use_fast=False) + else: + tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}") except OSError as e: logger.error(f"Failed to load tokenizer for model {name}. Error: {e}") continue # Skip this model and continue with the next one in the loop @@ -326,6 +343,6 @@ logger.info("\nRun the following commands to generate the vocab files for testin for model in models: name = model["name"] - print(f"python3 convert-hf-to-gguf.py models/tokenizers/{name}/ --outfile models/ggml-vocab-{name}.gguf --vocab-only") # noqa: NP100 + print(f"python3 convert_hf_to_gguf.py models/tokenizers/{name}/ --outfile models/ggml-vocab-{name}.gguf --vocab-only") # noqa: NP100 logger.info("\n") diff --git a/convert-llama-ggml-to-gguf.py b/convert_llama_ggml_to_gguf.py similarity index 100% rename from convert-llama-ggml-to-gguf.py rename to convert_llama_ggml_to_gguf.py diff --git a/docs/HOWTO-add-model.md b/docs/HOWTO-add-model.md index 3eec077ea..87093cedd 100644 --- a/docs/HOWTO-add-model.md +++ b/docs/HOWTO-add-model.md @@ -17,7 +17,7 @@ Also, it is important to check that the examples and main ggml backends (CUDA, M ### 1. Convert the model to GGUF This step is done in python with a `convert` script using the [gguf](https://pypi.org/project/gguf/) library. -Depending on the model architecture, you can use either [convert-hf-to-gguf.py](../convert-hf-to-gguf.py) or [examples/convert-legacy-llama.py](../examples/convert-legacy-llama.py) (for `llama/llama2` models in `.pth` format). +Depending on the model architecture, you can use either [convert_hf_to_gguf.py](../convert_hf_to_gguf.py) or [examples/convert_legacy_llama.py](../examples/convert_legacy_llama.py) (for `llama/llama2` models in `.pth` format). The convert script reads the model configuration, tokenizer, tensor names+data and converts them to GGUF metadata and tensors. diff --git a/examples/batched/batched.cpp b/examples/batched/batched.cpp index 62d9b144d..2442e954d 100644 --- a/examples/batched/batched.cpp +++ b/examples/batched/batched.cpp @@ -93,14 +93,34 @@ int main(int argc, char ** argv) { // create a llama_batch // we use this object to submit token data for decoding - llama_batch batch = llama_batch_init(std::max(tokens_list.size(), (size_t)n_parallel), 0, 1); + llama_batch batch = llama_batch_init(std::max(tokens_list.size(), (size_t) n_parallel), 0, n_parallel); + + std::vector seq_ids(n_parallel, 0); + for (int32_t i = 0; i < n_parallel; ++i) { + seq_ids[i] = i; + } // evaluate the initial prompt for (size_t i = 0; i < tokens_list.size(); ++i) { - llama_batch_add(batch, tokens_list[i], i, { 0 }, false); + llama_batch_add(batch, tokens_list[i], i, seq_ids, false); } GGML_ASSERT(batch.n_tokens == (int) tokens_list.size()); + if (llama_model_has_encoder(model)) { + if (llama_encode(ctx, batch)) { + LOG_TEE("%s : failed to eval\n", __func__); + return 1; + } + + llama_token decoder_start_token_id = llama_model_decoder_start_token(model); + if (decoder_start_token_id == -1) { + decoder_start_token_id = llama_token_bos(model); + } + + llama_batch_clear(batch); + llama_batch_add(batch, decoder_start_token_id, 0, seq_ids, false); + } + // llama_decode will output logits only for the last token of the prompt batch.logits[batch.n_tokens - 1] = true; @@ -109,11 +129,11 @@ int main(int argc, char ** argv) { return 1; } - // assign the system KV cache to all parallel sequences - // this way, the parallel sequences will "reuse" the prompt tokens without having to copy them - for (int32_t i = 1; i < n_parallel; ++i) { - llama_kv_cache_seq_cp(ctx, 0, i, -1, -1); - } + //// assign the system KV cache to all parallel sequences + //// this way, the parallel sequences will "reuse" the prompt tokens without having to copy them + //for (int32_t i = 1; i < n_parallel; ++i) { + // llama_kv_cache_seq_cp(ctx, 0, i, -1, -1); + //} if (n_parallel > 1) { LOG_TEE("\n\n%s: generating %d sequences ...\n", __func__, n_parallel); diff --git a/examples/convert-legacy-llama.py b/examples/convert_legacy_llama.py similarity index 100% rename from examples/convert-legacy-llama.py rename to examples/convert_legacy_llama.py diff --git a/examples/embedding/README.md b/examples/embedding/README.md index 86df18958..e3705b454 100644 --- a/examples/embedding/README.md +++ b/examples/embedding/README.md @@ -58,4 +58,3 @@ The above command will output space-separated float values. ```powershell embedding.exe -p 'Castle<#sep#>Stronghold<#sep#>Dog<#sep#>Cat' --embd-separator '<#sep#>' --embd-normalize 2 --embd-output-format '' -m './path/to/model.gguf' --n-gpu-layers 99 --log-disable 2>/dev/null ``` - diff --git a/examples/finetune/convert-finetune-checkpoint-to-gguf.py b/examples/finetune/convert_finetune_checkpoint_to_gguf.py similarity index 100% rename from examples/finetune/convert-finetune-checkpoint-to-gguf.py rename to examples/finetune/convert_finetune_checkpoint_to_gguf.py diff --git a/examples/infill/infill.cpp b/examples/infill/infill.cpp index 1556a2fb7..f1d82d363 100644 --- a/examples/infill/infill.cpp +++ b/examples/infill/infill.cpp @@ -660,4 +660,3 @@ int main(int argc, char ** argv) { return 0; } - diff --git a/examples/json-schema-pydantic-example.py b/examples/json_schema_pydantic_example.py similarity index 98% rename from examples/json-schema-pydantic-example.py rename to examples/json_schema_pydantic_example.py index 2a24f8118..c7ca7b8d9 100644 --- a/examples/json-schema-pydantic-example.py +++ b/examples/json_schema_pydantic_example.py @@ -1,7 +1,7 @@ # Usage: #! ./llama-server -m some-model.gguf & #! pip install pydantic -#! python json-schema-pydantic-example.py +#! python json_schema_pydantic_example.py from pydantic import BaseModel, Extra, TypeAdapter from annotated_types import MinLen diff --git a/examples/llava/MobileVLM-README.md b/examples/llava/MobileVLM-README.md index f6c619c87..06a65fba4 100644 --- a/examples/llava/MobileVLM-README.md +++ b/examples/llava/MobileVLM-README.md @@ -30,16 +30,16 @@ git clone https://huggingface.co/mtgv/MobileVLM-1.7B git clone https://huggingface.co/openai/clip-vit-large-patch14-336 ``` -2. Use `llava-surgery.py` to split the LLaVA model to LLaMA and multimodel projector constituents: +2. Use `llava_surgery.py` to split the LLaVA model to LLaMA and multimodel projector constituents: ```sh -python ./examples/llava/llava-surgery.py -m path/to/MobileVLM-1.7B +python ./examples/llava/llava_surgery.py -m path/to/MobileVLM-1.7B ``` -3. Use `convert-image-encoder-to-gguf.py` with `--projector-type ldp` (for **V2** please use `--projector-type ldpv2`) to convert the LLaVA image encoder to GGUF: +3. Use `convert_image_encoder_to_gguf.py` with `--projector-type ldp` (for **V2** please use `--projector-type ldpv2`) to convert the LLaVA image encoder to GGUF: ```sh -python ./examples/llava/convert-image-encoder-to-gguf \ +python ./examples/llava/convert_image_encoder_to_gguf \ -m path/to/clip-vit-large-patch14-336 \ --llava-projector path/to/MobileVLM-1.7B/llava.projector \ --output-dir path/to/MobileVLM-1.7B \ @@ -47,17 +47,17 @@ python ./examples/llava/convert-image-encoder-to-gguf \ ``` ```sh -python ./examples/llava/convert-image-encoder-to-gguf \ +python ./examples/llava/convert_image_encoder_to_gguf \ -m path/to/clip-vit-large-patch14-336 \ --llava-projector path/to/MobileVLM-1.7B_V2/llava.projector \ --output-dir path/to/MobileVLM-1.7B_V2 \ --projector-type ldpv2 ``` -4. Use `examples/convert-legacy-llama.py` to convert the LLaMA part of LLaVA to GGUF: +4. Use `examples/convert_legacy_llama.py` to convert the LLaMA part of LLaVA to GGUF: ```sh -python ./examples/convert-legacy-llama.py path/to/MobileVLM-1.7B +python ./examples/convert_legacy_llama.py path/to/MobileVLM-1.7B ``` 5. Use `quantize` to convert LLaMA part's DataType from `fp16` to `q4_k` diff --git a/examples/llava/README.md b/examples/llava/README.md index f4554de67..012451361 100644 --- a/examples/llava/README.md +++ b/examples/llava/README.md @@ -38,22 +38,22 @@ git clone https://huggingface.co/openai/clip-vit-large-patch14-336 pip install -r examples/llava/requirements.txt ``` -3. Use `llava-surgery.py` to split the LLaVA model to LLaMA and multimodel projector constituents: +3. Use `llava_surgery.py` to split the LLaVA model to LLaMA and multimodel projector constituents: ```sh -python ./examples/llava/llava-surgery.py -m ../llava-v1.5-7b +python ./examples/llava/llava_surgery.py -m ../llava-v1.5-7b ``` -4. Use `convert-image-encoder-to-gguf.py` to convert the LLaVA image encoder to GGUF: +4. Use `convert_image_encoder_to_gguf.py` to convert the LLaVA image encoder to GGUF: ```sh -python ./examples/llava/convert-image-encoder-to-gguf.py -m ../clip-vit-large-patch14-336 --llava-projector ../llava-v1.5-7b/llava.projector --output-dir ../llava-v1.5-7b +python ./examples/llava/convert_image_encoder_to_gguf.py -m ../clip-vit-large-patch14-336 --llava-projector ../llava-v1.5-7b/llava.projector --output-dir ../llava-v1.5-7b ``` -5. Use `examples/convert-legacy-llama.py` to convert the LLaMA part of LLaVA to GGUF: +5. Use `examples/convert_legacy_llama.py` to convert the LLaMA part of LLaVA to GGUF: ```sh -python ./examples/convert-legacy-llama.py ../llava-v1.5-7b --skip-unknown +python ./examples/convert_legacy_llama.py ../llava-v1.5-7b --skip-unknown ``` Now both the LLaMA part and the image encoder are in the `llava-v1.5-7b` directory. @@ -70,9 +70,9 @@ git clone https://huggingface.co/liuhaotian/llava-v1.6-vicuna-7b pip install -r examples/llava/requirements.txt ``` -3) Use `llava-surgery-v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models: +3) Use `llava_surgery_v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models: ```console -python examples/llava/llava-surgery-v2.py -C -m ../llava-v1.6-vicuna-7b/ +python examples/llava/llava_surgery_v2.py -C -m ../llava-v1.6-vicuna-7b/ ``` - you will find a llava.projector and a llava.clip file in your model directory @@ -86,13 +86,13 @@ curl -s -q https://huggingface.co/cmp-nct/llava-1.6-gguf/raw/main/config_vit.jso 5) Create the visual gguf model: ```console -python ./examples/llava/convert-image-encoder-to-gguf.py -m vit --llava-projector vit/llava.projector --output-dir vit --clip-model-is-vision +python ./examples/llava/convert_image_encoder_to_gguf.py -m vit --llava-projector vit/llava.projector --output-dir vit --clip-model-is-vision ``` - This is similar to llava-1.5, the difference is that we tell the encoder that we are working with the pure vision model part of CLIP 6) Then convert the model to gguf format: ```console -python ./examples/convert-legacy-llama.py ../llava-v1.6-vicuna-7b/ --skip-unknown +python ./examples/convert_legacy_llama.py ../llava-v1.6-vicuna-7b/ --skip-unknown ``` 7) And finally we can run the llava cli using the 1.6 model version: diff --git a/examples/llava/convert-image-encoder-to-gguf.py b/examples/llava/convert_image_encoder_to_gguf.py similarity index 100% rename from examples/llava/convert-image-encoder-to-gguf.py rename to examples/llava/convert_image_encoder_to_gguf.py diff --git a/examples/llava/llava-surgery.py b/examples/llava/llava_surgery.py similarity index 100% rename from examples/llava/llava-surgery.py rename to examples/llava/llava_surgery.py diff --git a/examples/llava/llava-surgery-v2.py b/examples/llava/llava_surgery_v2.py similarity index 100% rename from examples/llava/llava-surgery-v2.py rename to examples/llava/llava_surgery_v2.py diff --git a/examples/llava/requirements.txt b/examples/llava/requirements.txt index 17cb4d5e5..4713f0a34 100644 --- a/examples/llava/requirements.txt +++ b/examples/llava/requirements.txt @@ -1,3 +1,3 @@ --r ../../requirements/requirements-convert-legacy-llama.txt +-r ../../requirements/requirements-convert_legacy_llama.txt pillow~=10.2.0 -torch~=2.1.1 +torch~=2.2.1 diff --git a/examples/lookup/README.md b/examples/lookup/README.md index 5bfb0de93..71c345c03 100644 --- a/examples/lookup/README.md +++ b/examples/lookup/README.md @@ -10,4 +10,3 @@ More info: https://github.com/ggerganov/llama.cpp/pull/4484 https://github.com/ggerganov/llama.cpp/issues/4226 - diff --git a/examples/main-cmake-pkg/.gitignore b/examples/main-cmake-pkg/.gitignore index e32c11c7f..67c01d64c 100644 --- a/examples/main-cmake-pkg/.gitignore +++ b/examples/main-cmake-pkg/.gitignore @@ -48,4 +48,3 @@ build*/ out/ tmp/ - diff --git a/examples/main/main.cpp b/examples/main/main.cpp index 002f136fd..6dcdea9bd 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -38,7 +38,8 @@ static gpt_params * g_params; static std::vector * g_input_tokens; static std::ostringstream * g_output_ss; static std::vector * g_output_tokens; -static bool is_interacting = false; +static bool is_interacting = false; +static bool need_insert_eot = false; static bool file_exists(const std::string & path) { std::ifstream f(path.c_str()); @@ -100,7 +101,8 @@ static void write_logfile( static void sigint_handler(int signo) { if (signo == SIGINT) { if (!is_interacting && g_params->interactive) { - is_interacting = true; + is_interacting = true; + need_insert_eot = true; } else { console::cleanup(); printf("\n"); @@ -225,7 +227,14 @@ int main(int argc, char ** argv) { __func__, n_ctx_train, n_ctx); } - LOG_TEE("%s: chat template example: %s\n", __func__, llama_chat_format_example(model, params.chat_template).c_str()); + // print chat template example in conversation mode + if (params.conversation) { + if (params.enable_chat_template) { + LOG_TEE("%s: chat template example: %s\n", __func__, llama_chat_format_example(model, params.chat_template).c_str()); + } else { + LOG_TEE("%s: in-suffix/prefix is specified, chat template will be disabled\n", __func__); + } + } // print system information { @@ -256,13 +265,15 @@ int main(int argc, char ** argv) { } const bool add_bos = llama_should_add_bos_token(model); - GGML_ASSERT(llama_add_eos_token(model) != 1); + if (!llama_model_has_encoder(model)) { + GGML_ASSERT(llama_add_eos_token(model) != 1); + } LOG("add_bos: %d\n", add_bos); std::vector embd_inp; { - auto prompt = (params.conversation && params.enable_chat_template) + auto prompt = (params.conversation && params.enable_chat_template && !params.prompt.empty()) ? chat_add_and_format(model, chat_msgs, "system", params.prompt) // format the system prompt in conversation mode : params.prompt; if (params.interactive_first || !params.prompt.empty() || session_tokens.empty()) { @@ -518,6 +529,24 @@ int main(int argc, char ** argv) { exit(1); } + if (llama_model_has_encoder(model)) { + int enc_input_size = embd_inp.size(); + llama_token * enc_input_buf = embd_inp.data(); + + if (llama_encode(ctx, llama_batch_get_one(enc_input_buf, enc_input_size, 0, 0))) { + LOG_TEE("%s : failed to eval\n", __func__); + return 1; + } + + llama_token decoder_start_token_id = llama_model_decoder_start_token(model); + if (decoder_start_token_id == -1) { + decoder_start_token_id = llama_token_bos(model); + } + + embd_inp.clear(); + embd_inp.push_back(decoder_start_token_id); + } + while ((n_remain != 0 && !is_antiprompt) || params.interactive) { // predict if (!embd.empty()) { @@ -886,6 +915,13 @@ int main(int argc, char ** argv) { LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp).c_str()); + // if user stop generation mid-way, we must add EOT to finish model's last response + if (need_insert_eot && format_chat) { + llama_token eot = llama_token_eot(model); + embd_inp.push_back(eot == -1 ? llama_token_eos(model) : eot); + need_insert_eot = false; + } + embd_inp.insert(embd_inp.end(), line_pfx.begin(), line_pfx.end()); embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end()); embd_inp.insert(embd_inp.end(), line_sfx.begin(), line_sfx.end()); diff --git a/examples/passkey/README.md b/examples/passkey/README.md index a48a6283a..2b8e910f9 100644 --- a/examples/passkey/README.md +++ b/examples/passkey/README.md @@ -1,5 +1,8 @@ # llama.cpp/example/passkey +A passkey retrieval task is an evaluation method used to measure a language +models ability to recall information from long contexts. + See the following PRs for more info: - https://github.com/ggerganov/llama.cpp/pull/3856 diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index 7fbc06f72..3269dfe19 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -1992,6 +1992,12 @@ int main(int argc, char ** argv) { params.n_batch = std::min(params.n_batch, n_kv); } else { params.n_batch = std::min(params.n_batch, params.n_ctx); + if (params.kl_divergence) { + params.n_parallel = 1; + } else { + // ensure there's at least enough seq_ids for HellaSwag + params.n_parallel = std::max(4, params.n_parallel); + } } if (params.ppl_stride > 0) { @@ -2016,9 +2022,6 @@ int main(int argc, char ** argv) { llama_model * model; llama_context * ctx; - // ensure there's at least enough seq_ids for HellaSwag - params.n_parallel = std::max(4, params.n_parallel); - // load the model and apply lora adapter, if any std::tie(model, ctx) = llama_init_from_gpt_params(params); if (model == NULL) { diff --git a/examples/pydantic-models-to-grammar-examples.py b/examples/pydantic_models_to_grammar_examples.py similarity index 100% rename from examples/pydantic-models-to-grammar-examples.py rename to examples/pydantic_models_to_grammar_examples.py diff --git a/examples/regex-to-grammar.py b/examples/regex_to_grammar.py similarity index 100% rename from examples/regex-to-grammar.py rename to examples/regex_to_grammar.py diff --git a/examples/server/README.md b/examples/server/README.md index 4fab006bb..aa4cbbe63 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -375,7 +375,7 @@ Notice that each `probs` is an array of length `n_probs`. - `default_generation_settings` - the default generation settings for the `/completion` endpoint, which has the same fields as the `generation_settings` response object from the `/completion` endpoint. - `total_slots` - the total number of slots for process requests (defined by `--parallel` option) -- **POST** `/v1/chat/completions`: OpenAI-compatible Chat Completions API. Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only model with [supported chat template](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) can be used optimally with this endpoint. By default, ChatML template will be used. +- **POST** `/v1/chat/completions`: OpenAI-compatible Chat Completions API. Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only models with a [supported chat template](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) can be used optimally with this endpoint. By default, the ChatML template will be used. *Options:* diff --git a/examples/server/tests/features/passkey.feature b/examples/server/tests/features/passkey.feature index 1bde7aab8..6a5a84e6a 100644 --- a/examples/server/tests/features/passkey.feature +++ b/examples/server/tests/features/passkey.feature @@ -52,4 +52,3 @@ Feature: Passkey / Self-extend with context shift #| TheBloke/Llama-2-7B-GGUF | llama-2-7b.Q2_K.gguf | 4096 | 3 | 16384 | 512 | 4 | 512 | 500 | 300 | 1234 | 5 | 1234 | #| TheBloke/Mixtral-8x7B-v0.1-GGUF | mixtral-8x7b-v0.1.Q2_K.gguf | 32768 | 2 | 16384 | 512 | 4 | 512 | 500 | 100 | 0987 | 5 | 0 # 987 | - diff --git a/examples/server/themes/buttons-top/index.html b/examples/server/themes/buttons-top/index.html index 6af30d307..8334bcde5 100644 --- a/examples/server/themes/buttons-top/index.html +++ b/examples/server/themes/buttons-top/index.html @@ -1054,4 +1054,3 @@ - diff --git a/examples/server/themes/wild/index.html b/examples/server/themes/wild/index.html index 772e716cd..8361c5774 100644 --- a/examples/server/themes/wild/index.html +++ b/examples/server/themes/wild/index.html @@ -1058,4 +1058,3 @@ - diff --git a/examples/server-embd.py b/examples/server_embd.py similarity index 99% rename from examples/server-embd.py rename to examples/server_embd.py index 118e04271..a9a36a44c 100644 --- a/examples/server-embd.py +++ b/examples/server_embd.py @@ -31,4 +31,3 @@ for i in range(n-1): embedding2 = np.array(result[j]) similarity = np.dot(embedding1, embedding2) / (np.linalg.norm(embedding1) * np.linalg.norm(embedding2)) print(f"Similarity between {i} and {j}: {similarity:.2f}") - diff --git a/examples/sycl/run-llama2.sh b/examples/sycl/run-llama2.sh index da0e4aaba..111366fb0 100755 --- a/examples/sycl/run-llama2.sh +++ b/examples/sycl/run-llama2.sh @@ -34,4 +34,3 @@ fi #use multiple GPUs with same max compute units #ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0 - diff --git a/examples/sycl/win-build-sycl.bat b/examples/sycl/win-build-sycl.bat index cdae5a528..17dd1ff5c 100644 --- a/examples/sycl/win-build-sycl.bat +++ b/examples/sycl/win-build-sycl.bat @@ -31,4 +31,3 @@ exit /B 0 :ERROR echo comomand error: %errorlevel% exit /B %errorlevel% - diff --git a/examples/sycl/win-run-llama2.bat b/examples/sycl/win-run-llama2.bat index 1d4d7d2cd..f0385cdf0 100644 --- a/examples/sycl/win-run-llama2.bat +++ b/examples/sycl/win-run-llama2.bat @@ -7,5 +7,3 @@ set INPUT2="Building a website can be done in 10 simple steps:\nStep 1:" .\build\bin\main.exe -m models\llama-2-7b.Q4_0.gguf -p %INPUT2% -n 400 -e -ngl 33 -s 0 - - diff --git a/examples/tokenize/tokenize.cpp b/examples/tokenize/tokenize.cpp index 54c9834af..0180c87d8 100644 --- a/examples/tokenize/tokenize.cpp +++ b/examples/tokenize/tokenize.cpp @@ -30,6 +30,7 @@ static void print_usage_information(const char * argv0, FILE * stream) { fprintf(stream, " --stdin read prompt from standard input.\n"); fprintf(stream, " --no-bos do not ever add a BOS token to the prompt, even if normally the model uses a BOS token.\n"); fprintf(stream, " --log-disable disable logs. Makes stderr quiet when loading the model.\n"); + fprintf(stream, " --show-count print the total number of tokens.\n"); } static void llama_log_callback_null(ggml_log_level level, const char * text, void * user_data) { @@ -195,6 +196,7 @@ int main(int raw_argc, char ** raw_argv) { bool printing_ids = false; bool no_bos = false; bool disable_logging = false; + bool show_token_count = false; const char * model_path = NULL; const char * prompt_path = NULL; const char * prompt_arg = NULL; @@ -249,6 +251,9 @@ int main(int raw_argc, char ** raw_argv) { else if (arg == "--log-disable") { disable_logging = true; } + else if (arg == "--show-count") { + show_token_count = true; + } else { fprintf(stderr, "Error: unknown option '%s'\n", argv[iarg].c_str()); return 1; @@ -384,6 +389,9 @@ int main(int raw_argc, char ** raw_argv) { printf("]\n"); } + if (show_token_count) { + printf("Total number of tokens: %ld\n", tokens.size()); + } // silence valgrind llama_free(ctx); llama_free_model(model); diff --git a/examples/train-text-from-scratch/convert-train-checkpoint-to-gguf.py b/examples/train-text-from-scratch/convert_train_checkpoint_to_gguf.py similarity index 100% rename from examples/train-text-from-scratch/convert-train-checkpoint-to-gguf.py rename to examples/train-text-from-scratch/convert_train_checkpoint_to_gguf.py diff --git a/ggml/include/ggml-metal.h b/ggml/include/ggml-metal.h index a5c542189..3c1de33a1 100644 --- a/ggml/include/ggml-metal.h +++ b/ggml/include/ggml-metal.h @@ -63,4 +63,3 @@ GGML_API void ggml_backend_metal_capture_next_compute(ggml_backend_t backend); #ifdef __cplusplus } #endif - diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh index 472f4ace1..4ff06b871 100644 --- a/ggml/src/ggml-cuda/common.cuh +++ b/ggml/src/ggml-cuda/common.cuh @@ -227,6 +227,10 @@ typedef float2 dfloat2; #define RDNA2 #endif +#if defined(__gfx1010__) || defined(__gfx1012__) +#define RDNA1 +#endif + #ifndef __has_builtin #define __has_builtin(x) 0 #endif diff --git a/ggml/src/ggml-cuda/cpy.cu b/ggml/src/ggml-cuda/cpy.cu index 12d741f01..3db57034b 100644 --- a/ggml/src/ggml-cuda/cpy.cu +++ b/ggml/src/ggml-cuda/cpy.cu @@ -487,4 +487,3 @@ void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) { GGML_ASSERT(false); } } - diff --git a/ggml/src/ggml-cuda/fattn-common.cuh b/ggml/src/ggml-cuda/fattn-common.cuh index 650780fd2..f24312dd0 100644 --- a/ggml/src/ggml-cuda/fattn-common.cuh +++ b/ggml/src/ggml-cuda/fattn-common.cuh @@ -68,7 +68,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_0( const int iqs4 = k_KQ % QI4_0; const int shift = k_KQ & (QI8_1/2); - const int v = (get_int_from_uint8(K_q4_0[ib].qs, iqs4) >> shift) & 0x0F0F0F0F; + const int v = (get_int_b2(K_q4_0[ib].qs, iqs4) >> shift) & 0x0F0F0F0F; const int u = Q_q8[k_KQ_0/WARP_SIZE]; const int sumi = ggml_cuda_dp4a(v, u, 0); @@ -108,7 +108,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_1( const int iqs4 = k_KQ % QI4_1; const int shift = k_KQ & (QI8_1/2); - const int v = (get_int_from_uint8_aligned(K_q4_1[ib].qs, iqs4) >> shift) & 0x0F0F0F0F; + const int v = (get_int_b4(K_q4_1[ib].qs, iqs4) >> shift) & 0x0F0F0F0F; const int u = Q_q8[k_KQ_0/WARP_SIZE]; const int sumi = ggml_cuda_dp4a(v, u, 0); @@ -153,8 +153,8 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_0( const int iqs8 = k_KQ % QI8_1; const int shift = k_KQ & (QI8_1/2); - int v = (get_int_from_uint8(K_q5_0[ib].qs, iqs4) >> shift) & 0x0F0F0F0F; - const int vh = get_int_from_uint8(K_q5_0[ib].qh, 0) >> (iqs8 * QI5_0); + int v = (get_int_b2(K_q5_0[ib].qs, iqs4) >> shift) & 0x0F0F0F0F; + const int vh = get_int_b2(K_q5_0[ib].qh, 0) >> (iqs8 * QI5_0); v |= (vh << 4) & 0x00000010; // 0 -> 4 v |= (vh << 11) & 0x00001000; // 1 -> 12 v |= (vh << 18) & 0x00100000; // 2 -> 20 @@ -200,8 +200,8 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_1( const int iqs8 = k_KQ % QI8_1; const int shift = k_KQ & (QI8_1/2); - int v = (get_int_from_uint8(K_q5_1[ib].qs, iqs4) >> shift) & 0x0F0F0F0F; - const int vh = get_int_from_uint8(K_q5_1[ib].qh, 0) >> (iqs8 * QI5_1); + int v = (get_int_b2(K_q5_1[ib].qs, iqs4) >> shift) & 0x0F0F0F0F; + const int vh = get_int_b2(K_q5_1[ib].qh, 0) >> (iqs8 * QI5_1); v |= (vh << 4) & 0x00000010; // 0 -> 4 v |= (vh << 11) & 0x00001000; // 1 -> 12 v |= (vh << 18) & 0x00100000; // 2 -> 20 @@ -249,7 +249,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q8_0( const int ib = k_KQ / QI8_0; const int iqs = k_KQ % QI8_0; - const int v = get_int_from_int8(K_q8_0[ib].qs, iqs); + const int v = get_int_b2(K_q8_0[ib].qs, iqs); T Q_d; if (std::is_same::value) { @@ -408,7 +408,7 @@ static __device__ __forceinline__ T dequantize_1_q5_0(const void * __restrict__ const T d = x[ib].d; const int ql0 = x[ib].qs[iqs]; - const int qh0 = get_int_from_uint8(x[ib].qh, 0); + const int qh0 = get_int_b2(x[ib].qh, 0); const int ql = ((ql0 >> (4*shift)) & 0x0F); const int qh = ((qh0 >> idq) << 4) & 0x10; const int q = (ql | qh) - 16; @@ -433,7 +433,7 @@ static __device__ __forceinline__ T dequantize_1_q5_1(const void * __restrict__ const half2 dm = x[ib].dm; const int ql0 = x[ib].qs[iqs]; - const int qh0 = get_int_from_uint8_aligned(x[ib].qh, 0); + const int qh0 = get_int_b4(x[ib].qh, 0); const int ql = ((ql0 >> (4*shift)) & 0x0F); const int qh = ((qh0 >> idq) << 4) & 0x10; const int q = (ql | qh); diff --git a/ggml/src/ggml-cuda/mmq.cu b/ggml/src/ggml-cuda/mmq.cu index 7d951035a..8d6e93e9d 100644 --- a/ggml/src/ggml-cuda/mmq.cu +++ b/ggml/src/ggml-cuda/mmq.cu @@ -59,6 +59,12 @@ void ggml_cuda_op_mul_mat_q( case GGML_TYPE_Q6_K: mul_mat_q_case(ctx, args, stream); break; + case GGML_TYPE_IQ4_XS: + mul_mat_q_case(ctx, args, stream); + break; + case GGML_TYPE_IQ4_NL: + mul_mat_q_case(ctx, args, stream); + break; default: GGML_ASSERT(false); break; @@ -89,6 +95,8 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) { case GGML_TYPE_Q4_K: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: + case GGML_TYPE_IQ4_XS: + case GGML_TYPE_IQ4_NL: mmq_supported = true; break; default: diff --git a/ggml/src/ggml-cuda/mmq.cuh b/ggml/src/ggml-cuda/mmq.cuh index 0b357fc0a..9ba1ed3d9 100644 --- a/ggml/src/ggml-cuda/mmq.cuh +++ b/ggml/src/ggml-cuda/mmq.cuh @@ -61,12 +61,16 @@ static constexpr __device__ int get_mmq_x_max_device() { } static constexpr int get_mmq_y_host(const int cc) { - return int8_mma_available(cc) || cc >= CC_VOLTA ? 128 : 64; + return cc >= CC_OFFSET_AMD ? (cc == CC_RDNA1 ? 64 : 128) : (cc >= CC_VOLTA ? 128 : 64); } static constexpr __device__ int get_mmq_y_device() { #if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA1) + return 64; +#else return 128; +#endif // defined RDNA1 #else #if __CUDA_ARCH__ >= CC_VOLTA return 128; @@ -89,15 +93,17 @@ static constexpr __device__ int get_mmq_y_device() { static constexpr __host__ __device__ tile_x_sizes mmq_get_dp4a_tile_x_sizes(ggml_type type, int mmq_y) { return type == GGML_TYPE_Q4_0 ? MMQ_DP4A_TXS_Q4_0 : - type == GGML_TYPE_Q4_1 ? MMQ_DP4A_TXS_Q4_1 : - type == GGML_TYPE_Q5_0 ? MMQ_DP4A_TXS_Q5_0 : - type == GGML_TYPE_Q5_1 ? MMQ_DP4A_TXS_Q5_1 : - type == GGML_TYPE_Q8_0 ? MMQ_DP4A_TXS_Q8_0 : - type == GGML_TYPE_Q2_K ? MMQ_DP4A_TXS_Q2_K : - type == GGML_TYPE_Q3_K ? MMQ_DP4A_TXS_Q3_K : - type == GGML_TYPE_Q4_K ? MMQ_DP4A_TXS_Q4_K : - type == GGML_TYPE_Q5_K ? MMQ_DP4A_TXS_Q5_K : - type == GGML_TYPE_Q6_K ? MMQ_DP4A_TXS_Q6_K : + type == GGML_TYPE_Q4_1 ? MMQ_DP4A_TXS_Q4_1 : + type == GGML_TYPE_Q5_0 ? MMQ_DP4A_TXS_Q5_0 : + type == GGML_TYPE_Q5_1 ? MMQ_DP4A_TXS_Q5_1 : + type == GGML_TYPE_Q8_0 ? MMQ_DP4A_TXS_Q8_0 : + type == GGML_TYPE_Q2_K ? MMQ_DP4A_TXS_Q2_K : + type == GGML_TYPE_Q3_K ? MMQ_DP4A_TXS_Q3_K : + type == GGML_TYPE_Q4_K ? MMQ_DP4A_TXS_Q4_K : + type == GGML_TYPE_Q5_K ? MMQ_DP4A_TXS_Q5_K : + type == GGML_TYPE_Q6_K ? MMQ_DP4A_TXS_Q6_K : + type == GGML_TYPE_IQ4_XS ? MMQ_DP4A_TXS_Q5_0 : + type == GGML_TYPE_IQ4_NL ? MMQ_DP4A_TXS_Q5_0 : tile_x_sizes{0, 0, 0}; } @@ -125,15 +131,17 @@ static_assert(MMQ_MMA_TILE_X_K_Q6_K % 8 == 4, "Wrong padding."); static constexpr __host__ __device__ int mmq_get_mma_tile_x_k(ggml_type type) { return type == GGML_TYPE_Q4_0 ? MMQ_MMA_TILE_X_K_Q4_0 : - type == GGML_TYPE_Q4_1 ? MMQ_MMA_TILE_X_K_Q4_1 : - type == GGML_TYPE_Q5_0 ? MMQ_MMA_TILE_X_K_Q5_0 : - type == GGML_TYPE_Q5_1 ? MMQ_MMA_TILE_X_K_Q5_1 : - type == GGML_TYPE_Q8_0 ? MMQ_MMA_TILE_X_K_Q8_0 : - type == GGML_TYPE_Q2_K ? MMQ_MMA_TILE_X_K_Q2_K : - type == GGML_TYPE_Q3_K ? MMQ_MMA_TILE_X_K_Q3_K : - type == GGML_TYPE_Q4_K ? MMQ_MMA_TILE_X_K_Q4_K : - type == GGML_TYPE_Q5_K ? MMQ_MMA_TILE_X_K_Q5_K : - type == GGML_TYPE_Q6_K ? MMQ_MMA_TILE_X_K_Q6_K : + type == GGML_TYPE_Q4_1 ? MMQ_MMA_TILE_X_K_Q4_1 : + type == GGML_TYPE_Q5_0 ? MMQ_MMA_TILE_X_K_Q5_0 : + type == GGML_TYPE_Q5_1 ? MMQ_MMA_TILE_X_K_Q5_1 : + type == GGML_TYPE_Q8_0 ? MMQ_MMA_TILE_X_K_Q8_0 : + type == GGML_TYPE_Q2_K ? MMQ_MMA_TILE_X_K_Q2_K : + type == GGML_TYPE_Q3_K ? MMQ_MMA_TILE_X_K_Q3_K : + type == GGML_TYPE_Q4_K ? MMQ_MMA_TILE_X_K_Q4_K : + type == GGML_TYPE_Q5_K ? MMQ_MMA_TILE_X_K_Q5_K : + type == GGML_TYPE_Q6_K ? MMQ_MMA_TILE_X_K_Q6_K : + type == GGML_TYPE_IQ4_XS ? MMQ_MMA_TILE_X_K_Q5_0 : + type == GGML_TYPE_IQ4_NL ? MMQ_MMA_TILE_X_K_Q5_0 : 0; } @@ -182,9 +190,9 @@ template static __device__ __forceinlin const block_q4_0 * bxi = (const block_q4_0 *) x + kbx0 + i*stride + kbx; #ifdef INT8_MMA_AVAILABLE - x_qs[i*MMQ_MMA_TILE_X_K_Q4_0 + threadIdx.x] = get_int_from_uint8(bxi->qs, kqsx); + x_qs[i*MMQ_MMA_TILE_X_K_Q4_0 + threadIdx.x] = get_int_b2(bxi->qs, kqsx); #else - x_qs[i*(WARP_SIZE + 1) + threadIdx.x] = get_int_from_uint8(bxi->qs, kqsx); + x_qs[i*(WARP_SIZE + 1) + threadIdx.x] = get_int_b2(bxi->qs, kqsx); #endif // INT8_MMA_AVAILABLE } @@ -345,9 +353,9 @@ template static __device__ __forceinlin const block_q4_1 * bxi = (const block_q4_1 *) x + kbx0 + i*stride + kbx; #ifdef INT8_MMA_AVAILABLE - x_qs[i*MMQ_MMA_TILE_X_K_Q4_1 + threadIdx.x] = get_int_from_uint8_aligned(bxi->qs, kqsx); + x_qs[i*MMQ_MMA_TILE_X_K_Q4_1 + threadIdx.x] = get_int_b4(bxi->qs, kqsx); #else - x_qs[i*(WARP_SIZE + 1) + threadIdx.x] = get_int_from_uint8_aligned(bxi->qs, kqsx); + x_qs[i*(WARP_SIZE + 1) + threadIdx.x] = get_int_b4(bxi->qs, kqsx); #endif // INT8_MMA_AVAILABLE } @@ -506,8 +514,8 @@ template static __device__ __forceinlin const block_q5_0 * bxi = (const block_q5_0 *) x + kbx0 + i*stride + kbx; - const int ql = get_int_from_uint8(bxi->qs, kqsx); - const int qh = get_int_from_uint8(bxi->qh, 0) >> (4 * (threadIdx.x % QI5_0)); + const int ql = get_int_b2(bxi->qs, kqsx); + const int qh = get_int_b2(bxi->qh, 0) >> (4 * (threadIdx.x % QI5_0)); int qs0 = (ql >> 0) & 0x0F0F0F0F; qs0 |= (qh << 4) & 0x00000010; // 0 -> 4 @@ -671,8 +679,8 @@ template static __device__ __forceinlin const block_q5_1 * bxi = (const block_q5_1 *) x + kbx0 + i*stride + kbx; - const int ql = get_int_from_uint8_aligned(bxi->qs, kqsx); - const int qh = get_int_from_uint8_aligned(bxi->qh, 0) >> (4 * (threadIdx.x % QI5_1)); + const int ql = get_int_b4(bxi->qs, kqsx); + const int qh = get_int_b4(bxi->qh, 0) >> (4 * (threadIdx.x % QI5_1)); int qs0 = (ql >> 0) & 0x0F0F0F0F; qs0 |= (qh << 4) & 0x00000010; // 0 -> 4 @@ -836,9 +844,9 @@ template static __device__ __forceinlin const block_q8_0 * bxi = (const block_q8_0 *) x + kbx0 + i*stride + kbx; #ifdef INT8_MMA_AVAILABLE - x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + threadIdx.x] = get_int_from_int8(bxi->qs, kqsx); + x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + threadIdx.x] = get_int_b2(bxi->qs, kqsx); #else - x_qs[i*(WARP_SIZE + 1) + threadIdx.x] = get_int_from_int8(bxi->qs, kqsx); + x_qs[i*(WARP_SIZE + 1) + threadIdx.x] = get_int_b2(bxi->qs, kqsx); #endif // INT8_MMA_AVAILABLE } @@ -981,7 +989,7 @@ template static __device__ __forceinlin const block_q2_K * bxi = (const block_q2_K *) x + kbx0 + i*stride + kbx; - const int x_ql_0 = get_int_from_uint8(bxi->qs, kqsx); + const int x_ql_0 = get_int_b2(bxi->qs, kqsx); #pragma unroll for (int l = 0; l < QR2_K; ++l) { @@ -1163,8 +1171,8 @@ template static __device__ __forceinlin const block_q3_K * bxi = (const block_q3_K *) x + kbx0 + i*stride + kbx; - const int x_ql_0 = get_int_from_uint8(bxi->qs, kqsx); - const int x_qh_0 = get_int_from_uint8(bxi->hmask, kqsx % (QI3_K/2)) >> (4 * (kqsx / (QI3_K/2))); + const int x_ql_0 = get_int_b2(bxi->qs, kqsx); + const int x_qh_0 = get_int_b2(bxi->hmask, kqsx % (QI3_K/2)) >> (4 * (kqsx / (QI3_K/2))); #pragma unroll for (int l = 0; l < QR3_K; ++l) { @@ -1222,11 +1230,11 @@ template static __device__ __forceinlin const int ksc_low = ksc % (QI3_K/8); const int shift_low = 4 * (ksc / (QI3_K/8)); - const int sc_low = (get_int_from_uint8(bxi->scales, ksc_low) >> shift_low) & 0x0F0F0F0F; + const int sc_low = (get_int_b2(bxi->scales, ksc_low) >> shift_low) & 0x0F0F0F0F; const int ksc_high = QI3_K/8; const int shift_high = 2 * ksc; - const int sc_high = ((get_int_from_uint8(bxi->scales, ksc_high) >> shift_high) << 4) & 0x30303030; + const int sc_high = ((get_int_b2(bxi->scales, ksc_high) >> shift_high) << 4) & 0x30303030; const int sc = __vsubss4(sc_low | sc_high, 0x20202020); @@ -1390,9 +1398,9 @@ template static __device__ __forceinlin const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride + kbx; #ifdef INT8_MMA_AVAILABLE - x_qs[i*MMQ_MMA_TILE_X_K_Q4_K + threadIdx.x] = get_int_from_uint8_aligned(bxi->qs, kqsx); + x_qs[i*MMQ_MMA_TILE_X_K_Q4_K + threadIdx.x] = get_int_b4(bxi->qs, kqsx); #else - x_qs[i*(WARP_SIZE + 1) + threadIdx.x] = get_int_from_uint8_aligned(bxi->qs, kqsx); + x_qs[i*(WARP_SIZE + 1) + threadIdx.x] = get_int_b4(bxi->qs, kqsx); #endif // INT8_MMA_AVAILABLE } @@ -1607,11 +1615,11 @@ template static __device__ __forceinlin const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride + kbx; const int ky = QR5_K*kqsx; - const int ql = get_int_from_uint8_aligned(bxi->qs, kqsx); + const int ql = get_int_b4(bxi->qs, kqsx); const int ql0 = (ql >> 0) & 0x0F0F0F0F; const int ql1 = (ql >> 4) & 0x0F0F0F0F; - const int qh = get_int_from_uint8_aligned(bxi->qh, kqsx % (QI5_K/4)); + const int qh = get_int_b4(bxi->qh, kqsx % (QI5_K/4)); const int qh0 = ((qh >> (2 * (kqsx / (QI5_K/4)) + 0)) << 4) & 0x10101010; const int qh1 = ((qh >> (2 * (kqsx / (QI5_K/4)) + 1)) << 4) & 0x10101010; @@ -1829,11 +1837,11 @@ template static __device__ __forceinlin const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride + kbx; const int ky = QR6_K*kqsx; - const int ql = get_int_from_uint8(bxi->ql, kqsx); + const int ql = get_int_b2(bxi->ql, kqsx); const int ql0 = (ql >> 0) & 0x0F0F0F0F; const int ql1 = (ql >> 4) & 0x0F0F0F0F; - const int qh = get_int_from_uint8(bxi->qh, (QI6_K/4) * (kqsx / (QI6_K/2)) + kqsx % (QI6_K/4)); + const int qh = get_int_b2(bxi->qh, (QI6_K/4) * (kqsx / (QI6_K/2)) + kqsx % (QI6_K/4)); const int qh0 = ((qh >> (2 * ((kqsx % (QI6_K/2)) / (QI6_K/4)))) << 4) & 0x30303030; const int qh1 = (qh >> (2 * ((kqsx % (QI6_K/2)) / (QI6_K/4)))) & 0x30303030; @@ -1880,9 +1888,9 @@ template static __device__ __forceinlin const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride + (threadIdx.x % (WARP_SIZE/8)) / 4; #ifdef INT8_MMA_AVAILABLE - x_sc[i*MMQ_MMA_TILE_X_K_Q6_K + threadIdx.x % (WARP_SIZE/8)] = get_int_from_int8(bxi->scales, threadIdx.x % (QI6_K/8)); + x_sc[i*MMQ_MMA_TILE_X_K_Q6_K + threadIdx.x % (WARP_SIZE/8)] = get_int_b2(bxi->scales, threadIdx.x % (QI6_K/8)); #else - x_sc[i*(WARP_SIZE/8) + i/8 + threadIdx.x % (WARP_SIZE/8)] = get_int_from_int8(bxi->scales, threadIdx.x % (QI6_K/8)); + x_sc[i*(WARP_SIZE/8) + i/8 + threadIdx.x % (WARP_SIZE/8)] = get_int_b2(bxi->scales, threadIdx.x % (QI6_K/8)); #endif // INT8_MMA_AVAILABLE } } @@ -2015,6 +2023,124 @@ static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma( #endif // INT8_MMA_AVAILABLE } +template static __device__ __forceinline__ void load_tiles_iq4_nl( + const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) { + +#ifdef INT8_MMA_AVAILABLE + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + WARP_SIZE*2); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ4_NL, mmq_y); + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + txs.qs); +#endif // INT8_MMA_AVAILABLE + + const int kbx = threadIdx.x / QI4_NL; + const int kqsx = threadIdx.x % QI4_NL; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { + int i = i0 + threadIdx.y; + + if (need_check) { + i = min(i, i_max); + } + + const block_iq4_nl * bxi = (const block_iq4_nl *) x + kbx0 + i*stride + kbx; + + const int aux_q4 = get_int_b2(bxi->qs, kqsx); + const int2 v = get_int_from_table_16(aux_q4); + const int k0 = 8 * (threadIdx.x / 4) + threadIdx.x % 4; +#ifdef INT8_MMA_AVAILABLE + x_qs[i*MMQ_MMA_TILE_X_K_Q5_0 + k0 + 0] = v.x; + x_qs[i*MMQ_MMA_TILE_X_K_Q5_0 + k0 + 4] = v.y; +#else + x_qs[i*(2*WARP_SIZE + 1) + k0 + 0] = v.x; + x_qs[i*(2*WARP_SIZE + 1) + k0 + 4] = v.y; +#endif // INT8_MMA_AVAILABLE + } + + const int blocks_per_tile_x_row = WARP_SIZE / QI4_NL; + const int kbxd = threadIdx.x % blocks_per_tile_x_row; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_NL) { + int i = i0 + threadIdx.y * QI4_NL + threadIdx.x / blocks_per_tile_x_row; + + if (need_check) { + i = min(i, i_max); + } + + const block_iq4_nl * bxi = (const block_iq4_nl *) x + kbx0 + i*stride + kbxd; + +#ifdef INT8_MMA_AVAILABLE + x_df[i*MMQ_MMA_TILE_X_K_Q5_0 + kbxd] = __half2float(bxi->d); +#else + x_df[i*(WARP_SIZE/4) + i/4 + kbxd] = __half2float(bxi->d); +#endif // INT8_MMA_AVAILABLE + } +} + +template static __device__ __forceinline__ void load_tiles_iq4_xs( + const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) { + +#ifdef INT8_MMA_AVAILABLE + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + WARP_SIZE*2); +#else + constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ4_XS, mmq_y); + int * x_qs = (int *) x_tile; + float * x_df = (float *) (x_qs + txs.qs); +#endif // INT8_MMA_AVAILABLE + + const int kbx = 0; // threadIdx.x / QI4_XS + const int kqsx = threadIdx.x; // threadIdx.x % QI4_XS + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { + int i = i0 + threadIdx.y; + + if (need_check) { + i = min(i, i_max); + } + + const block_iq4_xs * bxi = (const block_iq4_xs *) x + kbx0 + i*stride + kbx; + + const int aux_q4 = get_int_b4(bxi->qs, kqsx); + const int2 v = get_int_from_table_16(aux_q4); + const int k0 = 8 * (threadIdx.x / 4) + threadIdx.x % 4; +#ifdef INT8_MMA_AVAILABLE + x_qs[i*MMQ_MMA_TILE_X_K_Q5_0 + k0 + 0] = v.x; + x_qs[i*MMQ_MMA_TILE_X_K_Q5_0 + k0 + 4] = v.y; +#else + x_qs[i*(2*WARP_SIZE + 1) + k0 + 0] = v.x; + x_qs[i*(2*WARP_SIZE + 1) + k0 + 4] = v.y; +#endif // INT8_MMA_AVAILABLE + } + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 4) { + int i = i0 + threadIdx.y * 4 + threadIdx.x / (WARP_SIZE/4); + + if (need_check) { + i = min(i, i_max); + } + + const block_iq4_xs * bxi = (const block_iq4_xs *) x + kbx0 + i*stride; + + const float d = __half2float(bxi->d); + + const int ls = ((bxi->scales_l[(threadIdx.x % 8)/2] >> (4*(threadIdx.x % 2))) & 0x0F) + | (((bxi->scales_h >> (2*(threadIdx.x % 8))) & 0x03) << 4); + +#ifdef INT8_MMA_AVAILABLE + x_df[i*MMQ_MMA_TILE_X_K_Q5_0 + threadIdx.x % 8] = d * (ls - 32); +#else + x_df[i*(WARP_SIZE/4) + i/4 + threadIdx.x % 8] = d * (ls - 32); +#endif // INT8_MMA_AVAILABLE + } +} + template static __device__ __forceinline__ void mmq_write_back_dp4a( const float * __restrict__ sum, float * __restrict__ dst, const int & stride, const int & i_max, const int & j_max) { @@ -2164,6 +2290,22 @@ struct mmq_type_traits { static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q6_K_q8_1_dp4a; }; +template +struct mmq_type_traits { + static constexpr int vdr = VDR_IQ4_NL_Q8_1_MMQ; + static constexpr load_tiles_mmq_t load_tiles = load_tiles_iq4_nl; + static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q5_0_q8_1_mma; + static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q5_0_q8_1_dp4a; +}; + +template +struct mmq_type_traits { + static constexpr int vdr = VDR_IQ4_XS_Q8_1_MMQ; + static constexpr load_tiles_mmq_t load_tiles = load_tiles_iq4_xs; + static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q5_0_q8_1_mma; + static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q5_0_q8_1_dp4a; +}; + static bool mmq_need_sum(const ggml_type type_x) { switch (type_x) { case GGML_TYPE_Q4_0: @@ -2181,6 +2323,8 @@ static bool mmq_need_sum(const ggml_type type_x) { case GGML_TYPE_Q5_K: return true; case GGML_TYPE_Q6_K: + case GGML_TYPE_IQ4_XS: + case GGML_TYPE_IQ4_NL: return false; default: GGML_ASSERT(false); @@ -2302,8 +2446,11 @@ static __global__ void mul_mat_q( const int nty = (ne01 + mmq_y - 1) / mmq_y; // Number of tiles y // kbc == k block continuous, current index in continuous ijk space. - int64_t kbc = GGML_PAD((int64_t) blockIdx.x *blocks_per_ne00*ntx*nty / gridDim.x, blocks_per_warp); - const int64_t kbc_stop = GGML_PAD((int64_t)(blockIdx.x + 1)*blocks_per_ne00*ntx*nty / gridDim.x, blocks_per_warp); + int64_t kbc = (int64_t) blockIdx.x *blocks_per_ne00*ntx*nty / gridDim.x; + int64_t kbc_stop = (int64_t)(blockIdx.x + 1)*blocks_per_ne00*ntx*nty / gridDim.x; + + kbc -= (kbc % blocks_per_ne00) % blocks_per_warp; + kbc_stop -= (kbc_stop % blocks_per_ne00) % blocks_per_warp; // kb0 == k index when doing the matrix multiplication for an output tile. int kb0_start = kbc % blocks_per_ne00; @@ -2359,8 +2506,11 @@ static __global__ void mul_mat_q_stream_k_fixup( const int bidx_stop = (blockIdx.y*nty + blockIdx.x + 1) * block_num_mmq / (gridDim.y*gridDim.x) + 