diff --git a/.devops/full-cuda.Dockerfile b/.devops/full-cuda.Dockerfile index 059fd2695..c01006efe 100644 --- a/.devops/full-cuda.Dockerfile +++ b/.devops/full-cuda.Dockerfile @@ -31,6 +31,6 @@ ENV LLAMA_CUDA=1 # Enable cURL ENV LLAMA_CURL=1 -RUN make +RUN make -j$(nproc) ENTRYPOINT ["/app/.devops/tools.sh"] diff --git a/.devops/full-rocm.Dockerfile b/.devops/full-rocm.Dockerfile index 6ecf3bcc7..0314d469b 100644 --- a/.devops/full-rocm.Dockerfile +++ b/.devops/full-rocm.Dockerfile @@ -45,6 +45,6 @@ ENV LLAMA_CURL=1 RUN apt-get update && \ apt-get install -y libcurl4-openssl-dev -RUN make +RUN make -j$(nproc) ENTRYPOINT ["/app/.devops/tools.sh"] diff --git a/.devops/full.Dockerfile b/.devops/full.Dockerfile index 432fb5dad..6d5943a2f 100644 --- a/.devops/full.Dockerfile +++ b/.devops/full.Dockerfile @@ -18,7 +18,7 @@ COPY . . ENV LLAMA_CURL=1 -RUN make +RUN make -j$(nproc) ENV LC_ALL=C.utf8 diff --git a/.devops/main-cuda.Dockerfile b/.devops/main-cuda.Dockerfile index b937a4829..23f428944 100644 --- a/.devops/main-cuda.Dockerfile +++ b/.devops/main-cuda.Dockerfile @@ -23,7 +23,7 @@ ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH} # Enable CUDA ENV LLAMA_CUDA=1 -RUN make +RUN make -j$(nproc) FROM ${BASE_CUDA_RUN_CONTAINER} as runtime diff --git a/.devops/main-intel.Dockerfile b/.devops/main-intel.Dockerfile index 274b91b71..7516c8313 100644 --- a/.devops/main-intel.Dockerfile +++ b/.devops/main-intel.Dockerfile @@ -2,6 +2,14 @@ ARG ONEAPI_VERSION=2024.0.1-devel-ubuntu22.04 FROM intel/oneapi-basekit:$ONEAPI_VERSION as build +RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | tee /usr/share/keyrings/intel-oneapi-archive-keyring.gpg > /dev/null && \ + echo "deb [signed-by=/usr/share/keyrings/intel-oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main " | tee /etc/apt/sources.list.d/oneAPI.list && \ + chmod 644 /usr/share/keyrings/intel-oneapi-archive-keyring.gpg && \ + rm /etc/apt/sources.list.d/intel-graphics.list && \ + wget -O- https://repositories.intel.com/graphics/intel-graphics.key | gpg --dearmor | tee /usr/share/keyrings/intel-graphics.gpg > /dev/null && \ + echo "deb [arch=amd64,i386 signed-by=/usr/share/keyrings/intel-graphics.gpg] https://repositories.intel.com/graphics/ubuntu jammy arc" | tee /etc/apt/sources.list.d/intel.gpu.jammy.list && \ + chmod 644 /usr/share/keyrings/intel-graphics.gpg + ARG LLAMA_SYCL_F16=OFF RUN apt-get update && \ apt-get install -y git diff --git a/.devops/main-rocm.Dockerfile b/.devops/main-rocm.Dockerfile index 0a706dc73..37576d68e 100644 --- a/.devops/main-rocm.Dockerfile +++ b/.devops/main-rocm.Dockerfile @@ -40,6 +40,6 @@ ENV LLAMA_HIPBLAS=1 ENV CC=/opt/rocm/llvm/bin/clang ENV CXX=/opt/rocm/llvm/bin/clang++ -RUN make +RUN make -j$(nproc) ENTRYPOINT [ "/app/main" ] diff --git a/.devops/main.Dockerfile b/.devops/main.Dockerfile index 3ab1decd6..763d75fce 100644 --- a/.devops/main.Dockerfile +++ b/.devops/main.Dockerfile @@ -9,7 +9,7 @@ WORKDIR /app COPY . . -RUN make +RUN make -j$(nproc) FROM ubuntu:$UBUNTU_VERSION as runtime diff --git a/.devops/server-cuda.Dockerfile b/.devops/server-cuda.Dockerfile index 59a52ba21..7f5228185 100644 --- a/.devops/server-cuda.Dockerfile +++ b/.devops/server-cuda.Dockerfile @@ -25,7 +25,7 @@ ENV LLAMA_CUDA=1 # Enable cURL ENV LLAMA_CURL=1 -RUN make +RUN make -j$(nproc) FROM ${BASE_CUDA_RUN_CONTAINER} as runtime diff --git a/.devops/server-intel.Dockerfile b/.devops/server-intel.Dockerfile index a8e451fa9..13d00b737 100644 --- a/.devops/server-intel.Dockerfile +++ b/.devops/server-intel.Dockerfile @@ -2,6 +2,14 @@ ARG ONEAPI_VERSION=2024.0.1-devel-ubuntu22.04 FROM intel/oneapi-basekit:$ONEAPI_VERSION as build +RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | tee /usr/share/keyrings/intel-oneapi-archive-keyring.gpg > /dev/null && \ + echo "deb [signed-by=/usr/share/keyrings/intel-oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main " | tee /etc/apt/sources.list.d/oneAPI.list && \ + chmod 644 /usr/share/keyrings/intel-oneapi-archive-keyring.gpg && \ + rm /etc/apt/sources.list.d/intel-graphics.list && \ + wget -O- https://repositories.intel.com/graphics/intel-graphics.key | gpg --dearmor | tee /usr/share/keyrings/intel-graphics.gpg > /dev/null && \ + echo "deb [arch=amd64,i386 signed-by=/usr/share/keyrings/intel-graphics.gpg] https://repositories.intel.com/graphics/ubuntu jammy arc" | tee /etc/apt/sources.list.d/intel.gpu.jammy.list && \ + chmod 644 /usr/share/keyrings/intel-graphics.gpg + ARG LLAMA_SYCL_F16=OFF RUN apt-get update && \ apt-get install -y git libcurl4-openssl-dev @@ -19,6 +27,14 @@ RUN if [ "${LLAMA_SYCL_F16}" = "ON" ]; then \ FROM intel/oneapi-basekit:$ONEAPI_VERSION as runtime +RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | tee /usr/share/keyrings/intel-oneapi-archive-keyring.gpg > /dev/null && \ + echo "deb [signed-by=/usr/share/keyrings/intel-oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main " | tee /etc/apt/sources.list.d/oneAPI.list && \ + chmod 644 /usr/share/keyrings/intel-oneapi-archive-keyring.gpg && \ + rm /etc/apt/sources.list.d/intel-graphics.list && \ + wget -O- https://repositories.intel.com/graphics/intel-graphics.key | gpg --dearmor | tee /usr/share/keyrings/intel-graphics.gpg > /dev/null && \ + echo "deb [arch=amd64,i386 signed-by=/usr/share/keyrings/intel-graphics.gpg] https://repositories.intel.com/graphics/ubuntu jammy arc" | tee /etc/apt/sources.list.d/intel.gpu.jammy.list && \ + chmod 644 /usr/share/keyrings/intel-graphics.gpg + RUN apt-get update && \ apt-get install -y libcurl4-openssl-dev diff --git a/.devops/server-rocm.Dockerfile b/.devops/server-rocm.Dockerfile index c02a31dd8..a6b76dee8 100644 --- a/.devops/server-rocm.Dockerfile +++ b/.devops/server-rocm.Dockerfile @@ -45,6 +45,6 @@ ENV LLAMA_CURL=1 RUN apt-get update && \ apt-get install -y libcurl4-openssl-dev -RUN make +RUN make -j$(nproc) ENTRYPOINT [ "/app/server" ] diff --git a/.devops/server.Dockerfile b/.devops/server.Dockerfile index be964e0e8..0d09d3627 100644 --- a/.devops/server.Dockerfile +++ b/.devops/server.Dockerfile @@ -11,7 +11,7 @@ COPY . . ENV LLAMA_CURL=1 -RUN make +RUN make -j$(nproc) FROM ubuntu:$UBUNTU_VERSION as runtime diff --git a/.devops/tools.sh b/.devops/tools.sh index 3a7d274e4..97424c3aa 100755 --- a/.devops/tools.sh +++ b/.devops/tools.sh @@ -8,7 +8,7 @@ arg1="$1" shift if [[ "$arg1" == '--convert' || "$arg1" == '-c' ]]; then - python3 ./convert.py "$@" + python3 ./convert-hf-to-gguf.py "$@" elif [[ "$arg1" == '--quantize' || "$arg1" == '-q' ]]; then ./quantize "$@" elif [[ "$arg1" == '--run' || "$arg1" == '-r' ]]; then diff --git a/.github/ISSUE_TEMPLATE/01-bug-low.yml b/.github/ISSUE_TEMPLATE/01-bug-low.yml new file mode 100644 index 000000000..bfb9d9a06 --- /dev/null +++ b/.github/ISSUE_TEMPLATE/01-bug-low.yml @@ -0,0 +1,50 @@ +name: Low Severity Bugs +description: Used to report low severity bugs in llama.cpp (e.g. cosmetic issues, non critical UI glitches) +title: "Bug: " +labels: ["bug-unconfirmed", "low severity"] +body: + - type: markdown + attributes: + value: | + Thanks for taking the time to fill out this bug report! + Please include information about your system, the steps to reproduce the bug, + and the version of llama.cpp that you are using. + If possible, please provide a minimal code example that reproduces the bug. + - type: textarea + id: what-happened + attributes: + label: What happened? + description: Also tell us, what did you expect to happen? + placeholder: Tell us what you see! + validations: + required: true + - type: textarea + id: version + attributes: + label: Name and Version + description: Which executable and which version of our software are you running? (use `--version` to get a version string) + placeholder: | + $./main --version + version: 2999 (42b4109e) + built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu + validations: + required: true + - type: dropdown + id: operating-system + attributes: + label: What operating system are you seeing the problem on? + multiple: true + options: + - Linux + - Mac + - Windows + - BSD + - Other? (Please let us know in description) + validations: + required: false + - type: textarea + id: logs + attributes: + label: Relevant log output + description: Please copy and paste any relevant log output. This will be automatically formatted into code, so no need for backticks. + render: shell diff --git a/.github/ISSUE_TEMPLATE/02-bug-medium.yml b/.github/ISSUE_TEMPLATE/02-bug-medium.yml new file mode 100644 index 000000000..e8297eea0 --- /dev/null +++ b/.github/ISSUE_TEMPLATE/02-bug-medium.yml @@ -0,0 +1,50 @@ +name: Medium Severity Bug +description: Used to report medium severity bugs in llama.cpp (e.g. Malfunctioning Features but generally still useable) +title: "Bug: " +labels: ["bug-unconfirmed", "medium severity"] +body: + - type: markdown + attributes: + value: | + Thanks for taking the time to fill out this bug report! + Please include information about your system, the steps to reproduce the bug, + and the version of llama.cpp that you are using. + If possible, please provide a minimal code example that reproduces the bug. + - type: textarea + id: what-happened + attributes: + label: What happened? + description: Also tell us, what did you expect to happen? + placeholder: Tell us what you see! + validations: + required: true + - type: textarea + id: version + attributes: + label: Name and Version + description: Which executable and which version of our software are you running? (use `--version` to get a version string) + placeholder: | + $./main --version + version: 2999 (42b4109e) + built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu + validations: + required: true + - type: dropdown + id: operating-system + attributes: + label: What operating system are you seeing the problem on? + multiple: true + options: + - Linux + - Mac + - Windows + - BSD + - Other? (Please let us know in description) + validations: + required: false + - type: textarea + id: logs + attributes: + label: Relevant log output + description: Please copy and paste any relevant log output. This will be automatically formatted into code, so no need for backticks. + render: shell diff --git a/.github/ISSUE_TEMPLATE/03-bug-high.yml b/.github/ISSUE_TEMPLATE/03-bug-high.yml new file mode 100644 index 000000000..3c9d50d16 --- /dev/null +++ b/.github/ISSUE_TEMPLATE/03-bug-high.yml @@ -0,0 +1,50 @@ +name: High Severity Bug +description: Used to report high severity bugs in llama.cpp (e.g. Malfunctioning features hindering important common workflow) +title: "Bug: " +labels: ["bug-unconfirmed", "high severity"] +body: + - type: markdown + attributes: + value: | + Thanks for taking the time to fill out this bug report! + Please include information about your system, the steps to reproduce the bug, + and the version of llama.cpp that you are using. + If possible, please provide a minimal code example that reproduces the bug. + - type: textarea + id: what-happened + attributes: + label: What happened? + description: Also tell us, what did you expect to happen? + placeholder: Tell us what you see! + validations: + required: true + - type: textarea + id: version + attributes: + label: Name and Version + description: Which executable and which version of our software are you running? (use `--version` to get a version string) + placeholder: | + $./main --version + version: 2999 (42b4109e) + built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu + validations: + required: true + - type: dropdown + id: operating-system + attributes: + label: What operating system are you seeing the problem on? + multiple: true + options: + - Linux + - Mac + - Windows + - BSD + - Other? (Please let us know in description) + validations: + required: false + - type: textarea + id: logs + attributes: + label: Relevant log output + description: Please copy and paste any relevant log output. This will be automatically formatted into code, so no need for backticks. + render: shell diff --git a/.github/ISSUE_TEMPLATE/04-bug-critical.yml b/.github/ISSUE_TEMPLATE/04-bug-critical.yml new file mode 100644 index 000000000..d089d5fa1 --- /dev/null +++ b/.github/ISSUE_TEMPLATE/04-bug-critical.yml @@ -0,0 +1,50 @@ +name: Critical Severity Bug +description: Used to report critical severity bugs in llama.cpp (e.g. Crashing, Corrupted, Dataloss) +title: "Bug: " +labels: ["bug-unconfirmed", "critical severity"] +body: + - type: markdown + attributes: + value: | + Thanks for taking the time to fill out this bug report! + Please include information about your system, the steps to reproduce the bug, + and the version of llama.cpp that you are using. + If possible, please provide a minimal code example that reproduces the bug. + - type: textarea + id: what-happened + attributes: + label: What happened? + description: Also tell us, what did you expect to happen? + placeholder: Tell us what you see! + validations: + required: true + - type: textarea + id: version + attributes: + label: Name and Version + description: Which executable and which version of our software are you running? (use `--version` to get a version string) + placeholder: | + $./main --version + version: 2999 (42b4109e) + built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu + validations: + required: true + - type: dropdown + id: operating-system + attributes: + label: What operating system are you seeing the problem on? + multiple: true + options: + - Linux + - Mac + - Windows + - BSD + - Other? (Please let us know in description) + validations: + required: false + - type: textarea + id: logs + attributes: + label: Relevant log output + description: Please copy and paste any relevant log output. This will be automatically formatted into code, so no need for backticks. + render: shell diff --git a/.github/ISSUE_TEMPLATE/05-enhancement.yml b/.github/ISSUE_TEMPLATE/05-enhancement.yml new file mode 100644 index 000000000..58fca7318 --- /dev/null +++ b/.github/ISSUE_TEMPLATE/05-enhancement.yml @@ -0,0 +1,51 @@ +name: Enhancement +description: Used to request enhancements for llama.cpp +title: "Feature Request: " +labels: ["enhancement"] +body: + - type: markdown + attributes: + value: | + [Please post your idea first in Discussion if there is not yet a consensus for this enhancement request. This will help to keep this issue tracker focused on enhancements that the community has agreed needs to be implemented.](https://github.com/ggerganov/llama.cpp/discussions/categories/ideas) + + - type: checkboxes + id: prerequisites + attributes: + label: Prerequisites + description: Please confirm the following before submitting your enhancement request. + options: + - label: I am running the latest code. Mention the version if possible as well. + required: true + - label: I carefully followed the [README.md](https://github.com/ggerganov/llama.cpp/blob/master/README.md). + required: true + - label: I searched using keywords relevant to my issue to make sure that I am creating a new issue that is not already open (or closed). + required: true + - label: I reviewed the [Discussions](https://github.com/ggerganov/llama.cpp/discussions), and have a new and useful enhancement to share. + required: true + + - type: textarea + id: feature-description + attributes: + label: Feature Description + description: Please provide a detailed written description of what you were trying to do, and what you expected `llama.cpp` to do as an enhancement. + placeholder: Detailed description of the enhancement + validations: + required: true + + - type: textarea + id: motivation + attributes: + label: Motivation + description: Please provide a detailed written description of reasons why this feature is necessary and how it is useful to `llama.cpp` users. + placeholder: Explanation of why this feature is needed and its benefits + validations: + required: true + + - type: textarea + id: possible-implementation + attributes: + label: Possible Implementation + description: If you have an idea as to how it can be implemented, please write a detailed description. Feel free to give links to external sources or share visuals that might be helpful to understand the details better. + placeholder: Detailed description of potential implementation + validations: + required: false diff --git a/.github/ISSUE_TEMPLATE/06-research.yml b/.github/ISSUE_TEMPLATE/06-research.yml new file mode 100644 index 000000000..3ae4e9f8c --- /dev/null +++ b/.github/ISSUE_TEMPLATE/06-research.yml @@ -0,0 +1,52 @@ +name: Research +description: Track new technical research area +title: "Research: " +labels: ["research 🔬"] +body: + - type: markdown + attributes: + value: | + Don't forget to check for any [duplicate research issue tickets](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aopen+is%3Aissue+label%3A%22research+%F0%9F%94%AC%22) + + - type: checkboxes + id: research-stage + attributes: + label: Research Stage + description: Track general state of this research ticket + options: + - label: Background Research (Let's try to avoid reinventing the wheel) + - label: Hypothesis Formed (How do you think this will work and it's effect?) + - label: Strategy / Implementation Forming + - label: Analysis of results + - label: Debrief / Documentation (So people in the future can learn from us) + + - type: textarea + id: background + attributes: + label: Previous existing literature and research + description: Whats the current state of the art and whats the motivation for this research? + + - type: textarea + id: hypothesis + attributes: + label: Hypothesis + description: How do you think this will work and it's effect? + + - type: textarea + id: implementation + attributes: + label: Implementation + description: Got an approach? e.g. a PR ready to go? + + - type: textarea + id: analysis + attributes: + label: Analysis + description: How does the proposed implementation behave? + + - type: textarea + id: logs + attributes: + label: Relevant log output + description: Please copy and paste any relevant log output. This will be automatically formatted into code, so no need for backticks. + render: shell diff --git a/.github/ISSUE_TEMPLATE/07-refactor.yml b/.github/ISSUE_TEMPLATE/07-refactor.yml new file mode 100644 index 000000000..3a68d3d53 --- /dev/null +++ b/.github/ISSUE_TEMPLATE/07-refactor.yml @@ -0,0 +1,28 @@ +name: Refactor (Maintainers) +description: Used to track refactoring opportunities +title: "Refactor: " +labels: ["refactor"] +body: + - type: markdown + attributes: + value: | + Don't forget to [check for existing refactor issue tickets](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aopen+is%3Aissue+label%3Arefactoring) in case it's already covered. + Also you may want to check [Pull request refactor label as well](https://github.com/ggerganov/llama.cpp/pulls?q=is%3Aopen+is%3Apr+label%3Arefactoring) for duplicates too. + + - type: textarea + id: background-description + attributes: + label: Background Description + description: Please provide a detailed written description of the pain points you are trying to solve. + placeholder: Detailed description behind your motivation to request refactor + validations: + required: true + + - type: textarea + id: possible-approaches + attributes: + label: Possible Refactor Approaches + description: If you have some idea of possible approaches to solve this problem. You may want to make it a todo list. + placeholder: Your idea of possible refactoring opportunity/approaches + validations: + required: false diff --git a/.github/ISSUE_TEMPLATE/bug.md b/.github/ISSUE_TEMPLATE/bug.md deleted file mode 100644 index 49812832c..000000000 --- a/.github/ISSUE_TEMPLATE/bug.md +++ /dev/null @@ -1,11 +0,0 @@ ---- -name: Bug template -about: Used to report bugs in llama.cpp -labels: ["bug-unconfirmed"] -assignees: '' - ---- - -Please include information about your system, the steps to reproduce the bug, and the version of llama.cpp that you are using. If possible, please provide a minimal code example that reproduces the bug. - -If the bug concerns the server, please try to reproduce it first using the [server test scenario framework](https://github.com/ggerganov/llama.cpp/tree/master/examples/server/tests). diff --git a/.github/ISSUE_TEMPLATE/config.yml b/.github/ISSUE_TEMPLATE/config.yml new file mode 100644 index 000000000..c88134dbb --- /dev/null +++ b/.github/ISSUE_TEMPLATE/config.yml @@ -0,0 +1,13 @@ +blank_issues_enabled: true +contact_links: + - name: Got an idea? + url: https://github.com/ggerganov/llama.cpp/discussions/categories/ideas + about: Pop it there. It may then become an enhancement ticket. + - name: Got a question? + url: https://github.com/ggerganov/llama.cpp/discussions/categories/q-a + about: Ask a question there! + - name: Want to contribute? + url: https://github.com/ggerganov/llama.cpp/wiki/contribute + about: Head to the contribution guide page of the wiki for areas you can help with + + diff --git a/.github/ISSUE_TEMPLATE/enhancement.md b/.github/ISSUE_TEMPLATE/enhancement.md deleted file mode 100644 index dcffda750..000000000 --- a/.github/ISSUE_TEMPLATE/enhancement.md +++ /dev/null @@ -1,28 +0,0 @@ ---- -name: Enhancement template -about: Used to request enhancements for llama.cpp -labels: ["enhancement"] -assignees: '' - ---- - -# Prerequisites - -Please answer the following questions for yourself before submitting an issue. - -- [ ] I am running the latest code. Development is very rapid so there are no tagged versions as of now. -- [ ] I carefully followed the [README.md](https://github.com/ggerganov/llama.cpp/blob/master/README.md). -- [ ] I [searched using keywords relevant to my issue](https://docs.github.com/en/issues/tracking-your-work-with-issues/filtering-and-searching-issues-and-pull-requests) to make sure that I am creating a new issue that is not already open (or closed). -- [ ] I reviewed the [Discussions](https://github.com/ggerganov/llama.cpp/discussions), and have a new bug or useful enhancement to share. - -# Feature Description - -Please provide a detailed written description of what you were trying to do, and what you expected `llama.cpp` to do as an enhancement. - -# Motivation - -Please provide a detailed written description of reasons why this feature is necessary and how it is useful to `llama.cpp` users. - -# Possible Implementation - -If you have an idea as to how it can be implemented, please write a detailed description. Feel free to give links to external sources or share visuals that might be helpful to understand the details better. diff --git a/.github/labeler.yml b/.github/labeler.yml index a67f78044..97d739b58 100644 --- a/.github/labeler.yml +++ b/.github/labeler.yml @@ -1,5 +1,16 @@ # https://github.com/actions/labeler - +Kompute: + - changed-files: + - any-glob-to-any-file: + - ggml-kompute.h + - ggml-kompute.cpp + - README-kompute.md +Apple Metal: + - changed-files: + - any-glob-to-any-file: + - ggml-metal.h + - ggml-metal.cpp + - README-metal.md SYCL: - changed-files: - any-glob-to-any-file: @@ -9,6 +20,7 @@ SYCL: Nvidia GPU: - changed-files: - any-glob-to-any-file: + - ggml-cuda.h - ggml-cuda/** Vulkan: - changed-files: diff --git a/CMakeLists.txt b/CMakeLists.txt index ef02ff669..60cf7bdc4 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -72,6 +72,7 @@ else() set(INS_ENB ON) endif() +option(LLAMA_SVE "llama: enable SVE" OFF) option(LLAMA_AVX "llama: enable AVX" ${INS_ENB}) option(LLAMA_AVX2 "llama: enable AVX2" ${INS_ENB}) option(LLAMA_AVX512 "llama: enable AVX512" OFF) @@ -627,6 +628,10 @@ if (LLAMA_SYCL) add_compile_definitions(GGML_SYCL_F16) endif() + if (LLAMA_CUDA_FORCE_MMQ) + add_compile_definitions(GGML_SYCL_FORCE_MMQ) + endif() + add_compile_options(-I./) #include DPCT add_compile_options(-I/${SYCL_INCLUDE_DIR}) @@ -1040,6 +1045,9 @@ if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" OR CMAKE_GENERATOR_PLATFORM_LWR STR # Raspberry Pi 3, 4, Zero 2 (32-bit) list(APPEND ARCH_FLAGS -mno-unaligned-access) endif() + if (LLAMA_SVE) + list(APPEND ARCH_FLAGS -march=armv8.6-a+sve) + endif() endif() elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND @@ -1306,7 +1314,7 @@ set_target_properties(llama PROPERTIES PUBLIC_HEADER ${CMAKE_CURRENT_SOURCE_DIR} install(TARGETS llama LIBRARY PUBLIC_HEADER) install( - FILES convert.py + FILES convert-hf-to-gguf.py PERMISSIONS OWNER_READ OWNER_WRITE diff --git a/CMakePresets.json b/CMakePresets.json index ad1af7ecc..e2b7a79e3 100644 --- a/CMakePresets.json +++ b/CMakePresets.json @@ -1,4 +1,4 @@ -{ +{ "version": 4, "configurePresets": [ { @@ -40,6 +40,10 @@ { "name": "arm64-windows-msvc-debug" , "inherits": [ "base", "arm64-windows-msvc", "debug" ] }, { "name": "arm64-windows-msvc-release", "inherits": [ "base", "arm64-windows-msvc", "release" ] }, - { "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc", "release", "static" ] } + { "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc", "release", "static" ] }, + + { "name": "x64-windows-msvc-debug" , "inherits": [ "base", "debug" ] }, + { "name": "x64-windows-msvc-release", "inherits": [ "base", "release" ] }, + { "name": "x64-windows-msvc+static-release", "inherits": [ "base", "release", "static" ] } ] } diff --git a/Makefile b/Makefile index fe63cbd60..aced0e370 100644 --- a/Makefile +++ b/Makefile @@ -441,6 +441,9 @@ endif # JETSON_EOL_MODULE_DETECT ifdef LLAMA_DEBUG MK_NVCCFLAGS += -lineinfo endif # LLAMA_DEBUG +ifdef LLAMA_CUDA_DEBUG + MK_NVCCFLAGS += --device-debug +endif # LLAMA_CUDA_DEBUG ifdef LLAMA_CUDA_NVCC NVCC = $(CCACHE) $(LLAMA_CUDA_NVCC) else @@ -568,6 +571,7 @@ ifdef LLAMA_HIP_UMA MK_CPPFLAGS += -DGGML_HIP_UMA endif # LLAMA_HIP_UMA MK_LDFLAGS += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib + MK_LDFLAGS += -L$(ROCM_PATH)/lib64 -Wl,-rpath=$(ROCM_PATH)/lib64 MK_LDFLAGS += -lhipblas -lamdhip64 -lrocblas HIPFLAGS += $(addprefix --offload-arch=,$(AMDGPU_TARGETS)) HIPFLAGS += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X) diff --git a/README-sycl.md b/README-sycl.md index cfa248a95..37f0306dc 100644 --- a/README-sycl.md +++ b/README-sycl.md @@ -54,10 +54,10 @@ It has the similar design of other llama.cpp BLAS-based paths such as *OpenBLAS, ## OS -| OS | Status | Verified | -|---------|---------|------------------------------------| -| Linux | Support | Ubuntu 22.04, Fedora Silverblue 39 | -| Windows | Support | Windows 11 | +| OS | Status | Verified | +|---------|---------|------------------------------------------------| +| Linux | Support | Ubuntu 22.04, Fedora Silverblue 39, Arch Linux | +| Windows | Support | Windows 11 | ## Hardware @@ -70,7 +70,7 @@ It has the similar design of other llama.cpp BLAS-based paths such as *OpenBLAS, |-------------------------------|---------|---------------------------------------| | Intel Data Center Max Series | Support | Max 1550, 1100 | | Intel Data Center Flex Series | Support | Flex 170 | -| Intel Arc Series | Support | Arc 770, 730M | +| Intel Arc Series | Support | Arc 770, 730M, Arc A750 | | Intel built-in Arc GPU | Support | built-in Arc GPU in Meteor Lake | | Intel iGPU | Support | iGPU in i5-1250P, i7-1260P, i7-1165G7 | diff --git a/README.md b/README.md index 2ee267fdf..e8bf1e246 100644 --- a/README.md +++ b/README.md @@ -2,7 +2,9 @@ ![llama](https://user-images.githubusercontent.com/1991296/230134379-7181e485-c521-4d23-a0d6-f7b3b61ba524.png) -[![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT) [![Server](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml/badge.svg?branch=master&event=schedule)](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml) +[![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT) +[![Server](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml/badge.svg?branch=master&event=schedule)](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml) +[![Conan Center](https://shields.io/conan/v/llama-cpp)](https://conan.io/center/llama-cpp) [Roadmap](https://github.com/users/ggerganov/projects/7) / [Project status](https://github.com/ggerganov/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggerganov/llama.cpp/discussions/205) / [ggml](https://github.com/ggerganov/ggml) @@ -20,7 +22,8 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) ### Hot topics -- **Initial Flash-Attention support: https://github.com/ggerganov/llama.cpp/pull/5021** +- **`convert.py` has been deprecated and moved to `examples/convert-legacy-llama.py`, please use `convert-hf-to-gguf.py`** https://github.com/ggerganov/llama.cpp/pull/7430 +- Initial Flash-Attention support: https://github.com/ggerganov/llama.cpp/pull/5021 - BPE pre-tokenization support has been added: https://github.com/ggerganov/llama.cpp/pull/6920 - MoE memory layout has been updated - reconvert models for `mmap` support and regenerate `imatrix` https://github.com/ggerganov/llama.cpp/pull/6387 - Model sharding instructions using `gguf-split` https://github.com/ggerganov/llama.cpp/discussions/6404 @@ -200,9 +203,14 @@ Unless otherwise noted these projects are open-source with permissive licensing: - [KodiBot](https://github.com/firatkiral/kodibot) (GPL) - [eva](https://github.com/ylsdamxssjxxdd/eva) (MIT) - [AI Sublime Text plugin](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (MIT) +- [AIKit](https://github.com/sozercan/aikit) (MIT) *(to have a project listed here, it should clearly state that it depends on `llama.cpp`)* +**Tools:** + +- [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from HuggingFace Hub and convert them to GGML + --- Here is a typical run using LLaMA v2 13B on M2 Ultra: @@ -311,8 +319,6 @@ In order to build llama.cpp you have four different options. make ``` - **Note**: for `Debug` builds, run `make LLAMA_DEBUG=1` - - On Windows: 1. Download the latest fortran version of [w64devkit](https://github.com/skeeto/w64devkit/releases). @@ -324,23 +330,32 @@ In order to build llama.cpp you have four different options. make ``` + - Notes: + - For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `make -j 8` will run 8 jobs in parallel. + - For faster repeated compilation, install [ccache](https://ccache.dev/). + - For debug builds, run `make LLAMA_DEBUG=1` + - Using `CMake`: - ```bash - cmake -B build - cmake --build build --config Release - ``` + ```bash + cmake -B build + cmake --build build --config Release + ``` - **Note**: for `Debug` builds, there are two cases: + **Notes**: - - Single-config generators (e.g. default = `Unix Makefiles`; note that they just ignore the `--config` flag): + - For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `cmake --build build --config Release -j 8` will run 8 jobs in parallel. + - For faster repeated compilation, install [ccache](https://ccache.dev/). + - For debug builds, there are two cases: + + 1. Single-config generators (e.g. default = `Unix Makefiles`; note that they just ignore the `--config` flag): ```bash cmake -B build -DCMAKE_BUILD_TYPE=Debug cmake --build build ``` - - Multi-config generators (`-G` param set to Visual Studio, XCode...): + 2. Multi-config generators (`-G` param set to Visual Studio, XCode...): ```bash cmake -B build -G "Xcode" @@ -375,6 +390,14 @@ In order to build llama.cpp you have four different options. CLBLAST support for use OpenCL GPU acceleration in FreeBSD. Please read the instructions for use and activate this options in this document below. +### Homebrew + +On Mac and Linux, the homebrew package manager can be used via +``` +brew install llama.cpp +``` +The formula is automatically updated with new `llama.cpp` releases. More info: https://github.com/ggerganov/llama.cpp/discussions/7668 + ### Metal Build On MacOS, Metal is enabled by default. Using Metal makes the computation run on the GPU. @@ -473,7 +496,8 @@ Building the program with BLAS support may lead to some performance improvements |--------------------------------|------------------------|---------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| | LLAMA_CUDA_FORCE_DMMV | Boolean | false | Force the use of dequantization + matrix vector multiplication kernels instead of using kernels that do matrix vector multiplication on quantized data. By default the decision is made based on compute capability (MMVQ for 6.1/Pascal/GTX 1000 or higher). Does not affect k-quants. | | LLAMA_CUDA_DMMV_X | Positive integer >= 32 | 32 | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. | - | LLAMA_CUDA_MMV_Y | Positive integer | 1 | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. | + | LLAMA_CUDA_MMV_Y | Positive integer | 1 | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. | + | LLAMA_CUDA_FORCE_MMQ | Boolean | false | Force the use of dequantization + matrix multiplication kernels instead of leveraging Math libraries. | | | LLAMA_CUDA_F16 | Boolean | false | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs. | | LLAMA_CUDA_KQUANTS_ITER | 1 or 2 | 2 | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. | | LLAMA_CUDA_PEER_MAX_BATCH_SIZE | Positive integer | 128 | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial. | @@ -692,7 +716,8 @@ Building the program with BLAS support may lead to some performance improvements To obtain the official LLaMA 2 weights please see the Obtaining and using the Facebook LLaMA 2 model section. There is also a large selection of pre-quantized `gguf` models available on Hugging Face. -Note: `convert.py` does not support LLaMA 3, you can use `convert-hf-to-gguf.py` with LLaMA 3 downloaded from Hugging Face. +Note: `convert.py` has been moved to `examples/convert-legacy-llama.py` and shouldn't be used for anything other than `Llama/Llama2/Mistral` models and their derievatives. +It does not support LLaMA 3, you can use `convert-hf-to-gguf.py` with LLaMA 3 downloaded from Hugging Face. ```bash # obtain the official LLaMA model weights and place them in ./models @@ -709,10 +734,10 @@ ls ./models python3 -m pip install -r requirements.txt # convert the model to ggml FP16 format -python3 convert.py models/mymodel/ +python3 convert-hf-to-gguf.py models/mymodel/ # [Optional] for models using BPE tokenizers -python convert.py models/mymodel/ --vocab-type bpe +python convert-hf-to-gguf.py models/mymodel/ --vocab-type bpe # quantize the model to 4-bits (using Q4_K_M method) ./quantize ./models/mymodel/ggml-model-f16.gguf ./models/mymodel/ggml-model-Q4_K_M.gguf Q4_K_M diff --git a/ci/run.sh b/ci/run.sh index 940299025..3fc5f48b2 100755 --- a/ci/run.sh +++ b/ci/run.sh @@ -287,7 +287,7 @@ function gg_run_open_llama_7b_v2 { (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DLLAMA_CUDA=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log (time make -j ) 2>&1 | tee -a $OUT/${ci}-make.log - python3 ../convert.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf + python3 ../examples/convert-legacy-llama.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf model_f16="${path_models}/ggml-model-f16.gguf" model_q8_0="${path_models}/ggml-model-q8_0.gguf" diff --git a/common/common.cpp b/common/common.cpp index 1a284c56d..0c573c57f 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -1156,6 +1156,10 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa params.interactive_specials = true; return true; } + if (arg == "--special") { + params.special = true; + return true; + } if (arg == "--embedding") { params.embedding = true; return true; @@ -1614,6 +1618,7 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param printf(" -h, --help show this help message and exit\n"); printf(" --version show version and build info\n"); printf(" -i, --interactive run in interactive mode\n"); + printf(" --special special tokens output enabled\n"); printf(" --interactive-specials allow special tokens in user text, in interactive mode\n"); printf(" --interactive-first run in interactive mode and wait for input right away\n"); printf(" -cnv, --conversation run in conversation mode (does not print special tokens and suffix/prefix)\n"); @@ -2135,11 +2140,15 @@ bool fs_create_directory_with_parents(const std::string & path) { std::string fs_get_cache_directory() { std::string cache_directory = ""; + auto ensure_trailing_slash = [](std::string p) { + // Make sure to add trailing slash + if (p.back() != DIRECTORY_SEPARATOR) { + p += DIRECTORY_SEPARATOR; + } + return p; + }; if (getenv("LLAMA_CACHE")) { cache_directory = std::getenv("LLAMA_CACHE"); - if (cache_directory.back() != DIRECTORY_SEPARATOR) { - cache_directory += DIRECTORY_SEPARATOR; - } } else { #ifdef __linux__ if (std::getenv("XDG_CACHE_HOME")) { @@ -2150,12 +2159,12 @@ std::string fs_get_cache_directory() { #elif defined(__APPLE__) cache_directory = std::getenv("HOME") + std::string("/Library/Caches/"); #elif defined(_WIN32) - cache_directory = std::getenv("APPDATA"); + cache_directory = std::getenv("LOCALAPPDATA"); #endif // __linux__ + cache_directory = ensure_trailing_slash(cache_directory); cache_directory += "llama.cpp"; - cache_directory += DIRECTORY_SEPARATOR; } - return cache_directory; + return ensure_trailing_slash(cache_directory); } @@ -3137,6 +3146,7 @@ void yaml_dump_non_result_info(FILE * stream, const gpt_params & params, const l fprintf(stream, "cpu_has_fma: %s\n", ggml_cpu_has_fma() ? "true" : "false"); fprintf(stream, "cpu_has_gpublas: %s\n", ggml_cpu_has_gpublas() ? "true" : "false"); fprintf(stream, "cpu_has_neon: %s\n", ggml_cpu_has_neon() ? "true" : "false"); + fprintf(stream, "cpu_has_sve: %s\n", ggml_cpu_has_sve() ? "true" : "false"); fprintf(stream, "cpu_has_f16c: %s\n", ggml_cpu_has_f16c() ? "true" : "false"); fprintf(stream, "cpu_has_fp16_va: %s\n", ggml_cpu_has_fp16_va() ? "true" : "false"); fprintf(stream, "cpu_has_wasm_simd: %s\n", ggml_cpu_has_wasm_simd() ? "true" : "false"); diff --git a/common/common.h b/common/common.h index 039042c81..079fae033 100644 --- a/common/common.h +++ b/common/common.h @@ -156,6 +156,7 @@ struct gpt_params { bool use_color = false; // use color to distinguish generations and inputs bool interactive = false; // interactive mode bool interactive_specials = false; // whether to allow special tokens from user, during interactive mode + bool special = false; // enable special token output bool conversation = false; // conversation mode (does not print special tokens and suffix/prefix) bool chatml = false; // chatml mode (used for models trained on chatml syntax) bool prompt_cache_all = false; // save user input and generations to prompt cache diff --git a/common/train.cpp b/common/train.cpp index 2d41a1d29..fef1e57c9 100644 --- a/common/train.cpp +++ b/common/train.cpp @@ -1052,7 +1052,7 @@ struct train_params_common get_default_train_params_common() { params.custom_n_ctx = false; - params.use_flash = true; + params.use_flash = false; params.use_checkpointing = true; params.sample_start = ""; diff --git a/convert-hf-to-gguf-update.py b/convert-hf-to-gguf-update.py index 1923b88ba..84b72348d 100755 --- a/convert-hf-to-gguf-update.py +++ b/convert-hf-to-gguf-update.py @@ -81,6 +81,7 @@ models = [ {"name": "jina-v2-en", "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-en", }, # WPM! {"name": "jina-v2-es", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-es", }, {"name": "jina-v2-de", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-de", }, + {"name": "smaug-bpe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/abacusai/Smaug-Llama-3-70B-Instruct", }, ] diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index 5a00a5e89..ad071b974 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -25,8 +25,6 @@ if 'NO_LOCAL_GGUF' not in os.environ: sys.path.insert(1, str(Path(__file__).parent / 'gguf-py')) import gguf -from convert import LlamaHfVocab - logger = logging.getLogger("hf-to-gguf") @@ -313,11 +311,10 @@ class Model: data = data.astype(np.float32) data_qtype = gguf.GGMLQuantizationType.F32 - block_size, type_size = gguf.GGML_QUANT_SIZES[data_qtype] + shape = gguf.quant_shape_from_byte_shape(data.shape, data_qtype) if data.dtype == np.uint8 else data.shape + # reverse shape to make it similar to the internal ggml dimension order - shape_str = f"""{{{', '.join(str(n) for n in reversed( - (*data.shape[:-1], data.shape[-1] * data.dtype.itemsize // type_size * block_size)) - )}}}""" + shape_str = f"{{{', '.join(str(n) for n in reversed(shape))}}}" # n_dims is implicit in the shape logger.info(f"{f'%-{max_name_len}s' % f'{new_name},'} {old_dtype} --> {data_qtype.name}, shape = {shape_str}") @@ -474,6 +471,9 @@ class Model: if chkhsh == "27949a2493fc4a9f53f5b9b029c82689cfbe5d3a1929bb25e043089e28466de6": # ref: https://huggingface.co/jinaai/jina-embeddings-v2-base-de res = "jina-v2-de" + if chkhsh == "c136ed14d01c2745d4f60a9596ae66800e2b61fa45643e72436041855ad4089d": + # ref: https://huggingface.co/abacusai/Smaug-Llama-3-70B-Instruct + res = "smaug-bpe" if res is None: logger.warning("\n") @@ -632,7 +632,7 @@ class Model: special_vocab.add_to_gguf(self.gguf_writer) def _set_vocab_llama_hf(self): - vocab = LlamaHfVocab(self.dir_model) + vocab = gguf.LlamaHfVocab(self.dir_model) tokens = [] scores = [] toktypes = [] @@ -1315,6 +1315,17 @@ class LlamaModel(Model): self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR) self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"]) + tokenizer_config_file = self.dir_model / 'tokenizer_config.json' + if tokenizer_config_file.is_file(): + with open(tokenizer_config_file, "r", encoding="utf-8") as f: + tokenizer_config_json = json.load(f) + if "add_prefix_space" in tokenizer_config_json: + self.gguf_writer.add_add_space_prefix(tokenizer_config_json["add_prefix_space"]) + + # Apply to granite small models only + if self.hparams.get("vocab_size", 32000) == 49152: + self.gguf_writer.add_add_bos_token(False) + @staticmethod def permute(weights: Tensor, n_head: int, n_head_kv: int | None): if n_head_kv is not None and n_head != n_head_kv: @@ -1329,9 +1340,9 @@ class LlamaModel(Model): n_head = self.hparams["num_attention_heads"] n_kv_head = self.hparams.get("num_key_value_heads") - if name.endswith("q_proj.weight"): + if name.endswith(("q_proj.weight", "q_proj.bias")): data_torch = LlamaModel.permute(data_torch, n_head, n_head) - if name.endswith("k_proj.weight"): + if name.endswith(("k_proj.weight", "k_proj.bias")): data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head) # process the experts separately @@ -2393,7 +2404,8 @@ class CommandR2Model(Model): # max_position_embeddings = 8192 in config.json but model was actually # trained on 128k context length - self.hparams["max_position_embeddings"] = self.hparams["model_max_length"] + # aya-23 models don't have model_max_length specified + self.hparams["max_position_embeddings"] = self.find_hparam(["model_max_length", "max_position_embeddings"]) def set_gguf_parameters(self): super().set_gguf_parameters() @@ -2466,6 +2478,236 @@ class JinaBertV2Model(BertModel): self.gguf_writer.add_add_eos_token(True) +@Model.register("ArcticForCausalLM") +class ArcticModel(Model): + model_arch = gguf.MODEL_ARCH.ARCTIC + + def set_vocab(self): + # The reason for using a custom implementation here is that the + # snowflake-arctic-instruct model redefined tokens 31998 and 31999 from + # tokenizer.model and used them as BOS and EOS instead of adding new tokens. + from sentencepiece import SentencePieceProcessor + + tokenizer_path = self.dir_model / 'tokenizer.model' + + if not tokenizer_path.is_file(): + logger.error(f'Error: Missing {tokenizer_path}') + sys.exit(1) + + # Read the whole vocabulary from the tokenizer.model file + tokenizer = SentencePieceProcessor() + tokenizer.LoadFromFile(str(tokenizer_path)) + + vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size()) + + tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)] + scores: list[float] = [-10000.0] * vocab_size + toktypes: list[int] = [SentencePieceTokenTypes.UNKNOWN] * vocab_size + + for token_id in range(tokenizer.vocab_size()): + + piece = tokenizer.IdToPiece(token_id) + text = piece.encode("utf-8") + score = tokenizer.GetScore(token_id) + + toktype = SentencePieceTokenTypes.NORMAL + if tokenizer.IsUnknown(token_id): + toktype = SentencePieceTokenTypes.UNKNOWN + elif tokenizer.IsControl(token_id): + toktype = SentencePieceTokenTypes.CONTROL + elif tokenizer.IsUnused(token_id): + toktype = SentencePieceTokenTypes.UNUSED + elif tokenizer.IsByte(token_id): + toktype = SentencePieceTokenTypes.BYTE + + tokens[token_id] = text + scores[token_id] = score + toktypes[token_id] = toktype + + # Use the added_tokens_decoder field from tokeniser_config.json as the source + # of information about added/redefined tokens and modify them accordingly. + tokenizer_config_file = self.dir_model / 'tokenizer_config.json' + if tokenizer_config_file.is_file(): + with open(tokenizer_config_file, "r", encoding="utf-8") as f: + tokenizer_config_json = json.load(f) + + if "added_tokens_decoder" in tokenizer_config_json: + added_tokens_decoder = tokenizer_config_json["added_tokens_decoder"] + for token_id, token_json in added_tokens_decoder.items(): + token_id = int(token_id) + if (token_id >= vocab_size): + logger.debug(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}') + continue + + token_content = token_json["content"] + token_type = SentencePieceTokenTypes.USER_DEFINED + token_score = -10000.0 + + # Map unk_token to UNKNOWN, other special tokens to CONTROL + # Set the score to 0.0 as in the original tokenizer.model + if ("special" in token_json) and token_json["special"]: + if token_content == tokenizer_config_json["unk_token"]: + token_type = SentencePieceTokenTypes.UNKNOWN + else: + token_type = SentencePieceTokenTypes.CONTROL + token_score = 0.0 + + logger.info(f"Setting added token {token_id} to '{token_content}' (type: {token_type}, score: {token_score:.2f})") + tokens[token_id] = token_content.encode("utf-8") + toktypes[token_id] = token_type + scores[token_id] = token_score + + self.gguf_writer.add_tokenizer_model("llama") + self.gguf_writer.add_tokenizer_pre("default") + self.gguf_writer.add_token_list(tokens) + self.gguf_writer.add_token_scores(scores) + self.gguf_writer.add_token_types(toktypes) + + special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens)) + special_vocab.add_to_gguf(self.gguf_writer) + + def set_gguf_parameters(self): + super().set_gguf_parameters() + hparams = self.hparams + self.gguf_writer.add_vocab_size(hparams["vocab_size"]) + self.gguf_writer.add_rope_dimension_count(hparams["hidden_size"] // hparams["num_attention_heads"]) + + _experts: list[dict[str, Tensor]] | None = None + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + n_head = self.hparams["num_attention_heads"] + n_kv_head = self.hparams.get("num_key_value_heads") + + if name.endswith("q_proj.weight"): + data_torch = LlamaModel.permute(data_torch, n_head, n_head) + if name.endswith("k_proj.weight"): + data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head) + + # process the experts separately + if name.find("block_sparse_moe.experts") != -1: + n_experts = self.hparams["num_local_experts"] + + assert bid is not None + + if self._experts is None: + self._experts = [{} for _ in range(self.block_count)] + + self._experts[bid][name] = data_torch + + if len(self._experts[bid]) >= n_experts * 3: + tensors: list[tuple[str, Tensor]] = [] + + # merge the experts into a single 3d tensor + for wid in ["w1", "w2", "w3"]: + datas: list[Tensor] = [] + + for xid in range(n_experts): + ename = f"model.layers.{bid}.block_sparse_moe.experts.{xid}.{wid}.weight" + datas.append(self._experts[bid][ename]) + del self._experts[bid][ename] + + data_torch = torch.stack(datas, dim=0) + + merged_name = f"layers.{bid}.feed_forward.experts.