# -*- coding: utf-8; mode: tcl; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4 -*- vim:fenc=utf-8:ft=tcl:et:sw=4:ts=4:sts=4

PortSystem              1.0
PortGroup               github 1.0
PortGroup               cmake 1.1
PortGroup               legacysupport 1.1

github.setup            ggerganov llama.cpp 0.3.0 v
github.tarball_from     archive
revision                0
categories              llm
maintainers             {i0ntempest @i0ntempest} openmaintainer
license                 MIT

description             LLM inference in C/C++
long_description        The main goal of ${name} is to enable LLM inference with minimal\
                        setup and state-of-the-art performance on a wide variety of hardware\
                         - locally and in the cloud.

set source_distfile     ${distfiles}
set nightly_tag_distfile \
                        nightly-tag-${version}.txt
master_sites-append     https://github.com/ggml-org/llama.cpp/releases/download/v${version}/nightly-tag.txt?dummy=:nightly_tag
distfiles-append        ${nightly_tag_distfile}:nightly_tag

checksums               ${source_distfile} \
                        rmd160  19299769e173a2a7a59f28c20be4c317cb6508ed \
                        sha256  d94c02d86db22d68692f6bb5b3854763d5091e52142868dc7251995517c666d1 \
                        size    36951021 \
                        ${nightly_tag_distfile} \
                        rmd160  356208a087e093b9a0066b009874b1c8bd01ff4f \
                        sha256  46637e1a90db3d9912a7722806c7657cc73a3965e21c0016cc6b087b3be84205 \
                        size    7

extract.only            ${source_distfile}

# error: 'filesystem' file not found on 10.14
legacysupport.newest_darwin_requires_legacy \
                        18
legacysupport.use_mp_libcxx \
                        yes

depends_build-append    path:bin/pkg-config:pkgconfig

depends_lib-append      port:curl

# fix metal backend for AMD discrete GPUs
# see https://github.com/ggml-org/llama.cpp/issues/19563 and https://github.com/ggml-org/llama.cpp/pull/19567
# also see (related but unused for now) https://github.com/Cerid-AI/quenchforge/blob/main/patches/llama.cpp
patchfiles-append       patch-19567.diff
patch.args              -p1

compiler.cxx_standard   2017

# cmake relies on git for version info. We need to set them manually.
configure.args-append   -DGGML_LTO=ON \
                        -DGGML_CCACHE=OFF \
                        -DGGML_OPENMP=OFF \
                        -DLLAMA_CURL=ON \
                        -DLLAMA_BUILD_IS_DEV=OFF \
                        -DLLAMA_BUILD_TESTS=OFF \
                        -DGGML_METAL=OFF \
                        -DGGML_METAL_EMBED_LIBRARY=OFF

pre-configure {
    set fd [open ${distpath}/${nightly_tag_distfile} r]
    set nightly_tag [string trim [read ${fd}]]
    close ${fd}

    if {![regexp {^b([0-9]+)$} ${nightly_tag} -> llama_build_number]} {
        return -code error \
            "Invalid llama.cpp nightly tag: ${nightly_tag}"
    }

    set git-commit [exec curl -fsSL \
                            -H "Accept: application/vnd.github.sha" \
                            https://api.github.com/repos/ggml-org/llama.cpp/commits/${github.tag_prefix}${version} | \
                            cut -c1-7]

    configure.args-append \
        -DLLAMA_BUILD_COMMIT=${git-commit} \
        -DLLAMA_BUILD_NUMBER=${llama_build_number}
}

variant blas description {Uses BLAS, improves performance} {
    configure.args-append \
                        -DGGML_BLAS=ON
    if {${os.platform} eq "darwin" && ${os.subplatform} eq "macosx"} {
        configure.args-append \
                        -DGGML_ACCELLERATE=ON \
                        -DGGML_BLAS_VENDOR=Apple
    } else {
        configure.args-append \
                        -DGGML_ACCELLERATE=OFF \
                        -DGGML_BLAS_VENDOR=OpenBLAS