1; for (int bidx = bidx_start; bidx < bidx_stop; ++bidx) { - const int64_t kbc = GGML_PAD((int64_t) bidx *blocks_per_ne00*ntx*nty / block_num_mmq, blocks_per_warp); - const int64_t kbc_stop = GGML_PAD((int64_t)(bidx + 1)*blocks_per_ne00*ntx*nty / block_num_mmq, blocks_per_warp); + int64_t kbc = (int64_t) bidx *blocks_per_ne00*ntx*nty / block_num_mmq; + int64_t kbc_stop = (int64_t)(bidx + 1)*blocks_per_ne00*ntx*nty / block_num_mmq; + + kbc -= (kbc % blocks_per_ne00) % blocks_per_warp; + kbc_stop -= (kbc_stop % blocks_per_ne00) % blocks_per_warp; // Skip fixup tile if the MMQ CUDA block never wrote anything to it: if (kbc == kbc_stop || kbc_stop % blocks_per_ne00 == 0) { @@ -2599,6 +2749,8 @@ extern DECL_MMQ_CASE(GGML_TYPE_Q3_K); extern DECL_MMQ_CASE(GGML_TYPE_Q4_K); extern DECL_MMQ_CASE(GGML_TYPE_Q5_K); extern DECL_MMQ_CASE(GGML_TYPE_Q6_K); +extern DECL_MMQ_CASE(GGML_TYPE_IQ4_NL); +extern DECL_MMQ_CASE(GGML_TYPE_IQ4_XS); // ------------------------------------------------------------------------------------------------------------------------- diff --git a/ggml/src/ggml-cuda/template-instances/generate_cu_files.py b/ggml/src/ggml-cuda/template-instances/generate_cu_files.py index ea58d0968..ffeb3c27d 100755 --- a/ggml/src/ggml-cuda/template-instances/generate_cu_files.py +++ b/ggml/src/ggml-cuda/template-instances/generate_cu_files.py @@ -22,7 +22,8 @@ SOURCE_FATTN_WMMA_CASE = "DECL_FATTN_WMMA_F16_CASE({head_size}, {cols_per_block} TYPES_MMQ = [ "GGML_TYPE_Q4_0", "GGML_TYPE_Q4_1", "GGML_TYPE_Q5_0", "GGML_TYPE_Q5_1", "GGML_TYPE_Q8_0", - "GGML_TYPE_Q2_K", "GGML_TYPE_Q3_K", "GGML_TYPE_Q4_K", "GGML_TYPE_Q5_K", "GGML_TYPE_Q6_K" + "GGML_TYPE_Q2_K", "GGML_TYPE_Q3_K", "GGML_TYPE_Q4_K", "GGML_TYPE_Q5_K", "GGML_TYPE_Q6_K", + "GGML_TYPE_IQ4_NL", "GGML_TYPE_IQ4_XS" ] SOURCE_MMQ = """// This file has been autogenerated by generate_cu_files.py, do not edit manually. diff --git a/ggml/src/ggml-cuda/template-instances/mmq-instance-iq4_nl.cu b/ggml/src/ggml-cuda/template-instances/mmq-instance-iq4_nl.cu new file mode 100644 index 000000000..eb02fab00 --- /dev/null +++ b/ggml/src/ggml-cuda/template-instances/mmq-instance-iq4_nl.cu @@ -0,0 +1,5 @@ +// This file has been autogenerated by generate_cu_files.py, do not edit manually. + +#include "../mmq.cuh" + +DECL_MMQ_CASE(GGML_TYPE_IQ4_NL); diff --git a/ggml/src/ggml-cuda/template-instances/mmq-instance-iq4_xs.cu b/ggml/src/ggml-cuda/template-instances/mmq-instance-iq4_xs.cu new file mode 100644 index 000000000..1eb3b7430 --- /dev/null +++ b/ggml/src/ggml-cuda/template-instances/mmq-instance-iq4_xs.cu @@ -0,0 +1,5 @@ +// This file has been autogenerated by generate_cu_files.py, do not edit manually. + +#include "../mmq.cuh" + +DECL_MMQ_CASE(GGML_TYPE_IQ4_XS); diff --git a/ggml/src/ggml-cuda/vecdotq.cuh b/ggml/src/ggml-cuda/vecdotq.cuh index 3f3a87c75..1d510484a 100644 --- a/ggml/src/ggml-cuda/vecdotq.cuh +++ b/ggml/src/ggml-cuda/vecdotq.cuh @@ -1,36 +1,8 @@ #include "common.cuh" #include -static __device__ __forceinline__ int get_int_from_int8(const int8_t * x8, const int & i32) { - const uint16_t * x16 = (const uint16_t *) (x8 + sizeof(int) * i32); // assume at least 2 byte alignment - - int x32 = 0; - x32 |= x16[0] << 0; - x32 |= x16[1] << 16; - - return x32; -} - -static __device__ __forceinline__ int get_int_from_uint8(const uint8_t * x8, const int & i32) { - const uint16_t * x16 = (const uint16_t *) (x8 + sizeof(int) * i32); // assume at least 2 byte alignment - - int x32 = 0; - x32 |= x16[0] << 0; - x32 |= x16[1] << 16; - - return x32; -} - -static __device__ __forceinline__ int get_int_from_int8_aligned(const int8_t * x8, const int & i32) { - return *((const int *) (x8 + sizeof(int) * i32)); // assume at least 4 byte alignment -} - -static __device__ __forceinline__ int get_int_from_uint8_aligned(const uint8_t * x8, const int & i32) { - return *((const int *) (x8 + sizeof(int) * i32)); // assume at least 4 byte alignment -} - static __device__ __forceinline__ int get_int_b2(const void * x, const int & i32) { - const uint16_t * x16 = (const uint16_t *) x; + const uint16_t * x16 = (const uint16_t *) x; // assume at least 2 byte alignment int x32 = x16[2*i32 + 0] << 0; x32 |= x16[2*i32 + 1] << 16; @@ -768,6 +740,7 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1( } #define VDR_IQ2_XXS_Q8_1_MMVQ 2 +#define VDR_IQ2_XXS_Q8_1_MMQ 2 static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) { @@ -802,6 +775,7 @@ static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1( } #define VDR_IQ2_XS_Q8_1_MMVQ 2 +#define VDR_IQ2_XS_Q8_1_MMQ 2 static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) { @@ -840,6 +814,7 @@ static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1( } #define VDR_IQ2_S_Q8_1_MMVQ 2 +#define VDR_IQ2_S_Q8_1_MMQ 2 static __device__ __forceinline__ float vec_dot_iq2_s_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) { @@ -887,6 +862,7 @@ static __device__ __forceinline__ float vec_dot_iq2_s_q8_1( } #define VDR_IQ3_XXS_Q8_1_MMVQ 2 +#define VDR_IQ3_XXS_Q8_1_MMQ 2 static __device__ __forceinline__ float vec_dot_iq3_xxs_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) { @@ -921,6 +897,7 @@ static __device__ __forceinline__ float vec_dot_iq3_xxs_q8_1( } #define VDR_IQ3_S_Q8_1_MMVQ 2 +#define VDR_IQ3_S_Q8_1_MMQ 2 // TODO: don't use lookup table for signs static __device__ __forceinline__ float vec_dot_iq3_s_q8_1( @@ -962,6 +939,9 @@ static __device__ __forceinline__ float vec_dot_iq3_s_q8_1( return d * sumi; } +#define VDR_IQ1_S_Q8_1_MMVQ 1 +#define VDR_IQ1_S_Q8_1_MMQ 1 + static __device__ __forceinline__ float vec_dot_iq1_s_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) { const block_iq1_s * bq1 = (const block_iq1_s *) vbq + kbx; @@ -992,6 +972,9 @@ static __device__ __forceinline__ float vec_dot_iq1_s_q8_1( return d1q * (ds.x*sumi + ds.y*delta); } +#define VDR_IQ1_M_Q8_1_MMVQ 1 +#define VDR_IQ1_M_Q8_1_MMQ 1 + static __device__ __forceinline__ float vec_dot_iq1_m_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) { @@ -1051,6 +1034,7 @@ static __device__ __forceinline__ int2 get_int_from_table_16(const int & q4) { } #define VDR_IQ4_NL_Q8_1_MMVQ 2 +#define VDR_IQ4_NL_Q8_1_MMQ 4 static __device__ __forceinline__ float vec_dot_iq4_nl_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) { @@ -1074,6 +1058,7 @@ static __device__ __forceinline__ float vec_dot_iq4_nl_q8_1( } #define VDR_IQ4_XS_Q8_1_MMVQ 4 +#define VDR_IQ4_XS_Q8_1_MMQ 4 static __device__ __forceinline__ float vec_dot_iq4_xs_q8_1( const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) { diff --git a/ggml/src/ggml-metal.metal b/ggml/src/ggml-metal.metal index e2796fd60..c3503479b 100644 --- a/ggml/src/ggml-metal.metal +++ b/ggml/src/ggml-metal.metal @@ -6537,4 +6537,3 @@ template [[host_name("kernel_mul_mv_id_iq3_s_f32")]] kernel kernel_mul_mv_id_t template [[host_name("kernel_mul_mv_id_iq2_s_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>; template [[host_name("kernel_mul_mv_id_iq4_nl_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>; template [[host_name("kernel_mul_mv_id_iq4_xs_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>; - diff --git a/ggml/src/ggml-quants.h b/ggml/src/ggml-quants.h index 4d436a8f0..30983b872 100644 --- a/ggml/src/ggml-quants.h +++ b/ggml/src/ggml-quants.h @@ -130,4 +130,3 @@ void iq3xs_free_impl(int grid_size); #ifdef __cplusplus } #endif - diff --git a/ggml/src/ggml-sycl.cpp b/ggml/src/ggml-sycl.cpp index 76bad57e2..21006cd7b 100644 --- a/ggml/src/ggml-sycl.cpp +++ b/ggml/src/ggml-sycl.cpp @@ -49,7 +49,7 @@ bool ggml_backend_is_sycl(ggml_backend_t backend); int ggml_backend_sycl_get_device(ggml_backend_t backend); static bool ggml_backend_buffer_is_sycl_split(ggml_backend_buffer_t buffer); static inline int get_sycl_env(const char *env_name, int default_val); -static inline int get_work_group_size(const sycl::device& device); + void dev2dev_memcpy(sycl::queue &q_dst, sycl::queue &q_src, void *ptr_dst, const void *ptr_src, size_t size) { @@ -892,117 +892,6 @@ static void diag_mask_inf_f32(const float * x, float * dst, const int ncols, con dst[i] = x[i] - (col > n_past + row % rows_per_channel) * FLT_MAX; } - -template -static void soft_max_f32(const float * x, const float * mask, float * dst, const int ncols_par, - const int nrows_y, const float scale, const float max_bias, const float m0, - const float m1, uint32_t n_head_log2, const sycl::nd_item<3> &item_ct1, float *buf) { - const int ncols = ncols_template == 0 ? ncols_par : ncols_template; - - const int tid = item_ct1.get_local_id(2); - const int rowx = item_ct1.get_group(2); - const int rowy = rowx % nrows_y; // broadcast the mask (y) in the row dimension - - const int block_size = block_size_template == 0 ? item_ct1.get_local_range(2) : block_size_template; - - const int warp_id = item_ct1.get_local_id(2) / WARP_SIZE; - const int lane_id = item_ct1.get_local_id(2) % WARP_SIZE; - - float slope = 1.0f; - - // ALiBi - if (max_bias > 0.0f) { - const uint32_t h = rowx/nrows_y; // head index - - const float base = h < n_head_log2 ? m0 : m1; - const int exp = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1; - - slope = sycl::pow(base, float(exp)); - } - - float * vals = vals_smem ? buf + WARP_SIZE : dst + rowx*ncols; - float max_val = -INFINITY; - - for (int col0 = 0; col0 < ncols; col0 += block_size) { - const int col = col0 + tid; - - if (ncols_template == 0 && col >= ncols) { - break; - } - - const int ix = rowx*ncols + col; - const int iy = rowy*ncols + col; - - const float val = x[ix]*scale + (mask ? slope*mask[iy] : 0.0f); - - vals[col] = val; - max_val = sycl::max(max_val, val); - } - - // find the max value in the block - max_val = warp_reduce_max(max_val, item_ct1); - if (block_size > WARP_SIZE) { - if (warp_id == 0) { - buf[lane_id] = -INFINITY; - } - item_ct1.barrier(sycl::access::fence_space::local_space); - - if (lane_id == 0) { - buf[warp_id] = max_val; - } - item_ct1.barrier(sycl::access::fence_space::local_space); - - max_val = buf[lane_id]; - max_val = warp_reduce_max(max_val, item_ct1); - } - - float tmp = 0.f; - -#pragma unroll - for (int col0 = 0; col0 < ncols; col0 += block_size) { - const int col = col0 + tid; - if (ncols_template == 0 && col >= ncols) { - break; - } - - const float val = sycl::native::exp(vals[col] - max_val); - tmp += val; - vals[col] = val; - } - - // find the sum of exps in the block - tmp = warp_reduce_sum(tmp, item_ct1); - if (block_size > WARP_SIZE) { - item_ct1.barrier(sycl::access::fence_space::local_space); - if (warp_id == 0) { - buf[lane_id] = 0.f; - } - item_ct1.barrier(sycl::access::fence_space::local_space); - - if (lane_id == 0) { - buf[warp_id] = tmp; - } - item_ct1.barrier(sycl::access::fence_space::local_space); - - tmp = buf[lane_id]; - tmp = warp_reduce_sum(tmp, item_ct1); - } - - const float inv_sum = 1.f / tmp; - -#pragma unroll - for (int col0 = 0; col0 < ncols; col0 += block_size) { - const int col = col0 + tid; - - if (ncols_template == 0 && col >= ncols) { - return; - } - - const int idst = rowx*ncols + col; - dst[idst] = vals[col] * inv_sum; - } -} - static void scale_f32(const float * x, float * dst, const float scale, const int k, const sycl::nd_item<3> &item_ct1) { const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + @@ -1890,106 +1779,6 @@ static void diag_mask_inf_f32_sycl(const float *x, float *dst, }); } -template -static void soft_max_f32_submitter(const float * x, const float * mask, float * dst, const int ncols_par, - const int nrows_y, const float scale, const float max_bias, const float m0, - const float m1, uint32_t n_head_log2, sycl::range<3> block_nums, sycl::range<3> block_dims, - const size_t n_local_scratch, queue_ptr stream) { - stream->submit([&](sycl::handler &cgh) { - sycl::local_accessor local_buf_acc(n_local_scratch, cgh); - - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { - soft_max_f32(x, mask, dst, ncols_par, - nrows_y, scale, max_bias, m0, - m1, n_head_log2, item_ct1, - local_buf_acc.get_pointer()); - }); - }); -} - -static void soft_max_f32_sycl(const float * x, const float * mask, - float * dst, const int ncols_x, const int nrows_x, - const int nrows_y, const float scale, const float max_bias, - queue_ptr stream) { - int nth = WARP_SIZE; - int max_block_size = get_work_group_size(stream->get_device()); - while (nth < ncols_x && nth < max_block_size) nth *= 2; - if (nth>max_block_size) nth = max_block_size; - - const sycl::range<3> block_dims(1, 1, nth); - const sycl::range<3> block_nums(1, 1, nrows_x); - const size_t n_local_scratch = (GGML_PAD(ncols_x, WARP_SIZE) + WARP_SIZE); - - const uint32_t n_head_kv = nrows_x/nrows_y; - const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head_kv)); - - const float m0 = powf(2.0f, -(max_bias ) / n_head_log2); - const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2); - - const size_t local_mem_size = stream->get_device().get_info(); - if (n_local_scratch*sizeof(float) < local_mem_size) { - if (ncols_x > max_block_size) { - soft_max_f32_submitter(x, mask, dst, ncols_x, nrows_y, scale, - max_bias, m0, m1, n_head_log2, block_nums, - block_dims, n_local_scratch, stream); - return; - } - switch (ncols_x) { - case 32: - soft_max_f32_submitter(x, mask, dst, ncols_x, nrows_y, scale, - max_bias, m0, m1, n_head_log2, block_nums, - block_dims, n_local_scratch, stream); - break; - case 64: - soft_max_f32_submitter(x, mask, dst, ncols_x, nrows_y, scale, - max_bias, m0, m1, n_head_log2, block_nums, - block_dims, n_local_scratch, stream); - break; - case 128: - soft_max_f32_submitter(x, mask, dst, ncols_x, nrows_y, scale, - max_bias, m0, m1, n_head_log2, block_nums, - block_dims, n_local_scratch, stream); - break; - case 256: - soft_max_f32_submitter(x, mask, dst, ncols_x, nrows_y, scale, - max_bias, m0, m1, n_head_log2, block_nums, - block_dims, n_local_scratch, stream); - break; - case 512: - soft_max_f32_submitter(x, mask, dst, ncols_x, nrows_y, scale, - max_bias, m0, m1, n_head_log2, block_nums, - block_dims, n_local_scratch, stream); - break; - case 1024: - soft_max_f32_submitter(x, mask, dst, ncols_x, nrows_y, scale, - max_bias, m0, m1, n_head_log2, block_nums, - block_dims, n_local_scratch, stream); - break; - case 2048: - soft_max_f32_submitter(x, mask, dst, ncols_x, nrows_y, scale, - max_bias, m0, m1, n_head_log2, block_nums, - block_dims, n_local_scratch, stream); - break; - case 4096: - soft_max_f32_submitter(x, mask, dst, ncols_x, nrows_y, scale, - max_bias, m0, m1, n_head_log2, block_nums, - block_dims, n_local_scratch, stream); - break; - default: - soft_max_f32_submitter(x, mask, dst, ncols_x, nrows_y, scale, - max_bias, m0, m1, n_head_log2, block_nums, - block_dims, n_local_scratch, stream); - break; - } - } else { - soft_max_f32_submitter(x, mask, dst, ncols_x, nrows_y, scale, - max_bias, m0, m1, n_head_log2, block_nums, - block_dims, WARP_SIZE, stream); - } -} - template static void im2col_sycl(const float *x, T *dst, int IW, int IH, int OW, int OH, int KW, int KH, int IC, @@ -2156,6 +1945,8 @@ static ggml_sycl_device_info ggml_sycl_init() { info.devices[i].cc = 100 * prop.get_major_version() + 10 * prop.get_minor_version(); + + info.max_work_group_sizes[i] = prop.get_max_work_group_size(); } for (int id = 0; id < info.device_count; ++id) { @@ -3007,33 +2798,6 @@ inline void ggml_sycl_op_diag_mask_inf(ggml_backend_sycl_context & ctx, const gg (void) src1_dd; } -inline void ggml_sycl_op_soft_max(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, - const ggml_tensor *src1, ggml_tensor *dst, - const float *src0_dd, const float *src1_dd, - float *dst_dd, - const queue_ptr &main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - -#pragma message("TODO: add ggml_sycl_op_soft_max() F16 src1 support") -#pragma message("ref: https://github.com/ggerganov/llama.cpp/pull/5021") - GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F32); // src1 contains mask and it is optional - - const int64_t ne00 = src0->ne[0]; - const int64_t nrows_x = ggml_nrows(src0); - const int64_t nrows_y = src0->ne[1]; - - float scale = 1.0f; - float max_bias = 0.0f; - - memcpy(&scale, dst->op_params + 0, sizeof(float)); - memcpy(&max_bias, dst->op_params + 1, sizeof(float)); - - soft_max_f32_sycl(src0_dd, src1 ? src1_dd : nullptr, dst_dd, ne00, - nrows_x, nrows_y, scale, max_bias, main_stream); -} - inline void ggml_sycl_op_scale(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, const float *src0_dd, const float *src1_dd, float *dst_dd, @@ -3729,10 +3493,6 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, SYCL_CHECK(ggml_sycl_set_device(ctx.device)); queue_ptr main_stream = ctx.stream();; - bool no_mixed_dtypes = main_stream->get_backend() == sycl::backend::ext_oneapi_cuda || - main_stream->get_backend() == sycl::backend::ext_oneapi_hip; - - void * src0_ddq = src0->data; sycl::half *src0_as_f16 = (sycl::half *)src0_ddq; float * src1_ddf = (float *) src1->data; @@ -3750,15 +3510,10 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, sycl::half *src1_f16 = src1->type == GGML_TYPE_F16 ? (sycl::half *)src1_ddf : src1_f16_alloc.get(); - ggml_sycl_pool_alloc dst_f16(ctx.pool()); char * dst_t; dpct::library_data_t cu_compute_type = dpct::library_data_t::real_float; dpct::library_data_t cu_data_type = dpct::library_data_t::real_float; - if (no_mixed_dtypes) { - cu_compute_type = dpct::library_data_t::real_half; - cu_data_type = dpct::library_data_t::real_half; - } // dst strides size_t nbd2 = dst->nb[2]; @@ -3767,26 +3522,10 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, const float alpha_f32 = 1.0f; const float beta_f32 = 0.0f; - const sycl::half alpha_f16 = 1.0f; - const sycl::half beta_f16 = 0.0f; - const void * alpha = &alpha_f32; const void * beta = &beta_f32; - if (no_mixed_dtypes) { - alpha = &alpha_f16; - beta = &beta_f16; - } - - // TODO: Renable (dst->op_params[0] =! GGML_PREC_DEFAULT) pathway - // when oneMKL open source supports half, half, float, float: datatypes dst_t = (char *) dst_ddf; - if (no_mixed_dtypes) { - dst_t = (char *) dst_f16.alloc(ne_dst); - - nbd2 /= sizeof(float) / sizeof(sycl::half); - nbd3 /= sizeof(float) / sizeof(sycl::half); - } GGML_ASSERT(ne12 % ne02 == 0); GGML_ASSERT(ne13 % ne03 == 0); @@ -3848,11 +3587,6 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, (void **)(ptrs_dst.get() + 0 * ne23), cu_data_type, ne01, ne23, cu_compute_type))); } - - if (no_mixed_dtypes) { - const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16); - to_fp32_sycl(dst_f16.get(), dst_ddf, ne_dst, main_stream); - } } catch (sycl::exception const &exc) { std::cerr << exc.what() << "Exception caught at file:" << __FILE__ @@ -5530,7 +5264,8 @@ GGML_CALL static bool ggml_backend_sycl_supports_op(ggml_backend_t backend, cons case GGML_OP_CONCAT: { ggml_type src0_type = op->src[0]->type; - return src0_type != GGML_TYPE_I32 && src0_type != GGML_TYPE_I16; + int dim = op->op_params[0]; + return ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]) && src0_type != GGML_TYPE_I32 && src0_type != GGML_TYPE_I16 && dim == 2; } break; case GGML_OP_DUP: case GGML_OP_NONE: diff --git a/ggml/src/ggml-sycl/backend.hpp b/ggml/src/ggml-sycl/backend.hpp index 3afa33919..2a789edfc 100644 --- a/ggml/src/ggml-sycl/backend.hpp +++ b/ggml/src/ggml-sycl/backend.hpp @@ -21,5 +21,6 @@ #include "mmvq.hpp" #include "rope.hpp" #include "norm.hpp" +#include "softmax.hpp" #endif // GGML_SYCL_BACKEND_HPP diff --git a/ggml/src/ggml-sycl/common.hpp b/ggml/src/ggml-sycl/common.hpp index dfd4a7c2c..9a1c161b6 100644 --- a/ggml/src/ggml-sycl/common.hpp +++ b/ggml/src/ggml-sycl/common.hpp @@ -47,10 +47,6 @@ static int g_ggml_sycl_debug = 0; } \ }() -// #define DEBUG_SYCL_MALLOC - -static int g_work_group_size = 0; -// typedef sycl::half ggml_fp16_t; #define __SYCL_ARCH__ DPCT_COMPATIBILITY_TEMP #define VER_4VEC 610 // todo for hardward optimize. @@ -193,6 +189,8 @@ struct ggml_sycl_device_info { sycl_device_info devices[GGML_SYCL_MAX_DEVICES] = {}; std::array default_tensor_split = {}; + + int max_work_group_sizes[GGML_SYCL_MAX_DEVICES] = {0}; }; const ggml_sycl_device_info & ggml_sycl_info(); @@ -295,15 +293,6 @@ struct ggml_backend_sycl_context { } }; -// common host functions - -static inline int get_work_group_size(const sycl::device& device) { - dpct::device_info prop; - dpct::get_device_info(prop, device); - return prop.get_max_work_group_size(); -} - - // common device functions static __dpct_inline__ float warp_reduce_sum(float x, @@ -351,4 +340,10 @@ static __dpct_inline__ float warp_reduce_max(float x, return x; } +// Helper for vec loading aligned data +template +inline sycl::vec vec_aligned_load(const Tp* aligned_ptr) { + return *reinterpret_cast*>(aligned_ptr); +} + #endif // GGML_SYCL_COMMON_HPP diff --git a/ggml/src/ggml-sycl/convert.cpp b/ggml/src/ggml-sycl/convert.cpp index ce9de2b42..a15271b51 100644 --- a/ggml/src/ggml-sycl/convert.cpp +++ b/ggml/src/ggml-sycl/convert.cpp @@ -152,12 +152,15 @@ static void dequantize_row_q4_K_sycl(const void *vx, dst_t *y, const int k, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor scale_local_acc(sycl::range<1>(12), cgh); + cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), [=](sycl::nd_item<3> item_ct1) { - dequantize_block_q4_K(vx, y, item_ct1); + dequantize_block_q4_K(vx, y, scale_local_acc.get_pointer(), item_ct1); }); + }); } } diff --git a/ggml/src/ggml-sycl/dequantize.hpp b/ggml/src/ggml-sycl/dequantize.hpp index b6080d83a..ed8ad098b 100644 --- a/ggml/src/ggml-sycl/dequantize.hpp +++ b/ggml/src/ggml-sycl/dequantize.hpp @@ -293,7 +293,8 @@ static void dequantize_block_q3_K(const void * __restrict__ vx, dst_t * __restri #if QK_K == 256 static inline void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) { if (j < 4) { - d = q[j] & 63; m = q[j + 4] & 63; + d = q[j] & 63; + m = q[j + 4] & 63; } else { d = (q[j+4] & 0xF) | ((q[j-4] >> 6) << 4); m = (q[j+4] >> 4) | ((q[j-0] >> 6) << 4); @@ -303,7 +304,7 @@ static inline void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8 template static void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restrict__ yy, - const sycl::nd_item<3> &item_ct1) { + uint8_t* scales_local, const sycl::nd_item<3> &item_ct1) { const block_q4_K * x = (const block_q4_K *) vx; const int i = item_ct1.get_group(2); @@ -318,19 +319,26 @@ static void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restri dst_t * y = yy + i*QK_K + 64*il + n*ir; - const float dall = x[i].dm[0]; - const float dmin = x[i].dm[1]; + const sycl::half2 dm = x[i].dm; + const float dall = dm[0]; + const float dmin = dm[1]; - const uint8_t * q = x[i].qs + 32*il + n*ir; + if (tid < 12) + scales_local[tid] = x[i].scales[tid]; + item_ct1.barrier(sycl::access::fence_space::local_space); uint8_t sc, m; - get_scale_min_k4(is + 0, x[i].scales, sc, m); - const float d1 = dall * sc; const float m1 = dmin * m; - get_scale_min_k4(is + 1, x[i].scales, sc, m); - const float d2 = dall * sc; const float m2 = dmin * m; + get_scale_min_k4(is + 0, scales_local, sc, m); + const float d1 = dall * sc; + const float m1 = dmin * m; + get_scale_min_k4(is + 1, scales_local, sc, m); + const float d2 = dall * sc; + const float m2 = dmin * m; + + sycl::vec q_vec = vec_aligned_load(x[i].qs + 32*il + n*ir); for (int l = 0; l < n; ++l) { - y[l + 0] = d1 * (q[l] & 0xF) - m1; - y[l +32] = d2 * (q[l] >> 4) - m2; + y[l + 0] = d1 * (q_vec[l] & 0xF) - m1; + y[l +32] = d2 * (q_vec[l] >> 4) - m2; } #else const int tid = item_ct1.get_local_id(2); diff --git a/ggml/src/ggml-sycl/dmmv.cpp b/ggml/src/ggml-sycl/dmmv.cpp index 927819281..70a94fc16 100644 --- a/ggml/src/ggml-sycl/dmmv.cpp +++ b/ggml/src/ggml-sycl/dmmv.cpp @@ -3,6 +3,7 @@ #include "dequantize.hpp" #include "presets.hpp" + static void convert_f16(const void * vx, const int ib, const int iqs, dfloat2 & v){ const sycl::half *x = (const sycl::half *)vx; @@ -227,7 +228,7 @@ static void dequantize_mul_mat_vec_q2_k(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { + for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -346,7 +347,7 @@ static void dequantize_mul_mat_vec_q3_k(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { + for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -499,7 +500,7 @@ static void dequantize_mul_mat_vec_q4_k(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { + for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -633,7 +634,7 @@ static void dequantize_mul_mat_vec_q5_k(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { + for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -748,7 +749,7 @@ static void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx, const floa // sum up partial sums and write back result #pragma unroll - for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { + for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -873,10 +874,10 @@ static void dequantize_mul_mat_vec_q2_K_sycl(const void *vx, const float *y, const int ny = 2; // very slightly faster than 1 even when K_QUANTS_PER_ITERATION = 2 const int block_num_y = (nrows + ny - 1) / ny; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, ny, WARP_SIZE); + const sycl::range<3> block_dims(1, ny, QK_WARP_SIZE); stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { dequantize_mul_mat_vec_q2_k(vx, y, dst, ncols, nrows, item_ct1); }); } @@ -889,10 +890,10 @@ static void dequantize_mul_mat_vec_q3_K_sycl(const void *vx, const float *y, const int ny = 2 / K_QUANTS_PER_ITERATION; const int block_num_y = (nrows + ny - 1) / ny; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, ny, WARP_SIZE); + const sycl::range<3> block_dims(1, ny, QK_WARP_SIZE); stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { dequantize_mul_mat_vec_q3_k(vx, y, dst, ncols, nrows, item_ct1); }); } @@ -905,10 +906,10 @@ static void dequantize_mul_mat_vec_q4_K_sycl(const void *vx, const float *y, const int ny = 2 / K_QUANTS_PER_ITERATION; const int block_num_y = (nrows + ny - 1) / ny; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, ny, WARP_SIZE); + const sycl::range<3> block_dims(1, ny, QK_WARP_SIZE); stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { dequantize_mul_mat_vec_q4_k(vx, y, dst, ncols, nrows, item_ct1); }); } @@ -918,10 +919,10 @@ static void dequantize_mul_mat_vec_q5_K_sycl(const void *vx, const float *y, const int nrows, dpct::queue_ptr stream) { GGML_ASSERT(ncols % QK_K == 0); - const sycl::range<3> block_dims(1, 1, WARP_SIZE); + const sycl::range<3> block_dims(1, 1, QK_WARP_SIZE); stream->parallel_for( sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { dequantize_mul_mat_vec_q5_k(vx, y, dst, ncols, item_ct1); }); } @@ -934,10 +935,10 @@ static void dequantize_mul_mat_vec_q6_K_sycl(const void *vx, const float *y, const int ny = 2 / K_QUANTS_PER_ITERATION; const int block_num_y = (nrows + ny - 1) / ny; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, ny, WARP_SIZE); + const sycl::range<3> block_dims(1, ny, QK_WARP_SIZE); stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { dequantize_mul_mat_vec_q6_k(vx, y, dst, ncols, nrows, item_ct1); }); } diff --git a/ggml/src/ggml-sycl/dpct/helper.hpp b/ggml/src/ggml-sycl/dpct/helper.hpp index 1ff297218..31df1cb9e 100644 --- a/ggml/src/ggml-sycl/dpct/helper.hpp +++ b/ggml/src/ggml-sycl/dpct/helper.hpp @@ -255,7 +255,7 @@ namespace dpct void set_pitch(size_t pitch) { _pitch = pitch; } size_t get_x() { return _x; } - void set_x(size_t x) { _x = x; }; + void set_x(size_t x) { _x = x; } size_t get_y() { return _y; } void set_y(size_t y) { _y = y; } @@ -1056,7 +1056,7 @@ namespace dpct #error "Only support Windows and Linux." #endif next_free = mapped_address_space; - }; + } public: using buffer_id_t = int; @@ -1077,7 +1077,7 @@ namespace dpct #else #error "Only support Windows and Linux." #endif - }; + } mem_mgr(const mem_mgr &) = delete; mem_mgr &operator=(const mem_mgr &) = delete; @@ -2426,6 +2426,7 @@ namespace dpct b, ldb, beta, c, ldc, batch_size); break; } +#endif case detail::get_type_combination_id( library_data_t::real_int8, library_data_t::real_int8, library_data_t::real_int32, library_data_t::real_int32): @@ -2458,7 +2459,6 @@ namespace dpct batch_size); break; } -#endif case detail::get_type_combination_id( library_data_t::real_half, library_data_t::real_half, library_data_t::real_half, library_data_t::real_float): @@ -2595,6 +2595,7 @@ namespace dpct stride_c, batch_size); break; } +#endif case detail::get_type_combination_id( library_data_t::real_int8, library_data_t::real_int8, library_data_t::real_int32, library_data_t::real_int32): @@ -2623,7 +2624,6 @@ namespace dpct beta, c, ldc, stride_c, batch_size); break; } -#endif case detail::get_type_combination_id( library_data_t::real_half, library_data_t::real_half, library_data_t::real_half, library_data_t::real_float): diff --git a/ggml/src/ggml-sycl/norm.cpp b/ggml/src/ggml-sycl/norm.cpp index a77f7852c..e0c5dfeca 100644 --- a/ggml/src/ggml-sycl/norm.cpp +++ b/ggml/src/ggml-sycl/norm.cpp @@ -57,6 +57,7 @@ static void group_norm_f32(const float* x, float* dst, const int group_size, con const int nwarps = nthreads / WARP_SIZE; assert(nwarps % WARP_SIZE == 0); start += item_ct1.get_local_id(2); + int nreduce = nwarps / WARP_SIZE; if (end >= ne_elements) { end = ne_elements; @@ -87,7 +88,6 @@ static void group_norm_f32(const float* x, float* dst, const int group_size, con */ item_ct1.barrier(); tmp = 0.f; - int nreduce = nwarps / WARP_SIZE; for (size_t i = 0; i < nreduce; i += 1) { tmp += s_sum[lane_id + i * WARP_SIZE]; @@ -122,7 +122,11 @@ static void group_norm_f32(const float* x, float* dst, const int group_size, con better performance if there is no access to global memory. */ item_ct1.barrier(); - tmp = s_sum[lane_id]; + tmp = 0.f; + for (size_t i = 0; i < nreduce; i += 1) + { + tmp += s_sum[lane_id + i * WARP_SIZE]; + } tmp = warp_reduce_sum(tmp, item_ct1); } @@ -181,7 +185,7 @@ static void rms_norm_f32(const float* x, float* dst, const int ncols, const floa static void norm_f32_sycl(const float* x, float* dst, const int ncols, const int nrows, const float eps, - queue_ptr stream) { + queue_ptr stream, int device) { GGML_ASSERT(ncols % WARP_SIZE == 0); if (ncols < 1024) { const sycl::range<3> block_dims(1, 1, WARP_SIZE); @@ -197,7 +201,7 @@ static void norm_f32_sycl(const float* x, float* dst, const int ncols, }); } else { - const int work_group_size = get_work_group_size(stream->get_device()); + const int work_group_size = ggml_sycl_info().max_work_group_sizes[device]; const sycl::range<3> block_dims(1, 1, work_group_size); /* DPCT1049:17: The work-group size passed to the SYCL kernel may exceed @@ -222,7 +226,7 @@ static void norm_f32_sycl(const float* x, float* dst, const int ncols, static void group_norm_f32_sycl(const float* x, float* dst, const int num_groups, const int group_size, - const int ne_elements, queue_ptr stream) { + const int ne_elements, queue_ptr stream, int device) { static const float eps = 1e-6f; if (group_size < 1024) { const sycl::range<3> block_dims(1, 1, WARP_SIZE); @@ -240,7 +244,7 @@ static void group_norm_f32_sycl(const float* x, float* dst, }); } else { - const int work_group_size = get_work_group_size(stream->get_device()); + const int work_group_size = ggml_sycl_info().max_work_group_sizes[device]; const sycl::range<3> block_dims(1, 1, work_group_size); /* DPCT1049:18: The work-group size passed to the SYCL kernel may exceed @@ -269,7 +273,7 @@ static void group_norm_f32_sycl(const float* x, float* dst, static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols, const int nrows, const float eps, - queue_ptr stream) { + queue_ptr stream, int device) { GGML_ASSERT(ncols % WARP_SIZE == 0); // printf("%s ncols=%d, nrows=%d, WARP_SIZE=%d\n", __func__, ncols, nrows, WARP_SIZE); if (ncols < 1024) { @@ -286,7 +290,7 @@ static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols, }); } else { - const int work_group_size = get_work_group_size(stream->get_device()); + const int work_group_size = ggml_sycl_info().max_work_group_sizes[device]; const sycl::range<3> block_dims(1, 1, work_group_size); /* DPCT1049:19: The work-group size passed to the SYCL kernel may exceed @@ -322,7 +326,7 @@ void ggml_sycl_op_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0, float eps; memcpy(&eps, dst->op_params, sizeof(float)); - norm_f32_sycl(src0_dd, dst_dd, ne00, nrows, eps, main_stream); + norm_f32_sycl(src0_dd, dst_dd, ne00, nrows, eps, main_stream, ctx.device); (void)src1; (void)dst; @@ -340,7 +344,7 @@ void ggml_sycl_op_group_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* int num_groups = dst->op_params[0]; int group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups); - group_norm_f32_sycl(src0_dd, dst_dd, num_groups, group_size, src0->ne[0] * src0->ne[1] * src0->ne[2], main_stream); + group_norm_f32_sycl(src0_dd, dst_dd, num_groups, group_size, src0->ne[0] * src0->ne[1] * src0->ne[2], main_stream, ctx.device); (void)src1; (void)dst; @@ -362,7 +366,7 @@ void ggml_sycl_op_rms_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* sr float eps; memcpy(&eps, dst->op_params, sizeof(float)); - rms_norm_f32_sycl(src0_dd, dst_dd, ne00, nrows, eps, main_stream); + rms_norm_f32_sycl(src0_dd, dst_dd, ne00, nrows, eps, main_stream, ctx.device); (void)src1; (void)dst; diff --git a/ggml/src/ggml-sycl/presets.hpp b/ggml/src/ggml-sycl/presets.hpp index c09c75dc7..15ddcac1f 100644 --- a/ggml/src/ggml-sycl/presets.hpp +++ b/ggml/src/ggml-sycl/presets.hpp @@ -62,4 +62,5 @@ static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUA #define MUL_MAT_SRC1_COL_STRIDE 128 +#define QK_WARP_SIZE 32 #endif // GGML_SYCL_PRESETS_HPP diff --git a/ggml/src/ggml-sycl/softmax.cpp b/ggml/src/ggml-sycl/softmax.cpp new file mode 100644 index 000000000..e624b6ba3 --- /dev/null +++ b/ggml/src/ggml-sycl/softmax.cpp @@ -0,0 +1,250 @@ +#include "norm.hpp" + +template +static void soft_max_f32(const float * x, const float * mask, float * dst, const int ncols_par, + const int nrows_y, const float scale, const float max_bias, const float m0, + const float m1, uint32_t n_head_log2, const sycl::nd_item<3> &item_ct1, float *buf) { + const int ncols = ncols_template == 0 ? ncols_par : ncols_template; + + const int tid = item_ct1.get_local_id(2); + const int rowx = item_ct1.get_group(2); + const int rowy = rowx % nrows_y; // broadcast the mask (y) in the row dimension + + const int block_size = block_size_template == 0 ? item_ct1.get_local_range(2) : block_size_template; + + const int warp_id = item_ct1.get_local_id(2) / WARP_SIZE; + const int lane_id = item_ct1.get_local_id(2) % WARP_SIZE; + const int nthreads = block_size; + const int nwarps = nthreads / WARP_SIZE; + int nreduce = nwarps / WARP_SIZE; + float slope = 1.0f; + + // ALiBi + if (max_bias > 0.0f) { + const uint32_t h = rowx/nrows_y; // head index + + const float base = h < n_head_log2 ? m0 : m1; + const int exp = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1; + + slope = sycl::pow(base, float(exp)); + } + + float *vals = vals_smem ? buf + std::max(nwarps, WARP_SIZE) : dst + rowx * ncols; + float max_val = -INFINITY; + + for (int col0 = 0; col0 < ncols; col0 += block_size) { + const int col = col0 + tid; + + if (ncols_template == 0 && col >= ncols) { + break; + } + + const int ix = rowx*ncols + col; + const int iy = rowy*ncols + col; + + const float val = x[ix]*scale + (mask ? slope*mask[iy] : 0.0f); + + vals[col] = val; + max_val = sycl::max(max_val, val); + } + + // find the max value in the block + max_val = warp_reduce_max(max_val, item_ct1); + if (block_size > WARP_SIZE) { + if (warp_id == 0) { + buf[lane_id] = -INFINITY; + for (size_t i = 1; i < nreduce; i += 1) + buf[lane_id + i * WARP_SIZE] = -INFINITY; + } + item_ct1.barrier(sycl::access::fence_space::local_space); + + if (lane_id == 0) { + buf[warp_id] = max_val; + } + item_ct1.barrier(sycl::access::fence_space::local_space); + max_val = buf[lane_id]; + for (size_t i = 1; i < nreduce; i += 1) + { + max_val = std::max(max_val, buf[lane_id + i * WARP_SIZE]); + } + max_val = warp_reduce_max(max_val, item_ct1); + } + + float tmp = 0.f; +#pragma unroll + for (int col0 = 0; col0 < ncols; col0 += block_size) { + const int col = col0 + tid; + if (ncols_template == 0 && col >= ncols) { + break; + } + + const float val = sycl::native::exp(vals[col] - max_val); + tmp += val; + vals[col] = val; + } + + // find the sum of exps in the block + tmp = warp_reduce_sum(tmp, item_ct1); + if (block_size > WARP_SIZE) { + item_ct1.barrier(sycl::access::fence_space::local_space); + if (warp_id == 0) { + buf[lane_id] = 0.f; + for (size_t i = 1; i < nreduce; i += 1) + buf[lane_id + i * WARP_SIZE] = 0.f; + } + item_ct1.barrier(sycl::access::fence_space::local_space); + + if (lane_id == 0) { + buf[warp_id] = tmp; + } + item_ct1.barrier(sycl::access::fence_space::local_space); + + tmp = buf[lane_id]; + for (size_t i = 1; i < nreduce; i += 1) + { + tmp += buf[lane_id + i * WARP_SIZE]; + } + tmp = warp_reduce_sum(tmp, item_ct1); + } + + const float inv_sum = 1.f / tmp; + +#pragma unroll + for (int col0 = 0; col0 < ncols; col0 += block_size) { + const int col = col0 + tid; + + if (ncols_template == 0 && col >= ncols) { + return; + } + + const int idst = rowx*ncols + col; + dst[idst] = vals[col] * inv_sum; + } +} + +template +static void soft_max_f32_submitter(const float * x, const float * mask, float * dst, const int ncols_par, + const int nrows_y, const float scale, const float max_bias, const float m0, + const float m1, uint32_t n_head_log2, sycl::range<3> block_nums, sycl::range<3> block_dims, + const size_t n_local_scratch, queue_ptr stream) { + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor local_buf_acc(n_local_scratch, cgh); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { + soft_max_f32(x, mask, dst, ncols_par, + nrows_y, scale, max_bias, m0, + m1, n_head_log2, item_ct1, + local_buf_acc.get_pointer()); + }); + }); +} + +static void soft_max_f32_sycl(const float * x, const float * mask, + float * dst, const int ncols_x, const int nrows_x, + const int nrows_y, const float scale, const float max_bias, + queue_ptr stream, int device) { + int nth = WARP_SIZE; + int max_block_size = ggml_sycl_info().max_work_group_sizes[device]; + while (nth < ncols_x && nth < max_block_size) nth *= 2; + if (nth>max_block_size) nth = max_block_size; + + const sycl::range<3> block_dims(1, 1, nth); + const sycl::range<3> block_nums(1, 1, nrows_x); + const size_t n_local_scratch = (GGML_PAD(ncols_x, WARP_SIZE) + WARP_SIZE); + + const uint32_t n_head_kv = nrows_x/nrows_y; + const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head_kv)); + + const float m0 = powf(2.0f, -(max_bias ) / n_head_log2); + const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2); + + const size_t local_mem_size = stream->get_device().get_info(); + if (n_local_scratch*sizeof(float) < local_mem_size) { + if (ncols_x > max_block_size) { + soft_max_f32_submitter(x, mask, dst, ncols_x, nrows_y, scale, + max_bias, m0, m1, n_head_log2, block_nums, + block_dims, n_local_scratch, stream); + return; + } + switch (ncols_x) { + case 32: + soft_max_f32_submitter(x, mask, dst, ncols_x, nrows_y, scale, + max_bias, m0, m1, n_head_log2, block_nums, + block_dims, n_local_scratch, stream); + break; + case 64: + soft_max_f32_submitter(x, mask, dst, ncols_x, nrows_y, scale, + max_bias, m0, m1, n_head_log2, block_nums, + block_dims, n_local_scratch, stream); + break; + case 128: + soft_max_f32_submitter(x, mask, dst, ncols_x, nrows_y, scale, + max_bias, m0, m1, n_head_log2, block_nums, + block_dims, n_local_scratch, stream); + break; + case 256: + soft_max_f32_submitter(x, mask, dst, ncols_x, nrows_y, scale, + max_bias, m0, m1, n_head_log2, block_nums, + block_dims, n_local_scratch, stream); + break; + case 512: + soft_max_f32_submitter(x, mask, dst, ncols_x, nrows_y, scale, + max_bias, m0, m1, n_head_log2, block_nums, + block_dims, n_local_scratch, stream); + break; + case 1024: + soft_max_f32_submitter(x, mask, dst, ncols_x, nrows_y, scale, + max_bias, m0, m1, n_head_log2, block_nums, + block_dims, n_local_scratch, stream); + break; + case 2048: + soft_max_f32_submitter(x, mask, dst, ncols_x, nrows_y, scale, + max_bias, m0, m1, n_head_log2, block_nums, + block_dims, n_local_scratch, stream); + break; + case 4096: + soft_max_f32_submitter(x, mask, dst, ncols_x, nrows_y, scale, + max_bias, m0, m1, n_head_log2, block_nums, + block_dims, n_local_scratch, stream); + break; + default: + soft_max_f32_submitter(x, mask, dst, ncols_x, nrows_y, scale, + max_bias, m0, m1, n_head_log2, block_nums, + block_dims, n_local_scratch, stream); + break; + } + } else { + soft_max_f32_submitter(x, mask, dst, ncols_x, nrows_y, scale, + max_bias, m0, m1, n_head_log2, block_nums, + block_dims, WARP_SIZE, stream); + } +} + +void ggml_sycl_op_soft_max(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, + const ggml_tensor *src1, ggml_tensor *dst, + const float *src0_dd, const float *src1_dd, + float *dst_dd, + const queue_ptr &main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + +#pragma message("TODO: add ggml_sycl_op_soft_max() F16 src1 support") +#pragma message("ref: https://github.com/ggerganov/llama.cpp/pull/5021") + GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F32); // src1 contains mask and it is optional + + const int64_t ne00 = src0->ne[0]; + const int64_t nrows_x = ggml_nrows(src0); + const int64_t nrows_y = src0->ne[1]; + + float scale = 1.0f; + float max_bias = 0.0f; + + memcpy(&scale, dst->op_params + 0, sizeof(float)); + memcpy(&max_bias, dst->op_params + 1, sizeof(float)); + + soft_max_f32_sycl(src0_dd, src1 ? src1_dd : nullptr, dst_dd, ne00, + nrows_x, nrows_y, scale, max_bias, main_stream, ctx.device); +} diff --git a/ggml/src/ggml-sycl/softmax.hpp b/ggml/src/ggml-sycl/softmax.hpp new file mode 100644 index 000000000..bdb8f712e --- /dev/null +++ b/ggml/src/ggml-sycl/softmax.hpp @@ -0,0 +1,24 @@ +// +// MIT license +// Copyright (C) 2024 Intel Corporation +// SPDX-License-Identifier: MIT +// + +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// + +#ifndef GGML_SYCL_SOFTMAX_HPP +#define GGML_SYCL_SOFTMAX_HPP + +#include "common.hpp" + +void ggml_sycl_op_soft_max(ggml_backend_sycl_context &ctx, const ggml_tensor *src0, + const ggml_tensor *src1, ggml_tensor *dst, + const float *src0_dd, const float *src1_dd, + float *dst_dd, + const queue_ptr &main_stream); + +#endif // GGML_SYCL_SOFTMAX_HPP diff --git a/ggml/src/ggml-vulkan-shaders.hpp b/ggml/src/ggml-vulkan-shaders.hpp index 01ff66f71..f0c4c6baf 100644 --- a/ggml/src/ggml-vulkan-shaders.hpp +++ b/ggml/src/ggml-vulkan-shaders.hpp @@ -144954,4 +144954,3 @@ unsigned char sum_rows_f32_data[] = { }; const uint64_t sum_rows_f32_len = 2112; - diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index 78f142a2e..816046ba6 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -5336,7 +5336,7 @@ void ggml_mul_mat_set_prec( as -> [cols, rows, n_expert] ids -> [n_experts_used, n_tokens] (i32) b -> [cols, n_expert_used, n_tokens] - c -> [cols, n_expert_used, n_tokens] + c -> [rows, n_expert_used, n_tokens] in b, n_experts_used can be broadcasted to match the n_expert_used of ids diff --git a/gguf-py/README.md b/gguf-py/README.md index a04c22759..bc46d6e1d 100644 --- a/gguf-py/README.md +++ b/gguf-py/README.md @@ -3,7 +3,7 @@ This is a Python package for writing binary files in the [GGUF](https://github.com/ggerganov/ggml/pull/302) (GGML Universal File) format. -See [convert-llama-hf-to-gguf.py](https://github.com/ggerganov/llama.cpp/blob/master/convert-hf-to-gguf.py) +See [convert_hf_to_gguf.py](https://github.com/ggerganov/llama.cpp/blob/master/convert_hf_to_gguf.py) as an example for its usage. ## Installation @@ -15,13 +15,13 @@ pip install gguf [examples/writer.