{wid}.weight" + + new_name = self.map_tensor_name(merged_name) + + tensors.append((new_name, data_torch)) + return tensors + else: + return [] + + return [(self.map_tensor_name(name), data_torch)] + + def write_tensors(self): + super().write_tensors() + + if self._experts is not None: + # flatten `list[dict[str, Tensor]]` into `list[str]` + experts = [k for d in self._experts for k in d.keys()] + if len(experts) > 0: + raise ValueError(f"Unprocessed experts: {experts}") + + +@Model.register("DeepseekV2ForCausalLM") +class DeepseekV2Model(Model): + model_arch = gguf.MODEL_ARCH.DEEPSEEK2 + + def set_vocab(self): + self._set_vocab_gpt2() + + def set_gguf_parameters(self): + super().set_gguf_parameters() + hparams = self.hparams + + self.gguf_writer.add_leading_dense_block_count(hparams["first_k_dense_replace"]) + self.gguf_writer.add_vocab_size(hparams["vocab_size"]) + if "q_lora_rank" in hparams and hparams["q_lora_rank"] is not None: + self.gguf_writer.add_q_lora_rank(hparams["q_lora_rank"]) + self.gguf_writer.add_kv_lora_rank(hparams["kv_lora_rank"]) + self.gguf_writer.add_key_length(hparams["qk_nope_head_dim"] + hparams["qk_rope_head_dim"]) + self.gguf_writer.add_value_length(hparams["v_head_dim"]) + self.gguf_writer.add_expert_feed_forward_length(hparams["moe_intermediate_size"]) + self.gguf_writer.add_expert_count(hparams["n_routed_experts"]) + self.gguf_writer.add_expert_shared_count(hparams["n_shared_experts"]) + self.gguf_writer.add_expert_weights_scale(hparams["routed_scaling_factor"]) + self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"]) + + if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]: + if self.hparams["rope_scaling"].get("type") == "yarn": + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN) + self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"]) + self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["rope_scaling"]["original_max_position_embeddings"]) + self.gguf_writer.add_rope_scaling_yarn_log_mul(0.1 * hparams["rope_scaling"]["mscale_all_dim"]) + + _experts: list[dict[str, Tensor]] | None = None + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + # process the experts separately + if name.find("mlp.experts") != -1: + n_experts = self.hparams["n_routed_experts"] + assert bid is not None + + if self._experts is None: + self._experts = [{} for _ in range(self.block_count)] + + self._experts[bid][name] = data_torch + + if len(self._experts[bid]) >= n_experts * 3: + tensors: list[tuple[str, Tensor]] = [] + + # merge the experts into a single 3d tensor + for w_name in ["down_proj", "gate_proj", "up_proj"]: + datas: list[Tensor] = [] + + for xid in range(n_experts): + ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight" + datas.append(self._experts[bid][ename]) + del self._experts[bid][ename] + + data_torch = torch.stack(datas, dim=0) + + merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight" + + new_name = self.map_tensor_name(merged_name) + + tensors.append((new_name, data_torch)) + return tensors + else: + return [] + + return [(self.map_tensor_name(name), data_torch)] + + def write_tensors(self): + super().write_tensors() + + if self._experts is not None: + # flatten `list[dict[str, Tensor]]` into `list[str]` + experts = [k for d in self._experts for k in d.keys()] + if len(experts) > 0: + raise ValueError(f"Unprocessed experts: {experts}") + + ###### CONVERSION LOGIC ###### @@ -2598,7 +2840,12 @@ def main() -> None: hparams = Model.load_hparams(dir_model) with torch.inference_mode(): - model_class = Model.from_model_architecture(hparams["architectures"][0]) + try: + model_class = Model.from_model_architecture(hparams["architectures"][0]) + except NotImplementedError: + logger.error(f"Model {hparams['architectures'][0]} is not supported") + sys.exit(1) + model_instance = model_class(dir_model, ftype_map[args.outtype], fname_out, args.bigendian, args.use_temp_file, args.no_lazy) logger.info("Set model parameters") diff --git a/docs/HOWTO-add-model.md b/docs/HOWTO-add-model.md index 48769cdf6..138124248 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.py](../convert.py) or [convert-hf-to-gguf.py](../convert-hf-to-gguf.py). +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/convert.py b/examples/convert-legacy-llama.py similarity index 82% rename from convert.py rename to examples/convert-legacy-llama.py index da1247957..fd8401015 100755 --- a/convert.py +++ b/examples/convert-legacy-llama.py @@ -24,14 +24,16 @@ from abc import ABC, abstractmethod from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor from dataclasses import dataclass from pathlib import Path -from typing import TYPE_CHECKING, Any, Callable, ClassVar, IO, Iterable, Literal, Protocol, TypeVar, runtime_checkable, Optional +from typing import TYPE_CHECKING, Any, Callable, IO, Iterable, Literal, TypeVar, Optional import numpy as np -from sentencepiece import SentencePieceProcessor if 'NO_LOCAL_GGUF' not in os.environ: - sys.path.insert(1, str(Path(__file__).parent / 'gguf-py')) + # use .parent.parent since we are in "examples" directory + sys.path.insert(1, str(Path(__file__).parent.parent / 'gguf-py')) + import gguf +from gguf import BaseVocab, Vocab, NoVocab, BpeVocab, SentencePieceVocab, LlamaHfVocab if TYPE_CHECKING: from typing_extensions import Self, TypeAlias @@ -380,306 +382,6 @@ class Metadata: return metadata -# -# vocab -# - - -@runtime_checkable -class BaseVocab(Protocol): - tokenizer_model: ClassVar[str] - name: ClassVar[str] - - -class NoVocab(BaseVocab): - tokenizer_model = "no_vocab" - name = "no_vocab" - - def __repr__(self) -> str: - return "" - - -@runtime_checkable -class Vocab(BaseVocab, Protocol): - vocab_size: int - added_tokens_dict: dict[str, int] - added_tokens_list: list[str] - fname_tokenizer: Path - - def __init__(self, base_path: Path): ... - def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: ... - - -class BpeVocab(Vocab): - tokenizer_model = "gpt2" - name = "bpe" - - def __init__(self, base_path: Path): - added_tokens: dict[str, int] = {} - - if (fname_tokenizer := base_path / 'vocab.json').exists(): - # "slow" tokenizer - with open(fname_tokenizer, encoding="utf-8") as f: - self.vocab = json.load(f) - - try: - # FIXME: Verify that added tokens here _cannot_ overlap with the main vocab. - with open(base_path / ADDED_TOKENS_FILE, encoding="utf-8") as f: - added_tokens = json.load(f) - except FileNotFoundError: - pass - else: - # "fast" tokenizer - fname_tokenizer = base_path / FAST_TOKENIZER_FILE - - # if this fails, FileNotFoundError propagates to caller - with open(fname_tokenizer, encoding="utf-8") as f: - tokenizer_json = json.load(f) - - tokenizer_model: dict[str, Any] = tokenizer_json['model'] - if ( - tokenizer_model['type'] != 'BPE' or tokenizer_model.get('byte_fallback', False) - or tokenizer_json['decoder']['type'] != 'ByteLevel' - ): - raise FileNotFoundError('Cannot find GPT-2 BPE tokenizer') - - self.vocab = tokenizer_model["vocab"] - - if (added := tokenizer_json.get('added_tokens')) is not None: - # Added tokens here can be duplicates of the main vocabulary. - added_tokens = {item['content']: item['id'] - for item in added - if item['content'] not in self.vocab} - - vocab_size = len(self.vocab) - expected_ids = list(range(vocab_size, vocab_size + len(added_tokens))) - actual_ids = sorted(added_tokens.values()) - if expected_ids != actual_ids: - expected_end_id = vocab_size + len(actual_ids) - 1 - raise ValueError(f"Expected the {len(actual_ids)} added token ID(s) to be sequential in the range " - f"{vocab_size} - {expected_end_id}; got {actual_ids}") - - items = sorted(added_tokens.items(), key=lambda text_idx: text_idx[1]) - self.added_tokens_dict = added_tokens - self.added_tokens_list = [text for (text, idx) in items] - self.vocab_size_base = vocab_size - self.vocab_size = self.vocab_size_base + len(self.added_tokens_list) - self.fname_tokenizer = fname_tokenizer - - def bpe_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: - reverse_vocab = {id: encoded_tok for encoded_tok, id in self.vocab.items()} - - for i, _ in enumerate(self.vocab): - yield reverse_vocab[i], 0.0, gguf.TokenType.NORMAL - - def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: - for text in self.added_tokens_list: - score = -1000.0 - yield text.encode("utf-8"), score, gguf.TokenType.CONTROL - - def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: - yield from self.bpe_tokens() - yield from self.added_tokens() - - def __repr__(self) -> str: - return f"" - - -class SentencePieceVocab(Vocab): - tokenizer_model = "llama" - name = "spm" - - def __init__(self, base_path: Path): - added_tokens: dict[str, int] = {} - if (fname_tokenizer := base_path / 'tokenizer.model').exists(): - # normal location - try: - with open(base_path / ADDED_TOKENS_FILE, encoding="utf-8") as f: - added_tokens = json.load(f) - except FileNotFoundError: - pass - elif not (fname_tokenizer := base_path.parent / 'tokenizer.model').exists(): - # not found in alternate location either - raise FileNotFoundError('Cannot find tokenizer.model') - - self.sentencepiece_tokenizer = SentencePieceProcessor() - self.sentencepiece_tokenizer.LoadFromFile(str(fname_tokenizer)) - vocab_size = self.sentencepiece_tokenizer.vocab_size() - - new_tokens = {id: piece for piece, id in added_tokens.items() if id >= vocab_size} - expected_new_ids = list(range(vocab_size, vocab_size + len(new_tokens))) - actual_new_ids = sorted(new_tokens.keys()) - - if expected_new_ids != actual_new_ids: - raise ValueError(f"Expected new token IDs {expected_new_ids} to be sequential; got {actual_new_ids}") - - # Token pieces that were added to the base vocabulary. - self.added_tokens_dict = added_tokens - self.added_tokens_list = [new_tokens[id] for id in actual_new_ids] - self.vocab_size_base = vocab_size - self.vocab_size = self.vocab_size_base + len(self.added_tokens_list) - self.fname_tokenizer = fname_tokenizer - - def sentencepiece_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: - tokenizer = self.sentencepiece_tokenizer - for i in range(tokenizer.vocab_size()): - piece = tokenizer.IdToPiece(i) - text = piece.encode("utf-8") - score: float = tokenizer.GetScore(i) - - toktype = gguf.TokenType.NORMAL - if tokenizer.IsUnknown(i): - toktype = gguf.TokenType.UNKNOWN - if tokenizer.IsControl(i): - toktype = gguf.TokenType.CONTROL - - # NOTE: I think added_tokens are user defined. - # ref: https://github.com/google/sentencepiece/blob/master/src/sentencepiece_model.proto - # if tokenizer.is_user_defined(i): toktype = gguf.TokenType.USER_DEFINED - - if tokenizer.IsUnused(i): - toktype = gguf.TokenType.UNUSED - if tokenizer.IsByte(i): - toktype = gguf.TokenType.BYTE - - yield text, score, toktype - - def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: - for text in self.added_tokens_list: - score = -1000.0 - yield text.encode("utf-8"), score, gguf.TokenType.USER_DEFINED - - def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: - yield from self.sentencepiece_tokens() - yield from self.added_tokens() - - def __repr__(self) -> str: - return f"" - - -class LlamaHfVocab(Vocab): - tokenizer_model = "llama" - name = "hfft" - - def __init__(self, base_path: Path): - fname_tokenizer = base_path / FAST_TOKENIZER_FILE - # if this fails, FileNotFoundError propagates to caller - with open(fname_tokenizer, encoding='utf-8') as f: - tokenizer_json = json.load(f) - - # pre-check so we know if we need transformers - tokenizer_model: dict[str, Any] = tokenizer_json['model'] - is_llama3 = ( - tokenizer_model['type'] == 'BPE' and tokenizer_model.get('ignore_merges', False) - and not tokenizer_model.get('byte_fallback', True) - ) - if is_llama3: - raise TypeError('Llama 3 must be converted with BpeVocab') - - if not is_llama3 and ( - tokenizer_model['type'] != 'BPE' or not tokenizer_model.get('byte_fallback', False) - or tokenizer_json['decoder']['type'] != 'Sequence' - ): - raise FileNotFoundError('Cannot find Llama BPE tokenizer') - - try: - from transformers import AutoTokenizer - except ImportError as e: - raise ImportError( - "To use LlamaHfVocab, please install the `transformers` package. " - "You can install it with `pip install transformers`." - ) from e - - # Allow the tokenizer to default to slow or fast versions. - # Explicitly set tokenizer to use local paths. - self.tokenizer = AutoTokenizer.from_pretrained( - base_path, - cache_dir=base_path, - local_files_only=True, - ) - assert self.tokenizer.is_fast # assume tokenizer.json is used - - # Initialize lists and dictionaries for added tokens - self.added_tokens_list = [] - self.added_tokens_dict = dict() - self.added_tokens_ids = set() - - # Process added tokens - for tok, tokidx in sorted( - self.tokenizer.get_added_vocab().items(), key=lambda x: x[1] - ): - # Only consider added tokens that are not in the base vocabulary - if tokidx >= self.tokenizer.vocab_size: - self.added_tokens_list.append(tok) - self.added_tokens_dict[tok] = tokidx - self.added_tokens_ids.add(tokidx) - - # Store special tokens and their IDs - self.specials = { - tok: self.tokenizer.get_vocab()[tok] - for tok in self.tokenizer.all_special_tokens - } - self.special_ids = set(self.tokenizer.all_special_ids) - - # Set vocabulary sizes - self.vocab_size_base = self.tokenizer.vocab_size - self.vocab_size = self.vocab_size_base + len(self.added_tokens_list) - - self.fname_tokenizer = fname_tokenizer - - def hf_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: - reverse_vocab = { - id: encoded_tok for encoded_tok, id in self.tokenizer.get_vocab().items() - } - - for token_id in range(self.vocab_size_base): - # Skip processing added tokens here - if token_id in self.added_tokens_ids: - continue - - # Convert token text to bytes - token_text = reverse_vocab[token_id].encode("utf-8") - - # Yield token text, score, and type - yield token_text, self.get_token_score(token_id), self.get_token_type( - token_id, token_text, self.special_ids # Reuse already stored special IDs - ) - - def get_token_type(self, token_id: int, token_text: bytes, special_ids: set[int]) -> gguf.TokenType: - # Special case for byte tokens - if re.fullmatch(br"<0x[0-9A-Fa-f]{2}>", token_text): - return gguf.TokenType.BYTE - - # Determine token type based on whether it's a special token - return gguf.TokenType.CONTROL if token_id in special_ids else gguf.TokenType.NORMAL - - def get_token_score(self, token_id: int) -> float: - # Placeholder for actual logic to determine the token's score - # This needs to be implemented based on specific requirements - return -1000.0 # Default score - - def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: - for text in self.added_tokens_list: - if text in self.specials: - toktype = self.get_token_type(self.specials[text], b'', self.special_ids) - score = self.get_token_score(self.specials[text]) - else: - toktype = gguf.TokenType.USER_DEFINED - score = -1000.0 - - yield text.encode("utf-8"), score, toktype - - def has_newline_token(self): - return "<0x0A>" in self.tokenizer.vocab or "\n" in self.tokenizer.vocab - - def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: - yield from self.hf_tokens() - yield from self.added_tokens() - - def __repr__(self) -> str: - return f"" - - # # data loading # TODO: reuse (probably move to gguf.py?) diff --git a/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp b/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp index 746c3fbef..8ca9f8915 100644 --- a/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp +++ b/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp @@ -774,7 +774,7 @@ static struct train_params get_default_train_params() { params.samples_start_after_nl = false; params.use_adam = true; - params.use_flash = true; + params.use_flash = false; params.use_scratch = true; // only adam diff --git a/examples/llama-bench/llama-bench.cpp b/examples/llama-bench/llama-bench.cpp index 82a001a0a..460dc5e75 100644 --- a/examples/llama-bench/llama-bench.cpp +++ b/examples/llama-bench/llama-bench.cpp @@ -178,6 +178,7 @@ struct cmd_params { std::vector type_v; std::vector n_threads; std::vector n_gpu_layers; + std::vector rpc_servers; std::vector split_mode; std::vector main_gpu; std::vector no_kv_offload; @@ -212,6 +213,7 @@ static const cmd_params cmd_params_defaults = { /* type_v */ {GGML_TYPE_F16}, /* n_threads */ {cpu_get_num_math()}, /* n_gpu_layers */ {99}, + /* rpc_servers */ {""}, /* split_mode */ {LLAMA_SPLIT_MODE_LAYER}, /* main_gpu */ {0}, /* no_kv_offload */ {false}, @@ -241,6 +243,7 @@ static void print_usage(int /* argc */, char ** argv) { printf(" -ctv, --cache-type-v (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str()); printf(" -t, --threads (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str()); printf(" -ngl, --n-gpu-layers (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str()); + printf(" -rpc, --rpc (default: %s)\n", join(cmd_params_defaults.rpc_servers, ",").c_str()); printf(" -sm, --split-mode (default: %s)\n", join(transform_to_str(cmd_params_defaults.split_mode, split_mode_str), ",").c_str()); printf(" -mg, --main-gpu (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str()); printf(" -nkvo, --no-kv-offload <0|1> (default: %s)\n", join(cmd_params_defaults.no_kv_offload, ",").c_str()); @@ -308,32 +311,28 @@ static cmd_params parse_cmd_params(int argc, char** argv) { if (arg == "-h" || arg == "--help") { print_usage(argc, argv); exit(0); - } - else if (arg == "-m" || arg == "--model") { + } else if (arg == "-m" || arg == "--model") { if (++i >= argc) { invalid_param = true; break; } auto p = split(argv[i], split_delim); params.model.insert(params.model.end(), p.begin(), p.end()); - } - else if (arg == "-p" || arg == "--n-prompt") { + } else if (arg == "-p" || arg == "--n-prompt") { if (++i >= argc) { invalid_param = true; break; } auto p = split(argv[i], split_delim); params.n_prompt.insert(params.n_prompt.end(), p.begin(), p.end()); - } - else if (arg == "-n" || arg == "--n-gen") { + } else if (arg == "-n" || arg == "--n-gen") { if (++i >= argc) { invalid_param = true; break; } auto p = split(argv[i], split_delim); params.n_gen.insert(params.n_gen.end(), p.begin(), p.end()); - } - else if (arg == "-pg") { + } else if (arg == "-pg") { if (++i >= argc) { invalid_param = true; break; @@ -344,24 +343,21 @@ static cmd_params parse_cmd_params(int argc, char** argv) { break; } params.n_pg.push_back({ std::stoi(p[0]), std::stoi(p[1]) }); - } - else if (arg == "-b" || arg == "--batch-size") { + } else if (arg == "-b" || arg == "--batch-size") { if (++i >= argc) { invalid_param = true; break; } auto p = split(argv[i], split_delim); params.n_batch.insert(params.n_batch.end(), p.begin(), p.end()); - } - else if (arg == "-ub" || arg == "--ubatch-size") { + } else if (arg == "-ub" || arg == "--ubatch-size") { if (++i >= argc) { invalid_param = true; break; } auto p = split(argv[i], split_delim); params.n_ubatch.insert(params.n_ubatch.end(), p.begin(), p.end()); - } - else if (arg == "-ctk" || arg == "--cache-type-k") { + } else if (arg == "-ctk" || arg == "--cache-type-k") { if (++i >= argc) { invalid_param = true; break; @@ -377,8 +373,7 @@ static cmd_params parse_cmd_params(int argc, char** argv) { types.push_back(gt); } params.type_k.insert(params.type_k.end(), types.begin(), types.end()); - } - else if (arg == "-ctv" || arg == "--cache-type-v") { + } else if (arg == "-ctv" || arg == "--cache-type-v") { if (++i >= argc) { invalid_param = true; break; @@ -394,24 +389,27 @@ static cmd_params parse_cmd_params(int argc, char** argv) { types.push_back(gt); } params.type_v.insert(params.type_v.end(), types.begin(), types.end()); - } - else if (arg == "-t" || arg == "--threads") { + } else if (arg == "-t" || arg == "--threads") { if (++i >= argc) { invalid_param = true; break; } auto p = split(argv[i], split_delim); params.n_threads.insert(params.n_threads.end(), p.begin(), p.end()); - } - else if (arg == "-ngl" || arg == "--n-gpu-layers") { + } else if (arg == "-ngl" || arg == "--n-gpu-layers") { if (++i >= argc) { invalid_param = true; break; } auto p = split(argv[i], split_delim); params.n_gpu_layers.insert(params.n_gpu_layers.end(), p.begin(), p.end()); - } - else if (arg == "-sm" || arg == "--split-mode") { + } else if (arg == "-rpc" || arg == "--rpc") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.rpc_servers.push_back(argv[i]); + } else if (arg == "-sm" || arg == "--split-mode") { if (++i >= argc) { invalid_param = true; break; @@ -422,42 +420,35 @@ static cmd_params parse_cmd_params(int argc, char** argv) { llama_split_mode mode; if (m == "none") { mode = LLAMA_SPLIT_MODE_NONE; - } - else if (m == "layer") { + } else if (m == "layer") { mode = LLAMA_SPLIT_MODE_LAYER; - } - else if (m == "row") { + } else if (m == "row") { mode = LLAMA_SPLIT_MODE_ROW; - } - else { + } else { invalid_param = true; break; } modes.push_back(mode); } params.split_mode.insert(params.split_mode.end(), modes.begin(), modes.end()); - } - else if (arg == "-mg" || arg == "--main-gpu") { + } else if (arg == "-mg" || arg == "--main-gpu") { if (++i >= argc) { invalid_param = true; break; } params.main_gpu = split(argv[i], split_delim); - } - else if (arg == "-nkvo" || arg == "--no-kv-offload") { + } else if (arg == "-nkvo" || arg == "--no-kv-offload") { if (++i >= argc) { invalid_param = true; break; } auto p = split(argv[i], split_delim); params.no_kv_offload.insert(params.no_kv_offload.end(), p.begin(), p.end()); - } - else if (arg == "--numa") { + } else if (arg == "--numa") { if (++i >= argc) { invalid_param = true; break; - } - else { + } else { std::string value(argv[i]); /**/ if (value == "distribute" || value == "") { params.numa = GGML_NUMA_STRATEGY_DISTRIBUTE; } else if (value == "isolate") { params.numa = GGML_NUMA_STRATEGY_ISOLATE; } @@ -465,15 +456,13 @@ static cmd_params parse_cmd_params(int argc, char** argv) { else { invalid_param = true; break; } } - } - else if (arg == "-mt" || arg == "--max-threads") { + } else if (arg == "-mt" || arg == "--max-threads") { if (++i >= argc) { invalid_param = true; break; } params.cpuparams.n_threads = std::stoi(argv[i]); - } - else if (arg == "-C" || arg == "--cpu-mask") { + } else if (arg == "-C" || arg == "--cpu-mask") { if (++i >= argc) { invalid_param = true; break; @@ -481,8 +470,7 @@ static cmd_params parse_cmd_params(int argc, char** argv) { std::string mask = argv[i]; params.cpuparams.mask_valid = true; invalid_param = !parse_cpu_mask(mask, params.cpuparams.cpumask); - } - else if (arg == "--prio") { + } else if (arg == "--prio") { if (++i >= argc) { invalid_param = true; break; @@ -581,6 +569,7 @@ static cmd_params parse_cmd_params(int argc, char** argv) { if (params.type_k.empty()) { params.type_k = cmd_params_defaults.type_k; } if (params.type_v.empty()) { params.type_v = cmd_params_defaults.type_v; } if (params.n_gpu_layers.empty()) { params.n_gpu_layers = cmd_params_defaults.n_gpu_layers; } + if (params.rpc_servers.empty()) { params.rpc_servers = cmd_params_defaults.rpc_servers; } if (params.split_mode.empty()) { params.split_mode = cmd_params_defaults.split_mode; } if (params.main_gpu.empty()) { params.main_gpu = cmd_params_defaults.main_gpu; } if (params.no_kv_offload.empty()){ params.no_kv_offload = cmd_params_defaults.no_kv_offload; } @@ -603,6 +592,7 @@ struct cmd_params_instance { ggml_type type_v; int n_threads; int n_gpu_layers; + std::string rpc_servers; llama_split_mode split_mode; int main_gpu; bool no_kv_offload; @@ -615,6 +605,9 @@ struct cmd_params_instance { llama_model_params mparams = llama_model_default_params(); mparams.n_gpu_layers = n_gpu_layers; + if (!rpc_servers.empty()) { + mparams.rpc_servers = rpc_servers.c_str(); + } mparams.split_mode = split_mode; mparams.main_gpu = main_gpu; mparams.tensor_split = tensor_split.data(); @@ -626,6 +619,7 @@ struct cmd_params_instance { bool equal_mparams(const cmd_params_instance & other) const { return model == other.model && n_gpu_layers == other.n_gpu_layers && + rpc_servers == other.rpc_servers && split_mode == other.split_mode && main_gpu == other.main_gpu && use_mmap == other.use_mmap && @@ -654,6 +648,7 @@ static std::vector get_cmd_params_instances(const cmd_param // this ordering minimizes the number of times that each model needs to be reloaded for (const auto & m : params.model) for (const auto & nl : params.n_gpu_layers) + for (const auto & rpc : params.rpc_servers) for (const auto & sm : params.split_mode) for (const auto & mg : params.main_gpu) for (const auto & ts : params.tensor_split) @@ -680,6 +675,7 @@ static std::vector get_cmd_params_instances(const cmd_param /* .type_v = */ tv, /* .n_threads = */ nt, /* .n_gpu_layers = */ nl, + /* .rpc_servers = */ rpc, /* .split_mode = */ sm, /* .main_gpu = */ mg, /* .no_kv_offload= */ nkvo, @@ -705,6 +701,7 @@ static std::vector get_cmd_params_instances(const cmd_param /* .type_v = */ tv, /* .n_threads = */ nt, /* .n_gpu_layers = */ nl, + /* .rpc_servers = */ rpc, /* .split_mode = */ sm, /* .main_gpu = */ mg, /* .no_kv_offload= */ nkvo, @@ -730,6 +727,7 @@ static std::vector get_cmd_params_instances(const cmd_param /* .type_v = */ tv, /* .n_threads = */ nt, /* .n_gpu_layers = */ nl, + /* .rpc_servers = */ rpc, /* .split_mode = */ sm, /* .main_gpu = */ mg, /* .no_kv_offload= */ nkvo, @@ -754,6 +752,7 @@ struct test { static const bool kompute; static const bool metal; static const bool sycl; + static const bool rpc; static const bool gpu_blas; static const bool blas; static const std::string cpu_info; @@ -852,6 +851,9 @@ struct test { if (sycl) { return GGML_SYCL_NAME; } + if (rpc) { + return "RPC"; + } if (gpu_blas) { return "GPU BLAS"; } @@ -865,7 +867,7 @@ struct test { static const std::vector & get_fields() { static const std::vector fields = { "build_commit", "build_number", - "cuda", "opencl", "vulkan", "kompute", "metal", "sycl", "gpu_blas", "blas", + "cuda", "opencl", "vulkan", "kompute", "metal", "sycl", "rpc", "gpu_blas", "blas", "cpu_info", "gpu_info", "model_filename", "model_type", "model_size", "model_n_params", "n_batch", "n_ubatch", @@ -921,7 +923,7 @@ struct test { std::vector values = { build_commit, std::to_string(build_number), std::to_string(cuda), std::to_string(opencl), std::to_string(vulkan), std::to_string(vulkan), - std::to_string(metal), std::to_string(sycl), std::to_string(gpu_blas), std::to_string(blas), + std::to_string(metal), std::to_string(sycl), std::to_string(rpc), std::to_string(gpu_blas), std::to_string(blas), cpu_info, gpu_info, model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params), std::to_string(n_batch), std::to_string(n_ubatch), @@ -956,6 +958,7 @@ const bool test::metal = !!ggml_cpu_has_metal(); const bool test::gpu_blas = !!ggml_cpu_has_gpublas(); const bool test::blas = !!ggml_cpu_has_blas(); const bool test::sycl = !!ggml_cpu_has_sycl(); +const bool test::rpc = !!ggml_cpu_has_rpc(); const std::string test::cpu_info = get_cpu_info(); const std::string test::gpu_info = get_gpu_info(); diff --git a/examples/llama.android/app/build.gradle.kts b/examples/llama.android/app/build.gradle.kts index d42140efe..8d1b37195 100644 --- a/examples/llama.android/app/build.gradle.kts +++ b/examples/llama.android/app/build.gradle.kts @@ -7,8 +7,6 @@ android { namespace = "com.example.llama" compileSdk = 34 - ndkVersion = "26.1.10909125" - defaultConfig { applicationId = "com.example.llama" minSdk = 33 @@ -20,17 +18,6 @@ android { vectorDrawables { useSupportLibrary = true } - ndk { - // Add NDK properties if wanted, e.g. - // abiFilters += listOf("arm64-v8a") - } - externalNativeBuild { - cmake { - arguments += "-DCMAKE_BUILD_TYPE=Release" - cppFlags += listOf() - arguments += listOf() - } - } } buildTypes { @@ -55,17 +42,6 @@ android { composeOptions { kotlinCompilerExtensionVersion = "1.5.1" } - packaging { - resources { - excludes += "/META-INF/{AL2.0,LGPL2.1}" - } - } - externalNativeBuild { - cmake { - path = file("src/main/cpp/CMakeLists.txt") - version = "3.22.1" - } - } } dependencies { @@ -78,6 +54,7 @@ dependencies { implementation("androidx.compose.ui:ui-graphics") implementation("androidx.compose.ui:ui-tooling-preview") implementation("androidx.compose.material3:material3") + implementation(project(":llama")) testImplementation("junit:junit:4.13.2") androidTestImplementation("androidx.test.ext:junit:1.1.5") androidTestImplementation("androidx.test.espresso:espresso-core:3.5.1") diff --git a/examples/llama.android/app/src/main/java/com/example/llama/MainViewModel.kt b/examples/llama.android/app/src/main/java/com/example/llama/MainViewModel.kt index be95e2221..45ac29938 100644 --- a/examples/llama.android/app/src/main/java/com/example/llama/MainViewModel.kt +++ b/examples/llama.android/app/src/main/java/com/example/llama/MainViewModel.kt @@ -1,5 +1,6 @@ package com.example.llama +import android.llama.cpp.LLamaAndroid import android.util.Log import androidx.compose.runtime.getValue import androidx.compose.runtime.mutableStateOf @@ -9,7 +10,7 @@ import androidx.lifecycle.viewModelScope import kotlinx.coroutines.flow.catch import kotlinx.coroutines.launch -class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() { +class MainViewModel(private val llamaAndroid: LLamaAndroid = LLamaAndroid.instance()): ViewModel() { companion object { @JvmStatic private val NanosPerSecond = 1_000_000_000.0 @@ -28,7 +29,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() { viewModelScope.launch { try { - llm.unload() + llamaAndroid.unload() } catch (exc: IllegalStateException) { messages += exc.message!! } @@ -44,7 +45,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() { messages += "" viewModelScope.launch { - llm.send(text) + llamaAndroid.send(text) .catch { Log.e(tag, "send() failed", it) messages += it.message!! @@ -57,7 +58,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() { viewModelScope.launch { try { val start = System.nanoTime() - val warmupResult = llm.bench(pp, tg, pl, nr) + val warmupResult = llamaAndroid.bench(pp, tg, pl, nr) val end = System.nanoTime() messages += warmupResult @@ -70,7 +71,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() { return@launch } - messages += llm.bench(512, 128, 1, 3) + messages += llamaAndroid.bench(512, 128, 1, 3) } catch (exc: IllegalStateException) { Log.e(tag, "bench() failed", exc) messages += exc.message!! @@ -81,7 +82,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() { fun load(pathToModel: String) { viewModelScope.launch { try { - llm.load(pathToModel) + llamaAndroid.load(pathToModel) messages += "Loaded $pathToModel" } catch (exc: IllegalStateException) { Log.e(tag, "load() failed", exc) diff --git a/examples/llama.android/build.gradle.kts b/examples/llama.android/build.gradle.kts index 50ebc8211..acd1ada7d 100644 --- a/examples/llama.android/build.gradle.kts +++ b/examples/llama.android/build.gradle.kts @@ -2,4 +2,5 @@ plugins { id("com.android.application") version "8.2.0" apply false id("org.jetbrains.kotlin.android") version "1.9.0" apply false + id("com.android.library") version "8.2.0" apply false } diff --git a/examples/llama.android/llama/.gitignore b/examples/llama.android/llama/.gitignore new file mode 100644 index 000000000..796b96d1c --- /dev/null +++ b/examples/llama.android/llama/.gitignore @@ -0,0 +1 @@ +/build diff --git a/examples/llama.android/app/src/main/cpp/CMakeLists.txt b/examples/llama.android/llama/CMakeLists.txt similarity index 98% rename from examples/llama.android/app/src/main/cpp/CMakeLists.txt rename to examples/llama.android/llama/CMakeLists.txt index 4536974a5..a5618cac0 100644 --- a/examples/llama.android/app/src/main/cpp/CMakeLists.txt +++ b/examples/llama.android/llama/CMakeLists.txt @@ -42,7 +42,7 @@ add_subdirectory(../../../../../../ build-llama) # used in the AndroidManifest.xml file. add_library(${CMAKE_PROJECT_NAME} SHARED # List C/C++ source files with relative paths to this CMakeLists.txt. - llama-android.cpp) + llama-android.cpp) # Specifies libraries CMake should link to your target library. You # can link libraries from various origins, such as libraries defined in this diff --git a/examples/llama.android/llama/build.gradle.kts b/examples/llama.android/llama/build.gradle.kts new file mode 100644 index 000000000..0a3806172 --- /dev/null +++ b/examples/llama.android/llama/build.gradle.kts @@ -0,0 +1,68 @@ +plugins { + id("com.android.library") + id("org.jetbrains.kotlin.android") +} + +android { + namespace = "android.llama.cpp" + compileSdk = 34 + + defaultConfig { + minSdk = 33 + + testInstrumentationRunner = "androidx.test.runner.AndroidJUnitRunner" + consumerProguardFiles("consumer-rules.pro") + ndk { + // Add NDK properties if wanted, e.g. + // abiFilters += listOf("arm64-v8a") + } + externalNativeBuild { + cmake { + arguments += "-DCMAKE_BUILD_TYPE=Release" + cppFlags += listOf() + arguments += listOf() + + cppFlags("") + } + } + } + + buildTypes { + release { + isMinifyEnabled = false + proguardFiles( + getDefaultProguardFile("proguard-android-optimize.txt"), + "proguard-rules.pro" + ) + } + } + externalNativeBuild { + cmake { + path("src/main/cpp/CMakeLists.txt") + version = "3.22.1" + } + } + compileOptions { + sourceCompatibility = JavaVersion.VERSION_1_8 + targetCompatibility = JavaVersion.VERSION_1_8 + } + kotlinOptions { + jvmTarget = "1.8" + } + + packaging { + resources { + excludes += "/META-INF/{AL2.0,LGPL2.1}" + } + } +} + +dependencies { + + implementation("androidx.core:core-ktx:1.12.0") + implementation("androidx.appcompat:appcompat:1.6.1") + implementation("com.google.android.material:material:1.11.0") + testImplementation("junit:junit:4.13.2") + androidTestImplementation("androidx.test.ext:junit:1.1.5") + androidTestImplementation("androidx.test.espresso:espresso-core:3.5.1") +} diff --git a/examples/llama.android/llama/consumer-rules.pro b/examples/llama.android/llama/consumer-rules.pro new file mode 100644 index 000000000..e69de29bb diff --git a/examples/llama.android/llama/proguard-rules.pro b/examples/llama.android/llama/proguard-rules.pro new file mode 100644 index 000000000..f1b424510 --- /dev/null +++ b/examples/llama.android/llama/proguard-rules.pro @@ -0,0 +1,21 @@ +# Add project specific ProGuard rules here. +# You can control the set of applied configuration files using the +# proguardFiles setting in build.gradle. +# +# For more details, see +# http://developer.android.com/guide/developing/tools/proguard.html + +# If your project uses WebView with JS, uncomment the following +# and specify the fully qualified class name to the JavaScript interface +# class: +#-keepclassmembers class fqcn.of.javascript.interface.for.webview { +# public *; +#} + +# Uncomment this to preserve the line number information for +# debugging stack traces. +#-keepattributes SourceFile,LineNumberTable + +# If you keep the line number information, uncomment this to +# hide the original source file name. +#-renamesourcefileattribute SourceFile diff --git a/examples/llama.android/llama/src/androidTest/java/android/llama/cpp/ExampleInstrumentedTest.kt b/examples/llama.android/llama/src/androidTest/java/android/llama/cpp/ExampleInstrumentedTest.kt new file mode 100644 index 000000000..05d6ab5d2 --- /dev/null +++ b/examples/llama.android/llama/src/androidTest/java/android/llama/cpp/ExampleInstrumentedTest.kt @@ -0,0 +1,24 @@ +package android.llama.cpp + +import androidx.test.platform.app.InstrumentationRegistry +import androidx.test.ext.junit.runners.AndroidJUnit4 + +import org.junit.Test +import org.junit.runner.RunWith + +import org.junit.Assert.* + +/** + * Instrumented test, which will execute on an Android device. + * + * See [testing documentation](http://d.android.com/tools/testing). + */ +@RunWith(AndroidJUnit4::class) +class ExampleInstrumentedTest { + @Test + fun useAppContext() { + // Context of the app under test. + val appContext = InstrumentationRegistry.getInstrumentation().targetContext + assertEquals("android.llama.cpp.test", appContext.packageName) + } +} diff --git a/examples/llama.android/llama/src/main/AndroidManifest.xml b/examples/llama.android/llama/src/main/AndroidManifest.xml new file mode 100644 index 000000000..8bdb7e14b --- /dev/null +++ b/examples/llama.android/llama/src/main/AndroidManifest.xml @@ -0,0 +1,4 @@ + + + + diff --git a/examples/llama.android/llama/src/main/cpp/CMakeLists.txt b/examples/llama.android/llama/src/main/cpp/CMakeLists.txt new file mode 100644 index 000000000..42ebaad49 --- /dev/null +++ b/examples/llama.android/llama/src/main/cpp/CMakeLists.txt @@ -0,0 +1,49 @@ +# For more information about using CMake with Android Studio, read the +# documentation: https://d.android.com/studio/projects/add-native-code.html. +# For more examples on how to use CMake, see https://github.com/android/ndk-samples. + +# Sets the minimum CMake version required for this project. +cmake_minimum_required(VERSION 3.22.1) + +# Declares the project name. The project name can be accessed via ${ PROJECT_NAME}, +# Since this is the top level CMakeLists.txt, the project name is also accessible +# with ${CMAKE_PROJECT_NAME} (both CMake variables are in-sync within the top level +# build script scope). +project("llama-android") + +include(FetchContent) +FetchContent_Declare( + llama + GIT_REPOSITORY https://github.com/ggerganov/llama.cpp + GIT_TAG master +) + +# Also provides "common" +FetchContent_MakeAvailable(llama) + +# Creates and names a library, sets it as either STATIC +# or SHARED, and provides the relative paths to its source code. +# You can define multiple libraries, and CMake builds them for you. +# Gradle automatically packages shared libraries with your APK. +# +# In this top level CMakeLists.txt, ${CMAKE_PROJECT_NAME} is used to define +# the target library name; in the sub-module's CMakeLists.txt, ${PROJECT_NAME} +# is preferred for the same purpose. +# +# In order to load a library into your app from Java/Kotlin, you must call +# System.loadLibrary() and pass the name of the library defined here; +# for GameActivity/NativeActivity derived applications, the same library name must be +# used in the AndroidManifest.xml file. +add_library(${CMAKE_PROJECT_NAME} SHARED + # List C/C++ source files with relative paths to this CMakeLists.txt. + llama-android.cpp) + +# Specifies libraries CMake should link to your target library. You +# can link libraries from various origins, such as libraries defined in this +# build script, prebuilt third-party libraries, or Android system libraries. +target_link_libraries(${CMAKE_PROJECT_NAME} + # List libraries link to the target library + llama + common + android + log) diff --git a/examples/llama.android/app/src/main/cpp/llama-android.cpp b/examples/llama.android/llama/src/main/cpp/llama-android.cpp similarity index 92% rename from examples/llama.android/app/src/main/cpp/llama-android.cpp rename to examples/llama.android/llama/src/main/cpp/llama-android.cpp index 4af9de303..874158ef0 100644 --- a/examples/llama.android/app/src/main/cpp/llama-android.cpp +++ b/examples/llama.android/llama/src/main/cpp/llama-android.cpp @@ -81,7 +81,7 @@ static void log_callback(ggml_log_level level, const char * fmt, void * data) { extern "C" JNIEXPORT jlong JNICALL -Java_com_example_llama_Llm_load_1model(JNIEnv *env, jobject, jstring filename) { +Java_android_llama_cpp_LLamaAndroid_load_1model(JNIEnv *env, jobject, jstring filename) { llama_model_params model_params = llama_model_default_params(); auto path_to_model = env->GetStringUTFChars(filename, 0); @@ -101,13 +101,13 @@ Java_com_example_llama_Llm_load_1model(JNIEnv *env, jobject, jstring filename) { extern "C" JNIEXPORT void JNICALL -Java_com_example_llama_Llm_free_1model(JNIEnv *, jobject, jlong model) { +Java_android_llama_cpp_LLamaAndroid_free_1model(JNIEnv *, jobject, jlong model) { llama_free_model(reinterpret_cast(model)); } extern "C" JNIEXPORT jlong JNICALL -Java_com_example_llama_Llm_new_1context(JNIEnv *env, jobject, jlong jmodel) { +Java_android_llama_cpp_LLamaAndroid_new_1context(JNIEnv *env, jobject, jlong jmodel) { auto model = reinterpret_cast(jmodel); if (!model) { @@ -139,25 +139,25 @@ Java_com_example_llama_Llm_new_1context(JNIEnv *env, jobject, jlong jmodel) { extern "C" JNIEXPORT void JNICALL -Java_com_example_llama_Llm_free_1context(JNIEnv *, jobject, jlong context) { +Java_android_llama_cpp_LLamaAndroid_free_1context(JNIEnv *, jobject, jlong context) { llama_free(reinterpret_cast(context)); } extern "C" JNIEXPORT void JNICALL -Java_com_example_llama_Llm_backend_1free(JNIEnv *, jobject) { +Java_android_llama_cpp_LLamaAndroid_backend_1free(JNIEnv *, jobject) { llama_backend_free(); } extern "C" JNIEXPORT void JNICALL -Java_com_example_llama_Llm_log_1to_1android(JNIEnv *, jobject) { +Java_android_llama_cpp_LLamaAndroid_log_1to_1android(JNIEnv *, jobject) { llama_log_set(log_callback, NULL); } extern "C" JNIEXPORT jstring JNICALL -Java_com_example_llama_Llm_bench_1model( +Java_android_llama_cpp_LLamaAndroid_bench_1model( JNIEnv *env, jobject, jlong context_pointer, @@ -271,13 +271,13 @@ Java_com_example_llama_Llm_bench_1model( extern "C" JNIEXPORT void JNICALL -Java_com_example_llama_Llm_free_1batch(JNIEnv *, jobject, jlong batch_pointer) { +Java_android_llama_cpp_LLamaAndroid_free_1batch(JNIEnv *, jobject, jlong batch_pointer) { llama_batch_free(*reinterpret_cast(batch_pointer)); } extern "C" JNIEXPORT jlong JNICALL -Java_com_example_llama_Llm_new_1batch(JNIEnv *, jobject, jint n_tokens, jint embd, jint n_seq_max) { +Java_android_llama_cpp_LLamaAndroid_new_1batch(JNIEnv *, jobject, jint n_tokens, jint embd, jint n_seq_max) { // Source: Copy of llama.cpp:llama_batch_init but heap-allocated. @@ -313,19 +313,19 @@ Java_com_example_llama_Llm_new_1batch(JNIEnv *, jobject, jint n_tokens, jint emb extern "C" JNIEXPORT void JNICALL -Java_com_example_llama_Llm_backend_1init(JNIEnv *, jobject) { +Java_android_llama_cpp_LLamaAndroid_backend_1init(JNIEnv *, jobject) { llama_backend_init(); } extern "C" JNIEXPORT jstring JNICALL -Java_com_example_llama_Llm_system_1info(JNIEnv *env, jobject) { +Java_android_llama_cpp_LLamaAndroid_system_1info(JNIEnv *env, jobject) { return env->NewStringUTF(llama_print_system_info()); } extern "C" JNIEXPORT jint JNICALL -Java_com_example_llama_Llm_completion_1init( +Java_android_llama_cpp_LLamaAndroid_completion_1init( JNIEnv *env, jobject, jlong context_pointer, @@ -376,7 +376,7 @@ Java_com_example_llama_Llm_completion_1init( extern "C" JNIEXPORT jstring JNICALL -Java_com_example_llama_Llm_completion_1loop( +Java_android_llama_cpp_LLamaAndroid_completion_1loop( JNIEnv * env, jobject, jlong context_pointer, @@ -438,6 +438,6 @@ Java_com_example_llama_Llm_completion_1loop( extern "C" JNIEXPORT void JNICALL -Java_com_example_llama_Llm_kv_1cache_1clear(JNIEnv *, jobject, jlong context) { +Java_android_llama_cpp_LLamaAndroid_kv_1cache_1clear(JNIEnv *, jobject, jlong context) { llama_kv_cache_clear(reinterpret_cast(context)); } diff --git a/examples/llama.android/app/src/main/java/com/example/llama/Llm.kt b/examples/llama.android/llama/src/main/java/android/llama/cpp/LLamaAndroid.kt similarity index 97% rename from examples/llama.android/app/src/main/java/com/example/llama/Llm.kt rename to examples/llama.android/llama/src/main/java/android/llama/cpp/LLamaAndroid.kt index d86afee37..6c63e54e0 100644 --- a/examples/llama.android/app/src/main/java/com/example/llama/Llm.kt +++ b/examples/llama.android/llama/src/main/java/android/llama/cpp/LLamaAndroid.kt @@ -1,4 +1,4 @@ -package com.example.llama +package android.llama.cpp import android.util.Log import kotlinx.coroutines.CoroutineDispatcher @@ -10,7 +10,7 @@ import kotlinx.coroutines.withContext import java.util.concurrent.Executors import kotlin.concurrent.thread -class Llm { +class LLamaAndroid { private val tag: String? = this::class.simpleName private val threadLocalState: ThreadLocal = ThreadLocal.withInitial { State.Idle } @@ -165,8 +165,8 @@ class Llm { } // Enforce only one instance of Llm. - private val _instance: Llm = Llm() + private val _instance: LLamaAndroid = LLamaAndroid() - fun instance(): Llm = _instance + fun instance(): LLamaAndroid = _instance } } diff --git a/examples/llama.android/llama/src/test/java/android/llama/cpp/ExampleUnitTest.kt b/examples/llama.android/llama/src/test/java/android/llama/cpp/ExampleUnitTest.kt new file mode 100644 index 000000000..cbbb974d3 --- /dev/null +++ b/examples/llama.android/llama/src/test/java/android/llama/cpp/ExampleUnitTest.kt @@ -0,0 +1,17 @@ +package android.llama.cpp + +import org.junit.Test + +import org.junit.Assert.* + +/** + * Example local unit test, which will execute on the development machine (host). + * + * See [testing documentation](http://d.android.com/tools/testing). + */ +class ExampleUnitTest { + @Test + fun addition_isCorrect() { + assertEquals(4, 2 + 2) + } +} diff --git a/examples/llama.android/settings.gradle.kts b/examples/llama.android/settings.gradle.kts index 2ba32c4fa..c7c1a034a 100644 --- a/examples/llama.android/settings.gradle.kts +++ b/examples/llama.android/settings.gradle.kts @@ -15,3 +15,4 @@ dependencyResolutionManagement { rootProject.name = "LlamaAndroid" include(":app") +include(":llama") diff --git a/examples/llava/MobileVLM-README.md b/examples/llava/MobileVLM-README.md index 413e433dd..74f021dec 100644 --- a/examples/llava/MobileVLM-README.md +++ b/examples/llava/MobileVLM-README.md @@ -54,10 +54,10 @@ python ./examples/llava/convert-image-encoder-to-gguf \ --projector-type ldpv2 ``` -4. Use `convert.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 ./convert.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 4fb0cf381..8d1ae5270 100644 --- a/examples/llava/README.md +++ b/examples/llava/README.md @@ -50,10 +50,10 @@ python ./examples/llava/llava-surgery.py -m ../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 `convert.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 ./convert.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. @@ -92,7 +92,7 @@ python ./examples/llava/convert-image-encoder-to-gguf.py -m vit --llava-projecto 6) Then convert the model to gguf format: ```console -python ./convert.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/clip.h b/examples/llava/clip.h index 45bdad689..ca3631384 100644 --- a/examples/llava/clip.h +++ b/examples/llava/clip.h @@ -68,7 +68,7 @@ CLIP_API bool clip_image_load_from_file(const char * fname, struct clip_image_u8 /** interpret bytes as an image file with length bytes_length, and use the result to populate img */ CLIP_API bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length, struct clip_image_u8 * img); -/** preprocess img and store the result in res_imgs, pad_to_square may be overriden to false depending on model configuration */ +/** preprocess img and store the result in res_imgs, pad_to_square may be overridden to false depending on model configuration */ CLIP_API bool clip_image_preprocess(struct clip_ctx * ctx, const struct clip_image_u8 * img, struct clip_image_f32_batch * res_imgs ); CLIP_API struct ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx); diff --git a/examples/llava/requirements.txt b/examples/llava/requirements.txt index f80f727a7..17cb4d5e5 100644 --- a/examples/llava/requirements.txt +++ b/examples/llava/requirements.txt @@ -1,3 +1,3 @@ --r ../../requirements/requirements-convert.txt +-r ../../requirements/requirements-convert-legacy-llama.txt pillow~=10.2.0 torch~=2.1.1 diff --git a/examples/main/main.cpp b/examples/main/main.cpp index 42a94a8c3..23415766b 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -767,18 +767,26 @@ int main(int argc, char ** argv) { // display text if (input_echo && display) { for (auto id : embd) { - const std::string token_str = llama_token_to_piece(ctx, id, !params.conversation); - printf("%s", token_str.c_str()); + const std::string token_str = llama_token_to_piece(ctx, id, params.special); + // Console/Stream Output + fprintf(stdout, "%s", token_str.c_str()); + + // Record Displayed Tokens To Log + // Note: Generated tokens are created one by one hence this check if (embd.size() > 1) { + // Incoming Requested Tokens input_tokens.push_back(id); } else { + // Outgoing Generated Tokens output_tokens.push_back(id); output_ss << token_str; } + + fflush(stdout); } - fflush(stdout); } + // reset color to default if there is no pending user input if (input_echo && (int) embd_inp.size() == n_consumed) { console::set_display(console::reset); diff --git a/examples/make-ggml.py b/examples/make-ggml.py deleted file mode 100755 index c73485ebf..000000000 --- a/examples/make-ggml.py +++ /dev/null @@ -1,98 +0,0 @@ -#!/usr/bin/env python3 -""" -This script converts Hugging Face Llama, StarCoder, Falcon, Baichuan, and GPT-NeoX models to GGUF and quantizes them. - -Usage: -python make-ggml.py {model_dir_or_hf_repo_name} --model_type {model_type} [--outname {output_name} (Optional)] [--outdir {output_directory} (Optional)] [--quants {quant_types} (Optional)] [--keep_fp16 (Optional)] - -Arguments: -- model: (Required) The directory of the downloaded Hugging Face model or the name of the Hugging Face model repository. If the model directory does not exist, it will be downloaded from the Hugging Face model hub. -- --model_type: (Required) The type of the model to be converted. Choose from llama, starcoder, falcon, baichuan, or gptneox. -- --outname: (Optional) The name of the output model. If not specified, the last part of the model directory path or the Hugging Face model repo name will be used. -- --outdir: (Optional) The directory where the output model(s) will be stored. If not specified, '../models/{outname}' will be used. -- --quants: (Optional) The types of quantization to apply. This should be a space-separated list. The default is 'Q4_K_M Q5_K_S'. -- --keep_fp16: (Optional) If specified, the FP16 model will not be deleted after the quantized models are created. - -Old quant types (some base model types require these): -- Q4_0: small, very high quality loss - legacy, prefer using Q3_K_M -- Q4_1: small, substantial quality loss - legacy, prefer using Q3_K_L -- Q5_0: medium, balanced quality - legacy, prefer using Q4_K_M -- Q5_1: medium, low quality loss - legacy, prefer using Q5_K_M - -New quant types (recommended): -- Q2_K: smallest, extreme quality loss - not recommended -- Q3_K: alias for Q3_K_M -- Q3_K_S: very small, very high quality loss -- Q3_K_M: very small, very high quality loss -- Q3_K_L: small, substantial quality loss -- Q4_K: alias for Q4_K_M -- Q4_K_S: small, significant quality loss -- Q4_K_M: medium, balanced quality - recommended -- Q5_K: alias for Q5_K_M -- Q5_K_S: large, low quality loss - recommended -- Q5_K_M: large, very low quality loss - recommended -- Q6_K: very large, extremely low quality loss -- Q8_0: very large, extremely low quality loss - not recommended -- F16: extremely large, virtually no quality loss - not recommended -- F32: absolutely huge, lossless - not recommended -""" -import subprocess -subprocess.run(f"pip install huggingface-hub==0.16.4", shell=True, check=True) - -import argparse -import os -from huggingface_hub import snapshot_download - -def main(model, model_type, outname, outdir, quants, keep_fp16): - if not os.path.isdir(model): - print(f"Model not found at {model}. Downloading...") - try: - if outname is None: - outname = model.split('/')[-1] - model = snapshot_download(repo_id=model, cache_dir='../models/hf_cache') - except Exception as e: - raise Exception(f"Could not download the model: {e}") - - if outdir is None: - outdir = f'../models/{outname}' - - if not os.path.isfile(f"{model}/config.json"): - raise Exception(f"Could not find config.json in {model}") - - os.makedirs(outdir, exist_ok=True) - - print("Building llama.cpp") - subprocess.run(f"cd .. && make quantize", shell=True, check=True) - - fp16 = f"{outdir}/{outname}.gguf.fp16.bin" - - print(f"Making unquantised GGUF at {fp16}") - if not os.path.isfile(fp16): - if model_type != "llama": - subprocess.run(f"python3 ../convert-{model_type}-hf-to-gguf.py {model} 1 --outfile {fp16}", shell=True, check=True) - else: - subprocess.run(f"python3 ../convert.py {model} --outtype f16 --outfile {fp16}", shell=True, check=True) - else: - print(f"Unquantised GGML already exists at: {fp16}") - - print("Making quants") - for type in quants: - outfile = f"{outdir}/{outname}.gguf.{type}.bin" - print(f"Making {type} : {outfile}") - subprocess.run(f"../quantize {fp16} {outfile} {type}", shell=True, check=True) - - if not keep_fp16: - os.remove(fp16) - -if __name__ == "__main__": - parser = argparse.ArgumentParser(description='Convert/Quantize HF models to GGUF. If you have the HF model downloaded already, pass the path to the model dir. Otherwise, pass the Hugging Face model repo name. You need to be in the /examples folder for it to work.') - parser.add_argument('model', help='Downloaded model dir or Hugging Face model repo name') - parser.add_argument('--model_type', required=True, choices=['llama', 'starcoder', 'falcon', 'baichuan', 'gptneox'], help='Type of the model to be converted. Choose from llama, starcoder, falcon, baichuan, or gptneox.') - parser.add_argument('--outname', default=None, help='Output model(s) name') - parser.add_argument('--outdir', default=None, help='Output directory') - parser.add_argument('--quants', nargs='*', default=["Q4_K_M", "Q5_K_S"], help='Quant types') - parser.add_argument('--keep_fp16', action='store_true', help='Keep fp16 model', default=False) - - args = parser.parse_args() - - main(args.model, args.model_type, args.outname, args.outdir, args.quants, args.keep_fp16) diff --git a/examples/server/public/index.html b/examples/server/public/index.html index 2961999f2..4c5a34d90 100644 --- a/examples/server/public/index.html +++ b/examples/server/public/index.html @@ -594,7 +594,7 @@ message = html`<${Probabilities} data=${data} />` } else { const text = isArrayMessage ? - data.map(msg => msg.content).join('').replace(/^\s+/, '') : + data.map(msg => msg.content).join('') : data; message = isCompletionMode ? text : @@ -877,19 +877,30 @@ // poor mans markdown replacement const Markdownish = (params) => { - const md = params.text - .replace(/&/g, '&') - .replace(//g, '>') - .replace(/(^|\n)#{1,6} ([^\n]*)(?=([^`]*`[^`]*`)*[^`]*$)/g, '$1

