ollama3.patch

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 940c5916..7e9cf5ae 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -1,5 +1,8 @@
 cmake_minimum_required(VERSION 3.21)
 
+SET(CMAKE_C_COMPILER "/usr/local/opt/llvm@19/bin/clang")
+SET(CMAKE_CXX_COMPILER "/usr/local/opt/llvm@19/bin/clang++")
+
 project(Ollama C CXX)
 
 include(CheckLanguage)
@@ -20,6 +23,9 @@ set(GGML_BACKEND_DL ON)
 set(GGML_BACKEND_SHARED ON)
 set(GGML_SCHED_MAX_COPIES 4)
 
+set(GGML_METAL ON)
+set(GGML_METAL_EMBED_LIBRARY ON)
+
 set(GGML_LLAMAFILE ON)
 set(GGML_CUDA_PEER_MAX_BATCH_SIZE 128)
 set(GGML_CUDA_GRAPHS ON)
diff --git a/discover/gpu_darwin.go b/discover/gpu_darwin.go
index dd5bf6e2..50b6bf0c 100644
--- a/discover/gpu_darwin.go
+++ b/discover/gpu_darwin.go
@@ -6,6 +6,12 @@ package discover
 #cgo CFLAGS: -x objective-c
 #cgo LDFLAGS: -framework Foundation -framework CoreGraphics -framework Metal
 #include "gpu_info_darwin.h"
+
+#include <Metal/Metal.h>
+
+int isMetalSupported() {
+	return MTLCreateSystemDefaultDevice() != NULL;
+}
 */
 import "C"
 
@@ -23,7 +29,7 @@ const (
 
 func GetGPUInfo() GpuInfoList {
 	mem, _ := GetCPUMem()
-	if runtime.GOARCH == "amd64" {
+	if runtime.GOARCH == "amd64" && C.isMetalSupported() == 0 {
 		return []GpuInfo{
 			{
 				Library: "cpu",
diff --git a/llama/llama.cpp/src/llama-mmap.cpp b/llama/llama.cpp/src/llama-mmap.cpp
index a9932633..040fe93d 100644
--- a/llama/llama.cpp/src/llama-mmap.cpp
+++ b/llama/llama.cpp/src/llama-mmap.cpp
@@ -4,6 +4,8 @@
 
 #include "ggml.h"
 
+#include <errno.h>
+
 #include <cstring>
 #include <climits>
 #include <stdexcept>
diff --git a/llama/llama.go b/llama/llama.go
index a20f2357..25d6fefa 100644
--- a/llama/llama.go
+++ b/llama/llama.go
@@ -60,14 +60,15 @@ func BackendInit() {
 
 func PrintSystemInfo() string {
 	var compiler string
-	switch C.get_compiler() {
+	/*switch C.get_compiler() {
 	case C.COMP_UNKNOWN:
 		compiler = "cgo(unknown_compiler)"
 	case C.COMP_GCC:
 		compiler = "cgo(gcc)"
 	case C.COMP_CLANG:
 		compiler = "cgo(clang)"
-	}
+	}*/
+	compiler = "cgo(clang)"
 	return C.GoString(C.llama_print_system_info()) + compiler
 }
 
diff --git a/ml/backend/ggml/ggml/src/CMakeLists.txt b/ml/backend/ggml/ggml/src/CMakeLists.txt
index 72b488dd..4ae893c6 100644
--- a/ml/backend/ggml/ggml/src/CMakeLists.txt
+++ b/ml/backend/ggml/ggml/src/CMakeLists.txt
@@ -301,6 +301,7 @@ else ()
     ggml_add_cpu_backend_variant_impl("")
 endif()
 
+
 ggml_add_backend(BLAS)
 ggml_add_backend(CANN)
 ggml_add_backend(CUDA)
diff --git a/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal.m b/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal.m
index 318addec..74efff23 100644
--- a/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal.m
+++ b/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal.m
@@ -60,13 +60,13 @@ static id<MTLDevice> ggml_backend_metal_device_acq(struct ggml_backend_metal_dev
     if (ctx->mtl_device == nil) {
         ctx->mtl_device = MTLCreateSystemDefaultDevice();
 
-        ctx->has_simdgroup_reduction  = [ctx->mtl_device supportsFamily:MTLGPUFamilyApple7];
+        ctx->has_simdgroup_reduction  = false; // [ctx->mtl_device supportsFamily:MTLGPUFamilyApple7];
         ctx->has_simdgroup_reduction |= [ctx->mtl_device supportsFamily:MTLGPUFamilyMetal3_GGML];
 
-        ctx->has_simdgroup_mm = [ctx->mtl_device supportsFamily:MTLGPUFamilyApple7];
+        ctx->has_simdgroup_mm = false; // [ctx->mtl_device supportsFamily:MTLGPUFamilyApple7];
 
         ctx->has_bfloat  = [ctx->mtl_device supportsFamily:MTLGPUFamilyMetal3_GGML];
-        ctx->has_bfloat |= [ctx->mtl_device supportsFamily:MTLGPUFamilyApple6];
+        //ctx->has_bfloat |= [ctx->mtl_device supportsFamily:MTLGPUFamilyApple6];
 
 #if defined(GGML_METAL_USE_BF16)
         ctx->use_bfloat = ctx->has_bfloat;
@@ -2251,7 +2251,7 @@ static void ggml_metal_encode_node(
                 } else
                 // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
                 // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
-                if ([device supportsFamily:MTLGPUFamilyApple7] &&
+                if (false /*[device supportsFamily:MTLGPUFamilyApple7]*/ &&
                         !ggml_is_transposed(src0) &&
                         !ggml_is_transposed(src1) &&
                         src1t == GGML_TYPE_F32 &&
@@ -2594,7 +2594,7 @@ static void ggml_metal_encode_node(
                 // TODO: for now, always use mat-vec kernels until we figure out how to improve the
                 //       indirect matrix multiplication
                 // !!!
-                if ([device supportsFamily:MTLGPUFamilyApple7] &&
+                if (false /*[device supportsFamily:MTLGPUFamilyApple7]*/ &&
                         ne00 % 32 == 0 && ne00 >= 64 &&
                         dst_rows > dst_rows_min) {
                     // some Metal matrix data types require aligned pointers