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/examples/writer.py) — Generates `example.gguf` in the current directory to demonstrate generating a GGUF file. Note that this file cannot be used as a model. -[scripts/gguf-dump.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/scripts/gguf-dump.py) — Dumps a GGUF file's metadata to the console. +[scripts/gguf_dump.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/scripts/gguf_dump.py) — Dumps a GGUF file's metadata to the console. -[scripts/gguf-set-metadata.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/scripts/gguf-set-metadata.py) — Allows changing simple metadata values in a GGUF file by key. +[scripts/gguf_set_metadata.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/scripts/gguf_set_metadata.py) — Allows changing simple metadata values in a GGUF file by key. -[scripts/gguf-convert-endian.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/scripts/gguf-convert-endian.py) — Allows converting the endianness of GGUF files. +[scripts/gguf_convert_endian.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/scripts/gguf_convert_endian.py) — Allows converting the endianness of GGUF files. -[scripts/gguf-new-metadata.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/scripts/gguf-new-metadata.py) — Copies a GGUF file with added/modified/removed metadata values. +[scripts/gguf_new_metadata.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/scripts/gguf_new_metadata.py) — Copies a GGUF file with added/modified/removed metadata values. ## Development Maintainers who participate in development of this package are advised to install it in editable mode: diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index e87c58266..d78fb1e44 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -160,10 +160,12 @@ class MODEL_ARCH(IntEnum): COMMAND_R = auto() DBRX = auto() OLMO = auto() + OPENELM = auto() ARCTIC = auto() DEEPSEEK2 = auto() BITNET = auto() T5 = auto() + JAIS = auto() class MODEL_TENSOR(IntEnum): @@ -284,10 +286,12 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.COMMAND_R: "command-r", MODEL_ARCH.DBRX: "dbrx", MODEL_ARCH.OLMO: "olmo", + MODEL_ARCH.OPENELM: "openelm", MODEL_ARCH.ARCTIC: "arctic", MODEL_ARCH.DEEPSEEK2: "deepseek2", MODEL_ARCH.BITNET: "bitnet", MODEL_ARCH.T5: "t5", + MODEL_ARCH.JAIS: "jais", } TENSOR_NAMES: dict[MODEL_TENSOR, str] = { @@ -859,6 +863,19 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_DOWN, MODEL_TENSOR.FFN_UP, ], + MODEL_ARCH.OPENELM: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_Q_NORM, + MODEL_TENSOR.ATTN_K_NORM, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], MODEL_ARCH.ARCTIC: [ MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.OUTPUT_NORM, @@ -954,6 +971,18 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.ENC_FFN_UP, MODEL_TENSOR.ENC_OUTPUT_NORM, ], + MODEL_ARCH.JAIS: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_UP, + ], # TODO } diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py index 75a8b2636..cf9554162 100644 --- a/gguf-py/gguf/gguf_writer.py +++ b/gguf-py/gguf/gguf_writer.py @@ -480,8 +480,11 @@ class GGUFWriter: def add_leading_dense_block_count(self, length: int) -> None: self.add_uint32(Keys.LLM.LEADING_DENSE_BLOCK_COUNT.format(arch=self.arch), length) - def add_feed_forward_length(self, length: int) -> None: - self.add_uint32(Keys.LLM.FEED_FORWARD_LENGTH.format(arch=self.arch), length) + def add_feed_forward_length(self, length: int | Sequence[int]) -> None: + if isinstance(length, int): + self.add_uint32(Keys.LLM.FEED_FORWARD_LENGTH.format(arch=self.arch), length) + else: + self.add_array(Keys.LLM.FEED_FORWARD_LENGTH.format(arch=self.arch), length) def add_expert_feed_forward_length(self, length: int) -> None: self.add_uint32(Keys.LLM.EXPERT_FEED_FORWARD_LENGTH.format(arch=self.arch), length) @@ -495,11 +498,17 @@ class GGUFWriter: def add_decoder_start_token_id(self, id: int) -> None: self.add_uint32(Keys.LLM.DECODER_START_TOKEN_ID.format(arch=self.arch), id) - def add_head_count(self, count: int) -> None: - self.add_uint32(Keys.Attention.HEAD_COUNT.format(arch=self.arch), count) + def add_head_count(self, count: int | Sequence[int]) -> None: + if isinstance(count, int): + self.add_uint32(Keys.Attention.HEAD_COUNT.format(arch=self.arch), count) + else: + self.add_array(Keys.Attention.HEAD_COUNT.format(arch=self.arch), count) - def add_head_count_kv(self, count: int) -> None: - self.add_uint32(Keys.Attention.HEAD_COUNT_KV.format(arch=self.arch), count) + def add_head_count_kv(self, count: int | Sequence[int]) -> None: + if isinstance(count, int): + self.add_uint32(Keys.Attention.HEAD_COUNT_KV.format(arch=self.arch), count) + else: + self.add_array(Keys.Attention.HEAD_COUNT_KV.format(arch=self.arch), count) def add_key_length(self, length: int) -> None: self.add_uint32(Keys.Attention.KEY_LENGTH.format(arch=self.arch), length) diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py index 0bed43939..6054f6159 100644 --- a/gguf-py/gguf/tensor_mapping.py +++ b/gguf-py/gguf/tensor_mapping.py @@ -10,7 +10,7 @@ class TensorNameMap: # Token embeddings MODEL_TENSOR.TOKEN_EMBD: ( "gpt_neox.embed_in", # gptneox - "transformer.wte", # gpt2 gpt-j mpt refact qwen dbrx + "transformer.wte", # gpt2 gpt-j mpt refact qwen dbrx jais "transformer.word_embeddings", # falcon "word_embeddings", # bloom "model.embed_tokens", # llama-hf @@ -24,6 +24,7 @@ class TensorNameMap: "backbone.embedding", # mamba "backbone.embeddings", # mamba-hf "transformer.in_out_embed", # Grok + "transformer.token_embeddings", # openelm "shared", # t5 ), @@ -37,6 +38,7 @@ class TensorNameMap: "word_embeddings_layernorm", # bloom "embeddings.LayerNorm", # bert "emb_ln", # nomic-bert + "transformer.norm", # openelm ), # Position embeddings @@ -49,7 +51,7 @@ class TensorNameMap: # Output MODEL_TENSOR.OUTPUT: ( "embed_out", # gptneox - "lm_head", # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx + "lm_head", # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx jais "output", # llama-pth bloom internlm2 "word_embeddings_for_head", # persimmon "lm_head.linear", # phi2 @@ -58,7 +60,7 @@ class TensorNameMap: # Output norm MODEL_TENSOR.OUTPUT_NORM: ( "gpt_neox.final_layer_norm", # gptneox - "transformer.ln_f", # gpt2 gpt-j falcon + "transformer.ln_f", # gpt2 gpt-j falcon jais "model.norm", # llama-hf baichuan internlm2 "norm", # llama-pth "transformer.norm_f", # mpt dbrx @@ -69,6 +71,7 @@ class TensorNameMap: "model.norm_f", # mamba-qbert "backbone.norm_f", # mamba "transformer.rms_norm", # Grok + "transformer.norm", # openelm ), # Rope frequencies @@ -81,7 +84,7 @@ class TensorNameMap: # Attention norm MODEL_TENSOR.ATTN_NORM: ( "gpt_neox.layers.{bid}.input_layernorm", # gptneox - "transformer.h.{bid}.ln_1", # gpt2 gpt-j refact qwen + "transformer.h.{bid}.ln_1", # gpt2 gpt-j refact qwen jais "transformer.blocks.{bid}.norm_1", # mpt "transformer.h.{bid}.input_layernorm", # falcon7b "h.{bid}.input_layernorm", # bloom @@ -98,6 +101,7 @@ class TensorNameMap: "backbone.layers.{bid}.norm", # mamba "transformer.decoder_layer.{bid}.rms_norm", # Grok "transformer.blocks.{bid}.norm_attn_norm.norm_1", # dbrx + "transformer.layers.{bid}.attn_norm", # openelm ), # Attention norm 2 @@ -109,7 +113,7 @@ class TensorNameMap: # Attention query-key-value MODEL_TENSOR.ATTN_QKV: ( "gpt_neox.layers.{bid}.attention.query_key_value", # gptneox - "transformer.h.{bid}.attn.c_attn", # gpt2 qwen + "transformer.h.{bid}.attn.c_attn", # gpt2 qwen jais "transformer.blocks.{bid}.attn.Wqkv", # mpt "transformer.blocks.{bid}.norm_attn_norm.attn.Wqkv", # dbrx "transformer.h.{bid}.self_attention.query_key_value", # falcon @@ -119,7 +123,8 @@ class TensorNameMap: "h.{bid}.attn.c_attn", # gpt2 "transformer.h.{bid}.mixer.Wqkv", # phi2 "encoder.layers.{bid}.attn.Wqkv", # nomic-bert - "model.layers.{bid}.self_attn.qkv_proj" # phi3 + "model.layers.{bid}.self_attn.qkv_proj", # phi3 + "transformer.layers.{bid}.attn.qkv_proj", # openelm ), # Attention query @@ -160,7 +165,7 @@ class TensorNameMap: # Attention output MODEL_TENSOR.ATTN_OUT: ( "gpt_neox.layers.{bid}.attention.dense", # gptneox - "transformer.h.{bid}.attn.c_proj", # gpt2 refact qwen + "transformer.h.{bid}.attn.c_proj", # gpt2 refact qwen jais "transformer.blocks.{bid}.attn.out_proj", # mpt "transformer.h.{bid}.self_attention.dense", # falcon "h.{bid}.self_attention.dense", # bloom @@ -177,6 +182,7 @@ class TensorNameMap: "encoder.layers.{bid}.attn.out_proj", # nomic-bert "transformer.decoder_layer.{bid}.multi_head_attention.linear", # Grok "transformer.blocks.{bid}.norm_attn_norm.attn.out_proj", # dbrx + "transformer.layers.{bid}.attn.out_proj", # openelm ), # Attention output norm @@ -202,7 +208,7 @@ class TensorNameMap: # Feed-forward norm MODEL_TENSOR.FFN_NORM: ( "gpt_neox.layers.{bid}.post_attention_layernorm", # gptneox - "transformer.h.{bid}.ln_2", # gpt2 refact qwen + "transformer.h.{bid}.ln_2", # gpt2 refact qwen jais "h.{bid}.post_attention_layernorm", # bloom "transformer.blocks.{bid}.norm_2", # mpt "model.layers.{bid}.post_attention_layernorm", # llama-hf @@ -212,6 +218,7 @@ class TensorNameMap: "h.{bid}.ln_2", # gpt2 "model.layers.{bid}.ffn_norm", # internlm2 "transformer.decoder_layer.{bid}.rms_norm_2", # Grok + "transformer.layers.{bid}.ffn_norm", # openelm ), # Post feed-forward norm @@ -239,7 +246,7 @@ class TensorNameMap: # Feed-forward up MODEL_TENSOR.FFN_UP: ( "gpt_neox.layers.{bid}.mlp.dense_h_to_4h", # gptneox - "transformer.h.{bid}.mlp.c_fc", # gpt2 + "transformer.h.{bid}.mlp.c_fc", # gpt2 jais "transformer.blocks.{bid}.ffn.up_proj", # mpt "transformer.h.{bid}.mlp.dense_h_to_4h", # falcon "h.{bid}.mlp.dense_h_to_4h", # bloom @@ -285,6 +292,7 @@ class TensorNameMap: "model.layers.{bid}.mlp.gate_proj", # llama-hf refact "layers.{bid}.feed_forward.w1", # llama-pth "transformer.h.{bid}.mlp.w2", # qwen + "transformer.h.{bid}.mlp.c_fc2", # jais "model.layers.layers.{bid}.mlp.gate_proj", # plamo "model.layers.{bid}.feed_forward.w1", # internlm2 "encoder.layers.{bid}.mlp.fc12", # nomic-bert @@ -308,7 +316,7 @@ class TensorNameMap: # Feed-forward down MODEL_TENSOR.FFN_DOWN: ( "gpt_neox.layers.{bid}.mlp.dense_4h_to_h", # gptneox - "transformer.h.{bid}.mlp.c_proj", # gpt2 refact qwen + "transformer.h.{bid}.mlp.c_proj", # gpt2 refact qwen jais "transformer.blocks.{bid}.ffn.down_proj", # mpt "transformer.h.{bid}.mlp.dense_4h_to_h", # falcon "h.{bid}.mlp.dense_4h_to_h", # bloom @@ -326,6 +334,7 @@ class TensorNameMap: "encoder.layers.{bid}.mlp.fc2", # nomic-bert "model.layers.{bid}.mlp.c_proj", # starcoder2 "encoder.layer.{bid}.mlp.wo", # jina-bert-v2 + "transformer.layers.{bid}.ffn.proj_2", # openelm "model.layers.{bid}.residual_mlp.w2", # arctic "encoder.layer.{bid}.mlp.down_layer", # jina-bert-v2 ), @@ -347,7 +356,8 @@ class TensorNameMap: "model.layers.{bid}.self_attn.q_layernorm", # persimmon "model.layers.{bid}.self_attn.q_norm", # cohere "transformer.blocks.{bid}.attn.q_ln", # sea-lion - "encoder.layer.{bid}.attention.self.layer_norm_q" # jina-bert-v2 + "encoder.layer.{bid}.attention.self.layer_norm_q", # jina-bert-v2 + "transformer.layers.{bid}.attn.q_norm", # openelm ), MODEL_TENSOR.ATTN_K_NORM: ( @@ -355,7 +365,8 @@ class TensorNameMap: "model.layers.{bid}.self_attn.k_layernorm", # persimmon "model.layers.{bid}.self_attn.k_norm", # cohere "transformer.blocks.{bid}.attn.k_ln", # sea-lion - "encoder.layer.{bid}.attention.self.layer_norm_k" # jina-bert-v2 + "encoder.layer.{bid}.attention.self.layer_norm_k", # jina-bert-v2 + "transformer.layers.{bid}.attn.k_norm", # openelm ), MODEL_TENSOR.ROPE_FREQS: ( diff --git a/gguf-py/scripts/__init__.py b/gguf-py/scripts/__init__.py index 1ad45639a..f9d29cb69 100644 --- a/gguf-py/scripts/__init__.py +++ b/gguf-py/scripts/__init__.py @@ -1,13 +1,4 @@ -import os - -from importlib import import_module - - -os.environ["NO_LOCAL_GGUF"] = "TRUE" - -gguf_convert_endian_entrypoint = import_module("scripts.gguf-convert-endian").main -gguf_dump_entrypoint = import_module("scripts.gguf-dump").main -gguf_set_metadata_entrypoint = import_module("scripts.gguf-set-metadata").main -gguf_new_metadata_entrypoint = import_module("scripts.gguf-new-metadata").main - -del import_module, os +from .gguf_convert_endian import main as gguf_convert_endian_entrypoint +from .gguf_dump import main as gguf_dump_entrypoint +from .gguf_set_metadata import main as gguf_set_metadata_entrypoint +from .gguf_new_metadata import main as gguf_new_metadata_entrypoint diff --git a/gguf-py/scripts/gguf-convert-endian.py b/gguf-py/scripts/gguf_convert_endian.py similarity index 100% rename from gguf-py/scripts/gguf-convert-endian.py rename to gguf-py/scripts/gguf_convert_endian.py diff --git a/gguf-py/scripts/gguf-dump.py b/gguf-py/scripts/gguf_dump.py similarity index 100% rename from gguf-py/scripts/gguf-dump.py rename to gguf-py/scripts/gguf_dump.py diff --git a/gguf-py/scripts/gguf-new-metadata.py b/gguf-py/scripts/gguf_new_metadata.py similarity index 100% rename from gguf-py/scripts/gguf-new-metadata.py rename to gguf-py/scripts/gguf_new_metadata.py diff --git a/gguf-py/scripts/gguf-set-metadata.py b/gguf-py/scripts/gguf_set_metadata.py similarity index 100% rename from gguf-py/scripts/gguf-set-metadata.py rename to gguf-py/scripts/gguf_set_metadata.py diff --git a/include/llama.h b/include/llama.h index 8a2e262eb..a80f49e3d 100644 --- a/include/llama.h +++ b/include/llama.h @@ -89,6 +89,7 @@ extern "C" { LLAMA_VOCAB_PRE_TYPE_SMAUG = 14, LLAMA_VOCAB_PRE_TYPE_PORO = 15, LLAMA_VOCAB_PRE_TYPE_VIKING = 16, + LLAMA_VOCAB_PRE_TYPE_JAIS = 17, }; // note: these values should be synchronized with ggml_rope @@ -179,6 +180,12 @@ extern "C" { LLAMA_POOLING_TYPE_LAST = 3, }; + enum llama_attention_type { + LLAMA_ATTENTION_TYPE_UNSPECIFIED = -1, + LLAMA_ATTENTION_TYPE_CAUSAL = 0, + LLAMA_ATTENTION_TYPE_NON_CAUSAL = 1, + }; + enum llama_split_mode { LLAMA_SPLIT_MODE_NONE = 0, // single GPU LLAMA_SPLIT_MODE_LAYER = 1, // split layers and KV across GPUs @@ -296,6 +303,7 @@ extern "C" { enum llama_rope_scaling_type rope_scaling_type; // RoPE scaling type, from `enum llama_rope_scaling_type` enum llama_pooling_type pooling_type; // whether to pool (sum) embedding results by sequence id + enum llama_attention_type attention_type; // attention type to use for embeddings // ref: https://github.com/ggerganov/llama.cpp/pull/2054 float rope_freq_base; // RoPE base frequency, 0 = from model @@ -484,6 +492,13 @@ extern "C" { // Get a llama model tensor LLAMA_API struct ggml_tensor * llama_get_model_tensor(struct llama_model * model, const char * name); + // Returns true if the model contains an encoder that requires llama_encode() call + LLAMA_API bool llama_model_has_encoder(const struct llama_model * model); + + // For encoder-decoder models, this function returns id of the token that must be provided + // to the decoder to start generating output sequence. For other models, it returns -1. + LLAMA_API llama_token llama_model_decoder_start_token(const struct llama_model * model); + // Returns 0 on success LLAMA_API uint32_t llama_model_quantize( const char * fname_inp, @@ -771,6 +786,14 @@ extern "C" { // Frees a batch of tokens allocated with llama_batch_init() LLAMA_API void llama_batch_free(struct llama_batch batch); + // Processes a batch of tokens with the ecoder part of the encoder-decoder model. + // Stores the encoder output internally for later use by the decoder cross-attention layers. + // 0 - success + // < 0 - error + LLAMA_API int32_t llama_encode( + struct llama_context * ctx, + struct llama_batch batch); + // Positive return values does not mean a fatal error, but rather a warning. // 0 - success // 1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context) diff --git a/models/ggml-vocab-command-r.gguf.inp b/models/ggml-vocab-command-r.gguf.inp index 0a89107c6..9baf7d77a 100644 --- a/models/ggml-vocab-command-r.gguf.inp +++ b/models/ggml-vocab-command-r.gguf.inp @@ -73,6 +73,8 @@ __ggml_vocab_test__ __ggml_vocab_test__ Hello, y'all! How are you 😁 ?我想在apple工作1314151天～ __ggml_vocab_test__ +!!!!!! +__ggml_vocab_test__ 3 __ggml_vocab_test__ 33 @@ -91,6 +93,10 @@ __ggml_vocab_test__ __ggml_vocab_test__ 333333333 __ggml_vocab_test__ +Cửa Việt +__ggml_vocab_test__ + discards +__ggml_vocab_test__ diff --git a/models/ggml-vocab-command-r.gguf.out b/models/ggml-vocab-command-r.gguf.out index cc4277daa..3f6b41888 100644 --- a/models/ggml-vocab-command-r.gguf.out +++ b/models/ggml-vocab-command-r.gguf.out @@ -31,6 +31,7 @@ 206 1857 14 4515 28339 19 1770 14 1954 8 4070 1955 1933 80503 231 5691 12081 13336 2648 29325 14315 24 26 24 27 24 28 24 5123 18372 + 57178 10251 26 26 26 26 26 26 @@ -40,4 +41,6 @@ 26 26 26 26 26 26 26 26 26 26 26 26 26 26 26 26 26 26 26 26 26 26 26 26 + 42 30719 12584 + 3642 4388 127731 51628 205 57788 18494 97469 126134 206 2226 256 230 1737 18258 16 80503 122 35927 2226 242 112 57462 1737 54457 223165 106230 2096 16 48389 11254 107 255 2226 107 255 228 26 228 26 26 228 26 26 26 228 26 26 26 26 228 26 26 26 26 26 228 26 26 26 26 26 26 228 26 26 26 26 26 26 26 228 26 26 26 26 26 26 26 26 228 26 21 26 228 26 2271 26 228 26 3834 26 182018 230 174833 38111 249 86325 241 38111 245 86325 232 38111 252 38111 123 38111 261 165 24629 38111 261 38111 103 174833 38111 235 188568 231 5691 12081 13336 2648 29325 14315 24 26 24 27 24 28 24 5123 18372 8391 158343 3512 40071 2196 3236 8750 1764 37097 41168 29721 32797 25646 3802 4975 4975 116167 57178 10251 154048 27292 1767 5125 2632 2155 91 2378 1919 1914 2782 19 2155 3354 1933 5470 38 2155 52 2068 5470 1767 4961 3059 1894 19 2155 43 1933 3026 2725 23186 38 2930 14 20676 1671 14 83 51 diff --git a/models/ggml-vocab-qwen2.gguf.inp b/models/ggml-vocab-qwen2.gguf.inp index 0a89107c6..9baf7d77a 100644 --- a/models/ggml-vocab-qwen2.gguf.inp +++ b/models/ggml-vocab-qwen2.gguf.inp @@ -73,6 +73,8 @@ __ggml_vocab_test__ __ggml_vocab_test__ Hello, y'all! How are you 😁 ?我想在apple工作1314151天～ __ggml_vocab_test__ +!!!!!! +__ggml_vocab_test__ 3 __ggml_vocab_test__ 33 @@ -91,6 +93,10 @@ __ggml_vocab_test__ __ggml_vocab_test__ 333333333 __ggml_vocab_test__ +Cửa Việt +__ggml_vocab_test__ + discards +__ggml_vocab_test__ diff --git a/models/ggml-vocab-qwen2.gguf.out b/models/ggml-vocab-qwen2.gguf.out index 401a510e8..18b4b45cd 100644 --- a/models/ggml-vocab-qwen2.gguf.out +++ b/models/ggml-vocab-qwen2.gguf.out @@ -31,6 +31,7 @@ 198 284 6 11385 9707 11 379 64848 0 2585 525 498 26525 223 937 104100 18493 22377 99257 16 18 16 19 16 20 16 35727 21216 + 17085 2928 18 18 18 18 18 18 @@ -40,4 +41,6 @@ 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 + 34 90063 128324 + 2560 2347 198 4710 14731 65497 7847 1572 2303 78672 10947 145836 320 8252 8 26525 114 378 235 149921 30543 320 35673 99066 97534 8 25521 227 11162 99 247 149955 220 18 220 18 18 220 18 18 18 220 18 18 18 18 220 18 18 18 18 18 220 18 18 18 18 18 18 220 18 18 18 18 18 18 18 220 18 18 18 18 18 18 18 18 220 18 13 18 220 18 496 18 220 18 1112 18 220 146394 97529 241 44258 233 146568 44258 224 147603 20879 115 146280 44258 223 146280 147272 97529 227 144534 937 104100 18493 22377 99257 16 18 16 19 16 20 16 35727 21216 55460 53237 18658 14144 1456 13073 63471 33594 3038 133178 79012 3355 4605 4605 13874 13874 73594 3014 3014 28149 17085 2928 26610 7646 358 3003 1012 364 83 813 566 594 1052 11 364 787 498 2704 30 364 44 537 2704 358 3278 1281 432 11 364 35 498 1075 1045 15243 30 1205 6 42612 264 63866 43 diff --git a/models/ggml-vocab-refact.gguf.inp b/models/ggml-vocab-refact.gguf.inp index 0a89107c6..9baf7d77a 100644 --- a/models/ggml-vocab-refact.gguf.inp +++ b/models/ggml-vocab-refact.gguf.inp @@ -73,6 +73,8 @@ __ggml_vocab_test__ __ggml_vocab_test__ Hello, y'all! How are you 😁 ?我想在apple工作1314151天～ __ggml_vocab_test__ +!!!!!! +__ggml_vocab_test__ 3 __ggml_vocab_test__ 33 @@ -91,6 +93,10 @@ __ggml_vocab_test__ __ggml_vocab_test__ 333333333 __ggml_vocab_test__ +Cửa Việt +__ggml_vocab_test__ + discards +__ggml_vocab_test__ diff --git a/models/ggml-vocab-refact.gguf.out b/models/ggml-vocab-refact.gguf.out index 06b15c090..63d8305c3 100644 --- a/models/ggml-vocab-refact.gguf.out +++ b/models/ggml-vocab-refact.gguf.out @@ -31,6 +31,7 @@ 203 280 25 34666 8279 30 533 25 464 19 4971 884 844 18458 228 1018 4982 13368 2909 9513 17827 35 37 35 38 35 39 35 11873 47838 + 9163 3202 37 37 37 37 37 37 @@ -40,4 +41,6 @@ 37 37 37 37 37 37 37 37 37 37 37 37 37 37 37 37 37 37 37 37 37 37 37 37 + 53 33934 83 33217 17102 102 + 1214 12258 334 719 8878 202 10885 4222 16104 28570 203 3807 253 227 308 4382 27 18458 133 46113 44967 123 13868 308 12565 19775 33071 40824 733 27 41889 5945 118 252 3807 118 252 225 37 225 37 37 225 37 37 37 225 37 37 37 37 225 37 37 37 37 37 225 37 37 37 37 37 37 225 37 37 37 37 37 37 37 225 37 37 37 37 37 37 37 37 225 37 32 37 225 37 497 37 225 37 1179 37 225 14574 227 14574 133 14574 246 30457 238 14574 242 30457 229 14574 249 14574 134 14574 258 30457 228 14574 258 14574 114 14574 133 14574 232 36628 228 1018 4982 13368 2909 9513 17827 35 37 35 38 35 39 35 11873 47838 20921 16623 13028 8372 1039 9446 40242 13852 2053 8949 12531 1520 10700 5881 9592 13299 914 31753 31359 9163 3202 35472 10397 439 4763 2583 330 102 1455 938 1182 2017 30 330 613 844 3654 49 330 63 646 3654 439 4621 1930 561 30 330 54 844 2124 1629 35993 49 2688 25 7709 312 25 94 62 diff --git a/poetry.lock b/poetry.lock new file mode 100644 index 000000000..eb6baa6c7 --- /dev/null +++ b/poetry.lock @@ -0,0 +1,1197 @@ +# This file is automatically @generated by Poetry 1.7.1 and should not be changed by hand. + +[[package]] +name = "atomicwrites" +version = "1.4.1" +description = "Atomic file writes." +optional = false +python-versions = ">=2.7, !=3.0.*, !=3.1.*, !=3.2.*, !=3.3.*" +files = [ + {file = "atomicwrites-1.4.1.tar.gz", hash = "sha256:81b2c9071a49367a7f770170e5eec8cb66567cfbbc8c73d20ce5ca4a8d71cf11"}, +] + +[[package]] +name = "attrs" +version = "23.2.0" +description = "Classes Without Boilerplate" +optional = false +python-versions = ">=3.7" +files = [ + {file = "attrs-23.2.0-py3-none-any.whl", hash = "sha256:99b87a485a5820b23b879f04c2305b44b951b502fd64be915879d77a7e8fc6f1"}, + {file = "attrs-23.2.0.tar.gz", hash = "sha256:935dc3b529c262f6cf76e50877d35a4bd3c1de194fd41f47a2b7ae8f19971f30"}, +] + +[package.extras] +cov = ["attrs[tests]", "coverage[toml] (>=5.3)"] +dev = ["attrs[tests]", "pre-commit"] +docs = ["furo", "myst-parser", "sphinx", "sphinx-notfound-page", "sphinxcontrib-towncrier", "towncrier", "zope-interface"] +tests = ["attrs[tests-no-zope]", "zope-interface"] +tests-mypy = ["mypy (>=1.6)", "pytest-mypy-plugins"] +tests-no-zope = ["attrs[tests-mypy]", "cloudpickle", "hypothesis", "pympler", "pytest (>=4.3.0)", "pytest-xdist[psutil]"] + +[[package]] +name = "certifi" +version = "2024.2.2" +description = "Python package for providing Mozilla's CA Bundle." +optional = false +python-versions = ">=3.6" +files = [ + {file = "certifi-2024.2.2-py3-none-any.whl", hash = "sha256:dc383c07b76109f368f6106eee2b593b04a011ea4d55f652c6ca24a754d1cdd1"}, + {file = "certifi-2024.2.2.tar.gz", hash = "sha256:0569859f95fc761b18b45ef421b1290a0f65f147e92a1e5eb3e635f9a5e4e66f"}, +] + +[[package]] +name = "charset-normalizer" +version = "3.3.2" +description = "The Real First Universal Charset Detector. Open, modern and actively maintained alternative to Chardet." +optional = false +python-versions = ">=3.7.0" +files = [ + {file = "charset-normalizer-3.3.2.tar.gz", hash = "sha256:f30c3cb33b24454a82faecaf01b19c18562b1e89558fb6c56de4d9118a032fd5"}, + {file = "charset_normalizer-3.3.2-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:25baf083bf6f6b341f4121c2f3c548875ee6f5339300e08be3f2b2ba1721cdd3"}, + {file = "charset_normalizer-3.3.2-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:06435b539f889b1f6f4ac1758871aae42dc3a8c0e24ac9e60c2384973ad73027"}, + {file = "charset_normalizer-3.3.2-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:9063e24fdb1e498ab71cb7419e24622516c4a04476b17a2dab57e8baa30d6e03"}, + {file = "charset_normalizer-3.3.2-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:6897af51655e3691ff853668779c7bad41579facacf5fd7253b0133308cf000d"}, + {file = "charset_normalizer-3.3.2-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:1d3193f4a680c64b4b6a9115943538edb896edc190f0b222e73761716519268e"}, + {file = "charset_normalizer-3.3.2-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:cd70574b12bb8a4d2aaa0094515df2463cb429d8536cfb6c7ce983246983e5a6"}, + {file = "charset_normalizer-3.3.2-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:8465322196c8b4d7ab6d1e049e4c5cb460d0394da4a27d23cc242fbf0034b6b5"}, + {file = "charset_normalizer-3.3.2-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:a9a8e9031d613fd2009c182b69c7b2c1ef8239a0efb1df3f7c8da66d5dd3d537"}, + {file = "charset_normalizer-3.3.2-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:beb58fe5cdb101e3a055192ac291b7a21e3b7ef4f67fa1d74e331a7f2124341c"}, + {file = "charset_normalizer-3.3.2-cp310-cp310-musllinux_1_1_i686.whl", hash = "sha256:e06ed3eb3218bc64786f7db41917d4e686cc4856944f53d5bdf83a6884432e12"}, + {file = "charset_normalizer-3.3.2-cp310-cp310-musllinux_1_1_ppc64le.whl", hash = "sha256:2e81c7b9c8979ce92ed306c249d46894776a909505d8f5a4ba55b14206e3222f"}, + {file = "charset_normalizer-3.3.2-cp310-cp310-musllinux_1_1_s390x.whl", hash = "sha256:572c3763a264ba47b3cf708a44ce965d98555f618ca42c926a9c1616d8f34269"}, + {file = "charset_normalizer-3.3.2-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:fd1abc0d89e30cc4e02e4064dc67fcc51bd941eb395c502aac3ec19fab46b519"}, + {file = "charset_normalizer-3.3.2-cp310-cp310-win32.whl", hash = "sha256:3d47fa203a7bd9c5b6cee4736ee84ca03b8ef23193c0d1ca99b5089f72645c73"}, + {file = "charset_normalizer-3.3.2-cp310-cp310-win_amd64.whl", hash = "sha256:10955842570876604d404661fbccbc9c7e684caf432c09c715ec38fbae45ae09"}, + {file = "charset_normalizer-3.3.2-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:802fe99cca7457642125a8a88a084cef28ff0cf9407060f7b93dca5aa25480db"}, + {file = "charset_normalizer-3.3.2-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:573f6eac48f4769d667c4442081b1794f52919e7edada77495aaed9236d13a96"}, + {file = "charset_normalizer-3.3.2-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:549a3a73da901d5bc3ce8d24e0600d1fa85524c10287f6004fbab87672bf3e1e"}, + {file = "charset_normalizer-3.3.2-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:f27273b60488abe721a075bcca6d7f3964f9f6f067c8c4c605743023d7d3944f"}, + {file = "charset_normalizer-3.3.2-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:1ceae2f17a9c33cb48e3263960dc5fc8005351ee19db217e9b1bb15d28c02574"}, + {file = "charset_normalizer-3.3.2-cp311-cp311-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:65f6f63034100ead094b8744b3b97965785388f308a64cf8d7c34f2f2e5be0c4"}, + {file = "charset_normalizer-3.3.2-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:753f10e867343b4511128c6ed8c82f7bec3bd026875576dfd88483c5c73b2fd8"}, + {file = "charset_normalizer-3.3.2-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:4a78b2b446bd7c934f5dcedc588903fb2f5eec172f3d29e52a9096a43722adfc"}, + {file = "charset_normalizer-3.3.2-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:e537484df0d8f426ce2afb2d0f8e1c3d0b114b83f8850e5f2fbea0e797bd82ae"}, + {file = "charset_normalizer-3.3.2-cp311-cp311-musllinux_1_1_i686.whl", hash = "sha256:eb6904c354526e758fda7167b33005998fb68c46fbc10e013ca97f21ca5c8887"}, + {file = "charset_normalizer-3.3.2-cp311-cp311-musllinux_1_1_ppc64le.whl", hash = "sha256:deb6be0ac38ece9ba87dea880e438f25ca3eddfac8b002a2ec3d9183a454e8ae"}, + {file = "charset_normalizer-3.3.2-cp311-cp311-musllinux_1_1_s390x.whl", hash = "sha256:4ab2fe47fae9e0f9dee8c04187ce5d09f48eabe611be8259444906793ab7cbce"}, + {file = "charset_normalizer-3.3.2-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:80402cd6ee291dcb72644d6eac93785fe2c8b9cb30893c1af5b8fdd753b9d40f"}, + {file = "charset_normalizer-3.3.2-cp311-cp311-win32.whl", hash = "sha256:7cd13a2e3ddeed6913a65e66e94b51d80a041145a026c27e6bb76c31a853c6ab"}, + {file = "charset_normalizer-3.3.2-cp311-cp311-win_amd64.whl", hash = "sha256:663946639d296df6a2bb2aa51b60a2454ca1cb29835324c640dafb5ff2131a77"}, + {file = "charset_normalizer-3.3.2-cp312-cp312-macosx_10_9_universal2.whl", hash = "sha256:0b2b64d2bb6d3fb9112bafa732def486049e63de9618b5843bcdd081d8144cd8"}, + {file = "charset_normalizer-3.3.2-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:ddbb2551d7e0102e7252db79ba445cdab71b26640817ab1e3e3648dad515003b"}, + {file = "charset_normalizer-3.3.2-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:55086ee1064215781fff39a1af09518bc9255b50d6333f2e4c74ca09fac6a8f6"}, + {file = "charset_normalizer-3.3.2-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:8f4a014bc36d3c57402e2977dada34f9c12300af536839dc38c0beab8878f38a"}, + {file = "charset_normalizer-3.3.2-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:a10af20b82360ab00827f916a6058451b723b4e65030c5a18577c8b2de5b3389"}, + {file = "charset_normalizer-3.3.2-cp312-cp312-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:8d756e44e94489e49571086ef83b2bb8ce311e730092d2c34ca8f7d925cb20aa"}, + {file = "charset_normalizer-3.3.2-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:90d558489962fd4918143277a773316e56c72da56ec7aa3dc3dbbe20fdfed15b"}, + {file = "charset_normalizer-3.3.2-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:6ac7ffc7ad6d040517be39eb591cac5ff87416c2537df6ba3cba3bae290c0fed"}, + {file = "charset_normalizer-3.3.2-cp312-cp312-musllinux_1_1_aarch64.whl", hash = "sha256:7ed9e526742851e8d5cc9e6cf41427dfc6068d4f5a3bb03659444b4cabf6bc26"}, + {file = "charset_normalizer-3.3.2-cp312-cp312-musllinux_1_1_i686.whl", hash = "sha256:8bdb58ff7ba23002a4c5808d608e4e6c687175724f54a5dade5fa8c67b604e4d"}, + {file = "charset_normalizer-3.3.2-cp312-cp312-musllinux_1_1_ppc64le.whl", hash = "sha256:6b3251890fff30ee142c44144871185dbe13b11bab478a88887a639655be1068"}, + {file = "charset_normalizer-3.3.2-cp312-cp312-musllinux_1_1_s390x.whl", hash = "sha256:b4a23f61ce87adf89be746c8a8974fe1c823c891d8f86eb218bb957c924bb143"}, + {file = "charset_normalizer-3.3.2-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:efcb3f6676480691518c177e3b465bcddf57cea040302f9f4e6e191af91174d4"}, + {file = "charset_normalizer-3.3.2-cp312-cp312-win32.whl", hash = "sha256:d965bba47ddeec8cd560687584e88cf699fd28f192ceb452d1d7ee807c5597b7"}, + {file = "charset_normalizer-3.3.2-cp312-cp312-win_amd64.whl", hash = "sha256:96b02a3dc4381e5494fad39be677abcb5e6634bf7b4fa83a6dd3112607547001"}, + {file = "charset_normalizer-3.3.2-cp37-cp37m-macosx_10_9_x86_64.whl", hash = "sha256:95f2a5796329323b8f0512e09dbb7a1860c46a39da62ecb2324f116fa8fdc85c"}, + {file = "charset_normalizer-3.3.2-cp37-cp37m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:c002b4ffc0be611f0d9da932eb0f704fe2602a9a949d1f738e4c34c75b0863d5"}, + {file = "charset_normalizer-3.3.2-cp37-cp37m-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:a981a536974bbc7a512cf44ed14938cf01030a99e9b3a06dd59578882f06f985"}, + {file = "charset_normalizer-3.3.2-cp37-cp37m-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:3287761bc4ee9e33561a7e058c72ac0938c4f57fe49a09eae428fd88aafe7bb6"}, + {file = "charset_normalizer-3.3.2-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:42cb296636fcc8b0644486d15c12376cb9fa75443e00fb25de0b8602e64c1714"}, + {file = "charset_normalizer-3.3.2-cp37-cp37m-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:0a55554a2fa0d408816b3b5cedf0045f4b8e1a6065aec45849de2d6f3f8e9786"}, + {file = "charset_normalizer-3.3.2-cp37-cp37m-musllinux_1_1_aarch64.whl", hash = "sha256:c083af607d2515612056a31f0a8d9e0fcb5876b7bfc0abad3ecd275bc4ebc2d5"}, + {file = "charset_normalizer-3.3.2-cp37-cp37m-musllinux_1_1_i686.whl", hash = "sha256:87d1351268731db79e0f8e745d92493ee2841c974128ef629dc518b937d9194c"}, + {file = "charset_normalizer-3.3.2-cp37-cp37m-musllinux_1_1_ppc64le.whl", hash = "sha256:bd8f7df7d12c2db9fab40bdd87a7c09b1530128315d047a086fa3ae3435cb3a8"}, + {file = "charset_normalizer-3.3.2-cp37-cp37m-musllinux_1_1_s390x.whl", hash = "sha256:c180f51afb394e165eafe4ac2936a14bee3eb10debc9d9e4db8958fe36afe711"}, + {file = "charset_normalizer-3.3.2-cp37-cp37m-musllinux_1_1_x86_64.whl", hash = "sha256:8c622a5fe39a48f78944a87d4fb8a53ee07344641b0562c540d840748571b811"}, + {file = "charset_normalizer-3.3.2-cp37-cp37m-win32.whl", hash = "sha256:db364eca23f876da6f9e16c9da0df51aa4f104a972735574842618b8c6d999d4"}, + {file = "charset_normalizer-3.3.2-cp37-cp37m-win_amd64.whl", hash = "sha256:86216b5cee4b06df986d214f664305142d9c76df9b6512be2738aa72a2048f99"}, + {file = "charset_normalizer-3.3.2-cp38-cp38-macosx_10_9_universal2.whl", hash = "sha256:6463effa3186ea09411d50efc7d85360b38d5f09b870c48e4600f63af490e56a"}, + {file = "charset_normalizer-3.3.2-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:6c4caeef8fa63d06bd437cd4bdcf3ffefe6738fb1b25951440d80dc7df8c03ac"}, + {file = "charset_normalizer-3.3.2-cp38-cp38-macosx_11_0_arm64.whl", hash = "sha256:37e55c8e51c236f95b033f6fb391d7d7970ba5fe7ff453dad675e88cf303377a"}, + {file = "charset_normalizer-3.3.2-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:fb69256e180cb6c8a894fee62b3afebae785babc1ee98b81cdf68bbca1987f33"}, + {file = "charset_normalizer-3.3.2-cp38-cp38-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:ae5f4161f18c61806f411a13b0310bea87f987c7d2ecdbdaad0e94eb2e404238"}, + {file = "charset_normalizer-3.3.2-cp38-cp38-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:b2b0a0c0517616b6869869f8c581d4eb2dd83a4d79e0ebcb7d373ef9956aeb0a"}, + {file = "charset_normalizer-3.3.2-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:45485e01ff4d3630ec0d9617310448a8702f70e9c01906b0d0118bdf9d124cf2"}, + {file = "charset_normalizer-3.3.2-cp38-cp38-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:eb00ed941194665c332bf8e078baf037d6c35d7c4f3102ea2d4f16ca94a26dc8"}, + {file = "charset_normalizer-3.3.2-cp38-cp38-musllinux_1_1_aarch64.whl", hash = "sha256:2127566c664442652f024c837091890cb1942c30937add288223dc895793f898"}, + {file = "charset_normalizer-3.3.2-cp38-cp38-musllinux_1_1_i686.whl", hash = "sha256:a50aebfa173e157099939b17f18600f72f84eed3049e743b68ad15bd69b6bf99"}, + {file = "charset_normalizer-3.3.2-cp38-cp38-musllinux_1_1_ppc64le.whl", hash = "sha256:4d0d1650369165a14e14e1e47b372cfcb31d6ab44e6e33cb2d4e57265290044d"}, + {file = "charset_normalizer-3.3.2-cp38-cp38-musllinux_1_1_s390x.whl", hash = "sha256:923c0c831b7cfcb071580d3f46c4baf50f174be571576556269530f4bbd79d04"}, + {file = "charset_normalizer-3.3.2-cp38-cp38-musllinux_1_1_x86_64.whl", hash = "sha256:06a81e93cd441c56a9b65d8e1d043daeb97a3d0856d177d5c90ba85acb3db087"}, + {file = "charset_normalizer-3.3.2-cp38-cp38-win32.whl", hash = "sha256:6ef1d82a3af9d3eecdba2321dc1b3c238245d890843e040e41e470ffa64c3e25"}, + {file = "charset_normalizer-3.3.2-cp38-cp38-win_amd64.whl", hash = "sha256:eb8821e09e916165e160797a6c17edda0679379a4be5c716c260e836e122f54b"}, + {file = "charset_normalizer-3.3.2-cp39-cp39-macosx_10_9_universal2.whl", hash = "sha256:c235ebd9baae02f1b77bcea61bce332cb4331dc3617d254df3323aa01ab47bd4"}, + {file = "charset_normalizer-3.3.2-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:5b4c145409bef602a690e7cfad0a15a55c13320ff7a3ad7ca59c13bb8ba4d45d"}, + {file = "charset_normalizer-3.3.2-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:68d1f8a9e9e37c1223b656399be5d6b448dea850bed7d0f87a8311f1ff3dabb0"}, + {file = "charset_normalizer-3.3.2-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:22afcb9f253dac0696b5a4be4a1c0f8762f8239e21b99680099abd9b2b1b2269"}, + {file = "charset_normalizer-3.3.2-cp39-cp39-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:e27ad930a842b4c5eb8ac0016b0a54f5aebbe679340c26101df33424142c143c"}, + {file = "charset_normalizer-3.3.2-cp39-cp39-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:1f79682fbe303db92bc2b1136016a38a42e835d932bab5b3b1bfcfbf0640e519"}, + {file = "charset_normalizer-3.3.2-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:b261ccdec7821281dade748d088bb6e9b69e6d15b30652b74cbbac25e280b796"}, + {file = "charset_normalizer-3.3.2-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:122c7fa62b130ed55f8f285bfd56d5f4b4a5b503609d181f9ad85e55c89f4185"}, + {file = "charset_normalizer-3.3.2-cp39-cp39-musllinux_1_1_aarch64.whl", hash = "sha256:d0eccceffcb53201b5bfebb52600a5fb483a20b61da9dbc885f8b103cbe7598c"}, + {file = "charset_normalizer-3.3.2-cp39-cp39-musllinux_1_1_i686.whl", hash = "sha256:9f96df6923e21816da7e0ad3fd47dd8f94b2a5ce594e00677c0013018b813458"}, + {file = "charset_normalizer-3.3.2-cp39-cp39-musllinux_1_1_ppc64le.whl", hash = "sha256:7f04c839ed0b6b98b1a7501a002144b76c18fb1c1850c8b98d458ac269e26ed2"}, + {file = "charset_normalizer-3.3.2-cp39-cp39-musllinux_1_1_s390x.whl", hash = "sha256:34d1c8da1e78d2e001f363791c98a272bb734000fcef47a491c1e3b0505657a8"}, + {file = "charset_normalizer-3.3.2-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:ff8fa367d09b717b2a17a052544193ad76cd49979c805768879cb63d9ca50561"}, + {file = "charset_normalizer-3.3.2-cp39-cp39-win32.whl", hash = "sha256:aed38f6e4fb3f5d6bf81bfa990a07806be9d83cf7bacef998ab1a9bd660a581f"}, + {file = "charset_normalizer-3.3.2-cp39-cp39-win_amd64.whl", hash = "sha256:b01b88d45a6fcb69667cd6d2f7a9aeb4bf53760d7fc536bf679ec94fe9f3ff3d"}, + {file = "charset_normalizer-3.3.2-py3-none-any.whl", hash = "sha256:3e4d1f6587322d2788836a99c69062fbb091331ec940e02d12d179c1d53e25fc"}, +] + +[[package]] +name = "colorama" +version = "0.4.6" +description = "Cross-platform colored terminal text." +optional = false +python-versions = "!=3.0.*,!=3.1.*,!=3.2.*,!=3.3.*,!=3.4.*,!=3.5.*,!=3.6.*,>=2.7" +files = [ + {file = "colorama-0.4.6-py2.py3-none-any.whl", hash = "sha256:4f1d9991f5acc0ca119f9d443620b77f9d6b33703e51011c16baf57afb285fc6"}, + {file = "colorama-0.4.6.tar.gz", hash = "sha256:08695f5cb7ed6e0531a20572697297273c47b8cae5a63ffc6d6ed5c201be6e44"}, +] + +[[package]] +name = "filelock" +version = "3.13.1" +description = "A platform independent file lock." +optional = false +python-versions = ">=3.8" +files = [ + {file = "filelock-3.13.1-py3-none-any.whl", hash = "sha256:57dbda9b35157b05fb3e58ee91448612eb674172fab98ee235ccb0b5bee19a1c"}, + {file = "filelock-3.13.1.tar.gz", hash = "sha256:521f5f56c50f8426f5e03ad3b281b490a87ef15bc6c526f168290f0c7148d44e"}, +] + +[package.extras] +docs = ["furo (>=2023.9.10)", "sphinx (>=7.2.6)", "sphinx-autodoc-typehints (>=1.24)"] +testing = ["covdefaults (>=2.3)", "coverage (>=7.3.2)", "diff-cover (>=8)", "pytest (>=7.4.3)", "pytest-cov (>=4.1)", "pytest-mock (>=3.12)", "pytest-timeout (>=2.2)"] +typing = ["typing-extensions (>=4.8)"] + +[[package]] +name = "fsspec" +version = "2024.2.0" +description = "File-system specification" +optional = false +python-versions = ">=3.8" +files = [ + {file = "fsspec-2024.2.0-py3-none-any.whl", hash = "sha256:817f969556fa5916bc682e02ca2045f96ff7f586d45110fcb76022063ad2c7d8"}, + {file = "fsspec-2024.2.0.tar.gz", hash = "sha256:b6ad1a679f760dda52b1168c859d01b7b80648ea6f7f7c7f5a8a91dc3f3ecb84"}, +] + +[package.extras] +abfs = ["adlfs"] +adl = ["adlfs"] +arrow = ["pyarrow (>=1)"] +dask = ["dask", "distributed"] +devel = ["pytest", "pytest-cov"] +dropbox = ["dropbox", "dropboxdrivefs", "requests"] +full = ["adlfs", "aiohttp (!=4.0.0a0,!=4.0.0a1)", "dask", "distributed", "dropbox", "dropboxdrivefs", "fusepy", "gcsfs", "libarchive-c", "ocifs", "panel", "paramiko", "pyarrow (>=1)", "pygit2", "requests", "s3fs", "smbprotocol", "tqdm"] +fuse = ["fusepy"] +gcs = ["gcsfs"] +git = ["pygit2"] +github = ["requests"] +gs = ["gcsfs"] +gui = ["panel"] +hdfs = ["pyarrow (>=1)"] +http = ["aiohttp (!=4.0.0a0,!