$2

') - .replace(/\*\*(.*?)\*\*(?=([^`]*`[^`]*`)*[^`]*$)/g, '$1') - .replace(/__(.*?)__(?=([^`]*`[^`]*`)*[^`]*$)/g, '$1') - .replace(/\*(.*?)\*(?=([^`]*`[^`]*`)*[^`]*$)/g, '$1') - .replace(/_(.*?)_(?=([^`]*`[^`]*`)*[^`]*$)/g, '$1') - .replace(/```.*?\n([\s\S]*?)```/g, '
$1
') - .replace(/`(.*?)`/g, '$1') - .replace(/\n/gim, '
'); - return html``; + const chunks = params.text.split('```'); + + for (let i = 0; i < chunks.length; i++) { + if (i % 2 === 0) { // outside code block + chunks[i] = chunks[i] + .replace(/&/g, '&') + .replace(//g, '>') + .replace(/(^|\n)#{1,6} ([^\n]*)(?=([^`]*`[^`]*`)*[^`]*$)/g, '$1

$2

') + .replace(/\*\*(.*?)\*\*(?=([^`]*`[^`]*`)*[^`]*$)/g, '$1') + .replace(/__(.*?)__(?=([^`]*`[^`]*`)*[^`]*$)/g, '$1') + .replace(/\*(.*?)\*(?=([^`]*`[^`]*`)*[^`]*$)/g, '$1') + .replace(/_(.*?)_(?=([^`]*`[^`]*`)*[^`]*$)/g, '$1') + .replace(/```.*?\n([\s\S]*?)```/g, '
$1
') + .replace(/`(.*?)`/g, '$1') + .replace(/\n/gim, '
'); + } else { // inside code block + chunks[i] = `
${chunks[i]}
`; + } + } + + const restoredText = chunks.join(''); + + return html``; }; const ModelGenerationInfo = (params) => { @@ -903,6 +914,7 @@ ` } + // simple popover impl const Popover = (props) => { const isOpen = useSignal(false); @@ -1054,4 +1066,3 @@ - diff --git a/examples/server/public_simplechat/index.html b/examples/server/public_simplechat/index.html index 1eb390b85..1a1a34208 100644 --- a/examples/server/public_simplechat/index.html +++ b/examples/server/public_simplechat/index.html @@ -1,7 +1,7 @@ - SimpleChat (LlamaCPP, ...) + SimpleChat LlamaCppEtal @@ -30,20 +30,17 @@
- +
-

Enter the system prompt above, before entering/submitting any user query.

-

Enter your text to the ai assistant below.

-

Use shift+enter for inserting enter.

-

Refresh the page to start over fresh.

+

You need to have javascript enabled.

- +
diff --git a/examples/server/public_simplechat/readme.md b/examples/server/public_simplechat/readme.md index 5ac8258f2..de0dfc99d 100644 --- a/examples/server/public_simplechat/readme.md +++ b/examples/server/public_simplechat/readme.md @@ -14,11 +14,15 @@ own system prompts. The UI follows a responsive web design so that the layout can adapt to available display space in a usable enough manner, in general. -NOTE: Given that the idea is for basic minimal testing, it doesnt bother with any model context length and -culling of old messages from the chat. +Allows developer/end-user to control some of the behaviour by updating gMe members from browser's devel-tool +console. -NOTE: It doesnt set any parameters other than temperature for now. However if someone wants they can update -the js file as needed. +NOTE: Given that the idea is for basic minimal testing, it doesnt bother with any model context length and +culling of old messages from the chat by default. However by enabling the sliding window chat logic, a crude +form of old messages culling can be achieved. + +NOTE: It doesnt set any parameters other than temperature and max_tokens for now. However if someone wants +they can update the js file or equivalent member in gMe as needed. ## usage @@ -43,11 +47,33 @@ next run this web front end in examples/server/public_simplechat ### using the front end Open this simple web front end from your local browser + * http://127.0.0.1:PORT/index.html Once inside + * Select between chat and completion mode. By default it is set to chat mode. + +* In completion mode + * logic by default doesnt insert any role specific "ROLE: " prefix wrt each role's message. + If the model requires any prefix wrt user role messages, then the end user has to + explicitly add the needed prefix, when they enter their chat message. + Similarly if the model requires any prefix to trigger assistant/ai-model response, + then the end user needs to enter the same. + This keeps the logic simple, while still giving flexibility to the end user to + manage any templating/tagging requirement wrt their messages to the model. + * the logic doesnt insert newline at the begining and end wrt the prompt message generated. + However if the chat being sent to /completions end point has more than one role's message, + then insert newline when moving from one role's message to the next role's message, so + that it can be clearly identified/distinguished. + * given that /completions endpoint normally doesnt add additional chat-templating of its + own, the above ensures that end user can create a custom single/multi message combo with + any tags/special-tokens related chat templating to test out model handshake. Or enduser + can use it just for normal completion related/based query. + * If you want to provide a system prompt, then ideally enter it first, before entering any user query. + Normally Completion mode doesnt need system prompt, while Chat mode can generate better/interesting + responses with a suitable system prompt. * if chat.add_system_begin is used * you cant change the system prompt, after it is has been submitted once along with user query. * you cant set a system prompt, after you have submitted any user query @@ -55,27 +81,121 @@ Once inside * one can change the system prompt any time during chat, by changing the contents of system prompt. * inturn the updated/changed system prompt will be inserted into the chat session. * this allows for the subsequent user chatting to be driven by the new system prompt set above. + * Enter your query and either press enter or click on the submit button. If you want to insert enter (\n) as part of your chat/query to ai model, use shift+enter. + * Wait for the logic to communicate with the server and get the response. * the user is not allowed to enter any fresh query during this time. * the user input box will be disabled and a working message will be shown in it. + * just refresh the page, to reset wrt the chat history and or system prompt and start afresh. + * Using NewChat one can start independent chat sessions. * two independent chat sessions are setup by default. ## Devel note +### Reason behind this + +The idea is to be easy enough to use for basic purposes, while also being simple and easily discernable +by developers who may not be from web frontend background (so inturn may not be familiar with template / +end-use-specific-language-extensions driven flows) so that they can use it to explore/experiment things. + +And given that the idea is also to help explore/experiment for developers, some flexibility is provided +to change behaviour easily using the devel-tools/console, for now. And skeletal logic has been implemented +to explore some of the end points and ideas/implications around them. + + +### General + +Me/gMe consolidates the settings which control the behaviour into one object. +One can see the current settings, as well as change/update them using browsers devel-tool/console. + + bCompletionFreshChatAlways - whether Completion mode collates complete/sliding-window history when + communicating with the server or only sends the latest user query/message. + + bCompletionInsertStandardRolePrefix - whether Completion mode inserts role related prefix wrt the + messages that get inserted into prompt field wrt /Completion endpoint. + + chatRequestOptions - maintains the list of options/fields to send along with chat request, + irrespective of whether /chat/completions or /completions endpoint. + + If you want to add additional options/fields to send to the server/ai-model, and or + modify the existing options value or remove them, for now you can update this global var + using browser's development-tools/console. + + iRecentUserMsgCnt - a simple minded SlidingWindow to limit context window load at Ai Model end. + This is disabled by default. However if enabled, then in addition to latest system message, only + the last/latest iRecentUserMsgCnt user messages after the latest system prompt and its responses + from the ai model will be sent to the ai-model, when querying for a new response. IE if enabled, + only user messages after the latest system message/prompt will be considered. + + This specified sliding window user message count also includes the latest user query. + <0 : Send entire chat history to server + 0 : Send only the system message if any to the server + >0 : Send the latest chat history from the latest system prompt, limited to specified cnt. + + +By using gMe's iRecentUserMsgCnt and chatRequestOptions.max_tokens one can try to control the +implications of loading of the ai-model's context window by chat history, wrt chat response to +some extent in a simple crude way. + + Sometimes the browser may be stuborn with caching of the file, so your updates to html/css/js may not be visible. Also remember that just refreshing/reloading page in browser or for that matter clearing site data, dont directly override site caching in all cases. Worst case you may have to change port. Or in dev tools of browser, you may be able to disable caching fully. + Concept of multiple chat sessions with different servers, as well as saving and restoring of those across browser usage sessions, can be woven around the SimpleChat/MultiChatUI class and its instances relatively easily, however given the current goal of keeping this simple, it has not been added, for now. + By switching between chat.add_system_begin/anytime, one can control whether one can change the system prompt, anytime during the conversation or only at the beginning. + + +read_json_early, is to experiment with reading json response data early on, if available, +so that user can be shown generated data, as and when it is being generated, rather than +at the end when full data is available. + + the server flow doesnt seem to be sending back data early, atleast for request (inc options) + that is currently sent. + + if able to read json data early on in future, as and when ai model is generating data, then + this helper needs to indirectly update the chat div with the recieved data, without waiting + for the overall data to be available. + + +### Default setup + +By default things are setup to try and make the user experience a bit better, if possible. +However a developer when testing the server of ai-model may want to change these value. + +Using iRecentUserMsgCnt reduce chat history context sent to the server/ai-model to be +just the system-prompt, prev-user-request-and-ai-response and cur-user-request, instead of +full chat history. This way if there is any response with garbage/repeatation, it doesnt +mess with things beyond the next question/request/query, in some ways. + +Set max_tokens to 1024, so that a relatively large previous reponse doesnt eat up the space +available wrt next query-response. However dont forget that the server when started should +also be started with a model context size of 1k or more, to be on safe side. + + The /completions endpoint of examples/server doesnt take max_tokens, instead it takes the + internal n_predict, for now add the same here on the client side, maybe later add max_tokens + to /completions endpoint handling code on server side. + +Frequency and presence penalty fields are set to 1.2 in the set of fields sent to server +along with the user query. So that the model is partly set to try avoid repeating text in +its response. + +A end-user can change these behaviour by editing gMe from browser's devel-tool/console. + + +## At the end + +Also a thank you to all open source and open model developers, who strive for the common good. diff --git a/examples/server/public_simplechat/simplechat.css b/examples/server/public_simplechat/simplechat.css index d45f50a95..20c738b12 100644 --- a/examples/server/public_simplechat/simplechat.css +++ b/examples/server/public_simplechat/simplechat.css @@ -48,6 +48,13 @@ button { flex-direction: column; } +.ul1 { + padding-inline-start: 2vw; +} +.ul2 { + padding-inline-start: 2vw; +} + * { margin: 0.6vmin; } diff --git a/examples/server/public_simplechat/simplechat.js b/examples/server/public_simplechat/simplechat.js index 3fc4dbc20..0c48da879 100644 --- a/examples/server/public_simplechat/simplechat.js +++ b/examples/server/public_simplechat/simplechat.js @@ -14,23 +14,86 @@ class ApiEP { } let gUsageMsg = ` -

Enter the system prompt above, before entering/submitting any user query.

-

Enter your text to the ai assistant below.

-

Use shift+enter for inserting enter.

-

Refresh the page to start over fresh.