        depends_lib-append \
                        path:lib/libopenblas.dylib:OpenBLAS
    }
}

variant openmp description {enable parallelism support using OpenMP} {
    compiler.openmp_version \
                        4.5
    compiler.blacklist-append \
                        {macports-clang-[0-9].*}
    configure.args-replace \
                        -DGGML_OPENMP=OFF \
                        -DGGML_OPENMP=ON
    if {[string match *clang* ${configure.compiler}]} {
        configure.ldflags-append \
                        -L${prefix}/lib/libomp -lomp
    }
}

variant metal description {Enable Metal support} {
    configure.args-append \
                        -DGGML_METAL_MACOSX_VERSION_MIN=${macos_version_major}
    configure.args-replace \
                        -DGGML_METAL=OFF -DGGML_METAL=ON
    configure.args-replace \
                        -DGGML_METAL_EMBED_LIBRARY=OFF -DGGML_METAL_EMBED_LIBRARY=ON
}

variant vulkan description {Enable Vulkan support via MoltenVK} conflicts metal {
    depends_build-append \
                        port:vulkan-headers \
                        port:spirv-headers
    depends_lib-append \
                        path:lib/libMoltenVK.dylib:MoltenVK-latest \
                        path:bin/glslc:shaderc \
                        path:bin/glslang:glslang \
                        port:vulkan-loader
    configure.args-append \
                        -DGGML_VULKAN=ON
                        
}

variant model_converters description {install extra model conversion Python scripts} {
    post-destroot {
        xinstall -d ${destroot}${prefix}/share/${name}
        xinstall -m 644 {*}[glob ${worksrcpath}/convert*.py] ${destroot}${prefix}/share/${name}
    }

    notes-append "
        Model conversion scripts have been installed to:

            ${prefix}/share/${name}/

        These scripts should be executed in a Python virtual environment with at least\
        these dependencies installed:
            numpy
            pytorch
            transformers
            gguf
    "
}

# This can replace the variant above once we have all the dependencies updated and fixed
# See: https://github.com/macports/macports-ports/pull/27526
#variant model_converters description {install extra model conversion Python scripts} {
#    set ::python_branch   3.12
#    set ::python_version  [string map {. ""} ${::python_branch}]
#    depends_run-append    port:python${::python_version} \
#                          port:py${::python_version}-numpy \
#                          port:py${::python_version}-pytorch \
#                          port:py${::python_version}-transformers \
#                          port:py${::python_version}-gguf
#
#    post-patch {
#        reinplace "s|#!/usr/bin/env python3|#!${prefix}/bin/python${::python_branch}|" {*}[glob ${worksrcpath}/convert*.py]
#    }
#
#    post-destroot {
#        xinstall -d ${destroot}${prefix}/libexec/${name}
#        xinstall -m 755 {*}[glob ${worksrcpath}/convert*.py] ${destroot}${prefix}/libexec/${name}
#        foreach file [glob -directory ${destroot}${prefix}/libexec/${name} *.py] {
#            set filebasename [file rootname [file tail $file]]
#            set link ${destroot}${prefix}/bin/[lindex [split ${name} .] 0]-[string map {_ -} ${filebasename}]
#            set target [string replace ${file} 0 [string length ${destroot}]-1]
#            ui_debug "Creating symlink: ${link} => ${target}"
#            ln -s ${target} ${link}
#        }
#    }
#}

variant native description {Force local build and optimize for CPU} {
    configure.args-append \
                        -DGGML_NATIVE=ON
}

default_variants        +blas +openmp

# error: use of undeclared identifier 'MTLGPUFamilyApple7' on 10.14
if {${os.platform} eq "darwin" && ${os.subplatform} eq "macosx" && \
    (${os.major} >= 20 && ${configure.sdk_version} >= 11)} {
    default_variants    +metal
}