=4.0.0a1)"] +libarchive = ["libarchive-c"] +oci = ["ocifs"] +s3 = ["s3fs"] +sftp = ["paramiko"] +smb = ["smbprotocol"] +ssh = ["paramiko"] +tqdm = ["tqdm"] + +[[package]] +name = "gguf" +version = "0.7.0" +description = "Read and write ML models in GGUF for GGML" +optional = false +python-versions = ">=3.8" +files = [] +develop = false + +[package.dependencies] +numpy = ">=1.17" + +[package.source] +type = "directory" +url = "gguf-py" + +[[package]] +name = "huggingface-hub" +version = "0.20.3" +description = "Client library to download and publish models, datasets and other repos on the huggingface.co hub" +optional = false +python-versions = ">=3.8.0" +files = [ + {file = "huggingface_hub-0.20.3-py3-none-any.whl", hash = "sha256:d988ae4f00d3e307b0c80c6a05ca6dbb7edba8bba3079f74cda7d9c2e562a7b6"}, + {file = "huggingface_hub-0.20.3.tar.gz", hash = "sha256:94e7f8e074475fbc67d6a71957b678e1b4a74ff1b64a644fd6cbb83da962d05d"}, +] + +[package.dependencies] +filelock = "*" +fsspec = ">=2023.5.0" +packaging = ">=20.9" +pyyaml = ">=5.1" +requests = "*" +tqdm = ">=4.42.1" +typing-extensions = ">=3.7.4.3" + +[package.extras] +all = ["InquirerPy (==0.3.4)", "Jinja2", "Pillow", "aiohttp", "gradio", "jedi", "mypy (==1.5.1)", "numpy", "pydantic (>1.1,<2.0)", "pydantic (>1.1,<3.0)", "pytest", "pytest-asyncio", "pytest-cov", "pytest-env", "pytest-rerunfailures", "pytest-vcr", "pytest-xdist", "ruff (>=0.1.3)", "soundfile", "types-PyYAML", "types-requests", "types-simplejson", "types-toml", "types-tqdm", "types-urllib3", "typing-extensions (>=4.8.0)", "urllib3 (<2.0)"] +cli = ["InquirerPy (==0.3.4)"] +dev = ["InquirerPy (==0.3.4)", "Jinja2", "Pillow", "aiohttp", "gradio", "jedi", "mypy (==1.5.1)", "numpy", "pydantic (>1.1,<2.0)", "pydantic (>1.1,<3.0)", "pytest", "pytest-asyncio", "pytest-cov", "pytest-env", "pytest-rerunfailures", "pytest-vcr", "pytest-xdist", "ruff (>=0.1.3)", "soundfile", "types-PyYAML", "types-requests", "types-simplejson", "types-toml", "types-tqdm", "types-urllib3", "typing-extensions (>=4.8.0)", "urllib3 (<2.0)"] +fastai = ["fastai (>=2.4)", "fastcore (>=1.3.27)", "toml"] +inference = ["aiohttp", "pydantic (>1.1,<2.0)", "pydantic (>1.1,<3.0)"] +quality = ["mypy (==1.5.1)", "ruff (>=0.1.3)"] +tensorflow = ["graphviz", "pydot", "tensorflow"] +testing = ["InquirerPy (==0.3.4)", "Jinja2", "Pillow", "aiohttp", "gradio", "jedi", "numpy", "pydantic (>1.1,<2.0)", "pydantic (>1.1,<3.0)", "pytest", "pytest-asyncio", "pytest-cov", "pytest-env", "pytest-rerunfailures", "pytest-vcr", "pytest-xdist", "soundfile", "urllib3 (<2.0)"] +torch = ["torch"] +typing = ["types-PyYAML", "types-requests", "types-simplejson", "types-toml", "types-tqdm", "types-urllib3", "typing-extensions (>=4.8.0)"] + +[[package]] +name = "idna" +version = "3.6" +description = "Internationalized Domain Names in Applications (IDNA)" +optional = false +python-versions = ">=3.5" +files = [ + {file = "idna-3.6-py3-none-any.whl", hash = "sha256:c05567e9c24a6b9faaa835c4821bad0590fbb9d5779e7caa6e1cc4978e7eb24f"}, + {file = "idna-3.6.tar.gz", hash = "sha256:9ecdbbd083b06798ae1e86adcbfe8ab1479cf864e4ee30fe4e46a003d12491ca"}, +] + +[[package]] +name = "jinja2" +version = "3.1.3" +description = "A very fast and expressive template engine." +optional = false +python-versions = ">=3.7" +files = [ + {file = "Jinja2-3.1.3-py3-none-any.whl", hash = "sha256:7d6d50dd97d52cbc355597bd845fabfbac3f551e1f99619e39a35ce8c370b5fa"}, + {file = "Jinja2-3.1.3.tar.gz", hash = "sha256:ac8bd6544d4bb2c9792bf3a159e80bba8fda7f07e81bc3aed565432d5925ba90"}, +] + +[package.dependencies] +MarkupSafe = ">=2.0" + +[package.extras] +i18n = ["Babel (>=2.7)"] + +[[package]] +name = "markupsafe" +version = "2.1.5" +description = "Safely add untrusted strings to HTML/XML markup." +optional = false +python-versions = ">=3.7" +files = [ + {file = "MarkupSafe-2.1.5-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:a17a92de5231666cfbe003f0e4b9b3a7ae3afb1ec2845aadc2bacc93ff85febc"}, + {file = "MarkupSafe-2.1.5-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:72b6be590cc35924b02c78ef34b467da4ba07e4e0f0454a2c5907f473fc50ce5"}, + {file = "MarkupSafe-2.1.5-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:e61659ba32cf2cf1481e575d0462554625196a1f2fc06a1c777d3f48e8865d46"}, + {file = "MarkupSafe-2.1.5-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:2174c595a0d73a3080ca3257b40096db99799265e1c27cc5a610743acd86d62f"}, + {file = "MarkupSafe-2.1.5-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:ae2ad8ae6ebee9d2d94b17fb62763125f3f374c25618198f40cbb8b525411900"}, + {file = "MarkupSafe-2.1.5-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:075202fa5b72c86ad32dc7d0b56024ebdbcf2048c0ba09f1cde31bfdd57bcfff"}, + {file = "MarkupSafe-2.1.5-cp310-cp310-musllinux_1_1_i686.whl", hash = "sha256:598e3276b64aff0e7b3451b72e94fa3c238d452e7ddcd893c3ab324717456bad"}, + {file = "MarkupSafe-2.1.5-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:fce659a462a1be54d2ffcacea5e3ba2d74daa74f30f5f143fe0c58636e355fdd"}, + {file = "MarkupSafe-2.1.5-cp310-cp310-win32.whl", hash = "sha256:d9fad5155d72433c921b782e58892377c44bd6252b5af2f67f16b194987338a4"}, + {file = "MarkupSafe-2.1.5-cp310-cp310-win_amd64.whl", hash = "sha256:bf50cd79a75d181c9181df03572cdce0fbb75cc353bc350712073108cba98de5"}, + {file = "MarkupSafe-2.1.5-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:629ddd2ca402ae6dbedfceeba9c46d5f7b2a61d9749597d4307f943ef198fc1f"}, + {file = "MarkupSafe-2.1.5-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:5b7b716f97b52c5a14bffdf688f971b2d5ef4029127f1ad7a513973cfd818df2"}, + {file = "MarkupSafe-2.1.5-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:6ec585f69cec0aa07d945b20805be741395e28ac1627333b1c5b0105962ffced"}, + {file = "MarkupSafe-2.1.5-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:b91c037585eba9095565a3556f611e3cbfaa42ca1e865f7b8015fe5c7336d5a5"}, + {file = "MarkupSafe-2.1.5-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:7502934a33b54030eaf1194c21c692a534196063db72176b0c4028e140f8f32c"}, + {file = "MarkupSafe-2.1.5-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:0e397ac966fdf721b2c528cf028494e86172b4feba51d65f81ffd65c63798f3f"}, + {file = "MarkupSafe-2.1.5-cp311-cp311-musllinux_1_1_i686.whl", hash = "sha256:c061bb86a71b42465156a3ee7bd58c8c2ceacdbeb95d05a99893e08b8467359a"}, + {file = "MarkupSafe-2.1.5-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:3a57fdd7ce31c7ff06cdfbf31dafa96cc533c21e443d57f5b1ecc6cdc668ec7f"}, + {file = "MarkupSafe-2.1.5-cp311-cp311-win32.whl", hash = "sha256:397081c1a0bfb5124355710fe79478cdbeb39626492b15d399526ae53422b906"}, + {file = "MarkupSafe-2.1.5-cp311-cp311-win_amd64.whl", hash = "sha256:2b7c57a4dfc4f16f7142221afe5ba4e093e09e728ca65c51f5620c9aaeb9a617"}, + {file = "MarkupSafe-2.1.5-cp312-cp312-macosx_10_9_universal2.whl", hash = "sha256:8dec4936e9c3100156f8a2dc89c4b88d5c435175ff03413b443469c7c8c5f4d1"}, + {file = "MarkupSafe-2.1.5-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:3c6b973f22eb18a789b1460b4b91bf04ae3f0c4234a0a6aa6b0a92f6f7b951d4"}, + {file = "MarkupSafe-2.1.5-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:ac07bad82163452a6884fe8fa0963fb98c2346ba78d779ec06bd7a6262132aee"}, + {file = "MarkupSafe-2.1.5-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:f5dfb42c4604dddc8e4305050aa6deb084540643ed5804d7455b5df8fe16f5e5"}, + {file = "MarkupSafe-2.1.5-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:ea3d8a3d18833cf4304cd2fc9cbb1efe188ca9b5efef2bdac7adc20594a0e46b"}, + {file = "MarkupSafe-2.1.5-cp312-cp312-musllinux_1_1_aarch64.whl", hash = "sha256:d050b3361367a06d752db6ead6e7edeb0009be66bc3bae0ee9d97fb326badc2a"}, + {file = "MarkupSafe-2.1.5-cp312-cp312-musllinux_1_1_i686.whl", hash = "sha256:bec0a414d016ac1a18862a519e54b2fd0fc8bbfd6890376898a6c0891dd82e9f"}, + {file = "MarkupSafe-2.1.5-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:58c98fee265677f63a4385256a6d7683ab1832f3ddd1e66fe948d5880c21a169"}, + {file = "MarkupSafe-2.1.5-cp312-cp312-win32.whl", hash = "sha256:8590b4ae07a35970728874632fed7bd57b26b0102df2d2b233b6d9d82f6c62ad"}, + {file = "MarkupSafe-2.1.5-cp312-cp312-win_amd64.whl", hash = "sha256:823b65d8706e32ad2df51ed89496147a42a2a6e01c13cfb6ffb8b1e92bc910bb"}, + {file = "MarkupSafe-2.1.5-cp37-cp37m-macosx_10_9_x86_64.whl", hash = "sha256:c8b29db45f8fe46ad280a7294f5c3ec36dbac9491f2d1c17345be8e69cc5928f"}, + {file = "MarkupSafe-2.1.5-cp37-cp37m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:ec6a563cff360b50eed26f13adc43e61bc0c04d94b8be985e6fb24b81f6dcfdf"}, + {file = "MarkupSafe-2.1.5-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a549b9c31bec33820e885335b451286e2969a2d9e24879f83fe904a5ce59d70a"}, + {file = "MarkupSafe-2.1.5-cp37-cp37m-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:4f11aa001c540f62c6166c7726f71f7573b52c68c31f014c25cc7901deea0b52"}, + {file = "MarkupSafe-2.1.5-cp37-cp37m-musllinux_1_1_aarch64.whl", hash = "sha256:7b2e5a267c855eea6b4283940daa6e88a285f5f2a67f2220203786dfa59b37e9"}, + {file = "MarkupSafe-2.1.5-cp37-cp37m-musllinux_1_1_i686.whl", hash = "sha256:2d2d793e36e230fd32babe143b04cec8a8b3eb8a3122d2aceb4a371e6b09b8df"}, + {file = "MarkupSafe-2.1.5-cp37-cp37m-musllinux_1_1_x86_64.whl", hash = "sha256:ce409136744f6521e39fd8e2a24c53fa18ad67aa5bc7c2cf83645cce5b5c4e50"}, + {file = "MarkupSafe-2.1.5-cp37-cp37m-win32.whl", hash = "sha256:4096e9de5c6fdf43fb4f04c26fb114f61ef0bf2e5604b6ee3019d51b69e8c371"}, + {file = "MarkupSafe-2.1.5-cp37-cp37m-win_amd64.whl", hash = "sha256:4275d846e41ecefa46e2015117a9f491e57a71ddd59bbead77e904dc02b1bed2"}, + {file = "MarkupSafe-2.1.5-cp38-cp38-macosx_10_9_universal2.whl", hash = "sha256:656f7526c69fac7f600bd1f400991cc282b417d17539a1b228617081106feb4a"}, + {file = "MarkupSafe-2.1.5-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:97cafb1f3cbcd3fd2b6fbfb99ae11cdb14deea0736fc2b0952ee177f2b813a46"}, + {file = "MarkupSafe-2.1.5-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:1f3fbcb7ef1f16e48246f704ab79d79da8a46891e2da03f8783a5b6fa41a9532"}, + {file = "MarkupSafe-2.1.5-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:fa9db3f79de01457b03d4f01b34cf91bc0048eb2c3846ff26f66687c2f6d16ab"}, + {file = "MarkupSafe-2.1.5-cp38-cp38-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:ffee1f21e5ef0d712f9033568f8344d5da8cc2869dbd08d87c84656e6a2d2f68"}, + {file = "MarkupSafe-2.1.5-cp38-cp38-musllinux_1_1_aarch64.whl", hash = "sha256:5dedb4db619ba5a2787a94d877bc8ffc0566f92a01c0ef214865e54ecc9ee5e0"}, + {file = "MarkupSafe-2.1.5-cp38-cp38-musllinux_1_1_i686.whl", hash = "sha256:30b600cf0a7ac9234b2638fbc0fb6158ba5bdcdf46aeb631ead21248b9affbc4"}, + {file = "MarkupSafe-2.1.5-cp38-cp38-musllinux_1_1_x86_64.whl", hash = "sha256:8dd717634f5a044f860435c1d8c16a270ddf0ef8588d4887037c5028b859b0c3"}, + {file = "MarkupSafe-2.1.5-cp38-cp38-win32.whl", hash = "sha256:daa4ee5a243f0f20d528d939d06670a298dd39b1ad5f8a72a4275124a7819eff"}, + {file = "MarkupSafe-2.1.5-cp38-cp38-win_amd64.whl", hash = "sha256:619bc166c4f2de5caa5a633b8b7326fbe98e0ccbfacabd87268a2b15ff73a029"}, + {file = "MarkupSafe-2.1.5-cp39-cp39-macosx_10_9_universal2.whl", hash = "sha256:7a68b554d356a91cce1236aa7682dc01df0edba8d043fd1ce607c49dd3c1edcf"}, + {file = "MarkupSafe-2.1.5-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:db0b55e0f3cc0be60c1f19efdde9a637c32740486004f20d1cff53c3c0ece4d2"}, + {file = "MarkupSafe-2.1.5-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:3e53af139f8579a6d5f7b76549125f0d94d7e630761a2111bc431fd820e163b8"}, + {file = "MarkupSafe-2.1.5-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:17b950fccb810b3293638215058e432159d2b71005c74371d784862b7e4683f3"}, + {file = "MarkupSafe-2.1.5-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:4c31f53cdae6ecfa91a77820e8b151dba54ab528ba65dfd235c80b086d68a465"}, + {file = "MarkupSafe-2.1.5-cp39-cp39-musllinux_1_1_aarch64.whl", hash = "sha256:bff1b4290a66b490a2f4719358c0cdcd9bafb6b8f061e45c7a2460866bf50c2e"}, + {file = "MarkupSafe-2.1.5-cp39-cp39-musllinux_1_1_i686.whl", hash = "sha256:bc1667f8b83f48511b94671e0e441401371dfd0f0a795c7daa4a3cd1dde55bea"}, + {file = "MarkupSafe-2.1.5-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:5049256f536511ee3f7e1b3f87d1d1209d327e818e6ae1365e8653d7e3abb6a6"}, + {file = "MarkupSafe-2.1.5-cp39-cp39-win32.whl", hash = "sha256:00e046b6dd71aa03a41079792f8473dc494d564611a8f89bbbd7cb93295ebdcf"}, + {file = "MarkupSafe-2.1.5-cp39-cp39-win_amd64.whl", hash = "sha256:fa173ec60341d6bb97a89f5ea19c85c5643c1e7dedebc22f5181eb73573142c5"}, + {file = "MarkupSafe-2.1.5.tar.gz", hash = "sha256:d283d37a890ba4c1ae73ffadf8046435c76e7bc2247bbb63c00bd1a709c6544b"}, +] + +[[package]] +name = "more-itertools" +version = "10.2.0" +description = "More routines for operating on iterables, beyond itertools" +optional = false +python-versions = ">=3.8" +files = [ + {file = "more-itertools-10.2.0.tar.gz", hash = "sha256:8fccb480c43d3e99a00087634c06dd02b0d50fbf088b380de5a41a015ec239e1"}, + {file = "more_itertools-10.2.0-py3-none-any.whl", hash = "sha256:686b06abe565edfab151cb8fd385a05651e1fdf8f0a14191e4439283421f8684"}, +] + +[[package]] +name = "mpmath" +version = "1.3.0" +description = "Python library for arbitrary-precision floating-point arithmetic" +optional = false +python-versions = "*" +files = [ + {file = "mpmath-1.3.0-py3-none-any.whl", hash = "sha256:a0b2b9fe80bbcd81a6647ff13108738cfb482d481d826cc0e02f5b35e5c88d2c"}, + {file = "mpmath-1.3.0.tar.gz", hash = "sha256:7a28eb2a9774d00c7bc92411c19a89209d5da7c4c9a9e227be8330a23a25b91f"}, +] + +[package.extras] +develop = ["codecov", "pycodestyle", "pytest (>=4.6)", "pytest-cov", "wheel"] +docs = ["sphinx"] +gmpy = ["gmpy2 (>=2.1.0a4)"] +tests = ["pytest (>=4.6)"] + +[[package]] +name = "networkx" +version = "3.2.1" +description = "Python package for creating and manipulating graphs and networks" +optional = false +python-versions = ">=3.9" +files = [ + {file = "networkx-3.2.1-py3-none-any.whl", hash = "sha256:f18c69adc97877c42332c170849c96cefa91881c99a7cb3e95b7c659ebdc1ec2"}, + {file = "networkx-3.2.1.tar.gz", hash = "sha256:9f1bb5cf3409bf324e0a722c20bdb4c20ee39bf1c30ce8ae499c8502b0b5e0c6"}, +] + +[package.extras] +default = ["matplotlib (>=3.5)", "numpy (>=1.22)", "pandas (>=1.4)", "scipy (>=1.9,!=1.11.0,!=1.11.1)"] +developer = ["changelist (==0.4)", "mypy (>=1.1)", "pre-commit (>=3.2)", "rtoml"] +doc = ["nb2plots (>=0.7)", "nbconvert (<7.9)", "numpydoc (>=1.6)", "pillow (>=9.4)", "pydata-sphinx-theme (>=0.14)", "sphinx (>=7)", "sphinx-gallery (>=0.14)", "texext (>=0.6.7)"] +extra = ["lxml (>=4.6)", "pydot (>=1.4.2)", "pygraphviz (>=1.11)", "sympy (>=1.10)"] +test = ["pytest (>=7.2)", "pytest-cov (>=4.0)"] + +[[package]] +name = "numpy" +version = "1.26.4" +description = "Fundamental package for array computing in Python" +optional = false +python-versions = ">=3.9" +files = [ + {file = "numpy-1.26.4-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:9ff0f4f29c51e2803569d7a51c2304de5554655a60c5d776e35b4a41413830d0"}, + {file = "numpy-1.26.4-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:2e4ee3380d6de9c9ec04745830fd9e2eccb3e6cf790d39d7b98ffd19b0dd754a"}, + {file = "numpy-1.26.4-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d209d8969599b27ad20994c8e41936ee0964e6da07478d6c35016bc386b66ad4"}, + {file = "numpy-1.26.4-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ffa75af20b44f8dba823498024771d5ac50620e6915abac414251bd971b4529f"}, + {file = "numpy-1.26.4-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:62b8e4b1e28009ef2846b4c7852046736bab361f7aeadeb6a5b89ebec3c7055a"}, + {file = "numpy-1.26.4-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:a4abb4f9001ad2858e7ac189089c42178fcce737e4169dc61321660f1a96c7d2"}, + {file = "numpy-1.26.4-cp310-cp310-win32.whl", hash = "sha256:bfe25acf8b437eb2a8b2d49d443800a5f18508cd811fea3181723922a8a82b07"}, + {file = "numpy-1.26.4-cp310-cp310-win_amd64.whl", hash = "sha256:b97fe8060236edf3662adfc2c633f56a08ae30560c56310562cb4f95500022d5"}, + {file = "numpy-1.26.4-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:4c66707fabe114439db9068ee468c26bbdf909cac0fb58686a42a24de1760c71"}, + {file = "numpy-1.26.4-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:edd8b5fe47dab091176d21bb6de568acdd906d1887a4584a15a9a96a1dca06ef"}, + {file = "numpy-1.26.4-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:7ab55401287bfec946ced39700c053796e7cc0e3acbef09993a9ad2adba6ca6e"}, + {file = "numpy-1.26.4-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:666dbfb6ec68962c033a450943ded891bed2d54e6755e35e5835d63f4f6931d5"}, + {file = "numpy-1.26.4-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:96ff0b2ad353d8f990b63294c8986f1ec3cb19d749234014f4e7eb0112ceba5a"}, + {file = "numpy-1.26.4-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:60dedbb91afcbfdc9bc0b1f3f402804070deed7392c23eb7a7f07fa857868e8a"}, + {file = "numpy-1.26.4-cp311-cp311-win32.whl", hash = "sha256:1af303d6b2210eb850fcf03064d364652b7120803a0b872f5211f5234b399f20"}, + {file = "numpy-1.26.4-cp311-cp311-win_amd64.whl", hash = "sha256:cd25bcecc4974d09257ffcd1f098ee778f7834c3ad767fe5db785be9a4aa9cb2"}, + {file = "numpy-1.26.4-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:b3ce300f3644fb06443ee2222c2201dd3a89ea6040541412b8fa189341847218"}, + {file = "numpy-1.26.4-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:03a8c78d01d9781b28a6989f6fa1bb2c4f2d51201cf99d3dd875df6fbd96b23b"}, + {file = "numpy-1.26.4-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:9fad7dcb1aac3c7f0584a5a8133e3a43eeb2fe127f47e3632d43d677c66c102b"}, + {file = "numpy-1.26.4-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:675d61ffbfa78604709862923189bad94014bef562cc35cf61d3a07bba02a7ed"}, + {file = "numpy-1.26.4-cp312-cp312-musllinux_1_1_aarch64.whl", hash = "sha256:ab47dbe5cc8210f55aa58e4805fe224dac469cde56b9f731a4c098b91917159a"}, + {file = "numpy-1.26.4-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:1dda2e7b4ec9dd512f84935c5f126c8bd8b9f2fc001e9f54af255e8c5f16b0e0"}, + {file = "numpy-1.26.4-cp312-cp312-win32.whl", hash = "sha256:50193e430acfc1346175fcbdaa28ffec49947a06918b7b92130744e81e640110"}, + {file = "numpy-1.26.4-cp312-cp312-win_amd64.whl", hash = "sha256:08beddf13648eb95f8d867350f6a018a4be2e5ad54c8d8caed89ebca558b2818"}, + {file = "numpy-1.26.4-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:7349ab0fa0c429c82442a27a9673fc802ffdb7c7775fad780226cb234965e53c"}, + {file = "numpy-1.26.4-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:52b8b60467cd7dd1e9ed082188b4e6bb35aa5cdd01777621a1658910745b90be"}, + {file = "numpy-1.26.4-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d5241e0a80d808d70546c697135da2c613f30e28251ff8307eb72ba696945764"}, + {file = "numpy-1.26.4-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:f870204a840a60da0b12273ef34f7051e98c3b5961b61b0c2c1be6dfd64fbcd3"}, + {file = "numpy-1.26.4-cp39-cp39-musllinux_1_1_aarch64.whl", hash = "sha256:679b0076f67ecc0138fd2ede3a8fd196dddc2ad3254069bcb9faf9a79b1cebcd"}, + {file = "numpy-1.26.4-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:47711010ad8555514b434df65f7d7b076bb8261df1ca9bb78f53d3b2db02e95c"}, + {file = "numpy-1.26.4-cp39-cp39-win32.whl", hash = "sha256:a354325ee03388678242a4d7ebcd08b5c727033fcff3b2f536aea978e15ee9e6"}, + {file = "numpy-1.26.4-cp39-cp39-win_amd64.whl", hash = "sha256:3373d5d70a5fe74a2c1bb6d2cfd9609ecf686d47a2d7b1d37a8f3b6bf6003aea"}, + {file = "numpy-1.26.4-pp39-pypy39_pp73-macosx_10_9_x86_64.whl", hash = "sha256:afedb719a9dcfc7eaf2287b839d8198e06dcd4cb5d276a3df279231138e83d30"}, + {file = "numpy-1.26.4-pp39-pypy39_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:95a7476c59002f2f6c590b9b7b998306fba6a5aa646b1e22ddfeaf8f78c3a29c"}, + {file = "numpy-1.26.4-pp39-pypy39_pp73-win_amd64.whl", hash = "sha256:7e50d0a0cc3189f9cb0aeb3a6a6af18c16f59f004b866cd2be1c14b36134a4a0"}, + {file = "numpy-1.26.4.tar.gz", hash = "sha256:2a02aba9ed12e4ac4eb3ea9421c420301a0c6460d9830d74a9df87efa4912010"}, +] + +[[package]] +name = "packaging" +version = "23.2" +description = "Core utilities for Python packages" +optional = false +python-versions = ">=3.7" +files = [ + {file = "packaging-23.2-py3-none-any.whl", hash = "sha256:8c491190033a9af7e1d931d0b5dacc2ef47509b34dd0de67ed209b5203fc88c7"}, + {file = "packaging-23.2.tar.gz", hash = "sha256:048fb0e9405036518eaaf48a55953c750c11e1a1b68e0dd1a9d62ed0c092cfc5"}, +] + +[[package]] +name = "pluggy" +version = "0.13.1" +description = "plugin and hook calling mechanisms for python" +optional = false +python-versions = ">=2.7, !=3.0.*, !=3.1.*, !=3.2.*, !=3.3.*" +files = [ + {file = "pluggy-0.13.1-py2.py3-none-any.whl", hash = "sha256:966c145cd83c96502c3c3868f50408687b38434af77734af1e9ca461a4081d2d"}, + {file = "pluggy-0.13.1.tar.gz", hash = "sha256:15b2acde666561e1298d71b523007ed7364de07029219b604cf808bfa1c765b0"}, +] + +[package.extras] +dev = ["pre-commit", "tox"] + +[[package]] +name = "protobuf" +version = "4.25.3" +description = "" +optional = false +python-versions = ">=3.8" +files = [ + {file = "protobuf-4.25.3-cp310-abi3-win32.whl", hash = "sha256:d4198877797a83cbfe9bffa3803602bbe1625dc30d8a097365dbc762e5790faa"}, + {file = "protobuf-4.25.3-cp310-abi3-win_amd64.whl", hash = "sha256:209ba4cc916bab46f64e56b85b090607a676f66b473e6b762e6f1d9d591eb2e8"}, + {file = "protobuf-4.25.3-cp37-abi3-macosx_10_9_universal2.whl", hash = "sha256:f1279ab38ecbfae7e456a108c5c0681e4956d5b1090027c1de0f934dfdb4b35c"}, + {file = "protobuf-4.25.3-cp37-abi3-manylinux2014_aarch64.whl", hash = "sha256:e7cb0ae90dd83727f0c0718634ed56837bfeeee29a5f82a7514c03ee1364c019"}, + {file = "protobuf-4.25.3-cp37-abi3-manylinux2014_x86_64.whl", hash = "sha256:7c8daa26095f82482307bc717364e7c13f4f1c99659be82890dcfc215194554d"}, + {file = "protobuf-4.25.3-cp38-cp38-win32.whl", hash = "sha256:f4f118245c4a087776e0a8408be33cf09f6c547442c00395fbfb116fac2f8ac2"}, + {file = "protobuf-4.25.3-cp38-cp38-win_amd64.whl", hash = "sha256:c053062984e61144385022e53678fbded7aea14ebb3e0305ae3592fb219ccfa4"}, + {file = "protobuf-4.25.3-cp39-cp39-win32.whl", hash = "sha256:19b270aeaa0099f16d3ca02628546b8baefe2955bbe23224aaf856134eccf1e4"}, + {file = "protobuf-4.25.3-cp39-cp39-win_amd64.whl", hash = "sha256:e3c97a1555fd6388f857770ff8b9703083de6bf1f9274a002a332d65fbb56c8c"}, + {file = "protobuf-4.25.3-py3-none-any.whl", hash = "sha256:f0700d54bcf45424477e46a9f0944155b46fb0639d69728739c0e47bab83f2b9"}, + {file = "protobuf-4.25.3.tar.gz", hash = "sha256:25b5d0b42fd000320bd7830b349e3b696435f3b329810427a6bcce6a5492cc5c"}, +] + +[[package]] +name = "py" +version = "1.11.0" +description = "library with cross-python path, ini-parsing, io, code, log facilities" +optional = false +python-versions = ">=2.7, !=3.0.*, !=3.1.*, !=3.2.*, !=3.3.*, !=3.4.*" +files = [ + {file = "py-1.11.0-py2.py3-none-any.whl", hash = "sha256:607c53218732647dff4acdfcd50cb62615cedf612e72d1724fb1a0cc6405b378"}, + {file = "py-1.11.0.tar.gz", hash = "sha256:51c75c4126074b472f746a24399ad32f6053d1b34b68d2fa41e558e6f4a98719"}, +] + +[[package]] +name = "pytest" +version = "5.4.3" +description = "pytest: simple powerful testing with Python" +optional = false +python-versions = ">=3.5" +files = [ + {file = "pytest-5.4.3-py3-none-any.whl", hash = "sha256:5c0db86b698e8f170ba4582a492248919255fcd4c79b1ee64ace34301fb589a1"}, + {file = "pytest-5.4.3.tar.gz", hash = "sha256:7979331bfcba207414f5e1263b5a0f8f521d0f457318836a7355531ed1a4c7d8"}, +] + +[package.dependencies] +atomicwrites = {version = ">=1.0", markers = "sys_platform == \"win32\""} +attrs = ">=17.4.0" +colorama = {version = "*", markers = "sys_platform == \"win32\""} +more-itertools = ">=4.0.0" +packaging = "*" +pluggy = ">=0.12,<1.0" +py = ">=1.5.0" +wcwidth = "*" + +[package.extras] +checkqa-mypy = ["mypy (==v0.761)"] +testing = ["argcomplete", "hypothesis (>=3.56)", "mock", "nose", "requests", "xmlschema"] + +[[package]] +name = "pyyaml" +version = "6.0.1" +description = "YAML parser and emitter for Python" +optional = false +python-versions = ">=3.6" +files = [ + {file = "PyYAML-6.0.1-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:d858aa552c999bc8a8d57426ed01e40bef403cd8ccdd0fc5f6f04a00414cac2a"}, + {file = "PyYAML-6.0.1-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:fd66fc5d0da6d9815ba2cebeb4205f95818ff4b79c3ebe268e75d961704af52f"}, + {file = "PyYAML-6.0.1-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:69b023b2b4daa7548bcfbd4aa3da05b3a74b772db9e23b982788168117739938"}, + {file = "PyYAML-6.0.1-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:81e0b275a9ecc9c0c0c07b4b90ba548307583c125f54d5b6946cfee6360c733d"}, + {file = "PyYAML-6.0.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ba336e390cd8e4d1739f42dfe9bb83a3cc2e80f567d8805e11b46f4a943f5515"}, + {file = "PyYAML-6.0.1-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:326c013efe8048858a6d312ddd31d56e468118ad4cdeda36c719bf5bb6192290"}, + {file = "PyYAML-6.0.1-cp310-cp310-win32.whl", hash = "sha256:bd4af7373a854424dabd882decdc5579653d7868b8fb26dc7d0e99f823aa5924"}, + {file = "PyYAML-6.0.1-cp310-cp310-win_amd64.whl", hash = "sha256:fd1592b3fdf65fff2ad0004b5e363300ef59ced41c2e6b3a99d4089fa8c5435d"}, + {file = "PyYAML-6.0.1-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:6965a7bc3cf88e5a1c3bd2e0b5c22f8d677dc88a455344035f03399034eb3007"}, + {file = "PyYAML-6.0.1-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:f003ed9ad21d6a4713f0a9b5a7a0a79e08dd0f221aff4525a2be4c346ee60aab"}, + {file = "PyYAML-6.0.1-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:42f8152b8dbc4fe7d96729ec2b99c7097d656dc1213a3229ca5383f973a5ed6d"}, + {file = "PyYAML-6.0.1-cp311-cp311-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:062582fca9fabdd2c8b54a3ef1c978d786e0f6b3a1510e0ac93ef59e0ddae2bc"}, + {file = "PyYAML-6.0.1-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:d2b04aac4d386b172d5b9692e2d2da8de7bfb6c387fa4f801fbf6fb2e6ba4673"}, + {file = "PyYAML-6.0.1-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:e7d73685e87afe9f3b36c799222440d6cf362062f78be1013661b00c5c6f678b"}, + {file = "PyYAML-6.0.1-cp311-cp311-win32.whl", hash = "sha256:1635fd110e8d85d55237ab316b5b011de701ea0f29d07611174a1b42f1444741"}, + {file = "PyYAML-6.0.1-cp311-cp311-win_amd64.whl", hash = "sha256:bf07ee2fef7014951eeb99f56f39c9bb4af143d8aa3c21b1677805985307da34"}, + {file = "PyYAML-6.0.1-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:855fb52b0dc35af121542a76b9a84f8d1cd886ea97c84703eaa6d88e37a2ad28"}, + {file = "PyYAML-6.0.1-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:40df9b996c2b73138957fe23a16a4f0ba614f4c0efce1e9406a184b6d07fa3a9"}, + {file = "PyYAML-6.0.1-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:a08c6f0fe150303c1c6b71ebcd7213c2858041a7e01975da3a99aed1e7a378ef"}, + {file = "PyYAML-6.0.1-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:6c22bec3fbe2524cde73d7ada88f6566758a8f7227bfbf93a408a9d86bcc12a0"}, + {file = "PyYAML-6.0.1-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:8d4e9c88387b0f5c7d5f281e55304de64cf7f9c0021a3525bd3b1c542da3b0e4"}, + {file = "PyYAML-6.0.1-cp312-cp312-win32.whl", hash = "sha256:d483d2cdf104e7c9fa60c544d92981f12ad66a457afae824d146093b8c294c54"}, + {file = "PyYAML-6.0.1-cp312-cp312-win_amd64.whl", hash = "sha256:0d3304d8c0adc42be59c5f8a4d9e3d7379e6955ad754aa9d6ab7a398b59dd1df"}, + {file = "PyYAML-6.0.1-cp36-cp36m-macosx_10_9_x86_64.whl", hash = "sha256:50550eb667afee136e9a77d6dc71ae76a44df8b3e51e41b77f6de2932bfe0f47"}, + {file = "PyYAML-6.0.1-cp36-cp36m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:1fe35611261b29bd1de0070f0b2f47cb6ff71fa6595c077e42bd0c419fa27b98"}, + {file = "PyYAML-6.0.1-cp36-cp36m-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:704219a11b772aea0d8ecd7058d0082713c3562b4e271b849ad7dc4a5c90c13c"}, + {file = "PyYAML-6.0.1-cp36-cp36m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:afd7e57eddb1a54f0f1a974bc4391af8bcce0b444685d936840f125cf046d5bd"}, + {file = "PyYAML-6.0.1-cp36-cp36m-win32.whl", hash = "sha256:fca0e3a251908a499833aa292323f32437106001d436eca0e6e7833256674585"}, + {file = "PyYAML-6.0.1-cp36-cp36m-win_amd64.whl", hash = "sha256:f22ac1c3cac4dbc50079e965eba2c1058622631e526bd9afd45fedd49ba781fa"}, + {file = "PyYAML-6.0.1-cp37-cp37m-macosx_10_9_x86_64.whl", hash = "sha256:b1275ad35a5d18c62a7220633c913e1b42d44b46ee12554e5fd39c70a243d6a3"}, + {file = "PyYAML-6.0.1-cp37-cp37m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:18aeb1bf9a78867dc38b259769503436b7c72f7a1f1f4c93ff9a17de54319b27"}, + {file = "PyYAML-6.0.1-cp37-cp37m-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:596106435fa6ad000c2991a98fa58eeb8656ef2325d7e158344fb33864ed87e3"}, + {file = "PyYAML-6.0.1-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:baa90d3f661d43131ca170712d903e6295d1f7a0f595074f151c0aed377c9b9c"}, + {file = "PyYAML-6.0.1-cp37-cp37m-win32.whl", hash = "sha256:9046c58c4395dff28dd494285c82ba00b546adfc7ef001486fbf0324bc174fba"}, + {file = "PyYAML-6.0.1-cp37-cp37m-win_amd64.whl", hash = "sha256:4fb147e7a67ef577a588a0e2c17b6db51dda102c71de36f8549b6816a96e1867"}, + {file = "PyYAML-6.0.1-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:1d4c7e777c441b20e32f52bd377e0c409713e8bb1386e1099c2415f26e479595"}, + {file = "PyYAML-6.0.1-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:a0cd17c15d3bb3fa06978b4e8958dcdc6e0174ccea823003a106c7d4d7899ac5"}, + {file = "PyYAML-6.0.1-cp38-cp38-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:28c119d996beec18c05208a8bd78cbe4007878c6dd15091efb73a30e90539696"}, + {file = "PyYAML-6.0.1-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:7e07cbde391ba96ab58e532ff4803f79c4129397514e1413a7dc761ccd755735"}, + {file = "PyYAML-6.0.1-cp38-cp38-musllinux_1_1_x86_64.whl", hash = "sha256:49a183be227561de579b4a36efbb21b3eab9651dd81b1858589f796549873dd6"}, + {file = "PyYAML-6.0.1-cp38-cp38-win32.whl", hash = "sha256:184c5108a2aca3c5b3d3bf9395d50893a7ab82a38004c8f61c258d4428e80206"}, + {file = "PyYAML-6.0.1-cp38-cp38-win_amd64.whl", hash = "sha256:1e2722cc9fbb45d9b87631ac70924c11d3a401b2d7f410cc0e3bbf249f2dca62"}, + {file = "PyYAML-6.0.1-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:9eb6caa9a297fc2c2fb8862bc5370d0303ddba53ba97e71f08023b6cd73d16a8"}, + {file = "PyYAML-6.0.1-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:c8098ddcc2a85b61647b2590f825f3db38891662cfc2fc776415143f599bb859"}, + {file = "PyYAML-6.0.1-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:5773183b6446b2c99bb77e77595dd486303b4faab2b086e7b17bc6bef28865f6"}, + {file = "PyYAML-6.0.1-cp39-cp39-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:b786eecbdf8499b9ca1d697215862083bd6d2a99965554781d0d8d1ad31e13a0"}, + {file = "PyYAML-6.0.1-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:bc1bf2925a1ecd43da378f4db9e4f799775d6367bdb94671027b73b393a7c42c"}, + {file = "PyYAML-6.0.1-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:04ac92ad1925b2cff1db0cfebffb6ffc43457495c9b3c39d3fcae417d7125dc5"}, + {file = "PyYAML-6.0.1-cp39-cp39-win32.whl", hash = "sha256:faca3bdcf85b2fc05d06ff3fbc1f83e1391b3e724afa3feba7d13eeab355484c"}, + {file = "PyYAML-6.0.1-cp39-cp39-win_amd64.whl", hash = "sha256:510c9deebc5c0225e8c96813043e62b680ba2f9c50a08d3724c7f28a747d1486"}, + {file = "PyYAML-6.0.1.tar.gz", hash = "sha256:bfdf460b1736c775f2ba9f6a92bca30bc2095067b8a9d77876d1fad6cc3b4a43"}, +] + +[[package]] +name = "regex" +version = "2023.12.25" +description = "Alternative regular expression module, to replace re." +optional = false +python-versions = ">=3.7" +files = [ + {file = "regex-2023.12.25-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:0694219a1d54336fd0445ea382d49d36882415c0134ee1e8332afd1529f0baa5"}, + {file = "regex-2023.12.25-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:b014333bd0217ad3d54c143de9d4b9a3ca1c5a29a6d0d554952ea071cff0f1f8"}, + {file = "regex-2023.12.25-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:d865984b3f71f6d0af64d0d88f5733521698f6c16f445bb09ce746c92c97c586"}, + {file = "regex-2023.12.25-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:1e0eabac536b4cc7f57a5f3d095bfa557860ab912f25965e08fe1545e2ed8b4c"}, + {file = "regex-2023.12.25-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:c25a8ad70e716f96e13a637802813f65d8a6760ef48672aa3502f4c24ea8b400"}, + {file = "regex-2023.12.25-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:a9b6d73353f777630626f403b0652055ebfe8ff142a44ec2cf18ae470395766e"}, + {file = "regex-2023.12.25-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a9cc99d6946d750eb75827cb53c4371b8b0fe89c733a94b1573c9dd16ea6c9e4"}, + {file = "regex-2023.12.25-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:88d1f7bef20c721359d8675f7d9f8e414ec5003d8f642fdfd8087777ff7f94b5"}, + {file = "regex-2023.12.25-cp310-cp310-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_12_x86_64.manylinux2010_x86_64.whl", hash = "sha256:cb3fe77aec8f1995611f966d0c656fdce398317f850d0e6e7aebdfe61f40e1cd"}, + {file = "regex-2023.12.25-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:7aa47c2e9ea33a4a2a05f40fcd3ea36d73853a2aae7b4feab6fc85f8bf2c9704"}, + {file = "regex-2023.12.25-cp310-cp310-musllinux_1_1_i686.whl", hash = "sha256:df26481f0c7a3f8739fecb3e81bc9da3fcfae34d6c094563b9d4670b047312e1"}, + {file = "regex-2023.12.25-cp310-cp310-musllinux_1_1_ppc64le.whl", hash = "sha256:c40281f7d70baf6e0db0c2f7472b31609f5bc2748fe7275ea65a0b4601d9b392"}, + {file = "regex-2023.12.25-cp310-cp310-musllinux_1_1_s390x.whl", hash = "sha256:d94a1db462d5690ebf6ae86d11c5e420042b9898af5dcf278bd97d6bda065423"}, + {file = "regex-2023.12.25-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:ba1b30765a55acf15dce3f364e4928b80858fa8f979ad41f862358939bdd1f2f"}, + {file = "regex-2023.12.25-cp310-cp310-win32.whl", hash = "sha256:150c39f5b964e4d7dba46a7962a088fbc91f06e606f023ce57bb347a3b2d4630"}, + {file = "regex-2023.12.25-cp310-cp310-win_amd64.whl", hash = "sha256:09da66917262d9481c719599116c7dc0c321ffcec4b1f510c4f8a066f8768105"}, + {file = "regex-2023.12.25-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:1b9d811f72210fa9306aeb88385b8f8bcef0dfbf3873410413c00aa94c56c2b6"}, + {file = "regex-2023.12.25-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:d902a43085a308cef32c0d3aea962524b725403fd9373dea18110904003bac97"}, + {file = "regex-2023.12.25-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:d166eafc19f4718df38887b2bbe1467a4f74a9830e8605089ea7a30dd4da8887"}, + {file = "regex-2023.12.25-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:c7ad32824b7f02bb3c9f80306d405a1d9b7bb89362d68b3c5a9be53836caebdb"}, + {file = "regex-2023.12.25-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:636ba0a77de609d6510235b7f0e77ec494d2657108f777e8765efc060094c98c"}, + {file = "regex-2023.12.25-cp311-cp311-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:0fda75704357805eb953a3ee15a2b240694a9a514548cd49b3c5124b4e2ad01b"}, + {file = "regex-2023.12.25-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:f72cbae7f6b01591f90814250e636065850c5926751af02bb48da94dfced7baa"}, + {file = "regex-2023.12.25-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:db2a0b1857f18b11e3b0e54ddfefc96af46b0896fb678c85f63fb8c37518b3e7"}, + {file = "regex-2023.12.25-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:7502534e55c7c36c0978c91ba6f61703faf7ce733715ca48f499d3dbbd7657e0"}, + {file = "regex-2023.12.25-cp311-cp311-musllinux_1_1_i686.whl", hash = "sha256:e8c7e08bb566de4faaf11984af13f6bcf6a08f327b13631d41d62592681d24fe"}, + {file = "regex-2023.12.25-cp311-cp311-musllinux_1_1_ppc64le.whl", hash = "sha256:283fc8eed679758de38fe493b7d7d84a198b558942b03f017b1f94dda8efae80"}, + {file = "regex-2023.12.25-cp311-cp311-musllinux_1_1_s390x.whl", hash = "sha256:f44dd4d68697559d007462b0a3a1d9acd61d97072b71f6d1968daef26bc744bd"}, + {file = "regex-2023.12.25-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:67d3ccfc590e5e7197750fcb3a2915b416a53e2de847a728cfa60141054123d4"}, + {file = "regex-2023.12.25-cp311-cp311-win32.whl", hash = "sha256:68191f80a9bad283432385961d9efe09d783bcd36ed35a60fb1ff3f1ec2efe87"}, + {file = "regex-2023.12.25-cp311-cp311-win_amd64.whl", hash = "sha256:7d2af3f6b8419661a0c421584cfe8aaec1c0e435ce7e47ee2a97e344b98f794f"}, + {file = "regex-2023.12.25-cp312-cp312-macosx_10_9_universal2.whl", hash = "sha256:8a0ccf52bb37d1a700375a6b395bff5dd15c50acb745f7db30415bae3c2b0715"}, + {file = "regex-2023.12.25-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:c3c4a78615b7762740531c27cf46e2f388d8d727d0c0c739e72048beb26c8a9d"}, + {file = "regex-2023.12.25-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:ad83e7545b4ab69216cef4cc47e344d19622e28aabec61574b20257c65466d6a"}, + {file = "regex-2023.12.25-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:b7a635871143661feccce3979e1727c4e094f2bdfd3ec4b90dfd4f16f571a87a"}, + {file = "regex-2023.12.25-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:d498eea3f581fbe1b34b59c697512a8baef88212f92e4c7830fcc1499f5b45a5"}, + {file = "regex-2023.12.25-cp312-cp312-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:43f7cd5754d02a56ae4ebb91b33461dc67be8e3e0153f593c509e21d219c5060"}, + {file = "regex-2023.12.25-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:51f4b32f793812714fd5307222a7f77e739b9bc566dc94a18126aba3b92b98a3"}, + {file = "regex-2023.12.25-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:ba99d8077424501b9616b43a2d208095746fb1284fc5ba490139651f971d39d9"}, + {file = "regex-2023.12.25-cp312-cp312-musllinux_1_1_aarch64.whl", hash = "sha256:4bfc2b16e3ba8850e0e262467275dd4d62f0d045e0e9eda2bc65078c0110a11f"}, + {file = "regex-2023.12.25-cp312-cp312-musllinux_1_1_i686.whl", hash = "sha256:8c2c19dae8a3eb0ea45a8448356ed561be843b13cbc34b840922ddf565498c1c"}, + {file = "regex-2023.12.25-cp312-cp312-musllinux_1_1_ppc64le.whl", hash = "sha256:60080bb3d8617d96f0fb7e19796384cc2467447ef1c491694850ebd3670bc457"}, + {file = "regex-2023.12.25-cp312-cp312-musllinux_1_1_s390x.whl", hash = "sha256:b77e27b79448e34c2c51c09836033056a0547aa360c45eeeb67803da7b0eedaf"}, + {file = "regex-2023.12.25-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:518440c991f514331f4850a63560321f833979d145d7d81186dbe2f19e27ae3d"}, + {file = "regex-2023.12.25-cp312-cp312-win32.whl", hash = "sha256:e2610e9406d3b0073636a3a2e80db05a02f0c3169b5632022b4e81c0364bcda5"}, + {file = "regex-2023.12.25-cp312-cp312-win_amd64.whl", hash = "sha256:cc37b9aeebab425f11f27e5e9e6cf580be7206c6582a64467a14dda211abc232"}, + {file = "regex-2023.12.25-cp37-cp37m-macosx_10_9_x86_64.whl", hash = "sha256:da695d75ac97cb1cd725adac136d25ca687da4536154cdc2815f576e4da11c69"}, + {file = "regex-2023.12.25-cp37-cp37m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d126361607b33c4eb7b36debc173bf25d7805847346dd4d99b5499e1fef52bc7"}, + {file = "regex-2023.12.25-cp37-cp37m-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:4719bb05094d7d8563a450cf8738d2e1061420f79cfcc1fa7f0a44744c4d8f73"}, + {file = "regex-2023.12.25-cp37-cp37m-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:5dd58946bce44b53b06d94aa95560d0b243eb2fe64227cba50017a8d8b3cd3e2"}, + {file = "regex-2023.12.25-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:22a86d9fff2009302c440b9d799ef2fe322416d2d58fc124b926aa89365ec482"}, + {file = "regex-2023.12.25-cp37-cp37m-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:2aae8101919e8aa05ecfe6322b278f41ce2994c4a430303c4cd163fef746e04f"}, + {file = "regex-2023.12.25-cp37-cp37m-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_12_x86_64.manylinux2010_x86_64.whl", hash = "sha256:e692296c4cc2873967771345a876bcfc1c547e8dd695c6b89342488b0ea55cd8"}, + {file = "regex-2023.12.25-cp37-cp37m-musllinux_1_1_aarch64.whl", hash = "sha256:263ef5cc10979837f243950637fffb06e8daed7f1ac1e39d5910fd29929e489a"}, + {file = "regex-2023.12.25-cp37-cp37m-musllinux_1_1_i686.whl", hash = "sha256:d6f7e255e5fa94642a0724e35406e6cb7001c09d476ab5fce002f652b36d0c39"}, + {file = "regex-2023.12.25-cp37-cp37m-musllinux_1_1_ppc64le.whl", hash = "sha256:88ad44e220e22b63b0f8f81f007e8abbb92874d8ced66f32571ef8beb0643b2b"}, + {file = "regex-2023.12.25-cp37-cp37m-musllinux_1_1_s390x.whl", hash = "sha256:3a17d3ede18f9cedcbe23d2daa8a2cd6f59fe2bf082c567e43083bba3fb00347"}, + {file = "regex-2023.12.25-cp37-cp37m-musllinux_1_1_x86_64.whl", hash = "sha256:d15b274f9e15b1a0b7a45d2ac86d1f634d983ca40d6b886721626c47a400bf39"}, + {file = "regex-2023.12.25-cp37-cp37m-win32.whl", hash = "sha256:ed19b3a05ae0c97dd8f75a5d8f21f7723a8c33bbc555da6bbe1f96c470139d3c"}, + {file = "regex-2023.12.25-cp37-cp37m-win_amd64.whl", hash = "sha256:a6d1047952c0b8104a1d371f88f4ab62e6275567d4458c1e26e9627ad489b445"}, + {file = "regex-2023.12.25-cp38-cp38-macosx_10_9_universal2.whl", hash = "sha256:b43523d7bc2abd757119dbfb38af91b5735eea45537ec6ec3a5ec3f9562a1c53"}, + {file = "regex-2023.12.25-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:efb2d82f33b2212898f1659fb1c2e9ac30493ac41e4d53123da374c3b5541e64"}, + {file = "regex-2023.12.25-cp38-cp38-macosx_11_0_arm64.whl", hash = "sha256:b7fca9205b59c1a3d5031f7e64ed627a1074730a51c2a80e97653e3e9fa0d415"}, + {file = "regex-2023.12.25-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:086dd15e9435b393ae06f96ab69ab2d333f5d65cbe65ca5a3ef0ec9564dfe770"}, + {file = "regex-2023.12.25-cp38-cp38-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:e81469f7d01efed9b53740aedd26085f20d49da65f9c1f41e822a33992cb1590"}, + {file = "regex-2023.12.25-cp38-cp38-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:34e4af5b27232f68042aa40a91c3b9bb4da0eeb31b7632e0091afc4310afe6cb"}, + {file = "regex-2023.12.25-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:9852b76ab558e45b20bf1893b59af64a28bd3820b0c2efc80e0a70a4a3ea51c1"}, + {file = "regex-2023.12.25-cp38-cp38-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:ff100b203092af77d1a5a7abe085b3506b7eaaf9abf65b73b7d6905b6cb76988"}, + {file = "regex-2023.12.25-cp38-cp38-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_12_x86_64.manylinux2010_x86_64.whl", hash = "sha256:cc038b2d8b1470364b1888a98fd22d616fba2b6309c5b5f181ad4483e0017861"}, + {file = "regex-2023.12.25-cp38-cp38-musllinux_1_1_aarch64.whl", hash = "sha256:094ba386bb5c01e54e14434d4caabf6583334090865b23ef58e0424a6286d3dc"}, + {file = "regex-2023.12.25-cp38-cp38-musllinux_1_1_i686.whl", hash = "sha256:5cd05d0f57846d8ba4b71d9c00f6f37d6b97d5e5ef8b3c3840426a475c8f70f4"}, + {file = "regex-2023.12.25-cp38-cp38-musllinux_1_1_ppc64le.whl", hash = "sha256:9aa1a67bbf0f957bbe096375887b2505f5d8ae16bf04488e8b0f334c36e31360"}, + {file = "regex-2023.12.25-cp38-cp38-musllinux_1_1_s390x.whl", hash = "sha256:98a2636994f943b871786c9e82bfe7883ecdaba2ef5df54e1450fa9869d1f756"}, + {file = "regex-2023.12.25-cp38-cp38-musllinux_1_1_x86_64.whl", hash = "sha256:37f8e93a81fc5e5bd8db7e10e62dc64261bcd88f8d7e6640aaebe9bc180d9ce2"}, + {file = "regex-2023.12.25-cp38-cp38-win32.whl", hash = "sha256:d78bd484930c1da2b9679290a41cdb25cc127d783768a0369d6b449e72f88beb"}, + {file = "regex-2023.12.25-cp38-cp38-win_amd64.whl", hash = "sha256:b521dcecebc5b978b447f0f69b5b7f3840eac454862270406a39837ffae4e697"}, + {file = "regex-2023.12.25-cp39-cp39-macosx_10_9_universal2.whl", hash = "sha256:f7bc09bc9c29ebead055bcba136a67378f03d66bf359e87d0f7c759d6d4ffa31"}, + {file = "regex-2023.12.25-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:e14b73607d6231f3cc4622809c196b540a6a44e903bcfad940779c80dffa7be7"}, + {file = "regex-2023.12.25-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:9eda5f7a50141291beda3edd00abc2d4a5b16c29c92daf8d5bd76934150f3edc"}, + {file = "regex-2023.12.25-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:cc6bb9aa69aacf0f6032c307da718f61a40cf970849e471254e0e91c56ffca95"}, + {file = "regex-2023.12.25-cp39-cp39-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:298dc6354d414bc921581be85695d18912bea163a8b23cac9a2562bbcd5088b1"}, + {file = "regex-2023.12.25-cp39-cp39-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:2f4e475a80ecbd15896a976aa0b386c5525d0ed34d5c600b6d3ebac0a67c7ddf"}, + {file = "regex-2023.12.25-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:531ac6cf22b53e0696f8e1d56ce2396311254eb806111ddd3922c9d937151dae"}, + {file = "regex-2023.12.25-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:22f3470f7524b6da61e2020672df2f3063676aff444db1daa283c2ea4ed259d6"}, + {file = "regex-2023.12.25-cp39-cp39-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_12_x86_64.manylinux2010_x86_64.whl", hash = "sha256:89723d2112697feaa320c9d351e5f5e7b841e83f8b143dba8e2d2b5f04e10923"}, + {file = "regex-2023.12.25-cp39-cp39-musllinux_1_1_aarch64.whl", hash = "sha256:0ecf44ddf9171cd7566ef1768047f6e66975788258b1c6c6ca78098b95cf9a3d"}, + {file = "regex-2023.12.25-cp39-cp39-musllinux_1_1_i686.whl", hash = "sha256:905466ad1702ed4acfd67a902af50b8db1feeb9781436372261808df7a2a7bca"}, + {file = "regex-2023.12.25-cp39-cp39-musllinux_1_1_ppc64le.whl", hash = "sha256:4558410b7a5607a645e9804a3e9dd509af12fb72b9825b13791a37cd417d73a5"}, + {file = "regex-2023.12.25-cp39-cp39-musllinux_1_1_s390x.whl", hash = "sha256:7e316026cc1095f2a3e8cc012822c99f413b702eaa2ca5408a513609488cb62f"}, + {file = "regex-2023.12.25-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:3b1de218d5375cd6ac4b5493e0b9f3df2be331e86520f23382f216c137913d20"}, + {file = "regex-2023.12.25-cp39-cp39-win32.whl", hash = "sha256:11a963f8e25ab5c61348d090bf1b07f1953929c13bd2309a0662e9ff680763c9"}, + {file = "regex-2023.12.25-cp39-cp39-win_amd64.whl", hash = "sha256:e693e233ac92ba83a87024e1d32b5f9ab15ca55ddd916d878146f4e3406b5c91"}, + {file = "regex-2023.12.25.tar.gz", hash = "sha256:29171aa128da69afdf4bde412d5bedc335f2ca8fcfe4489038577d05f16181e5"}, +] + +[[package]] +name = "requests" +version = "2.31.0" +description = "Python HTTP for Humans." +optional = false +python-versions = ">=3.7" +files = [ + {file = "requests-2.31.0-py3-none-any.whl", hash = "sha256:58cd2187c01e70e6e26505bca751777aa9f2ee0b7f4300988b709f44e013003f"}, + {file = "requests-2.31.0.tar.gz", hash = "sha256:942c5a758f98d790eaed1a29cb6eefc7ffb0d1cf7af05c3d2791656dbd6ad1e1"}, +] + +[package.dependencies] +certifi = ">=2017.4.17" +charset-normalizer = ">=2,<4" +idna = ">=2.5,<4" +urllib3 = ">=1.21.1,<3" + +[package.extras] +socks = ["PySocks (>=1.5.6,!