+

Usage

+
    +
  • Set system prompt above, to try control ai response charactersitic, if model supports same.
    • +
      +
    • Completion mode normally wont have a system prompt.
      • +
    +
  • Enter your query to ai assistant below.
    • +
      +
    • Completion mode doesnt insert user/role: prefix implicitly.
      • +
    • Use shift+enter for inserting enter/newline.
      • +
    +
  • Default ContextWindow = [System, Last Query+Resp, Cur Query].
    • +
      +
    • experiment iRecentUserMsgCnt, max_tokens, model ctxt window to expand
      • +
    +
`; +/** @typedef {{role: string, content: string}[]} ChatMessages */ + class SimpleChat { constructor() { /** * Maintain in a form suitable for common LLM web service chat/completions' messages entry - * @type {{role: string, content: string}[]} + * @type {ChatMessages} */ this.xchat = []; this.iLastSys = -1; } + clear() { + this.xchat = []; + this.iLastSys = -1; + } + + /** + * Recent chat messages. + * If iRecentUserMsgCnt < 0 + * Then return the full chat history + * Else + * Return chat messages from latest going back till the last/latest system prompt. + * While keeping track that the number of user queries/messages doesnt exceed iRecentUserMsgCnt. + * @param {number} iRecentUserMsgCnt + */ + recent_chat(iRecentUserMsgCnt) { + if (iRecentUserMsgCnt < 0) { + return this.xchat; + } + if (iRecentUserMsgCnt == 0) { + console.warn("WARN:SimpleChat:SC:RecentChat:iRecentUsermsgCnt of 0 means no user message/query sent"); + } + /** @type{ChatMessages} */ + let rchat = []; + let sysMsg = this.get_system_latest(); + if (sysMsg.length != 0) { + rchat.push({role: Roles.System, content: sysMsg}); + } + let iUserCnt = 0; + let iStart = this.xchat.length; + for(let i=this.xchat.length-1; i > this.iLastSys; i--) { + if (iUserCnt >= iRecentUserMsgCnt) { + break; + } + let msg = this.xchat[i]; + if (msg.role == Roles.User) { + iStart = i; + iUserCnt += 1; + } + } + for(let i = iStart; i < this.xchat.length; i++) { + let msg = this.xchat[i]; + if (msg.role == Roles.System) { + continue; + } + rchat.push({role: msg.role, content: msg.content}); + } + return rchat; + } + /** * Add an entry into xchat * @param {string} role @@ -57,7 +120,7 @@ class SimpleChat { div.replaceChildren(); } let last = undefined; - for(const x of this.xchat) { + for(const x of this.recent_chat(gMe.iRecentUserMsgCnt)) { let entry = document.createElement("p"); entry.className = `role-${x.role}`; entry.innerText = `${x.role}: ${x.content}`; @@ -69,17 +132,21 @@ class SimpleChat { } else { if (bClear) { div.innerHTML = gUsageMsg; + gMe.show_info(div); } } } /** - * Add needed fields wrt json object to be sent wrt LLM web services completions endpoint + * Add needed fields wrt json object to be sent wrt LLM web services completions endpoint. + * The needed fields/options are picked from a global object. * Convert the json into string. * @param {Object} obj */ request_jsonstr(obj) { - obj["temperature"] = 0.7; + for(let k in gMe.chatRequestOptions) { + obj[k] = gMe.chatRequestOptions[k]; + } return JSON.stringify(obj); } @@ -88,18 +155,27 @@ class SimpleChat { */ request_messages_jsonstr() { let req = { - messages: this.xchat, + messages: this.recent_chat(gMe.iRecentUserMsgCnt), } return this.request_jsonstr(req); } /** * Return a string form of json object suitable for /completions + * @param {boolean} bInsertStandardRolePrefix Insert ": " as prefix wrt each role's message */ - request_prompt_jsonstr() { + request_prompt_jsonstr(bInsertStandardRolePrefix) { let prompt = ""; - for(const chat of this.xchat) { - prompt += `${chat.role}: ${chat.content}\n`; + let iCnt = 0; + for(const chat of this.recent_chat(gMe.iRecentUserMsgCnt)) { + iCnt += 1; + if (iCnt > 1) { + prompt += "\n"; + } + if (bInsertStandardRolePrefix) { + prompt += `${chat.role}: `; + } + prompt += `${chat.content}`; } let req = { prompt: prompt, @@ -171,7 +247,6 @@ let gChatURL = { 'chat': `${gBaseURL}/chat/completions`, 'completion': `${gBaseURL}/completions`, } -const gbCompletionFreshChatAlways = true; /** @@ -291,6 +366,8 @@ class MultiChatUI { // allow user to insert enter into their message using shift+enter. // while just pressing enter key will lead to submitting. if ((ev.key === "Enter") && (!ev.shiftKey)) { + let value = this.elInUser.value; + this.elInUser.value = value.substring(0,value.length-1); this.elBtnUser.click(); ev.preventDefault(); } @@ -321,6 +398,29 @@ class MultiChatUI { } } + /** + * Try read json response early, if available. + * @param {Response} resp + */ + async read_json_early(resp) { + if (!resp.body) { + throw Error("ERRR:SimpleChat:MCUI:ReadJsonEarly:No body..."); + } + let tdUtf8 = new TextDecoder("utf-8"); + let rr = resp.body.getReader(); + let gotBody = ""; + while(true) { + let { value: cur, done: done} = await rr.read(); + let curBody = tdUtf8.decode(cur); + console.debug("DBUG:SC:PART:", curBody); + gotBody += curBody; + if (done) { + break; + } + } + return JSON.parse(gotBody); + } + /** * Handle user query submit request, wrt specified chat session. * @param {string} chatId @@ -330,6 +430,14 @@ class MultiChatUI { let chat = this.simpleChats[chatId]; + // In completion mode, if configured, clear any previous chat history. + // So if user wants to simulate a multi-chat based completion query, + // they will have to enter the full thing, as a suitable multiline + // user input/query. + if ((apiEP == ApiEP.Completion) && (gMe.bCompletionFreshChatAlways)) { + chat.clear(); + } + chat.add_system_anytime(this.elInSystem.value, chatId); let content = this.elInUser.value; @@ -344,7 +452,7 @@ class MultiChatUI { if (apiEP == ApiEP.Chat) { theBody = chat.request_messages_jsonstr(); } else { - theBody = chat.request_prompt_jsonstr(); + theBody = chat.request_prompt_jsonstr(gMe.bCompletionInsertStandardRolePrefix); } this.elInUser.value = "working..."; @@ -359,6 +467,7 @@ class MultiChatUI { }); let respBody = await resp.json(); + //let respBody = await this.read_json_early(resp); console.debug(`DBUG:SimpleChat:MCUI:${chatId}:HandleUserSubmit:RespBody:${JSON.stringify(respBody)}`); let assistantMsg; if (apiEP == ApiEP.Chat) { @@ -376,13 +485,6 @@ class MultiChatUI { } else { console.debug(`DBUG:SimpleChat:MCUI:HandleUserSubmit:ChatId has changed:[${chatId}] [${this.curChatId}]`); } - // Purposefully clear at end rather than begin of this function - // so that one can switch from chat to completion mode and sequece - // in a completion mode with multiple user-assistant chat data - // from before to be sent/occur once. - if ((apiEP == ApiEP.Completion) && (gbCompletionFreshChatAlways)) { - chat.xchat.length = 0; - } this.ui_reset_userinput(); } @@ -462,17 +564,66 @@ class MultiChatUI { } -let gMuitChat; -const gChatIds = [ "Default", "Other" ]; +class Me { + + constructor() { + this.defaultChatIds = [ "Default", "Other" ]; + this.multiChat = new MultiChatUI(); + this.bCompletionFreshChatAlways = true; + this.bCompletionInsertStandardRolePrefix = false; + this.iRecentUserMsgCnt = 2; + // Add needed fields wrt json object to be sent wrt LLM web services completions endpoint. + this.chatRequestOptions = { + "temperature": 0.7, + "max_tokens": 1024, + "frequency_penalty": 1.2, + "presence_penalty": 1.2, + "n_predict": 1024 + }; + } + + /** + * @param {HTMLDivElement} elDiv + */ + show_info(elDiv) { + + var p = document.createElement("p"); + p.innerText = "Settings (devel-tools-console gMe)"; + p.className = "role-system"; + elDiv.appendChild(p); + + var p = document.createElement("p"); + p.innerText = `bCompletionFreshChatAlways:${this.bCompletionFreshChatAlways}`; + elDiv.appendChild(p); + + p = document.createElement("p"); + p.innerText = `bCompletionInsertStandardRolePrefix:${this.bCompletionInsertStandardRolePrefix}`; + elDiv.appendChild(p); + + p = document.createElement("p"); + p.innerText = `iRecentUserMsgCnt:${this.iRecentUserMsgCnt}`; + elDiv.appendChild(p); + + p = document.createElement("p"); + p.innerText = `chatRequestOptions:${JSON.stringify(this.chatRequestOptions)}`; + elDiv.appendChild(p); + + } + +} + + +/** @type {Me} */ +let gMe; function startme() { console.log("INFO:SimpleChat:StartMe:Starting..."); - gMuitChat = new MultiChatUI(); - for (let cid of gChatIds) { - gMuitChat.new_chat_session(cid); + gMe = new Me(); + for (let cid of gMe.defaultChatIds) { + gMe.multiChat.new_chat_session(cid); } - gMuitChat.setup_ui(gChatIds[0]); - gMuitChat.show_sessions(); + gMe.multiChat.setup_ui(gMe.defaultChatIds[0], true); + gMe.multiChat.show_sessions(); } document.addEventListener("DOMContentLoaded", startme); diff --git a/examples/sycl/win-build-sycl.bat b/examples/sycl/win-build-sycl.bat index 1b0dc41ba..b8037aae8 100644 --- a/examples/sycl/win-build-sycl.bat +++ b/examples/sycl/win-build-sycl.bat @@ -13,10 +13,10 @@ if %errorlevel% neq 0 goto ERROR :: for FP16 :: faster for long-prompt inference -:: cmake -G "MinGW Makefiles" .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx -DCMAKE_BUILD_TYPE=Release -DLLAMA_SYCL_F16=ON +:: cmake -G "MinGW Makefiles" .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=Release -DLLAMA_SYCL_F16=ON :: for FP32 -cmake -G "MinGW Makefiles" .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx -DCMAKE_BUILD_TYPE=Release +cmake -G "MinGW Makefiles" .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=Release if %errorlevel% neq 0 goto ERROR :: build example/main only :: make main diff --git a/examples/tokenize/tokenize.cpp b/examples/tokenize/tokenize.cpp index 8b1baea80..54c9834af 100644 --- a/examples/tokenize/tokenize.cpp +++ b/examples/tokenize/tokenize.cpp @@ -3,40 +3,390 @@ #include #include +#include #include #include -int main(int argc, char ** argv) { - if (argc < 3 || argv[1][0] == '-') { - printf("usage: %s MODEL_PATH PROMPT [--ids]\n" , argv[0]); +#if defined(_WIN32) +#define WIN32_LEAN_AND_MEAN +#include +#include // For CommandLineToArgvW +#endif + +static void print_usage_information(const char * argv0, FILE * stream) { + fprintf(stream, "usage: %s [options]\n\n", argv0); + fprintf(stream, "The tokenize program tokenizes a prompt using a given model,\n"); + fprintf(stream, "and prints the resulting tokens to standard output.\n\n"); + fprintf(stream, "It needs a model file, a prompt, and optionally other flags\n"); + fprintf(stream, "to control the behavior of the tokenizer.\n\n"); + fprintf(stream, " The possible options are:\n"); + fprintf(stream, "\n"); + fprintf(stream, " -h, --help print this help and exit\n"); + fprintf(stream, " -m MODEL_PATH, --model MODEL_PATH path to model.\n"); + fprintf(stream, " --ids if given, only print numerical token IDs, and not token strings.\n"); + fprintf(stream, " The output format looks like [1, 2, 3], i.e. parseable by Python.\n"); + fprintf(stream, " -f PROMPT_FNAME, --file PROMPT_FNAME read prompt from a file.\n"); + fprintf(stream, " -p PROMPT, --prompt PROMPT read prompt from the argument.\n"); + 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"); +} + +static void llama_log_callback_null(ggml_log_level level, const char * text, void * user_data) { + (void) level; + (void) text; + (void) user_data; +} + +static std::string read_prompt_from_file(const char * filepath, bool & success) { + success = false; + + std::ifstream in(filepath, std::ios::binary); + if (!in) { + fprintf(stderr, "%s: could not open file '%s' for reading: %s\n", __func__, filepath, strerror(errno)); + return std::string(); + } + // do not assume the file is seekable (e.g. /dev/stdin) + std::stringstream buffer; + buffer << in.rdbuf(); + if (in.fail()) { + fprintf(stderr, "%s: could not read the entire file '%s': %s\n", __func__, filepath, strerror(errno)); + return std::string(); + } + + success = true; + return buffer.str(); +} + +// +// Function: ingest_args(...) -> vector +// +// Takes argc and argv arguments, and converts them to a vector of UTF-8 encoded +// strings, as an STL vector. +// +// In particular, it handles character encoding shenanigans on Windows. +// +// Note: raw_argc and raw_argv are not actually read at all on Windows. +// On Windows we call GetCommandLineW to get the arguments in wchar_t +// format, ignoring the regular argc/argv arguments to main(). +// +// TODO: potential opportunity to roll common stuff into common/console.cpp +// in relation to Windows wchar_t shenanigans. +static std::vector ingest_args(int raw_argc, char ** raw_argv) { + std::vector argv; + + // Handle Windows, if given non-ASCII arguments. + // We convert wchar_t arguments into UTF-8 char* on this platform. + // Lets you invoke 'tokenize' on Windows cmd.exe with non-ASCII characters + // without throwing tantrums. +#if defined(_WIN32) + int argc; + const LPWSTR cmdline_wargv = GetCommandLineW(); + LPWSTR * wargv = CommandLineToArgvW(cmdline_wargv, &argc); + + // silence unused arg warnings + (void) raw_argc; + (void) raw_argv; + + for (int i = 0; i < argc; ++i) { + int length_needed = WideCharToMultiByte(CP_UTF8, 0, wargv[i], wcslen(wargv[i]), 0, 0, NULL, NULL); + char * output_buf = (char *) calloc(length_needed+1, sizeof(char)); + GGML_ASSERT(output_buf); + + WideCharToMultiByte(CP_UTF8, 0, wargv[i], wcslen(wargv[i]), output_buf, length_needed, NULL, NULL); + output_buf[length_needed] = '\0'; + + argv.push_back(output_buf); + free(output_buf); + } + + LocalFree((HLOCAL) wargv); +#else + int argc = raw_argc; + for (int i = 0; i < argc; ++i) { + argv.push_back(raw_argv[i]); + } +#endif + + GGML_ASSERT((unsigned int) argc == argv.size()); + + return argv; +} + +// +// Function: write_utf8_cstr_to_stdout(const char *) -> +// +// writes a string to standard output; taking into account that on Windows +// to display correctly you have to use special handling. Works even if the +// user has not set a unicode code page on a Windows cmd.exe. +// +// In case of invalid UTF-8, invalid_utf8 is set to true on Windows, and something +// a human-readable is written instead. +// +// On non-Windows systems, simply printfs() the string. +static void write_utf8_cstr_to_stdout(const char * str, bool & invalid_utf8) { + invalid_utf8 = false; + +#if defined(_WIN32) + // Are we in a console? + HANDLE hConsole = GetStdHandle(STD_OUTPUT_HANDLE); + DWORD dwMode = 0; + + // According to Microsoft docs: + // "WriteConsole fails if it is used with a standard handle that is redirected to a file." + // Also according to the docs, you can use GetConsoleMode to check for that. + if (hConsole == INVALID_HANDLE_VALUE || !GetConsoleMode(hConsole, &dwMode)) { + printf("%s", str); + return; + } + + // MultiByteToWideChar reports an error if str is empty, don't report + // them as invalid_utf8. + if (*str == 0) { + return; + } + int length_needed = MultiByteToWideChar(CP_UTF8, MB_ERR_INVALID_CHARS, str, strlen(str), NULL, 0); + if (length_needed == 0) { + DWORD err = GetLastError(); + if (err == ERROR_NO_UNICODE_TRANSLATION) { + invalid_utf8 = true; + int len = strlen(str); + printf("<"); + for (int i = 0; i < len; ++i) { + if (i > 0) { + printf(" "); + } + printf("%02x", (uint8_t) str[i]); + } + printf(">"); + return; + } + GGML_ASSERT(false && "MultiByteToWideChar() failed in an unexpected way."); + } + + LPWSTR wstr = (LPWSTR) calloc(length_needed+1, sizeof(*wstr)); + GGML_ASSERT(wstr); + + MultiByteToWideChar(CP_UTF8, 0, str, strlen(str), wstr, length_needed); + WriteConsoleW(hConsole, wstr, length_needed, NULL, NULL); + + free(wstr); +#else + // TODO: reporting invalid_utf8 would be useful on non-Windows too. + // printf will silently just write bad unicode. + printf("%s", str); +#endif +} + +int main(int raw_argc, char ** raw_argv) { + const std::vector argv = ingest_args(raw_argc, raw_argv); + const int argc = argv.size(); + + if (argc <= 1) { + print_usage_information(argv[0].c_str(), stderr); return 1; } - const char * model_path = argv[1]; - const char * prompt = argv[2]; + ////// + // Read out all the command line arguments. + ////// - const bool printing_ids = argc > 3 && std::string(argv[3]) == "--ids"; + // variables where to put any arguments we see. + bool printing_ids = false; + bool no_bos = false; + bool disable_logging = false; + const char * model_path = NULL; + const char * prompt_path = NULL; + const char * prompt_arg = NULL; + + // track which arguments were explicitly given + // used for sanity checking down the line + bool model_path_set = false; + bool prompt_path_set = false; + bool prompt_set = false; + bool stdin_set = false; + + int iarg = 1; + for (; iarg < argc; ++iarg) { + std::string arg{argv[iarg]}; + if (arg == "-h" || arg == "--help") { + print_usage_information(argv[0].c_str(), stdout); + return 0; + } + else if (arg == "--ids") { + printing_ids = true; + } + else if (arg == "-m" || arg == "--model") { + if (model_path_set) { + fprintf(stderr, "Error: -m or --model specified multiple times.\n"); + return 1; + } + model_path = argv[++iarg].c_str(); + model_path_set = true; + } + else if (arg == "--no-bos") { + no_bos = true; + } + else if (arg == "-p" || arg == "--prompt") { + if (prompt_set) { + fprintf(stderr, "Error: -p or --prompt specified multiple times.\n"); + return 1; + } + prompt_arg = argv[++iarg].c_str(); + prompt_set = true; + } + else if (arg == "-f" || arg == "--file") { + if (prompt_path_set) { + fprintf(stderr, "Error: -f or --file specified multiple times.\n"); + return 1; + } + prompt_path = argv[++iarg].c_str(); + prompt_path_set = true; + } + else if (arg == "--stdin") { + stdin_set = true; + } + else if (arg == "--log-disable") { + disable_logging = true; + } + else { + fprintf(stderr, "Error: unknown option '%s'\n", argv[iarg].c_str()); + return 1; + } + } + + ////// + // Sanity check the command line arguments. + ////// + + // Check that we have the required stuff set. + if (model_path_set && model_path == NULL) { + fprintf(stderr, "Error: --model requires an argument.\n"); + return 1; + } + if (!model_path_set) { + fprintf(stderr, "Error: must specify --model.\n"); + return 1; + } + if (prompt_path_set && prompt_path == NULL) { + fprintf(stderr, "Error: --file requires an argument.\n"); + return 1; + } + if (prompt_set && prompt_arg == NULL) { + fprintf(stderr, "Error: --prompt requires an argument.\n"); + return 1; + } + const int prompts_set = !!(prompt_path_set) + !!(prompt_set) + !!(stdin_set); + if (prompts_set > 1) { + fprintf(stderr, "Error: --stdin, --file and --prompt are mutually exclusive.\n"); + return 1; + } + // Must have some prompt. + if (prompts_set == 0) { + fprintf(stderr, "Error: must specify one of: --stdin, --file or --prompt.\n"); + return 1; + } + + GGML_ASSERT(model_path); + GGML_ASSERT(prompt_path || prompt_arg || stdin_set); + + ////// + // Figure out where will the prompt come from. + ////// + + std::string prompt; + if (prompt_path_set) { + bool success = false; + prompt = read_prompt_from_file(prompt_path, success); + if (!success) { + return 1; + } + } else if (prompt_set) { + prompt = prompt_arg; + } else { + GGML_ASSERT(stdin_set); + // we read stdin *after* loading model (early exit if model cannot + // be loaded, which can be a nicer user experience) + } + + ////// + // Start actually doing the tokenizing stuff. + ////// + +#ifdef LOG_DISABLE_LOGS + disable_logging = true; +#endif + + if (disable_logging) { + llama_log_set(llama_log_callback_null, NULL); + } llama_backend_init(); llama_model_params model_params = llama_model_default_params(); model_params.vocab_only = true; llama_model * model = llama_load_model_from_file(model_path, model_params); + if (!model) { + fprintf(stderr, "Error: could not load model from file '%s'.\n", model_path); + return 1; + } llama_context_params ctx_params = llama_context_default_params(); llama_context * ctx = llama_new_context_with_model(model, ctx_params); + if (!ctx) { + fprintf(stderr, "Error: could not create context.\n"); + return 1; + } + + // read entire prompt from stdin? + if (stdin_set) { + GGML_ASSERT(!prompt_path_set && !prompt_set); + + std::stringstream stdin_buffer; + stdin_buffer << std::cin.rdbuf(); + if (std::cin.fail()) { + fprintf(stderr, "Error: could not read the entire standard input.\n"); + return 1; + } + + prompt = stdin_buffer.str(); + } + + const bool model_wants_add_bos = llama_should_add_bos_token(model); + const bool add_bos = model_wants_add_bos && !no_bos; std::vector tokens; + tokens = ::llama_tokenize(model, prompt, add_bos, true); - tokens = ::llama_tokenize(model, prompt, true, true); + if (printing_ids) { + printf("["); + } for (int i = 0; i < (int) tokens.size(); i++) { if (printing_ids) { - printf("%d\n", tokens[i]); + if (i > 0) { + printf(", "); + } + printf("%d", tokens[i]); } else { - printf("%6d -> '%s'\n", tokens[i], llama_token_to_piece(ctx, tokens[i]).c_str()); + bool invalid_utf8 = false; + printf("%6d -> '", tokens[i]); + write_utf8_cstr_to_stdout(llama_token_to_piece(ctx, tokens[i]).c_str(), invalid_utf8); + if (invalid_utf8) { + printf("' (utf-8 decode failure)\n"); + } else { + printf("'\n"); + } } } + if (printing_ids) { + printf("]\n"); + } + + // silence valgrind + llama_free(ctx); + llama_free_model(model); + return 0; } diff --git a/flake.lock b/flake.lock index c9ead0bf7..fd6e2a5f6 100644 --- a/flake.lock +++ b/flake.lock @@ -5,11 +5,11 @@ "nixpkgs-lib": "nixpkgs-lib" }, "locked": { - "lastModified": 1714641030, - "narHash": "sha256-yzcRNDoyVP7+SCNX0wmuDju1NUCt8Dz9+lyUXEI0dbI=", + "lastModified": 1715865404, + "narHash": "sha256-/GJvTdTpuDjNn84j82cU6bXztE0MSkdnTWClUCRub78=", "owner": "hercules-ci", "repo": "flake-parts", - "rev": "e5d10a24b66c3ea8f150e47dfdb0416ab7c3390e", + "rev": "8dc45382d5206bd292f9c2768b8058a8fd8311d9", "type": "github" }, "original": { @@ -20,11 +20,11 @@ }, "nixpkgs": { "locked": { - "lastModified": 1714635257, - "narHash": "sha256-4cPymbty65RvF1DWQfc+Bc8B233A1BWxJnNULJKQ1EY=", + "lastModified": 1716509168, + "narHash": "sha256-4zSIhSRRIoEBwjbPm3YiGtbd8HDWzFxJjw5DYSDy1n8=", "owner": "NixOS", "repo": "nixpkgs", - "rev": "63c3a29ca82437c87573e4c6919b09a24ea61b0f", + "rev": "bfb7a882678e518398ce9a31a881538679f6f092", "type": "github" }, "original": { diff --git a/ggml-cuda.cu b/ggml-cuda.cu index b82167cbf..1172f7b2f 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -119,6 +119,20 @@ int ggml_cuda_get_device() { return id; } +static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device) { + ggml_cuda_set_device(device); +#if defined(GGML_USE_HIPBLAS) && defined(GGML_HIP_UMA) + auto res = hipMallocManaged(ptr, size); + if (res == hipSuccess) { + // if error we "need" to know why... + CUDA_CHECK(hipMemAdvise(*ptr, size, hipMemAdviseSetCoarseGrain, device)); + } + return res; +#else + return cudaMalloc(ptr, size); +#endif +} + static ggml_cuda_device_info ggml_cuda_init() { #ifdef __HIP_PLATFORM_AMD__ // Workaround for a rocBLAS bug when using multiple graphics cards: @@ -271,7 +285,7 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool { size_t look_ahead_size = (size_t) (1.05 * size); look_ahead_size = 256 * ((look_ahead_size + 255)/256); ggml_cuda_set_device(device); - CUDA_CHECK(cudaMalloc((void **) &ptr, look_ahead_size)); + CUDA_CHECK(ggml_cuda_device_malloc(&ptr, look_ahead_size, device)); *actual_size = look_ahead_size; pool_size += look_ahead_size; #ifdef DEBUG_CUDA_MALLOC @@ -537,7 +551,7 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffe size = std::max(size, (size_t)1); // cudaMalloc returns null for size 0 void * dev_ptr; - cudaError_t err = cudaMalloc(&dev_ptr, size); + cudaError_t err = ggml_cuda_device_malloc(&dev_ptr, size, buft_ctx->device); if (err != cudaSuccess) { // clear the error cudaGetLastError(); @@ -798,7 +812,7 @@ GGML_CALL static void ggml_backend_cuda_split_buffer_init_tensor(ggml_backend_bu // currently, init_tensor cannot fail, it needs to be fixed in ggml-backend first ggml_cuda_set_device(id); char * buf; - CUDA_CHECK(cudaMalloc(&buf, size)); + CUDA_CHECK(ggml_cuda_device_malloc((void**)&buf, size, id)); // set padding to 0 to avoid possible NaN values if (size > original_size) { @@ -1856,7 +1870,7 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co } } #else - if (r2 == 1 && r3 == 1 && src0->nb[2]*src0->ne[2] == src0->nb[3] && src1->nb[2]*src1->ne[2] == src1->nb[3]) { + if (r2 == 1 && r3 == 1 && ggml_is_contiguous_2(src0) && ggml_is_contiguous_2(src1)) { // there is no broadcast and src0, src1 are contiguous across dims 2, 3 // use cublasGemmStridedBatchedEx CUBLAS_CHECK( @@ -2510,9 +2524,9 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t bool use_cuda_graph = true; bool cuda_graph_update_required = false; - // pointer to CUDA cpy kernel, which is required to identify + // vector of pointers to CUDA cpy kernels, which are required to identify // kernel parameters which need updated in the graph for each token - void * ggml_cuda_cpy_fn_ptr = nullptr; + std::vector ggml_cuda_cpy_fn_ptrs; if (cuda_ctx->cuda_graph->graph == nullptr) { if (ggml_cuda_info().devices[cuda_ctx->device].cc < CC_AMPERE) { @@ -2588,9 +2602,10 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t if (node->op == GGML_OP_CPY) { // store the copy op parameter which changes with each token. cuda_ctx->cuda_graph->updated_kernel_arg.push_back((char **) &(node->src[1]->data)); - if (ggml_cuda_cpy_fn_ptr == nullptr) { - // store a pointer to the copy op CUDA kernel to identify it later - ggml_cuda_cpy_fn_ptr = ggml_cuda_cpy_fn(node->src[0], node->src[1]); + // store a pointer to each copy op CUDA kernel to identify it later + void * ptr = ggml_cuda_cpy_fn(node->src[0], node->src[1]); + if (std::find(ggml_cuda_cpy_fn_ptrs.begin(), ggml_cuda_cpy_fn_ptrs.end(), ptr) == ggml_cuda_cpy_fn_ptrs.end()) { + ggml_cuda_cpy_fn_ptrs.push_back(ptr); } } @@ -2720,7 +2735,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t if (!cuda_graph_update_required) { // on update steps, the live parameters will already be captured int k = 0; for (size_t i = 0; i < cuda_ctx->cuda_graph->num_nodes; i++) { - if (cuda_ctx->cuda_graph->params[i].func == ggml_cuda_cpy_fn_ptr) { + if(count(ggml_cuda_cpy_fn_ptrs.begin(), ggml_cuda_cpy_fn_ptrs.end(), cuda_ctx->cuda_graph->params[i].func) > 0) { char ** updated_kernel_arg_ptr = cuda_ctx->cuda_graph->updated_kernel_arg.at(k++); cuda_ctx->cuda_graph->params[i].kernelParams[1] = updated_kernel_arg_ptr; CUDA_CHECK(cudaGraphKernelNodeSetParams(cuda_ctx->cuda_graph->nodes[i], &cuda_ctx->cuda_graph->params[i])); @@ -2871,7 +2886,9 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons case GGML_OP_CONT: case GGML_OP_DIAG_MASK_INF: case GGML_OP_SOFT_MAX: + return true; case GGML_OP_ROPE: + return ggml_is_contiguous(op->src[0]); case GGML_OP_IM2COL: case GGML_OP_POOL_2D: case GGML_OP_SUM_ROWS: diff --git a/ggml-cuda/common.cuh b/ggml-cuda/common.cuh index 8f6fd71cf..22872ca5c 100644 --- a/ggml-cuda/common.cuh +++ b/ggml-cuda/common.cuh @@ -79,13 +79,8 @@ #define cudaHostRegisterReadOnly hipHostRegisterReadOnly #define cudaHostUnregister hipHostUnregister #define cudaLaunchHostFunc hipLaunchHostFunc -#ifdef GGML_HIP_UMA -#define cudaMalloc hipMallocManaged -#define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size) -#else #define cudaMalloc hipMalloc #define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size, hipHostMallocDefault) -#endif #define cudaMemcpy hipMemcpy #define cudaMemcpyAsync hipMemcpyAsync #define cudaMemcpyPeerAsync hipMemcpyPeerAsync diff --git a/ggml-cuda/concat.cu b/ggml-cuda/concat.cu index 2941d2f17..dac10ec36 100644 --- a/ggml-cuda/concat.cu +++ b/ggml-cuda/concat.cu @@ -1,15 +1,69 @@ #include "concat.cuh" -static __global__ void concat_f32(const float * x,const float * y, float * dst, const int ne0, const int ne02) { +// contiguous kernels +static __global__ void concat_f32_dim0(const float * x, const float * y, float * dst, const int ne0, const int ne00) { int nidx = threadIdx.x + blockIdx.x * blockDim.x; if (nidx >= ne0) { return; } - // operation + int offset_dst = nidx + blockIdx.y * ne0 + blockIdx.z * ne0 * gridDim.y; + + if (nidx < ne00) { // src0 + int offset_src = + nidx + + blockIdx.y * ne00 + + blockIdx.z * ne00 * gridDim.y; + dst[offset_dst] = x[offset_src]; + } else { + int offset_src = + (nidx - ne00) + + blockIdx.y * (ne0 - ne00) + + blockIdx.z * (ne0 - ne00) * gridDim.y; + dst[offset_dst] = y[offset_src]; + } +} + +static __global__ void concat_f32_dim1(const float * x, const float * y, float * dst, const int ne0, const int ne01) { + int nidx = threadIdx.x + blockIdx.x * blockDim.x; + if (nidx >= ne0) { + return; + } + + int offset_dst = + nidx + + blockIdx.y * ne0 + + blockIdx.z * ne0 * gridDim.y; + + if (blockIdx.y < ne01) { // src0 + int offset_src = + nidx + + blockIdx.y * ne0 + + blockIdx.z * ne0 * ne01; + dst[offset_dst] = x[offset_src]; + } else { + int offset_src = + nidx + + (blockIdx.y - ne01) * ne0 + + blockIdx.z * ne0 * (gridDim.y - ne01); + dst[offset_dst] = y[offset_src]; + } +} + +static __global__ void concat_f32_dim2(const float * x, const float * y, float * dst, const int ne0, const int ne02) { + int nidx = threadIdx.x + blockIdx.x * blockDim.x; + if (nidx >= ne0) { + return; + } + + int offset_dst = + nidx + + blockIdx.y * ne0 + + blockIdx.z * ne0 * gridDim.y; + if (blockIdx.z < ne02) { // src0 int offset_src = nidx + @@ -25,25 +79,118 @@ static __global__ void concat_f32(const float * x,const float * y, float * dst, } } -static void concat_f32_cuda(const float * x, const float * y, float * dst, const int ne0, int ne1, int ne2, int ne02, cudaStream_t stream) { +static void concat_f32_cuda(const float * x, const float * y, float * dst, int ne00, int ne01, int ne02, int ne0, int ne1, int ne2, int dim, cudaStream_t stream) { int num_blocks = (ne0 + CUDA_CONCAT_BLOCK_SIZE - 1) / CUDA_CONCAT_BLOCK_SIZE; dim3 gridDim(num_blocks, ne1, ne2); - concat_f32<<>>(x, y, dst, ne0, ne02); + if (dim == 0) { + concat_f32_dim0<<>>(x, y, dst, ne0, ne00); + return; + } + if (dim == 1) { + concat_f32_dim1<<>>(x, y, dst, ne0, ne01); + return; + } + concat_f32_dim2<<>>(x, y, dst, ne0, ne02); } +// non-contiguous kernel (slow) +static __global__ void concat_f32_non_cont( + const char * src0, + const char * src1, + char * dst, + int64_t ne00, + int64_t ne01, + int64_t ne02, + int64_t ne03, + uint64_t nb00, + uint64_t nb01, + uint64_t nb02, + uint64_t nb03, + int64_t /*ne10*/, + int64_t /*ne11*/, + int64_t /*ne12*/, + int64_t /*ne13*/, + uint64_t nb10, + uint64_t nb11, + uint64_t nb12, + uint64_t nb13, + int64_t ne0, + int64_t /*ne1*/, + int64_t /*ne2*/, + int64_t /*ne3*/, + uint64_t nb0, + uint64_t nb1, + uint64_t nb2, + uint64_t nb3, + int32_t dim) { + const int64_t i3 = blockIdx.z; + const int64_t i2 = blockIdx.y; + const int64_t i1 = blockIdx.x; + + int64_t o[4] = {0, 0, 0, 0}; + o[dim] = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03)); + + const float * x; + + for (int i0 = threadIdx.x; i0 < ne0; i0 += blockDim.x) { + if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) { + x = (const float *)(src0 + (i3 )*nb03 + (i2 )*nb02 + (i1 )*nb01 + (i0 )*nb00); + } else { + x = (const float *)(src1 + (i3 - o[3])*nb13 + (i2 - o[2])*nb12 + (i1 - o[1])*nb11 + (i0 - o[0])*nb10); + } + + float * y = (float *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + *y = *x; + } +} + + void ggml_cuda_op_concat(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { const ggml_tensor * src0 = dst->src[0]; const ggml_tensor * src1 = dst->src[1]; - const float * src0_d = (const float *)src0->data; - const float * src1_d = (const float *)src1->data; - float * dst_d = (float *)dst->data; + cudaStream_t stream = ctx.stream(); + const int32_t dim = ((int32_t *) dst->op_params)[0]; + GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT(src1->type == GGML_TYPE_F32); - GGML_ASSERT(dst->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); - for (int i3 = 0; i3 < dst->ne[3]; i3++) { - concat_f32_cuda(src0_d + i3 * (src0->nb[3] / 4), src1_d + i3 * (src1->nb[3] / 4), dst_d + i3 * (dst->nb[3] / 4), dst->ne[0], dst->ne[1], dst->ne[2], src0->ne[2], stream); + if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) { + const float * src0_d = (const float *)src0->data; + const float * src1_d = (const float *)src1->data; + + float * dst_d = (float *)dst->data; + + if (dim != 3) { + for (int i3 = 0; i3 < dst->ne[3]; i3++) { + concat_f32_cuda( + src0_d + i3 * (src0->nb[3] / 4), + src1_d + i3 * (src1->nb[3] / 4), + dst_d + i3 * ( dst->nb[3] / 4), + src0->ne[0], src0->ne[1], src0->ne[2], + dst->ne[0], dst->ne[1], dst->ne[2], dim, stream); + } + } else { + const size_t size0 = ggml_nbytes(src0); + const size_t size1 = ggml_nbytes(src1); + + CUDA_CHECK(cudaMemcpyAsync(dst_d, src0_d, size0, cudaMemcpyDeviceToDevice, stream)); + CUDA_CHECK(cudaMemcpyAsync(dst_d + size0/4, src1_d, size1, cudaMemcpyDeviceToDevice, stream)); + } + } else { + dim3 grid_dim(dst->ne[1], dst->ne[2], dst->ne[3]); + concat_f32_non_cont<<>>( + (const char *)src0->data, + (const char *)src1->data, + ( char *)dst->data, + src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], + src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3], + src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3], + src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3], + dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], + dst->nb[0], dst->nb[1], dst->nb[2], dst->nb[3], dim); } } diff --git a/ggml-cuda/norm.cu b/ggml-cuda/norm.cu index 86f774534..30866d512 100644 --- a/ggml-cuda/norm.cu +++ b/ggml-cuda/norm.cu @@ -170,6 +170,8 @@ void ggml_cuda_op_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { float * dst_d = (float *)dst->data; cudaStream_t stream = ctx.stream(); + GGML_ASSERT(ggml_is_contiguous(src0)); + GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F32); @@ -188,6 +190,8 @@ void ggml_cuda_op_group_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) float * dst_d = (float *)dst->data; cudaStream_t stream = ctx.stream(); + GGML_ASSERT(ggml_is_contiguous(src0)); + GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F32); @@ -202,6 +206,8 @@ void ggml_cuda_op_rms_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { float * dst_d = (float *)dst->data; cudaStream_t stream = ctx.stream(); + GGML_ASSERT(ggml_is_contiguous(src0)); + GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F32); diff --git a/ggml-cuda/rope.cu b/ggml-cuda/rope.cu index 50f2cf415..0dd07977e 100644 --- a/ggml-cuda/rope.cu +++ b/ggml-cuda/rope.cu @@ -61,7 +61,7 @@ static __global__ void rope( template static __global__ void rope_neox( const T * x, T * dst, int ncols, int n_dims, const int32_t * pos, float freq_scale, int p_delta_rows, - float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, float inv_ndims, const float * freq_factors + float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, const float * freq_factors ) { const int col = 2*(blockDim.y*blockIdx.y + threadIdx.y); @@ -85,15 +85,13 @@ static __global__ void rope_neox( const int i = row*ncols + ib*n_dims + ic/2; const int i2 = row/p_delta_rows; - float cur_rot = inv_ndims * ic - ib; - const int p = has_pos ? pos[i2] : 0; const float freq_factor = has_freq_facs ? freq_factors[ic/2] : 1.0f; - const float theta_base = p*freq_scale*powf(theta_scale, col/2.0f)/freq_factor; + const float theta_base = p*powf(theta_scale, col/2.0f)/freq_factor; float cos_theta, sin_theta; - rope_yarn(theta_base, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, &cos_theta, &sin_theta); + rope_yarn(theta_base, freq_scale, corr_dims, ic, ext_factor, attn_factor, &cos_theta, &sin_theta); const float x0 = x[i + 0]; const float x1 = x[i + n_dims/2]; @@ -174,30 +172,29 @@ static void rope_neox_cuda( const dim3 block_nums(nrows, num_blocks_x, 1); const float theta_scale = powf(freq_base, -2.0f/n_dims); - const float inv_ndims = -1.0f / n_dims; if (pos == nullptr) { if (freq_factors == nullptr) { rope_neox<<>>( x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims, - theta_scale, inv_ndims, freq_factors + theta_scale, freq_factors ); } else { rope_neox<<>>( x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims, - theta_scale, inv_ndims, freq_factors + theta_scale, freq_factors ); } } else { if (freq_factors == nullptr) { rope_neox<<>>( x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims, - theta_scale, inv_ndims, freq_factors + theta_scale, freq_factors ); } else { rope_neox<<>>( x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims, - theta_scale, inv_ndims, freq_factors + theta_scale, freq_factors ); } } @@ -254,6 +251,7 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { float * dst_d = (float *)dst->data; cudaStream_t stream = ctx.stream(); + GGML_ASSERT(ggml_is_contiguous(src0)); GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16); GGML_ASSERT( dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); GGML_ASSERT(src0->type == dst->type); diff --git a/ggml-impl.h b/ggml-impl.h index 362d40f4d..5e77471f3 100644 --- a/ggml-impl.h +++ b/ggml-impl.h @@ -144,6 +144,10 @@ extern "C" { #endif #endif +#if defined(__ARM_FEATURE_SVE) +#include +#endif + // 16-bit float // on Arm, we use __fp16 // on x86, we use uint16_t diff --git a/ggml-kompute.cpp b/ggml-kompute.cpp index 6c6058b2a..0c51c322f 100644 --- a/ggml-kompute.cpp +++ b/ggml-kompute.cpp @@ -1597,7 +1597,6 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml { GGML_ASSERT(ne00 == ne10); - // TODO: assert that dim2 and dim3 are contiguous GGML_ASSERT(ne12 % ne02 == 0); GGML_ASSERT(ne13 % ne03 == 0); diff --git a/ggml-metal.m b/ggml-metal.m index c9e570dbf..079912952 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -35,6 +35,10 @@ enum ggml_metal_kernel_type { GGML_METAL_KERNEL_TYPE_MUL_ROW, GGML_METAL_KERNEL_TYPE_DIV, GGML_METAL_KERNEL_TYPE_DIV_ROW, + GGML_METAL_KERNEL_TYPE_REPEAT_F32, + GGML_METAL_KERNEL_TYPE_REPEAT_F16, + GGML_METAL_KERNEL_TYPE_REPEAT_I32, + GGML_METAL_KERNEL_TYPE_REPEAT_I16, GGML_METAL_KERNEL_TYPE_SCALE, GGML_METAL_KERNEL_TYPE_SCALE_4, GGML_METAL_KERNEL_TYPE_CLAMP, @@ -184,9 +188,9 @@ enum ggml_metal_kernel_type { GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128, - GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256, + //GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256, // https://github.com/ggerganov/llama.cpp/issues/7261 GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128, - GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256, + //GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256, // https://github.com/ggerganov/llama.cpp/issues/7261 GGML_METAL_KERNEL_TYPE_CPY_F32_F16, GGML_METAL_KERNEL_TYPE_CPY_F32_F32, GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0, @@ -485,6 +489,10 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) { GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_ROW, mul_row, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV, div, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV_ROW, div_row, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_F32, repeat_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_F16, repeat_f16, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_I32, repeat_i32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_I16, repeat_i16, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE, scale, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE_4, scale_4, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CLAMP, clamp, true); @@ -634,9 +642,9 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) { GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96, flash_attn_ext_f16_h96, ctx->support_simdgroup_mm); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112, flash_attn_ext_f16_h112, ctx->support_simdgroup_mm); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128, flash_attn_ext_f16_h128, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256, flash_attn_ext_f16_h256, ctx->support_simdgroup_mm); + //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256, flash_attn_ext_f16_h256, ctx->support_simdgroup_mm); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128, flash_attn_ext_vec_f16_h128, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256, flash_attn_ext_vec_f16_h256, ctx->support_simdgroup_reduction); + //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256, flash_attn_ext_vec_f16_h256, ctx->support_simdgroup_reduction); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F16, cpy_f32_f16, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F32, cpy_f32_f32, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0, cpy_f32_q8_0, true); @@ -746,6 +754,7 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const case GGML_OP_ACC: case GGML_OP_MUL: case GGML_OP_DIV: + case GGML_OP_REPEAT: case GGML_OP_SCALE: case GGML_OP_CLAMP: case GGML_OP_SQR: @@ -770,6 +779,9 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const case GGML_OP_LEAKY_RELU: return true; case GGML_OP_FLASH_ATTN_EXT: + if (op->src[0]->ne[0] == 256) { + return false; + } return ctx->support_simdgroup_mm; // TODO: over-restricted for vec-kernels case GGML_OP_MUL_MAT: case GGML_OP_MUL_MAT_ID: @@ -976,10 +988,10 @@ static enum ggml_status ggml_metal_graph_compute( switch (dst->op) { case GGML_OP_CONCAT: { - const int64_t nb = ne00; - id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CONCAT].pipeline; + const int32_t dim = ((int32_t *) dst->op_params)[0]; + [encoder setComputePipelineState:pipeline]; [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; @@ -1008,7 +1020,7 @@ static enum ggml_status ggml_metal_graph_compute( [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:24]; [encoder setBytes:&nb2 length:sizeof(nb2) atIndex:25]; [encoder setBytes:&nb3 length:sizeof(nb3) atIndex:26]; - [encoder setBytes:&nb length:sizeof(nb) atIndex:27]; + [encoder setBytes:&dim length:sizeof(dim) atIndex:27]; const int nth = MIN(1024, ne0); @@ -1018,11 +1030,14 @@ static enum ggml_status ggml_metal_graph_compute( case GGML_OP_MUL: case GGML_OP_DIV: { + GGML_ASSERT(src0t == GGML_TYPE_F32); + GGML_ASSERT(src1t == GGML_TYPE_F32); + const size_t offs = 0; bool bcast_row = false; - int64_t nb = ne00; + int64_t nb = ne00; // used by the "row" kernels id pipeline = nil; @@ -1091,6 +1106,42 @@ static enum ggml_status ggml_metal_graph_compute( [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; } } break; + case GGML_OP_REPEAT: + { + id pipeline; + + switch (src0t) { + case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_F32].pipeline; break; + case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_F16].pipeline; break; + case GGML_TYPE_I32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_I32].pipeline; break; + case GGML_TYPE_I16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_I16].pipeline; break; + default: GGML_ASSERT(false); + } + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2]; + [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3]; + [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4]; + [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5]; + [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6]; + [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7]; + [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8]; + [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9]; + [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:10]; + [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:11]; + [encoder setBytes:&ne2 length:sizeof(ne2) atIndex:12]; + [encoder setBytes:&ne3 length:sizeof(ne3) atIndex:13]; + [encoder setBytes:&nb0 length:sizeof(nb0) atIndex:14]; + [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:15]; + [encoder setBytes:&nb2 