llama-cpp: Add new b4570 version

recipes/llama-cpp/all/conandata.yml

@@ -1,17 +1,17 @@
 sources:
+  "b4570":
+    url: "https://github.com/ggerganov/llama.cpp/archive/refs/tags/b4570.tar.gz"
+    sha256: "35bfe07807fd0cf30710023765b9a7ab6c1003f27ef907ce9cea2c5464411430"
   "b4079":
     url: "https://github.com/ggerganov/llama.cpp/archive/refs/tags/b4079.tar.gz"
     sha256: "79093413dcdbd30f83b800aeb958c87369fdfdaf4e5603b094185898ff404a32"
   "b3542":
     url: "https://github.com/ggerganov/llama.cpp/archive/refs/tags/b3542.tar.gz"
     sha256: "6f8b23d930400fce5708d2c85022ef33f1083af8f6ac395abefadacee0942e78"
-  "b3040":
-    url: "https://github.com/ggerganov/llama.cpp/archive/refs/tags/b3040.tar.gz"
-    sha256: "020e040139660eb40113503bb1057d5387677d249b990e008e04821532f7cd62"
 patches:
+  "b4570":
+    - patch_file: "patches/b4570-001-curl-patch-targets.patch"
   "b4079":
     - patch_file: "patches/b4079-001-curl-patch-targets.patch"
   "b3542":
     - patch_file: "patches/b3542-001-curl-patch-targets.patch"
-  "b3040":
-    - patch_file: "patches/b3040-001-curl-patch-targets.patch"

recipes/llama-cpp/all/conanfile.py

@@ -1,15 +1,15 @@
 import os
+import textwrap
 
 from conan import ConanFile
 from conan.errors import ConanInvalidConfiguration
 from conan.tools.apple import is_apple_os
 from conan.tools.build import check_min_cppstd, cross_building
 from conan.tools.cmake import CMake, CMakeDeps, CMakeToolchain, cmake_layout
-from conan.tools.files import copy, get, rmdir, apply_conandata_patches, export_conandata_patches
+from conan.tools.files import save, copy, get, rmdir, apply_conandata_patches, export_conandata_patches
 from conan.tools.scm import Version
 
-
-required_conan_version = ">=1.53.0"
+required_conan_version = ">=2.0.9"
 
 
 class LlamaCppConan(ConanFile):
@@ -37,36 +37,37 @@ class LlamaCppConan(ConanFile):
         "with_curl": False,
     }
 
+    implements = ["auto_shared_fpic"]
+
     @property
-    def _min_cppstd(self):
-        return "11"
+    def _is_new_llama(self):
+        # Structure of llama.cpp libraries was changed after b4079
+        return Version(self.version) >= "b4570"
 
     @property
-    def _compilers_minimum_version(self):
-        return {
-            "gcc": "8"
-        }
+    def _cuda_build_module(self):
+        # Adding this to the package info is necessary if we want consumers of llama to link correctly when
+        # they activate the CUDA option. In the future, when we have a CUDA recipe this could be removed.
+        cuda_target = "ggml-cuda" if self._is_new_llama else "ggml"
+        return textwrap.dedent(f"""\
+            find_dependency(CUDAToolkit REQUIRED)
+            if (WIN32)
+                # As of CUDA 12.3.1, Windows does not offer a static cublas library
+                target_link_libraries({cuda_target} INTERFACE CUDA::cudart_static CUDA::cublas CUDA::cublasLt CUDA::cuda_driver)
+            else ()
+                target_link_libraries({cuda_target} INTERFACE CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static CUDA::cuda_driver)
+            endif()
+        """)
 
     def export_sources(self):
         export_conandata_patches(self)
-        copy(self, "cmake/*", dst=self.export_sources_folder, src=self.recipe_folder)
-
-    def config_options(self):
-        if self.settings.os == "Windows":
-            del self.options.fPIC
-
-    def configure(self):
-        if self.options.shared:
-            self.options.rm_safe("fPIC")
 