=1.5.7)"] +use-chardet-on-py3 = ["chardet (>=3.0.2,<6)"] + +[[package]] +name = "safetensors" +version = "0.4.2" +description = "" +optional = false +python-versions = ">=3.7" +files = [ + {file = "safetensors-0.4.2-cp310-cp310-macosx_10_12_x86_64.whl", hash = "sha256:69d8bb8384dc2cb5b72c36c4d6980771b293d1a1377b378763f5e37b6bb8d133"}, + {file = "safetensors-0.4.2-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:3d420e19fcef96d0067f4de4699682b4bbd85fc8fea0bd45fcd961fdf3e8c82c"}, + {file = "safetensors-0.4.2-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:9ca54742122fa3c4821754adb67318e1cd25c3a22bbf0c5520d5176e77a099ac"}, + {file = "safetensors-0.4.2-cp310-cp310-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:8b47aa643afdfd66cf7ce4c184092ae734e15d10aba2c2948f24270211801c3c"}, + {file = "safetensors-0.4.2-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:d88a16bbc330f27e7f2d4caaf6fb061ad0b8a756ecc4033260b0378e128ce8a2"}, + {file = "safetensors-0.4.2-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:e9223b8ac21085db614a510eb3445e7083cae915a9202357555fa939695d4f57"}, + {file = "safetensors-0.4.2-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ce6cb86133dc8930a7ab5e7438545a7f205f7a1cdd5aaf108c1d0da6bdcfbc2b"}, + {file = "safetensors-0.4.2-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:b8a628e0ae2bbc334b62952c384aa5f41621d01850f8d67b04a96b9c39dd7326"}, + {file = "safetensors-0.4.2-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:88d6beb7f811a081e0e5f1d9669fdac816c45340c04b1eaf7ebfda0ce93ea403"}, + {file = "safetensors-0.4.2-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:b57fc5b1b54cb12d8690a58a4cf4b7144730d4bde9d98aa0e1dab6295a1cd579"}, + {file = "safetensors-0.4.2-cp310-none-win32.whl", hash = "sha256:9d87a1c98803c16cf113b9ba03f07b2dce5e8eabfd1811a7f7323fcaa2a1bf47"}, + {file = "safetensors-0.4.2-cp310-none-win_amd64.whl", hash = "sha256:18930ec1d1ecb526d3d9835abc2489b8f1530877518f0c541e77ef0b7abcbd99"}, + {file = "safetensors-0.4.2-cp311-cp311-macosx_10_12_x86_64.whl", hash = "sha256:c5dd2ed788730ed56b415d1a11c62026b8cc8c573f55a2092afb3ab383e94fff"}, + {file = "safetensors-0.4.2-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:cc41791b33efb9c83a59b731619f3d15f543dfe71f3a793cb8fbf9bd5d0d5d71"}, + {file = "safetensors-0.4.2-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:4c888bf71d5ca12a720f1ed87d407c4918afa022fb247a6546d8fac15b1f112b"}, + {file = "safetensors-0.4.2-cp311-cp311-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:e6b2feb4b47226a16a792e6fac3f49442714884a3d4c1008569d5068a3941be9"}, + {file = "safetensors-0.4.2-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:f41cc0ee4b838ae8f4d8364a1b162067693d11a3893f0863be8c228d40e4d0ee"}, + {file = "safetensors-0.4.2-cp311-cp311-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:51b7228e46c0a483c40ba4b9470dea00fb1ff8685026bb4766799000f6328ac2"}, + {file = "safetensors-0.4.2-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:02697f8f2be8ca3c37a4958702dbdb1864447ef765e18b5328a1617022dcf164"}, + {file = "safetensors-0.4.2-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:27fd8f65cf7c80e4280cae1ee6bcd85c483882f6580821abe71ee1a0d3dcfca7"}, + {file = "safetensors-0.4.2-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:c487b5f113b0924c9534a07dc034830fb4ef05ce9bb6d78cfe016a7dedfe281f"}, + {file = "safetensors-0.4.2-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:da7f6483f3fe67ff39b3a55552552c67930ea10a36e9f2539d36fc205273d767"}, + {file = "safetensors-0.4.2-cp311-none-win32.whl", hash = "sha256:52a7012f6cb9cb4a132760b6308daede18a9f5f8952ce08adc7c67a7d865c2d8"}, + {file = "safetensors-0.4.2-cp311-none-win_amd64.whl", hash = "sha256:4d1361a097ac430b310ce9eed8ed4746edee33ddafdfbb965debc8966fc34dc2"}, + {file = "safetensors-0.4.2-cp312-cp312-macosx_10_12_x86_64.whl", hash = "sha256:77af8aa0edcc2863760fd6febbfdb82e88fd75d0e60c1ce4ba57208ba5e4a89b"}, + {file = "safetensors-0.4.2-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:846666c1c5a8c8888d2dfda8d3921cb9cb8e2c5f78365be756c11021e75a0a2a"}, + {file = "safetensors-0.4.2-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:4f4bfc7ea19b446bfad41510d4b4c76101698c00caaa8a332c8edd8090a412ef"}, + {file = "safetensors-0.4.2-cp312-cp312-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:233436fd30f27ffeb3c3780d0b84f496518868445c7a8db003639a649cc98453"}, + {file = "safetensors-0.4.2-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:7a09237a795d11cd11f9dae505d170a29b5616151db1e10c14f892b11caadc7d"}, + {file = "safetensors-0.4.2-cp312-cp312-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:de01c9a3a3b7b69627d624ff69d9f11d28ce9908eea2fb6245adafa4b1d43df6"}, + {file = "safetensors-0.4.2-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:8c1f25c5069ee42a5bcffdc66c300a407941edd73f3239e9fdefd26216407391"}, + {file = "safetensors-0.4.2-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:7a73b3649456d09ca8506140d44484b63154a7378434cc1e8719f8056550b224"}, + {file = "safetensors-0.4.2-cp312-cp312-musllinux_1_1_aarch64.whl", hash = "sha256:e1625a8d07d046e968bd5c4961810aba1225984e4fb9243626f9d04a06ed3fee"}, + {file = "safetensors-0.4.2-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:8f74c86b25615cb24ad4cff765a2eefc09d71bf0fed97588cf585aad9c38fbb4"}, + {file = "safetensors-0.4.2-cp312-none-win32.whl", hash = "sha256:8523b9c5777d771bcde5c2389c03f1cdf7ebe8797432a1bd5e345efe25c55987"}, + {file = "safetensors-0.4.2-cp312-none-win_amd64.whl", hash = "sha256:dcff0243e1737a21f83d664c63fed89d1f532c23fc6830d0427279fabd789ccb"}, + {file = "safetensors-0.4.2-cp313-cp313-macosx_10_12_x86_64.whl", hash = "sha256:96ad3d7d472612e26cbe413922b4fb13933310f0511d346ea5cc9a1e856e52eb"}, + {file = "safetensors-0.4.2-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:88250922401b5ae4e37de929178caf46be47ed16c817b2237b81679bec07c120"}, + {file = "safetensors-0.4.2-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d40443554142fc0ab30652d5cc8554c4b7a613513bde00373e18afd5de8cbe4b"}, + {file = "safetensors-0.4.2-cp313-cp313-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:27f53f70106224d32d874aacecbeb4a6e4c5b16a1d2006d0e876d97229086d71"}, + {file = "safetensors-0.4.2-cp313-cp313-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:cc068afe23734dfb26ce19db0a7877499ddf73b1d55ceb762417e8da4a1b05fb"}, + {file = "safetensors-0.4.2-cp313-cp313-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:9be1918eb8d43a11a6f8806759fccfa0eeb0542b12924caba66af8a7800ad01a"}, + {file = "safetensors-0.4.2-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:41911087d20a7bbd78cb4ad4f98aab0c431533107584df6635d8b54b99945573"}, + {file = "safetensors-0.4.2-cp313-cp313-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:50771c662aab909f31e94d048e76861fd027d66076ea773eef2e66c717766e24"}, + {file = "safetensors-0.4.2-cp313-cp313-musllinux_1_1_aarch64.whl", hash = "sha256:13f2e57be007b7ea9329133d2399e6bdfcf1910f655440a4da17df3a45afcd30"}, + {file = "safetensors-0.4.2-cp313-cp313-musllinux_1_1_x86_64.whl", hash = "sha256:c772147e6395bc829842e0a98e1b30c67fe25d816299c28196488511d5a5e951"}, + {file = "safetensors-0.4.2-cp37-cp37m-macosx_10_12_x86_64.whl", hash = "sha256:36239a0060b537a3e8c473df78cffee14c3ec4f51d5f1a853af99371a2fb2a35"}, + {file = "safetensors-0.4.2-cp37-cp37m-macosx_11_0_arm64.whl", hash = "sha256:d0cbb7664fad2c307f95195f951b7059e95dc23e0e1822e5978c8b500098543c"}, + {file = "safetensors-0.4.2-cp37-cp37m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:2b3e55adb6bd9dc1c2a341e72f48f075953fa35d173dd8e29a95b3b02d0d1462"}, + {file = "safetensors-0.4.2-cp37-cp37m-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:42f743b3cca863fba53ca57a193f510e5ec359b97f38c282437716b6768e4a25"}, + {file = "safetensors-0.4.2-cp37-cp37m-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:04e6af4a6dbeb06c4e6e7d46cf9c716cbc4cc5ef62584fd8a7c0fe558562df45"}, + {file = "safetensors-0.4.2-cp37-cp37m-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:a492ba21b5c8f14ee5ec9b20f42ba969e53ca1f909a4d04aad736b66a341dcc2"}, + {file = "safetensors-0.4.2-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:b25b8233a1a85dc67e39838951cfb01595d792f3b7b644add63edb652992e030"}, + {file = "safetensors-0.4.2-cp37-cp37m-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:fd27e063fbdafe776f7b1714da59110e88f270e86db00788a8fd65f4eacfeba7"}, + {file = "safetensors-0.4.2-cp37-cp37m-musllinux_1_1_aarch64.whl", hash = "sha256:1b6fa399f251bbeb52029bf5a0ac2878d7705dd3612a2f8895b48e9c11f0367d"}, + {file = "safetensors-0.4.2-cp37-cp37m-musllinux_1_1_x86_64.whl", hash = "sha256:de642d46b459e4afd5c2020b26c0d6d869a171ea00411897d5776c127cac74f0"}, + {file = "safetensors-0.4.2-cp37-none-win32.whl", hash = "sha256:77b72d17754c93bb68f3598182f14d78776e0b9b31682ca5bb2c7c5bd9a75267"}, + {file = "safetensors-0.4.2-cp37-none-win_amd64.whl", hash = "sha256:d36ee3244d461cd655aeef493792c3bccf4875282f8407fd9af99e9a41cf2530"}, + {file = "safetensors-0.4.2-cp38-cp38-macosx_10_12_x86_64.whl", hash = "sha256:16b6b3884f7876c6b3b23a742428223a7170a5a9dac819d8c12a1569422c4b5a"}, + {file = "safetensors-0.4.2-cp38-cp38-macosx_11_0_arm64.whl", hash = "sha256:ee25d311493fbbe0be9d395faee46e9d79e8948f461e388ff39e59875ed9a350"}, + {file = "safetensors-0.4.2-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:eed8097968585cd752a1171f86fce9aa1d89a29033e5cd8bec5a502e29f6b7af"}, + {file = "safetensors-0.4.2-cp38-cp38-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:880e6865cf72cb67f9ab8d04a3c4b49dd95ae92fb1583929ce65aed94e1f685f"}, + {file = "safetensors-0.4.2-cp38-cp38-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:91290f83daf80ce6d1a7f629b244443c200060a80f908b29d879021409e5ea94"}, + {file = "safetensors-0.4.2-cp38-cp38-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:3517d568486ab3508a7acc360b82d7a4a3e26b86efdf210a9ecd9d233c40708a"}, + {file = "safetensors-0.4.2-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:e1f43a77eb38540f782999e5dc5645164fe9027d3f0194f6c9a5126168017efa"}, + {file = "safetensors-0.4.2-cp38-cp38-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:b684d9818aa5d63fddc65f7d0151968037d255d91adf74eba82125b41c680aaa"}, + {file = "safetensors-0.4.2-cp38-cp38-musllinux_1_1_aarch64.whl", hash = "sha256:ab1f5d84185f9fefaf21413efb764e4908057b8a9a0b987ede890c353490fd70"}, + {file = "safetensors-0.4.2-cp38-cp38-musllinux_1_1_x86_64.whl", hash = "sha256:2bd979642e6c3a517ef4b84ff36c2fee4015664fea05a61154fc565978347553"}, + {file = "safetensors-0.4.2-cp38-none-win32.whl", hash = "sha256:11be6e7afed29e5a5628f0aa6214e34bc194da73f558dc69fc7d56e07037422a"}, + {file = "safetensors-0.4.2-cp38-none-win_amd64.whl", hash = "sha256:2f7a6e5d29bd2cc340cffaa391fa437b1be9d21a2bd8b8724d2875d13a6ef2a9"}, + {file = "safetensors-0.4.2-cp39-cp39-macosx_10_12_x86_64.whl", hash = "sha256:a5a921b4fe6925f9942adff3ebae8c16e0487908c54586a5a42f35b59fd69794"}, + {file = "safetensors-0.4.2-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:b691727228c28f2d82d8a92b2bc26e7a1f129ee40b2f2a3185b5974e038ed47c"}, + {file = "safetensors-0.4.2-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:91ca1056decc4e981248786e87b2a202d4841ee5f99d433f1adf3d44d4bcfa0e"}, + {file = "safetensors-0.4.2-cp39-cp39-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:55969fd2e6fdb38dc221b0ab380668c21b0efa12a7562db9924759faa3c51757"}, + {file = "safetensors-0.4.2-cp39-cp39-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:6ae429bfaecc10ab5fe78c93009b3d1656c1581da560041e700eadb497dbe7a4"}, + {file = "safetensors-0.4.2-cp39-cp39-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:4ff88f194fe4ac50b463a4a6f0c03af9ad72eb5d24ec6d6730af59522e37fedb"}, + {file = "safetensors-0.4.2-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a80cb48d0a447f8dd18e61813efa7d3f8f8d52edf0f05806abc0c59b83431f57"}, + {file = "safetensors-0.4.2-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:b286fb7adfee70a4189898ac2342b8a67d5f493e6b21b0af89ca8eac1b967cbf"}, + {file = "safetensors-0.4.2-cp39-cp39-musllinux_1_1_aarch64.whl", hash = "sha256:0ceeff9ddbab4f78738489eb6682867ae946178776f33699737b2129b5394dc1"}, + {file = "safetensors-0.4.2-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:a26fae748a7488cb3aac381eddfa818c42052c87b5e689fb4c6e82ed58cec209"}, + {file = "safetensors-0.4.2-cp39-none-win32.whl", hash = "sha256:039a42ab33c9d68b39706fd38f1922ace26866eff246bf20271edb619f5f848b"}, + {file = "safetensors-0.4.2-cp39-none-win_amd64.whl", hash = "sha256:b3a3e1f5b85859e398773f064943b62a4059f225008a2a8ee6add1edcf77cacf"}, + {file = "safetensors-0.4.2-pp310-pypy310_pp73-macosx_10_12_x86_64.whl", hash = "sha256:4e70d442ad17e8b153ef9095bf48ea64f15a66bf26dc2b6ca94660c154edbc24"}, + {file = "safetensors-0.4.2-pp310-pypy310_pp73-macosx_11_0_arm64.whl", hash = "sha256:b90f1d9809caf4ff395951b4703295a68d12907f6945bbc3129e934ff8ae46f6"}, + {file = "safetensors-0.4.2-pp310-pypy310_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:8c7ac9ad3728838006598e296b3ae9f27d80b489effd4685b92d97b3fc4c98f6"}, + {file = "safetensors-0.4.2-pp310-pypy310_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:de5730d77e6ff7f4c7039e20913661ad0ea2f86c09e71c039e73dfdd1f394f08"}, + {file = "safetensors-0.4.2-pp310-pypy310_pp73-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:44feb8cb156d6803dcd19fc6b81b27235f29b877660605a6ac35e1da7d64f0e4"}, + {file = "safetensors-0.4.2-pp310-pypy310_pp73-musllinux_1_1_aarch64.whl", hash = "sha256:523a241c33e7c827ab9a3a23760d75c7d062f43dfe55b6b019409f89b0fb52d1"}, + {file = "safetensors-0.4.2-pp310-pypy310_pp73-musllinux_1_1_x86_64.whl", hash = "sha256:fb18300e8eb74291225214f26c9a8ae2110fd61a6c9b5a2ff4c4e0eb1bb9a998"}, + {file = "safetensors-0.4.2-pp37-pypy37_pp73-macosx_10_12_x86_64.whl", hash = "sha256:fe5437ff9fb116e44f2ab558981249ae63f978392b4576e62fcfe167d353edbc"}, + {file = "safetensors-0.4.2-pp37-pypy37_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d9304a0934ced5a5d272f39de36291dc141dfc152d277f03fb4d65f2fb2ffa7c"}, + {file = "safetensors-0.4.2-pp37-pypy37_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:160ba1b1e11cf874602c233ab80a14f588571d09556cbc3586900121d622b5ed"}, + {file = "safetensors-0.4.2-pp37-pypy37_pp73-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:04fcd6fcf7d9c13c7e5dc7e08de5e492ee4daa8f4ad74b4d8299d3eb0224292f"}, + {file = "safetensors-0.4.2-pp37-pypy37_pp73-musllinux_1_1_aarch64.whl", hash = "sha256:906d14c4a677d35834fb0f3a5455ef8305e1bba10a5e0f2e0f357b3d1ad989f2"}, + {file = "safetensors-0.4.2-pp37-pypy37_pp73-musllinux_1_1_x86_64.whl", hash = "sha256:df3fcdec0cd543084610d1f09c65cdb10fb3079f79bceddc092b0d187c6a265b"}, + {file = "safetensors-0.4.2-pp38-pypy38_pp73-macosx_10_12_x86_64.whl", hash = "sha256:5ca76f13fb1cef242ea3ad2cb37388e7d005994f42af8b44bee56ba48b2d45ce"}, + {file = "safetensors-0.4.2-pp38-pypy38_pp73-macosx_11_0_arm64.whl", hash = "sha256:278a1a3414c020785decdcd741c578725721274d2f9f787fcc930882e83b89cc"}, + {file = "safetensors-0.4.2-pp38-pypy38_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:05b5a461cc68ecd42d9d546e5e1268a39d8ede7934a68d1ce17c3c659cb829d6"}, + {file = "safetensors-0.4.2-pp38-pypy38_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:c2341411412a41671d25e26bed59ec121e46bf4fadb8132895e610411c4b9681"}, + {file = "safetensors-0.4.2-pp38-pypy38_pp73-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:3497ac3895acf17c5f98197f1fa4769f09c5e7ede07fcb102f1c201e663e052c"}, + {file = "safetensors-0.4.2-pp38-pypy38_pp73-musllinux_1_1_aarch64.whl", hash = "sha256:01b5e71d3754d2201294f1eb7a6d59cce3a5702ff96d83d226571b2ca2183837"}, + {file = "safetensors-0.4.2-pp38-pypy38_pp73-musllinux_1_1_x86_64.whl", hash = "sha256:3627dbd1ea488dd8046a0491de5087f3c0d641e7acc80c0189a33c69398f1cd1"}, + {file = "safetensors-0.4.2-pp39-pypy39_pp73-macosx_10_12_x86_64.whl", hash = "sha256:9d56f0ef53afad26ec54ceede78a43e9a23a076dadbbda7b44d304c591abf4c1"}, + {file = "safetensors-0.4.2-pp39-pypy39_pp73-macosx_11_0_arm64.whl", hash = "sha256:b259ca73d42daf658a1bda463f1f83885ae4d93a60869be80d7f7dfcc9d8bbb5"}, + {file = "safetensors-0.4.2-pp39-pypy39_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:1ebc3cd401e4eb54e7c0a70346be565e81942d9a41fafd5f4bf7ab3a55d10378"}, + {file = "safetensors-0.4.2-pp39-pypy39_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:5bc384a0309b706aa0425c93abb0390508a61bf029ce99c7d9df4220f25871a5"}, + {file = "safetensors-0.4.2-pp39-pypy39_pp73-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:af2d8f7235d8a08fbccfb8394387890e7fa38942b349a94e6eff13c52ac98087"}, + {file = "safetensors-0.4.2-pp39-pypy39_pp73-musllinux_1_1_aarch64.whl", hash = "sha256:0911315bbcc5289087d063c2c2c7ccd711ea97a7e557a7bce005ac2cf80146aa"}, + {file = "safetensors-0.4.2-pp39-pypy39_pp73-musllinux_1_1_x86_64.whl", hash = "sha256:1efe31673be91832d73439a2af426743e1395fc9ef7b081914e9e1d567bd7b5f"}, + {file = "safetensors-0.4.2.tar.gz", hash = "sha256:acc85dcb09ec5e8aa787f588d7ad4d55c103f31e4ff060e17d92cc0e8b8cac73"}, +] + +[package.extras] +all = ["safetensors[jax]", "safetensors[numpy]", "safetensors[paddlepaddle]", "safetensors[pinned-tf]", "safetensors[quality]", "safetensors[testing]", "safetensors[torch]"] +dev = ["safetensors[all]"] +jax = ["flax (>=0.6.3)", "jax (>=0.3.25)", "jaxlib (>=0.3.25)", "safetensors[numpy]"] +mlx = ["mlx (>=0.0.9)"] +numpy = ["numpy (>=1.21.6)"] +paddlepaddle = ["paddlepaddle (>=2.4.1)", "safetensors[numpy]"] +pinned-tf = ["safetensors[numpy]", "tensorflow (==2.11.0)"] +quality = ["black (==22.3)", "click (==8.0.4)", "flake8 (>=3.8.3)", "isort (>=5.5.4)"] +tensorflow = ["safetensors[numpy]", "tensorflow (>=2.11.0)"] +testing = ["h5py (>=3.7.0)", "huggingface_hub (>=0.12.1)", "hypothesis (>=6.70.2)", "pytest (>=7.2.0)", "pytest-benchmark (>=4.0.0)", "safetensors[numpy]", "setuptools_rust (>=1.5.2)"] +torch = ["safetensors[numpy]", "torch (>=1.10)"] + +[[package]] +name = "sentencepiece" +version = "0.1.99" +description = "SentencePiece python wrapper" +optional = false +python-versions = "*" +files = [ + {file = "sentencepiece-0.1.99-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:0eb528e70571b7c02723e5804322469b82fe7ea418c96051d0286c0fa028db73"}, + {file = "sentencepiece-0.1.99-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:77d7fafb2c4e4659cbdf303929503f37a26eabc4ff31d3a79bf1c5a1b338caa7"}, + {file = "sentencepiece-0.1.99-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:be9cf5b9e404c245aeb3d3723c737ba7a8f5d4ba262ef233a431fa6c45f732a0"}, + {file = "sentencepiece-0.1.99-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:baed1a26464998f9710d20e52607c29ffd4293e7c71c6a1f83f51ad0911ec12c"}, + {file = "sentencepiece-0.1.99-cp310-cp310-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:9832f08bb372d4c8b567612f8eab9e36e268dff645f1c28f9f8e851be705f6d1"}, + {file = "sentencepiece-0.1.99-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:019e7535108e309dae2b253a75834fc3128240aa87c00eb80732078cdc182588"}, + {file = "sentencepiece-0.1.99-cp310-cp310-win32.whl", hash = "sha256:fa16a830416bb823fa2a52cbdd474d1f7f3bba527fd2304fb4b140dad31bb9bc"}, + {file = "sentencepiece-0.1.99-cp310-cp310-win_amd64.whl", hash = "sha256:14b0eccb7b641d4591c3e12ae44cab537d68352e4d3b6424944f0c447d2348d5"}, + {file = "sentencepiece-0.1.99-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:6d3c56f24183a1e8bd61043ff2c58dfecdc68a5dd8955dc13bab83afd5f76b81"}, + {file = "sentencepiece-0.1.99-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:ed6ea1819fd612c989999e44a51bf556d0ef6abfb553080b9be3d347e18bcfb7"}, + {file = "sentencepiece-0.1.99-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:a2a0260cd1fb7bd8b4d4f39dc2444a8d5fd4e0a0c4d5c899810ef1abf99b2d45"}, + {file = "sentencepiece-0.1.99-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:8a1abff4d1ff81c77cac3cc6fefa34fa4b8b371e5ee51cb7e8d1ebc996d05983"}, + {file = "sentencepiece-0.1.99-cp311-cp311-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:004e6a621d4bc88978eecb6ea7959264239a17b70f2cbc348033d8195c9808ec"}, + {file = "sentencepiece-0.1.99-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:db361e03342c41680afae5807590bc88aa0e17cfd1a42696a160e4005fcda03b"}, + {file = "sentencepiece-0.1.99-cp311-cp311-win32.whl", hash = "sha256:2d95e19168875b70df62916eb55428a0cbcb834ac51d5a7e664eda74def9e1e0"}, + {file = "sentencepiece-0.1.99-cp311-cp311-win_amd64.whl", hash = "sha256:f90d73a6f81248a909f55d8e6ef56fec32d559e1e9af045f0b0322637cb8e5c7"}, + {file = "sentencepiece-0.1.99-cp36-cp36m-macosx_10_9_x86_64.whl", hash = "sha256:62e24c81e74bd87a6e0d63c51beb6527e4c0add67e1a17bac18bcd2076afcfeb"}, + {file = "sentencepiece-0.1.99-cp36-cp36m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:57efcc2d51caff20d9573567d9fd3f854d9efe613ed58a439c78c9f93101384a"}, + {file = "sentencepiece-0.1.99-cp36-cp36m-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:6a904c46197993bd1e95b93a6e373dca2f170379d64441041e2e628ad4afb16f"}, + {file = "sentencepiece-0.1.99-cp36-cp36m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:d89adf59854741c0d465f0e1525b388c0d174f611cc04af54153c5c4f36088c4"}, + {file = "sentencepiece-0.1.99-cp36-cp36m-win32.whl", hash = "sha256:47c378146928690d1bc106fdf0da768cebd03b65dd8405aa3dd88f9c81e35dba"}, + {file = "sentencepiece-0.1.99-cp36-cp36m-win_amd64.whl", hash = "sha256:9ba142e7a90dd6d823c44f9870abdad45e6c63958eb60fe44cca6828d3b69da2"}, + {file = "sentencepiece-0.1.99-cp37-cp37m-macosx_10_9_x86_64.whl", hash = "sha256:b7b1a9ae4d7c6f1f867e63370cca25cc17b6f4886729595b885ee07a58d3cec3"}, + {file = "sentencepiece-0.1.99-cp37-cp37m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d0f644c9d4d35c096a538507b2163e6191512460035bf51358794a78515b74f7"}, + {file = "sentencepiece-0.1.99-cp37-cp37m-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:c8843d23a0f686d85e569bd6dcd0dd0e0cbc03731e63497ca6d5bacd18df8b85"}, + {file = "sentencepiece-0.1.99-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:33e6f690a1caebb4867a2e367afa1918ad35be257ecdb3455d2bbd787936f155"}, + {file = "sentencepiece-0.1.99-cp37-cp37m-win32.whl", hash = "sha256:8a321866c2f85da7beac74a824b4ad6ddc2a4c9bccd9382529506d48f744a12c"}, + {file = "sentencepiece-0.1.99-cp37-cp37m-win_amd64.whl", hash = "sha256:c42f753bcfb7661c122a15b20be7f684b61fc8592c89c870adf52382ea72262d"}, + {file = "sentencepiece-0.1.99-cp38-cp38-macosx_10_9_universal2.whl", hash = "sha256:85b476406da69c70586f0bb682fcca4c9b40e5059814f2db92303ea4585c650c"}, + {file = "sentencepiece-0.1.99-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:cfbcfe13c69d3f87b7fcd5da168df7290a6d006329be71f90ba4f56bc77f8561"}, + {file = "sentencepiece-0.1.99-cp38-cp38-macosx_11_0_arm64.whl", hash = "sha256:445b0ec381af1cd4eef95243e7180c63d9c384443c16c4c47a28196bd1cda937"}, + {file = "sentencepiece-0.1.99-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:c6890ea0f2b4703f62d0bf27932e35808b1f679bdb05c7eeb3812b935ba02001"}, + {file = "sentencepiece-0.1.99-cp38-cp38-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:fb71af492b0eefbf9f2501bec97bcd043b6812ab000d119eaf4bd33f9e283d03"}, + {file = "sentencepiece-0.1.99-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:27b866b5bd3ddd54166bbcbf5c8d7dd2e0b397fac8537991c7f544220b1f67bc"}, + {file = "sentencepiece-0.1.99-cp38-cp38-win32.whl", hash = "sha256:b133e8a499eac49c581c3c76e9bdd08c338cc1939e441fee6f92c0ccb5f1f8be"}, + {file = "sentencepiece-0.1.99-cp38-cp38-win_amd64.whl", hash = "sha256:0eaf3591dd0690a87f44f4df129cf8d05d8a4029b5b6709b489b8e27f9a9bcff"}, + {file = "sentencepiece-0.1.99-cp39-cp39-macosx_10_9_universal2.whl", hash = "sha256:38efeda9bbfb55052d482a009c6a37e52f42ebffcea9d3a98a61de7aee356a28"}, + {file = "sentencepiece-0.1.99-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:6c030b081dc1e1bcc9fadc314b19b740715d3d566ad73a482da20d7d46fd444c"}, + {file = "sentencepiece-0.1.99-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:84dbe53e02e4f8a2e45d2ac3e430d5c83182142658e25edd76539b7648928727"}, + {file = "sentencepiece-0.1.99-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:0b0f55d0a0ee1719b4b04221fe0c9f0c3461dc3dabd77a035fa2f4788eb3ef9a"}, + {file = "sentencepiece-0.1.99-cp39-cp39-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:18e800f206cd235dc27dc749299e05853a4e4332e8d3dfd81bf13d0e5b9007d9"}, + {file = "sentencepiece-0.1.99-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:2ae1c40cda8f9d5b0423cfa98542735c0235e7597d79caf318855cdf971b2280"}, + {file = "sentencepiece-0.1.99-cp39-cp39-win32.whl", hash = "sha256:c84ce33af12ca222d14a1cdd37bd76a69401e32bc68fe61c67ef6b59402f4ab8"}, + {file = "sentencepiece-0.1.99-cp39-cp39-win_amd64.whl", hash = "sha256:350e5c74d739973f1c9643edb80f7cc904dc948578bcb1d43c6f2b173e5d18dd"}, + {file = "sentencepiece-0.1.99.tar.gz", hash = "sha256:189c48f5cb2949288f97ccdb97f0473098d9c3dcf5a3d99d4eabe719ec27297f"}, +] + +[[package]] +name = "sympy" +version = "1.12" +description = "Computer algebra system (CAS) in Python" +optional = false +python-versions = ">=3.8" +files = [ + {file = "sympy-1.12-py3-none-any.whl", hash = "sha256:c3588cd4295d0c0f603d0f2ae780587e64e2efeedb3521e46b9bb1d08d184fa5"}, + {file = "sympy-1.12.tar.gz", hash = "sha256:ebf595c8dac3e0fdc4152c51878b498396ec7f30e7a914d6071e674d49420fb8"}, +] + +[package.dependencies] +mpmath = ">=0.19" + +[[package]] +name = "tokenizers" +version = "0.15.2" +description = "" +optional = false +python-versions = ">=3.7" +files = [ + {file = "tokenizers-0.15.2-cp310-cp310-macosx_10_12_x86_64.whl", hash = "sha256:52f6130c9cbf70544287575a985bf44ae1bda2da7e8c24e97716080593638012"}, + {file = "tokenizers-0.15.2-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:054c1cc9c6d68f7ffa4e810b3d5131e0ba511b6e4be34157aa08ee54c2f8d9ee"}, + {file = "tokenizers-0.15.2-cp310-cp310-manylinux_2_12_i686.manylinux2010_i686.whl", hash = "sha256:a9b9b070fdad06e347563b88c278995735292ded1132f8657084989a4c84a6d5"}, + {file = "tokenizers-0.15.2-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:ea621a7eef4b70e1f7a4e84dd989ae3f0eeb50fc8690254eacc08acb623e82f1"}, + {file = "tokenizers-0.15.2-cp310-cp310-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:cf7fd9a5141634fa3aa8d6b7be362e6ae1b4cda60da81388fa533e0b552c98fd"}, + {file = "tokenizers-0.15.2-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:44f2a832cd0825295f7179eaf173381dc45230f9227ec4b44378322d900447c9"}, + {file = "tokenizers-0.15.2-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:8b9ec69247a23747669ec4b0ca10f8e3dfb3545d550258129bd62291aabe8605"}, + {file = "tokenizers-0.15.2-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:40b6a4c78da863ff26dbd5ad9a8ecc33d8a8d97b535172601cf00aee9d7ce9ce"}, + {file = "tokenizers-0.15.2-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:5ab2a4d21dcf76af60e05af8063138849eb1d6553a0d059f6534357bce8ba364"}, + {file = "tokenizers-0.15.2-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:a47acfac7e511f6bbfcf2d3fb8c26979c780a91e06fb5b9a43831b2c0153d024"}, + {file = "tokenizers-0.15.2-cp310-none-win32.whl", hash = "sha256:064ff87bb6acdbd693666de9a4b692add41308a2c0ec0770d6385737117215f2"}, + {file = "tokenizers-0.15.2-cp310-none-win_amd64.whl", hash = "sha256:3b919afe4df7eb6ac7cafd2bd14fb507d3f408db7a68c43117f579c984a73843"}, + {file = "tokenizers-0.15.2-cp311-cp311-macosx_10_12_x86_64.whl", hash = "sha256:89cd1cb93e4b12ff39bb2d626ad77e35209de9309a71e4d3d4672667b4b256e7"}, + {file = "tokenizers-0.15.2-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:cfed5c64e5be23d7ee0f0e98081a25c2a46b0b77ce99a4f0605b1ec43dd481fa"}, + {file = "tokenizers-0.15.2-cp311-cp311-manylinux_2_12_i686.manylinux2010_i686.whl", hash = "sha256:a907d76dcfda37023ba203ab4ceeb21bc5683436ebefbd895a0841fd52f6f6f2"}, + {file = "tokenizers-0.15.2-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:20ea60479de6fc7b8ae756b4b097572372d7e4032e2521c1bbf3d90c90a99ff0"}, + {file = "tokenizers-0.15.2-cp311-cp311-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:48e2b9335be2bc0171df9281385c2ed06a15f5cf121c44094338306ab7b33f2c"}, + {file = "tokenizers-0.15.2-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:112a1dd436d2cc06e6ffdc0b06d55ac019a35a63afd26475205cb4b1bf0bfbff"}, + {file = "tokenizers-0.15.2-cp311-cp311-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:4620cca5c2817177ee8706f860364cc3a8845bc1e291aaf661fb899e5d1c45b0"}, + {file = "tokenizers-0.15.2-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ccd73a82751c523b3fc31ff8194702e4af4db21dc20e55b30ecc2079c5d43cb7"}, + {file = "tokenizers-0.15.2-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:107089f135b4ae7817affe6264f8c7a5c5b4fd9a90f9439ed495f54fcea56fb4"}, + {file = "tokenizers-0.15.2-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:0ff110ecc57b7aa4a594396525a3451ad70988e517237fe91c540997c4e50e29"}, + {file = "tokenizers-0.15.2-cp311-none-win32.whl", hash = "sha256:6d76f00f5c32da36c61f41c58346a4fa7f0a61be02f4301fd30ad59834977cc3"}, + {file = "tokenizers-0.15.2-cp311-none-win_amd64.whl", hash = "sha256:cc90102ed17271cf0a1262babe5939e0134b3890345d11a19c3145184b706055"}, + {file = "tokenizers-0.15.2-cp312-cp312-macosx_10_12_x86_64.whl", hash = "sha256:f86593c18d2e6248e72fb91c77d413a815153b8ea4e31f7cd443bdf28e467670"}, + {file = "tokenizers-0.15.2-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:0774bccc6608eca23eb9d620196687c8b2360624619623cf4ba9dc9bd53e8b51"}, + {file = "tokenizers-0.15.2-cp312-cp312-manylinux_2_12_i686.manylinux2010_i686.whl", hash = "sha256:d0222c5b7c9b26c0b4822a82f6a7011de0a9d3060e1da176f66274b70f846b98"}, + {file = "tokenizers-0.15.2-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:3835738be1de66624fff2f4f6f6684775da4e9c00bde053be7564cbf3545cc66"}, + {file = "tokenizers-0.15.2-cp312-cp312-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:0143e7d9dcd811855c1ce1ab9bf5d96d29bf5e528fd6c7824d0465741e8c10fd"}, + {file = "tokenizers-0.15.2-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:db35825f6d54215f6b6009a7ff3eedee0848c99a6271c870d2826fbbedf31a38"}, + {file = "tokenizers-0.15.2-cp312-cp312-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:3f5e64b0389a2be47091d8cc53c87859783b837ea1a06edd9d8e04004df55a5c"}, + {file = "tokenizers-0.15.2-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:9e0480c452217edd35eca56fafe2029fb4d368b7c0475f8dfa3c5c9c400a7456"}, + {file = "tokenizers-0.15.2-cp312-cp312-musllinux_1_1_aarch64.whl", hash = "sha256:a33ab881c8fe70474980577e033d0bc9a27b7ab8272896e500708b212995d834"}, + {file = "tokenizers-0.15.2-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:a308a607ca9de2c64c1b9ba79ec9a403969715a1b8ba5f998a676826f1a7039d"}, + {file = "tokenizers-0.15.2-cp312-none-win32.whl", hash = "sha256:b8fcfa81bcb9447df582c5bc96a031e6df4da2a774b8080d4f02c0c16b42be0b"}, + {file = "tokenizers-0.15.2-cp312-none-win_amd64.whl", hash = "sha256:38d7ab43c6825abfc0b661d95f39c7f8af2449364f01d331f3b51c94dcff7221"}, + {file = "tokenizers-0.15.2-cp313-cp313-macosx_10_12_x86_64.whl", hash = "sha256:38bfb0204ff3246ca4d5e726e8cc8403bfc931090151e6eede54d0e0cf162ef0"}, + {file = "tokenizers-0.15.2-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:9c861d35e8286a53e06e9e28d030b5a05bcbf5ac9d7229e561e53c352a85b1fc"}, + {file = "tokenizers-0.15.2-cp313-cp313-manylinux_2_12_i686.manylinux2010_i686.whl", hash = "sha256:936bf3842db5b2048eaa53dade907b1160f318e7c90c74bfab86f1e47720bdd6"}, + {file = "tokenizers-0.15.2-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:620beacc3373277700d0e27718aa8b25f7b383eb8001fba94ee00aeea1459d89"}, + {file = "tokenizers-0.15.2-cp313-cp313-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:2735ecbbf37e52db4ea970e539fd2d450d213517b77745114f92867f3fc246eb"}, + {file = "tokenizers-0.15.2-cp313-cp313-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:473c83c5e2359bb81b0b6fde870b41b2764fcdd36d997485e07e72cc3a62264a"}, + {file = "tokenizers-0.15.2-cp313-cp313-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:968fa1fb3c27398b28a4eca1cbd1e19355c4d3a6007f7398d48826bbe3a0f728"}, + {file = "tokenizers-0.15.2-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:865c60ae6eaebdde7da66191ee9b7db52e542ed8ee9d2c653b6d190a9351b980"}, + {file = "tokenizers-0.15.2-cp313-cp313-musllinux_1_1_aarch64.whl", hash = "sha256:7c0d8b52664ab2d4a8d6686eb5effc68b78608a9008f086a122a7b2996befbab"}, + {file = "tokenizers-0.15.2-cp313-cp313-musllinux_1_1_x86_64.whl", hash = "sha256:f33dfbdec3784093a9aebb3680d1f91336c56d86cc70ddf88708251da1fe9064"}, + {file = "tokenizers-0.15.2-cp37-cp37m-macosx_10_12_x86_64.whl", hash = "sha256:d44ba80988ff9424e33e0a49445072ac7029d8c0e1601ad25a0ca5f41ed0c1d6"}, + {file = "tokenizers-0.15.2-cp37-cp37m-macosx_11_0_arm64.whl", hash = "sha256:dce74266919b892f82b1b86025a613956ea0ea62a4843d4c4237be2c5498ed3a"}, + {file = "tokenizers-0.15.2-cp37-cp37m-manylinux_2_12_i686.manylinux2010_i686.whl", hash = "sha256:0ef06b9707baeb98b316577acb04f4852239d856b93e9ec3a299622f6084e4be"}, + {file = "tokenizers-0.15.2-cp37-cp37m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:c73e2e74bbb07910da0d37c326869f34113137b23eadad3fc00856e6b3d9930c"}, + {file = "tokenizers-0.15.2-cp37-cp37m-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:4eeb12daf02a59e29f578a865f55d87cd103ce62bd8a3a5874f8fdeaa82e336b"}, + {file = "tokenizers-0.15.2-cp37-cp37m-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:9ba9f6895af58487ca4f54e8a664a322f16c26bbb442effd01087eba391a719e"}, + {file = "tokenizers-0.15.2-cp37-cp37m-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:ccec77aa7150e38eec6878a493bf8c263ff1fa8a62404e16c6203c64c1f16a26"}, + {file = "tokenizers-0.15.2-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:f3f40604f5042ff210ba82743dda2b6aa3e55aa12df4e9f2378ee01a17e2855e"}, + {file = "tokenizers-0.15.2-cp37-cp37m-musllinux_1_1_aarch64.whl", hash = "sha256:5645938a42d78c4885086767c70923abad047163d809c16da75d6b290cb30bbe"}, + {file = "tokenizers-0.15.2-cp37-cp37m-musllinux_1_1_x86_64.whl", hash = "sha256:05a77cbfebe28a61ab5c3891f9939cc24798b63fa236d84e5f29f3a85a200c00"}, + {file = "tokenizers-0.15.2-cp37-none-win32.whl", hash = "sha256:361abdc068e8afe9c5b818769a48624687fb6aaed49636ee39bec4e95e1a215b"}, + {file = "tokenizers-0.15.2-cp37-none-win_amd64.whl", hash = "sha256:7ef789f83eb0f9baeb4d09a86cd639c0a5518528f9992f38b28e819df397eb06"}, + {file = "tokenizers-0.15.2-cp38-cp38-macosx_10_12_x86_64.whl", hash = "sha256:4fe1f74a902bee74a3b25aff180fbfbf4f8b444ab37c4d496af7afd13a784ed2"}, + {file = "tokenizers-0.15.2-cp38-cp38-macosx_11_0_arm64.whl", hash = "sha256:4c4b89038a684f40a6b15d6b09f49650ac64d951ad0f2a3ea9169687bbf2a8ba"}, + {file = "tokenizers-0.15.2-cp38-cp38-manylinux_2_12_i686.manylinux2010_i686.whl", hash = "sha256:d05a1b06f986d41aed5f2de464c003004b2df8aaf66f2b7628254bcbfb72a438"}, + {file = "tokenizers-0.15.2-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:508711a108684111ec8af89d3a9e9e08755247eda27d0ba5e3c50e9da1600f6d"}, + {file = "tokenizers-0.15.2-cp38-cp38-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:daa348f02d15160cb35439098ac96e3a53bacf35885072611cd9e5be7d333daa"}, + {file = "tokenizers-0.15.2-cp38-cp38-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:494fdbe5932d3416de2a85fc2470b797e6f3226c12845cadf054dd906afd0442"}, + {file = "tokenizers-0.15.2-cp38-cp38-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:c2d60f5246f4da9373f75ff18d64c69cbf60c3bca597290cea01059c336d2470"}, + {file = "tokenizers-0.15.2-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:93268e788825f52de4c7bdcb6ebc1fcd4a5442c02e730faa9b6b08f23ead0e24"}, + {file = "tokenizers-0.15.2-cp38-cp38-musllinux_1_1_aarch64.whl", hash = "sha256:6fc7083ab404019fc9acafe78662c192673c1e696bd598d16dc005bd663a5cf9"}, + {file = "tokenizers-0.15.2-cp38-cp38-musllinux_1_1_x86_64.whl", hash = "sha256:41e39b41e5531d6b2122a77532dbea60e171ef87a3820b5a3888daa847df4153"}, + {file = "tokenizers-0.15.2-cp38-none-win32.whl", hash = "sha256:06cd0487b1cbfabefb2cc52fbd6b1f8d4c37799bd6c6e1641281adaa6b2504a7"}, + {file = "tokenizers-0.15.2-cp38-none-win_amd64.whl", hash = "sha256:5179c271aa5de9c71712e31cb5a79e436ecd0d7532a408fa42a8dbfa4bc23fd9"}, + {file = "tokenizers-0.15.2-cp39-cp39-macosx_10_12_x86_64.whl", hash = "sha256:82f8652a74cc107052328b87ea8b34291c0f55b96d8fb261b3880216a9f9e48e"}, + {file = "tokenizers-0.15.2-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:02458bee6f5f3139f1ebbb6d042b283af712c0981f5bc50edf771d6b762d5e4f"}, + {file = "tokenizers-0.15.2-cp39-cp39-manylinux_2_12_i686.manylinux2010_i686.whl", hash = "sha256:c9a09cd26cca2e1c349f91aa665309ddb48d71636370749414fbf67bc83c5343"}, + {file = "tokenizers-0.15.2-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:158be8ea8554e5ed69acc1ce3fbb23a06060bd4bbb09029431ad6b9a466a7121"}, + {file = "tokenizers-0.15.2-cp39-cp39-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:1ddba9a2b0c8c81633eca0bb2e1aa5b3a15362b1277f1ae64176d0f6eba78ab1"}, + {file = "tokenizers-0.15.2-cp39-cp39-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:3ef5dd1d39797044642dbe53eb2bc56435308432e9c7907728da74c69ee2adca"}, + {file = "tokenizers-0.15.2-cp39-cp39-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:454c203164e07a860dbeb3b1f4a733be52b0edbb4dd2e5bd75023ffa8b49403a"}, + {file = "tokenizers-0.15.2-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:0cf6b7f1d4dc59af960e6ffdc4faffe6460bbfa8dce27a58bf75755ffdb2526d"}, + {file = "tokenizers-0.15.2-cp39-cp39-musllinux_1_1_aarch64.whl", hash = "sha256:2ef09bbc16519f6c25d0c7fc0c6a33a6f62923e263c9d7cca4e58b8c61572afb"}, + {file = "tokenizers-0.15.2-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:c9a2ebdd2ad4ec7a68e7615086e633857c85e2f18025bd05d2a4399e6c5f7169"}, + {file = "tokenizers-0.15.2-cp39-none-win32.whl", hash = "sha256:918fbb0eab96fe08e72a8c2b5461e9cce95585d82a58688e7f01c2bd546c79d0"}, + {file = "tokenizers-0.15.2-cp39-none-win_amd64.whl", hash = "sha256:524e60da0135e106b254bd71f0659be9f89d83f006ea9093ce4d1fab498c6d0d"}, + {file = "tokenizers-0.15.2-pp310-pypy310_pp73-macosx_10_12_x86_64.whl", hash = "sha256:6a9b648a58281c4672212fab04e60648fde574877d0139cd4b4f93fe28ca8944"}, + {file = "tokenizers-0.15.2-pp310-pypy310_pp73-macosx_11_0_arm64.whl", hash = "sha256:7c7d18b733be6bbca8a55084027f7be428c947ddf871c500ee603e375013ffba"}, + {file = "tokenizers-0.15.2-pp310-pypy310_pp73-manylinux_2_12_i686.manylinux2010_i686.whl", hash = "sha256:13ca3611de8d9ddfbc4dc39ef54ab1d2d4aaa114ac8727dfdc6a6ec4be017378"}, + {file = "tokenizers-0.15.2-pp310-pypy310_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:237d1bf3361cf2e6463e6c140628e6406766e8b27274f5fcc62c747ae3c6f094"}, + {file = "tokenizers-0.15.2-pp310-pypy310_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:67a0fe1e49e60c664915e9fb6b0cb19bac082ab1f309188230e4b2920230edb3"}, + {file = "tokenizers-0.15.2-pp310-pypy310_pp73-musllinux_1_1_aarch64.whl", hash = "sha256:4e022fe65e99230b8fd89ebdfea138c24421f91c1a4f4781a8f5016fd5cdfb4d"}, + {file = "tokenizers-0.15.2-pp310-pypy310_pp73-musllinux_1_1_x86_64.whl", hash = "sha256:d857be2df69763362ac699f8b251a8cd3fac9d21893de129bc788f8baaef2693"}, + {file = "tokenizers-0.15.2-pp37-pypy37_pp73-macosx_10_12_x86_64.whl", hash = "sha256:708bb3e4283177236309e698da5fcd0879ce8fd37457d7c266d16b550bcbbd18"}, + {file = "tokenizers-0.15.2-pp37-pypy37_pp73-manylinux_2_12_i686.manylinux2010_i686.whl", hash = "sha256:64c35e09e9899b72a76e762f9854e8750213f67567787d45f37ce06daf57ca78"}, + {file = "tokenizers-0.15.2-pp37-pypy37_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:c1257f4394be0d3b00de8c9e840ca5601d0a4a8438361ce9c2b05c7d25f6057b"}, + {file = "tokenizers-0.15.2-pp37-pypy37_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:02272fe48280e0293a04245ca5d919b2c94a48b408b55e858feae9618138aeda"}, + {file = "tokenizers-0.15.2-pp37-pypy37_pp73-musllinux_1_1_aarch64.whl", hash = "sha256:dc3ad9ebc76eabe8b1d7c04d38be884b8f9d60c0cdc09b0aa4e3bcf746de0388"}, + {file = "tokenizers-0.15.2-pp37-pypy37_pp73-musllinux_1_1_x86_64.whl", hash = "sha256:32e16bdeffa7c4f46bf2152172ca511808b952701d13e7c18833c0b73cb5c23f"}, + {file = "tokenizers-0.15.2-pp38-pypy38_pp73-macosx_10_12_x86_64.whl", hash = "sha256:fb16ba563d59003028b678d2361a27f7e4ae0ab29c7a80690efa20d829c81fdb"}, + {file = "tokenizers-0.15.2-pp38-pypy38_pp73-macosx_11_0_arm64.whl", hash = "sha256:2277c36d2d6cdb7876c274547921a42425b6810d38354327dd65a8009acf870c"}, + {file = "tokenizers-0.15.2-pp38-pypy38_pp73-manylinux_2_12_i686.manylinux2010_i686.whl", hash = "sha256:1cf75d32e8d250781940d07f7eece253f2fe9ecdb1dc7ba6e3833fa17b82fcbc"}, + {file = "tokenizers-0.15.2-pp38-pypy38_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:f1b3b31884dc8e9b21508bb76da80ebf7308fdb947a17affce815665d5c4d028"}, + {file = "tokenizers-0.15.2-pp38-pypy38_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:b10122d8d8e30afb43bb1fe21a3619f62c3e2574bff2699cf8af8b0b6c5dc4a3"}, + {file = "tokenizers-0.15.2-pp38-pypy38_pp73-musllinux_1_1_aarch64.whl", hash = "sha256:d88b96ff0fe8e91f6ef01ba50b0d71db5017fa4e3b1d99681cec89a85faf7bf7"}, + {file = "tokenizers-0.15.2-pp38-pypy38_pp73-musllinux_1_1_x86_64.whl", hash = "sha256:37aaec5a52e959892870a7c47cef80c53797c0db9149d458460f4f31e2fb250e"}, + {file = "tokenizers-0.15.2-pp39-pypy39_pp73-macosx_10_12_x86_64.whl", hash = "sha256:e2ea752f2b0fe96eb6e2f3adbbf4d72aaa1272079b0dfa1145507bd6a5d537e6"}, + {file = "tokenizers-0.15.2-pp39-pypy39_pp73-macosx_11_0_arm64.whl", hash = "sha256:4b19a808d8799fda23504a5cd31d2f58e6f52f140380082b352f877017d6342b"}, + {file = "tokenizers-0.15.2-pp39-pypy39_pp73-manylinux_2_12_i686.manylinux2010_i686.whl", hash = "sha256:64c86e5e068ac8b19204419ed8ca90f9d25db20578f5881e337d203b314f4104"}, + {file = "tokenizers-0.15.2-pp39-pypy39_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:de19c4dc503c612847edf833c82e9f73cd79926a384af9d801dcf93f110cea4e"}, + {file = "tokenizers-0.15.2-pp39-pypy39_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ea09acd2fe3324174063d61ad620dec3bcf042b495515f27f638270a7d466e8b"}, + {file = "tokenizers-0.15.2-pp39-pypy39_pp73-musllinux_1_1_aarch64.whl", hash = "sha256:cf27fd43472e07b57cf420eee1e814549203d56de00b5af8659cb99885472f1f"}, + {file = "tokenizers-0.15.2-pp39-pypy39_pp73-musllinux_1_1_x86_64.whl", hash = "sha256:7ca22bd897537a0080521445d91a58886c8c04084a6a19e6c78c586e0cfa92a5"}, + {file = "tokenizers-0.15.2.tar.gz", hash = "sha256:e6e9c6e019dd5484be5beafc775ae6c925f4c69a3487040ed09b45e13df2cb91"}, +] + +[package.dependencies] +huggingface_hub = ">=0.16.4,<1.0" + +[package.extras] +dev = ["tokenizers[testing]"] +docs = ["setuptools_rust", "sphinx", "sphinx_rtd_theme"] +testing = ["black (==22.3)", "datasets", "numpy", "pytest", "requests"] + +[[package]] +name = "torch" +version = "2.2.1+cpu" +description = "Tensors and Dynamic neural networks in Python with strong GPU acceleration" +optional = false +python-versions = ">=3.8.0" +files = [ + {file = "torch-2.2.1+cpu-cp310-cp310-linux_x86_64.whl", hash = "sha256:5d82422cf04797f1b2a8574b64a916070ec83eef58ad4900615ee0218d7b8b8e"}, + {file = "torch-2.2.1+cpu-cp310-cp310-win_amd64.whl", hash = "sha256:f8914dd0f5f0e5c66fdecd9559403eea9feac82d1ea639b672fde0073c6addbd"}, + {file = "torch-2.2.1+cpu-cp311-cp311-linux_x86_64.whl", hash = "sha256:6bc973d5632374b92b4b293817b4d2ff8c8ce1c784c748b471dba1fffcd9c333"}, + {file = "torch-2.2.1+cpu-cp311-cp311-win_amd64.whl", hash = "sha256:abdec34b0ade8fca0520055e72c3094425ae0ef210718e9c0278121cd3608c32"}, + {file = "torch-2.2.1+cpu-cp312-cp312-linux_x86_64.whl", hash = "sha256:d7339580135da4105c1244a8621faa076990409afeab5a7b642c3c1ee70a5622"}, + {file = "torch-2.2.1+cpu-cp312-cp312-win_amd64.whl", hash = "sha256:039128fcb5548122465b15f679b8831c47d14f0d6c28c1f1b631f8019c104720"}, + {file = "torch-2.2.1+cpu-cp38-cp38-linux_x86_64.whl", hash = "sha256:2b447f7bb50b393b4544b4036d587e39ab524d4353e77c197f6a2727f22b0d47"}, + {file = "torch-2.2.1+cpu-cp38-cp38-win_amd64.whl", hash = "sha256:2ccdf3e5f71e6426ea9e34d21c3cc333b29d4f48299b981d28aeb5112b5495e1"}, + {file = "torch-2.2.1+cpu-cp39-cp39-linux_x86_64.whl", hash = "sha256:2fb340b289760040a16a77a6d70b8a48961abba1822e6f58705c97c80befa03e"}, + {file = "torch-2.2.1+cpu-cp39-cp39-win_amd64.whl", hash = "sha256:e03dc4654ecceeb5b03f0a6f60b342c0e0d267b3ebc61e4f672cace1df8cd930"}, +] + +[package.dependencies] +filelock = "*" +fsspec = "*" +jinja2 = "*" +networkx = "*" +sympy = "*" +typing-extensions = ">=4.8.0" + +[package.extras] +opt-einsum = ["opt-einsum (>=3.3)"] +optree = ["optree (>=0.9.1)"] + +[package.source] +type = "legacy" +url = "https://download.pytorch.org/whl/cpu" +reference = "pytorch" + +[[package]] +name = "tqdm" +version = "4.66.2" +description = "Fast, Extensible Progress Meter" +optional = false +python-versions = ">=3.7" +files = [ + {file = "tqdm-4.66.2-py3-none-any.whl", hash = "sha256:1ee4f8a893eb9bef51c6e35730cebf234d5d0b6bd112b0271e10ed7c24a02bd9"}, + {file = "tqdm-4.66.2.tar.gz", hash = "sha256:6cd52cdf0fef0e0f543299cfc96fec90d7b8a7e88745f411ec33eb44d5ed3531"}, +] + +[package.dependencies] +colorama = {version = "*", markers = "platform_system == \"Windows\""} + +[package.extras] +dev = ["pytest (>=6)", "pytest-cov", "pytest-timeout", "pytest-xdist"] +notebook = ["ipywidgets (>=6)"] +slack = ["slack-sdk"] +telegram = ["requests"] + +[[package]] +name = "transformers" +version = "4.38.1" +description = "State-of-the-art Machine Learning for JAX, PyTorch and TensorFlow" +optional = false +python-versions = ">=3.8.0" +files = [ + {file = "transformers-4.38.1-py3-none-any.whl", hash = "sha256:a7a9265fb060183e9d975cbbadc4d531b10281589c43f6d07563f86322728973"}, + {file = "transformers-4.38.1.tar.gz", hash = "sha256:86dc84ccbe36123647e84cbd50fc31618c109a41e6be92514b064ab55bf1304c"}, +] + +[package.dependencies] +filelock = "*" +huggingface-hub = ">=0.19.3,<1.0" +numpy = ">=1.17" +packaging = ">=20.0" +pyyaml = ">=5.1" +regex = "!