length:sizeof(nb2) atIndex:16]; + [encoder setBytes:&nb3 length:sizeof(nb3) atIndex:17]; + + const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne0); + + [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; + } break; case GGML_OP_ACC: { GGML_ASSERT(src0t == GGML_TYPE_F32); @@ -1468,7 +1519,6 @@ static enum ggml_status ggml_metal_graph_compute( { GGML_ASSERT(ne00 == ne10); - // TODO: assert that dim2 and dim3 are contiguous GGML_ASSERT(ne12 % ne02 == 0); GGML_ASSERT(ne13 % ne03 == 0); @@ -2136,6 +2186,7 @@ static enum ggml_status ggml_metal_graph_compute( case GGML_OP_RMS_NORM: { GGML_ASSERT(ne00 % 4 == 0); + GGML_ASSERT(ggml_is_contiguous_1(src0)); float eps; memcpy(&eps, dst->op_params, sizeof(float)); @@ -2163,6 +2214,7 @@ static enum ggml_status ggml_metal_graph_compute( case GGML_OP_GROUP_NORM: { GGML_ASSERT(ne00 % 4 == 0); + GGML_ASSERT(ggml_is_contiguous(src0)); //float eps; //memcpy(&eps, dst->op_params, sizeof(float)); @@ -2196,6 +2248,8 @@ static enum ggml_status ggml_metal_graph_compute( } break; case GGML_OP_NORM: { + GGML_ASSERT(ggml_is_contiguous_1(src0)); + float eps; memcpy(&eps, dst->op_params, sizeof(float)); @@ -2573,7 +2627,7 @@ static enum ggml_status ggml_metal_graph_compute( case 96: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96 ].pipeline; break; case 112: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112].pipeline; break; case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128].pipeline; break; - case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256].pipeline; break; + //case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256].pipeline; break; default: { GGML_METAL_LOG_ERROR("unsupported size: %lld\n", ne00); @@ -2586,7 +2640,7 @@ static enum ggml_status ggml_metal_graph_compute( switch (ne00) { case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128].pipeline; break; - case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256].pipeline; break; + //case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256].pipeline; break; default: { GGML_METAL_LOG_ERROR("unsupported size: %lld\n", ne00); diff --git a/ggml-metal.metal b/ggml-metal.metal index 8ff70d7a7..0cb85e1a5 100644 --- a/ggml-metal.metal +++ b/ggml-metal.metal @@ -168,6 +168,53 @@ kernel void kernel_div( } } +template +kernel void kernel_repeat( + device const char * src0, + device char * dst, + constant int64_t & ne00, + constant int64_t & ne01, + constant int64_t & ne02, + constant int64_t & ne03, + constant uint64_t & nb00, + constant uint64_t & nb01, + constant uint64_t & nb02, + constant uint64_t & nb03, + constant int64_t & ne0, + constant int64_t & ne1, + constant int64_t & ne2, + constant int64_t & ne3, + constant uint64_t & nb0, + constant uint64_t & nb1, + constant uint64_t & nb2, + constant uint64_t & nb3, + uint3 tgpig[[threadgroup_position_in_grid]], + uint3 tpitg[[thread_position_in_threadgroup]], + uint3 ntg[[threads_per_threadgroup]]) { + const int64_t i3 = tgpig.z; + const int64_t i2 = tgpig.y; + const int64_t i1 = tgpig.x; + + const int64_t i03 = i3 % ne03; + const int64_t i02 = i2 % ne02; + const int64_t i01 = i1 % ne01; + + device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01; + device char * dst_ptr = dst + i3*nb3 + i2*nb2 + i1*nb1 ; + + for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) { + const int i00 = i0 % ne00; + *((device T *)(dst_ptr + i0*nb0)) = *((device T *)(src0_ptr + i00*nb00)); + } +} + +typedef decltype(kernel_repeat) kernel_repeat_t; + +template [[host_name("kernel_repeat_f32")]] kernel kernel_repeat_t kernel_repeat; +template [[host_name("kernel_repeat_f16")]] kernel kernel_repeat_t kernel_repeat; +template [[host_name("kernel_repeat_i32")]] kernel kernel_repeat_t kernel_repeat; +template [[host_name("kernel_repeat_i16")]] kernel kernel_repeat_t kernel_repeat; + // assumption: src1 is a row // broadcast src1 into src0 kernel void kernel_add_row( @@ -1720,13 +1767,13 @@ kernel void kernel_rope( const int64_t p = pos[i2]; - const float theta_0 = (float)p; + const float theta_base = (float)p; const float inv_ndims = -1.f/n_dims; if (!is_neox) { for (int64_t i0 = 2*tiitg; i0 < ne0; i0 += 2*tptg.x) { + const float theta = theta_base * pow(freq_base, inv_ndims*i0); - const float theta = theta_0 * pow(freq_base, inv_ndims*i0); float cos_theta, sin_theta; rope_yarn(theta, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta); @@ -1742,18 +1789,14 @@ kernel void kernel_rope( } else { for (int64_t ic = 2*tiitg; ic < ne0; ic += 2*tptg.x) { if (ic < n_dims) { - const int64_t ib = 0; + const int64_t i0 = ic/2; - // simplified from `(ib * n_dims + ic) * inv_ndims` - const float cur_rot = inv_ndims*ic - ib; - const float freq_factor = src2 != src0 ? src2[ic/2] : 1.0f; + const float freq_factor = src2 != src0 ? src2[i0] : 1.0f; - const float theta = theta_0 * pow(freq_base, cur_rot) / freq_factor; + const float theta = theta_base * pow(freq_base, inv_ndims*ic); float cos_theta, sin_theta; - rope_yarn(theta, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, &cos_theta, &sin_theta); - - const int64_t i0 = ib*n_dims + ic/2; + rope_yarn(theta/freq_factor, freq_scale, corr_dims, ic, ext_factor, attn_factor, &cos_theta, &sin_theta); device const T * const src = (device T *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); device T * dst_data = (device T *)((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); @@ -2418,7 +2461,7 @@ template [[host_name("kernel_flash_attn_ext_f16_h80" )]] kernel flash_attn_ext_f template [[host_name("kernel_flash_attn_ext_f16_h96" )]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<96>; template [[host_name("kernel_flash_attn_ext_f16_h112")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<112>; template [[host_name("kernel_flash_attn_ext_f16_h128")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<128>; -template [[host_name("kernel_flash_attn_ext_f16_h256")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<256>; +//template [[host_name("kernel_flash_attn_ext_f16_h256")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<256>; template // head size, queries per threadgroup, cache items per threadgroup kernel void kernel_flash_attn_ext_vec_f16( @@ -2696,7 +2739,7 @@ kernel void kernel_flash_attn_ext_vec_f16( } template [[host_name("kernel_flash_attn_ext_vec_f16_h128")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_vec_f16<128>; -template [[host_name("kernel_flash_attn_ext_vec_f16_h256")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_vec_f16<256>; +//template [[host_name("kernel_flash_attn_ext_vec_f16_h256")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_vec_f16<256>; kernel void kernel_cpy_f16_f16( device const half * src0, @@ -3319,31 +3362,30 @@ kernel void kernel_concat( constant uint64_t & nb1, constant uint64_t & nb2, constant uint64_t & nb3, + constant int32_t & dim, uint3 tgpig[[threadgroup_position_in_grid]], uint3 tpitg[[thread_position_in_threadgroup]], uint3 ntg[[threads_per_threadgroup]]) { - const int64_t i03 = tgpig.z; - const int64_t i02 = tgpig.y; - const int64_t i01 = tgpig.x; + const int64_t i3 = tgpig.z; + const int64_t i2 = tgpig.y; + const int64_t i1 = tgpig.x; - const int64_t i13 = i03 % ne13; - const int64_t i12 = i02 % ne12; - const int64_t i11 = i01 % ne11; + int64_t o[4] = {0, 0, 0, 0}; + o[dim] = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03)); - device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01 + tpitg.x*nb00; - device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11 + tpitg.x*nb10; - device char * dst_ptr = dst + i03*nb3 + i02*nb2 + i01*nb1 + tpitg.x*nb0; + device const float * x; for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) { - if (i02 < ne02) { - ((device float *)dst_ptr)[0] = ((device float *)src0_ptr)[0]; - src0_ptr += ntg.x*nb00; + if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) { + x = (device const float *)(src0 + (i3 )*nb03 + (i2 )*nb02 + (i1 )*nb01 + (i0 )*nb00); } else { - ((device float *)dst_ptr)[0] = ((device float *)src1_ptr)[0]; - src1_ptr += ntg.x*nb10; + x = (device const float *)(src1 + (i3 - o[3])*nb13 + (i2 - o[2])*nb12 + (i1 - o[1])*nb11 + (i0 - o[0])*nb10); } - dst_ptr += ntg.x*nb0; + + device float * y = (device float *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + *y = *x; } } diff --git a/ggml-quants.c b/ggml-quants.c index bb01ce93c..9f864e5c4 100644 --- a/ggml-quants.c +++ b/ggml-quants.c @@ -3813,7 +3813,44 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r return; } #endif -#if defined(__ARM_NEON) +#if defined(__ARM_FEATURE_SVE) + const svbool_t ptrueh = svptrue_pat_b8(SV_VL16); + const svbool_t ptruel = svnot_b_z(svptrue_b8(), ptrueh); + + svfloat32_t sumv0 = svdup_n_f32(0.0f); + svfloat32_t sumv1 = svdup_n_f32(0.0f); + + assert(nb % 2 == 0); // TODO: handle odd nb + + for (int i = 0; i < nb; i += 2) { + const block_q4_0 * restrict x0 = &x[i + 0]; + const block_q4_0 * restrict x1 = &x[i + 1]; + const block_q8_0 * restrict y0 = &y[i + 0]; + const block_q8_0 * restrict y1 = &y[i + 1]; + + // load x + const svuint8_t qx0r = svld1rq_u8(svptrue_b8(), x0->qs); + const svuint8_t qx1r = svld1rq_u8(svptrue_b8(), x1->qs); + + // 4-bit -> 8-bit + const svint8_t qx0 = svreinterpret_s8_u8(svlsr_n_u8_m(ptruel, svand_n_u8_m(ptrueh, qx0r, 0x0F), 0x04)); + const svint8_t qx1 = svreinterpret_s8_u8(svlsr_n_u8_m(ptruel, svand_n_u8_m(ptrueh, qx1r, 0x0F), 0x04)); + + // sub 8 + const svint8_t qx0s = svsub_n_s8_x(svptrue_b8(), qx0, 8); + const svint8_t qx1s = svsub_n_s8_x(svptrue_b8(), qx1, 8); + + // load y + const svint8_t qy0 = svld1_s8(svptrue_b8(), y0->qs); + const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs); + + // dot product + sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); + sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); + } + + *s = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1)); +#elif defined(__ARM_NEON) float32x4_t sumv0 = vdupq_n_f32(0.0f); float32x4_t sumv1 = vdupq_n_f32(0.0f); @@ -5384,7 +5421,32 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r return; } #endif -#if defined(__ARM_NEON) +#if defined(__ARM_FEATURE_SVE) + svfloat32_t sumv0 = svdup_n_f32(0.0f); + svfloat32_t sumv1 = svdup_n_f32(0.0f); + + assert(nb % 2 == 0); // TODO: handle odd nb + + for (int i = 0; i < nb; i += 2) { + const block_q8_0 * restrict x0 = &x[i + 0]; + const block_q8_0 * restrict x1 = &x[i + 1]; + const block_q8_0 * restrict y0 = &y[i + 0]; + const block_q8_0 * restrict y1 = &y[i + 1]; + + // load x + const svint8_t qx0 = svld1_s8(svptrue_b8(), x0->qs); + const svint8_t qx1 = svld1_s8(svptrue_b8(), x1->qs); + + // load y + const svint8_t qy0 = svld1_s8(svptrue_b8(), y0->qs); + const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs); + + sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx0, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); + sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx1, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); + } + + *s = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1)); +#elif defined(__ARM_NEON) float32x4_t sumv0 = vdupq_n_f32(0.0f); float32x4_t sumv1 = vdupq_n_f32(0.0f); @@ -6026,6 +6088,7 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r const uint8_t * restrict q2 = x[i].qs; const int8_t * restrict q8 = y[i].qs; + const __m128i mins_and_scales = __lsx_vld((const __m128i*)x[i].scales, 0); const __m128i scales8 = __lsx_vand_v(mins_and_scales, m4); const __m128i mins8 = __lsx_vand_v(__lsx_vsrli_h(mins_and_scales, 4), m4); @@ -6745,6 +6808,8 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r for (int i = 0; i < nb; ++i) { const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const uint8_t * restrict q3 = x[i].qs; + const int8_t * restrict q8 = y[i].qs; // Set up scales memcpy(aux, x[i].scales, 12); __m128i scales128 = lsx_set_w( @@ -6766,29 +6831,32 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r int bit = 0; int is = 0; + __m256i xvbit; - const uint8_t * restrict q3 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; for (int j = 0; j < QK_K/128; ++j) { // load low 2 bits const __m256i q3bits = __lasx_xvld((const __m256i*)q3, 0); q3 += 32; + xvbit = __lasx_xvreplgr2vr_h(bit); // prepare low and high bits const __m256i q3l_0 = __lasx_xvand_v(q3bits, m3); - const __m256i q3h_0 = __lasx_xvslli_h(__lasx_xvsrli_h(__lasx_xvandn_v(hbits, __lasx_xvslli_h(mone, bit)), bit), 2); + const __m256i q3h_0 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2); ++bit; + xvbit = __lasx_xvreplgr2vr_h(bit); const __m256i q3l_1 = __lasx_xvand_v(__lasx_xvsrli_h(q3bits, 2), m3); - const __m256i q3h_1 = __lasx_xvslli_h(__lasx_xvsrli_h(__lasx_xvandn_v(hbits, __lasx_xvslli_h(mone, bit)), bit), 2); + const __m256i q3h_1 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2); ++bit; + xvbit = __lasx_xvreplgr2vr_h(bit); const __m256i q3l_2 = __lasx_xvand_v(__lasx_xvsrli_h(q3bits, 4), m3); - const __m256i q3h_2 = __lasx_xvslli_h(__lasx_xvsrli_h(__lasx_xvandn_v(hbits, __lasx_xvslli_h(mone, bit)), bit), 2); + const __m256i q3h_2 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2); ++bit; + xvbit = __lasx_xvreplgr2vr_h(bit); const __m256i q3l_3 = __lasx_xvand_v(__lasx_xvsrli_h(q3bits, 6), m3); - const __m256i q3h_3 = __lasx_xvslli_h(__lasx_xvsrli_h(__lasx_xvandn_v(hbits, __lasx_xvslli_h(mone, bit)), bit), 2); + const __m256i q3h_3 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2); ++bit; // load Q8 quants @@ -7337,6 +7405,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r *s = vec_extract(vsumf0, 0); #elif defined __loongarch_asx + GGML_UNUSED(kmask1); + GGML_UNUSED(kmask2); + GGML_UNUSED(kmask3); const __m256i m4 = __lasx_xvreplgr2vr_b(0xF); @@ -7349,6 +7420,11 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); memcpy(utmp, x[i].scales, 12); + utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4); + const uint32_t uaux = utmp[1] & kmask1; + utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4); + utmp[2] = uaux; + utmp[0] &= kmask1; const uint8_t * restrict q4 = x[i].qs; const int8_t * restrict q8 = y[i].qs; @@ -7388,16 +7464,17 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r __m256 vd = __lasx_xvreplfr2vr_s(d); acc = __lasx_xvfmadd_s(vd, __lasx_xvffint_s_w(sumi), acc); + } acc_m = __lsx_vfadd_s(acc_m, (__m128)__lsx_vpermi_w((__m128i)acc_m, (__m128i)acc_m, 0xee)); __m128i tmp1 = __lsx_vinsgr2vr_w(__lsx_vldi(0), __lsx_vpickve2gr_w((__m128i)acc_m, 1), 0); acc_m = __lsx_vfadd_s(acc_m, (__m128)tmp1); + ft_union fi; fi.i = __lsx_vpickve2gr_w(acc_m, 0); *s = hsum_float_8(acc) + fi.f ; - #else const uint8_t * scales = (const uint8_t*)&utmp[0]; @@ -7935,6 +8012,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r *s = vec_extract(vsumf0, 0); #elif defined __loongarch_asx + GGML_UNUSED(kmask1); + GGML_UNUSED(kmask2); + GGML_UNUSED(kmask3); const __m256i m4 = __lasx_xvreplgr2vr_b(0xF); const __m128i mzero = __lsx_vldi(0); @@ -7953,6 +8033,11 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); memcpy(utmp, x[i].scales, 12); + utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4); + const uint32_t uaux = utmp[1] & kmask1; + utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4); + utmp[2] = uaux; + utmp[0] &= kmask1; const __m256i mins_and_scales = lasx_extu8_16(lsx_set_w(utmp[3], utmp[2], utmp[1], utmp[0])); @@ -7971,6 +8056,7 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r __m256i sumi = __lasx_xvldi(0); int bit = 0; + __m256i xvbit; for (int j = 0; j < QK_K/64; ++j) { @@ -7979,13 +8065,15 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r const __m256i q5bits = __lasx_xvld((const __m256i*)q5, 0); q5 += 32; + xvbit = __lasx_xvreplgr2vr_h(bit++); const __m256i q5l_0 = __lasx_xvand_v(q5bits, m4); - const __m256i q5h_0 = __lasx_xvslli_h(__lasx_xvsrli_h(__lasx_xvand_v(hbits, hmask), bit++), 4); + const __m256i q5h_0 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvand_v(hbits, hmask), xvbit), 4); const __m256i q5_0 = __lasx_xvadd_b(q5l_0, q5h_0); hmask = __lasx_xvslli_h(hmask, 1); + xvbit = __lasx_xvreplgr2vr_h(bit++); const __m256i q5l_1 = __lasx_xvand_v(__lasx_xvsrli_h(q5bits, 4), m4); - const __m256i q5h_1 = __lasx_xvslli_h(__lasx_xvsrli_h(__lasx_xvand_v(hbits, hmask), bit++), 4); + const __m256i q5h_1 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvand_v(hbits, hmask), xvbit), 4); const __m256i q5_1 = __lasx_xvadd_b(q5l_1, q5h_1); hmask = __lasx_xvslli_h(hmask, 1); @@ -7999,10 +8087,12 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r p16_1 = lasx_madd_h(scale_1, p16_1); sumi = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p16_0, p16_1)); + } __m256 vd = __lasx_xvreplfr2vr_s(d); acc = __lasx_xvfmadd_s(vd, __lasx_xvffint_s_w(sumi), acc); + } *s = hsum_float_8(acc) + summs; diff --git a/ggml-rpc.cpp b/ggml-rpc.cpp index cc1d3ace1..49a20df4b 100644 --- a/ggml-rpc.cpp +++ b/ggml-rpc.cpp @@ -6,6 +6,7 @@ #include #include #include +#include #include #include #ifdef _WIN32 @@ -47,6 +48,7 @@ struct socket_t { sockfd_t fd; socket_t(sockfd_t fd) : fd(fd) {} ~socket_t() { + GGML_PRINT_DEBUG("[%s] closing socket %d\n", __func__, this->fd); #ifdef _WIN32 closesocket(this->fd); #else @@ -97,7 +99,7 @@ static ggml_guid_t ggml_backend_rpc_guid() { } struct ggml_backend_rpc_buffer_type_context { - std::shared_ptr sock; + std::string endpoint; std::string name; size_t alignment; size_t max_size; @@ -106,8 +108,6 @@ struct ggml_backend_rpc_buffer_type_context { struct ggml_backend_rpc_context { std::string endpoint; std::string name; - std::shared_ptr sock; - ggml_backend_buffer_type_t buft; }; struct ggml_backend_rpc_buffer_context { @@ -231,14 +231,13 @@ static bool recv_data(sockfd_t sockfd, void * data, size_t size) { return true; } -static bool parse_endpoint(const char * endpoint, std::string & host, int & port) { - std::string str(endpoint); - size_t pos = str.find(':'); +static bool parse_endpoint(const std::string & endpoint, std::string & host, int & port) { + size_t pos = endpoint.find(':'); if (pos == std::string::npos) { return false; } - host = str.substr(0, pos); - port = std::stoi(str.substr(pos + 1)); + host = endpoint.substr(0, pos); + port = std::stoi(endpoint.substr(pos + 1)); return true; } @@ -273,6 +272,44 @@ static bool send_rpc_cmd(const std::shared_ptr & sock, enum rpc_cmd cm // RPC client-side implementation +static std::shared_ptr get_socket(const std::string & endpoint) { + static std::mutex mutex; + std::lock_guard lock(mutex); + static std::unordered_map> sockets; + static bool initialized = false; + + auto it = sockets.find(endpoint); + if (it != sockets.end()) { + if (auto sock = it->second.lock()) { + return sock; + } + } + std::string host; + int port; + if (!parse_endpoint(endpoint, host, port)) { + return nullptr; + } +#ifdef _WIN32 + if (!initialized) { + WSADATA wsaData; + int res = WSAStartup(MAKEWORD(2, 2), &wsaData); + if (res != 0) { + return nullptr; + } + initialized = true; + } +#else + UNUSED(initialized); +#endif + auto sock = socket_connect(host.c_str(), port); + if (sock == nullptr) { + return nullptr; + } + GGML_PRINT_DEBUG("[%s] connected to %s, sockfd=%d\n", __func__, endpoint.c_str(), sock->fd); + sockets[endpoint] = sock; + return sock; +} + GGML_CALL static const char * ggml_backend_rpc_buffer_get_name(ggml_backend_buffer_t buffer) { ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context; return ctx->name.c_str(); @@ -442,7 +479,8 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_rpc_buffer_type_alloc_buffer std::vector input(input_size, 0); memcpy(input.data(), &size, sizeof(size)); std::vector output; - bool status = send_rpc_cmd(buft_ctx->sock, ALLOC_BUFFER, input, output); + auto sock = get_socket(buft_ctx->endpoint); + bool status = send_rpc_cmd(sock, ALLOC_BUFFER, input, output); GGML_ASSERT(status); GGML_ASSERT(output.size() == 2*sizeof(uint64_t)); // output serialization format: | remote_ptr (8 bytes) | remote_size (8 bytes) | @@ -453,7 +491,7 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_rpc_buffer_type_alloc_buffer if (remote_ptr != 0) { ggml_backend_buffer_t buffer = ggml_backend_buffer_init(buft, ggml_backend_rpc_buffer_interface, - new ggml_backend_rpc_buffer_context{buft_ctx->sock, {}, remote_ptr, "RPC"}, + new ggml_backend_rpc_buffer_context{sock, {}, remote_ptr, "RPC"}, remote_size); return buffer; } else { @@ -508,7 +546,7 @@ GGML_CALL static bool ggml_backend_rpc_buffer_type_supports_backend(ggml_backend } ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)buft->context; ggml_backend_rpc_context * rpc_ctx = (ggml_backend_rpc_context *)backend->context; - return buft_ctx->sock == rpc_ctx->sock; + return buft_ctx->endpoint == rpc_ctx->endpoint; } static ggml_backend_buffer_type_i ggml_backend_rpc_buffer_type_interface = { @@ -521,7 +559,6 @@ static ggml_backend_buffer_type_i ggml_backend_rpc_buffer_type_interface = { /* .is_host = */ NULL, }; - GGML_CALL static const char * ggml_backend_rpc_name(ggml_backend_t backend) { ggml_backend_rpc_context * rpc_ctx = (ggml_backend_rpc_context *)backend->context; @@ -530,16 +567,13 @@ GGML_CALL static const char * ggml_backend_rpc_name(ggml_backend_t backend) { GGML_CALL static void ggml_backend_rpc_free(ggml_backend_t backend) { ggml_backend_rpc_context * rpc_ctx = (ggml_backend_rpc_context *)backend->context; - ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)rpc_ctx->buft->context; - delete buft_ctx; - delete rpc_ctx->buft; delete rpc_ctx; delete backend; } GGML_CALL static ggml_backend_buffer_type_t ggml_backend_rpc_get_default_buffer_type(ggml_backend_t backend) { ggml_backend_rpc_context * ctx = (ggml_backend_rpc_context *)backend->context; - return ctx->buft; + return ggml_backend_rpc_buffer_type(ctx->endpoint.c_str()); } GGML_CALL static void ggml_backend_rpc_synchronize(ggml_backend_t backend) { @@ -590,7 +624,8 @@ GGML_CALL static enum ggml_status ggml_backend_rpc_graph_compute(ggml_backend_t std::vector input; serialize_graph(cgraph, input); std::vector output; - bool status = send_rpc_cmd(rpc_ctx->sock, GRAPH_COMPUTE, input, output); + auto sock = get_socket(rpc_ctx->endpoint); + bool status = send_rpc_cmd(sock, GRAPH_COMPUTE, input, output); GGML_ASSERT(status); GGML_ASSERT(output.size() == 1); return (enum ggml_status)output[0]; @@ -624,65 +659,48 @@ static ggml_backend_i ggml_backend_rpc_interface = { /* .event_synchronize = */ NULL, }; -static std::unordered_map instances; - GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint) { - ggml_backend_t backend = ggml_backend_rpc_init(endpoint); - return backend != nullptr ? ggml_backend_rpc_get_default_buffer_type(backend) : nullptr; -} - -GGML_CALL ggml_backend_t ggml_backend_rpc_init(const char * endpoint) { - std::string endpoint_str(endpoint); - if (instances.find(endpoint_str) != instances.end()) { - return instances[endpoint_str]; + static std::mutex mutex; + std::lock_guard lock(mutex); + // NOTE: buffer types are allocated and never freed; this is by design + static std::unordered_map buft_map; + auto it = buft_map.find(endpoint); + if (it != buft_map.end()) { + return it->second; } -#ifdef _WIN32 - { - WSADATA wsaData; - int res = WSAStartup(MAKEWORD(2, 2), &wsaData); - if (res != 0) { - return nullptr; - } - } -#endif - fprintf(stderr, "Connecting to %s\n", endpoint); - std::string host; - int port; - if (!parse_endpoint(endpoint, host, port)) { - return nullptr; - } - auto sock = socket_connect(host.c_str(), port); + auto sock = get_socket(endpoint); if (sock == nullptr) { return nullptr; } size_t alignment = get_alignment(sock); size_t max_size = get_max_size(sock); ggml_backend_rpc_buffer_type_context * buft_ctx = new ggml_backend_rpc_buffer_type_context { - /* .sock = */ sock, - /* .name = */ "RPC" + std::to_string(sock->fd), + /* .endpoint = */ endpoint, + /* .name = */ "RPC[" + std::string(endpoint) + "]", /* .alignment = */ alignment, - /* .max_size = */ max_size + /* .max_size = */ max_size }; ggml_backend_buffer_type_t buft = new ggml_backend_buffer_type { /* .iface = */ ggml_backend_rpc_buffer_type_interface, /* .context = */ buft_ctx }; + buft_map[endpoint] = buft; + return buft; +} +GGML_CALL ggml_backend_t ggml_backend_rpc_init(const char * endpoint) { ggml_backend_rpc_context * ctx = new ggml_backend_rpc_context { - /* .endpoint = */ endpoint, - /* .name = */ "RPC" + std::to_string(sock->fd), - /* .sock = */ sock, - /* .buft = */ buft + /* .endpoint = */ endpoint, + /* .name = */ "RPC", }; - instances[endpoint] = new ggml_backend { + ggml_backend_t backend = new ggml_backend { /* .guid = */ ggml_backend_rpc_guid(), /* .interface = */ ggml_backend_rpc_interface, /* .context = */ ctx }; - - return instances[endpoint]; + return backend; } GGML_API GGML_CALL bool ggml_backend_is_rpc(ggml_backend_t backend) { @@ -706,14 +724,13 @@ static void get_device_memory(const std::shared_ptr & sock, size_t * f } GGML_API GGML_CALL void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total) { - ggml_backend_t backend = ggml_backend_rpc_init(endpoint); - if (backend == nullptr) { + auto sock = get_socket(endpoint); + if (sock == nullptr) { *free = 0; *total = 0; return; } - ggml_backend_rpc_context * ctx = (ggml_backend_rpc_context *)backend->context; - get_device_memory(ctx->sock, free, total); + get_device_memory(sock, free, total); } // RPC server-side implementation diff --git a/ggml-sycl.cpp b/ggml-sycl.cpp index 496ec61c3..5cd97e4ff 100644 --- a/ggml-sycl.cpp +++ b/ggml-sycl.cpp @@ -2944,6 +2944,57 @@ namespace dpct using shared_memory = detail::device_memory; + template + inline T atomic_fetch_add(T *addr, T operand) { + auto atm = + sycl::atomic_ref(addr[0]); + return atm.fetch_add(operand); + } + + template + inline T1 atomic_fetch_add(T1 *addr, T2 operand) { + auto atm = + sycl::atomic_ref(addr[0]); + return atm.fetch_add(operand); + } + + template + inline T atomic_fetch_add(T *addr, T operand, + sycl::memory_order memoryOrder) { + switch (memoryOrder) { + case sycl::memory_order::relaxed: + return atomic_fetch_add(addr, operand); + case sycl::memory_order::acq_rel: + return atomic_fetch_add(addr, operand); + case sycl::memory_order::seq_cst: + return atomic_fetch_add(addr, operand); + default: + assert(false && "Invalid memory_order for atomics. Valid memory_order for " + "atomics are: sycl::memory_order::relaxed, " + "sycl::memory_order::acq_rel, sycl::memory_order::seq_cst!"); + } + } + + template + inline T1 atomic_fetch_add(T1 *addr, T2 operand, + sycl::memory_order memoryOrder) { + atomic_fetch_add(addr, operand, memoryOrder); + } + } // COPY from DPCT head files #define GGML_COMMON_DECL_SYCL @@ -2971,20 +3022,19 @@ 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. +#define VER_4VEC 130 //todo for hardward optimize. #define VER_GEN9 700 //todo for hardward optimize. #define VER_GEN12 1000000 //todo for hardward optimize. #define VER_GEN13 (VER_GEN12 + 1030) //todo for hardward optimize. #define GGML_SYCL_MAX_NODES 8192 //TODO: adapt to hardwares - -//define for XMX in Intel GPU -//TODO: currently, it's not used for XMX really. -#define SYCL_USE_XMX +#if !defined(GGML_SYCL_FORCE_MMQ) + #define SYCL_USE_XMX +#endif // max batch size to use MMQ kernels when tensor cores are available -#define XMX_MAX_BATCH_SIZE 32 +#define MMQ_MAX_BATCH_SIZE 32 #if defined(_MSC_VER) @@ -3060,6 +3110,7 @@ void ggml_sycl_get_device_description(int device, char * description, size_t d bool ggml_backend_is_sycl(ggml_backend_t backend); int ggml_backend_sycl_get_device(ggml_backend_t backend); int get_main_device(); +static bool ggml_backend_buffer_is_sycl_split(ggml_backend_buffer_t buffer); void print_ggml_tensor(const char*name, struct ggml_tensor *src); void log_tensor_with_cnt(const char* name, struct ggml_tensor * src, int stop_cnt); @@ -8830,12 +8881,11 @@ static void rope( dst[i + 1] = x0*sin_theta + x1*cos_theta; } -template +template static void rope_neox( const T * x, T * dst, int ncols, int n_dims, const int32_t * pos, float freq_scale, int p_delta_rows, - float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, float inv_ndims -, - const sycl::nd_item<3> &item_ct1) { + float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, float inv_ndims, + const float * freq_factors, const sycl::nd_item<3> &item_ct1) { const int col = 2 * (item_ct1.get_local_range(1) * item_ct1.get_group(1) + item_ct1.get_local_id(1)); @@ -8863,8 +8913,10 @@ static void rope_neox( float cur_rot = inv_ndims * ic - ib; const int p = has_pos ? pos[i2] : 0; + const float freq_factor = has_freq_facs ? freq_factors[ic/2] : 1.0f; + const float theta_base = - p * freq_scale * dpct::pow(theta_scale, col / 2.0f); + p * freq_scale * dpct::pow(theta_scale, col / 2.0f)/freq_factor; float cos_theta, sin_theta; rope_yarn(theta_base, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, &cos_theta, &sin_theta); @@ -12413,7 +12465,7 @@ static void rope_neox_sycl(const T *x, T *dst, int ncols, int n_dims, int nrows, const int32_t *pos, float freq_scale, int p_delta_rows, float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, - dpct::queue_ptr stream) { + const float * freq_factors, dpct::queue_ptr stream) { GGML_ASSERT(ncols % 2 == 0); const sycl::range<3> block_dims(1, SYCL_ROPE_BLOCK_SIZE, 1); const int num_blocks_x = (ncols + 2*SYCL_ROPE_BLOCK_SIZE - 1) / (2*SYCL_ROPE_BLOCK_SIZE); @@ -12423,38 +12475,48 @@ static void rope_neox_sycl(const T *x, T *dst, int ncols, int n_dims, int nrows, const float inv_ndims = -1.0f / n_dims; if (pos == nullptr) { - /* - DPCT1049:42: The work-group size passed to the SYCL kernel may exceed - the limit. To get the device limit, query - info::device::max_work_group_size. Adjust the work-group size if needed. - */ dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { - rope_neox(x, dst, ncols, n_dims, pos, freq_scale, - p_delta_rows, ext_factor, attn_factor, - corr_dims, theta_scale, inv_ndims, - item_ct1); - }); + if (freq_factors == nullptr) { + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + rope_neox(x, dst, ncols, n_dims, pos, freq_scale, + p_delta_rows, ext_factor, attn_factor, + corr_dims, theta_scale, inv_ndims, freq_factors, + item_ct1); + }); + } else { + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + rope_neox(x, dst, ncols, n_dims, pos, freq_scale, + p_delta_rows, ext_factor, attn_factor, + corr_dims, theta_scale, inv_ndims, freq_factors, + item_ct1); + }); + } } else { - /* - DPCT1049:43: The work-group size passed to the SYCL kernel may exceed - the limit. To get the device limit, query - info::device::max_work_group_size. Adjust the work-group size if needed. - */ dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { - rope_neox(x, dst, ncols, n_dims, pos, freq_scale, - p_delta_rows, ext_factor, attn_factor, - corr_dims, theta_scale, inv_ndims, item_ct1); - }); + if (freq_factors == nullptr) { + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + rope_neox(x, dst, ncols, n_dims, pos, freq_scale, + p_delta_rows, ext_factor, attn_factor, + corr_dims, theta_scale, inv_ndims, freq_factors, item_ct1); + }); + } else { + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + rope_neox(x, dst, ncols, n_dims, pos, freq_scale, + p_delta_rows, ext_factor, attn_factor, + corr_dims, theta_scale, inv_ndims, freq_factors, item_ct1); + }); + } } } @@ -13501,6 +13563,10 @@ inline void ggml_sycl_op_concat(const ggml_tensor *src0, const float *src0_dd, const float *src1_dd, float *dst_dd, const dpct::queue_ptr &main_stream) { +#pragma message("TODO: generalize concat kernel for dim != 2") +#pragma message(" https://github.com/ggerganov/llama.cpp/pull/7563") + int dim = dst->op_params[0]; + GGML_ASSERT(dim == 2); GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT(src1->type == GGML_TYPE_F32); @@ -13986,9 +14052,7 @@ inline void ggml_sycl_op_rope(const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, const float *src0_dd, const float *src1_dd, float *dst_dd, const dpct::queue_ptr &main_stream) { -#pragma message("TODO: implement phi3 frequency factors support") -#pragma message(" https://github.com/ggerganov/llama.cpp/pull/7225") - GGML_ASSERT(dst->src[2] == nullptr && "phi3 frequency factors not implemented yet"); + const ggml_tensor * src2 = dst->src[2]; GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16); GGML_ASSERT( dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); @@ -14014,6 +14078,7 @@ inline void ggml_sycl_op_rope(const ggml_tensor *src0, const ggml_tensor *src1, memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float)); memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float)); + const float * freq_factors = nullptr; const int32_t * pos = nullptr; if ((mode & 1) == 0) { GGML_ASSERT(src1->type == GGML_TYPE_I32); @@ -14024,6 +14089,16 @@ inline void ggml_sycl_op_rope(const ggml_tensor *src0, const ggml_tensor *src1, const bool is_neox = mode & 2; const bool is_glm = mode & 4; + if (is_neox) { + pos = (const int32_t *) src1_dd; + + if (src2 != nullptr) { + freq_factors = (const float *) src2->data; + } + } else { + GGML_ASSERT(src2 == nullptr && "TODO: freq_factors not implemented for !is_neox"); + } + rope_corr_dims corr_dims; ggml_rope_yarn_corr_dims(n_dims, n_orig_ctx, freq_base, beta_fast, beta_slow, corr_dims.v); @@ -14035,13 +14110,13 @@ inline void ggml_sycl_op_rope(const ggml_tensor *src0, const ggml_tensor *src1, if (src0->type == GGML_TYPE_F32) { rope_neox_sycl( (const float *)src0_dd, (float *)dst_dd, ne00, n_dims, nrows, pos, freq_scale, ne01, freq_base, ext_factor, - attn_factor, corr_dims, main_stream + attn_factor, corr_dims, freq_factors, main_stream ); } else if (src0->type == GGML_TYPE_F16) { rope_neox_sycl((const sycl::half *)src0_dd, (sycl::half *)dst_dd, ne00, n_dims, nrows, pos, freq_scale, ne01, freq_base, ext_factor, attn_factor, corr_dims, - main_stream); + freq_factors, main_stream); } else { GGML_ASSERT(false); } @@ -15108,7 +15183,7 @@ static void ggml_sycl_mul_mat_batched_sycl(const ggml_tensor *src0, const int64_t r2 = ne12/ne02; const int64_t r3 = ne13/ne03; - if (r2 == 1 && r3 == 1 && src0->nb[2]*src0->ne[2] == src0->nb[3] && src1->nb[2]*src1->ne[2] == src1->nb[3]) { + if (r2 == 1 && r3 == 1 && ggml_is_contiguous_2(src0) && ggml_is_contiguous_2(src1)) { // there is no broadcast and src0, src1 are contiguous across dims 2, 3 SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch( *g_sycl_handles[g_main_device], oneapi::mkl::transpose::trans, @@ -15173,6 +15248,29 @@ catch (sycl::exception const &exc) { std::exit(1); } +inline bool ggml_sycl_supports_mmq(enum ggml_type type) { + // TODO: accuracy issues in MMQ + return false; +} + +bool ggml_sycl_supports_dmmv(enum ggml_type type) { + switch (type) { + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + case GGML_TYPE_Q2_K: + case GGML_TYPE_Q3_K: + case GGML_TYPE_Q4_K: + case GGML_TYPE_Q5_K: + case GGML_TYPE_Q6_K: + case GGML_TYPE_F16: + return true; + default: + return false; + } +} static void ggml_sycl_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { const bool all_on_device = @@ -15189,75 +15287,42 @@ static void ggml_sycl_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1 } } + // check data types and tensor shapes for custom matrix multiplication kernels: + bool use_dequantize_mul_mat_vec = ggml_sycl_supports_dmmv(src0->type) + && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32 + && src0->ne[0] % GGML_SYCL_DMMV_X == 0 && src1->ne[1] == 1; + + bool use_mul_mat_vec_q = ggml_is_quantized(src0->type) + && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32 + && src1->ne[1] <= MMVQ_MAX_BATCH_SIZE; + + bool use_mul_mat_q = ggml_sycl_supports_mmq(src0->type) + && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32; + + // mmvq and mmq need the __dp4a instruction which is available for gen12+ + // Workaround in https://github.com/ggerganov/llama.cpp/commit/95f84d5ce8b449a9b16009434aca800df504a02e + use_mul_mat_q = use_mul_mat_q && (src0->type != GGML_TYPE_IQ2_XXS); #ifdef SYCL_USE_XMX - const bool use_xmx = true; -#else - const bool use_xmx = false; -#endif + use_mul_mat_q = use_mul_mat_q && (src1->ne[1] <= MMQ_MAX_BATCH_SIZE); +#endif // SYCL_USE_XMX - // debug helpers - //printf("src0: %8d %8d %8d %8d\n", src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3]); - //printf(" %8d %8d %8d %8d\n", src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3]); - //printf("src1: %8d %8d %8d %8d\n", src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3]); - //printf(" %8d %8d %8d %8d\n", src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3]); - //printf("src0 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name); - //printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name); - - if (!split && all_on_device && !use_xmx && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) { + if (!split && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) { // KQ single-batch - // GGML_SYCL_DEBUG("ggml_sycl_mul_mat_vec_p021\n"); ggml_sycl_mul_mat_vec_p021(src0, src1, dst); - } else if (!split && all_on_device && !use_xmx && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) { + } else if (!split && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) { // KQV single-batch - // GGML_SYCL_DEBUG("ggml_sycl_mul_mat_vec_nc\n"); ggml_sycl_mul_mat_vec_nc(src0, src1, dst); - } else if (!split && all_on_device && use_xmx && src0->type == GGML_TYPE_F16 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1)) { + } else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16) && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) { // KQ + KQV multi-batch - // GGML_SYCL_DEBUG("ggml_sycl_mul_mat_batched_sycl\n"); ggml_sycl_mul_mat_batched_sycl(src0, src1, dst); - } else if (src0->type == GGML_TYPE_F32) { - // GGML_SYCL_DEBUG("ggml_sycl_op_mul_mat\n"); - ggml_sycl_op_mul_mat(src0, src1, dst, ggml_sycl_op_mul_mat_sycl, false); - } else if (ggml_is_quantized(src0->type) || src0->type == GGML_TYPE_F16) { - // GGML_SYCL_DEBUG("ggml_is_quantized or GGML_TYPE_F16\n"); - if (src1->ne[1] == 1 && src0->ne[0] % GGML_SYCL_DMMV_X == 0) { -#ifdef GGML_SYCL_FORCE_DMMV - const bool use_mul_mat_vec_q = false; -#else - bool use_mul_mat_vec_q = min_compute_capability >= VER_4VEC && ggml_is_quantized(src0->type); - use_mul_mat_vec_q = use_mul_mat_vec_q || - (src0->type == GGML_TYPE_IQ2_XXS) || (src0->type == GGML_TYPE_IQ2_XS) || (src0->type == GGML_TYPE_IQ2_S) || - (src0->type == GGML_TYPE_IQ3_XXS) || (src0->type == GGML_TYPE_IQ3_S) || - (src0->type == GGML_TYPE_IQ4_NL) || (src0->type == GGML_TYPE_IQ4_XS) || - (src0->type == GGML_TYPE_IQ1_S) || (src0->type == GGML_TYPE_IQ1_M); - - -#endif // GGML_SYCL_FORCE_DMMV - - if (use_mul_mat_vec_q) { - // GGML_SYCL_DEBUG("ggml_sycl_mul_mat ggml_sycl_op_mul_mat_vec_q path\n"); - ggml_sycl_op_mul_mat(src0, src1, dst, ggml_sycl_op_mul_mat_vec_q, true); - } else { - // GGML_SYCL_DEBUG("ggml_sycl_mul_mat ggml_sycl_op_dequantize_mul_mat_vec path\n"); - ggml_sycl_op_mul_mat(src0, src1, dst, ggml_sycl_op_dequantize_mul_mat_vec, false); - } - } else { - bool use_mul_mat_q = min_compute_capability >= VER_4VEC && ggml_is_quantized(src0->type); - - if (use_xmx && min_compute_capability >= VER_GEN9 && src1->ne[1] > XMX_MAX_BATCH_SIZE) { - use_mul_mat_q = false; - } - - if (use_mul_mat_q) { - // GGML_SYCL_DEBUG("ggml_sycl_mul_mat ggml_sycl_op_mul_mat_q path\n"); - ggml_sycl_op_mul_mat(src0, src1, dst, ggml_sycl_op_mul_mat_q, true); - } else { - // GGML_SYCL_DEBUG("ggml_sycl_mul_mat ggml_sycl_op_mul_mat_sycl path\n"); - ggml_sycl_op_mul_mat(src0, src1, dst, ggml_sycl_op_mul_mat_sycl, false); - } - } + } else if (use_dequantize_mul_mat_vec) { + ggml_sycl_op_mul_mat(src0, src1, dst, ggml_sycl_op_dequantize_mul_mat_vec, false); + } else if (use_mul_mat_vec_q) { + ggml_sycl_op_mul_mat(src0, src1, dst, ggml_sycl_op_mul_mat_vec_q, true); + } else if (use_mul_mat_q) { + ggml_sycl_op_mul_mat(src0, src1, dst, ggml_sycl_op_mul_mat_q, true); } else { - GGML_ASSERT(false); + ggml_sycl_op_mul_mat(src0, src1, dst, ggml_sycl_op_mul_mat_sycl, false); } } @@ -15434,22 +15499,86 @@ static void ggml_sycl_mul_mat_id_sycl(ggml_tensor * dst) { } #endif +struct mmid_row_mapping { + int32_t i1; + int32_t i2; +}; + +__dpct_inline__ static void k_copy_src1_to_contiguous( + const char *__restrict__ src1_original, char *__restrict__ src1_contiguous, + int *__restrict__ cur_src1_row, mmid_row_mapping *__restrict__ row_mapping, + const char *__restrict ids, int64_t i02, size_t ids_nb1, size_t ids_nb0, + int64_t ne11, int64_t ne10, size_t nb11, size_t nb12, + const sycl::nd_item<3> &item_ct1, int &src1_row) { + int32_t iid1 = item_ct1.get_group(2); + int32_t id = item_ct1.get_group(1); + + const int32_t row_id_i = *(const int32_t *) (ids + iid1*ids_nb1 + id*ids_nb0); + + if (row_id_i != i02) { + return; + } + + const int64_t i11 = id % ne11; + const int64_t i12 = iid1; + + if (item_ct1.get_local_id(2) == 0) { + src1_row = + dpct::atomic_fetch_add( + cur_src1_row, 1); + row_mapping[src1_row] = {id, iid1}; + } + /* + DPCT1065:194: Consider replacing sycl::nd_item::barrier() with + sycl::nd_item::barrier(sycl::access::fence_space::local_space) for better + performance if there is no access to global memory. + */ + item_ct1.barrier(); + + const float * src1_row_original = (const float *)(src1_original + i11*nb11 + i12*nb12); + float * src1_row_contiguous = (float *)(src1_contiguous + src1_row*nb11); + +#pragma unroll + for (int i = item_ct1.get_local_id(2); i < ne10; + i += item_ct1.get_local_range(2)) { + src1_row_contiguous[i] = src1_row_original[i]; + } +} + +__dpct_inline__ static void k_copy_dst_from_contiguous( + char *__restrict__ dst_original, const char *__restrict__ dst_contiguous, + const mmid_row_mapping *__restrict__ row_mapping, int64_t ne0, size_t nb1, + size_t nb2, const sycl::nd_item<3> &item_ct1) { + int32_t i = item_ct1.get_group(2); + + const int32_t i1 = row_mapping[i].i1; + const int32_t i2 = row_mapping[i].i2; + + const float * dst_row_contiguous = (const float *)(dst_contiguous + i*nb1); + float * dst_row_original = (float *)(dst_original + i1*nb1 + i2*nb2); + +#pragma unroll + for (int j = item_ct1.get_local_id(2); j < ne0; + j += item_ct1.get_local_range(2)) { + dst_row_original[j] = dst_row_contiguous[j]; + } +} + static void ggml_sycl_mul_mat_id(const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst) try { - GGML_ASSERT(src0->backend != GGML_BACKEND_TYPE_GPU_SPLIT && - "mul_mat_id does not support split buffers"); + GGML_ASSERT(!ggml_backend_buffer_is_sycl_split(src0->buffer) && "mul_mat_id does not support split buffers"); + const ggml_tensor *ids = dst->src[2]; + GGML_TENSOR_BINARY_OP_LOCALS + const dpct::queue_ptr stream = g_syclStreams[g_main_device][0]; - const size_t nb11 = src1->nb[1]; - const size_t nb1 = dst->nb[1]; - - const int32_t id = ((int32_t *)dst->op_params)[0]; - const int32_t n_as = src0->ne[2]; + const int64_t n_as = ne02; + const int64_t n_ids = ids->ne[0]; std::vector ids_host(ggml_nbytes(ids)); - const char *ids_dev = (const char *)ids->data; + const char * ids_dev = (const char *) ids->data; SYCL_CHECK(CHECK_TRY_ERROR( stream->memcpy(ids_host.data(), ids_dev, ggml_nbytes(ids)))); @@ -15489,24 +15618,40 @@ static void ggml_sycl_mul_mat_id(const ggml_tensor *src0, src0_row.ne[2] = 1; src0_row.ne[3] = 1; - src0_row.nb[3] = src0->nb[2]; + src0_row.nb[3] = nb02; - if (src1->ne[1] == 1) { - for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) { - const int32_t row_id = - *(const int32_t *)(ids_host.data() + i01 * ids->nb[1] + - id * ids->nb[0]); + src1_row.ne[1] = 1; + src1_row.ne[2] = 1; + src1_row.ne[3] = 1; + src1_row.nb[2] = nb11; + src1_row.nb[3] = nb11; - GGML_ASSERT(row_id >= 0 && row_id < n_as); + dst_row.ne[1] = 1; + dst_row.ne[2] = 1; + dst_row.ne[3] = 1; + dst_row.nb[2] = nb1; + dst_row.nb[3] = nb1; + if (ne12 == 1) { + for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) { + for (int64_t id = 0; id < n_ids; id++) { + const int32_t i02 = *(const int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]); + GGML_ASSERT(i02 >= 0 && i02 < n_as); + + const int64_t i11 = id % ne11; + const int64_t i12 = iid1; + + const int64_t i1 = id; + const int64_t i2 = i12; src0_row_extra.data_device[g_main_device] = - src0_original + row_id * src0->nb[2]; + src0_original + i02*nb02; src1_row_extra.data_device[g_main_device] = - src1_original + i01 * src1->nb[1]; + src1_original + + i11*nb11 + i12*nb12; dst_row_extra.data_device[g_main_device] = - dst_original + i01 * dst->nb[1]; + dst_original + i1*nb1 + i2*nb2; ggml_sycl_mul_mat(&src0_row, &src1_row, &dst_row); + } } } else { sycl_pool_alloc src1_contiguous(sizeof(float)*ggml_nelements(src1)); @@ -15515,64 +15660,98 @@ static void ggml_sycl_mul_mat_id(const ggml_tensor *src0, src1_row_extra.data_device[g_main_device] = src1_contiguous.get(); dst_row_extra.data_device[g_main_device] = dst_contiguous.get(); - for (int32_t row_id = 0; row_id < n_as; ++row_id) { + for (int64_t i02 = 0; i02 < n_as; i02++) { int64_t num_src1_rows = 0; - for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) { - const int32_t row_id_i = *(const int32_t *) (ids_host.data() + i01*ids->nb[1] + id*ids->nb[0]); + for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) { + for (int64_t id = 0; id < n_ids; id++) { + const int32_t row_id_i = *(const int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]); - if (row_id_i != row_id) { - continue; + GGML_ASSERT(row_id_i >= 0 && row_id_i < n_as); + + if (row_id_i != i02) { + continue; + } + + num_src1_rows++; } - - GGML_ASSERT(row_id >= 0 && row_id < n_as); - - SYCL_CHECK(CHECK_TRY_ERROR( - stream->memcpy(src1_contiguous.get() + num_src1_rows * nb11, - src1_original + i01 * nb11, nb11))); - num_src1_rows++; } if (num_src1_rows == 0) { continue; } - src0_row_extra.data_device[g_main_device] = - src0_original + row_id * src0->nb[2]; + sycl_pool_alloc dev_cur_src1_row(1); + sycl_pool_alloc dev_row_mapping(num_src1_rows); + SYCL_CHECK(CHECK_TRY_ERROR( + stream->memset(dev_cur_src1_row.get(), 0, sizeof(int)))); + + { + sycl::range<3> block_dims(1, 1, std::min((unsigned int)ne10, 768u)); + sycl::range<3> grid_dims(1, n_ids, ids->ne[1]); + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor src1_row_acc(cgh); + + char *__restrict src1_contiguous_get = + src1_contiguous.get(); + int *__restrict dev_cur_src1_row_get = + dev_cur_src1_row.get(); + mmid_row_mapping *__restrict dev_row_mapping_get = + dev_row_mapping.get(); + size_t ids_nb_ct6 = ids->nb[1]; + size_t ids_nb_ct7 = ids->nb[0]; + + cgh.parallel_for( + sycl::nd_range<3>(grid_dims * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + k_copy_src1_to_contiguous( + src1_original, src1_contiguous_get, + dev_cur_src1_row_get, + dev_row_mapping_get, ids_dev, i02, + ids_nb_ct6, ids_nb_ct7, ne11, ne10, nb11, nb12, + item_ct1, src1_row_acc); + }); + }); + } + + src0_row_extra.data_device[g_main_device] = src0_original + i02*nb02; + + GGML_ASSERT(nb11 == sizeof(float)*ne10); + GGML_ASSERT(nb1 == sizeof(float)*ne0); src1_row.ne[1] = num_src1_rows; - dst_row.ne[1] = num_src1_rows; src1_row.nb[1] = nb11; src1_row.nb[2] = num_src1_rows*nb11; src1_row.nb[3] = num_src1_rows*nb11; + dst_row.ne[1] = num_src1_rows; dst_row.nb[1] = nb1; dst_row.nb[2] = num_src1_rows*nb1; dst_row.nb[3] = num_src1_rows*nb1; ggml_sycl_mul_mat(&src0_row, &src1_row, &dst_row); - num_src1_rows = 0; - for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) { - const int32_t row_id_i = *(const int32_t *) (ids_host.data() + i01*ids->nb[1] + id*ids->nb[0]); + { + sycl::range<3> block_dims(1, 1, std::min((unsigned int)ne0, 768u)); + sycl::range<3> grid_dims(1, 1, num_src1_rows); + stream->submit([&](sycl::handler &cgh) { + const char *__restrict dst_contiguous_get = + dst_contiguous.get(); + const mmid_row_mapping *__restrict dev_row_mapping_get = + dev_row_mapping.get(); - if (row_id_i != row_id) { - continue; - } - - GGML_ASSERT(row_id >= 0 && row_id < n_as); - - SYCL_CHECK(CHECK_TRY_ERROR(stream->memcpy( - dst_original + i01 * nb1, - dst_contiguous.get() + num_src1_rows * nb1, nb1))); - num_src1_rows++; + cgh.parallel_for( + sycl::nd_range<3>(grid_dims * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + k_copy_dst_from_contiguous(dst_original, + dst_contiguous_get, + dev_row_mapping_get, + ne0, nb1, nb2, item_ct1); + }); + }); } } } - - if (dst->backend == GGML_BACKEND_TYPE_CPU) { - SYCL_CHECK(CHECK_TRY_ERROR(stream->wait())); - } } catch (sycl::exception const &exc) { std::cerr << exc.what() << "Exception caught at file:" << __FILE__ @@ -16555,10 +16734,9 @@ GGML_CALL static const char * ggml_backend_sycl_split_buffer_get_name(ggml_backe UNUSED(buffer); } -// unused at the moment -//static bool ggml_backend_buffer_is_sycl_split(ggml_backend_buffer_t buffer) { -// return buffer->iface.get_name == ggml_backend_sycl_split_buffer_get_name; -//} +static bool ggml_backend_buffer_is_sycl_split(ggml_backend_buffer_t buffer) { + return buffer->iface.get_name == ggml_backend_sycl_split_buffer_get_name; +} GGML_CALL static void ggml_backend_sycl_split_buffer_free_buffer(ggml_backend_buffer_t buffer) { ggml_backend_sycl_split_buffer_context * ctx = (ggml_backend_sycl_split_buffer_context *)buffer->context; diff --git a/ggml-vulkan.cpp b/ggml-vulkan.cpp index 79ce1479f..92e622b04 100644 --- a/ggml-vulkan.cpp +++ b/ggml-vulkan.cpp @@ -6012,6 +6012,8 @@ static ggml_backend_buffer_type_i ggml_backend_vk_buffer_type_interface = { }; GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num) { + ggml_vk_instance_init(); + #ifdef GGML_VULKAN_DEBUG std::cerr << "ggml_backend_vk_buffer_type(" << dev_num << ")" << std::endl; #endif diff --git a/ggml.c b/ggml.c index 201f01958..264023d6b 100644 --- a/ggml.c +++ b/ggml.c @@ -60,6 +60,9 @@ typedef volatile LONG atomic_int; typedef atomic_int atomic_bool; +typedef atomic_int atomic_flag; + +#define ATOMIC_FLAG_INIT 0 static void atomic_store(atomic_int * ptr, LONG val) { InterlockedExchange(ptr, val); @@ -73,6 +76,12 @@ static LONG atomic_fetch_add(atomic_int * ptr, LONG inc) { static LONG atomic_fetch_sub(atomic_int * ptr, LONG dec) { return atomic_fetch_add(ptr, -(dec)); } +static atomic_bool atomic_flag_test_and_set(atomic_flag * ptr) { + return InterlockedExchange(ptr, 1); +} +static void atomic_flag_clear(atomic_flag * ptr) { + InterlockedExchange(ptr, 0); +} typedef HANDLE pthread_t; @@ -1567,11 +1576,11 @@ do { \ // F16 arithmetic is not supported by AVX, so we use F32 instead -#define GGML_F32Cx8 __m256 +#define GGML_F32Cx8 __m256 #define GGML_F32Cx8_ZERO (__m256)__lasx_xvldi(0) #define GGML_F32Cx8_SET1(x) (__m256)__lasx_xvreplgr2vr_w((x)) -static inline __m256 __lasx_f32cx8_load(ggml_fp16_t *x) { +static inline __m256 __lasx_f32cx8_load(const ggml_fp16_t * x) { float tmp[8]; for (int i = 0; i < 8; i++) { @@ -1580,13 +1589,14 @@ static inline __m256 __lasx_f32cx8_load(ggml_fp16_t *x) { return (__m256)__lasx_xvld(tmp, 0); } -static inline void __lasx_f32cx8_store(ggml_fp16_t *x, __m256 y) { +static inline void __lasx_f32cx8_store(ggml_fp16_t * x, __m256 y) { float arr[8]; __lasx_xvst(y, arr, 0); - for (int i = 0; i < 8; i++) + for (int i = 0; i < 8; i++) { x[i] = GGML_FP32_TO_FP16(arr[i]); + } } #define GGML_F32Cx8_LOAD(x) __lasx_f32cx8_load(x) #define GGML_F32Cx8_STORE(x, y) __lasx_f32cx8_store(x, y) @@ -1662,7 +1672,7 @@ static inline void __lasx_f32cx8_store(ggml_fp16_t *x, __m256 y) { #define GGML_F16_STEP 32 #define GGML_F16_EPR 4 -static inline __m128 __lsx_f16x4_load(ggml_fp16_t *x) { +static inline __m128 __lsx_f16x4_load(const ggml_fp16_t * x) { float tmp[4]; tmp[0] = GGML_FP16_TO_FP32(x[0]); @@ -1673,7 +1683,7 @@ static inline __m128 __lsx_f16x4_load(ggml_fp16_t *x) { return __lsx_vld(tmp, 0); } -static inline void __lsx_f16x4_store(ggml_fp16_t *x, __m128 y) { +static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) { float arr[4]; __lsx_vst(y, arr, 0); @@ -2380,32 +2390,27 @@ inline static __m512 ggml_v_expf(__m512 x) { const __m512 r = _mm512_set1_ps(0x1.8p23f); const __m512 z = _mm512_fmadd_ps(x, _mm512_set1_ps(0x1.715476p+0f), r); const __m512 n = _mm512_sub_ps(z, r); - const __m512 b = _mm512_fnmadd_ps(n, _mm512_set1_ps(0x1.7f7d1cp-20f), - _mm512_fnmadd_ps(n, _mm512_set1_ps(0x1.62e4p-1f), x)); - const __m512i e = _mm512_slli_epi32(_mm512_castps_si512(z), 23); - const __m512 k = _mm512_castsi512_ps(_mm512_add_epi32(e, _mm512_castps_si512(_mm512_set1_ps(1)))); - const __mmask16 c = _mm512_cmp_ps_mask(_mm512_abs_ps(n), _mm512_set1_ps(126), _CMP_GT_OQ); - const __m512 u = _mm512_mul_ps(b, b); - const __m512 j = _mm512_fmadd_ps(_mm512_fmadd_ps(_mm512_fmadd_ps(_mm512_set1_ps(0x1.0e4020p-7f), b, - _mm512_set1_ps(0x1.573e2ep-5f)), u, - _mm512_fmadd_ps(_mm512_set1_ps(0x1.555e66p-3f), b, - _mm512_set1_ps(0x1.fffdb6p-2f))), - u, _mm512_mul_ps(_mm512_set1_ps(0x1.ffffecp-1f), b)); - if (_mm512_kortestz(c, c)) - return _mm512_fmadd_ps(j, k, k); - const __m512i g = _mm512_and_si512( - _mm512_movm_epi32(_mm512_cmp_ps_mask(n, _mm512_setzero_ps(), _CMP_LE_OQ)), - _mm512_set1_epi32(0x82000000u)); - const __m512 s1 = - _mm512_castsi512_ps(_mm512_add_epi32(g, _mm512_set1_epi32(0x7f000000u))); - const __m512 s2 = _mm512_castsi512_ps(_mm512_sub_epi32(e, g)); + const __m512 b = + _mm512_fnmadd_ps(n, _mm512_set1_ps(0x1.7f7d1cp-20f), + _mm512_fnmadd_ps(n, _mm512_set1_ps(0x1.62e4p-1f), x)); const __mmask16 d = _mm512_cmp_ps_mask(_mm512_abs_ps(n), _mm512_set1_ps(192), _CMP_GT_OQ); - return _mm512_mask_blend_ps( - d, _mm512_mask_blend_ps( - c, _mm512_fmadd_ps(k, j, k), - _mm512_mul_ps(_mm512_fmadd_ps(s2, j, s2), s1)), - _mm512_mul_ps(s1, s1)); + const __m512 u = _mm512_mul_ps(b, b); + const __m512 j = _mm512_fmadd_ps( + _mm512_fmadd_ps(_mm512_fmadd_ps(_mm512_set1_ps(0x1.0e4020p-7f), b, + _mm512_set1_ps(0x1.573e2ep-5f)), + u, + _mm512_fmadd_ps(_mm512_set1_ps(0x1.555e66p-3f), b, + _mm512_set1_ps(0x1.fffdb6p-2f))), + u, + _mm512_fmadd_ps(_mm512_set1_ps(0x1.ffffecp-1f), b, _mm512_set1_ps(1.0F))); + const __m512 res = _mm512_scalef_ps(j, n); + if (_mm512_kortestz(d, d)) + return res; + const __m512 zero = _mm512_setzero_ps(); + const __m512 alt = _mm512_mask_blend_ps( + _mm512_cmp_ps_mask(n, zero, _CMP_LE_OQ), _mm512_set1_ps(INFINITY), zero); + return _mm512_mask_blend_ps(d, res, alt); } // computes silu x/(1+exp(-x)) in single precision vector @@ -2957,24 +2962,20 @@ struct ggml_state { // global state static struct ggml_state g_state; -static atomic_int g_state_barrier = 0; +static atomic_flag g_state_critical = ATOMIC_FLAG_INIT; // barrier via spin lock inline static void ggml_critical_section_start(void) { - int processing = atomic_fetch_add(&g_state_barrier, 1); - - while (processing > 0) { - // wait for other threads to finish - atomic_fetch_sub(&g_state_barrier, 1); - sched_yield(); // TODO: reconsider this - processing = atomic_fetch_add(&g_state_barrier, 1); + while (atomic_flag_test_and_set(&g_state_critical)) { + // spin + sched_yield(); } } // TODO: make this somehow automatically executed // some sort of "sentry" mechanism inline static void ggml_critical_section_end(void) { - atomic_fetch_sub(&g_state_barrier, 1); + atomic_flag_clear(&g_state_critical); } #if defined(__gnu_linux__) @@ -3290,7 +3291,11 @@ GGML_CALL bool ggml_is_contiguous(const struct ggml_tensor * tensor) { tensor->nb[3] == tensor->nb[2]*tensor->ne[2]; } -static inline bool ggml_is_contiguous_except_dim_1(const struct ggml_tensor * tensor) { +GGML_CALL bool ggml_is_contiguous_0(const struct ggml_tensor * tensor) { + return ggml_is_contiguous(tensor); +} + +GGML_CALL bool ggml_is_contiguous_1(const struct ggml_tensor * tensor) { static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function"); return @@ -3299,6 +3304,14 @@ static inline bool ggml_is_contiguous_except_dim_1(const struct ggml_tensor * te tensor->nb[3] == tensor->nb[2]*tensor->ne[2]; } +GGML_CALL bool ggml_is_contiguous_2(const struct ggml_tensor * tensor) { + static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function"); + + return + tensor->nb[0] == ggml_type_size(tensor->type) && + tensor->nb[3] == tensor->nb[2]*tensor->ne[2]; +} + GGML_CALL bool ggml_is_permuted(const struct ggml_tensor * tensor) { static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function"); @@ -4956,10 +4969,21 @@ struct ggml_tensor * ggml_repeat_back( // ggml_concat struct ggml_tensor * ggml_concat( - struct ggml_context* ctx, - struct ggml_tensor* a, - struct ggml_tensor* b) { - GGML_ASSERT(a->ne[0] == b->ne[0] && a->ne[1] == b->ne[1] && a->ne[3] == b->ne[3]); + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + int dim) { + GGML_ASSERT(dim >= 0 && dim < GGML_MAX_DIMS); + + int64_t ne[GGML_MAX_DIMS]; + for (int d = 0; d < GGML_MAX_DIMS; ++d) { + if (d == dim) { + ne[d] = a->ne[d] + b->ne[d]; + continue; + } + GGML_ASSERT(a->ne[d] == b->ne[d]); + ne[d] = a->ne[d]; + } bool is_node = false; @@ -4967,7 +4991,9 @@ struct ggml_tensor * ggml_concat( is_node = true; } - struct ggml_tensor * result = ggml_new_tensor_4d(ctx, a->type, a->ne[0], a->ne[1], a->ne[2] + b->ne[2], a->ne[3]); + struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, GGML_MAX_DIMS, ne); + + ggml_set_op_params_i32(result, 0, dim); result->op = GGML_OP_CONCAT; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; @@ -5087,6 +5113,7 @@ struct ggml_tensor * ggml_leaky_relu( } struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); + ggml_set_op_params(result, &negative_slope, sizeof(negative_slope)); result->op = GGML_OP_LEAKY_RELU; @@ -6452,6 +6479,16 @@ struct ggml_tensor * ggml_rope_custom_inplace( ); } +struct ggml_tensor * ggml_rope_xpos_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + int n_dims, + float base, + bool down) { + return ggml_rope_impl(ctx, a, b, NULL, n_dims, 0, 0, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, base, down, true); +} + // ggml_rope_back struct ggml_tensor * ggml_rope_back( @@ -11041,26 +11078,29 @@ static void ggml_compute_forward_concat_f32( GGML_ASSERT(nb00 == sizeof(float)); GGML_ASSERT(nb10 == sizeof(float)); + const int32_t dim = ggml_get_op_params_i32(dst, 0); + + GGML_ASSERT(dim >= 0 && dim < 4); + + int64_t o[4] = {0, 0, 0, 0}; + o[dim] = src0->ne[dim]; + + const float * x; + + // TODO: smarter multi-theading for (int i3 = 0; i3 < ne3; i3++) { for (int i2 = ith; i2 < ne2; i2 += nth) { - if (i2 < ne02) { // src0 - for (int i1 = 0; i1 < ne1; i1++) { - for (int i0 = 0; i0 < ne0; i0++) { - const float * x = (float *)((char *) src0->data + i0 * nb00 + i1 * nb01 + i2 * nb02 + i3 * nb03); - - float * y = (float *)((char *)dst->data + i0 * nb0 + i1 * nb1 + i2 * nb2 + i3 * nb3); - *y = *x; + for (int i1 = 0; i1 < ne1; i1++) { + for (int i0 = 0; i0 < ne0; i0++) { + if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) { + x = (const float *) ((const char *)src0->data + (i0 )*nb00 + (i1 )*nb01 + (i2 )*nb02 + (i3 )*nb03); + } else { + x = (const float *) ((const char *)src1->data + (i0 - o[0])*nb10 + (i1 - o[1])*nb11 + (i2 - o[2])*nb12 + (i3 - o[3])*nb13); } - } - } // src1 - else { - for (int i1 = 0; i1 < ne1; i1++) { - for (int i0 = 0; i0 < ne0; i0++) { - const float * x = (float *)((char *) src1->data + i0 * nb10 + i1 * nb11 + (i2 - ne02) * nb12 + i3 * nb13); - float * y = (float *)((char *)dst->data + i0 * nb0 + i1 * nb1 + i2 * nb2 + i3 * nb3); - *y = *x; - } + float * y = (float *)((char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3); + + *y = *x; } } } @@ -11068,8 +11108,8 @@ static void ggml_compute_forward_concat_f32( } static void ggml_compute_forward_concat( - const struct ggml_compute_params* params, - struct ggml_tensor* dst) { + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { const struct ggml_tensor * src0 = dst->src[0]; @@ -11462,8 +11502,8 @@ static void ggml_compute_forward_gelu_f32( const struct ggml_tensor * src0 = dst->src[0]; - GGML_ASSERT(ggml_is_contiguous_except_dim_1(src0)); - GGML_ASSERT(ggml_is_contiguous_except_dim_1(dst)); + GGML_ASSERT(ggml_is_contiguous_1(src0)); + GGML_ASSERT(ggml_is_contiguous_1(dst)); GGML_ASSERT(ggml_are_same_shape(src0, dst)); if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) { @@ -11525,8 +11565,8 @@ static void ggml_compute_forward_gelu_quick_f32( const struct ggml_tensor * src0 = dst->src[0]; - GGML_ASSERT(ggml_is_contiguous_except_dim_1(src0)); - GGML_ASSERT(ggml_is_contiguous_except_dim_1(dst)); + GGML_ASSERT(ggml_is_contiguous_1(src0)); + GGML_ASSERT(ggml_is_contiguous_1(dst)); GGML_ASSERT(ggml_are_same_shape(src0, dst)); if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) { @@ -11588,8 +11628,8 @@ static void ggml_compute_forward_silu_f32( const struct ggml_tensor * src0 = dst->src[0]; - GGML_ASSERT(ggml_is_contiguous_except_dim_1(src0)); - GGML_ASSERT(ggml_is_contiguous_except_dim_1(dst)); + GGML_ASSERT(ggml_is_contiguous_1(src0)); + GGML_ASSERT(ggml_is_contiguous_1(dst)); GGML_ASSERT(ggml_are_same_shape(src0, dst)); if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) { @@ -11700,9 +11740,9 @@ static void ggml_compute_forward_silu_back_f32( const struct ggml_tensor * src0 = dst->src[0]; const struct ggml_tensor * grad = dst->src[1]; - GGML_ASSERT(ggml_is_contiguous_except_dim_1(grad)); - GGML_ASSERT(ggml_is_contiguous_except_dim_1(src0)); - GGML_ASSERT(ggml_is_contiguous_except_dim_1(dst)); + GGML_ASSERT(ggml_is_contiguous_1(grad)); + GGML_ASSERT(ggml_is_contiguous_1(src0)); + GGML_ASSERT(ggml_is_contiguous_1(dst)); GGML_ASSERT(ggml_are_same_shape(src0, dst)); GGML_ASSERT(ggml_are_same_shape(src0, grad)); @@ -14400,7 +14440,7 @@ static void ggml_compute_forward_rope_f32( int ir = 0; const float theta_scale = powf(freq_base, -2.0f/n_dims); - const float inv_ndims = -1.f/n_dims; + float corr_dims[2]; ggml_rope_yarn_corr_dims(n_dims, n_orig_ctx, freq_base, beta_fast, beta_slow, corr_dims); @@ -14449,7 +14489,7 @@ static void ggml_compute_forward_rope_f32( const float cos_block_theta = cosf(block_theta); const float sin_block_theta = sinf(block_theta) * sin_sign; - theta_base *= theta_scale; + theta_base *= theta_scale; block_theta *= theta_scale; const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); @@ -14484,29 +14524,22 @@ static void ggml_compute_forward_rope_f32( dst_data[1] = x0*sin_theta*zeta + x1*cos_theta*zeta; } } else { - // TODO: this might be wrong for ne0 != n_dims - need double check - // it seems we have to rope just the first n_dims elements and do nothing with the rest - // ref: https://github.com/ml-explore/mlx/blob/dc2edc762c797e3b8de50b1dad4dc0a131691033/benchmarks/python/llama_jax_bench.py#L11-L26 - theta_base *= freq_scale; + // ref: https://github.com/jquesnelle/yarn/blob/master/scaled_rope/LlamaYaRNScaledRotaryEmbedding.py for (int64_t ic = 0; ic < ne0; ic += 2) { if (ic < n_dims) { - const int64_t ib = 0; + const int64_t i0 = ic/2; - // simplified from `(ib * n_dims + ic) * inv_ndims` - float cur_rot = inv_ndims * ic - ib; - float freq_factor = freq_factors ? freq_factors[ic/2] : 1.0f; + const float freq_factor = freq_factors ? freq_factors[i0] : 1.0f; float cos_theta, sin_theta; rope_yarn( - theta_base/freq_factor, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, + theta_base/freq_factor, freq_scale, corr_dims, ic, ext_factor, attn_factor, &cos_theta, &sin_theta ); - sin_theta *= sin_sign; + sin_theta *= sin_sign; theta_base *= theta_scale; - const int64_t i0 = ib*n_dims + ic/2; - const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); float * dst_data = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); @@ -14585,7 +14618,7 @@ static void ggml_compute_forward_rope_f16( int ir = 0; const float theta_scale = powf(freq_base, -2.0f/n_dims); - const float inv_ndims = -1.f/n_dims; + float corr_dims[2]; ggml_rope_yarn_corr_dims(n_dims, n_orig_ctx, freq_base, beta_fast, beta_slow, corr_dims); @@ -14634,7 +14667,7 @@ static void ggml_compute_forward_rope_f16( const float cos_block_theta = cosf(block_theta); const float sin_block_theta = sinf(block_theta) * sin_sign; - theta_base *= theta_scale; + theta_base *= theta_scale; block_theta *= theta_scale; const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); @@ -14665,29 +14698,22 @@ static void ggml_compute_forward_rope_f16( dst_data[1] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta); } } else { - // TODO: this might be wrong for ne0 != n_dims - need double check - // it seems we have to rope just the first n_dims elements and do nothing with the rest - // ref: https://github.com/ml-explore/mlx/blob/dc2edc762c797e3b8de50b1dad4dc0a131691033/benchmarks/python/llama_jax_bench.py#L11-L26 - theta_base *= freq_scale; + // ref: https://github.com/jquesnelle/yarn/blob/master/scaled_rope/LlamaYaRNScaledRotaryEmbedding.py for (int64_t ic = 0; ic < ne0; ic += 2) { if (ic < n_dims) { - const int64_t ib = 0; + const int64_t i0 = ic/2; - // simplified from `(ib * n_dims + ic) * inv_ndims` - float cur_rot = inv_ndims * ic - ib; - float freq_factor = freq_factors ? freq_factors[ic/2] : 1.0f; + const float freq_factor = freq_factors ? freq_factors[i0] : 1.0f; float cos_theta, sin_theta; rope_yarn( - theta_base/freq_factor, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, + theta_base/freq_factor, freq_scale, corr_dims, ic, ext_factor, attn_factor, &cos_theta, &sin_theta ); - sin_theta *= sin_sign; + sin_theta *= sin_sign; theta_base *= theta_scale; - const int64_t i0 = ib*n_dims + ic/2; - const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); @@ -23192,6 +23218,16 @@ int ggml_cpu_has_neon(void) { #endif } +int ggml_cpu_has_sve(void) { +#if defined(__ARM_FEATURE_SVE) + // TODO: Currently, SVE 256 bit is only supported. + GGML_ASSERT(svcntb() == QK8_0); + return 1; +#else + return 0; +#endif +} + int ggml_cpu_has_arm_fma(void) { #if defined(__ARM_FEATURE_FMA) return 1; @@ -23280,6 +23316,14 @@ int ggml_cpu_has_sycl(void) { #endif } +int ggml_cpu_has_rpc(void) { +#if defined(GGML_USE_RPC) + return 1; +#else + return 0; +#endif +} + int ggml_cpu_has_gpublas(void) { return ggml_cpu_has_cuda() || ggml_cpu_has_clblast() || ggml_cpu_has_vulkan() || ggml_cpu_has_kompute() || ggml_cpu_has_sycl(); diff --git a/ggml.h b/ggml.h index 5c63d825e..75b97a62f 100644 --- a/ggml.h +++ b/ggml.h @@ -274,7 +274,7 @@ #define GGML_UNREACHABLE() ((void) 0) #endif -#define GGML_N_CORES_MAX 16 +#define GGML_N_CORES_MAX 512 // used to copy the number of elements and stride in bytes of tensors into local variables. // main purpose is to reduce code duplication and improve readability. @@ -770,7 +770,6 @@ extern "C" { GGML_API enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype); GGML_API GGML_CALL bool ggml_is_transposed(const struct ggml_tensor * tensor); - GGML_API GGML_CALL bool ggml_is_contiguous(const struct ggml_tensor * tensor); GGML_API GGML_CALL bool ggml_is_permuted (const struct ggml_tensor * tensor); GGML_API GGML_CALL bool ggml_is_empty (const struct ggml_tensor * tensor); GGML_API bool ggml_is_scalar (const struct ggml_tensor * tensor); @@ -779,6 +778,11 @@ extern "C" { GGML_API bool ggml_is_3d (const struct ggml_tensor * tensor); GGML_API int ggml_n_dims (const struct ggml_tensor * tensor); // returns 1 for scalars + GGML_API GGML_CALL bool ggml_is_contiguous (const struct ggml_tensor * tensor); + GGML_API GGML_CALL bool ggml_is_contiguous_0(const struct ggml_tensor * tensor); // same as ggml_is_contiguous() + GGML_API GGML_CALL bool ggml_is_contiguous_1(const struct ggml_tensor * tensor); // contiguous for dims >= 1 + GGML_API GGML_CALL bool ggml_is_contiguous_2(const struct ggml_tensor * tensor); // contiguous for dims >= 2 + GGML_API bool ggml_are_same_shape (const struct ggml_tensor * t0, const struct ggml_tensor * t1); GGML_API bool ggml_are_same_stride(const struct ggml_tensor * t0, const struct ggml_tensor * t1); @@ -1021,12 +1025,13 @@ extern "C" { struct ggml_tensor * a, struct ggml_tensor * b); - // concat a and b on dim 2 + // concat a and b along dim // used in stable-diffusion GGML_API struct ggml_tensor * ggml_concat( struct ggml_context * ctx, struct ggml_tensor * a, - struct ggml_tensor * b); + struct ggml_tensor * b, + int dim); GGML_API struct ggml_tensor * ggml_abs( struct ggml_context * ctx, @@ -1561,6 +1566,14 @@ extern "C" { float beta_slow), "use ggml_rope_ext_inplace instead"); + struct ggml_tensor * ggml_rope_xpos_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + int n_dims, + float base, + bool down); + // compute correction dims for YaRN RoPE scaling GGML_CALL void ggml_rope_yarn_corr_dims( int n_dims, int n_orig_ctx, float freq_base, float beta_fast, float beta_slow, float dims[2]); @@ -2427,6 +2440,7 @@ extern "C" { GGML_API int ggml_cpu_has_avx512_bf16(void); GGML_API int ggml_cpu_has_fma (void); GGML_API int ggml_cpu_has_neon (void); + GGML_API int ggml_cpu_has_sve (void); GGML_API int ggml_cpu_has_arm_fma (void); GGML_API int ggml_cpu_has_metal (void); GGML_API int ggml_cpu_has_f16c (void); @@ -2441,6 +2455,7 @@ extern "C" { GGML_API int ggml_cpu_has_sse3 (void); GGML_API int ggml_cpu_has_ssse3 (void); GGML_API int ggml_cpu_has_sycl (void); + GGML_API int ggml_cpu_has_rpc (void); GGML_API int ggml_cpu_has_vsx (void); GGML_API int ggml_cpu_has_matmul_int8(void); diff --git a/ggml_vk_generate_shaders.py b/ggml_vk_generate_shaders.py index a8f7373df..7c85ca7ba 100644 --- a/ggml_vk_generate_shaders.py +++ b/ggml_vk_generate_shaders.py @@ -2670,14 +2670,12 @@ void main() { const uint i = row*p.ncols + ib*p.ndims + ic/2; const uint i2 = row/p.p_delta_rows; - const float cur_rot = p.inv_ndims * ic - ib; - const int pos = data_b[i2]; const float freq_factor = p.has_freq_facs != 0 ? data_freq_factors[ic/2] : 1.0f; const float theta_base = pos*p.freq_scale*pow(p.theta_scale, col/2.0f) / freq_factor; float cos_theta, sin_theta; - rope_yarn(theta_base, uint(cur_rot), cos_theta, sin_theta); + rope_yarn(theta_base, ic, cos_theta, sin_theta); const float x0 = float(data_a[i + 0]); const float x1 = float(data_a[i + p.ndims/2]); diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 67e23dcc1..55ec2cb5c 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -33,17 +33,21 @@ class Keys: FILE_TYPE = "general.file_type" class LLM: - VOCAB_SIZE = "{arch}.vocab_size" - CONTEXT_LENGTH = "{arch}.context_length" - EMBEDDING_LENGTH = "{arch}.embedding_length" - BLOCK_COUNT = "{arch}.block_count" - FEED_FORWARD_LENGTH = "{arch}.feed_forward_length" - USE_PARALLEL_RESIDUAL = "{arch}.use_parallel_residual" - TENSOR_DATA_LAYOUT = "{arch}.tensor_data_layout" - EXPERT_COUNT = "{arch}.expert_count" - EXPERT_USED_COUNT = "{arch}.expert_used_count" - POOLING_TYPE = "{arch}.pooling_type" - LOGIT_SCALE = "{arch}.logit_scale" + VOCAB_SIZE = "{arch}.vocab_size" + CONTEXT_LENGTH = "{arch}.context_length" + EMBEDDING_LENGTH = "{arch}.embedding_length" + BLOCK_COUNT = "{arch}.block_count" + LEADING_DENSE_BLOCK_COUNT = "{arch}.leading_dense_block_count" + FEED_FORWARD_LENGTH = "{arch}.feed_forward_length" + EXPERT_FEED_FORWARD_LENGTH = "{arch}.expert_feed_forward_length" + USE_PARALLEL_RESIDUAL = "{arch}.use_parallel_residual" + TENSOR_DATA_LAYOUT = "{arch}.tensor_data_layout" + EXPERT_COUNT = "{arch}.expert_count" + EXPERT_USED_COUNT = "{arch}.expert_used_count" + EXPERT_SHARED_COUNT = "{arch}.expert_shared_count" + EXPERT_WEIGHTS_SCALE = "{arch}.expert_weights_scale" + POOLING_TYPE = "{arch}.pooling_type" + LOGIT_SCALE = "{arch}.logit_scale" class Attention: HEAD_COUNT = "{arch}.attention.head_count" @@ -55,6 +59,8 @@ class Keys: LAYERNORM_EPS = "{arch}.attention.layer_norm_epsilon" LAYERNORM_RMS_EPS = "{arch}.attention.layer_norm_rms_epsilon" CAUSAL = "{arch}.attention.causal" + Q_LORA_RANK = "{arch}.attention.q_lora_rank" + KV_LORA_RANK = "{arch}.attention.kv_lora_rank" class Rope: DIMENSION_COUNT = "{arch}.rope.dimension_count" @@ -64,6 +70,7 @@ class Keys: SCALING_ATTN_FACTOR = "{arch}.rope.scaling.attn_factor" SCALING_ORIG_CTX_LEN = "{arch}.rope.scaling.original_context_length" SCALING_FINETUNED = "{arch}.rope.scaling.finetuned" + SCALING_YARN_LOG_MUL = "{arch}.rope.scaling.yarn_log_multiplier" class SSM: CONV_KERNEL = "{arch}.ssm.conv_kernel" @@ -139,6 +146,8 @@ class MODEL_ARCH(IntEnum): COMMAND_R = auto() DBRX = auto() OLMO = auto() + ARCTIC = auto() + DEEPSEEK2 = auto() class MODEL_TENSOR(IntEnum): @@ -167,6 +176,7 @@ class MODEL_TENSOR(IntEnum): FFN_DOWN = auto() FFN_UP = auto() FFN_ACT = auto() + FFN_NORM_EXP = auto() FFN_GATE_EXP = auto() FFN_DOWN_EXP = auto() FFN_UP_EXP = auto() @@ -183,6 +193,12 @@ class MODEL_TENSOR(IntEnum): SSM_A = auto() SSM_D = auto() SSM_OUT = auto() + ATTN_Q_A = auto() + ATTN_Q_B = auto() + ATTN_KV_A_MQA = auto() + ATTN_KV_B = auto() + ATTN_Q_A_NORM = auto() + ATTN_KV_A_NORM = auto() MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { @@ -218,6 +234,8 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.COMMAND_R: "command-r", MODEL_ARCH.DBRX: "dbrx", MODEL_ARCH.OLMO: "olmo", + MODEL_ARCH.ARCTIC: "arctic", + MODEL_ARCH.DEEPSEEK2: "deepseek2", } TENSOR_NAMES: dict[MODEL_TENSOR, str] = { @@ -251,6 +269,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = { MODEL_TENSOR.FFN_DOWN_SHEXP: "blk.{bid}.ffn_down_shexp", MODEL_TENSOR.FFN_UP_SHEXP: "blk.{bid}.ffn_up_shexp", MODEL_TENSOR.FFN_ACT: "blk.{bid}.ffn", + MODEL_TENSOR.FFN_NORM_EXP: "blk.{bid}.ffn_norm_exps", MODEL_TENSOR.FFN_GATE_EXP: "blk.{bid}.ffn_gate_exps", MODEL_TENSOR.FFN_DOWN_EXP: "blk.{bid}.ffn_down_exps", MODEL_TENSOR.FFN_UP_EXP: "blk.{bid}.ffn_up_exps", @@ -262,6 +281,12 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = { MODEL_TENSOR.SSM_A: "blk.{bid}.ssm_a", MODEL_TENSOR.SSM_D: "blk.{bid}.ssm_d", MODEL_TENSOR.SSM_OUT: "blk.{bid}.ssm_out", + MODEL_TENSOR.ATTN_Q_A: "blk.{bid}.attn_q_a", + MODEL_TENSOR.ATTN_Q_B: "blk.{bid}.attn_q_b", + MODEL_TENSOR.ATTN_KV_A_MQA: "blk.{bid}.attn_kv_a_mqa", + MODEL_TENSOR.ATTN_KV_B: "blk.{bid}.attn_kv_b", + MODEL_TENSOR.ATTN_Q_A_NORM: "blk.{bid}.attn_q_a_norm", + MODEL_TENSOR.ATTN_KV_A_NORM: "blk.{bid}.attn_kv_a_norm", } MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { @@ -732,6 +757,54 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_DOWN, MODEL_TENSOR.FFN_UP, ], + MODEL_ARCH.ARCTIC: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_ROT_EMBD, + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.FFN_NORM_EXP, + MODEL_TENSOR.FFN_GATE_EXP, + MODEL_TENSOR.FFN_DOWN_EXP, + MODEL_TENSOR.FFN_UP_EXP, + ], + MODEL_ARCH.DEEPSEEK2: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_Q_A, + MODEL_TENSOR.ATTN_Q_B, + MODEL_TENSOR.ATTN_KV_A_MQA, + MODEL_TENSOR.ATTN_KV_B, + MODEL_TENSOR.ATTN_Q_A_NORM, + MODEL_TENSOR.ATTN_KV_A_NORM, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_ROT_EMBD, + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.FFN_GATE_EXP, + MODEL_TENSOR.FFN_DOWN_EXP, + MODEL_TENSOR.FFN_UP_EXP, + MODEL_TENSOR.FFN_GATE_SHEXP, + MODEL_TENSOR.FFN_DOWN_SHEXP, + MODEL_TENSOR.FFN_UP_SHEXP, + ], # TODO } @@ -765,6 +838,10 @@ MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.ROPE_FREQS, MODEL_TENSOR.ATTN_ROT_EMBD, ], + MODEL_ARCH.DEEPSEEK2: [ + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_ROT_EMBD, + ], } # diff --git a/gguf-py/gguf/gguf_reader.py b/gguf-py/gguf/gguf_reader.py index 21b089f8a..e48bc00c3 100644 --- a/gguf-py/gguf/gguf_reader.py +++ b/gguf-py/gguf/gguf_reader.py @@ -12,6 +12,8 @@ from typing import Any, Literal, NamedTuple, TypeVar, Union import numpy as np import numpy.typing as npt +from .quants import quant_shape_to_byte_shape + if __name__ == "__main__": import sys from pathlib import Path @@ -251,6 +253,7 @@ class GGUFReader: tensor_names.add(tensor_name) ggml_type = GGMLQuantizationType(raw_dtype[0]) n_elems = int(np.prod(dims)) + np_dims = tuple(reversed(dims.tolist())) block_size, type_size = GGML_QUANT_SIZES[ggml_type] n_bytes = n_elems * type_size // block_size data_offs = int(start_offs + offset_tensor[0]) @@ -279,6 +282,7 @@ class GGUFReader: else: item_count = n_bytes item_type = np.uint8 + np_dims = quant_shape_to_byte_shape(np_dims, ggml_type) tensors.append(ReaderTensor( name = tensor_name, tensor_type = ggml_type, @@ -286,7 +290,7 @@ class GGUFReader: n_elements = n_elems, n_bytes = n_bytes, data_offset = data_offs, - data = self._get(data_offs, item_type, item_count), + data = self._get(data_offs, item_type, item_count).reshape(np_dims), field = field, )) self.tensors = tensors diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py index 8b41b54ea..b93747aff 100644 --- a/gguf-py/gguf/gguf_writer.py +++ b/gguf-py/gguf/gguf_writer.py @@ -13,7 +13,6 @@ from string import ascii_letters, digits import numpy as np from .constants import ( - GGML_QUANT_SIZES, GGUF_DEFAULT_ALIGNMENT, GGUF_MAGIC, GGUF_VERSION, @@ -26,6 +25,8 @@ from .constants import ( TokenType, ) +from .quants import quant_shape_from_byte_shape + logger = logging.getLogger(__name__) @@ -229,10 +230,7 @@ class GGUFWriter: else: dtype = raw_dtype if tensor_dtype == np.uint8: - block_size, type_size = GGML_QUANT_SIZES[raw_dtype] - if tensor_shape[-1] % type_size != 0: - raise ValueError(f"Quantized tensor row size ({tensor_shape[-1]}) is not a multiple of {dtype.name} type size ({type_size})") - tensor_shape = tuple(tensor_shape[:-1]) + (tensor_shape[-1] // type_size * block_size,) + tensor_shape = quant_shape_from_byte_shape(tensor_shape, raw_dtype) n_dims = len(tensor_shape) self.ti_data += self._pack("I", n_dims) for i in range(n_dims): @@ -376,9 +374,15 @@ class GGUFWriter: def add_block_count(self, length: int) -> None: self.add_uint32(Keys.LLM.BLOCK_COUNT.format(arch=self.arch), length) + def add_leading_dense_block_count(self, length: int) -> None: + self.add_uint32(Keys.LLM.LEADING_DENSE_BLOCK_COUNT.format(arch=self.arch), length) + def add_feed_forward_length(self, length: int) -> None: self.add_uint32(Keys.LLM.FEED_FORWARD_LENGTH.format(arch=self.arch), length) + def add_expert_feed_forward_length(self, length: int) -> None: + self.add_uint32(Keys.LLM.EXPERT_FEED_FORWARD_LENGTH.format(arch=self.arch), length) + def add_parallel_residual(self, use: bool) -> None: self.add_bool(Keys.LLM.USE_PARALLEL_RESIDUAL.format(arch=self.arch), use) @@ -409,6 +413,12 @@ class GGUFWriter: def add_expert_used_count(self, count: int) -> None: self.add_uint32(Keys.LLM.EXPERT_USED_COUNT.format(arch=self.arch), count) + def add_expert_shared_count(self, count: int) -> None: + self.add_uint32(Keys.LLM.EXPERT_SHARED_COUNT.format(arch=self.arch), count) + + def add_expert_weights_scale(self, value: float) -> None: + self.add_float32(Keys.LLM.EXPERT_WEIGHTS_SCALE.format(arch=self.arch), value) + def add_layer_norm_eps(self, value: float) -> None: self.add_float32(Keys.Attention.LAYERNORM_EPS.format(arch=self.arch), value) @@ -418,6 +428,12 @@ class GGUFWriter: def add_causal_attention(self, value: bool) -> None: self.add_bool(Keys.Attention.CAUSAL.format(arch=self.arch), value) + def add_q_lora_rank(self, length: int) -> None: + self.add_uint32(Keys.Attention.Q_LORA_RANK.format(arch=self.arch), length) + + def add_kv_lora_rank(self, length: int) -> None: + self.add_uint32(Keys.Attention.KV_LORA_RANK.format(arch=self.arch), length) + def add_pooling_type(self, value: PoolingType) -> None: self.add_uint32(Keys.LLM.POOLING_TYPE.format(arch=self.arch), value.value) @@ -442,6 +458,9 @@ class GGUFWriter: def add_rope_scaling_finetuned(self, value: bool) -> None: self.add_bool(Keys.Rope.SCALING_FINETUNED.format(arch=self.arch), value) + def add_rope_scaling_yarn_log_mul(self, value: float) -> None: + self.add_float32(Keys.Rope.SCALING_YARN_LOG_MUL.format(arch=self.arch), value) + def add_ssm_conv_kernel(self, value: int) -> None: self.add_uint32(Keys.SSM.CONV_KERNEL.format(arch=self.arch), value) diff --git a/gguf-py/gguf/quants.py b/gguf-py/gguf/quants.py index e7fc0eae3..b22eec166 100644 --- a/gguf-py/gguf/quants.py +++ b/gguf-py/gguf/quants.py @@ -1,5 +1,5 @@ from __future__ import annotations -from typing import Callable +from typing import Callable, Sequence from numpy.typing import DTypeLike @@ -9,6 +9,20 @@ from .lazy import LazyNumpyTensor import numpy as np +def quant_shape_to_byte_shape(shape: Sequence[int], quant_type: GGMLQuantizationType): + block_size, type_size = GGML_QUANT_SIZES[quant_type] + if shape[-1] % block_size != 0: + raise ValueError(f"Quantized tensor row size ({shape[-1]}) is not a multiple of {quant_type.name} block size ({block_size})") + return (*shape[:-1], shape[-1] // block_size * type_size) + + +def quant_shape_from_byte_shape(shape: Sequence[int], quant_type: GGMLQuantizationType): + block_size, type_size = GGML_QUANT_SIZES[quant_type] + if shape[-1] % type_size != 0: + raise ValueError(f"Quantized tensor bytes per row ({shape[-1]}) is not a multiple of {quant_type.name} type size ({type_size})") + return (*shape[:-1], shape[-1] // type_size * block_size) + + # same as ggml_compute_fp32_to_bf16 in ggml-impl.h def __compute_fp32_to_bf16(n: np.ndarray) -> np.ndarray: n = n.astype(np.float32, copy=False).view(np.int32) diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py index 8e1cac915..83e3c4c33 100644 --- a/gguf-py/gguf/tensor_mapping.py +++ b/gguf-py/gguf/tensor_mapping.py @@ -244,6 +244,7 @@ class TensorNameMap: "encoder.layers.{bid}.mlp.fc11", # nomic-bert "model.layers.{bid}.mlp.c_fc", # starcoder2 "encoder.layer.{bid}.mlp.gated_layers_v", # jina-bert-v2 + "model.layers.{bid}.residual_mlp.w3", # arctic ), MODEL_TENSOR.FFN_UP_EXP: ( @@ -255,6 +256,7 @@ class TensorNameMap: MODEL_TENSOR.FFN_UP_SHEXP: ( "model.layers.{bid}.mlp.shared_expert.up_proj", # qwen2moe + "model.layers.{bid}.mlp.shared_experts.up_proj", # deepseek2 ), # AWQ-activation gate @@ -272,6 +274,7 @@ class TensorNameMap: "encoder.layers.{bid}.mlp.fc12", # nomic-bert "encoder.layer.{bid}.mlp.gated_layers_w", # jina-bert-v2 "transformer.h.{bid}.mlp.linear_1", # refact + "model.layers.{bid}.residual_mlp.w1", # arctic ), MODEL_TENSOR.FFN_GATE_EXP: ( @@ -283,6 +286,7 @@ class TensorNameMap: MODEL_TENSOR.FFN_GATE_SHEXP: ( "model.layers.{bid}.mlp.shared_expert.gate_proj", # qwen2moe + "model.layers.{bid}.mlp.shared_experts.gate_proj", # deepseek2 ), # Feed-forward down @@ -306,6 +310,7 @@ class TensorNameMap: "encoder.layers.{bid}.mlp.fc2", # nomic-bert "model.layers.{bid}.mlp.c_proj", # starcoder2 "encoder.layer.{bid}.mlp.wo", # jina-bert-v2 + "model.layers.{bid}.residual_mlp.w2", # arctic ), MODEL_TENSOR.FFN_DOWN_EXP: ( @@ -317,6 +322,7 @@ class TensorNameMap: MODEL_TENSOR.FFN_DOWN_SHEXP: ( "model.layers.{bid}.mlp.shared_expert.down_proj", # qwen2moe + "model.layers.{bid}.mlp.shared_experts.down_proj", # deepseek2 ), MODEL_TENSOR.ATTN_Q_NORM: ( @@ -380,6 +386,42 @@ class TensorNameMap: "model.layers.{bid}.out_proj", "backbone.layers.{bid}.mixer.out_proj", ), + + MODEL_TENSOR.ATTN_Q_A: ( + "model.layers.{bid}.self_attn.q_a_proj", # deepseek2 + ), + + MODEL_TENSOR.ATTN_Q_B: ( + "model.layers.{bid}.self_attn.q_b_proj", # deepseek2 + ), + + MODEL_TENSOR.ATTN_KV_A_MQA: ( + "model.layers.{bid}.self_attn.kv_a_proj_with_mqa", # deepseek2 + ), + + MODEL_TENSOR.ATTN_KV_B: ( + "model.layers.{bid}.self_attn.kv_b_proj", # deepseek2 + ), + + MODEL_TENSOR.ATTN_Q_A_NORM: ( + "model.layers.{bid}.self_attn.q_a_layernorm", # deepseek2 + ), + + MODEL_TENSOR.ATTN_KV_A_NORM: ( + "model.layers.{bid}.self_attn.kv_a_layernorm", # deepseek2 + ), + } + + # architecture-specific block mappings + arch_block_mappings_cfg: dict[MODEL_ARCH, dict[MODEL_TENSOR, tuple[str, ...]]] = { + MODEL_ARCH.ARCTIC: { + MODEL_TENSOR.FFN_NORM: ( + "model.layers.{bid}.residual_layernorm", + ), + MODEL_TENSOR.FFN_NORM_EXP: ( + "model.layers.