     def validate(self):
-        if self.settings.compiler.cppstd:
-            check_min_cppstd(self, self._min_cppstd)
-        minimum_version = self._compilers_minimum_version.get(str(self.settings.compiler), False)
-        if minimum_version and Version(self.settings.get_safe("compiler.version")) < minimum_version:
-            raise ConanInvalidConfiguration(
-                f"{self.ref} requires {str(self.settings.compiler)}>={minimum_version}."
-            )
+        check_min_cppstd(self, 17 if self._is_new_llama else 11)
+
+    def validate_build(self):
+        if self._is_new_llama and self.settings.compiler == "msvc" and "arm" in self.settings.arch:
+            raise ConanInvalidConfiguration("llama-cpp does not support ARM architecture on msvc, it recommends to use clang instead")
 
     def layout(self):
         cmake_layout(self, src_folder="src")
@@ -83,14 +84,20 @@ def generate(self):
         deps.generate()
 
         tc = CMakeToolchain(self)
+        tc.variables["BUILD_SHARED_LIBS"] = bool(self.options.shared)
         tc.variables["LLAMA_STANDALONE"] = False
         tc.variables["LLAMA_BUILD_TESTS"] = False
         tc.variables["LLAMA_BUILD_EXAMPLES"] = self.options.get_safe("with_examples")
         tc.variables["LLAMA_CURL"] = self.options.get_safe("with_curl")
-        tc.variables["BUILD_SHARED_LIBS"] = bool(self.options.shared)
-        tc.variables["GGML_CUDA"] = self.options.get_safe("with_cuda")
-        if hasattr(self, "settings_build") and cross_building(self):
+        if cross_building(self):
             tc.variables["LLAMA_NATIVE"] = False
+            tc.variables["GGML_NATIVE_DEFAULT"] = False
+
+        tc.variables["GGML_BUILD_TESTS"] = False
+        # Follow with_examples when newer versions can compile examples,
+        # right now it tries to add_subdirectory to a non-existent folder
+        tc.variables["GGML_BUILD_EXAMPLES"] = False
+        tc.variables["GGML_CUDA"] = self.options.get_safe("with_cuda")
         tc.generate()
 
     def build(self):
@@ -112,25 +119,74 @@ def package(self):
         copy(self, "*common*.so", src=self.build_folder, dst=os.path.join(self.package_folder, "lib"), keep_path=False)
         copy(self, "*common*.dylib", src=self.build_folder, dst=os.path.join(self.package_folder, "lib"), keep_path=False)
         copy(self, "*common*.a", src=self.build_folder, dst=os.path.join(self.package_folder, "lib"), keep_path=False)
-        copy(self, "*.cmake", src=os.path.join(self.export_sources_folder, "cmake"), dst=os.path.join(self.package_folder, "lib", "cmake"))
+        if self.options.with_cuda and not self.options.shared:
+            save(self, os.path.join(self.package_folder, "lib", "cmake", "llama-cpp-cuda-static.cmake"), self._cuda_build_module)
+
+    def _get_backends(self):
+        results = ["cpu"]
+        if is_apple_os(self):
+            results.append("blas")
+            results.append("metal")
+        if self.options.with_cuda:
+            results.append("cuda")
+        return results
 
     def package_info(self):
-        self.cpp_info.components["common"].includedirs = [os.path.join("include", "common")]
-        self.cpp_info.components["common"].libs = ["common"]
-        self.cpp_info.components["common"].libdirs = ["lib"]
-        if self.version >= Version("b3240"):
-            self.cpp_info.components["common"].libs.append("ggml")
+        self.cpp_info.components["ggml"].libs = ["ggml"]
+        self.cpp_info.components["ggml"].resdirs = ["res"]
+        self.cpp_info.components["ggml"].set_property("cmake_target_name", "ggml::all")
+        if self.settings.os in ("Linux", "FreeBSD"):
+            self.cpp_info.components["ggml"].system_libs.append("dl")
 
         self.cpp_info.components["llama"].libs = ["llama"]
         self.cpp_info.components["llama"].resdirs = ["res"]
-        self.cpp_info.components["llama"].libdirs = ["lib"]
+        self.cpp_info.components["llama"].requires.append("ggml")
 
-        if self.options.with_cuda and not self.options.shared:
-            self.cpp_info.builddirs.append(os.path.join("lib", "cmake"))
-            module_path = os.path.join("lib", "cmake", "llama-cpp-cuda-static.cmake")
-            self.cpp_info.set_property("cmake_build_modules", [module_path])
+        self.cpp_info.components["common"].includedirs = [os.path.join("include", "common")]
+        self.cpp_info.components["common"].libs = ["common"]
+        self.cpp_info.components["common"].requires = ["llama"]
+
+        if self.options.with_curl:
+            self.cpp_info.components["common"].requires.append("libcurl::libcurl")
+            self.cpp_info.components["common"].defines.append("LLAMA_USE_CURL")
 
         if is_apple_os(self):
             self.cpp_info.components["common"].frameworks.extend(["Foundation", "Accelerate", "Metal"])
         elif self.settings.os in ("Linux", "FreeBSD"):
             self.cpp_info.components["common"].system_libs.extend(["dl", "m", "pthread", "gomp"])
+
+        if self.options.with_cuda and not self.options.shared:
+            self.cpp_info.builddirs.append(os.path.join("lib", "cmake"))
+            module_path = os.path.join("lib", "cmake", "llama-cpp-cuda-static.cmake")
+            self.cpp_info.set_property("cmake_build_modules", [module_path])
+
+        if self._is_new_llama:
+            self.cpp_info.components["ggml-base"].libs = ["ggml-base"]
+            self.cpp_info.components["ggml-base"].resdirs = ["res"]
+            self.cpp_info.components["ggml-base"].set_property("cmake_target_name", "ggml-base")
+
+            self.cpp_info.components["ggml"].requires = ["ggml-base"]
+            if self.settings.os in ("Linux", "FreeBSD"):
+                self.cpp_info.components["ggml-base"].system_libs.extend(["dl", "m", "pthread"])
+
+
+            if self.options.shared:
+                self.cpp_info.components["llama"].defines.append("LLAMA_SHARED")
+                self.cpp_info.components["ggml-base"].defines.append("GGML_SHARED")
+                self.cpp_info.components["ggml"].defines.append("GGML_SHARED")
+
+            backends = self._get_backends()
+            for backend in backends:
+                self.cpp_info.components[f"ggml-{backend}"].libs = [f"ggml-{backend}"]
+                self.cpp_info.components[f"ggml-{backend}"].resdirs = ["res"]
+                self.cpp_info.components[f"ggml-{backend}"].set_property("cmake_target_name", f"ggml-{backend}")
+                if self.options.shared:
+                    self.cpp_info.components[f"ggml-{backend}"].defines.append("GGML_BACKEND_SHARED")
+                self.cpp_info.components["ggml"].defines.append(f"GGML_USE_{backend.upper()}")
+                self.cpp_info.components["ggml"].requires.append(f"ggml-{backend}")
+
+            if is_apple_os(self):
+                if "blas" in backends:
+                    self.cpp_info.components["ggml-blas"].frameworks.append("Accelerate")
+                if "metal" in backends:
+                    self.cpp_info.components["ggml-metal"].frameworks.extend(["Metal", "MetalKit", "Foundation", "CoreFoundation"])

recipes/llama-cpp/all/patches/b4570-001-curl-patch-targets.patch

@@ -0,0 +1,15 @@
+diff --git a/common/CMakeLists.txt b/common/CMakeLists.txt
+index 24b7f87..e332bec 100644
+--- a/common/CMakeLists.txt
++++ b/common/CMakeLists.txt
+@@ -84,9 +84,7 @@ set(LLAMA_COMMON_EXTRA_LIBS build_info)
+ if (LLAMA_CURL)
+     find_package(CURL REQUIRED)
+     target_compile_definitions(${TARGET} PUBLIC LLAMA_USE_CURL)
+-    include_directories(${CURL_INCLUDE_DIRS})
+-    find_library(CURL_LIBRARY curl REQUIRED)
+-    set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} ${CURL_LIBRARY})
++    list(APPEND LLAMA_COMMON_EXTRA_LIBS CURL::libcurl)
+ endif ()
+
+ target_include_directories(${TARGET} PUBLIC .)

recipes/llama-cpp/config.yml

@@ -1,7 +1,7 @@
 versions:
+  "b4570":
+    folder: "all"
   "b4079":
     folder: "all"
   "b3542":
     folder: "all"
-  "b3040":
-    folder: "all"