=2019.12.17" +requests = "*" +safetensors = ">=0.4.1" +tokenizers = ">=0.14,<0.19" +tqdm = ">=4.27" + +[package.extras] +accelerate = ["accelerate (>=0.21.0)"] +agents = ["Pillow (>=10.0.1,<=15.0)", "accelerate (>=0.21.0)", "datasets (!=2.5.0)", "diffusers", "opencv-python", "sentencepiece (>=0.1.91,!=0.1.92)", "torch"] +all = ["Pillow (>=10.0.1,<=15.0)", "accelerate (>=0.21.0)", "av (==9.2.0)", "codecarbon (==1.2.0)", "decord (==0.6.0)", "flax (>=0.4.1,<=0.7.0)", "jax (>=0.4.1,<=0.4.13)", "jaxlib (>=0.4.1,<=0.4.13)", "kenlm", "keras-nlp (>=0.3.1)", "librosa", "onnxconverter-common", "optax (>=0.0.8,<=0.1.4)", "optuna", "phonemizer", "protobuf", "pyctcdecode (>=0.4.0)", "ray[tune] (>=2.7.0)", "sentencepiece (>=0.1.91,!=0.1.92)", "sigopt", "tensorflow (>=2.6,<2.16)", "tensorflow-text (<2.16)", "tf2onnx", "timm", "tokenizers (>=0.14,<0.19)", "torch", "torchaudio", "torchvision"] +audio = ["kenlm", "librosa", "phonemizer", "pyctcdecode (>=0.4.0)"] +codecarbon = ["codecarbon (==1.2.0)"] +deepspeed = ["accelerate (>=0.21.0)", "deepspeed (>=0.9.3)"] +deepspeed-testing = ["GitPython (<3.1.19)", "accelerate (>=0.21.0)", "beautifulsoup4", "cookiecutter (==1.7.3)", "datasets (!=2.5.0)", "deepspeed (>=0.9.3)", "dill (<0.3.5)", "evaluate (>=0.2.0)", "faiss-cpu", "hf-doc-builder (>=0.3.0)", "nltk", "optuna", "parameterized", "protobuf", "psutil", "pydantic", "pytest (>=7.2.0,<8.0.0)", "pytest-timeout", "pytest-xdist", "rjieba", "rouge-score (!=0.0.7,!=0.0.8,!=0.1,!=0.1.1)", "ruff (==0.1.5)", "sacrebleu (>=1.4.12,<2.0.0)", "sacremoses", "sentencepiece (>=0.1.91,!=0.1.92)", "tensorboard", "timeout-decorator"] +dev = ["GitPython (<3.1.19)", "Pillow (>=10.0.1,<=15.0)", "accelerate (>=0.21.0)", "av (==9.2.0)", "beautifulsoup4", "codecarbon (==1.2.0)", "cookiecutter (==1.7.3)", "datasets (!=2.5.0)", "decord (==0.6.0)", "dill (<0.3.5)", "evaluate (>=0.2.0)", "faiss-cpu", "flax (>=0.4.1,<=0.7.0)", "fugashi (>=1.0)", "hf-doc-builder", "hf-doc-builder (>=0.3.0)", "ipadic (>=1.0.0,<2.0)", "isort (>=5.5.4)", "jax (>=0.4.1,<=0.4.13)", "jaxlib (>=0.4.1,<=0.4.13)", "kenlm", "keras-nlp (>=0.3.1)", "librosa", "nltk", "onnxconverter-common", "optax (>=0.0.8,<=0.1.4)", "optuna", "parameterized", "phonemizer", "protobuf", "psutil", "pyctcdecode (>=0.4.0)", "pydantic", "pytest (>=7.2.0,<8.0.0)", "pytest-timeout", "pytest-xdist", "ray[tune] (>=2.7.0)", "rhoknp (>=1.1.0,<1.3.1)", "rjieba", "rouge-score (!=0.0.7,!=0.0.8,!=0.1,!=0.1.1)", "ruff (==0.1.5)", "sacrebleu (>=1.4.12,<2.0.0)", "sacremoses", "scikit-learn", "sentencepiece (>=0.1.91,!=0.1.92)", "sigopt", "sudachidict-core (>=20220729)", "sudachipy (>=0.6.6)", "tensorboard", "tensorflow (>=2.6,<2.16)", "tensorflow-text (<2.16)", "tf2onnx", "timeout-decorator", "timm", "tokenizers (>=0.14,<0.19)", "torch", "torchaudio", "torchvision", "unidic (>=1.0.2)", "unidic-lite (>=1.0.7)", "urllib3 (<2.0.0)"] +dev-tensorflow = ["GitPython (<3.1.19)", "Pillow (>=10.0.1,<=15.0)", "beautifulsoup4", "cookiecutter (==1.7.3)", "datasets (!=2.5.0)", "dill (<0.3.5)", "evaluate (>=0.2.0)", "faiss-cpu", "hf-doc-builder", "hf-doc-builder (>=0.3.0)", "isort (>=5.5.4)", "kenlm", "keras-nlp (>=0.3.1)", "librosa", "nltk", "onnxconverter-common", "onnxruntime (>=1.4.0)", "onnxruntime-tools (>=1.4.2)", "parameterized", "phonemizer", "protobuf", "psutil", "pyctcdecode (>=0.4.0)", "pydantic", "pytest (>=7.2.0,<8.0.0)", "pytest-timeout", "pytest-xdist", "rjieba", "rouge-score (!=0.0.7,!=0.0.8,!=0.1,!=0.1.1)", "ruff (==0.1.5)", "sacrebleu (>=1.4.12,<2.0.0)", "sacremoses", "scikit-learn", "sentencepiece (>=0.1.91,!=0.1.92)", "tensorboard", "tensorflow (>=2.6,<2.16)", "tensorflow-text (<2.16)", "tf2onnx", "timeout-decorator", "tokenizers (>=0.14,<0.19)", "urllib3 (<2.0.0)"] +dev-torch = ["GitPython (<3.1.19)", "Pillow (>=10.0.1,<=15.0)", "accelerate (>=0.21.0)", "beautifulsoup4", "codecarbon (==1.2.0)", "cookiecutter (==1.7.3)", "datasets (!=2.5.0)", "dill (<0.3.5)", "evaluate (>=0.2.0)", "faiss-cpu", "fugashi (>=1.0)", "hf-doc-builder", "hf-doc-builder (>=0.3.0)", "ipadic (>=1.0.0,<2.0)", "isort (>=5.5.4)", "kenlm", "librosa", "nltk", "onnxruntime (>=1.4.0)", "onnxruntime-tools (>=1.4.2)", "optuna", "parameterized", "phonemizer", "protobuf", "psutil", "pyctcdecode (>=0.4.0)", "pydantic", "pytest (>=7.2.0,<8.0.0)", "pytest-timeout", "pytest-xdist", "ray[tune] (>=2.7.0)", "rhoknp (>=1.1.0,<1.3.1)", "rjieba", "rouge-score (!=0.0.7,!=0.0.8,!=0.1,!=0.1.1)", "ruff (==0.1.5)", "sacrebleu (>=1.4.12,<2.0.0)", "sacremoses", "scikit-learn", "sentencepiece (>=0.1.91,!=0.1.92)", "sigopt", "sudachidict-core (>=20220729)", "sudachipy (>=0.6.6)", "tensorboard", "timeout-decorator", "timm", "tokenizers (>=0.14,<0.19)", "torch", "torchaudio", "torchvision", "unidic (>=1.0.2)", "unidic-lite (>=1.0.7)", "urllib3 (<2.0.0)"] +docs = ["Pillow (>=10.0.1,<=15.0)", "accelerate (>=0.21.0)", "av (==9.2.0)", "codecarbon (==1.2.0)", "decord (==0.6.0)", "flax (>=0.4.1,<=0.7.0)", "hf-doc-builder", "jax (>=0.4.1,<=0.4.13)", "jaxlib (>=0.4.1,<=0.4.13)", "kenlm", "keras-nlp (>=0.3.1)", "librosa", "onnxconverter-common", "optax (>=0.0.8,<=0.1.4)", "optuna", "phonemizer", "protobuf", "pyctcdecode (>=0.4.0)", "ray[tune] (>=2.7.0)", "sentencepiece (>=0.1.91,!=0.1.92)", "sigopt", "tensorflow (>=2.6,<2.16)", "tensorflow-text (<2.16)", "tf2onnx", "timm", "tokenizers (>=0.14,<0.19)", "torch", "torchaudio", "torchvision"] +docs-specific = ["hf-doc-builder"] +flax = ["flax (>=0.4.1,<=0.7.0)", "jax (>=0.4.1,<=0.4.13)", "jaxlib (>=0.4.1,<=0.4.13)", "optax (>=0.0.8,<=0.1.4)"] +flax-speech = ["kenlm", "librosa", "phonemizer", "pyctcdecode (>=0.4.0)"] +ftfy = ["ftfy"] +integrations = ["optuna", "ray[tune] (>=2.7.0)", "sigopt"] +ja = ["fugashi (>=1.0)", "ipadic (>=1.0.0,<2.0)", "rhoknp (>=1.1.0,<1.3.1)", "sudachidict-core (>=20220729)", "sudachipy (>=0.6.6)", "unidic (>=1.0.2)", "unidic-lite (>=1.0.7)"] +modelcreation = ["cookiecutter (==1.7.3)"] +natten = ["natten (>=0.14.6,<0.15.0)"] +onnx = ["onnxconverter-common", "onnxruntime (>=1.4.0)", "onnxruntime-tools (>=1.4.2)", "tf2onnx"] +onnxruntime = ["onnxruntime (>=1.4.0)", "onnxruntime-tools (>=1.4.2)"] +optuna = ["optuna"] +quality = ["GitPython (<3.1.19)", "datasets (!=2.5.0)", "hf-doc-builder (>=0.3.0)", "isort (>=5.5.4)", "ruff (==0.1.5)", "urllib3 (<2.0.0)"] +ray = ["ray[tune] (>=2.7.0)"] +retrieval = ["datasets (!=2.5.0)", "faiss-cpu"] +sagemaker = ["sagemaker (>=2.31.0)"] +sentencepiece = ["protobuf", "sentencepiece (>=0.1.91,!=0.1.92)"] +serving = ["fastapi", "pydantic", "starlette", "uvicorn"] +sigopt = ["sigopt"] +sklearn = ["scikit-learn"] +speech = ["kenlm", "librosa", "phonemizer", "pyctcdecode (>=0.4.0)", "torchaudio"] +testing = ["GitPython (<3.1.19)", "beautifulsoup4", "cookiecutter (==1.7.3)", "datasets (!=2.5.0)", "dill (<0.3.5)", "evaluate (>=0.2.0)", "faiss-cpu", "hf-doc-builder (>=0.3.0)", "nltk", "parameterized", "protobuf", "psutil", "pydantic", "pytest (>=7.2.0,<8.0.0)", "pytest-timeout", "pytest-xdist", "rjieba", "rouge-score (!=0.0.7,!=0.0.8,!=0.1,!=0.1.1)", "ruff (==0.1.5)", "sacrebleu (>=1.4.12,<2.0.0)", "sacremoses", "tensorboard", "timeout-decorator"] +tf = ["keras-nlp (>=0.3.1)", "onnxconverter-common", "tensorflow (>=2.6,<2.16)", "tensorflow-text (<2.16)", "tf2onnx"] +tf-cpu = ["keras-nlp (>=0.3.1)", "onnxconverter-common", "tensorflow-cpu (>=2.6,<2.16)", "tensorflow-text (<2.16)", "tf2onnx"] +tf-speech = ["kenlm", "librosa", "phonemizer", "pyctcdecode (>=0.4.0)"] +timm = ["timm"] +tokenizers = ["tokenizers (>=0.14,<0.19)"] +torch = ["accelerate (>=0.21.0)", "torch"] +torch-speech = ["kenlm", "librosa", "phonemizer", "pyctcdecode (>=0.4.0)", "torchaudio"] +torch-vision = ["Pillow (>=10.0.1,<=15.0)", "torchvision"] +torchhub = ["filelock", "huggingface-hub (>=0.19.3,<1.0)", "importlib-metadata", "numpy (>=1.17)", "packaging (>=20.0)", "protobuf", "regex (!=2019.12.17)", "requests", "sentencepiece (>=0.1.91,!=0.1.92)", "tokenizers (>=0.14,<0.19)", "torch", "tqdm (>=4.27)"] +video = ["av (==9.2.0)", "decord (==0.6.0)"] +vision = ["Pillow (>=10.0.1,<=15.0)"] + +[[package]] +name = "typing-extensions" +version = "4.9.0" +description = "Backported and Experimental Type Hints for Python 3.8+" +optional = false +python-versions = ">=3.8" +files = [ + {file = "typing_extensions-4.9.0-py3-none-any.whl", hash = "sha256:af72aea155e91adfc61c3ae9e0e342dbc0cba726d6cba4b6c72c1f34e47291cd"}, + {file = "typing_extensions-4.9.0.tar.gz", hash = "sha256:23478f88c37f27d76ac8aee6c905017a143b0b1b886c3c9f66bc2fd94f9f5783"}, +] + +[[package]] +name = "urllib3" +version = "2.2.1" +description = "HTTP library with thread-safe connection pooling, file post, and more." +optional = false +python-versions = ">=3.8" +files = [ + {file = "urllib3-2.2.1-py3-none-any.whl", hash = "sha256:450b20ec296a467077128bff42b73080516e71b56ff59a60a02bef2232c4fa9d"}, + {file = "urllib3-2.2.1.tar.gz", hash = "sha256:d0570876c61ab9e520d776c38acbbb5b05a776d3f9ff98a5c8fd5162a444cf19"}, +] + +[package.extras] +brotli = ["brotli (>=1.0.9)", "brotlicffi (>=0.8.0)"] +h2 = ["h2 (>=4,<5)"] +socks = ["pysocks (>=1.5.6,!=1.5.7,<2.0)"] +zstd = ["zstandard (>=0.18.0)"] + +[[package]] +name = "wcwidth" +version = "0.2.13" +description = "Measures the displayed width of unicode strings in a terminal" +optional = false +python-versions = "*" +files = [ + {file = "wcwidth-0.2.13-py2.py3-none-any.whl", hash = "sha256:3da69048e4540d84af32131829ff948f1e022c1c6bdb8d6102117aac784f6859"}, + {file = "wcwidth-0.2.13.tar.gz", hash = "sha256:72ea0c06399eb286d978fdedb6923a9eb47e1c486ce63e9b4e64fc18303972b5"}, +] + +[metadata] +lock-version = "2.0" +python-versions = ">=3.9" +content-hash = "c8c4cc87637266a7b85debcbafa8887c5ad81cc8ef40e98a3f52c7c50af05c03" diff --git a/pyproject.toml b/pyproject.toml new file mode 100644 index 000000000..25e2e20b2 --- /dev/null +++ b/pyproject.toml @@ -0,0 +1,44 @@ +[tool.poetry] +name = "llama-cpp-scripts" +version = "0.0.0" +description = "Scripts that ship with llama.cpp" +authors = ["GGML "] +readme = "README.md" +homepage = "https://ggml.ai" +repository = "https://github.com/ggerganov/llama.cpp" +keywords = ["ggml", "gguf", "llama.cpp"] +packages = [{ include = "*.py", from = "." }] +classifiers = [ + "Programming Language :: Python :: 3", + "License :: OSI Approved :: MIT License", + "Operating System :: OS Independent", +] + +[tool.poetry.dependencies] +python = ">=3.9" +numpy = "^1.25.0" +sentencepiece = ">=0.1.98,<0.2.0" +transformers = ">=4.35.2,<5.0.0" +protobuf = ">=4.21.0,<5.0.0" +gguf = { path = "./gguf-py" } +torch = { version = "^2.2.0", source = "pytorch" } + +[tool.poetry.dev-dependencies] +pytest = "^5.2" + + +# Force wheel + cpu +# For discussion and context see https://github.com/python-poetry/poetry#6409 +[[tool.poetry.source]] +name = "pytorch" +url = "https://download.pytorch.org/whl/cpu" +priority = "explicit" + +[build-system] +requires = ["poetry-core>=1.0.0"] +build-backend = "poetry.core.masonry.api" + +[tool.poetry.scripts] +llama-convert-hf-to-gguf = "convert_hf_to_gguf:main" +llama-convert-llama-ggml-to-gguf = "convert_llama_ggml_to_gguf:main" +llama-ggml-vk-generate-shaders = "ggml_vk_generate_shaders:main" diff --git a/requirements/requirements-convert_hf_to_gguf.txt b/requirements/requirements-convert_hf_to_gguf.txt new file mode 100644 index 000000000..653355c07 --- /dev/null +++ b/requirements/requirements-convert_hf_to_gguf.txt @@ -0,0 +1,2 @@ +-r ./requirements-convert_legacy_llama.txt +torch~=2.2.1 diff --git a/requirements/requirements-convert_hf_to_gguf_update.txt b/requirements/requirements-convert_hf_to_gguf_update.txt new file mode 100644 index 000000000..653355c07 --- /dev/null +++ b/requirements/requirements-convert_hf_to_gguf_update.txt @@ -0,0 +1,2 @@ +-r ./requirements-convert_legacy_llama.txt +torch~=2.2.1 diff --git a/requirements/requirements-convert_llama_ggml_to_gguf.txt b/requirements/requirements-convert_llama_ggml_to_gguf.txt new file mode 100644 index 000000000..afe2747d4 --- /dev/null +++ b/requirements/requirements-convert_llama_ggml_to_gguf.txt @@ -0,0 +1 @@ +-r ./requirements-convert_legacy_llama.txt diff --git a/src/llama.cpp b/src/llama.cpp index dc9c4c0ea..b4c5add7d 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -109,8 +109,10 @@ #define LLAMA_ATTRIBUTE_FORMAT(...) #endif +// bump if necessary #define LLAMA_MAX_NODES 8192 -#define LLAMA_MAX_EXPERTS 160 +#define LLAMA_MAX_LAYERS 256 +#define LLAMA_MAX_EXPERTS 160 // DeepSeekV2 // // logging @@ -248,10 +250,12 @@ enum llm_arch { LLM_ARCH_COMMAND_R, LLM_ARCH_DBRX, LLM_ARCH_OLMO, + LLM_ARCH_OPENELM, LLM_ARCH_ARCTIC, LLM_ARCH_DEEPSEEK2, LLM_ARCH_BITNET, LLM_ARCH_T5, + LLM_ARCH_JAIS, LLM_ARCH_UNKNOWN, }; @@ -289,10 +293,12 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_COMMAND_R, "command-r" }, { LLM_ARCH_DBRX, "dbrx" }, { LLM_ARCH_OLMO, "olmo" }, + { LLM_ARCH_OPENELM, "openelm" }, { LLM_ARCH_ARCTIC, "arctic" }, { LLM_ARCH_DEEPSEEK2, "deepseek2" }, { LLM_ARCH_BITNET, "bitnet" }, { LLM_ARCH_T5, "t5" }, + { LLM_ARCH_JAIS, "jais" }, { LLM_ARCH_UNKNOWN, "(unknown)" }, }; @@ -1156,6 +1162,22 @@ static const std::map> LLM_TENSOR_NA { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, }, }, + { + LLM_ARCH_OPENELM, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, + { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" }, + { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, { LLM_ARCH_ARCTIC, { @@ -1260,6 +1282,21 @@ static const std::map> LLM_TENSOR_NA { LLM_TENSOR_ENC_FFN_UP, "enc.blk.%d.ffn_up" }, }, }, + { + LLM_ARCH_JAIS, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + }, + }, { LLM_ARCH_UNKNOWN, { @@ -2055,14 +2092,23 @@ enum e_model { MODEL_17M, MODEL_22M, MODEL_33M, + MODEL_60M, MODEL_70M, + MODEL_80M, MODEL_109M, MODEL_137M, MODEL_160M, + MODEL_220M, + MODEL_250M, + MODEL_270M, MODEL_335M, MODEL_410M, + MODEL_450M, + MODEL_770M, + MODEL_780M, MODEL_0_5B, MODEL_1B, + MODEL_1_3B, MODEL_1_4B, MODEL_2B, MODEL_2_8B, @@ -2071,6 +2117,7 @@ enum e_model { MODEL_6_9B, MODEL_7B, MODEL_8B, + MODEL_11B, MODEL_12B, MODEL_13B, MODEL_14B, @@ -2111,17 +2158,19 @@ struct llama_hparams { uint32_t n_vocab; uint32_t n_ctx_train; // context size the model was trained on uint32_t n_embd; - uint32_t n_head; - uint32_t n_head_kv; uint32_t n_layer; uint32_t n_rot; uint32_t n_swa = 0; // sliding window attention (SWA) uint32_t n_embd_head_k; // dimension of keys (d_k). d_q is assumed to be the same, but there are n_head q heads, and only n_head_kv k-v heads uint32_t n_embd_head_v; // dimension of values (d_v) aka n_embd_head - uint32_t n_ff; uint32_t n_expert = 0; uint32_t n_expert_used = 0; uint32_t n_vocab_type = 0; // for BERT-style token types + uint32_t n_rel_attn_bkts = 0; + + std::array n_head_arr; + std::array n_head_kv_arr; + std::array n_ff_arr; uint32_t n_layer_dense_lead = 0; uint32_t n_lora_q = 0; @@ -2157,6 +2206,10 @@ struct llama_hparams { bool use_alibi = false; bool attn_soft_cap = false; + // needed by encoder-decoder models (e.g. T5, FLAN-T5) + // ref: https://github.com/ggerganov/llama.cpp/pull/8141 + llama_token dec_start_token_id = -1; + enum llama_pooling_type pooling_type = LLAMA_POOLING_TYPE_NONE; enum llama_rope_type rope_type = LLAMA_ROPE_TYPE_NONE; enum llama_rope_scaling_type rope_scaling_type_train = LLAMA_ROPE_SCALING_TYPE_NONE; @@ -2166,17 +2219,19 @@ struct llama_hparams { if (this->n_vocab != other.n_vocab) return true; if (this->n_ctx_train != other.n_ctx_train) return true; if (this->n_embd != other.n_embd) return true; - if (this->n_head != other.n_head) return true; - if (this->n_head_kv != other.n_head_kv) return true; if (this->n_layer != other.n_layer) return true; if (this->n_rot != other.n_rot) return true; if (this->n_swa != other.n_swa) return true; if (this->n_embd_head_k != other.n_embd_head_k) return true; if (this->n_embd_head_v != other.n_embd_head_v) return true; - if (this->n_ff != other.n_ff) return true; if (this->n_expert != other.n_expert) return true; if (this->n_expert_used != other.n_expert_used) return true; + if (this->n_head_arr != other.n_head_arr) return true; + if (this->n_head_kv_arr != other.n_head_kv_arr) return true; + if (this->n_ff_arr != other.n_ff_arr) return true; + + if (this->n_rel_attn_bkts != other.n_rel_attn_bkts) return true; if (this->n_layer_dense_lead != other.n_layer_dense_lead) return true; if (this->n_lora_q != other.n_lora_q) return true; if (this->n_lora_kv != other.n_lora_kv) return true; @@ -2192,6 +2247,8 @@ struct llama_hparams { if (this->ssm_d_state != other.ssm_d_state) return true; if (this->ssm_dt_rank != other.ssm_dt_rank) return true; + if (this->dec_start_token_id != other.dec_start_token_id) return true; + const float EPSILON = 1e-9f; if (!is_float_close(this->f_norm_eps, other.f_norm_eps, EPSILON)) return true; @@ -2205,18 +2262,53 @@ struct llama_hparams { return false; } - uint32_t n_gqa() const { + uint32_t n_head(uint32_t il = 0) const { + if (il < n_layer) { + return n_head_arr[il]; + } + + GGML_ASSERT(false); + return 0; + } + + uint32_t n_head_kv(uint32_t il = 0) const { + if (il < n_layer) { + return n_head_kv_arr[il]; + } + + GGML_ASSERT(false); + return 0; + } + + uint32_t n_ff(uint32_t il = 0) const { + if (il < n_layer) { + return n_ff_arr[il]; + } + + GGML_ASSERT(false); + return 0; + } + + uint32_t n_gqa(uint32_t il = 0) const { + const uint32_t n_head = this->n_head(il); + const uint32_t n_head_kv = this->n_head_kv(il); + if (n_head_kv == 0) { return 0; } + return n_head/n_head_kv; } - uint32_t n_embd_k_gqa() const { // dimension of key embeddings across all k-v heads + uint32_t n_embd_k_gqa(uint32_t il = 0) const { // dimension of key embeddings across all k-v heads + const uint32_t n_head_kv = this->n_head_kv(il); + return n_embd_head_k * n_head_kv; } - uint32_t n_embd_v_gqa() const { // dimension of value embeddings across all k-v heads + uint32_t n_embd_v_gqa(uint32_t il = 0) const { // dimension of value embeddings across all k-v heads + const uint32_t n_head_kv = this->n_head_kv(il); + return n_embd_head_v * n_head_kv; } @@ -2233,6 +2325,8 @@ struct llama_hparams { } }; +static_assert(std::is_trivially_copyable::value, "llama_hparams must be trivially copyable"); + struct llama_cparams { uint32_t n_ctx; // context size used during inference uint32_t n_batch; @@ -2264,6 +2358,7 @@ struct llama_cparams { void * cb_eval_user_data; }; +// TODO: separate into "llama_layer_enc" and "llama_layer_dec" struct llama_layer { // normalization struct ggml_tensor * attn_norm; @@ -2281,6 +2376,8 @@ struct llama_layer { struct ggml_tensor * attn_sub_norm; struct ggml_tensor * attn_post_norm; struct ggml_tensor * ffn_sub_norm; + struct ggml_tensor * attn_norm_cross; + struct ggml_tensor * attn_norm_enc; // attention struct ggml_tensor * wq; @@ -2292,6 +2389,14 @@ struct llama_layer { struct ggml_tensor * wq_b; struct ggml_tensor * wkv_a_mqa; struct ggml_tensor * wkv_b; + struct ggml_tensor * wq_cross; + struct ggml_tensor * wk_cross; + struct ggml_tensor * wv_cross; + struct ggml_tensor * wo_cross; + struct ggml_tensor * wq_enc; + struct ggml_tensor * wk_enc; + struct ggml_tensor * wv_enc; + struct ggml_tensor * wo_enc; // attention bias struct ggml_tensor * bq; @@ -2300,6 +2405,11 @@ struct llama_layer { struct ggml_tensor * bo; struct ggml_tensor * bqkv; + // relative position bias + struct ggml_tensor * attn_rel_b; + struct ggml_tensor * attn_rel_b_enc; + struct ggml_tensor * attn_rel_b_cross; + // normalization struct ggml_tensor * ffn_norm; struct ggml_tensor * ffn_norm_b; @@ -2307,11 +2417,15 @@ struct llama_layer { struct ggml_tensor * layer_out_norm; struct ggml_tensor * layer_out_norm_b; struct ggml_tensor * ffn_norm_exps; + struct ggml_tensor * ffn_norm_enc; // ff struct ggml_tensor * ffn_gate; // w1 struct ggml_tensor * ffn_down; // w2 struct ggml_tensor * ffn_up; // w3 + struct ggml_tensor * ffn_gate_enc; + struct ggml_tensor * ffn_down_enc; + struct ggml_tensor * ffn_up_enc; // ff MoE struct ggml_tensor * ffn_gate_inp; @@ -2550,6 +2664,7 @@ struct llama_model { struct ggml_tensor * output_norm_b; struct ggml_tensor * output; struct ggml_tensor * output_b; + struct ggml_tensor * output_norm_enc; std::vector layers; @@ -2678,6 +2793,13 @@ struct llama_context { // populated only when pooling_type != LLAMA_POOLING_TYPE_NONE std::map> embd_seq; + // whether we are computing encoder output or decoder output + bool is_encoding = false; + + // output of the encoder part of the encoder-decoder models + std::vector embd_enc; + std::vector> seq_ids_enc; + // memory buffers used to evaluate the model std::vector buf_compute_meta; ggml_backend_sched_t sched = nullptr; @@ -2698,6 +2820,9 @@ struct llama_context { struct ggml_tensor * inp_s_copy; // I32 [kv_size] struct ggml_tensor * inp_s_mask; // F32 [1, n_kv] struct ggml_tensor * inp_s_seq; // I32 [n_kv, n_batch] + struct ggml_tensor * inp_pos_bucket; // I32 [n_batch|n_kv, n_batch] + struct ggml_tensor * inp_embd_enc; // F32 [n_embd, n_outputs_enc] + struct ggml_tensor * inp_KQ_mask_cross; // F32 [n_outputs_enc, n_batch] // control vectors struct llama_control_vector cvec; @@ -2828,9 +2953,7 @@ static bool llama_kv_cache_init( const struct llama_hparams & hparams = model.hparams; - const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa() + hparams.n_embd_k_s(); - const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa() + hparams.n_embd_v_s(); - const int64_t n_layer = hparams.n_layer; + const int64_t n_layer = hparams.n_layer; cache.has_shift = false; @@ -2838,13 +2961,6 @@ static bool llama_kv_cache_init( cache.recurrent = model.arch == LLM_ARCH_MAMBA; cache.v_trans = !cparams.flash_attn; - // TODO: support mixed recurrent Transformer architectures - // NOTE: (!a || b) is a logical implication (a -> b) - GGML_ASSERT(!cache.recurrent || n_embd_k_gqa == hparams.n_embd_k_s()); - GGML_ASSERT(!cache.recurrent || n_embd_v_gqa == hparams.n_embd_v_s()); - GGML_ASSERT( cache.recurrent || n_embd_k_gqa == hparams.n_embd_k_gqa()); - GGML_ASSERT( cache.recurrent || n_embd_v_gqa == hparams.n_embd_v_gqa()); - cache.head = 0; cache.size = kv_size; cache.used = 0; @@ -2898,6 +3014,9 @@ static bool llama_kv_cache_init( cache.v_l.reserve(n_layer); for (int i = 0; i < (int) n_layer; i++) { + const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i) + hparams.n_embd_k_s(); + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(i) + hparams.n_embd_v_s(); + struct ggml_context * ctx = offload ? ctx_map.at(model.buft_layer[i].buft) : cache.ctxs.front(); ggml_tensor * k = ggml_new_tensor_1d(ctx, type_k, n_embd_k_gqa*kv_size); ggml_tensor * v = ggml_new_tensor_1d(ctx, type_v, n_embd_v_gqa*kv_size); @@ -3774,9 +3893,9 @@ struct llama_model_loader { bool get_arr(const std::string & key, std::vector & result, const bool required = true) { const int kid = gguf_find_key(meta, key.c_str()); - if (kid < 0) { + if (kid < 0 || gguf_get_kv_type(meta, kid) != GGUF_TYPE_ARRAY) { if (required) { - throw std::runtime_error(format("key not found in model: %s", key.c_str())); + throw std::runtime_error(format("array key not found in model: %s", key.c_str())); } return false; } @@ -3784,22 +3903,53 @@ struct llama_model_loader { struct GGUFMeta::ArrayInfo arr_info = GGUFMeta::GKV::get_kv(meta, kid); - if (arr_info.gt != GGUF_TYPE_FLOAT32 && arr_info.gt != GGUF_TYPE_INT32) { - throw std::runtime_error(format("%s is not a float32 or int32 array", key.c_str())); + switch (arr_info.gt) { + case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same::value)); break; + case GGUF_TYPE_INT32: GGML_ASSERT( + (std::is_same::value) || + (std::is_same::value)); break; + default: + throw std::runtime_error(format("%s is not a float32, int32 array", key.c_str())); } - // GGML_ASSERT(gguf_type_size(arr_info.gt) == sizeof(T)); - GGML_ASSERT((arr_info.gt != GGUF_TYPE_FLOAT32 || std::is_same::value)); - GGML_ASSERT((arr_info.gt != GGUF_TYPE_INT32 || std::is_same::value)); - result.resize(arr_info.length); result.assign((const T*)arr_info.data, (const T *)arr_info.data + arr_info.length); return true; } + template + bool get_arr(const std::string & key, std::array & result, const bool required = true) { + const int kid = gguf_find_key(meta, key.c_str()); + + if (kid < 0 || gguf_get_kv_type(meta, kid) != GGUF_TYPE_ARRAY) { + if (required) { + throw std::runtime_error(format("array key not found in model: %s", key.c_str())); + } + return false; + } + + struct GGUFMeta::ArrayInfo arr_info = + GGUFMeta::GKV::get_kv(meta, kid); + + switch (arr_info.gt) { + case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same::value)); break; + case GGUF_TYPE_INT32: GGML_ASSERT( + (std::is_same::value) || + (std::is_same::value)); break; + default: + throw std::runtime_error(format("%s is not a float32, int32 array", key.c_str())); + } + + GGML_ASSERT(arr_info.length <= N_MAX); + + std::copy((const T*)arr_info.data, (const T *)arr_info.data + arr_info.length, result.begin()); + + return true; + } + template - bool get_arr(const enum llm_kv kid, T& result, const bool required = true) { + bool get_arr(const enum llm_kv kid, T & result, const bool required = true) { return get_arr(llm_kv(kid), result, required); } @@ -3824,6 +3974,50 @@ struct llama_model_loader { return get_key(llm_kv(kid), result, required); } + // get array of n <= N_MAX elements, or a single element repeated n times + template + bool get_key_or_arr(const std::string & key, std::array & result, uint32_t n, const bool required = true) { + GGML_ASSERT(n <= N_MAX); + + const int kid = gguf_find_key(meta, key.c_str()); + + if (kid < 0) { + if (required) { + throw std::runtime_error(format("key not found in model: %s", key.c_str())); + } + return false; + } + + if (gguf_get_kv_type(meta, kid) == GGUF_TYPE_ARRAY) { + struct GGUFMeta::ArrayInfo arr_info = + GGUFMeta::GKV::get_kv(meta, kid); + + if (n != arr_info.length) { + throw std::runtime_error(format("key %s has wrong array length; expected %u, got %u", key.c_str(), n, (uint32_t) arr_info.length)); + } + + return get_arr(key, result, required); + } else { + T value; + + bool ok = get_key(key, value, required); + if (!ok) { + return false; + } + + for (uint32_t i = 0; i < n; i++) { + result[i] = value; + } + + return true; + } + } + + template + bool get_key_or_arr(const enum llm_kv kid, T & result, uint32_t n, const bool required = true) { + return get_key_or_arr(llm_kv(kid), result, n, required); + } + std::string get_arch_name() const { return arch_name; } @@ -4056,7 +4250,7 @@ struct llama_model_loader { #if defined(GGML_USE_CUDA) // 4 staging buffers for async uploads, each sized 1MB seems to be a good default for single NVMe drives. // NVMe raid configurations might require more / larger buffers. - constexpr size_t num_buffers = 4; + constexpr size_t n_buffers = 4; constexpr size_t buffer_size = 1 * 1024 * 1024; // 1MB std::vector host_buffers; @@ -4082,7 +4276,7 @@ struct llama_model_loader { // If the cuda backend is active create pinned memory buffers and events for synchronisation. if (cuda_backend) { - for (size_t idx = 0; idx < num_buffers; ++idx) { + for (size_t idx = 0; idx < n_buffers; ++idx) { host_buffers.emplace_back(ggml_backend_buft_alloc_buffer(llama_default_buffer_type_cpu(true), buffer_size)); host_ptrs.emplace_back(ggml_backend_buffer_get_base(host_buffers[idx])); events.emplace_back(ggml_backend_event_new(cuda_backend)); @@ -4163,7 +4357,7 @@ struct llama_model_loader { bytes_read += read_iteration; ++buffer_idx; - buffer_idx %= num_buffers; + buffer_idx %= n_buffers; } } else @@ -4199,7 +4393,7 @@ struct llama_model_loader { #if defined(GGML_USE_CUDA) // free temporary resources used for async cuda uploads if (cuda_backend) { - for (size_t idx = 0; idx < num_buffers;++idx) { + for (size_t idx = 0; idx < n_buffers;++idx) { ggml_backend_event_synchronize(events[idx]); ggml_backend_event_free(events[idx]); ggml_backend_buffer_free(host_buffers[idx]); @@ -4321,14 +4515,23 @@ static const char * llama_model_type_name(e_model type) { case MODEL_17M: return "17M"; case MODEL_22M: return "22M"; case MODEL_33M: return "33M"; + case MODEL_60M: return "60M"; case MODEL_70M: return "70M"; + case MODEL_80M: return "80M"; case MODEL_109M: return "109M"; case MODEL_137M: return "137M"; case MODEL_160M: return "160M"; + case MODEL_220M: return "220M"; + case MODEL_250M: return "250M"; + case MODEL_270M: return "270M"; case MODEL_335M: return "335M"; case MODEL_410M: return "410M"; + case MODEL_450M: return "450M"; + case MODEL_770M: return "770M"; + case MODEL_780M: return "780M"; case MODEL_0_5B: return "0.5B"; case MODEL_1B: return "1B"; + case MODEL_1_3B: return "1.3B"; case MODEL_1_4B: return "1.4B"; case MODEL_2B: return "2B"; case MODEL_2_8B: return "2.8B"; @@ -4337,6 +4540,7 @@ static const char * llama_model_type_name(e_model type) { case MODEL_6_9B: return "6.9B"; case MODEL_7B: return "7B"; case MODEL_8B: return "8B"; + case MODEL_11B: return "11B"; case MODEL_12B: return "12B"; case MODEL_13B: return "13B"; case MODEL_14B: return "14B"; @@ -4406,20 +4610,18 @@ static void llm_load_hparams( ml.get_key(LLM_KV_GENERAL_NAME, model.name, false); // get hparams kv - ml.get_key(LLM_KV_VOCAB_SIZE, hparams.n_vocab, false) || ml.get_arr_n(LLM_KV_TOKENIZER_LIST, hparams.n_vocab); + ml.get_key(LLM_KV_VOCAB_SIZE, hparams.n_vocab, false) || ml.get_arr_n(LLM_KV_TOKENIZER_LIST, hparams.n_vocab); // everything past this point is not vocab-related if (hparams.vocab_only) { return; } - ml.get_key(LLM_KV_CONTEXT_LENGTH, hparams.n_ctx_train); - ml.get_key(LLM_KV_EMBEDDING_LENGTH, hparams.n_embd); - ml.get_key(LLM_KV_FEED_FORWARD_LENGTH, hparams.n_ff); - ml.get_key(LLM_KV_ATTENTION_HEAD_COUNT, hparams.n_head); - ml.get_key(LLM_KV_BLOCK_COUNT, hparams.n_layer); - ml.get_key(LLM_KV_EXPERT_COUNT, hparams.n_expert, false); - ml.get_key(LLM_KV_EXPERT_USED_COUNT, hparams.n_expert_used, false); + ml.get_key(LLM_KV_CONTEXT_LENGTH, hparams.n_ctx_train); + ml.get_key(LLM_KV_EMBEDDING_LENGTH, hparams.n_embd); + ml.get_key(LLM_KV_BLOCK_COUNT, hparams.n_layer); + ml.get_key(LLM_KV_EXPERT_COUNT, hparams.n_expert, false); + ml.get_key(LLM_KV_EXPERT_USED_COUNT, hparams.n_expert_used, false); GGML_ASSERT(hparams.n_expert <= LLAMA_MAX_EXPERTS); GGML_ASSERT(hparams.n_expert_used <= hparams.n_expert); @@ -4429,9 +4631,18 @@ static void llm_load_hparams( GGML_ASSERT(hparams.n_expert_used == 0); } + // zero-out the per-layer hparams + std::fill(hparams.n_head_arr.begin(), hparams.n_head_arr.end(), 0); + std::fill(hparams.n_head_kv_arr.begin(), hparams.n_head_kv_arr.end(), 0); + std::fill(hparams.n_ff_arr.begin(), hparams.n_ff_arr.end(), 0); + + ml.get_key_or_arr(LLM_KV_FEED_FORWARD_LENGTH, hparams.n_ff_arr, hparams.n_layer); + ml.get_key_or_arr(LLM_KV_ATTENTION_HEAD_COUNT, hparams.n_head_arr, hparams.n_layer); + // n_head_kv is optional, default to n_head - hparams.n_head_kv = hparams.n_head; - ml.get_key(LLM_KV_ATTENTION_HEAD_COUNT_KV, hparams.n_head_kv, false); + hparams.n_head_kv_arr = hparams.n_head_arr; + + ml.get_key_or_arr(LLM_KV_ATTENTION_HEAD_COUNT_KV, hparams.n_head_kv_arr, hparams.n_layer, false); bool rope_finetuned = false; ml.get_key(LLM_KV_ROPE_SCALING_FINETUNED, rope_finetuned, false); @@ -4459,27 +4670,32 @@ static void llm_load_hparams( ml.get_key(LLM_KV_ROPE_SCALING_ATTN_FACTOR, hparams.rope_attn_factor, false); - // sanity check for n_rot (optional) - { - hparams.n_rot = (hparams.n_head == 0) ? 0 : hparams.n_embd / hparams.n_head; + // non-transformer models do not have attention heads + if (hparams.n_head() > 0) { + // sanity check for n_rot (optional) + hparams.n_rot = hparams.n_embd / hparams.n_head(); ml.get_key(LLM_KV_ROPE_DIMENSION_COUNT, hparams.n_rot, false); if (model.arch == LLM_ARCH_LLAMA || model.arch == LLM_ARCH_FALCON) { - if (hparams.n_rot != hparams.n_embd / hparams.n_head) { - throw std::runtime_error(format("invalid n_rot: %u, expected %u", hparams.n_rot, hparams.n_embd / hparams.n_head)); + if (hparams.n_rot != hparams.n_embd / hparams.n_head()) { + throw std::runtime_error(format("invalid n_rot: %u, expected %u", hparams.n_rot, hparams.n_embd / hparams.n_head())); } } // gpt-neox n_rot = rotary_pct * (n_embd / n_head) // gpt-j n_rot = rotary_dim + + hparams.n_embd_head_k = hparams.n_embd / hparams.n_head(); + ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH, hparams.n_embd_head_k, false); + + hparams.n_embd_head_v = hparams.n_embd / hparams.n_head(); + ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH, hparams.n_embd_head_v, false); + } else { + hparams.n_rot = 0; + hparams.n_embd_head_k = 0; + hparams.n_embd_head_v = 0; } - hparams.n_embd_head_k = (hparams.n_head == 0) ? 0 : hparams.n_embd / hparams.n_head; - ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH, hparams.n_embd_head_k, false); - - hparams.n_embd_head_v = (hparams.n_head == 0) ? 