{bid}.post_attention_layernorm", + ), + }, } mapping: dict[str, tuple[MODEL_TENSOR, str]] @@ -393,12 +435,14 @@ class TensorNameMap: self.mapping[tensor_name] = (tensor, tensor_name) for key in keys: self.mapping[key] = (tensor, tensor_name) + if arch in self.arch_block_mappings_cfg: + self.block_mappings_cfg.update(self.arch_block_mappings_cfg[arch]) for bid in range(n_blocks): for tensor, keys in self.block_mappings_cfg.items(): if tensor not in MODEL_TENSORS[arch]: continue # TODO: make this configurable - n_experts = 60 + n_experts = 160 for xid in range(n_experts): tensor_name = TENSOR_NAMES[tensor].format(bid = bid, xid = xid) self.mapping[tensor_name] = (tensor, tensor_name) diff --git a/gguf-py/gguf/vocab.py b/gguf-py/gguf/vocab.py index 3ba99be4f..dc5749913 100644 --- a/gguf-py/gguf/vocab.py +++ b/gguf-py/gguf/vocab.py @@ -1,10 +1,15 @@ from __future__ import annotations +import re import logging import json import os from pathlib import Path -from typing import Any, Callable, Sequence, Mapping, Iterable +from typing import Any, Callable, Sequence, Mapping, Iterable, Protocol, ClassVar, runtime_checkable + +from sentencepiece import SentencePieceProcessor + +import gguf from .gguf_writer import GGUFWriter @@ -163,3 +168,298 @@ class SpecialVocab: for typ in self.special_token_types: self._set_special_token(typ, config.get(f'{typ}_token_id')) return True + + +@runtime_checkable +class BaseVocab(Protocol): + tokenizer_model: ClassVar[str] + name: ClassVar[str] + + +@runtime_checkable +class Vocab(BaseVocab, Protocol): + vocab_size: int + added_tokens_dict: dict[str, int] + added_tokens_list: list[str] + fname_tokenizer: Path + + def __init__(self, base_path: Path): ... + def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: ... + + +class NoVocab(BaseVocab): + tokenizer_model = "no_vocab" + name = "no_vocab" + + def __repr__(self) -> str: + return "" + + +class BpeVocab(Vocab): + tokenizer_model = "gpt2" + name = "bpe" + + def __init__(self, base_path: Path): + added_tokens: dict[str, int] = {} + + if (fname_tokenizer := base_path / 'vocab.json').exists(): + # "slow" tokenizer + with open(fname_tokenizer, encoding="utf-8") as f: + self.vocab = json.load(f) + + try: + # FIXME: Verify that added tokens here _cannot_ overlap with the main vocab. + with open(base_path / 'added_tokens.json', encoding="utf-8") as f: + added_tokens = json.load(f) + except FileNotFoundError: + pass + else: + # "fast" tokenizer + fname_tokenizer = base_path / 'tokenizer.json' + + # if this fails, FileNotFoundError propagates to caller + with open(fname_tokenizer, encoding="utf-8") as f: + tokenizer_json = json.load(f) + + tokenizer_model: dict[str, Any] = tokenizer_json['model'] + if ( + tokenizer_model['type'] != 'BPE' or tokenizer_model.get('byte_fallback', False) + or tokenizer_json['decoder']['type'] != 'ByteLevel' + ): + raise FileNotFoundError('Cannot find GPT-2 BPE tokenizer') + + self.vocab = tokenizer_model["vocab"] + + if (added := tokenizer_json.get('added_tokens')) is not None: + # Added tokens here can be duplicates of the main vocabulary. + added_tokens = {item['content']: item['id'] + for item in added + if item['content'] not in self.vocab} + + vocab_size = len(self.vocab) + expected_ids = list(range(vocab_size, vocab_size + len(added_tokens))) + actual_ids = sorted(added_tokens.values()) + if expected_ids != actual_ids: + expected_end_id = vocab_size + len(actual_ids) - 1 + raise ValueError(f"Expected the {len(actual_ids)} added token ID(s) to be sequential in the range " + f"{vocab_size} - {expected_end_id}; got {actual_ids}") + + items = sorted(added_tokens.items(), key=lambda text_idx: text_idx[1]) + self.added_tokens_dict = added_tokens + self.added_tokens_list = [text for (text, idx) in items] + self.vocab_size_base = vocab_size + self.vocab_size = self.vocab_size_base + len(self.added_tokens_list) + self.fname_tokenizer = fname_tokenizer + + def bpe_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: + reverse_vocab = {id: encoded_tok for encoded_tok, id in self.vocab.items()} + + for i, _ in enumerate(self.vocab): + yield reverse_vocab[i], 0.0, gguf.TokenType.NORMAL + + def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: + for text in self.added_tokens_list: + score = -1000.0 + yield text.encode("utf-8"), score, gguf.TokenType.CONTROL + + def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: + yield from self.bpe_tokens() + yield from self.added_tokens() + + def __repr__(self) -> str: + return f"" + + +class SentencePieceVocab(Vocab): + tokenizer_model = "llama" + name = "spm" + + def __init__(self, base_path: Path): + added_tokens: dict[str, int] = {} + if (fname_tokenizer := base_path / 'tokenizer.model').exists(): + # normal location + try: + with open(base_path / 'added_tokens.json', encoding="utf-8") as f: + added_tokens = json.load(f) + except FileNotFoundError: + pass + elif not (fname_tokenizer := base_path.parent / 'tokenizer.model').exists(): + # not found in alternate location either + raise FileNotFoundError('Cannot find tokenizer.model') + + self.sentencepiece_tokenizer = SentencePieceProcessor() + self.sentencepiece_tokenizer.LoadFromFile(str(fname_tokenizer)) + vocab_size = self.sentencepiece_tokenizer.vocab_size() + + new_tokens = {id: piece for piece, id in added_tokens.items() if id >= vocab_size} + expected_new_ids = list(range(vocab_size, vocab_size + len(new_tokens))) + actual_new_ids = sorted(new_tokens.keys()) + + if expected_new_ids != actual_new_ids: + raise ValueError(f"Expected new token IDs {expected_new_ids} to be sequential; got {actual_new_ids}") + + # Token pieces that were added to the base vocabulary. + self.added_tokens_dict = added_tokens + self.added_tokens_list = [new_tokens[id] for id in actual_new_ids] + self.vocab_size_base = vocab_size + self.vocab_size = self.vocab_size_base + len(self.added_tokens_list) + self.fname_tokenizer = fname_tokenizer + + def sentencepiece_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: + tokenizer = self.sentencepiece_tokenizer + for i in range(tokenizer.vocab_size()): + piece = tokenizer.IdToPiece(i) + text = piece.encode("utf-8") + score: float = tokenizer.GetScore(i) + + toktype = gguf.TokenType.NORMAL + if tokenizer.IsUnknown(i): + toktype = gguf.TokenType.UNKNOWN + if tokenizer.IsControl(i): + toktype = gguf.TokenType.CONTROL + + # NOTE: I think added_tokens are user defined. + # ref: https://github.com/google/sentencepiece/blob/master/src/sentencepiece_model.proto + # if tokenizer.is_user_defined(i): toktype = gguf.TokenType.USER_DEFINED + + if tokenizer.IsUnused(i): + toktype = gguf.TokenType.UNUSED + if tokenizer.IsByte(i): + toktype = gguf.TokenType.BYTE + + yield text, score, toktype + + def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: + for text in self.added_tokens_list: + score = -1000.0 + yield text.encode("utf-8"), score, gguf.TokenType.USER_DEFINED + + def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: + yield from self.sentencepiece_tokens() + yield from self.added_tokens() + + def __repr__(self) -> str: + return f"" + + +class LlamaHfVocab(Vocab): + tokenizer_model = "llama" + name = "hfft" + + def __init__(self, base_path: Path): + fname_tokenizer = base_path / 'tokenizer.json' + # if this fails, FileNotFoundError propagates to caller + with open(fname_tokenizer, encoding='utf-8') as f: + tokenizer_json = json.load(f) + + # pre-check so we know if we need transformers + tokenizer_model: dict[str, Any] = tokenizer_json['model'] + is_llama3 = ( + tokenizer_model['type'] == 'BPE' and tokenizer_model.get('ignore_merges', False) + and not tokenizer_model.get('byte_fallback', True) + ) + if is_llama3: + raise TypeError('Llama 3 must be converted with BpeVocab') + + if not is_llama3 and ( + tokenizer_model['type'] != 'BPE' or not tokenizer_model.get('byte_fallback', False) + or tokenizer_json['decoder']['type'] != 'Sequence' + ): + raise FileNotFoundError('Cannot find Llama BPE tokenizer') + + try: + from transformers import AutoTokenizer + except ImportError as e: + raise ImportError( + "To use LlamaHfVocab, please install the `transformers` package. " + "You can install it with `pip install transformers`." + ) from e + + # Allow the tokenizer to default to slow or fast versions. + # Explicitly set tokenizer to use local paths. + self.tokenizer = AutoTokenizer.from_pretrained( + base_path, + cache_dir=base_path, + local_files_only=True, + ) + assert self.tokenizer.is_fast # assume tokenizer.json is used + + # Initialize lists and dictionaries for added tokens + self.added_tokens_list = [] + self.added_tokens_dict = dict() + self.added_tokens_ids = set() + + # Process added tokens + for tok, tokidx in sorted( + self.tokenizer.get_added_vocab().items(), key=lambda x: x[1] + ): + # Only consider added tokens that are not in the base vocabulary + if tokidx >= self.tokenizer.vocab_size: + self.added_tokens_list.append(tok) + self.added_tokens_dict[tok] = tokidx + self.added_tokens_ids.add(tokidx) + + # Store special tokens and their IDs + self.specials = { + tok: self.tokenizer.get_vocab()[tok] + for tok in self.tokenizer.all_special_tokens + } + self.special_ids = set(self.tokenizer.all_special_ids) + + # Set vocabulary sizes + self.vocab_size_base = self.tokenizer.vocab_size + self.vocab_size = self.vocab_size_base + len(self.added_tokens_list) + + self.fname_tokenizer = fname_tokenizer + + def hf_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: + reverse_vocab = { + id: encoded_tok for encoded_tok, id in self.tokenizer.get_vocab().items() + } + + for token_id in range(self.vocab_size_base): + # Skip processing added tokens here + if token_id in self.added_tokens_ids: + continue + + # Convert token text to bytes + token_text = reverse_vocab[token_id].encode("utf-8") + + # Yield token text, score, and type + yield token_text, self.get_token_score(token_id), self.get_token_type( + token_id, token_text, self.special_ids # Reuse already stored special IDs + ) + + def get_token_type(self, token_id: int, token_text: bytes, special_ids: set[int]) -> gguf.TokenType: + # Special case for byte tokens + if re.fullmatch(br"<0x[0-9A-Fa-f]{2}>", token_text): + return gguf.TokenType.BYTE + + # Determine token type based on whether it's a special token + return gguf.TokenType.CONTROL if token_id in special_ids else gguf.TokenType.NORMAL + + def get_token_score(self, token_id: int) -> float: + # Placeholder for actual logic to determine the token's score + # This needs to be implemented based on specific requirements + return -1000.0 # Default score + + def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: + for text in self.added_tokens_list: + if text in self.specials: + toktype = self.get_token_type(self.specials[text], b'', self.special_ids) + score = self.get_token_score(self.specials[text]) + else: + toktype = gguf.TokenType.USER_DEFINED + score = -1000.0 + + yield text.encode("utf-8"), score, toktype + + def has_newline_token(self): + return "<0x0A>" in self.tokenizer.vocab or "\n" in self.tokenizer.vocab + + def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: + yield from self.hf_tokens() + yield from self.added_tokens() + + def __repr__(self) -> str: + return f"" diff --git a/gguf-py/scripts/gguf-new-metadata.py b/gguf-py/scripts/gguf-new-metadata.py index 63d3c5d8f..21e91180c 100755 --- a/gguf-py/scripts/gguf-new-metadata.py +++ b/gguf-py/scripts/gguf-new-metadata.py @@ -118,9 +118,7 @@ def copy_with_new_metadata(reader: gguf.GGUFReader, writer: gguf.GGUFWriter, new for tensor in reader.tensors: total_bytes += tensor.n_bytes - # Dimensions are written in reverse order, so flip them first - shape = np.flipud(tensor.shape).tolist() - writer.add_tensor_info(tensor.name, shape, tensor.data.dtype, tensor.data.nbytes, tensor.tensor_type) + writer.add_tensor_info(tensor.name, tensor.data.shape, tensor.data.dtype, tensor.data.nbytes, tensor.tensor_type) bar = tqdm(desc="Writing", total=total_bytes, unit="byte", unit_scale=True) @@ -146,6 +144,7 @@ def main() -> None: parser.add_argument("--general-description", type=str, help="The models general.description", metavar='"Description ..."') parser.add_argument("--chat-template", type=str, help="Chat template string (or JSON string containing templates)", metavar='"{% ... %} ..."') parser.add_argument("--chat-template-config", type=Path, help="Config file containing chat template(s)", metavar='tokenizer_config.json') + parser.add_argument("--pre-tokenizer", type=str, help="The models tokenizer.ggml.pre", metavar='"pre tokenizer"') parser.add_argument("--remove-metadata", action="append", type=str, help="Remove metadata (by key name) from output model", metavar='general.url') parser.add_argument("--special-token", action="append", type=str, help="Special token by value", nargs=2, metavar=(' | '.join(token_names.keys()), '""')) parser.add_argument("--special-token-by-id", action="append", type=str, help="Special token by id", nargs=2, metavar=(' | '.join(token_names.keys()), '0')) @@ -174,6 +173,9 @@ def main() -> None: if template: new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE] = MetadataDetails(gguf.GGUFValueType.STRING, template) + if args.pre_tokenizer: + new_metadata[gguf.Keys.Tokenizer.PRE] = MetadataDetails(gguf.GGUFValueType.STRING, args.pre_tokenizer) + if remove_metadata: logger.warning('*** Warning *** Warning *** Warning **') logger.warning('* Most metadata is required for a fully functional GGUF file,') diff --git a/llama.cpp b/llama.cpp index 180d1153d..cd0ba2ede 100644 --- a/llama.cpp +++ b/llama.cpp @@ -103,7 +103,7 @@ #endif #define LLAMA_MAX_NODES 8192 -#define LLAMA_MAX_EXPERTS 60 +#define LLAMA_MAX_EXPERTS 160 // // logging @@ -221,6 +221,8 @@ enum llm_arch { LLM_ARCH_COMMAND_R, LLM_ARCH_DBRX, LLM_ARCH_OLMO, + LLM_ARCH_ARCTIC, + LLM_ARCH_DEEPSEEK2, LLM_ARCH_UNKNOWN, }; @@ -257,6 +259,8 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_COMMAND_R, "command-r" }, { LLM_ARCH_DBRX, "dbrx" }, { LLM_ARCH_OLMO, "olmo" }, + { LLM_ARCH_ARCTIC, "arctic" }, + { LLM_ARCH_DEEPSEEK2, "deepseek2" }, { LLM_ARCH_UNKNOWN, "(unknown)" }, }; @@ -277,11 +281,15 @@ enum llm_kv { LLM_KV_CONTEXT_LENGTH, LLM_KV_EMBEDDING_LENGTH, LLM_KV_BLOCK_COUNT, + LLM_KV_LEADING_DENSE_BLOCK_COUNT, LLM_KV_FEED_FORWARD_LENGTH, + LLM_KV_EXPERT_FEED_FORWARD_LENGTH, LLM_KV_USE_PARALLEL_RESIDUAL, LLM_KV_TENSOR_DATA_LAYOUT, LLM_KV_EXPERT_COUNT, LLM_KV_EXPERT_USED_COUNT, + LLM_KV_EXPERT_SHARED_COUNT, + LLM_KV_EXPERT_WEIGHTS_SCALE, LLM_KV_POOLING_TYPE, LLM_KV_LOGIT_SCALE, @@ -294,6 +302,8 @@ enum llm_kv { LLM_KV_ATTENTION_LAYERNORM_EPS, LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, LLM_KV_ATTENTION_CAUSAL, + LLM_KV_ATTENTION_Q_LORA_RANK, + LLM_KV_ATTENTION_KV_LORA_RANK, LLM_KV_ROPE_DIMENSION_COUNT, LLM_KV_ROPE_FREQ_BASE, @@ -303,6 +313,7 @@ enum llm_kv { LLM_KV_ROPE_SCALING_ATTN_FACTOR, LLM_KV_ROPE_SCALING_ORIG_CTX_LEN, LLM_KV_ROPE_SCALING_FINETUNED, + LLM_KV_ROPE_SCALING_YARN_LOG_MUL, LLM_KV_SPLIT_NO, LLM_KV_SPLIT_COUNT, @@ -351,17 +362,21 @@ static const std::map LLM_KV_NAMES = { { LLM_KV_GENERAL_SOURCE_URL, "general.source.url" }, { LLM_KV_GENERAL_SOURCE_HF_REPO, "general.source.huggingface.repository" }, - { LLM_KV_VOCAB_SIZE, "%s.vocab_size" }, - { LLM_KV_CONTEXT_LENGTH, "%s.context_length" }, - { LLM_KV_EMBEDDING_LENGTH, "%s.embedding_length" }, - { LLM_KV_BLOCK_COUNT, "%s.block_count" }, - { LLM_KV_FEED_FORWARD_LENGTH, "%s.feed_forward_length" }, - { LLM_KV_USE_PARALLEL_RESIDUAL, "%s.use_parallel_residual" }, - { LLM_KV_TENSOR_DATA_LAYOUT, "%s.tensor_data_layout" }, - { LLM_KV_EXPERT_COUNT, "%s.expert_count" }, - { LLM_KV_EXPERT_USED_COUNT, "%s.expert_used_count" }, - { LLM_KV_POOLING_TYPE , "%s.pooling_type" }, - { LLM_KV_LOGIT_SCALE, "%s.logit_scale" }, + { LLM_KV_VOCAB_SIZE, "%s.vocab_size" }, + { LLM_KV_CONTEXT_LENGTH, "%s.context_length" }, + { LLM_KV_EMBEDDING_LENGTH, "%s.embedding_length" }, + { LLM_KV_BLOCK_COUNT, "%s.block_count" }, + { LLM_KV_LEADING_DENSE_BLOCK_COUNT, "%s.leading_dense_block_count" }, + { LLM_KV_FEED_FORWARD_LENGTH, "%s.feed_forward_length" }, + { LLM_KV_EXPERT_FEED_FORWARD_LENGTH, "%s.expert_feed_forward_length" }, + { LLM_KV_USE_PARALLEL_RESIDUAL, "%s.use_parallel_residual" }, + { LLM_KV_TENSOR_DATA_LAYOUT, "%s.tensor_data_layout" }, + { LLM_KV_EXPERT_COUNT, "%s.expert_count" }, + { LLM_KV_EXPERT_USED_COUNT, "%s.expert_used_count" }, + { LLM_KV_EXPERT_SHARED_COUNT, "%s.expert_shared_count" }, + { LLM_KV_EXPERT_WEIGHTS_SCALE, "%s.expert_weights_scale" }, + { LLM_KV_POOLING_TYPE , "%s.pooling_type" }, + { LLM_KV_LOGIT_SCALE, "%s.logit_scale" }, { LLM_KV_ATTENTION_HEAD_COUNT, "%s.attention.head_count" }, { LLM_KV_ATTENTION_HEAD_COUNT_KV, "%s.attention.head_count_kv" }, @@ -372,6 +387,8 @@ static const std::map LLM_KV_NAMES = { { LLM_KV_ATTENTION_LAYERNORM_EPS, "%s.attention.layer_norm_epsilon" }, { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, "%s.attention.layer_norm_rms_epsilon" }, { LLM_KV_ATTENTION_CAUSAL, "%s.attention.causal" }, + { LLM_KV_ATTENTION_Q_LORA_RANK, "%s.attention.q_lora_rank" }, + { LLM_KV_ATTENTION_KV_LORA_RANK, "%s.attention.kv_lora_rank" }, { LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" }, { LLM_KV_ROPE_FREQ_BASE, "%s.rope.freq_base" }, @@ -381,6 +398,7 @@ static const std::map LLM_KV_NAMES = { { LLM_KV_ROPE_SCALING_ATTN_FACTOR, "%s.rope.scaling.attn_factor" }, { LLM_KV_ROPE_SCALING_ORIG_CTX_LEN, "%s.rope.scaling.original_context_length" }, { LLM_KV_ROPE_SCALING_FINETUNED, "%s.rope.scaling.finetuned" }, + { LLM_KV_ROPE_SCALING_YARN_LOG_MUL, "%s.rope.scaling.yarn_log_multiplier" }, { LLM_KV_SPLIT_NO, "split.no" }, { LLM_KV_SPLIT_COUNT, "split.count" }, @@ -455,6 +473,7 @@ enum llm_tensor { LLM_TENSOR_FFN_DOWN_EXP, // split experts for backward compatibility LLM_TENSOR_FFN_GATE_EXP, LLM_TENSOR_FFN_UP_EXP, + LLM_TENSOR_FFN_NORM_EXPS, LLM_TENSOR_FFN_DOWN_EXPS, // merged experts LLM_TENSOR_FFN_GATE_EXPS, LLM_TENSOR_FFN_UP_EXPS, @@ -471,6 +490,12 @@ enum llm_tensor { LLM_TENSOR_SSM_A, LLM_TENSOR_SSM_D, LLM_TENSOR_SSM_OUT, + LLM_TENSOR_ATTN_Q_A, + LLM_TENSOR_ATTN_Q_B, + LLM_TENSOR_ATTN_KV_A_MQA, + LLM_TENSOR_ATTN_KV_B, + LLM_TENSOR_ATTN_Q_A_NORM, + LLM_TENSOR_ATTN_KV_A_NORM, }; static const std::map> LLM_TENSOR_NAMES = { @@ -1032,6 +1057,57 @@ static const std::map> LLM_TENSOR_NA { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, }, }, + { + LLM_ARCH_ARCTIC, + { + { 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_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" }, + { 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_TENSOR_FFN_NORM_EXPS, "blk.%d.ffn_norm_exps" }, + { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" }, + { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" }, + { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" }, + }, + }, + { + LLM_ARCH_DEEPSEEK2, + { + { 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_Q_A_NORM, "blk.%d.attn_q_a_norm" }, + { LLM_TENSOR_ATTN_KV_A_NORM, "blk.%d.attn_kv_a_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_Q_A, "blk.%d.attn_q_a" }, + { LLM_TENSOR_ATTN_Q_B, "blk.%d.attn_q_b" }, + { LLM_TENSOR_ATTN_KV_A_MQA, "blk.%d.attn_kv_a_mqa" }, + { LLM_TENSOR_ATTN_KV_B, "blk.%d.attn_kv_b" }, + { 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_UP, "blk.%d.ffn_up" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" }, + { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" }, + { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" }, + { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" }, + { LLM_TENSOR_FFN_GATE_INP_SHEXP, "blk.%d.ffn_gate_inp_shexp" }, + { LLM_TENSOR_FFN_GATE_SHEXP, "blk.%d.ffn_gate_shexp" }, + { LLM_TENSOR_FFN_DOWN_SHEXP, "blk.%d.ffn_down_shexp" }, + { LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" }, + }, + }, { LLM_ARCH_UNKNOWN, { @@ -1626,12 +1702,13 @@ struct llama_mlock { }; using llama_mlocks = std::vector>; -static std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token, bool special) { +// NOTE: avoid ever using this except for building the token_to_piece caches +static std::string llama_token_to_piece(const struct llama_model * model, llama_token token, bool special) { std::vector result(8, 0); - const int n_tokens = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size(), special); + const int n_tokens = llama_token_to_piece(model, token, result.data(), result.size(), special); if (n_tokens < 0) { result.resize(-n_tokens); - int check = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size(), special); + int check = llama_token_to_piece(model, token, result.data(), result.size(), special); GGML_ASSERT(check == -n_tokens); } else { @@ -1716,6 +1793,7 @@ enum e_model { MODEL_13B, MODEL_14B, MODEL_15B, + MODEL_16B, MODEL_20B, MODEL_30B, MODEL_34B, @@ -1723,6 +1801,7 @@ enum e_model { MODEL_40B, MODEL_65B, MODEL_70B, + MODEL_236B, MODEL_314B, MODEL_SMALL, MODEL_MEDIUM, @@ -1732,6 +1811,7 @@ enum e_model { MODEL_8x7B, MODEL_8x22B, MODEL_16x12B, + MODEL_10B_128x3_66B, }; static const size_t kiB = 1024; @@ -1757,6 +1837,13 @@ struct llama_hparams { uint32_t n_expert_used = 0; uint32_t n_vocab_type = 0; // for BERT-style token types + uint32_t n_layer_dense_lead = 0; + uint32_t n_lora_q = 0; + uint32_t n_lora_kv = 0; + uint32_t n_ff_exp = 0; + uint32_t n_expert_shared = 0; + float expert_weights_scale = 0.0; + float f_norm_eps; float f_norm_rms_eps; @@ -1764,6 +1851,7 @@ struct llama_hparams { float rope_freq_base_train; float rope_freq_scale_train; uint32_t n_yarn_orig_ctx; + float rope_yarn_log_mul; // for State Space Models uint32_t ssm_d_conv = 0; @@ -1797,6 +1885,12 @@ struct llama_hparams { if (this->n_expert != other.n_expert) return true; if (this->n_expert_used != other.n_expert_used) 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; + if (this->n_ff_exp != other.n_ff_exp) return true; + if (this->n_expert_shared != other.n_expert_shared) return true; + if (this->rope_finetuned != other.rope_finetuned) return true; if (this->n_yarn_orig_ctx != other.n_yarn_orig_ctx) return true; @@ -1812,6 +1906,8 @@ struct llama_hparams { if (!is_float_close(this->rope_attn_factor, other.rope_attn_factor, EPSILON)) return true; if (!is_float_close(this->rope_freq_base_train, other.rope_freq_base_train, EPSILON)) return true; if (!is_float_close(this->rope_freq_scale_train, other.rope_freq_scale_train, EPSILON)) return true; + if (!is_float_close(this->expert_weights_scale, other.expert_weights_scale, EPSILON)) return true; + if (!is_float_close(this->rope_yarn_log_mul, other.rope_yarn_log_mul, EPSILON)) return true; return false; } @@ -1887,6 +1983,8 @@ struct llama_layer { struct ggml_tensor * attn_k_norm_b; struct ggml_tensor * attn_out_norm; struct ggml_tensor * attn_out_norm_b; + struct ggml_tensor * attn_q_a_norm; + struct ggml_tensor * attn_kv_a_norm; // attention struct ggml_tensor * wq; @@ -1894,6 +1992,10 @@ struct llama_layer { struct ggml_tensor * wv; struct ggml_tensor * wo; struct ggml_tensor * wqkv; + struct ggml_tensor * wq_a; + struct ggml_tensor * wq_b; + struct ggml_tensor * wkv_a_mqa; + struct ggml_tensor * wkv_b; // attention bias struct ggml_tensor * bq; @@ -1907,6 +2009,7 @@ struct llama_layer { struct ggml_tensor * ffn_norm_b; struct ggml_tensor * layer_out_norm; struct ggml_tensor * layer_out_norm_b; + struct ggml_tensor * ffn_norm_exps; // ff struct ggml_tensor * ffn_gate; // w1 @@ -1926,8 +2029,9 @@ struct llama_layer { struct ggml_tensor * ffn_up_shexp; // ff bias - struct ggml_tensor * ffn_down_b; // b2 - struct ggml_tensor * ffn_up_b; // b3 + struct ggml_tensor * ffn_gate_b = nullptr; + struct ggml_tensor * ffn_down_b = nullptr; // b2 + struct ggml_tensor * ffn_up_b = nullptr; // b3 struct ggml_tensor * ffn_act; // mamba proj @@ -2059,7 +2163,9 @@ struct llama_vocab { std::unordered_map token_to_id; std::vector id_to_token; - std::unordered_map special_tokens_cache; + std::vector cache_special_tokens; + std::vector cache_token_to_piece; // llama_token_to_piece(special = false); + std::vector cache_token_to_piece_special; // llama_token_to_piece(special = true); std::map, int> bpe_ranks; @@ -3784,47 +3890,50 @@ static std::string llama_model_ftype_name(llama_ftype ftype) { static const char * llama_model_type_name(e_model type) { switch (type) { - case MODEL_14M: return "14M"; - case MODEL_17M: return "17M"; - case MODEL_22M: return "22M"; - case MODEL_33M: return "33M"; - case MODEL_70M: return "70M"; - case MODEL_109M: return "109M"; - case MODEL_137M: return "137M"; - case MODEL_160M: return "160M"; - case MODEL_335M: return "335M"; - case MODEL_410M: return "410M"; - case MODEL_0_5B: return "0.5B"; - case MODEL_1B: return "1B"; - case MODEL_1_4B: return "1.4B"; - case MODEL_2B: return "2B"; - case MODEL_2_8B: return "2.8B"; - case MODEL_3B: return "3B"; - case MODEL_4B: return "4B"; - case MODEL_6_9B: return "6.9B"; - case MODEL_7B: return "7B"; - case MODEL_8B: return "8B"; - case MODEL_12B: return "12B"; - case MODEL_13B: return "13B"; - case MODEL_14B: return "14B"; - case MODEL_15B: return "15B"; - case MODEL_20B: return "20B"; - case MODEL_30B: return "30B"; - case MODEL_34B: return "34B"; - case MODEL_35B: return "35B"; - case MODEL_40B: return "40B"; - case MODEL_65B: return "65B"; - case MODEL_70B: return "70B"; - case MODEL_314B: return "314B"; - case MODEL_SMALL: return "0.1B"; - case MODEL_MEDIUM: return "0.4B"; - case MODEL_LARGE: return "0.8B"; - case MODEL_XL: return "1.5B"; - case MODEL_A2_7B: return "A2.7B"; - case MODEL_8x7B: return "8x7B"; - case MODEL_8x22B: return "8x22B"; - case MODEL_16x12B: return "16x12B"; - default: return "?B"; + case MODEL_14M: return "14M"; + case MODEL_17M: return "17M"; + case MODEL_22M: return "22M"; + case MODEL_33M: return "33M"; + case MODEL_70M: return "70M"; + case MODEL_109M: return "109M"; + case MODEL_137M: return "137M"; + case MODEL_160M: return "160M"; + case MODEL_335M: return "335M"; + case MODEL_410M: return "410M"; + case MODEL_0_5B: return "0.5B"; + case MODEL_1B: return "1B"; + case MODEL_1_4B: return "1.4B"; + case MODEL_2B: return "2B"; + case MODEL_2_8B: return "2.8B"; + case MODEL_3B: return "3B"; + case MODEL_4B: return "4B"; + case MODEL_6_9B: return "6.9B"; + case MODEL_7B: return "7B"; + case MODEL_8B: return "8B"; + case MODEL_12B: return "12B"; + case MODEL_13B: return "13B"; + case MODEL_14B: return "14B"; + case MODEL_15B: return "15B"; + case MODEL_16B: return "16B"; + case MODEL_20B: return "20B"; + case MODEL_30B: return "30B"; + case MODEL_34B: return "34B"; + case MODEL_35B: return "35B"; + case MODEL_40B: return "40B"; + case MODEL_65B: return "65B"; + case MODEL_70B: return "70B"; + case MODEL_236B: return "236B"; + case MODEL_314B: return "314B"; + case MODEL_SMALL: return "0.1B"; + case MODEL_MEDIUM: return "0.4B"; + case MODEL_LARGE: return "0.8B"; + case MODEL_XL: return "1.5B"; + case MODEL_A2_7B: return "A2.7B"; + case MODEL_8x7B: return "8x7B"; + case MODEL_8x22B: return "8x22B"; + case MODEL_16x12B: return "16x12B"; + case MODEL_10B_128x3_66B: return "10B+128x3.66B"; + default: return "?B"; } } @@ -3956,7 +4065,9 @@ static void llm_load_hparams( switch (hparams.n_layer) { case 22: model.type = e_model::MODEL_1B; break; case 26: model.type = e_model::MODEL_3B; break; - case 32: model.type = hparams.n_vocab < 40000 ? e_model::MODEL_7B : e_model::MODEL_8B; break; + // granite uses a vocab with len 49152 + case 32: model.type = hparams.n_vocab == 49152 ? e_model::MODEL_3B : (hparams.n_vocab < 40000 ? e_model::MODEL_7B : e_model::MODEL_8B); break; + case 36: model.type = e_model::MODEL_8B; break; // granite 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; @@ -4226,6 +4337,8 @@ static void llm_load_hparams( case 30: model.type = e_model::MODEL_3B; break; case 32: model.type = e_model::MODEL_7B; break; case 40: model.type = e_model::MODEL_15B; break; + case 52: model.type = e_model::MODEL_20B; break; // granite + case 88: model.type = e_model::MODEL_34B; break; // granite default: model.type = e_model::MODEL_UNKNOWN; } } break; @@ -4346,6 +4459,39 @@ static void llm_load_hparams( default: model.type = e_model::MODEL_UNKNOWN; } } break; + case LLM_ARCH_ARCTIC: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + + if (hparams.n_expert == 128) { + switch (hparams.n_layer) { + case 35: model.type = e_model::MODEL_10B_128x3_66B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } else { + model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_DEEPSEEK2: + { + bool is_lite = (hparams.n_layer == 27); + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT, hparams.n_layer_dense_lead); + if (!is_lite) { + ml.get_key(LLM_KV_ATTENTION_Q_LORA_RANK, hparams.n_lora_q); + } + ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK, hparams.n_lora_kv); + ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp); + ml.get_key(LLM_KV_EXPERT_SHARED_COUNT, hparams.n_expert_shared); + ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE, hparams.expert_weights_scale); + ml.get_key(LLM_KV_ROPE_SCALING_YARN_LOG_MUL, hparams.rope_yarn_log_mul); + + switch (hparams.n_layer) { + case 27: model.type = e_model::MODEL_16B; break; + case 60: model.type = e_model::MODEL_236B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; default: (void)0; } @@ -4452,15 +4598,14 @@ static void llm_load_vocab( vocab.special_cls_id = 101; vocab.special_mask_id = 103; vocab.add_space_prefix = false; - } else { - if (tokenizer_model == "gpt2") { - vocab.type = LLAMA_VOCAB_TYPE_BPE; - } else { - LLAMA_LOG_WARN("%s: unknown tokenizer: '%s'", __func__, tokenizer_model.c_str()); - LLAMA_LOG_WARN("%s: using default tokenizer: 'llama'", __func__); - vocab.type = LLAMA_VOCAB_TYPE_SPM; - return; + } else if (tokenizer_model == "gpt2") { + vocab.type = LLAMA_VOCAB_TYPE_BPE; + + const int add_space_prefix_keyidx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_ADD_PREFIX).c_str()); + if (add_space_prefix_keyidx != -1) { + vocab.add_space_prefix = gguf_get_val_bool(ctx, add_space_prefix_keyidx); } + // read bpe merges and populate bpe ranks const int merges_keyidx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_MERGES).c_str()); if (merges_keyidx == -1) { @@ -4494,6 +4639,8 @@ static void llm_load_vocab( vocab.special_pad_id = -1; vocab.special_cls_id = -1; vocab.special_mask_id = -1; + } else { + throw std::runtime_error(format("unknown tokenizer: '%s'", tokenizer_model.c_str())); } // for now, only BPE models have pre-tokenizers @@ -4555,6 +4702,9 @@ static void llm_load_vocab( } else if ( tokenizer_pre == "dbrx") { vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DBRX; + } else if ( + tokenizer_pre == "smaug-bpe") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_SMAUG; } else { throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str())); } @@ -4683,97 +4833,40 @@ static void llm_load_vocab( // build special tokens cache { - // TODO: It is unclear (to me) at this point, whether special tokes are guaranteed to be of a deterministic type, - // and will always be correctly labeled in 'added_tokens.json' etc. - // The assumption is, since special tokens aren't meant to be exposed to end user, they are designed - // to be unmatchable by the tokenizer, therefore tokens from the vocab, which are unmatchable by the tokenizer - // are special tokens. - // From testing, this appears to correlate 1:1 with special tokens. - // - - // Counting special tokens and verifying in only one direction - // is sufficient to detect difference in those two sets. - // - uint32_t special_tokens_count_by_type = 0; - uint32_t special_tokens_count_from_verification = 0; - - bool special_tokens_definition_mismatch = false; - - for (const auto & t : vocab.token_to_id) { - const auto & token = t.first; - const auto & id = t.second; - - // Count all non-normal tokens in the vocab while iterating + for (llama_vocab::id id = 0; id < (llama_vocab::id)n_vocab; ++id) { if (vocab.id_to_token[id].type != LLAMA_TOKEN_TYPE_NORMAL) { - special_tokens_count_by_type++; - } - - // Skip single character tokens - if (token.length() > 1) { - bool is_tokenizable = false; - - // Split token string representation in two, in all possible ways - // and check if both halves can be matched to a valid token - for (unsigned i = 1; i < token.length();) { - const auto left = token.substr(0, i); - const auto right = token.substr(i); - - // check if we didnt partition in the middle of a utf sequence - auto utf = utf8_len(left.at(left.length() - 1)); - - if (utf == 1) { - if (vocab.token_to_id.find(left) != vocab.token_to_id.end() && - vocab.token_to_id.find(right) != vocab.token_to_id.end() ) { - is_tokenizable = true; - break; - } - i++; - } else { - // skip over the rest of multibyte utf sequence - i += utf - 1; - } - } - - if (!is_tokenizable) { - // Some tokens are multibyte, but they are utf sequences with equivalent text length of 1 - // it's faster to re-filter them here, since there are way less candidates now - - // Calculate a total "utf" length of a token string representation - size_t utf8_str_len = 0; - for (unsigned i = 0; i < token.length();) { - utf8_str_len++; - i += utf8_len(token.at(i)); - } - - // And skip the ones which are one character - if (utf8_str_len > 1) { - // At this point what we have left are special tokens only - vocab.special_tokens_cache[token] = id; - - // Count manually found special tokens - special_tokens_count_from_verification++; - - // If this manually found special token is not marked as such, flag a mismatch - if (vocab.id_to_token[id].type == LLAMA_TOKEN_TYPE_NORMAL) { - special_tokens_definition_mismatch = true; - } - } - } + vocab.cache_special_tokens.push_back(id); } } - if (special_tokens_definition_mismatch || special_tokens_count_from_verification != special_tokens_count_by_type) { - LLAMA_LOG_WARN("%s: mismatch in special tokens definition ( %u/%zu vs %u/%zu ).\n", - __func__, - special_tokens_count_from_verification, vocab.id_to_token.size(), - special_tokens_count_by_type, vocab.id_to_token.size() - ); - } else { - LLAMA_LOG_INFO("%s: special tokens definition check successful ( %u/%zu ).\n", - __func__, - special_tokens_count_from_verification, vocab.id_to_token.size() - ); + std::sort( vocab.cache_special_tokens.begin(), vocab.cache_special_tokens.end(), + [&] (const llama_vocab::id a, const llama_vocab::id b) { + return vocab.id_to_token[a].text.size() > vocab.id_to_token[b].text.size(); + } + ); + + LLAMA_LOG_INFO("%s: special tokens cache size = %u\n", __func__, (uint32_t)vocab.cache_special_tokens.size()); + } + + // build token to piece caches + { + size_t size_cache = 0; + + std::vector cache_token_to_piece (n_vocab); + std::vector cache_token_to_piece_special(n_vocab); + + for (uint32_t id = 0; id < n_vocab; ++id) { + cache_token_to_piece[id] = llama_token_to_piece(&model, id, false); + cache_token_to_piece_special[id] = llama_token_to_piece(&model, id, true); + + size_cache += cache_token_to_piece[id].size(); + size_cache += cache_token_to_piece_special[id].size(); } + + std::swap(vocab.cache_token_to_piece, cache_token_to_piece); + std::swap(vocab.cache_token_to_piece_special, cache_token_to_piece_special); + + LLAMA_LOG_INFO("%s: token to piece cache size = %.4f MB\n", __func__, size_cache / 1024.0 / 1024.0); } } @@ -4854,6 +4947,16 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) { if (vocab.special_suffix_id != -1) { LLAMA_LOG_INFO( "%s: SUF token = %d '%s'\n", __func__, vocab.special_suffix_id, vocab.id_to_token[vocab.special_suffix_id].text.c_str() ); } if (vocab.special_middle_id != -1) { LLAMA_LOG_INFO( "%s: MID token = %d '%s'\n", __func__, vocab.special_middle_id, vocab.id_to_token[vocab.special_middle_id].text.c_str() ); } if (vocab.special_eot_id != -1) { LLAMA_LOG_INFO( "%s: EOT token = %d '%s'\n", __func__, vocab.special_eot_id, vocab.id_to_token[vocab.special_eot_id].text.c_str() ); } + + if (model.arch == LLM_ARCH_DEEPSEEK2) { + LLAMA_LOG_INFO("%s: n_layer_dense_lead = %d\n", __func__, hparams.n_layer_dense_lead); + LLAMA_LOG_INFO("%s: n_lora_q = %d\n", __func__, hparams.n_lora_q); + LLAMA_LOG_INFO("%s: n_lora_kv = %d\n", __func__, hparams.n_lora_kv); + LLAMA_LOG_INFO("%s: n_ff_exp = %d\n", __func__, hparams.n_ff_exp); + LLAMA_LOG_INFO("%s: n_expert_shared = %d\n", __func__, hparams.n_expert_shared); + LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n", __func__, hparams.expert_weights_scale); + LLAMA_LOG_INFO("%s: rope_yarn_log_mul = %.4f\n", __func__, hparams.rope_yarn_log_mul); + } } // Returns false if cancelled by progress_callback @@ -5010,8 +5113,6 @@ static bool llm_load_tensors( throw std::runtime_error("model has expert layers but no expert layers are used"); } - GGML_ASSERT(n_embd_gqa == n_embd_k_gqa); - 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); @@ -5065,6 +5166,11 @@ static bool llm_load_tensors( 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}); + + // optional MLP bias + layer.ffn_gate_b = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.ffn_down_b = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.ffn_up_b = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED); } else { layer.ffn_gate_inp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}); @@ -6132,6 +6238,110 @@ static bool llm_load_tensors( layer.ffn_up_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}); } } break; + case LLM_ARCH_ARCTIC: + { + 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 = 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 = 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_embd}); + layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_embd, n_embd}); + layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, 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_norm_exps = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM_EXPS, "weight", i), {n_embd}); + layer.ffn_gate_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff, n_expert}, false); + 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}); + } + } break; + case LLM_ARCH_DEEPSEEK2: + { + 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; + + 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 = 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}); + 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}); + } else { + 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.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) { + 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); + + // 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}); + 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 + 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}); + } + } + } break; default: throw std::runtime_error("unknown architecture"); } @@ -6586,6 +6796,8 @@ static struct ggml_tensor * llm_build_moe_ffn( int64_t n_expert_used, llm_ffn_op_type type_op, bool norm_w, + bool scale_w, + float w_scale, const llm_build_cb & cb, int il) { int64_t n_embd = cur->ne[0]; @@ -6617,6 +6829,10 @@ static struct ggml_tensor * llm_build_moe_ffn( weights = ggml_reshape_3d(ctx, weights, 1, n_expert_used, n_tokens); } + if (scale_w) { + weights = ggml_scale(ctx, weights, w_scale); + cb(weights, "ffn_moe_weights_scaled", il); + } cur = ggml_reshape_3d(ctx, cur, n_embd, 1, n_tokens); ggml_tensor * up = ggml_mul_mat_id(ctx, up_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens] @@ -7227,9 +7443,9 @@ struct llm_build_context { cb(cur, "ffn_norm", il); cur = llm_build_ffn(ctx0, cur, - model.layers[il].ffn_up, NULL, - model.layers[il].ffn_gate, NULL, - model.layers[il].ffn_down, NULL, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, + model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, LLM_FFN_SILU, LLM_FFN_PAR, cb, il); cb(cur, "ffn_out", il); @@ -7247,6 +7463,7 @@ struct llm_build_context { model.layers[il].ffn_down_exps, n_expert, n_expert_used, LLM_FFN_SILU, true, + false, 0.0, cb, il); cb(cur, "ffn_moe_out", il); } @@ -7728,6 +7945,7 @@ struct llm_build_context { model.layers[il].ffn_down_exps, n_expert, n_expert_used, LLM_FFN_GELU, true, + false, 0.0, cb, il); cb(cur, "ffn_moe_out", il); @@ -7871,6 +8089,7 @@ struct llm_build_context { model.layers[il].ffn_down_exps, n_expert, n_expert_used, LLM_FFN_SILU, true, + false, 0.0, cb, il); cb(cur, "ffn_moe_out", il); @@ -9009,6 +9228,7 @@ struct llm_build_context { model.layers[il].ffn_down_exps, n_expert, n_expert_used, LLM_FFN_SILU, false, + false, 0.0, cb, il); cb(cur, "ffn_moe_out", il); @@ -10793,6 +11013,374 @@ struct llm_build_context { return gf; } + + struct ggml_cgraph * build_arctic() { + 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; + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + 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) { + 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 + { + // compute Q and K and RoPE them + 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); + + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, + n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, + n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", 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(); + 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, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "ffn_norm", il); + + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up, NULL, + model.layers[il].ffn_gate, NULL, + model.layers[il].ffn_down, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, cb, il); + cb(cur, "ffn_out", il); + + struct ggml_tensor * ffn_out = ggml_add(ctx0, cur, ffn_inp); + cb(ffn_out, "ffn_out", il); + + // MoE + cur = llm_build_norm(ctx0, inpSA, hparams, + model.layers[il].ffn_norm_exps, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "ffn_norm_exps", il); + + cur = llm_build_moe_ffn(ctx0, cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + n_expert, n_expert_used, + LLM_FFN_SILU, true, + false, 0.0, + cb, il); + cb(cur, "ffn_moe_out", il); + + cur = ggml_add(ctx0, cur, ffn_out); + 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; + + 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_deepseek2() { + 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; + + bool is_lite = (hparams.n_layer == 27); + + // We have to pre-scale kq_scale and attn_factor to make the YaRN RoPE work correctly. + // See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation. + const float mscale = attn_factor * (1.0f + hparams.rope_yarn_log_mul * logf(1.0f / freq_scale)); + const float kq_scale = 1.0f*mscale*mscale/sqrtf(float(hparams.n_embd_head_k)); + const float attn_factor_scaled = 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale)); + + 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 kv_lora_rank = hparams.n_lora_kv; + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + // {n_embd, n_tokens} + 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) { + 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 * q = NULL; + if (!is_lite) { + // {n_embd, q_lora_rank} * {n_embd, n_tokens} -> {q_lora_rank, n_tokens} + q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur); + cb(q, "q", il); + + q = llm_build_norm(ctx0, q, hparams, + model.layers[il].attn_q_a_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(q, "q", il); + + // {q_lora_rank, n_head * hparams.n_embd_head_k} * {q_lora_rank, n_tokens} -> {n_head * hparams.n_embd_head_k, n_tokens} + q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q); + cb(q, "q", il); + } else { + q = ggml_mul_mat(ctx0, model.layers[il].wq, cur); + cb(q, "q", il); + } + + // split into {n_head * n_embd_head_qk_nope, n_tokens} + struct ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens, + ggml_row_size(q->type, hparams.n_embd_head_k), + ggml_row_size(q->type, hparams.n_embd_head_k * n_head), + 0); + cb(q_nope, "q_nope", il); + + // and {n_head * n_embd_head_qk_rope, n_tokens} + struct ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens, + ggml_row_size(q->type, hparams.n_embd_head_k), + ggml_row_size(q->type, hparams.n_embd_head_k * n_head), + ggml_row_size(q->type, n_embd_head_qk_nope)); + cb(q_pe, "q_pe", il); + + // {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens} + struct ggml_tensor * kv_pe_compresseed = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur); + cb(kv_pe_compresseed, "kv_pe_compresseed", il); + + // split into {kv_lora_rank, n_tokens} + struct ggml_tensor * kv_compressed = ggml_view_2d(ctx0, kv_pe_compresseed, kv_lora_rank, n_tokens, + kv_pe_compresseed->nb[1], + 0); + cb(kv_compressed, "kv_compressed", il); + + // and {n_embd_head_qk_rope, n_tokens} + struct ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_pe_compresseed, n_embd_head_qk_rope, 1, n_tokens, + kv_pe_compresseed->nb[1], + kv_pe_compresseed->nb[1], + ggml_row_size(kv_pe_compresseed->type, kv_lora_rank)); + cb(k_pe, "k_pe", il); + + kv_compressed = ggml_cont(ctx0, kv_compressed); // TODO: the CUDA backend does not support non-contiguous norm + kv_compressed = llm_build_norm(ctx0, kv_compressed, hparams, + model.layers[il].attn_kv_a_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(kv_compressed, "kv_compressed", il); + + // {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)} * {kv_lora_rank, n_tokens} -> {n_head * (n_embd_head_qk_nope + n_embd_head_v), n_tokens} + struct ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_compressed); + cb(kv, "kv", il); + + // split into {n_head * n_embd_head_qk_nope, n_tokens} + struct ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens, + ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v), + ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)), + 0); + cb(k_nope, "k_nope", il); + + // and {n_head * n_embd_head_v, n_tokens} + struct ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens, + ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)), + ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)*n_head), + ggml_row_size(kv->type, (n_embd_head_qk_nope))); + cb(v_states, "v_states", il); + + v_states = ggml_cont(ctx0, v_states); + cb(v_states, "v_states", il); + + v_states = ggml_view_2d(ctx0, v_states, hparams.n_embd_head_v * n_head, n_tokens, + ggml_row_size(kv->type, hparams.n_embd_head_v * n_head), + 0); + cb(v_states, "v_states", il); + + q_pe = ggml_cont(ctx0, q_pe); // TODO: the CUDA backend does not support non-contiguous RoPE + q_pe = ggml_rope_ext( + ctx0, q_pe, inp_pos, nullptr, + n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor_scaled, beta_fast, beta_slow + ); + cb(q_pe, "q_pe", il); + + // shared RoPE key + k_pe = ggml_cont(ctx0, k_pe); // TODO: the CUDA backend does not support non-contiguous RoPE + k_pe = ggml_rope_ext( + ctx0, k_pe, inp_pos, nullptr, + n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor_scaled, beta_fast, beta_slow + ); + cb(k_pe, "k_pe", il); + + struct ggml_tensor * q_states = ggml_concat(ctx0, q_nope, q_pe, 0); + cb(q_states, "q_states", il); + + struct ggml_tensor * k_states = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0); + cb(k_states, "k_states", il); + + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, + model.layers[il].wo, NULL, + k_states, v_states, q_states, KQ_mask, n_tokens, kv_head, n_kv, kq_scale, cb, 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); + + if ((uint32_t) il < hparams.n_layer_dense_lead) { + 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, + model.layers[il].ffn_gate, NULL, + model.layers[il].ffn_down, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, cb, il); + cb(cur, "ffn_out", il); + } else { + // MoE branch + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "ffn_norm", il); + + ggml_tensor * moe_out = + llm_build_moe_ffn(ctx0, cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + n_expert, n_expert_used, + LLM_FFN_SILU, false, + true, hparams.expert_weights_scale, + cb, il); + cb(moe_out, "ffn_moe_out", il); + + // FFN shared expert + { + ggml_tensor * ffn_shexp = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up_shexp, NULL, + model.layers[il].ffn_gate_shexp, NULL, + model.layers[il].ffn_down_shexp, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, cb, il); + cb(ffn_shexp, "ffn_shexp", il); + + cur = ggml_add(ctx0, moe_out, ffn_shexp); + cb(cur, "ffn_out", il); + } + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + 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; + } + }; static struct ggml_cgraph * llama_build_graph_defrag(llama_context & lctx, const std::vector & ids) { @@ -11007,6 +11595,14 @@ static struct ggml_cgraph * llama_build_graph( { result = llm.build_gptneox(); } break; + case LLM_ARCH_ARCTIC: + { + result = llm.build_arctic(); + } break; + case LLM_ARCH_DEEPSEEK2: + { + result = llm.build_deepseek2(); + } break; default: GGML_ASSERT(false); } @@ -12328,6 +12924,7 @@ struct llm_tokenizer_bpe { }); break; case LLAMA_VOCAB_PRE_TYPE_DBRX: + case LLAMA_VOCAB_PRE_TYPE_SMAUG: word_collection = unicode_regex_split(text, { // same as llama3 "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", @@ -12550,7 +13147,7 @@ struct llm_tokenizer_wpm { llm_tokenizer_wpm(const llama_vocab & vocab): vocab(vocab) {} void tokenize(const std::string & text, std::vector & output) { - auto * token_map = &vocab.token_to_id; + const auto & token_map = vocab.token_to_id; // normalize and split by whitespace std::vector words = preprocess(text); @@ -12565,108 +13162,89 @@ struct llm_tokenizer_wpm { } // prepend phantom space - std::string word1 = "\xe2\x96\x81" + word; - int n = word1.size(); + const std::string word1 = "\xe2\x96\x81" + word; + const int n = word1.size(); + + const size_t current_tokens = output.size(); // we're at the start of a new word - int i = 0; - bool match_any = false; - // move through character position in word - while (i < n) { + for (int i = 0; i < n; ++i) { // loop through possible match length bool match = false; for (int j = n; j > i; j--) { - auto it = token_map->find(word1.substr(i, j - i)); - if (it != token_map->end()) { + auto it = token_map.find(word1.substr(i, j - i)); + if (it != token_map.end()) { output.push_back(it->second); match = true; - match_any = true; - i = j; + i = j - 1; break; } } - // must be an unknown character - if (!match) { - i++; + if (!match) { // discard all + output.resize(current_tokens); + break; // and discard next tokens } } // we didn't find any matches for this word - if (!match_any) { + if (current_tokens == output.size()) { output.push_back(vocab.special_unk_id); } } } std::vector preprocess(const std::string & text) { - std::vector cpts_nfd = unicode_cpts_normalize_nfd(unicode_cpts_from_utf8(text)); + const std::vector cpts_nfd = unicode_cpts_normalize_nfd(unicode_cpts_from_utf8(text)); + std::vector words(1, ""); - // strip accents, strip control, uniformize whitespace, - // to lowercase, pad chinese characters, pad punctuation - std::string new_str = ""; - for (uint32_t code : cpts_nfd) { - const codepoint_flags flags = unicode_cpt_flags(code); - if (flags.is_accent_mark || flags.is_control) { + for (const char32_t cpt : cpts_nfd) { + const auto flags = unicode_cpt_flags(cpt); + + if (flags.is_whitespace) { + if (words.back().size()) { // finish previous word if any + words.emplace_back(); + } continue; } - code = unicode_tolower(code); - if (flags.is_separator || flags.is_whitespace) { //####FIXME: is_separator ? - code = ' '; + + assert (!flags.is_separator); + if (cpt == 0 || cpt == 0xFFFD || flags.is_control) { + continue; } - std::string s = unicode_cpt_to_utf8(code); - if (flags.is_punctuation || is_ascii_punct(code) || is_chinese_char(code)) { - new_str += " "; - new_str += s; - new_str += " "; + + const std::string s = unicode_cpt_to_utf8(unicode_tolower(cpt)); + if (flags.is_punctuation || ( cpt < 0x7F && flags.is_symbol ) || is_chinese_char(cpt)) { + if (words.back().size()) { // finish previous word if any + words.emplace_back(); + } + words.back() = s; // single char word + words.emplace_back(); // start a new word } else { - new_str += s; + words.back() += s; // append char to word } } - // split by whitespace - uint64_t l = 0; - uint64_t r = 0; - std::vector words; - while (r < new_str.size()) { - // if is whitespace - if (isspace(new_str[r], std::locale::classic())) { - if (r > l) words.push_back(new_str.substr(l, (r - l))); - l = r + 1; - r = l; - } else { - r += 1; - } - } - if (r > l) { - words.push_back(new_str.substr(l, (r - l))); + if (!words.back().size()) { + words.pop_back(); } + return words; } - bool is_ascii_punct(uint32_t code) { - if (code > 0xFF) { - return false; - } - auto c = char(static_cast(code)); - return ispunct(c, std::locale::classic()); - } - - bool is_chinese_char(uint32_t cpt) { - if ((cpt >= 0x4E00 && cpt <= 0x9FFF) || - (cpt >= 0x3400 && cpt <= 0x4DBF) || + static bool is_chinese_char(uint32_t cpt) { + return + (cpt >= 0x04E00 && cpt <= 0x09FFF) || + (cpt >= 0x03400 && cpt <= 0x04DBF) || (cpt >= 0x20000 && cpt <= 0x2A6DF) || (cpt >= 0x2A700 && cpt <= 0x2B73F) || (cpt >= 0x2B740 && cpt <= 0x2B81F) || (cpt >= 0x2B920 && cpt <= 0x2CEAF) || // this should be 0x2B820 but in hf rust code it is 0x2B920 - (cpt >= 0xF900 && cpt <= 0xFAFF) || - (cpt >= 0x2F800 && cpt <= 0x2FA1F) || - (cpt >= 0x3000 && cpt <= 0x303F) || - (cpt >= 0xFF00 && cpt <= 0xFFEF)) { - return true; // NOLINT - } - return false; + (cpt >= 0x0F900 && cpt <= 0x0FAFF) || + (cpt >= 0x2F800 && cpt <= 0x2FA1F); + //(cpt >= 0x3000 && cpt <= 0x303F) || + //(cpt >= 0xFF00 && cpt <= 0xFFEF); } const llama_vocab & vocab; @@ -12710,9 +13288,8 @@ struct fragment_buffer_variant { static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list & buffer) { // for each special token - for (const auto & st: vocab.special_tokens_cache) { - const auto & special_token = st.first; - const auto & special_id = st.second; + for (const llama_vocab::id special_id : vocab.cache_special_tokens) { + const auto & special_token = vocab.id_to_token[special_id].text; // for each text fragment std::forward_list::iterator it = buffer.begin(); @@ -12721,7 +13298,7 @@ static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list< // if a fragment is text ( not yet processed ) if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) { - auto * raw_text = &(fragment.raw_text); + auto & raw_text = fragment.raw_text; auto raw_text_base_offset = fragment.offset; auto raw_text_base_length = fragment.length; @@ -12731,7 +13308,7 @@ static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list< // find the first occurrence of a given special token in this fragment // passing offset argument only limit the "search area" but match coordinates // are still relative to the source full raw_text - auto match = raw_text->find(special_token, raw_text_base_offset); + auto match = raw_text.find(special_token, raw_text_base_offset); // no occurrences found, stop processing this fragment for a given special token if (match == std::string::npos) break; @@ -12750,7 +13327,7 @@ static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list< // left const int64_t left_reminder_offset = raw_text_base_offset + 0; const int64_t left_reminder_length = match - raw_text_base_offset; - buffer.emplace_after(it, (*raw_text), left_reminder_offset, left_reminder_length); + buffer.emplace_after(it, raw_text, left_reminder_offset, left_reminder_length); #ifdef PRETOKENIZERDEBUG LLAMA_LOG_WARN("FL: (%ld %ld) '%s'\n", left_reminder_offset, left_reminder_length, raw_text->substr(left_reminder_offset, left_reminder_length).c_str()); @@ -12766,7 +13343,7 @@ static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list< if (match + special_token.length() < raw_text_base_offset + raw_text_base_length) { const int64_t right_reminder_offset = match + special_token.length(); const int64_t right_reminder_length = raw_text_base_length - ((match - raw_text_base_offset) + special_token.length()); - buffer.emplace_after(it, (*raw_text), right_reminder_offset, right_reminder_length); + buffer.emplace_after(it, raw_text, right_reminder_offset, right_reminder_length); #ifdef PRETOKENIZERDEBUG LLAMA_LOG_WARN("FR: (%ld %ld) '%s'\n", right_reminder_offset, right_reminder_length, raw_text->substr(right_reminder_offset, right_reminder_length).c_str()); @@ -13870,7 +14447,7 @@ void llama_sample_repetition_penalties( void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * candidates, const struct llama_grammar * grammar) { GGML_ASSERT(ctx); - const int64_t t_start_sample_us = ggml_time_us(); + int64_t t_start_sample_us = ggml_time_us(); bool allow_eog = false; for (const auto & stack : grammar->stacks) { @@ -13882,12 +14459,13 @@ void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * c std::vector, llama_partial_utf8>> candidates_decoded; candidates_decoded.reserve(candidates->size); - std::vector candidates_grammar; + + std::vector candidates_grammar; candidates_grammar.reserve(candidates->size); for (size_t i = 0; i < candidates->size; ++i) { - const llama_token id = candidates->data[i].id; - const std::string piece = llama_token_to_piece(ctx, id, false); + const llama_token id = candidates->data[i].id; + const std::string & piece = ctx->model.vocab.cache_token_to_piece.at(id); if (llama_token_is_eog(&ctx->model, id)) { if (!allow_eog) { @@ -14087,7 +14665,7 @@ void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar GGML_ASSERT(false); } - const std::string piece = llama_token_to_piece(ctx, token, false); + const std::string & piece = ctx->model.vocab.cache_token_to_piece.at(token); // Note terminating 0 in decoded string const auto decoded = decode_utf8(piece, grammar->partial_utf8); @@ -16075,6 +16653,8 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) { case LLM_ARCH_XVERSE: case LLM_ARCH_COMMAND_R: case LLM_ARCH_OLMO: + case LLM_ARCH_ARCTIC: + case LLM_ARCH_DEEPSEEK2: return LLAMA_ROPE_TYPE_NORM; // the pairs of head values are offset by n_rot/2 @@ -17701,6 +18281,10 @@ bool llama_token_is_eog(const struct llama_model * model, llama_token token) { ); } +bool llama_token_is_control(const struct llama_model * model, llama_token token) { + return llama_is_control_token(model->vocab, token); +} + llama_token llama_token_bos(const struct llama_model * model) { return model->vocab.special_bos_id; } @@ -17772,7 +18356,16 @@ static std::string llama_decode_text(const std::string & text) { const auto cpts = unicode_cpts_from_utf8(text); for (const auto cpt : cpts) { - decoded_text += unicode_utf8_to_byte(unicode_cpt_to_utf8(cpt)); + const auto utf8 = unicode_cpt_to_utf8(cpt); + try { + decoded_text += unicode_utf8_to_byte(utf8); + } catch (const std::out_of_range & e) { + decoded_text += "[UNK_BYTE_0x"; + for (const auto c : utf8) { + decoded_text += format("%02x", (uint8_t) c); + } + decoded_text += text + "]"; + } } return decoded_text; @@ -17780,69 +18373,83 @@ static std::string llama_decode_text(const std::string & text) { // does not write null-terminator to buf int32_t llama_token_to_piece(const struct llama_model * model, llama_token token, char * buf, int32_t length, bool special) { + // if we have a cache - use it + { + const auto & cache = special ? model->vocab.cache_token_to_piece_special : model->vocab.cache_token_to_piece; + + if (!cache.empty()) { + const auto & res = cache.at(token); + if (length < (int) res.size()) { + return -(int) res.size(); + } + memcpy(buf, res.c_str(), res.size()); + return res.size(); + } + } + if (0 <= token && token < llama_n_vocab(model)) { switch (llama_vocab_get_type(model->vocab)) { - case LLAMA_VOCAB_TYPE_WPM: - case LLAMA_VOCAB_TYPE_SPM: { - // NOTE: we accept all unsupported token types, - // suppressing them like CONTROL tokens. - if (llama_is_normal_token(model->vocab, token)) { - std::string result = model->vocab.id_to_token[token].text; - llama_unescape_whitespace(result); - if (length < (int) result.length()) { - return -(int) result.length(); + case LLAMA_VOCAB_TYPE_WPM: + case LLAMA_VOCAB_TYPE_SPM: { + // NOTE: we accept all unsupported token types, + // suppressing them like CONTROL tokens. + if (llama_is_normal_token(model->vocab, token)) { + std::string result = model->vocab.id_to_token[token].text; + llama_unescape_whitespace(result); + if (length < (int) result.length()) { + return -(int) result.length(); + } + memcpy(buf, result.c_str(), result.length()); + return result.length(); + } else if ( + (llama_is_user_defined_token(model->vocab, token)) || + (llama_is_control_token (model->vocab, token) && special)) { + std::string result = model->vocab.id_to_token[token].text; + if (length < (int) result.length()) { + return -(int) result.length(); + } + memcpy(buf, result.c_str(), result.length()); + return result.length(); + } else if (llama_is_unknown_token(model->vocab, token)) { // NOLINT + if (length < 3) { + return -3; + } + memcpy(buf, "\xe2\x96\x85", 3); + return 3; + } else if (llama_is_byte_token(model->vocab, token)) { + if (length < 1) { + return -1; + } + buf[0] = llama_token_to_byte(model->vocab, token); + return 1; } - memcpy(buf, result.c_str(), result.length()); - return result.length(); - } else if ( - (llama_is_user_defined_token(model->vocab, token)) || - (llama_is_control_token (model->vocab, token) && special)) { - std::string result = model->vocab.id_to_token[token].text; - if (length < (int) result.length()) { - return -(int) result.length(); - } - memcpy(buf, result.c_str(), result.length()); - return result.length(); - } else if (llama_is_unknown_token(model->vocab, token)) { // NOLINT - if (length < 3) { - return -3; - } - memcpy(buf, "\xe2\x96\x85", 3); - return 3; - } else if (llama_is_byte_token(model->vocab, token)) { - if (length < 1) { - return -1; - } - buf[0] = llama_token_to_byte(model->vocab, token); - return 1; + break; } - break; - } - case LLAMA_VOCAB_TYPE_BPE: { - // NOTE: we accept all unsupported token types, - // suppressing them like CONTROL tokens. - if (llama_is_normal_token(model->vocab, token)) { - std::string result = model->vocab.id_to_token[token].text; - result = llama_decode_text(result); - if (length < (int) result.length()) { - return -(int) result.length(); + case LLAMA_VOCAB_TYPE_BPE: { + // NOTE: we accept all unsupported token types, + // suppressing them like CONTROL tokens. + if (llama_is_normal_token(model->vocab, token)) { + std::string result = model->vocab.id_to_token[token].text; + result = llama_decode_text(result); + if (length < (int) result.length()) { + return -(int) result.length(); + } + memcpy(buf, result.c_str(), result.length()); + return result.length(); + } else if ( + (llama_is_user_defined_token(model->vocab, token)) || + (llama_is_control_token (model->vocab, token) && special)) { + std::string result = model->vocab.id_to_token[token].text; + if (length < (int) result.length()) { + return -(int) result.length(); + } + memcpy(buf, result.c_str(), result.length()); + return result.length(); } - memcpy(buf, result.c_str(), result.length()); - return result.length(); - } else if ( - (llama_is_user_defined_token(model->vocab, token)) || - (llama_is_control_token (model->vocab, token) && special)) { - std::string result = model->vocab.id_to_token[token].text; - if (length < (int) result.length()) { - return -(int) result.length(); - } - memcpy(buf, result.c_str(), result.length()); - return result.length(); + break; } - break; - } - default: - GGML_ASSERT(false); + default: + GGML_ASSERT(false); } } return 0; @@ -18177,6 +18784,7 @@ const char * llama_print_system_info(void) { s += "AVX512_BF16 = " + std::to_string(ggml_cpu_has_avx512_bf16()) + " | "; s += "FMA = " + std::to_string(ggml_cpu_has_fma()) + " | "; s += "NEON = " + std::to_string(ggml_cpu_has_neon()) + " | "; + s += "SVE = " + std::to_string(ggml_cpu_has_sve()) + " | "; s += "ARM_FMA = " + std::to_string(ggml_cpu_has_arm_fma()) + " | "; s += "F16C = " + std::to_string(ggml_cpu_has_f16c()) + " | "; s += "FP16_VA = " + std::to_string(ggml_cpu_has_fp16_va()) + " | "; diff --git a/llama.h b/llama.h index 7aae52e29..e2bfbfa9b 100644 --- a/llama.h +++ b/llama.h @@ -85,6 +85,7 @@ extern "C" { LLAMA_VOCAB_PRE_TYPE_QWEN2 = 11, LLAMA_VOCAB_PRE_TYPE_OLMO = 12, LLAMA_VOCAB_PRE_TYPE_DBRX = 13, + LLAMA_VOCAB_PRE_TYPE_SMAUG = 14, }; // note: these values should be synchronized with ggml_rope @@ -264,6 +265,8 @@ extern "C" { bool check_tensors; // validate model tensor data }; + // NOTE: changing the default values of parameters marked as [EXPERIMENTAL] may cause crashes or incorrect results in certain configurations + // https://github.com/ggerganov/llama.cpp/pull/7544 struct llama_context_params { uint32_t seed; // RNG seed, -1 for random uint32_t n_ctx; // text context, 0 = from model @@ -290,14 +293,14 @@ extern "C" { ggml_backend_sched_eval_callback cb_eval; void * cb_eval_user_data; - enum ggml_type type_k; // data type for K cache - enum ggml_type type_v; // data type for V cache + enum ggml_type type_k; // data type for K cache [EXPERIMENTAL] + enum ggml_type type_v; // data type for V cache [EXPERIMENTAL] // Keep the booleans together to avoid misalignment during copy-by-value. bool logits_all; // the llama_decode() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead) bool embeddings; // if true, extract embeddings (together with logits) bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU - bool flash_attn; // whether to use flash attention + bool flash_attn; // whether to use flash attention [EXPERIMENTAL] // Abort callback // if it returns true, execution of llama_decode() will be aborted @@ -431,8 +434,8 @@ extern "C" { LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx); - LLAMA_API enum llama_vocab_type llama_vocab_type (const struct llama_model * model); - LLAMA_API enum llama_rope_type llama_rope_type (const struct llama_model * model); + LLAMA_API enum llama_vocab_type llama_vocab_type (const struct llama_model * model); + LLAMA_API enum llama_rope_type llama_rope_type (const struct llama_model * model); LLAMA_API int32_t llama_n_vocab (const struct llama_model * model); LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model); @@ -833,6 +836,9 @@ extern "C" { // Check if the token is supposed to end generation (end-of-generation, eg. EOS, EOT, etc.) LLAMA_API bool llama_token_is_eog(const struct llama_model * model, llama_token token); + // Identify if Token Id is a control token or a render-able token + LLAMA_API bool llama_token_is_control(const struct llama_model * model, llama_token token); + // Special tokens LLAMA_API llama_token llama_token_bos(const struct llama_model * model); // beginning-of-sentence LLAMA_API llama_token llama_token_eos(const struct llama_model * model); // end-of-sentence diff --git a/requirements.txt b/requirements.txt index 43f82dc2e..e5cfbf10b 100644 --- a/requirements.txt +++ b/requirements.txt @@ -4,7 +4,7 @@ # Package versions must stay compatible across all top-level python scripts. # --r ./requirements/requirements-convert.txt +-r ./requirements/requirements-convert-legacy-llama.txt -r ./requirements/requirements-convert-hf-to-gguf.txt -r ./requirements/requirements-convert-hf-to-gguf-update.txt diff --git a/requirements/requirements-convert-hf-to-gguf-update.txt b/requirements/requirements-convert-hf-to-gguf-update.txt index 6ac402610..6eacaf429 100644 --- a/requirements/requirements-convert-hf-to-gguf-update.txt +++ b/requirements/requirements-convert-hf-to-gguf-update.txt @@ -1,2 +1,2 @@ --r ./requirements-convert.txt +-r ./requirements-convert-legacy-llama.txt torch~=2.1.1 diff --git a/requirements/requirements-convert-hf-to-gguf.txt b/requirements/requirements-convert-hf-to-gguf.txt index 6ac402610..6eacaf429 100644 --- a/requirements/requirements-convert-hf-to-gguf.txt +++ b/requirements/requirements-convert-hf-to-gguf.txt @@ -1,2 +1,2 @@ --r ./requirements-convert.txt +-r ./requirements-convert-legacy-llama.txt torch~=2.1.1 diff --git a/requirements/requirements-convert.txt b/requirements/requirements-convert-legacy-llama.txt similarity index 100% rename from requirements/requirements-convert.txt rename to requirements/requirements-convert-legacy-llama.txt diff --git a/requirements/requirements-convert-llama-ggml-to-gguf.txt b/requirements/requirements-convert-llama-ggml-to-gguf.txt index a0f37cd1c..e80c29012 100644 --- a/requirements/requirements-convert-llama-ggml-to-gguf.txt +++ b/requirements/requirements-convert-llama-ggml-to-gguf.txt @@ -1 +1 @@ --r ./requirements-convert.txt +-r ./requirements-convert-legacy-llama.txt diff --git a/scripts/check-requirements.sh b/scripts/check-requirements.sh index 6a7400d3c..0c6afdd59 100755 --- a/scripts/check-requirements.sh +++ b/scripts/check-requirements.sh @@ -166,7 +166,7 @@ if (( do_cleanup )); then rm -rf -- "$all_venv" fi -check_convert_script convert.py +check_convert_script examples/convert-legacy-llama.py for py in convert-*.py; do # skip convert-hf-to-gguf-update.py # TODO: the check is failing for some reason: diff --git a/scripts/compare-llama-bench.py b/scripts/compare-llama-bench.py index 0ede9e67c..6016eb2c0 100755 --- a/scripts/compare-llama-bench.py +++ b/scripts/compare-llama-bench.py @@ -19,22 +19,22 @@ logger = logging.getLogger("compare-llama-bench") # Properties by which to differentiate results per commit: KEY_PROPERTIES = [ - "cpu_info", "gpu_info", "n_gpu_layers", "main_gpu", "cuda", "opencl", "metal", "gpu_blas", - "blas", "model_filename", "model_type", "model_size", "model_n_params", "n_batch", "n_threads", - "type_k", "type_v", "no_kv_offload", "tensor_split", "n_prompt", "n_gen" + "cpu_info", "gpu_info", "n_gpu_layers", "cuda", "opencl", "vulkan", "kompute", "metal", "sycl", "rpc", "gpu_blas", + "blas", "model_filename", "model_type", "model_size", "model_n_params", "n_batch", "n_ubatch", "embeddings", "n_threads", + "type_k", "type_v", "use_mmap", "no_kv_offload", "split_mode", "main_gpu", "tensor_split", "flash_attn", "n_prompt", "n_gen" ] # Properties that are boolean and are converted to Yes/No for the table: -BOOL_PROPERTIES = ["cuda", "opencl", "metal", "gpu_blas", "blas"] +BOOL_PROPERTIES = ["cuda", "opencl", "vulkan", "kompute", "metal", "sycl", "gpu_blas", "blas", "embeddings", "use_mmap", "no_kv_offload", "flash_attn"] # Header names for the table: PRETTY_NAMES = { - "cuda": "CUDA", "opencl": "OpenCL", "metal": "Metal", "gpu_blas": "GPU BLAS", "blas": "BLAS", - "cpu_info": "CPU", "gpu_info": "GPU", "model_filename": "File", "model_type": "Model", - "model_size": "Model Size [GiB]", "model_n_params": "Num. of Parameters", - "n_batch": "Batch size", "n_threads": "Threads", "type_k": "K type", "type_v": "V type", - "n_gpu_layers": "GPU layers", "main_gpu": "Main GPU", "no_kv_offload": "NKVO", - "tensor_split": "Tensor split" + "cuda": "CUDA", "opencl": "OpenCL", "vulkan": "Vulkan", "kompute": "Kompute", "metal": "Metal", "sycl": "SYCL", "rpc": "RPC", + "gpu_blas": "GPU BLAS", "blas": "BLAS", "cpu_info": "CPU", "gpu_info": "GPU", "model_filename": "File", "model_type": "Model", + "model_size": "Model Size [GiB]", "model_n_params": "Num. of Par.", "n_batch": "Batch size", "n_ubatch": "Microbatch size", + "n_threads": "Threads", "type_k": "K type", "type_v": "V type", "n_gpu_layers": "GPU layers", "split_mode": "Split mode", + "main_gpu": "Main GPU", "no_kv_offload": "NKVO", "flash_attn": "FlashAttention", "tensor_split": "Tensor split", + "use_mmap": "Use mmap", "embeddings": "Embeddings", } DEFAULT_SHOW = ["model_type"] # Always show these properties by default. diff --git a/scripts/convert-gg.sh b/scripts/convert-gg.sh index 01fda16fd..8a0168432 100755 --- a/scripts/convert-gg.sh +++ b/scripts/convert-gg.sh @@ -3,20 +3,20 @@ set -e # LLaMA v1 -python3 convert.py ../llama1/7B --outfile models/llama-7b/ggml-model-f16.gguf --outtype f16 -python3 convert.py ../llama1/13B --outfile models/llama-13b/ggml-model-f16.gguf --outtype f16 -python3 convert.py ../llama1/30B --outfile models/llama-30b/ggml-model-f16.gguf --outtype f16 -python3 convert.py ../llama1/65B --outfile models/llama-65b/ggml-model-f16.gguf --outtype f16 +python3 examples/convert-legacy-llama.py ../llama1/7B --outfile models/llama-7b/ggml-model-f16.gguf --outtype f16 +python3 examples/convert-legacy-llama.py ../llama1/13B --outfile models/llama-13b/ggml-model-f16.gguf --outtype f16 +python3 examples/convert-legacy-llama.py ../llama1/30B --outfile models/llama-30b/ggml-model-f16.gguf --outtype f16 +python3 examples/convert-legacy-llama.py ../llama1/65B --outfile models/llama-65b/ggml-model-f16.gguf --outtype f16 # LLaMA v2 -python3 convert.py ../llama2/llama-2-7b --outfile models/llama-7b-v2/ggml-model-f16.gguf --outtype f16 -python3 convert.py ../llama2/llama-2-13b --outfile models/llama-13b-v2/ggml-model-f16.gguf --outtype f16 -python3 convert.py ../llama2/llama-2-70b --outfile models/llama-70b-v2/ggml-model-f16.gguf --outtype f16 +python3 examples/convert-legacy-llama.py ../llama2/llama-2-7b --outfile models/llama-7b-v2/ggml-model-f16.gguf --outtype f16 +python3 examples/convert-legacy-llama.py ../llama2/llama-2-13b --outfile models/llama-13b-v2/ggml-model-f16.gguf --outtype f16 +python3 examples/convert-legacy-llama.py ../llama2/llama-2-70b --outfile models/llama-70b-v2/ggml-model-f16.gguf --outtype f16 # Code Llama -python3 convert.py ../codellama/CodeLlama-7b/ --outfile models/codellama-7b/ggml-model-f16.gguf --outtype f16 -python3 convert.py ../codellama/CodeLlama-13b/ --outfile models/codellama-13b/ggml-model-f16.gguf --outtype f16 -python3 convert.py ../codellama/CodeLlama-34b/ --outfile models/codellama-34b/ggml-model-f16.gguf --outtype f16 +python3 examples/convert-legacy-llama.py ../codellama/CodeLlama-7b/ --outfile models/codellama-7b/ggml-model-f16.gguf --outtype f16 +python3 examples/convert-legacy-llama.py ../codellama/CodeLlama-13b/ --outfile models/codellama-13b/ggml-model-f16.gguf --outtype f16 +python3 examples/convert-legacy-llama.py ../codellama/CodeLlama-34b/ --outfile models/codellama-34b/ggml-model-f16.gguf --outtype f16 # Falcon python3 convert-falcon-hf-to-gguf.py ../falcon/falcon-7b 1 diff --git a/scripts/pod-llama.sh b/scripts/pod-llama.sh index 2058ceabf..5dabbf60e 100644 --- a/scripts/pod-llama.sh +++ b/scripts/pod-llama.sh @@ -75,7 +75,7 @@ if [ "$1" -eq "1" ]; then cd /workspace/llama.cpp - python3 convert.py ./models/tinyllama-1b --outfile ./models/tinyllama-1b/ggml-model-f16.gguf --outtype f16 + python3 examples/convert-legacy-llama.py ./models/tinyllama-1b --outfile ./models/tinyllama-1b/ggml-model-f16.gguf --outtype f16 ./quantize ./models/tinyllama-1b/ggml-model-f16.gguf ./models/tinyllama-1b/ggml-model-q4_0.gguf q4_0 ./quantize ./models/tinyllama-1b/ggml-model-f16.gguf ./models/tinyllama-1b/ggml-model-q4_k.gguf q4_k @@ -90,7 +90,7 @@ if [ "$1" -eq "2" ]; then cd /workspace/llama.cpp - python3 convert.py ./models/codellama-7b --outfile ./models/codellama-7b/ggml-model-f16.gguf --outtype f16 + python3 examples/convert-legacy-llama.py ./models/codellama-7b --outfile ./models/codellama-7b/ggml-model-f16.gguf --outtype f16 ./quantize ./models/codellama-7b/ggml-model-f16.gguf ./models/codellama-7b/ggml-model-q4_0.gguf q4_0 ./quantize ./models/codellama-7b/ggml-model-f16.gguf ./models/codellama-7b/ggml-model-q4_k.gguf q4_k @@ -105,7 +105,7 @@ if [ "$1" -eq "3" ]; then cd /workspace/llama.cpp - python3 convert.py ./models/codellama-13b --outfile ./models/codellama-13b/ggml-model-f16.gguf --outtype f16 + python3 examples/convert-legacy-llama.py ./models/codellama-13b --outfile ./models/codellama-13b/ggml-model-f16.gguf --outtype f16 ./quantize ./models/codellama-13b/ggml-model-f16.gguf ./models/codellama-13b/ggml-model-q4_0.gguf q4_0 ./quantize ./models/codellama-13b/ggml-model-f16.gguf ./models/codellama-13b/ggml-model-q4_k.gguf q4_k @@ -120,7 +120,7 @@ if [ "$1" -eq "4" ]; then cd /workspace/llama.cpp - python3 convert.py ./models/codellama-34b --outfile ./models/codellama-34b/ggml-model-f16.gguf --outtype f16 + python3 examples/convert-legacy-llama.py ./models/codellama-34b --outfile ./models/codellama-34b/ggml-model-f16.gguf --outtype f16 ./quantize ./models/codellama-34b/ggml-model-f16.gguf ./models/codellama-34b/ggml-model-q4_0.gguf q4_0 ./quantize ./models/codellama-34b/ggml-model-f16.gguf ./models/codellama-34b/ggml-model-q4_k.gguf q4_k @@ -135,7 +135,7 @@ if [ "$1" -eq "5" ]; then cd /workspace/llama.cpp - python3 convert.py ./models/codellama-7b-instruct --outfile ./models/codellama-7b-instruct/ggml-model-f16.gguf --outtype f16 + python3 examples/convert-legacy-llama.py ./models/codellama-7b-instruct --outfile ./models/codellama-7b-instruct/ggml-model-f16.gguf --outtype f16 ./quantize ./models/codellama-7b-instruct/ggml-model-f16.gguf ./models/codellama-7b-instruct/ggml-model-q4_0.gguf q4_0 ./quantize ./models/codellama-7b-instruct/ggml-model-f16.gguf ./models/codellama-7b-instruct/ggml-model-q4_k.gguf q4_k @@ -150,7 +150,7 @@ if [ "$1" -eq "6" ]; then cd /workspace/llama.cpp - python3 convert.py ./models/codellama-13b-instruct --outfile ./models/codellama-13b-instruct/ggml-model-f16.gguf --outtype f16 + python3 examples/convert-legacy-llama.py ./models/codellama-13b-instruct --outfile ./models/codellama-13b-instruct/ggml-model-f16.gguf --outtype f16 ./quantize ./models/codellama-13b-instruct/ggml-model-f16.gguf ./models/codellama-13b-instruct/ggml-model-q4_0.gguf q4_0 ./quantize ./models/codellama-13b-instruct/ggml-model-f16.gguf ./models/codellama-13b-instruct/ggml-model-q4_k.gguf q4_k @@ -165,7 +165,7 @@ if [ "$1" -eq "7" ]; then cd /workspace/llama.cpp - python3 convert.py ./models/codellama-34b-instruct --outfile ./models/codellama-34b-instruct/ggml-model-f16.gguf --outtype f16 + python3 examples/convert-legacy-llama.py ./models/codellama-34b-instruct --outfile ./models/codellama-34b-instruct/ggml-model-f16.gguf --outtype f16 ./quantize ./models/codellama-34b-instruct/ggml-model-f16.gguf ./models/codellama-34b-instruct/ggml-model-q4_0.gguf q4_0 ./quantize ./models/codellama-34b-instruct/ggml-model-f16.gguf ./models/codellama-34b-instruct/ggml-model-q4_k.gguf q4_k diff --git a/scripts/sync-ggml-am.sh b/scripts/sync-ggml-am.sh index cf22afc41..3f8ddf37b 100755 --- a/scripts/sync-ggml-am.sh +++ b/scripts/sync-ggml-am.sh @@ -106,8 +106,6 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then # src/ggml-kompute.h -> ggml-kompute.h # src/ggml-metal.h -> ggml-metal.h # src/ggml-metal.m -> ggml-metal.m - # src/ggml-mpi.h -> ggml-mpi.h - # src/ggml-mpi.c -> ggml-mpi.c # src/ggml-opencl.cpp -> ggml-opencl.cpp # src/ggml-opencl.h -> ggml-opencl.h # src/ggml-quants.c -> ggml-quants.c @@ -145,8 +143,6 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then -e 's/src\/ggml-kompute\.h/ggml-kompute.h/g' \ -e 's/src\/ggml-metal\.h/ggml-metal.h/g' \ -e 's/src\/ggml-metal\.m/ggml-metal.m/g' \ - -e 's/src\/ggml-mpi\.h/ggml-mpi.h/g' \ - -e 's/src\/ggml-mpi\.c/ggml-mpi.c/g' \ -e 's/src\/ggml-opencl\.cpp/ggml-opencl.cpp/g' \ -e 's/src\/ggml-opencl\.h/ggml-opencl.h/g' \ -e 's/src\/ggml-quants\.c/ggml-quants.c/g' \ diff --git a/scripts/sync-ggml.last b/scripts/sync-ggml.last index 57bede67b..5042f82ae 100644 --- a/scripts/sync-ggml.last +++ b/scripts/sync-ggml.last @@ -1 +1 @@ -126d34985705a5a2222723c145cb4e125ac689f3 +2aae01fd9b8f9399f343cf18f46f38996ef52e2c diff --git a/scripts/sync-ggml.sh b/scripts/sync-ggml.sh index ec47fb27c..fbae6b7f8 100755 --- a/scripts/sync-ggml.sh +++ b/scripts/sync-ggml.sh @@ -14,8 +14,6 @@ cp -rpv ../ggml/src/ggml-kompute.h ./ggml-kompute.h cp -rpv ../ggml/src/ggml-metal.h ./ggml-metal.h cp -rpv ../ggml/src/ggml-metal.m ./ggml-metal.m cp -rpv ../ggml/src/ggml-metal.metal ./ggml-metal.metal -cp -rpv ../ggml/src/ggml-mpi.h ./ggml-mpi.h -cp -rpv ../ggml/src/ggml-mpi.c ./ggml-mpi.c cp -rpv ../ggml/src/ggml-opencl.cpp ./ggml-opencl.cpp cp -rpv ../ggml/src/ggml-opencl.h ./ggml-opencl.h cp -rpv ../ggml/src/ggml-quants.c ./ggml-quants.c diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt index 766a01752..cfa707315 100644 --- a/tests/CMakeLists.txt +++ b/tests/CMakeLists.txt @@ -129,8 +129,11 @@ llama_target_and_test(test-rope.cpp) llama_target_and_test(test-model-load-cancel.cpp LABEL "model") llama_target_and_test(test-autorelease.cpp LABEL "model") -llama_target_and_test(test-json-schema-to-grammar.cpp WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/..) -target_include_directories(test-json-schema-to-grammar PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/../examples/server) +# TODO: disabled on loongarch64 because the ggml-ci node lacks Python 3.8 +if (NOT ${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64") + llama_target_and_test(test-json-schema-to-grammar.cpp WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/..) + target_include_directories(test-json-schema-to-grammar PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/../examples/server) +endif() # dummy executable - not installed get_filename_component(TEST_TARGET test-c.c NAME_WE) diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp index de74585da..72edc64a7 100644 --- a/tests/test-backend-ops.cpp +++ b/tests/test-backend-ops.cpp @@ -1138,26 +1138,37 @@ struct test_soft_max : public test_case { // GGML_OP_ROPE struct test_rope : public test_case { const ggml_type type; - const std::array ne; + const std::array ne_a; int n_dims; int mode; int n_ctx; + float fs; // freq_scale + float ef; // ext_factor + float af; // attn_factor bool ff; + int v; // view (1 : non-contiguous a) std::string vars() override { - return VARS_TO_STR6(type, ne, n_dims, mode, n_ctx, ff); + return VARS_TO_STR10(type, ne_a, n_dims, mode, n_ctx, fs, ef, af, ff, v); } test_rope(ggml_type type = GGML_TYPE_F32, - std::array ne = {10, 10, 10, 1}, - int n_dims = 10, int mode = 0, int n_ctx = 512, bool ff = false) - : type(type), ne(ne), n_dims(n_dims), mode(mode), n_ctx(n_ctx), ff(ff) {} + std::array ne_a = {10, 10, 10, 1}, + int n_dims = 10, int mode = 0, int n_ctx = 512, float fs = 1.0f, float ef = 0.0f, float af = 0.0f, bool ff = false, int v = 0) + : type(type), ne_a(ne_a), n_dims(n_dims), mode(mode), n_ctx(n_ctx), fs(fs), ef(ef), af(af), ff(ff), v(v) {} ggml_tensor * build_graph(ggml_context * ctx) override { - ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); - ggml_tensor * pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, ne[2]); + ggml_tensor * a; + if (v & 1) { + auto ne = ne_a; ne[0] *= 2; ne[1] *= 4; ne[2] *= 3; + a = ggml_new_tensor(ctx, type, 4, ne.data()); + a = ggml_view_4d(ctx, a, ne_a[0], ne_a[1], ne_a[2], ne_a[3], a->nb[1], a->nb[2], a->nb[3], 0); + } else { + a = ggml_new_tensor(ctx, type, 4, ne_a.data()); + } + ggml_tensor * pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, ne_a[2]); ggml_tensor * freq = ff ? ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_dims/2) : nullptr; - ggml_tensor * out = ggml_rope_ext(ctx, a, pos, freq, n_dims, mode, n_ctx, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f); + ggml_tensor * out = ggml_rope_ext(ctx, a, pos, freq, n_dims, mode, n_ctx, 0, 10000.0f, fs, ef, af, 1.0f, 1.0f); return out; } @@ -1165,11 +1176,11 @@ struct test_rope : public test_case { for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) { if (t->type == GGML_TYPE_I32) { // pos - std::vector data(ne[2]); - for (int i = 0; i < ne[2]; i++) { + std::vector data(ne_a[2]); + for (int i = 0; i < ne_a[2]; i++) { data[i] = rand() % n_ctx; } - ggml_backend_tensor_set(t, data.data(), 0, ne[2] * sizeof(int)); + ggml_backend_tensor_set(t, data.data(), 0, ne_a[2] * sizeof(int)); } else { if (t->ne[0] == n_dims/2) { // frequency factors in the range [0.9f, 1.1f] @@ -1259,22 +1270,41 @@ struct test_im2col : public test_case { // GGML_OP_CONCAT struct test_concat : public test_case { const ggml_type type; - const std::array ne; - const int64_t b_ne2; + const std::array ne_a; + const int64_t ne_b_d; + const int dim; + const int v; // view (1 << 0: non-cont a, 1 << 1: non-cont b) std::string vars() override { - return VARS_TO_STR3(type, ne, b_ne2); + return VARS_TO_STR5(type, ne_a, ne_b_d, dim, v); } test_concat(ggml_type type = GGML_TYPE_F32, - std::array ne = {10, 10, 10, 10}, - int64_t b_ne2 = 10) - : type(type), ne(ne), b_ne2(b_ne2) {} + std::array ne_a = {10, 10, 10, 10}, + int64_t ne_b_d = 10, + int dim = 2, int v = 0) + : type(type), ne_a(ne_a), ne_b_d(ne_b_d), dim(dim), v(v) {} ggml_tensor * build_graph(ggml_context * ctx) override { - ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); - ggml_tensor * b = ggml_new_tensor_4d(ctx, type, ne[0], ne[1], b_ne2, ne[3]); - ggml_tensor * out = ggml_concat(ctx, a, b); + auto ne_b = ne_a; + ne_b[dim] = ne_b_d; + ggml_tensor * a; + if (v & 1) { + auto ne = ne_a; ne[0] *= 2; ne[1] *= 4; ne[2] *= 3; + a = ggml_new_tensor(ctx, type, 4, ne.data()); + a = ggml_view_4d(ctx, a, ne_a[0], ne_a[1], ne_a[2], ne_a[3], a->nb[1], a->nb[2], a->nb[3], 0); + } else { + a = ggml_new_tensor(ctx, type, 4, ne_a.data()); + } + ggml_tensor * b; + if (v & 2) { + auto ne = ne_b; ne[0] *= 3; ne[1] *= 2; ne[2] *= 4; + b = ggml_new_tensor(ctx, type, 4, ne.data()); + b = ggml_view_4d(ctx, b, ne_b[0], ne_b[1], ne_b[2], ne_b[3], b->nb[1], b->nb[2], b->nb[3], 0); + } else { + b = ggml_new_tensor(ctx, type, 4, ne_b.data()); + } + ggml_tensor * out = ggml_concat(ctx, a, b, dim); return out; } }; @@ -2194,25 +2224,47 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true, 0.1f, 0.0f)); test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true, 0.1f, 8.0f)); - for (ggml_type type : {GGML_TYPE_F32, GGML_TYPE_F16}) { - // TODO: ff not supported yet for !neox - test_cases.emplace_back(new test_rope(type, {128, 32, 10, 1}, 128, 0, 512, false)); // llama 7B - test_cases.emplace_back(new test_rope(type, {128, 40, 10, 1}, 128, 0, 512, false)); // llama 13B - test_cases.emplace_back(new test_rope(type, {128, 52, 10, 1}, 128, 0, 512, false)); // llama 30B - test_cases.emplace_back(new test_rope(type, {128, 64, 10, 1}, 128, 0, 512, false)); // llama 65B + { + bool all = true; - for (bool ff : {false, true}) { // freq_factors - test_cases.emplace_back(new test_rope(type, { 64, 1, 10, 1}, 64, 2, 512, ff)); // neox (falcon 7B) - test_cases.emplace_back(new test_rope(type, { 64, 71, 10, 1}, 64, 2, 512, ff)); // neox (falcon 7B) - test_cases.emplace_back(new test_rope(type, { 64, 8, 10, 1}, 64, 2, 512, ff)); // neox (falcon 40B) - test_cases.emplace_back(new test_rope(type, { 64, 128, 10, 1}, 64, 2, 512, ff)); // neox (falcon 40B) - test_cases.emplace_back(new test_rope(type, { 80, 32, 10, 1}, 20, 2, 512, ff)); // neox (stablelm) - test_cases.emplace_back(new test_rope(type, { 80, 32, 10, 1}, 32, 2, 512, ff)); // neox (phi-2) + for (float v : { 0, 1 }) { + for (float fs : { 1.0f, 1.4245f }) { + for (float ef : { 0.0f, 0.7465f }) { + for (float af : { 1.0f, 1.4245f }) { + for (ggml_type type : {GGML_TYPE_F32, GGML_TYPE_F16}) { + // TODO: ff not supported yet for !neox + test_cases.emplace_back(new test_rope(type, {128, 32, 10, 1}, 128, 0, 512, fs, ef, af, false, v)); // llama 7B + if (all) { + test_cases.emplace_back(new test_rope(type, {128, 40, 10, 1}, 128, 0, 512, fs, ef, af, false, v)); // llama 13B + test_cases.emplace_back(new test_rope(type, {128, 52, 10, 1}, 128, 0, 512, fs, ef, af, false, v)); // llama 30B + test_cases.emplace_back(new test_rope(type, {128, 64, 10, 1}, 128, 0, 512, fs, ef, af, false, v)); // llama 65B + } + + for (bool ff : {false, true}) { // freq_factors + if (all) { + test_cases.emplace_back(new test_rope(type, { 64, 1, 10, 1}, 64, 2, 512, fs, ef, af, ff, v)); // neox (falcon 7B) + test_cases.emplace_back(new test_rope(type, { 64, 71, 10, 1}, 64, 2, 512, fs, ef, af, ff, v)); // neox (falcon 7B) + test_cases.emplace_back(new test_rope(type, { 64, 8, 10, 1}, 64, 2, 512, fs, ef, af, ff, v)); // neox (falcon 40B) + test_cases.emplace_back(new test_rope(type, { 80, 32, 10, 1}, 20, 2, 512, fs, ef, af, ff, v)); // neox (stablelm) + test_cases.emplace_back(new test_rope(type, { 80, 32, 10, 1}, 32, 2, 512, fs, ef, af, ff, v)); // neox (phi-2) + } + + test_cases.emplace_back(new test_rope(type, { 64, 128, 10, 1}, 64, 2, 512, fs, ef, af, ff, v)); // neox (falcon 40B) + } + } + all = false; + } + } + } } } - test_cases.emplace_back(new test_concat(GGML_TYPE_F32)); - test_cases.emplace_back(new test_concat(GGML_TYPE_I32)); + for (int v : { 0, 1, 2, 3 }) { + for (int dim : { 0, 1, 2, 3, }) { + test_cases.emplace_back(new test_concat(GGML_TYPE_F32, {11, 12, 13, 14}, 7, dim, v)); + test_cases.emplace_back(new test_concat(GGML_TYPE_I32, {11, 12, 13, 14}, 7, dim, v)); + } + } for (ggml_sort_order order : {GGML_SORT_ORDER_ASC, GGML_SORT_ORDER_DESC}) { test_cases.emplace_back(new test_argsort(GGML_TYPE_F32, {8, 1, 1, 1}, order)); diff --git a/tests/test-tokenizer-0.sh b/tests/test-tokenizer-0.sh index 1fec8bbf1..4d2b83655 100755 --- a/tests/test-tokenizer-0.sh +++ b/tests/test-tokenizer-0.sh @@ -28,6 +28,8 @@ printf "Tokenizing using (cpp) llama.cpp ...\n" cat /tmp/test-tokenizer-0-$name-py.log | grep "tokenized in" cat /tmp/test-tokenizer-0-$name-cpp.log | grep "tokenized in" +set +e + diff $input.tok $input.tokcpp > /dev/null 2>&1 if [ $? -eq 0 ]; then diff --git a/tests/test-tokenizer-random.py b/tests/test-tokenizer-random.py index 7e1b656e5..ec1b2837c 100644 --- a/tests/test-tokenizer-random.py +++ b/tests/test-tokenizer-random.py @@ -167,8 +167,10 @@ def generator_random_special_tokens(tokenizer, iterations=100) -> Iterator[str]: for m in range(iterations): rand.seed(m) words = rand.choices(special_tokens, k=500) - if tokenizer.add_bos_token: # skip spam warning of double BOS - while words and words[0] == tokenizer.bos_token: + if words[0] == tokenizer.bos_token: # skip spam warning of double BOS + while len(words) > 1 and words[1] == tokenizer.bos_token: # leave one starting BOS + words.pop(0) + if tokenizer.add_bos_token: # drop all starting BOS words.pop(0) yield "".join(words) @@ -293,15 +295,17 @@ def main(argv: list[str] = None): model = LibLlamaModel(LibLlama(), args.vocab_file, mparams=dict(vocab_only=True), cparams=dict(n_ctx=4096)) tokenizer = AutoTokenizer.from_pretrained(args.dir_tokenizer) - tokenizer.add_bos_token = getattr(tokenizer, "add_bos_token", True) - tokenizer.add_eos_token = getattr(tokenizer, "add_eos_token", False) - def func_tokenize1(text: str): return model.tokenize(text, add_special=True, parse_special=True) def func_tokenize2(text: str): return tokenizer.encode(text, add_special_tokens=True) + ids = func_tokenize2("a") + assert 1 <= len(ids) <= 3 + add_bos_token = len(ids) > 1 and tokenizer.bos_token_id == ids[0] + tokenizer.add_bos_token = getattr(tokenizer, "add_bos_token", add_bos_token) + vocab = list(sorted(tokenizer.batch_decode(list(tokenizer.get_vocab().values()), skip_special_tokens=True))) test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_custom_text()) test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_custom_text_edge_cases()) @@ -324,8 +328,10 @@ if __name__ == "__main__": # import os # tokenizers = os.listdir(path_tokenizers) tokenizers = [ - "llama-spm", # SPM - "phi-3", # SPM + # "llama-spm", # SPM + # "phi-3", # SPM + "jina-v2-en", # WPM + "bert-bge", # WPM ] for tokenizer in tokenizers: