llama.cpp: rework (#5256)

* llama.cpp: rework

* fix link order

* fix ios frameworks

* add curl configs

packages/l/llama.cpp/xmake.lua

@@ -1,43 +1,105 @@
 package("llama.cpp")
     set_homepage("https://github.com/ggerganov/llama.cpp")
     set_description("Port of Facebook's LLaMA model in C/C++")
+    set_license("MIT")
 
-    add_urls("https://github.com/ggerganov/llama.cpp.git")
-    add_versions("2023.03.11", "7d9ed7b25fe17db3fc8848b5116d14682864ce8e")
+    add_urls("https://github.com/ggerganov/llama.cpp/archive/refs/tags/b$(version).tar.gz",
+             "https://github.com/ggerganov/llama.cpp.git")
 
-    if is_plat("macosx") then
-        add_frameworks("Accelerate")
-    elseif is_plat("linux") then
+    add_versions("3775", "405bae9d550cb3fbf36d6583377b951a346b548f5850987238fe024a16f45cad")
+
+    add_configs("curl", {description = "llama: use libcurl to download model from an URL", default = false, type = "boolean"})
+    add_configs("openmp", {description = "ggml: use OpenMP", default = false, type = "boolean"})
+    add_configs("cuda", {description = "ggml: use CUDA", default = false, type = "boolean"})
+    add_configs("vulkan", {description = "ggml: use Vulkan", default = false, type = "boolean"})
+    add_configs("blas", {description = "ggml: use BLAS", default = nil, type = "string", values = {"mkl", "openblas"}})
+
+    if is_plat("macosx", "iphoneos") then
+        add_frameworks("Accelerate", "Foundation", "Metal", "MetalKit")
+    elseif is_plat("linux", "bsd") then
         add_syslinks("pthread")
     end
 
-    on_install("linux", "macosx", function (package)
-        local configs = {}
-        io.writefile("xmake.lua", [[
-            add_rules("mode.release", "mode.debug")
-            target("llama")
-                set_kind("$(kind)")
-                add_files("*.c")
-                add_headerfiles("(*.h)")
-                set_languages("c11")
-                add_cflags("-pthread")
-                if is_plat("macosx") then
-                    add_defines("GGML_USE_ACCELERATE")
-                    add_frameworks("Accelerate")
-                end
-                if is_arch("x86_64", "x64", "i386", "x86") then
-                    add_vectorexts("avx", "avx2", "sse3")
-                    add_cflags("-mf16c")
-                elseif is_arch("arm.*") then
-                    add_vectorexts("neon")
+    add_links("llama", "ggml")
+
+    add_deps("cmake")
+
+    if on_check then
+        on_check("android", function (package)
+            local ndk = package:toolchain("ndk")
+            local ndkver = ndk:config("ndkver")
+            local ndk_sdkver = ndk:config("ndk_sdkver")
+            assert(ndkver and tonumber(ndkver) > 22, "package(llama.cpp) require ndkver > 22")
+            assert(ndk_sdkver and tonumber(ndk_sdkver) >= 24, "package(llama.cpp) require ndk api >= 24")
+        end)
+
+        on_check("windows", function (package)
+            if package:is_arch("arm.*") then
+                local vs_toolset = package:toolchain("msvc"):config("vs_toolset")
+                if vs_toolset then
+                    local vs_toolset_ver = import("core.base.semver").new(vs_toolset)
+                    local minor = vs_toolset_ver:minor()
+                    assert(minor and minor >= 30, "package(llama.cpp/arm64) require vs_toolset >= 14.3")
                 end
-        ]])
+            end
+        end)
+    end
+
+    on_load(function (package)
         if package:config("shared") then
-            configs.kind = "shared"
+            package:add("defines", "GGML_SHARED", "LLAMA_SHARED")
         end
-        import("package.tools.xmake").install(package, configs)
+
+        if package:config("curl") then
+            package:add("deps", "libcurl")
+        end
+        if package:config("openmp") then
+            package:add("deps", "openmp")
+        end
+        if package:config("cuda") then
+            package:add("deps", "cuda")
+        end
+        if package:config("vulkan") then
+            -- requires vulkan-1 and glslc
+            package:add("deps", "vulkansdk")
+            package:add("deps", "shaderc", {configs = {binaryonly = true}})
+        end
+        if package:config("blas") then
+            if is_subhost("windows") then
+                package:add("deps", "pkgconf")
+            else
+                package:add("deps", "pkg-config")
+            end
+        end
+    end)
+
+    on_install(function (package)
+        local configs = {
+            "-DLLAMA_ALL_WARNINGS=OFF",
+            "-DLLAMA_BUILD_TESTS=OFF",
+            "-DLLAMA_BUILD_EXAMPLES=OFF",
+            "-DLLAMA_BUILD_SERVER=OFF",
+            "-DGGML_ALL_WARNINGS=OFF",
+            "-DGGML_BUILD_TESTS=OFF",
+            "-DGGML_BUILD_EXAMPLES=OFF",
+            "-DGGML_CCACHE=OFF",
+        }
+        table.insert(configs, "-DCMAKE_CROSSCOMPILING=" .. (package:is_cross() and "ON" or "OFF"))
+        table.insert(configs, "-DCMAKE_BUILD_TYPE=" .. (package:is_debug() and "Debug" or "Release"))
+        table.insert(configs, "-DBUILD_SHARED_LIBS=" .. (package:config("shared") and "ON" or "OFF"))
+        table.insert(configs, "-DLLAMA_SANITIZE_ADDRESS=" .. (package:config("asan") and "ON" or "OFF"))
+        table.insert(configs, "-DGGML_LTO=" .. (package:config("lto") and "ON" or "OFF"))
+        table.insert(configs, "-DGGML_SANITIZE_ADDRESS=" .. (package:config("asan") and "ON" or "OFF"))
+
+        table.insert(configs, "-DLLAMA_CURL=" .. (package:config("curl") and "ON" or "OFF"))
+        table.insert(configs, "-DGGML_OPENMP=" .. (package:config("openmp") and "ON" or "OFF"))
+        table.insert(configs, "-DGGML_CUDA=" .. (package:config("cuda") and "ON" or "OFF"))
+        table.insert(configs, "-DGGML_VULKAN=" .. (package:config("vulkan") and "ON" or "OFF"))
+        table.insert(configs, "-DGGML_OPENBLAS=" .. (package:config("blas") and "ON" or "OFF"))
+        import("package.tools.cmake").install(package, configs)
     end)
 
     on_test(function (package)
         assert(package:has_cfuncs("ggml_time_us", {includes = "ggml.h"}))
+        assert(package:has_cfuncs("llama_backend_init", {includes = "llama.h"}))
     end)