0 : hparams.n_embd / hparams.n_head; - ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH, hparams.n_embd_head_v, false); - // arch-specific KVs switch (model.arch) { case LLM_ARCH_LLAMA: @@ -4502,7 +4718,7 @@ static void llm_load_hparams( case 40: model.type = e_model::MODEL_13B; break; case 48: model.type = e_model::MODEL_34B; break; case 60: model.type = e_model::MODEL_30B; break; - case 80: model.type = hparams.n_head == hparams.n_head_kv ? e_model::MODEL_65B : e_model::MODEL_70B; break; + case 80: model.type = hparams.n_head() == hparams.n_head_kv() ? e_model::MODEL_65B : e_model::MODEL_70B; break; default: model.type = e_model::MODEL_UNKNOWN; } } @@ -4671,7 +4887,7 @@ static void llm_load_hparams( switch (hparams.n_layer) { case 24: model.type = hparams.n_embd == 1024 ? e_model::MODEL_0_5B : e_model::MODEL_1B; break; case 32: model.type = e_model::MODEL_7B; break; - case 40: model.type = hparams.n_head == 20 ? e_model::MODEL_4B : e_model::MODEL_13B; break; + case 40: model.type = hparams.n_head() == 20 ? e_model::MODEL_4B : e_model::MODEL_13B; break; case 80: model.type = e_model::MODEL_70B; break; default: model.type = e_model::MODEL_UNKNOWN; } @@ -4863,46 +5079,58 @@ static void llm_load_hparams( default: model.type = e_model::MODEL_UNKNOWN; } } break; + case LLM_ARCH_OPENELM: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + + switch (hparams.n_layer) { + case 16: model.type = e_model::MODEL_270M; break; + case 20: model.type = e_model::MODEL_450M; break; + case 28: model.type = e_model::MODEL_1B; break; + case 36: model.type = e_model::MODEL_3B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; case LLM_ARCH_GPTNEOX: { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); ml.get_key(LLM_KV_USE_PARALLEL_RESIDUAL, hparams.use_par_res); switch (hparams.n_layer) { case 6: - switch (hparams.n_ff) { + switch (hparams.n_ff()) { case 512: model.type = e_model::MODEL_14M; break; case 2048: model.type = e_model::MODEL_70M; break; default: model.type = e_model::MODEL_UNKNOWN; } break; case 12: - switch (hparams.n_ff) { + switch (hparams.n_ff()) { case 3072: model.type = e_model::MODEL_160M; break; default: model.type = e_model::MODEL_UNKNOWN; } break; case 16: - switch (hparams.n_ff) { + switch (hparams.n_ff()) { case 8192: model.type = e_model::MODEL_1B; break; default: model.type = e_model::MODEL_UNKNOWN; } break; case 24: - switch (hparams.n_ff) { + switch (hparams.n_ff()) { case 4096: model.type = e_model::MODEL_410M; break; case 8192: model.type = e_model::MODEL_1_4B; break; default: model.type = e_model::MODEL_UNKNOWN; } break; case 32: - switch (hparams.n_ff) { + switch (hparams.n_ff()) { case 10240: model.type = e_model::MODEL_2_8B; break; case 16384: model.type = e_model::MODEL_6_9B; break; default: model.type = e_model::MODEL_UNKNOWN; } break; case 36: - switch (hparams.n_ff) { + switch (hparams.n_ff()) { case 20480: model.type = e_model::MODEL_12B; break; default: model.type = e_model::MODEL_UNKNOWN; } break; case 44: - switch (hparams.n_ff) { + switch (hparams.n_ff()) { case 24576: model.type = e_model::MODEL_20B; break; default: model.type = e_model::MODEL_UNKNOWN; } break; @@ -4951,6 +5179,50 @@ static void llm_load_hparams( default: model.type = e_model::MODEL_UNKNOWN; } } break; + case LLM_ARCH_T5: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + ml.get_key(LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, hparams.n_rel_attn_bkts); + + uint32_t dec_start_token_id; + if (ml.get_key(LLM_KV_DECODER_START_TOKEN_ID, dec_start_token_id, false)) { + hparams.dec_start_token_id = dec_start_token_id; + } + + switch (hparams.n_layer) { + case 6: model.type = e_model::MODEL_60M; break; // t5-small + case 8: model.type = e_model::MODEL_80M; break; // flan-t5-small + case 12: + switch (hparams.n_ff()) { + case 3072: model.type = e_model::MODEL_220M; break; // t5-base + case 2048: model.type = e_model::MODEL_250M; break; // flan-t5-base + default: model.type = e_model::MODEL_UNKNOWN; + } break; + case 24: + switch (hparams.n_ff()) { + case 4096: model.type = e_model::MODEL_770M; break; // t5-large + case 2816: model.type = e_model::MODEL_780M; break; // flan-t5-large + case 16384: model.type = e_model::MODEL_3B; break; // t5-3b + case 5120: model.type = e_model::MODEL_3B; break; // flan-t5-xl + case 65536: model.type = e_model::MODEL_11B; break; // t5-11b + case 10240: model.type = e_model::MODEL_11B; break; // flan-t5-xxl + default: model.type = e_model::MODEL_UNKNOWN; + } break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_JAIS: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + ml.get_key(LLM_KV_ATTENTION_MAX_ALIBI_BIAS, hparams.f_max_alibi_bias); + + switch (hparams.n_layer) { + case 24: model.type = e_model::MODEL_1_3B; break; + case 40: model.type = e_model::MODEL_13B; break; + /* TODO: add variants */ + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; default: (void)0; } @@ -5192,6 +5464,9 @@ static void llm_load_vocab( } else if ( tokenizer_pre == "viking") { vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_VIKING; + } else if ( + tokenizer_pre == "jais") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_JAIS; } else { throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str())); } @@ -5486,44 +5761,78 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) { const char * rope_scaling_type = LLAMA_ROPE_SCALING_TYPES.at(hparams.rope_scaling_type_train); + auto print_f = [](const std::function & f, uint32_t n) { + bool is_var = false; + + std::vector v; + for (uint32_t i = 0; i < n; ++i) { + v.push_back(f(i)); + if (v[i] != v[0]) { + is_var = true; + } + } + + std::stringstream ss; + + if (is_var) { + ss << "["; + for (uint32_t i = 0; i < n; ++i) { + ss << v[i]; + if (i < n - 1) { + ss << ", "; + } + } + ss << "]"; + } else { + ss << v[0]; + } + + return ss.str(); + }; + // hparams LLAMA_LOG_INFO("%s: format = %s\n", __func__, llama_file_version_name(ml.fver)); LLAMA_LOG_INFO("%s: arch = %s\n", __func__, LLM_ARCH_NAMES.at(model.arch)); LLAMA_LOG_INFO("%s: vocab type = %s\n", __func__, llama_model_vocab_type_name(vocab.type)); LLAMA_LOG_INFO("%s: n_vocab = %u\n", __func__, hparams.n_vocab); LLAMA_LOG_INFO("%s: n_merges = %u\n", __func__, (int) vocab.bpe_ranks.size()); - LLAMA_LOG_INFO("%s: n_ctx_train = %u\n", __func__, hparams.n_ctx_train); - LLAMA_LOG_INFO("%s: n_embd = %u\n", __func__, hparams.n_embd); - LLAMA_LOG_INFO("%s: n_head = %u\n", __func__, hparams.n_head); - LLAMA_LOG_INFO("%s: n_head_kv = %u\n", __func__, hparams.n_head_kv); - LLAMA_LOG_INFO("%s: n_layer = %u\n", __func__, hparams.n_layer); - LLAMA_LOG_INFO("%s: n_rot = %u\n", __func__, hparams.n_rot); - LLAMA_LOG_INFO("%s: n_swa = %u\n", __func__, hparams.n_swa); - LLAMA_LOG_INFO("%s: n_embd_head_k = %u\n", __func__, hparams.n_embd_head_k); - LLAMA_LOG_INFO("%s: n_embd_head_v = %u\n", __func__, hparams.n_embd_head_v); - LLAMA_LOG_INFO("%s: n_gqa = %u\n", __func__, hparams.n_gqa()); - LLAMA_LOG_INFO("%s: n_embd_k_gqa = %u\n", __func__, hparams.n_embd_k_gqa()); - LLAMA_LOG_INFO("%s: n_embd_v_gqa = %u\n", __func__, hparams.n_embd_v_gqa()); - LLAMA_LOG_INFO("%s: f_norm_eps = %.1e\n", __func__, hparams.f_norm_eps); - LLAMA_LOG_INFO("%s: f_norm_rms_eps = %.1e\n", __func__, hparams.f_norm_rms_eps); - LLAMA_LOG_INFO("%s: f_clamp_kqv = %.1e\n", __func__, hparams.f_clamp_kqv); - LLAMA_LOG_INFO("%s: f_max_alibi_bias = %.1e\n", __func__, hparams.f_max_alibi_bias); - LLAMA_LOG_INFO("%s: f_logit_scale = %.1e\n", __func__, hparams.f_logit_scale); - LLAMA_LOG_INFO("%s: n_ff = %u\n", __func__, hparams.n_ff); - LLAMA_LOG_INFO("%s: n_expert = %u\n", __func__, hparams.n_expert); - LLAMA_LOG_INFO("%s: n_expert_used = %u\n", __func__, hparams.n_expert_used); - LLAMA_LOG_INFO("%s: causal attn = %d\n", __func__, hparams.causal_attn); - LLAMA_LOG_INFO("%s: pooling type = %d\n", __func__, hparams.pooling_type); - LLAMA_LOG_INFO("%s: rope type = %d\n", __func__, hparams.rope_type); - LLAMA_LOG_INFO("%s: rope scaling = %s\n", __func__, rope_scaling_type); - LLAMA_LOG_INFO("%s: freq_base_train = %.1f\n", __func__, hparams.rope_freq_base_train); - LLAMA_LOG_INFO("%s: freq_scale_train = %g\n", __func__, hparams.rope_freq_scale_train); - LLAMA_LOG_INFO("%s: n_ctx_orig_yarn = %u\n", __func__, hparams.n_ctx_orig_yarn); - LLAMA_LOG_INFO("%s: rope_finetuned = %s\n", __func__, hparams.rope_finetuned ? "yes" : "unknown"); - LLAMA_LOG_INFO("%s: ssm_d_conv = %u\n", __func__, hparams.ssm_d_conv); - LLAMA_LOG_INFO("%s: ssm_d_inner = %u\n", __func__, hparams.ssm_d_inner); - LLAMA_LOG_INFO("%s: ssm_d_state = %u\n", __func__, hparams.ssm_d_state); - LLAMA_LOG_INFO("%s: ssm_dt_rank = %u\n", __func__, hparams.ssm_dt_rank); + LLAMA_LOG_INFO("%s: vocab_only = %d\n", __func__, hparams.vocab_only); + + if (!hparams.vocab_only) { + LLAMA_LOG_INFO("%s: n_ctx_train = %u\n", __func__, hparams.n_ctx_train); + LLAMA_LOG_INFO("%s: n_embd = %u\n", __func__, hparams.n_embd); + LLAMA_LOG_INFO("%s: n_layer = %u\n", __func__, hparams.n_layer); + LLAMA_LOG_INFO("%s: n_head = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_head(il); }, hparams.n_layer).c_str()); + LLAMA_LOG_INFO("%s: n_head_kv = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_head_kv(il); }, hparams.n_layer).c_str()); + LLAMA_LOG_INFO("%s: n_rot = %u\n", __func__, hparams.n_rot); + LLAMA_LOG_INFO("%s: n_swa = %u\n", __func__, hparams.n_swa); + LLAMA_LOG_INFO("%s: n_embd_head_k = %u\n", __func__, hparams.n_embd_head_k); + LLAMA_LOG_INFO("%s: n_embd_head_v = %u\n", __func__, hparams.n_embd_head_v); + LLAMA_LOG_INFO("%s: n_gqa = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_gqa(il); }, hparams.n_layer).c_str()); + LLAMA_LOG_INFO("%s: n_embd_k_gqa = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_embd_k_gqa(il); }, hparams.n_layer).c_str()); + LLAMA_LOG_INFO("%s: n_embd_v_gqa = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_embd_v_gqa(il); }, hparams.n_layer).c_str()); + LLAMA_LOG_INFO("%s: f_norm_eps = %.1e\n", __func__, hparams.f_norm_eps); + LLAMA_LOG_INFO("%s: f_norm_rms_eps = %.1e\n", __func__, hparams.f_norm_rms_eps); + LLAMA_LOG_INFO("%s: f_clamp_kqv = %.1e\n", __func__, hparams.f_clamp_kqv); + LLAMA_LOG_INFO("%s: f_max_alibi_bias = %.1e\n", __func__, hparams.f_max_alibi_bias); + LLAMA_LOG_INFO("%s: f_logit_scale = %.1e\n", __func__, hparams.f_logit_scale); + LLAMA_LOG_INFO("%s: n_ff = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_ff(il); }, hparams.n_layer).c_str()); + LLAMA_LOG_INFO("%s: n_expert = %u\n", __func__, hparams.n_expert); + LLAMA_LOG_INFO("%s: n_expert_used = %u\n", __func__, hparams.n_expert_used); + LLAMA_LOG_INFO("%s: causal attn = %d\n", __func__, hparams.causal_attn); + LLAMA_LOG_INFO("%s: pooling type = %d\n", __func__, hparams.pooling_type); + LLAMA_LOG_INFO("%s: rope type = %d\n", __func__, hparams.rope_type); + LLAMA_LOG_INFO("%s: rope scaling = %s\n", __func__, rope_scaling_type); + LLAMA_LOG_INFO("%s: freq_base_train = %.1f\n", __func__, hparams.rope_freq_base_train); + LLAMA_LOG_INFO("%s: freq_scale_train = %g\n", __func__, hparams.rope_freq_scale_train); + LLAMA_LOG_INFO("%s: n_ctx_orig_yarn = %u\n", __func__, hparams.n_ctx_orig_yarn); + LLAMA_LOG_INFO("%s: rope_finetuned = %s\n", __func__, hparams.rope_finetuned ? "yes" : "unknown"); + LLAMA_LOG_INFO("%s: ssm_d_conv = %u\n", __func__, hparams.ssm_d_conv); + LLAMA_LOG_INFO("%s: ssm_d_inner = %u\n", __func__, hparams.ssm_d_inner); + LLAMA_LOG_INFO("%s: ssm_d_state = %u\n", __func__, hparams.ssm_d_state); + LLAMA_LOG_INFO("%s: ssm_dt_rank = %u\n", __func__, hparams.ssm_dt_rank); + } + LLAMA_LOG_INFO("%s: model type = %s\n", __func__, llama_model_type_name(model.type)); LLAMA_LOG_INFO("%s: model ftype = %s\n", __func__, llama_model_ftype_name(model.ftype).c_str()); if (ml.n_elements >= 1e12) { @@ -5603,8 +5912,8 @@ static bool llm_load_tensors( model.main_gpu = main_gpu; model.n_gpu_layers = n_gpu_layers; - const int64_t n_layer = hparams.n_layer; - int64_t i_gpu_start = std::max((int64_t) hparams.n_layer - n_gpu_layers, (int64_t) 0); + const int n_layer = hparams.n_layer; + int i_gpu_start = std::max((int) hparams.n_layer - n_gpu_layers, (int) 0); bool use_mmap_buffer = true; #if defined(GGML_USE_CLBLAST) @@ -5623,7 +5932,7 @@ static bool llm_load_tensors( model.buft_layer.resize(n_layer); // assign cpu layers - for (int64_t i = 0; i < i_gpu_start; ++i) { + for (int i = 0; i < i_gpu_start; ++i) { model.buft_layer[i] = llama_default_buffer_type_cpu(true); } @@ -5653,7 +5962,7 @@ static bool llm_load_tensors( // assign the repeating layers to the devices according to the splits int act_gpu_layers = std::min(n_gpu_layers, (int)n_layer + 1); - for (int64_t i = i_gpu_start; i < n_layer; ++i) { + for (int i = i_gpu_start; i < n_layer; ++i) { int layer_gpu = std::upper_bound(splits.begin(), splits.begin() + device_count, float(i - i_gpu_start)/act_gpu_layers) - splits.begin(); model.buft_layer[i] = llama_default_buffer_type_offload(model, layer_gpu); } @@ -5673,7 +5982,7 @@ static bool llm_load_tensors( split_buft = llama_default_buffer_type_offload(model, main_gpu); } // assign the repeating layers - for (int64_t i = i_gpu_start; i < n_layer; ++i) { + for (int i = i_gpu_start; i < n_layer; ++i) { model.buft_layer[i] = { split_buft, llama_default_buffer_type_offload(model, main_gpu) @@ -5696,7 +6005,7 @@ static bool llm_load_tensors( buft_layer_count[model.buft_input.buft_matrix]++; buft_layer_count[model.buft_output.buft]++; buft_layer_count[model.buft_output.buft_matrix]++; - for (int64_t i = 0; i < n_layer; ++i) { + for (int i = 0; i < n_layer; ++i) { buft_layer_count[model.buft_layer[i].buft]++; buft_layer_count[model.buft_layer[i].buft_matrix]++; } @@ -5726,15 +6035,21 @@ static bool llm_load_tensors( // create tensors for the weights { - const int64_t n_embd = hparams.n_embd; - const int64_t n_embd_head = (hparams.n_head == 0) ? 0 : n_embd / hparams.n_head; - const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(); - const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(); - const int64_t n_embd_gqa = n_embd_v_gqa; - const int64_t n_vocab = hparams.n_vocab; - const int64_t n_vocab_type = hparams.n_vocab_type; - const int64_t n_ff = hparams.n_ff; - const int64_t n_expert = hparams.n_expert; + // note: cast to int64_t since we will use these for the tensor dimensions + const int64_t n_head = hparams.n_head(); + const int64_t n_head_kv = hparams.n_head_kv(); + const int64_t n_embd = hparams.n_embd; + const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(); + const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(); + const int64_t n_embd_head_k = hparams.n_embd_head_k; + const int64_t n_embd_head_v = hparams.n_embd_head_v; + const int64_t n_ff = hparams.n_ff(); + const int64_t n_embd_gqa = n_embd_v_gqa; + const int64_t n_vocab = hparams.n_vocab; + const int64_t n_vocab_type = hparams.n_vocab_type; + const int64_t n_expert = hparams.n_expert; + const int64_t n_expert_used = hparams.n_expert_used; + const int64_t n_ctx_train = hparams.n_ctx_train; if (n_expert > 0 && hparams.n_expert_used == 0) { throw std::runtime_error("model has expert layers but no expert layers are used"); @@ -5743,8 +6058,9 @@ static bool llm_load_tensors( ggml_context * ctx_input = ctx_map.at(model.buft_input.buft); ggml_context * ctx_output = ctx_map.at(model.buft_output.buft); ggml_context * ctx_output_split = ctx_map.at(model.buft_output.buft_matrix); - auto ctx_for_layer = [&](int i) { return ctx_map.at(model.buft_layer[i].buft); }; - auto ctx_for_layer_split = [&](int i) { return ctx_map.at(model.buft_layer[i].buft_matrix); }; + + auto ctx_for_layer = [&](int i) { return ctx_map.at(model.buft_layer[i].buft); }; + auto ctx_for_layer_split = [&](int i) { return ctx_map.at(model.buft_layer[i].buft_matrix); }; model.layers.resize(n_layer); @@ -5759,7 +6075,8 @@ static bool llm_load_tensors( // output { model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); - model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED); + model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED); + // if output is NULL, init from the input tok embed if (model.output == NULL) { model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); @@ -5842,6 +6159,7 @@ static bool llm_load_tensors( { model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED); + // if output is NULL, init from the input tok embed if (model.output == NULL) { model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); @@ -5865,9 +6183,9 @@ static bool llm_load_tensors( layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); - layer.ffn_gate_inp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}); - + layer.ffn_gate_inp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}); layer.ffn_gate_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff, n_expert}, llama_model_loader::TENSOR_NOT_REQUIRED); + if (layer.ffn_gate_exps) { layer.ffn_down_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff, n_embd, n_expert}); layer.ffn_up_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {n_embd, n_ff, n_expert}); @@ -5919,12 +6237,12 @@ static bool llm_load_tensors( auto & layer = model.layers[i]; - layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); + layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}); layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); - layer.attn_out_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}); + layer.attn_out_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}); layer.ffn_gate_inp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}); layer.ffn_gate_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff, n_expert}); @@ -5966,10 +6284,10 @@ static bool llm_load_tensors( // output { - model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); - model.output_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}); + model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); + model.output_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}); - model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED); + model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED); if (!model.output) { model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); // needs to be on GPU } @@ -5997,7 +6315,7 @@ static bool llm_load_tensors( case LLM_ARCH_STARCODER: { model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); - model.pos_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_POS_EMBD, "weight"), {n_embd, hparams.n_ctx_train}); + model.pos_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_POS_EMBD, "weight"), {n_embd, n_ctx_train}); // output { @@ -6023,8 +6341,8 @@ static bool llm_load_tensors( layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}); layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}); - layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); - layer.bo = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}); + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); + layer.bo = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}); layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}); @@ -6032,8 +6350,8 @@ static bool llm_load_tensors( layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}); layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}); - layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); - layer.ffn_up_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}); + layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); + layer.ffn_up_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}); } } break; case LLM_ARCH_BERT: @@ -6041,8 +6359,9 @@ static bool llm_load_tensors( { model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); model.type_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_TYPES, "weight"), {n_embd, n_vocab_type}); + if (model.arch == LLM_ARCH_BERT) { - model.pos_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_POS_EMBD, "weight"), {n_embd, hparams.n_ctx_train}); + model.pos_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_POS_EMBD, "weight"), {n_embd, n_ctx_train}); } model.tok_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}); @@ -6055,31 +6374,30 @@ static bool llm_load_tensors( auto & layer = model.layers[i]; if (model.arch == LLM_ARCH_BERT) { - layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}); - layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}); + layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}); + layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}); - layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}); - layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa}); + layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}); + layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa}); - layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}); - layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}); + layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}); + layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}); } else { layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}); } - layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); layer.attn_out_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}); layer.attn_out_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT_NORM, "bias", i), {n_embd}); - layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); - layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}); + layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); + layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}); if (model.arch == LLM_ARCH_BERT) { - layer.bo = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}); - layer.ffn_up_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}); - - layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}); + layer.bo = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}); + layer.ffn_up_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}); + layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}); } else { layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}); } @@ -6090,8 +6408,9 @@ static bool llm_load_tensors( } break; case LLM_ARCH_JINA_BERT_V2: { - model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); // word_embeddings - model.type_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_TYPES, "weight"), {n_embd, n_vocab_type}); //token_type_embeddings + model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); // word_embeddings + model.type_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_TYPES, "weight"), {n_embd, n_vocab_type}); // token_type_embeddings + model.tok_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}); // LayerNorm model.tok_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"), {n_embd}); //LayerNorm bias @@ -6101,38 +6420,38 @@ static bool llm_load_tensors( auto & layer = model.layers[i]; // JinaBertLayer - layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}); - layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}); + layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}); + layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}); layer.attn_q_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.attn_q_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.attn_q_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}); - layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa}); + layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}); + layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa}); layer.attn_k_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.attn_k_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.attn_k_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}); - layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}); + layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}); + layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}); - layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); //output_dens - layer.bo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}); //output_dens + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); //output_dens + layer.bo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}); //output_dens layer.attn_out_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}); //output_norm - layer.attn_out_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT_NORM, "bias", i), {n_embd}); + layer.attn_out_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT_NORM, "bias", i), {n_embd}); layer.attn_norm_2 = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM_2, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.attn_norm_2_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM_2, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.attn_norm_2_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM_2, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); - layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}); + layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); + layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}); - layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}); - layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}); + layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}); + layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}); - layer.layer_out_norm = ml.create_tensor(ctx_split, tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd}); - layer.layer_out_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_LAYER_OUT_NORM, "bias", i), {n_embd}); + layer.layer_out_norm = ml.create_tensor(ctx_split, tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd}); + layer.layer_out_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_LAYER_OUT_NORM, "bias", i), {n_embd}); } } break; case LLM_ARCH_BLOOM: @@ -6155,35 +6474,35 @@ static bool llm_load_tensors( auto & layer = model.layers[i]; layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); - layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}); + layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}); layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}); - layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}); + layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}); layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); - layer.bo = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}); + layer.bo = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}); layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); - layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}); + layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}); layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}); - layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}); + layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}); - layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); - layer.ffn_up_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}); + layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); + layer.ffn_up_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}); } } break; case LLM_ARCH_MPT: { model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); - model.pos_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_POS_EMBD, "weight"), {n_embd, hparams.n_ctx_train}, llama_model_loader::TENSOR_NOT_REQUIRED); + model.pos_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_POS_EMBD, "weight"), {n_embd, n_ctx_train}, llama_model_loader::TENSOR_NOT_REQUIRED); // output { - model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); - model.output_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED); + model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); + model.output_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED); - model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED); + model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED); if (!model.output) { model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); // needs to be on GPU } @@ -6254,8 +6573,8 @@ static bool llm_load_tensors( layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED); // optional q and k layernorms, present in StableLM 2 12B - layer.attn_q_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {hparams.n_embd_head_k, hparams.n_head}, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.attn_k_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {hparams.n_embd_head_k, hparams.n_head_kv}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.attn_q_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k, n_head}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.attn_k_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k, n_head_kv}, llama_model_loader::TENSOR_NOT_REQUIRED); // optional FFN norm, not present in StableLM 2 12B which uses parallel residual layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED); @@ -6366,21 +6685,23 @@ static bool llm_load_tensors( layer.ffn_gate_inp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}); - GGML_ASSERT(hparams.n_expert > 0); - GGML_ASSERT(hparams.n_expert_used > 0); + GGML_ASSERT(n_expert > 0); + GGML_ASSERT(n_expert_used > 0); // MoE branch - auto n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / hparams.n_expert_used; + const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used; + layer.ffn_gate_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert}); layer.ffn_down_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd, n_expert}); layer.ffn_up_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert}); // Shared expert branch - auto n_ff_shexp = hparams.n_ff_shexp ? hparams.n_ff_shexp : n_ff; + const int64_t n_ff_shexp = hparams.n_ff_shexp ? hparams.n_ff_shexp : n_ff; + layer.ffn_gate_inp_shexp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP_SHEXP, "weight", i), {n_embd}); - layer.ffn_gate_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_shexp}); - layer.ffn_down_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd}); - layer.ffn_up_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, n_ff_shexp}); + layer.ffn_gate_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), { n_embd, n_ff_shexp}); + layer.ffn_down_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd}); + layer.ffn_up_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), { n_embd, n_ff_shexp}); } } break; case LLM_ARCH_PHI2: @@ -6430,6 +6751,8 @@ static bool llm_load_tensors( } break; case LLM_ARCH_PHI3: { + const int64_t n_embd_head = n_embd / n_head; + model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }); // output @@ -6439,8 +6762,8 @@ static bool llm_load_tensors( } for (int i = 0; i < n_layer; ++i) { - ggml_context* ctx_layer = ctx_for_layer(i); - ggml_context* ctx_split = ctx_for_layer_split(i); + ggml_context * ctx_layer = ctx_for_layer(i); + ggml_context * ctx_split = ctx_for_layer_split(i); auto & layer = model.layers[i]; @@ -6489,7 +6812,7 @@ static bool llm_load_tensors( case LLM_ARCH_GPT2: { model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); - model.pos_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_POS_EMBD, "weight"), {n_embd, hparams.n_ctx_train}); + model.pos_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_POS_EMBD, "weight"), {n_embd, n_ctx_train}); // output { @@ -6626,12 +6949,7 @@ static bool llm_load_tensors( model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); // same as tok_embd, duplicated to allow offloading - const int64_t n_ff = hparams.n_ff; - const int64_t n_embd_head_k = hparams.n_embd_head_k; - const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(); - const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(); - - for (uint32_t i = 0; i < n_layer; ++i) { + for (int i = 0; i < n_layer; ++i) { ggml_context * ctx_layer = ctx_for_layer(i); ggml_context * ctx_split = ctx_for_layer_split(i); @@ -6639,10 +6957,10 @@ static bool llm_load_tensors( layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); - layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * hparams.n_head}); + layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head}); layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}); layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}); - layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * hparams.n_head, n_embd}); + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}); layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}); @@ -6658,12 +6976,7 @@ static bool llm_load_tensors( model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); // same as tok_embd, duplicated to allow offloading - const int64_t n_ff = hparams.n_ff; - const int64_t n_embd_head_k = hparams.n_embd_head_k; - const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(); - const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(); - - for (uint32_t i = 0; i < n_layer; ++i) { + for (int i = 0; i < n_layer; ++i) { ggml_context * ctx_layer = ctx_for_layer(i); ggml_context * ctx_split = ctx_for_layer_split(i); @@ -6671,10 +6984,10 @@ static bool llm_load_tensors( layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); - layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * hparams.n_head}); + layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head}); layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}); layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}); - layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * hparams.n_head, n_embd}); + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}); layer.attn_post_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}); layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); @@ -6738,6 +7051,7 @@ static bool llm_load_tensors( const int64_t d_inner = hparams.ssm_d_inner; const int64_t d_state = hparams.ssm_d_state; const int64_t dt_rank = hparams.ssm_dt_rank; + // only an expansion factor of 2 is supported for now GGML_ASSERT(2 * n_embd == d_inner); @@ -6788,15 +7102,20 @@ static bool llm_load_tensors( model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}); } + for (int i = 0; i < n_layer; ++i) { ggml_context * ctx_layer = ctx_for_layer(i); ggml_context * ctx_split = ctx_for_layer_split(i); + auto & layer = model.layers[i]; + layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); + layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}); layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}); layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}); layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); + layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}); layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}); @@ -6823,8 +7142,8 @@ static bool llm_load_tensors( layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); if (n_layer >= 64){ - layer.attn_q_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {hparams.n_embd_head_k, hparams.n_head}); - layer.attn_k_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {hparams.n_embd_head_k, hparams.n_head_kv}); + layer.attn_q_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k, n_head}); + layer.attn_k_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k, n_head_kv}); } layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}); @@ -6860,15 +7179,49 @@ static bool llm_load_tensors( layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}); layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); - layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}); layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}); layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); } } break; + case LLM_ARCH_OPENELM: + { + model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); + + // output + { + model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); + // init output from the input tok embed + model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); + } + + for (int i = 0; i < n_layer; ++i) { + const int64_t n_head = hparams.n_head(i); + const int64_t n_head_qkv = 2*hparams.n_head_kv(i) + n_head; + const int64_t n_ff = hparams.n_ff(i); + + ggml_context * ctx_layer = ctx_for_layer(i); + ggml_context * ctx_split = ctx_for_layer_split(i); + + auto & layer = model.layers[i]; + + layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); + + layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_head_qkv*n_embd_head_k}); + layer.attn_q_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}); + layer.attn_k_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}); + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head*n_embd_head_k, n_embd}); + + layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); + layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}); + layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}); + layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); + } + } break; case LLM_ARCH_GPTNEOX: { - model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); + model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); + // output { model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); @@ -6907,8 +7260,9 @@ static bool llm_load_tensors( // output { - model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); - model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED); + model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); + model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED); + // if output is NULL, init from the input tok embed if (model.output == NULL) { model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); @@ -6943,13 +7297,16 @@ static bool llm_load_tensors( } break; case LLM_ARCH_DEEPSEEK2: { - bool is_lite = (hparams.n_layer == 27); + const bool is_lite = (hparams.n_layer == 27); - const uint32_t n_embd_head_qk_rope = hparams.n_rot; - const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot; - const uint32_t q_lora_rank = hparams.n_lora_q; - const uint32_t kv_lora_rank = hparams.n_lora_kv; - const uint32_t n_ff_exp = hparams.n_ff_exp; + const int64_t n_embd_head_qk_rope = hparams.n_rot; + const int64_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot; + + const int64_t q_lora_rank = hparams.n_lora_q; + const int64_t kv_lora_rank = hparams.n_lora_kv; + + const int64_t n_ff_exp = hparams.n_ff_exp; + const int64_t n_expert_shared = hparams.n_expert_shared; model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); @@ -6969,29 +7326,31 @@ static bool llm_load_tensors( if (!is_lite) { layer.attn_q_a_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_A_NORM, "weight", i), {q_lora_rank}); } + layer.attn_kv_a_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_KV_A_NORM, "weight", i), {kv_lora_rank}); if (!is_lite) { - layer.wq_a = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q_A, "weight", i), {n_embd, q_lora_rank}); - layer.wq_b = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, hparams.n_head * hparams.n_embd_head_k}); + layer.wq_a = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q_A, "weight", i), {n_embd, q_lora_rank}); + layer.wq_b = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head * n_embd_head_k}); } else { - layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_k_gqa}); + layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_k_gqa}); } - layer.wkv_a_mqa = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + n_embd_head_qk_rope}); - layer.wkv_b = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_KV_B, "weight", i), {kv_lora_rank, hparams.n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)}); - layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {hparams.n_head * hparams.n_embd_head_v, n_embd}); + + layer.wkv_a_mqa = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + (n_embd_head_qk_rope)}); + layer.wkv_b = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_KV_B, "weight", i), {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)}); + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_head * ( n_embd_head_v), n_embd}); layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); - if ((uint32_t) i < hparams.n_layer_dense_lead) { + if (i < (int) hparams.n_layer_dense_lead) { layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}); layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}); layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); } else { layer.ffn_gate_inp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}); - GGML_ASSERT(hparams.n_expert > 0); - GGML_ASSERT(hparams.n_expert_used > 0); + GGML_ASSERT(n_expert > 0); + GGML_ASSERT(n_expert_used > 0); // MoE branch layer.ffn_gate_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert}); @@ -6999,9 +7358,9 @@ static bool llm_load_tensors( layer.ffn_up_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert}); // Shared expert branch - layer.ffn_gate_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * hparams.n_expert_shared}); - layer.ffn_down_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { n_ff_exp * hparams.n_expert_shared, n_embd}); - layer.ffn_up_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, n_ff_exp * hparams.n_expert_shared}); + layer.ffn_gate_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}); + layer.ffn_down_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { n_ff_exp * n_expert_shared, n_embd}); + layer.ffn_up_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}); } } } break; @@ -7020,27 +7379,126 @@ static bool llm_load_tensors( auto & layer = model.layers[i]; - layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); + layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); layer.attn_sub_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_SUB_NORM, "weight", i), {n_embd}); - layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}); - layer.wq_scale = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q, "scale", i), {1}); - layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}); - layer.wk_scale = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K, "scale", i), {1}); - layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}); - layer.wv_scale = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V, "scale", i), {1}); - layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); - layer.wo_scale = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "scale", i), {1}); + layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}); + layer.wq_scale = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q, "scale", i), {1}); + layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}); + layer.wk_scale = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K, "scale", i), {1}); + layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}); + layer.wv_scale = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V, "scale", i), {1}); + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); + layer.wo_scale = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "scale", i), {1}); - layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); + layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); layer.ffn_sub_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_SUB_NORM, "weight", i), {n_ff}); - layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}); - layer.ffn_gate_scale = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE, "scale", i), {1}); - layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}); - layer.ffn_down_scale = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "scale", i), {1}); - layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); - layer.ffn_up_scale = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "scale", i), {1}); + layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}); + layer.ffn_gate_scale = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE, "scale", i), {1}); + layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}); + layer.ffn_down_scale = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "scale", i), {1}); + layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); + layer.ffn_up_scale = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "scale", i), {1}); + } + } break; + case LLM_ARCH_T5: + { + const auto n_rel_attn_bkts = hparams.n_rel_attn_bkts; + + model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); + + // output + { + model.output_norm_enc = ml.create_tensor(ctx_output, tn(LLM_TENSOR_ENC_OUTPUT_NORM, "weight"), {n_embd}); + model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_DEC_OUTPUT_NORM, "weight"), {n_embd}); + + model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED); + // if output is NULL, init from the input tok embed + if (model.output == NULL) { + model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); + } + } + + for (int i = 0; i < n_layer; ++i) { + ggml_context * ctx_layer = ctx_for_layer(i); + ggml_context * ctx_split = ctx_for_layer_split(i); + + auto & layer = model.layers[i]; + + layer.attn_norm_enc = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ENC_ATTN_NORM, "weight", i), {n_embd}); + layer.attn_rel_b_enc = ml.create_tensor(ctx_input, tn(LLM_TENSOR_ENC_ATTN_REL_B, "weight", i), {n_head, n_rel_attn_bkts}, llama_model_loader::TENSOR_NOT_REQUIRED); + + layer.wq_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_ATTN_Q, "weight", i), {n_embd, n_embd_k_gqa}); + layer.wk_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}); + layer.wv_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}); + layer.wo_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_ATTN_OUT, "weight", i), {n_embd_v_gqa, n_embd}); + + layer.ffn_norm_enc = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ENC_FFN_NORM, "weight", i), {n_embd}); + layer.ffn_gate_enc = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ENC_FFN_GATE, "weight", i), {n_embd, n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.ffn_down_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_FFN_DOWN, "weight", i), { n_ff, n_embd}); + layer.ffn_up_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_FFN_UP, "weight", i), {n_embd, n_ff}); + + layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_DEC_ATTN_NORM, "weight", i), {n_embd}); + layer.attn_rel_b = ml.create_tensor(ctx_input, tn(LLM_TENSOR_DEC_ATTN_REL_B, "weight", i), {n_head, n_rel_attn_bkts}, llama_model_loader::TENSOR_NOT_REQUIRED); + + layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_ATTN_Q, "weight", i), {n_embd, n_embd_k_gqa}); + layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}); + layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}); + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_ATTN_OUT, "weight", i), {n_embd_v_gqa, n_embd}); + + layer.attn_norm_cross = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_DEC_CROSS_ATTN_NORM, "weight", i), {n_embd}); + // this tensor seems to be unused in HF transformers implementation + layer.attn_rel_b_cross = ml.create_tensor(ctx_input, tn(LLM_TENSOR_DEC_CROSS_ATTN_REL_B, "weight", i), {n_head, n_rel_attn_bkts}, llama_model_loader::TENSOR_NOT_REQUIRED); + + layer.wq_cross = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_CROSS_ATTN_Q, "weight", i), {n_embd, n_embd_k_gqa}); + layer.wk_cross = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_CROSS_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}); + layer.wv_cross = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_CROSS_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}); + layer.wo_cross = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_CROSS_ATTN_OUT, "weight", i), {n_embd_v_gqa, n_embd}); + + layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_DEC_FFN_NORM, "weight", i), {n_embd}); + layer.ffn_gate = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_DEC_FFN_GATE, "weight", i), {n_embd, n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_FFN_DOWN, "weight", i), { n_ff, n_embd}); + layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_FFN_UP, "weight", i), {n_embd, n_ff}); + } + } break; + case LLM_ARCH_JAIS: + { + model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); + + // Output + { + model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); + model.output_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}); + model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}); + } + + for (int i = 0; i < n_layer; ++i) { + ggml_context * ctx_layer = ctx_for_layer(i); + ggml_context * ctx_split = ctx_for_layer_split(i); + + auto & layer = model.layers[i]; + + layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); + layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}); + + layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}); + layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}); + + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); + layer.bo = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}); + + layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); + layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}); + + layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}); + layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}); + + layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}); + layer.ffn_gate_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}); + + layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); + layer.ffn_up_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}); } } break; default: @@ -7321,8 +7779,8 @@ static void llm_build_kv_store( int64_t il) { const int64_t n_ctx = cparams.n_ctx; - const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(); - const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(); + const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il); + const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(il); GGML_ASSERT(kv.size == n_ctx); @@ -7615,12 +8073,12 @@ static struct ggml_tensor * llm_build_kqv( const llm_build_cb & cb, int il) { const int64_t n_ctx = cparams.n_ctx; - const int64_t n_head = hparams.n_head; - const int64_t n_head_kv = hparams.n_head_kv; + const int64_t n_head = hparams.n_head(il); + const int64_t n_head_kv = hparams.n_head_kv(il); const int64_t n_embd_head_k = hparams.n_embd_head_k; - const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(); + const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il); const int64_t n_embd_head_v = hparams.n_embd_head_v; - const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(); + const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(il); struct ggml_tensor * q = ggml_permute(ctx, q_cur, 0, 2, 1, 3); cb(q, "q", il); @@ -7796,6 +8254,7 @@ struct llm_build_context { const int32_t n_tokens; const int32_t n_kv; // size of KV cache to consider (n_kv <= kv_self.size) const int32_t n_outputs; + const int32_t n_outputs_enc; const int32_t kv_head; // index of where we store new KV data in the cache const int32_t n_ctx_orig; @@ -7826,8 +8285,8 @@ struct llm_build_context { n_layer (hparams.n_layer), n_rot (hparams.n_rot), n_ctx (cparams.n_ctx), - n_head (hparams.n_head), - n_head_kv (hparams.n_head_kv), + n_head (hparams.n_head()), + n_head_kv (hparams.n_head_kv()), n_embd_head_k (hparams.n_embd_head_k), n_embd_k_gqa (hparams.n_embd_k_gqa()), n_embd_head_v (hparams.n_embd_head_v), @@ -7845,6 +8304,7 @@ struct llm_build_context { n_tokens (batch.n_tokens), n_kv (worst_case ? kv_self.size : kv_self.n), n_outputs (worst_case ? n_tokens : lctx.n_outputs), + n_outputs_enc (worst_case ? n_tokens : lctx.embd_enc.size() / hparams.n_embd), kv_head (worst_case ? (kv_self.recurrent ? 