Add ggml-backend-meta.cpp to GPU runtime build scripts

Upstream's ggml-backend.cpp now references ggml_backend_buffer_is_meta
(line 133, 2006) and ggml-alloc.c references ggml_backend_buft_is_meta
(line 1240). Both functions are defined in the new ggml-backend-meta.cpp
which upstream made part of ggml-base.

Without this, the runtime-built GPU DSOs (ggml-cuda.so/.dll, ggml-rocm.dll,
ggml-vulkan.dll) and the on-the-fly Metal dylib build would link with
undefined references.

Updated:
- llamafile/build-functions.sh (Linux CUDA + ROCm via cuda.sh / rocm.sh)
- llamafile/cuda.bat, llamafile/cuda_parallel.bat
- llamafile/rocm.bat, llamafile/rocm_parallel.bat
- llamafile/vulkan.bat
- llamafile/metal.c (yoink + extracted-files map + compile list)
- llamafile/BUILD.mk (add ggml-backend-meta.cpp.zip.o to LLAMAFILE_METAL_SOURCES)

Author: aittalam

llamafile/BUILD.mk

@@ -234,6 +234,7 @@ LLAMAFILE_METAL_SOURCES := \
 	o/$(MODE)/llama.cpp/ggml/src/ggml.c.zip.o \
 	o/$(MODE)/llama.cpp/ggml/src/ggml-alloc.c.zip.o \
 	o/$(MODE)/llama.cpp/ggml/src/ggml-backend.cpp.zip.o \
+	o/$(MODE)/llama.cpp/ggml/src/ggml-backend-meta.cpp.zip.o \
 	o/$(MODE)/llama.cpp/ggml/src/ggml-quants.c.zip.o \
 	o/$(MODE)/llama.cpp/ggml/src/ggml-threading.cpp.zip.o \
 	o/$(MODE)/llama.cpp/ggml/include/ggml.h.zip.o \

llamafile/build-functions.sh

@@ -175,6 +175,7 @@ compile_ggml_core() {
         $llama_cpp_dir/ggml/src/ggml.c \
         $llama_cpp_dir/ggml/src/ggml-alloc.c \
         $llama_cpp_dir/ggml/src/ggml-backend.cpp \
+        $llama_cpp_dir/ggml/src/ggml-backend-meta.cpp \
         $llama_cpp_dir/ggml/src/ggml-quants.c \
         $llama_cpp_dir/ggml/src/ggml-threading.cpp"
 

llamafile/cuda.bat

@@ -249,7 +249,7 @@ for %%f in (ggml.c ggml-alloc.c ggml-quants.c) do (
 )
 
 :: C++ sources
-for %%f in (ggml-backend.cpp ggml-threading.cpp) do (
+for %%f in (ggml-backend.cpp ggml-backend-meta.cpp ggml-threading.cpp) do (
     set "SRC=%GGML_SRC_DIR%\%%f"
     set "BASE=%%~nf"
     set "OBJ=%BUILD_DIR%\ggml-core-!BASE!.obj"

llamafile/cuda_parallel.bat

@@ -268,7 +268,7 @@ for %%f in (ggml.c ggml-alloc.c ggml-quants.c) do (
 )
 
 :: C++ sources
-for %%f in (ggml-backend.cpp ggml-threading.cpp) do (
+for %%f in (ggml-backend.cpp ggml-backend-meta.cpp ggml-threading.cpp) do (
     set "SRC=%GGML_SRC_DIR%\%%f"
     set "BASE=%%~nf"
     set "OBJ=%BUILD_DIR%\ggml-core-!BASE!.obj"

llamafile/metal.c

@@ -54,6 +54,7 @@ extern char **environ;
 __static_yoink("llama.cpp/ggml/src/ggml.c");
 __static_yoink("llama.cpp/ggml/src/ggml-alloc.c");
 __static_yoink("llama.cpp/ggml/src/ggml-backend.cpp");
+__static_yoink("llama.cpp/ggml/src/ggml-backend-meta.cpp");
 __static_yoink("llama.cpp/ggml/src/ggml-quants.c");
 __static_yoink("llama.cpp/ggml/src/ggml-threading.cpp");
 
@@ -108,6 +109,7 @@ static const struct MetalSource {
     {"/zip/llama.cpp/ggml/src/ggml.c", "ggml.c"},
     {"/zip/llama.cpp/ggml/src/ggml-alloc.c", "ggml-alloc.c"},
     {"/zip/llama.cpp/ggml/src/ggml-backend.cpp", "ggml-backend.cpp"},
+    {"/zip/llama.cpp/ggml/src/ggml-backend-meta.cpp", "ggml-backend-meta.cpp"},
     {"/zip/llama.cpp/ggml/src/ggml-quants.c", "ggml-quants.c"},
     {"/zip/llama.cpp/ggml/src/ggml-threading.cpp", "ggml-threading.cpp"},
 
@@ -379,6 +381,7 @@ static bool BuildMetal(const char *dso) {
             "ggml-alloc.c",
             "ggml-quants.c",
             "ggml-backend.cpp",
+            "ggml-backend-meta.cpp",
             "ggml-threading.cpp",
             "ggml-metal.cpp",
             "ggml-metal-device.cpp",

llamafile/rocm.bat

@@ -235,7 +235,7 @@ for %%f in (ggml.c ggml-alloc.c ggml-quants.c) do (
 )
 
 :: C++ sources
-for %%f in (ggml-backend.cpp ggml-threading.cpp) do (
+for %%f in (ggml-backend.cpp ggml-backend-meta.cpp ggml-threading.cpp) do (
     set "SRC=%GGML_SRC_DIR%\%%f"
     set "BASE=%%~nf"
     set "OBJ=%BUILD_DIR%\ggml-core-!BASE!.obj"

llamafile/rocm_parallel.bat

@@ -254,7 +254,7 @@ for %%f in (ggml.c ggml-alloc.c ggml-quants.c) do (
 )
 
 :: C++ sources
-for %%f in (ggml-backend.cpp ggml-threading.cpp) do (
+for %%f in (ggml-backend.cpp ggml-backend-meta.cpp ggml-threading.cpp) do (
     set "SRC=%GGML_SRC_DIR%\%%f"
     set "BASE=%%~nf"
     set "OBJ=%BUILD_DIR%\ggml-core-!BASE!.obj"

llamafile/vulkan.bat

@@ -292,7 +292,7 @@ for %%f in (ggml.c ggml-alloc.c ggml-quants.c) do (
 )
 
 :: C++ sources
-for %%f in (ggml-backend.cpp ggml-threading.cpp) do (
+for %%f in (ggml-backend.cpp ggml-backend-meta.cpp ggml-threading.cpp) do (
     set "SRC=%GGML_SRC_DIR%\%%f"
     set "BASE=%%~nf"
     set "OBJ=%BUILD_DIR%\ggml-core-!BASE!.obj"