0 : kv_self.size - n_tokens) : kv_self.head), n_ctx_orig (cparams.n_ctx_orig_yarn), flash_attn (cparams.flash_attn), @@ -7876,6 +8336,9 @@ struct llm_build_context { lctx.inp_s_copy = nullptr; lctx.inp_s_mask = nullptr; lctx.inp_s_seq = nullptr; + lctx.inp_pos_bucket = nullptr; + lctx.inp_embd_enc = nullptr; + lctx.inp_KQ_mask_cross = nullptr; } void free() { @@ -7895,6 +8358,8 @@ struct llm_build_context { ggml_set_input(lctx.inp_K_shift); for (int il = 0; il < n_layer; ++il) { + const int64_t n_head_kv = hparams.n_head_kv(il); + const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il); struct ggml_tensor * rope_factors = build_rope_factors(il); struct ggml_tensor * tmp = // we rotate only the first n_rot dimensions @@ -7954,6 +8419,9 @@ struct llm_build_context { } for (int il = 0; il < n_layer; ++il) { + const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il); + const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(il); + ggml_tensor * view_k_src = ggml_view_2d(ctx0, kv_self.k_l[il], n_embd_k_gqa, nm, ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa), @@ -8128,6 +8596,53 @@ struct llm_build_context { return gf; } + struct ggml_tensor * llm_build_pos_bucket(bool causal) { + if (causal) { + lctx.inp_pos_bucket = ggml_new_tensor_2d(ctx0, GGML_TYPE_I32, n_kv, n_tokens); + } else { + lctx.inp_pos_bucket = ggml_new_tensor_2d(ctx0, GGML_TYPE_I32, n_tokens, n_tokens); + } + + ggml_set_input(lctx.inp_pos_bucket); + cb(lctx.inp_pos_bucket, "pos_bucket", -1); + + return lctx.inp_pos_bucket; + } + + struct ggml_tensor * llm_build_pos_bias(struct ggml_tensor * pos_bucket, struct ggml_tensor * attn_rel_b) { + struct ggml_tensor * pos_bucket_1d = ggml_view_1d(ctx0, pos_bucket, pos_bucket->ne[0] * pos_bucket->ne[1], 0); + cb(pos_bucket_1d, "pos_bucket_1d", -1); + + struct ggml_tensor * pos_bias = ggml_get_rows(ctx0, attn_rel_b, pos_bucket_1d); + cb(pos_bias, "pos_bias", -1); + + pos_bias = ggml_view_3d(ctx0, pos_bias, pos_bias->ne[0], lctx.inp_pos_bucket->ne[0], lctx.inp_pos_bucket->ne[1], ggml_element_size(pos_bias) * pos_bias->ne[0], ggml_element_size(pos_bias) * pos_bias->ne[0] * lctx.inp_pos_bucket->ne[0], 0); + cb(pos_bias, "pos_bias", -1); + + pos_bias = ggml_permute(ctx0, pos_bias, 2, 0, 1, 3); + cb(pos_bias, "pos_bias", -1); + + pos_bias = ggml_cont(ctx0, pos_bias); + cb(pos_bias, "pos_bias", -1); + + return pos_bias; + } + + struct ggml_tensor * llm_build_inp_embd_enc() { + const int64_t n_embd = hparams.n_embd; + lctx.inp_embd_enc = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_outputs_enc); + ggml_set_input(lctx.inp_embd_enc); + cb(lctx.inp_embd_enc, "embd_enc", -1); + return lctx.inp_embd_enc; + } + + struct ggml_tensor * llm_build_inp_KQ_mask_cross() { + lctx.inp_KQ_mask_cross = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_outputs_enc, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); + ggml_set_input(lctx.inp_KQ_mask_cross); + cb(lctx.inp_KQ_mask_cross, "KQ_mask_cross", -1); + return lctx.inp_KQ_mask_cross; + } + struct ggml_cgraph * build_llama() { struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); @@ -11788,6 +12303,131 @@ struct llm_build_context { return gf; } + struct ggml_cgraph * build_openelm() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); + + const int64_t n_embd_head = hparams.n_embd_head_v; + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb); + + // inp_pos - contains the positions + struct ggml_tensor * inp_pos = build_inp_pos(); + + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); + + for (int il = 0; il < n_layer; ++il) { + const int64_t n_head = hparams.n_head(il); + const int64_t n_head_kv = hparams.n_head_kv(il); + const int64_t n_head_qkv = 2*n_head_kv + n_head; + + cur = inpL; + struct ggml_tensor * residual = cur; + + // norm + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "attn_norm", il); + + // self-attention + { + cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_reshape_3d(ctx0, cur, n_embd_head_k, n_head_qkv, n_tokens); + + struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, cur->nb[1], cur->nb[2], 0)); + cb(Qcur, "Qcur", il); + + struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, cur->nb[1], cur->nb[2], cur->nb[1]*n_head)); + cb(Kcur, "Kcur", il); + + struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, cur->nb[1], cur->nb[2], cur->nb[1]*(n_head+n_head_kv))); + cb(Vcur, "Vcur", il); + + Qcur = llm_build_norm(ctx0, Qcur, hparams, + model.layers[il].attn_q_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(Qcur, "Qcur", il); + + Kcur = llm_build_norm(ctx0, Kcur, hparams, + model.layers[il].attn_k_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(Kcur, "Kcur", il); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, NULL, n_rot, rope_type, n_ctx_orig, + freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, NULL, n_rot, rope_type, n_ctx_orig, + freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + Vcur = ggml_reshape_2d(ctx0, Vcur, n_embd_head * n_head_kv, n_tokens); + cb(Qcur, "Vcur", il); + + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, + model.layers[il].wo, NULL, + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + struct ggml_tensor * inp_out_ids = build_inp_out_ids(); + residual = ggml_get_rows(ctx0, residual, inp_out_ids); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + } + + struct ggml_tensor * ffn_inp = ggml_add(ctx0, residual, cur); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "ffn_norm", il); + + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, cb, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cur = lctx.cvec.apply_to(ctx0, cur, il); + cb(cur, "l_out", il); + + inpL = cur; + } + + cur = inpL; + + // norm + cur = llm_build_norm(ctx0, cur, hparams, + model.output_norm, NULL, + LLM_NORM_RMS, cb, -1); + cb(cur, "result_norm", -1); + + cur = ggml_mul_mat(ctx0, model.output, cur); + cb(cur, "result_output", -1); + + ggml_build_forward_expand(gf, cur); + + return gf; + } + struct ggml_cgraph * build_gptneox() { struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); @@ -12435,6 +13075,412 @@ struct llm_build_context { return gf; } + struct ggml_cgraph * build_t5() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); + + // mutable variable, needed during the last layer of the computation to skip unused tokens + int32_t n_tokens = this->n_tokens; + + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb); + + if (lctx.is_encoding) { + struct ggml_tensor * pos_bucket_enc = llm_build_pos_bucket(false); + + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask_enc = build_inp_KQ_mask(false); + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * inpSA = inpL; + + // norm + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm_enc, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "attn_norm", il); + + // self-attention + { + struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq_enc, cur); + cb(Qcur, "Qcur", il); + + struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk_enc, cur); + cb(Kcur, "Kcur", il); + + struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv_enc, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + + struct ggml_tensor * q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); + struct ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3)); + + struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q); + cb(kq, "kq", il); + + struct ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b_enc ? model.layers[il].attn_rel_b_enc : model.layers[0].attn_rel_b_enc; + struct ggml_tensor * pos_bias = llm_build_pos_bias(pos_bucket_enc, attn_rel_b); + struct ggml_tensor * kq_b = ggml_add(ctx0, kq, pos_bias); + cb(kq_b, "kq_b", il); + + kq = ggml_soft_max_ext(ctx0, kq_b, KQ_mask_enc, 1.0f, hparams.f_max_alibi_bias); + cb(kq, "kq_soft_max_ext", il); + + struct ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_tokens))); + cb(v, "v", il); + + struct ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_tokens, n_embd_head, n_head_kv), kq); + cb(kqv, "kqv", il); + + struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3); + cb(kqv_merged, "kqv_merged", il); + + cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens); + cb(cur, "kqv_merged_cont", il); + + ggml_build_forward_expand(gf, cur); + + cur = ggml_mul_mat(ctx0, model.layers[il].wo_enc, cur); + cb(cur, "kqv_out", il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + struct ggml_tensor * inp_out_ids = build_inp_out_ids(); + n_tokens = n_outputs; + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm_enc, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "ffn_norm", il); + + // T5 uses relu, flan-T5 uses gelu-gated + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up_enc, NULL, NULL, + model.layers[il].ffn_gate_enc, NULL, NULL, + model.layers[il].ffn_down_enc, NULL, NULL, + NULL, + model.layers[il].ffn_gate_enc ? LLM_FFN_GELU : LLM_FFN_RELU, + model.layers[il].ffn_gate_enc ? LLM_FFN_PAR : LLM_FFN_SEQ, + cb, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + ggml_tensor * layer_dir = lctx.cvec.tensor_for(il); + if (layer_dir != nullptr) { + cur = ggml_add(ctx0, cur, layer_dir); + } + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + cb(cur, "result_embd", -1); + + cur = llm_build_norm(ctx0, cur, hparams, + model.output_norm_enc, NULL, + LLM_NORM_RMS, cb, -1); + cb(cur, "result_norm", -1); + } else { + struct ggml_tensor * embd_enc = llm_build_inp_embd_enc(); + struct ggml_tensor * pos_bucket_dec = llm_build_pos_bucket(true); + + struct ggml_tensor * KQ_mask_dec = build_inp_KQ_mask(); + struct ggml_tensor * KQ_mask_cross = llm_build_inp_KQ_mask_cross(); + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * inpSA = inpL; + + // norm + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "attn_norm", il); + + // self-attention + { + struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + llm_build_kv_store(ctx0, hparams, cparams, kv_self, gf, Kcur, Vcur, n_tokens, kv_head, cb, il); + + struct ggml_tensor * k = + ggml_view_3d(ctx0, kv_self.k_l[il], + n_embd_head_k, n_kv, n_head_kv, + ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa), + ggml_row_size(kv_self.k_l[il]->type, n_embd_head_k), + 0); + cb(k, "k", il); + + struct ggml_tensor * v = + ggml_view_3d(ctx0, kv_self.v_l[il], + n_kv, n_embd_head_v, n_head_kv, + ggml_element_size(kv_self.v_l[il])*n_ctx, + ggml_element_size(kv_self.v_l[il])*n_ctx*n_embd_head_v, + 0); + cb(v, "v", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + + struct ggml_tensor * q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); + + struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q); + cb(kq, "kq", il); + + struct ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b ? model.layers[il].attn_rel_b : model.layers[0].attn_rel_b; + struct ggml_tensor * pos_bias = llm_build_pos_bias(pos_bucket_dec, attn_rel_b); + struct ggml_tensor * kq_b = ggml_add(ctx0, kq, pos_bias); + cb(kq_b, "kq_b", il); + + kq = ggml_soft_max_ext(ctx0, kq_b, KQ_mask_dec, 1.0f, hparams.f_max_alibi_bias); + cb(kq, "kq_soft_max_ext", il); + + struct ggml_tensor * kqv = ggml_mul_mat(ctx0, v, kq); + cb(kqv, "kqv", il); + + struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3); + cb(kqv_merged, "kqv_merged", il); + + cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens); + cb(cur, "kqv_merged_cont", il); + + ggml_build_forward_expand(gf, cur); + + cur = ggml_mul_mat(ctx0, model.layers[il].wo, cur); + cb(cur, "kqv_out", il); + } + + cur = ggml_add(ctx0, cur, inpSA); + cb(cur, "cross_inp", il); + + struct ggml_tensor * inpCA = cur; + + // norm + cur = llm_build_norm(ctx0, cur, hparams, + model.layers[il].attn_norm_cross, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "attn_norm_cross", il); + + // cross-attention + { + struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq_cross, cur); + cb(Qcur, "Qcur", il); + + struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk_cross, embd_enc); + cb(Kcur, "Kcur", il); + + struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv_cross, embd_enc); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_outputs_enc); + + struct ggml_tensor * q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); + struct ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3)); + + struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q); + cb(kq, "kq", il); + + kq = ggml_soft_max_ext(ctx0, kq, KQ_mask_cross, 1.0f, hparams.f_max_alibi_bias); + cb(kq, "kq_soft_max_ext", il); + + struct ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_outputs_enc))); + cb(v, "v", il); + + struct ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_outputs_enc, n_embd_head, n_head_kv), kq); + cb(kqv, "kqv", il); + + struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3); + cb(kqv_merged, "kqv_merged", il); + + cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens); + cb(cur, "kqv_merged_cont", il); + + ggml_build_forward_expand(gf, cur); + + cur = ggml_mul_mat(ctx0, model.layers[il].wo_cross, cur); + cb(cur, "kqv_out", il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + struct ggml_tensor * inp_out_ids = build_inp_out_ids(); + n_tokens = n_outputs; + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + inpCA = ggml_get_rows(ctx0, inpCA, inp_out_ids); + } + + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpCA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "ffn_norm", il); + + // T5 uses relu, flan-T5 uses gelu-gated + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + model.layers[il].ffn_gate_enc ? LLM_FFN_GELU : LLM_FFN_RELU, + model.layers[il].ffn_gate_enc ? LLM_FFN_PAR : LLM_FFN_SEQ, + cb, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + ggml_tensor * layer_dir = lctx.cvec.tensor_for(il); + if (layer_dir != nullptr) { + cur = ggml_add(ctx0, cur, layer_dir); + } + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + cb(cur, "result_embd", -1); + + cur = llm_build_norm(ctx0, cur, hparams, + model.output_norm, NULL, + LLM_NORM_RMS, cb, -1); + cb(cur, "result_norm", -1); + + // lm_head + cur = ggml_mul_mat(ctx0, model.output, cur); + cb(cur, "result_output", -1); + } + + ggml_build_forward_expand(gf, cur); + + return gf; + } + + struct ggml_cgraph * build_jais() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); + + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb); + + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); + + for (int il = 0; il < n_layer; ++il) { + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, cb, il); + cb(cur, "attn_norm", il); + + // self-attention + { + cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + + struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*cur->nb[0]*(n_embd))); + struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*cur->nb[0]*(n_embd))); + struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*cur->nb[0]*(n_embd + n_embd_gqa))); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, + model.layers[il].wo, model.layers[il].bo, + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/float(n_embd_head), cb, il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + struct ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + // add the input + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + // FF + { + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, cb, il); + cb(cur, "ffn_norm", il); + + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, cb, il); + cb(cur, "ffn_out", il); + } + + inpL = ggml_add(ctx0, cur, ffn_inp); + cb(inpL, "l_out", il); + } + + cur = llm_build_norm(ctx0, inpL, hparams, + model.output_norm, + model.output_norm_b, + LLM_NORM, cb, -1); + cb(cur, "result_norm", -1); + + cur = ggml_mul_mat(ctx0, model.output, cur); + + cb(cur, "result_output", -1); + + ggml_build_forward_expand(gf, cur); + + return gf; + } }; static struct ggml_cgraph * llama_build_graph_defrag(llama_context & lctx, const std::vector & ids) { @@ -12650,6 +13696,10 @@ static struct ggml_cgraph * llama_build_graph( { result = llm.build_olmo(); } break; + case LLM_ARCH_OPENELM: + { + result = llm.build_openelm(); + } break; case LLM_ARCH_GPTNEOX: { result = llm.build_gptneox(); @@ -12666,6 +13716,14 @@ static struct ggml_cgraph * llama_build_graph( { result = llm.build_bitnet(); } break; + case LLM_ARCH_T5: + { + result = llm.build_t5(); + } break; + case LLM_ARCH_JAIS: + { + result = llm.build_jais(); + } break; default: GGML_ASSERT(false); } @@ -12704,6 +13762,30 @@ static void llama_set_s_copy(llama_context & lctx) { } } +static int32_t llama_relative_position_bucket(llama_pos x, llama_pos y, uint64_t n_buckets, bool bidirectional) { + // TODO move to hparams if a T5 variant appears that uses a different value + const int64_t max_distance = 128; + + if (bidirectional) { + n_buckets >>= 1; + } + + const int64_t max_exact = n_buckets >> 1; + + int32_t relative_position = x - y; + int32_t relative_bucket = 0; + if (bidirectional) { + relative_bucket += (relative_position > 0) * n_buckets; + relative_position = abs(relative_position); + } else { + relative_position = -std::min(relative_position, 0); + } + int32_t relative_position_if_large = floorf(max_exact + logf(1.0 * relative_position / max_exact) * (n_buckets - max_exact) / log(1.0 * max_distance / max_exact)); + relative_position_if_large = std::min(relative_position_if_large, n_buckets - 1); + relative_bucket += (relative_position < max_exact ? relative_position : relative_position_if_large); + return relative_bucket; +} + static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { // // set input data @@ -12769,7 +13851,7 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { if (lctx.inp_KQ_mask) { // NOTE: hparams.causal_attn indicates the model is capable of generation and uses the kv cache. - if (cparams.causal_attn) { + if (cparams.causal_attn && !lctx.is_encoding) { const int64_t n_kv = kv_self.n; const int64_t n_tokens = batch.n_tokens; @@ -12822,7 +13904,7 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { } else { // when using kv cache, the mask needs to match the kv cache size const int64_t n_tokens = batch.n_tokens; - const int64_t n_stride = hparams.causal_attn ? kv_self.n : n_tokens; + const int64_t n_stride = hparams.causal_attn && !lctx.is_encoding ? kv_self.n : n_tokens; GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_KQ_mask->buffer)); @@ -12856,7 +13938,7 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { } } - if (cparams.pooling_type == LLAMA_POOLING_TYPE_MEAN) { + if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_MEAN) { const int64_t n_tokens = batch.n_tokens; GGML_ASSERT(lctx.inp_mean); @@ -12888,7 +13970,7 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { } } - if (cparams.pooling_type == LLAMA_POOLING_TYPE_CLS) { + if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_CLS) { const int64_t n_tokens = batch.n_tokens; GGML_ASSERT(lctx.inp_cls); @@ -12909,7 +13991,7 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { } } - if (cparams.pooling_type == LLAMA_POOLING_TYPE_LAST) { + if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_LAST) { const int64_t n_tokens = batch.n_tokens; GGML_ASSERT(lctx.inp_cls); @@ -12986,6 +14068,70 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { } } } + + if (lctx.inp_pos_bucket) { + const int64_t n_tokens = batch.n_tokens; + + GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_pos_bucket->buffer)); + + int32_t * data = (int32_t *) lctx.inp_pos_bucket->data; + + if (!lctx.is_encoding) { + const int64_t n_kv = kv_self.n; + for (int h = 0; h < 1; ++h) { + for (int j = 0; j < n_tokens; ++j) { + for (int i = 0; i < n_kv; ++i) { + data[h*(n_kv*n_tokens) + j*n_kv + i] = llama_relative_position_bucket(lctx.kv_self.cells[i].pos, batch.pos[j], hparams.n_rel_attn_bkts, lctx.is_encoding); + } + } + } + } else { + for (int h = 0; h < 1; ++h) { + for (int j = 0; j < n_tokens; ++j) { + for (int i = 0; i < n_tokens; ++i) { + data[h*(n_tokens*n_tokens) + j*n_tokens + i] = llama_relative_position_bucket(batch.pos[i], batch.pos[j], hparams.n_rel_attn_bkts, lctx.is_encoding); + } + } + } + } + } + + if (!lctx.is_encoding && lctx.inp_embd_enc) { + assert(lctx.inp_embd_enc->type == GGML_TYPE_F32); + assert((size_t) ggml_nelements(lctx.inp_embd_enc) == lctx.embd_enc.size()); + + ggml_backend_tensor_set(lctx.inp_embd_enc, lctx.embd_enc.data(), 0, ggml_nbytes(lctx.inp_embd_enc)); + } + + if (!lctx.is_encoding && lctx.inp_KQ_mask_cross) { + const int64_t n_output_enc = lctx.embd_enc.size() / hparams.n_embd; + const int64_t n_tokens = batch.n_tokens; + + GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_KQ_mask_cross->buffer)); + + float * data = (float *) lctx.inp_KQ_mask_cross->data; + + for (int h = 0; h < 1; ++h) { + for (int j = 0; j < n_tokens; ++j) { + for (int i = 0; i < n_output_enc; ++i) { + float f = -INFINITY; + for (int s = 0; s < batch.n_seq_id[j]; ++s) { + const llama_seq_id seq_id = batch.seq_id[j][s]; + if (lctx.seq_ids_enc[i].find(seq_id) != lctx.seq_ids_enc[i].end()) { + f = 0.0f; + } + } + data[h*(n_output_enc*n_tokens) + j*n_output_enc + i] = f; + } + } + + for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) { + for (int j = 0; j < n_output_enc; ++j) { + data[h*(n_output_enc*n_tokens) + i*n_output_enc + j] = -INFINITY; + } + } + } + } } // Make sure enough space is available for outputs. @@ -13002,7 +14148,7 @@ static size_t llama_output_reserve(llama_context & lctx, size_t n_outputs) { // TODO: use a per-batch flag for logits presence instead const bool has_logits = !cparams.embeddings; - const bool has_embd = cparams.embeddings && (cparams.pooling_type == LLAMA_POOLING_TYPE_NONE); + const bool has_embd = lctx.is_encoding || (cparams.embeddings && (cparams.pooling_type == LLAMA_POOLING_TYPE_NONE)); const size_t logits_size = has_logits ? n_vocab*n_outputs_max : 0; const size_t embd_size = has_embd ? n_embd*n_outputs_max : 0; @@ -13094,6 +14240,7 @@ static int llama_decode_internal( llama_context & lctx, llama_batch batch_all) { // TODO: rename back to batch + lctx.is_encoding = false; const uint32_t n_tokens_all = batch_all.n_tokens; if (n_tokens_all == 0) { @@ -13126,19 +14273,21 @@ static int llama_decode_internal( const auto n_ubatch = cparams.n_ubatch; + // TODO: simplify or deprecate std::vector pos; std::vector n_seq_id; std::vector seq_id_arr; std::vector> seq_id; + // this indicates we are doing pooled embedding, so we ignore batch.logits and output all tokens + const bool embd_pooled = cparams.embeddings && cparams.pooling_type != LLAMA_POOLING_TYPE_NONE; + // count outputs - if (cparams.embeddings && cparams.pooling_type != LLAMA_POOLING_TYPE_NONE) { - n_outputs = n_tokens_all; - } else if (batch_all.logits) { + if (batch_all.logits && !embd_pooled) { for (uint32_t i = 0; i < n_tokens_all; ++i) { n_outputs += batch_all.logits[i] != 0; } - } else if (lctx.logits_all) { + } else if (lctx.logits_all || embd_pooled) { n_outputs = n_tokens_all; } else { // keep last output only @@ -13184,7 +14333,7 @@ static int llama_decode_internal( { int32_t n_outputs_new = 0; - if (u_batch.logits) { + if (u_batch.logits && !embd_pooled) { for (uint32_t i = 0; i < n_tokens; i++) { n_outputs_new += u_batch.logits[i] != 0; } @@ -13389,6 +14538,138 @@ static int llama_decode_internal( return 0; } +// encode a batch of tokens by evaluating the encoder part of the transformer +// +// - lctx: llama context +// - batch: batch to evaluate +// +// return 0 on success +// return positive int on warning +// return negative int on error +// +static int llama_encode_internal( + llama_context & lctx, + llama_batch batch) { + + lctx.is_encoding = true; + + const uint32_t n_tokens = batch.n_tokens; + + if (n_tokens == 0) { + LLAMA_LOG_ERROR("%s: n_tokens == 0", __func__); + return -1; + } + + const auto & model = lctx.model; + const auto & hparams = model.hparams; + const auto & cparams = lctx.cparams; + + GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT + + // micro-batching is not possible for non-causal encoding, so we process the batch in a single shot + GGML_ASSERT(cparams.n_ubatch >= n_tokens && "encoder requires n_ubatch >= n_tokens"); + + if (lctx.t_compute_start_us == 0) { + lctx.t_compute_start_us = ggml_time_us(); + } + + lctx.n_queued_tokens += n_tokens; + + const int64_t n_embd = hparams.n_embd; + + // TODO: simplify or deprecate + std::vector pos; + std::vector n_seq_id; + std::vector seq_id_arr; + std::vector> seq_id; + + // reserve output buffer + if (llama_output_reserve(lctx, n_tokens) < n_tokens) { + LLAMA_LOG_ERROR("%s: could not reserve space for batch with %u outputs\n", __func__, n_tokens); + return -2; + }; + + for (uint32_t i = 0; i < n_tokens; ++i) { + lctx.output_ids[i] = i; + } + + lctx.inp_embd_enc = NULL; + lctx.n_outputs = n_tokens; + + const int n_threads = n_tokens == 1 ? cparams.n_threads : cparams.n_threads_batch; + GGML_ASSERT(n_threads > 0); + + // helpers for smoother batch API transition + // after deprecating the llama_eval calls, these will be removed + if (batch.pos == nullptr) { + pos.resize(n_tokens); + for (uint32_t i = 0; i < n_tokens; i++) { + pos[i] = batch.all_pos_0 + i*batch.all_pos_1; + } + + batch.pos = pos.data(); + } + + if (batch.seq_id == nullptr) { + n_seq_id.resize(n_tokens); + seq_id.resize(n_tokens); + seq_id_arr.resize(n_tokens); + for (uint32_t i = 0; i < n_tokens; i++) { + n_seq_id[i] = 1; + seq_id[i].resize(1); + seq_id[i][0] = batch.all_seq_id; + seq_id_arr[i] = seq_id[i].data(); + } + + batch.n_seq_id = n_seq_id.data(); + batch.seq_id = seq_id_arr.data(); + } + + ggml_backend_sched_reset(lctx.sched); + ggml_backend_sched_set_eval_callback(lctx.sched, lctx.cparams.cb_eval, lctx.cparams.cb_eval_user_data); + + ggml_cgraph * gf = llama_build_graph(lctx, batch, false); + + // the output embeddings after the final encoder normalization + struct ggml_tensor * embd = gf->nodes[gf->n_nodes - 1]; + + GGML_ASSERT(strcmp(embd->name, "result_norm") == 0); + + ggml_backend_sched_alloc_graph(lctx.sched, gf); + + llama_set_inputs(lctx, batch); + + llama_graph_compute(lctx, gf, n_threads); + + // extract embeddings + if (embd) { + ggml_backend_t backend_embd = ggml_backend_sched_get_tensor_backend(lctx.sched, embd); + GGML_ASSERT(backend_embd != nullptr); + + // extract token embeddings + GGML_ASSERT(lctx.embd != nullptr); + + lctx.embd_enc.resize(n_tokens*n_embd); + float * embd_out = lctx.embd_enc.data(); + + ggml_backend_tensor_get_async(backend_embd, embd, embd_out, 0, n_tokens*n_embd*sizeof(float)); + + // remember the sequence ids used during the encoding - needed for cross attention later + lctx.seq_ids_enc.resize(n_tokens); + for (uint32_t i = 0; i < n_tokens; i++) { + for (int s = 0; s < batch.n_seq_id[i]; s++) { + llama_seq_id seq_id = batch.seq_id[i][s]; + lctx.seq_ids_enc[i].insert(seq_id); + } + } + } + + // Reset state for the next token before backend sync, to allow the CPU activities in the reset to + // overlap with device computation. + ggml_backend_sched_reset(lctx.sched); + + return 0; +} // find holes from the beginning of the KV cache and fill them by moving data from the end of the cache static void llama_kv_cache_defrag_internal(struct llama_context & lctx) { @@ -14249,6 +15530,7 @@ struct llm_tokenizer_bpe { break; case LLAMA_VOCAB_PRE_TYPE_GPT2: case LLAMA_VOCAB_PRE_TYPE_OLMO: + case LLAMA_VOCAB_PRE_TYPE_JAIS: regex_exprs = { "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", }; @@ -14681,11 +15963,14 @@ struct llm_tokenizer_ugm { std::string normalized; normalize(text, &normalized); size_t input_len = normalized.size(); + if (input_len == 0) { + return; + } // initialize score_sum to -FLT_MAX so it will be always lower than sums of token scores - std::vector tokenization_results(input_len + 1, {0, 0, -FLT_MAX}); + std::vector tokenization_results(input_len + 1, {vocab.special_unk_id, 0, -FLT_MAX}); // at the beginning tokenization score is zero - tokenization_results[0] = { 0, 0, 0 }; + tokenization_results[0] = { vocab.special_unk_id, 0, 0 }; for (size_t input_offset = 0; input_offset < input_len;) { size_t prefix_offset = input_offset; @@ -14705,7 +15990,7 @@ struct llm_tokenizer_ugm { single_codepoint_token_found = true; } llama_token token_id = node->value; - const auto &token_data = vocab.id_to_token[token_id]; + const auto & token_data = vocab.id_to_token[token_id]; // we set the user-defined token scores to 0 to make them more likely to be selected // (normal token scores are log probabilities, so they are negative) @@ -14915,7 +16200,7 @@ private: size_t prefix_offset = input_offset; unicode_cpt_from_utf8(input, prefix_offset); return { &input[input_offset], prefix_offset - input_offset, prefix_offset - input_offset }; - } catch(std::invalid_argument & ex) { + } catch (std::invalid_argument & /*ex*/) { // if no, consume 1 byte and return U+FFFD - REPLACEMENT CHARACTER return { "\xEF\xBF\xBD", 3, 1 }; } @@ -16565,8 +17850,8 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n const llm_arch arch = qs.model.arch; const auto tn = LLM_TN(arch); - auto use_more_bits = [](int i_layer, int num_layers) -> bool { - return i_layer < num_layers/8 || i_layer >= 7*num_layers/8 || (i_layer - num_layers/8)%3 == 2; + auto use_more_bits = [](int i_layer, int n_layers) -> bool { + return i_layer < n_layers/8 || i_layer >= 7*n_layers/8 || (i_layer - n_layers/8)%3 == 2; }; const int n_expert = std::max(1, (int)qs.model.hparams.n_expert); auto layer_info = [n_expert] (int i_layer, int n_layer, const char * name) { @@ -17030,10 +18315,11 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s // sanity checks // - // - qs.n_attention_wv == 0 for Mamba models - // - qs.n_attention_wv == model.hparams.n_layer for Transformer models + // - qs.n_attention_wv == 0 for Mamba models + // - qs.n_attention_wv == model.hparams.n_layer for Transformer models + // - qs.n_attention_wv == 3 * model.hparams.n_layer for Encoder-Decoder models // - GGML_ASSERT((qs.n_attention_wv == 0 || qs.n_attention_wv == (int)model.hparams.n_layer) && "n_attention_wv is unexpected"); + GGML_ASSERT((qs.n_attention_wv == 0 || qs.n_attention_wv == (int)model.hparams.n_layer || qs.n_attention_wv == 3 * (int)model.hparams.n_layer) && "n_attention_wv is unexpected"); size_t total_size_org = 0; size_t total_size_new = 0; @@ -17158,6 +18444,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s quantize &= name.find("ssm_x.weight") == std::string::npos; quantize &= name.find("ssm_dt.weight") == std::string::npos; + // do not quantize relative position bias (T5) + quantize &= name.find("attn_rel_b.weight") == std::string::npos; + enum ggml_type new_type; void * new_data; size_t new_size; @@ -17606,6 +18895,7 @@ struct llama_context_params llama_context_default_params() { /*.n_threads_batch =*/ GGML_DEFAULT_N_THREADS, /*.rope_scaling_type =*/ LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED, /*.pooling_type =*/ LLAMA_POOLING_TYPE_UNSPECIFIED, + /*.attention_type =*/ LLAMA_ATTENTION_TYPE_UNSPECIFIED, /*.rope_freq_base =*/ 0.0f, /*.rope_freq_scale =*/ 0.0f, /*.yarn_ext_factor =*/ -1.0f, @@ -17846,7 +19136,6 @@ struct llama_context * llama_new_context_with_model( } cparams.yarn_attn_factor *= hparams.rope_attn_factor; - cparams.causal_attn = hparams.causal_attn; if (cparams.pooling_type == LLAMA_POOLING_TYPE_UNSPECIFIED) { if (hparams.pooling_type == LLAMA_POOLING_TYPE_UNSPECIFIED) { @@ -17856,6 +19145,12 @@ struct llama_context * llama_new_context_with_model( } } + if (params.attention_type == LLAMA_ATTENTION_TYPE_UNSPECIFIED) { + cparams.causal_attn = hparams.causal_attn; + } else { + cparams.causal_attn = params.attention_type == LLAMA_ATTENTION_TYPE_CAUSAL; + } + if (params.seed == LLAMA_DEFAULT_SEED) { params.seed = time(NULL); } @@ -18158,6 +19453,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) { case LLM_ARCH_MAMBA: case LLM_ARCH_JINA_BERT_V2: case LLM_ARCH_T5: + case LLM_ARCH_JAIS: return LLAMA_ROPE_TYPE_NONE; // use what we call a normal RoPE, operating on pairs of consecutive head values @@ -18192,6 +19488,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) { case LLM_ARCH_GEMMA: case LLM_ARCH_GEMMA2: case LLM_ARCH_STARCODER2: + case LLM_ARCH_OPENELM: case LLM_ARCH_GPTNEOX: return LLAMA_ROPE_TYPE_NEOX; @@ -18301,6 +19598,17 @@ struct ggml_tensor * llama_get_model_tensor(struct llama_model * model, const ch return it->second; } +bool llama_model_has_encoder(const struct llama_model * model) { + switch (model->arch) { + case LLM_ARCH_T5: return true; + default: return false; + } +} + +llama_token llama_model_decoder_start_token(const struct llama_model * model) { + return model->hparams.dec_start_token_id; +} + uint32_t llama_model_quantize( const char * fname_inp, const char * fname_out, @@ -18774,8 +20082,6 @@ static void llama_state_get_data_internal(struct llama_context * ctx, llama_data const auto & hparams = ctx->model.hparams; const uint32_t n_layer = hparams.n_layer; - const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa() + hparams.n_embd_k_s(); - const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa() + hparams.n_embd_v_s(); // NOTE: kv_size and kv_buf_size are mostly used for sanity checks const uint32_t kv_head = llama_kv_cache_cell_max(kv_self); @@ -18795,6 +20101,9 @@ static void llama_state_get_data_internal(struct llama_context * ctx, llama_data std::vector tmp_buf; for (int il = 0; il < (int) n_layer; ++il) { + const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s(); + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(); + const size_t k_size = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa*kv_head); tmp_buf.resize(k_size); @@ -18927,8 +20236,6 @@ size_t llama_state_set_data(struct llama_context * ctx, const uint8_t * src) { const auto & hparams = ctx->model.hparams; const uint32_t n_layer = hparams.n_layer; - const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa() + hparams.n_embd_k_s(); - const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa() + hparams.n_embd_v_s(); size_t kv_buf_size; uint32_t kv_head; @@ -18960,6 +20267,9 @@ size_t llama_state_set_data(struct llama_context * ctx, const uint8_t * src) { GGML_ASSERT(kv_self.total_size() >= kv_buf_size); for (int il = 0; il < (int) n_layer; ++il) { + const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s(); + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(); + const size_t k_size = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa*kv_head); ggml_backend_tensor_set(kv_self.k_l[il], inp, 0, k_size); @@ -19122,8 +20432,6 @@ size_t llama_state_seq_get_size(struct llama_context* ctx, llama_seq_id seq_id) const auto & hparams = ctx->model.hparams; const uint32_t n_layer = hparams.n_layer; - const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa() + hparams.n_embd_k_s(); - const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa() + hparams.n_embd_v_s(); for (uint32_t i = 0; i < kv_self.size; ++i) { const auto & cell = kv_self.cells[i]; @@ -19134,6 +20442,9 @@ size_t llama_state_seq_get_size(struct llama_context* ctx, llama_seq_id seq_id) } for (int il = 0; il < (int)n_layer; ++il) { + const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s(); + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(); + // types of keys and values s_cell_data_size += sizeof(int32_t) * 2; // k_size_row and v_size_el values of layer @@ -19208,14 +20519,15 @@ static size_t llama_state_seq_get_data_internal(struct llama_context * ctx, llam const auto & hparams = ctx->model.hparams; const uint32_t n_layer = hparams.n_layer; - const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa() + hparams.n_embd_k_s(); - const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa() + hparams.n_embd_v_s(); // Write the layer count data_ctx.write(&n_layer, sizeof(n_layer)); - // Write n_embd_v_gqa - data_ctx.write(&n_embd_v_gqa, sizeof(n_embd_v_gqa)); + // Write n_embd_v_gqa (reference value) + { + const uint32_t n_embd_v_gqa_ref = hparams.n_embd_v_gqa() + hparams.n_embd_k_s(); + data_ctx.write(&n_embd_v_gqa_ref, sizeof(n_embd_v_gqa_ref)); + } // Iterate the ranges and write all the pos (this is the token position in the prompt) for (const auto & range : cell_ranges) { @@ -19229,6 +20541,8 @@ static size_t llama_state_seq_get_data_internal(struct llama_context * ctx, llam // Get whole range at a time std::vector tmp_buf; for (int il = 0; il < (int)n_layer; ++il) { + const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s(); + // Write key type const int32_t k_type_i = (int32_t)kv_self.k_l[il]->type; data_ctx.write(&k_type_i, sizeof(k_type_i)); @@ -19249,6 +20563,8 @@ static size_t llama_state_seq_get_data_internal(struct llama_context * ctx, llam // TODO: simplify, reduce copy-paste if (!kv_self.v_trans) { for (int il = 0; il < (int)n_layer; ++il) { + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(); + // Write value type const int32_t v_type_i = (int32_t)kv_self.v_l[il]->type; data_ctx.write(&v_type_i, sizeof(v_type_i)); @@ -19269,6 +20585,8 @@ static size_t llama_state_seq_get_data_internal(struct llama_context * ctx, llam // For the values, they are transposed, so we also need the element size and get the element ranges from each row const uint32_t kv_size = kv_self.size; for (int il = 0; il < (int)n_layer; ++il) { + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(); + // Write value type const int32_t v_type_i = (int32_t)kv_self.v_l[il]->type; data_ctx.write(&v_type_i, sizeof(v_type_i)); @@ -19337,14 +20655,14 @@ size_t llama_state_seq_set_data(struct llama_context * ctx, const uint8_t * src, // Sanity check model compatibility const auto & hparams = ctx->model.hparams; const uint32_t n_layer = hparams.n_layer; - const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa() + hparams.n_embd_k_s(); - const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa() + hparams.n_embd_v_s(); + if (n_layer != n_layer_ref) { LLAMA_LOG_ERROR("%s: mismatched n_layer (%d != %d)\n", __func__, n_layer, n_layer_ref); return 0; } - if (n_embd_v_gqa != n_embd_v_gqa_ref) { - LLAMA_LOG_ERROR("%s: mismatched n_embd_v_gqa (%d != %d)\n", __func__, n_embd_v_gqa, n_embd_v_gqa_ref); + + if (hparams.n_embd_v_gqa() != n_embd_v_gqa_ref) { + LLAMA_LOG_ERROR("%s: mismatched n_embd_v_gqa (%d != %d)\n", __func__, hparams.n_embd_v_gqa(), n_embd_v_gqa_ref); return 0; } @@ -19384,6 +20702,8 @@ size_t llama_state_seq_set_data(struct llama_context * ctx, const uint8_t * src, // For each layer, read the keys for each cell, one row is one cell, read as one contiguous blo for (int il = 0; il < (int)n_layer; ++il) { + const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s(); + // Read type of key int32_t k_type_i_ref; memcpy(&k_type_i_ref, inp, sizeof(k_type_i_ref)); @@ -19416,6 +20736,8 @@ size_t llama_state_seq_set_data(struct llama_context * ctx, const uint8_t * src, // TODO: simplify, reduce copy-paste if (!kv_self.v_trans) { for (int il = 0; il < (int)n_layer; ++il) { + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(); + // Read type of value int32_t v_type_i_ref; memcpy(&v_type_i_ref, inp, sizeof(v_type_i_ref)); @@ -19447,6 +20769,8 @@ size_t llama_state_seq_set_data(struct llama_context * ctx, const uint8_t * src, } else { // For each layer, read the values for each cell (transposed) for (int il = 0; il < (int)n_layer; ++il) { + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(); + // Read type of value int32_t v_type_i_ref; memcpy(&v_type_i_ref, inp, sizeof(v_type_i_ref)); @@ -19658,6 +20982,17 @@ void llama_batch_free(struct llama_batch batch) { if (batch.logits) free(batch.logits); } +int32_t llama_encode( + struct llama_context * ctx, + struct llama_batch batch) { + const int ret = llama_encode_internal(*ctx, batch); + if (ret < 0) { + LLAMA_LOG_ERROR("%s: failed to encode, ret = %d\n", __func__, ret); + } + + return ret; +} + int32_t llama_decode( struct llama_context * ctx, struct llama_batch batch) { @@ -19899,7 +21234,7 @@ static std::string llama_decode_text(const std::string & text) { const auto utf8 = unicode_cpt_to_utf8(cpt); try { decoded_text += unicode_utf8_to_byte(utf8); - } catch (const std::out_of_range & e) { + } catch (const std::out_of_range & /*e*/) { decoded_text += "[UNK_BYTE_0x"; for (const auto c : utf8) { decoded_text += format("%02x", (uint8_t) c); diff --git a/src/unicode-data.cpp b/src/unicode-data.cpp index 4a939898b..02bdf7823 100644 --- a/src/unicode-data.cpp +++ b/src/unicode-data.cpp @@ -7030,4 +7030,3 @@ const std::vector unicode_ranges_nfd = { // start, last, nfd {0x02FA1C, 0x02FA1C, 0x009F3B}, {0x02FA1D, 0x02FA1D, 0x02A600}, }; -