sync : ggml (Metal F32 support + reduce ggml-alloc size) (#3192)

* sync : ggml (Metal F32 support + reduce ggml-alloc size)

ggml-ci

* llama-bench : fix ggml_cpu_has_metal() duplicate function

ggml-ci

1 files changed, 6 insertions, 6 deletions

diff --git a/ggml-alloc.c b/ggml-alloc.c
index a1f6e7bf..304964be 100644
--- a/ggml-alloc.c
+++ b/ggml-alloc.c
@@ -131,6 +131,10 @@ static bool ggml_allocr_is_own(struct ggml_allocr * alloc, const struct ggml_ten
     return ptr >= alloc->data && (char *)ptr < (char *)alloc->data + alloc->max_size;
 }
 
+static bool ggml_is_view(struct ggml_tensor * t) {
+    return t->view_src != NULL;
+}
+
 void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
 #ifdef GGML_ALLOCATOR_DEBUG
     GGML_ASSERT(!ggml_is_view(tensor)); // views generally get data pointer from one of their sources
@@ -338,8 +342,8 @@ static void free_vmem(void * base_addr, size_t size) {
 
 // allocate uncommitted virtual memory to measure the size of the graph
 static void alloc_measure_vmem(void ** base_addr, size_t * size) {
-    // 1TB for 64-bit, 1GB for 32-bit
-    *size = sizeof(void *) == 4 ? 1ULL<<30 : 1ULL<<40;
+    // 128GB for 64-bit, 1GB for 32-bit
+    *size = sizeof(void *) == 4 ? 1ULL<<30 : 1ULL<<37;
     do {
         *base_addr = alloc_vmem(*size);
         if (*base_addr != NULL) {
@@ -399,10 +403,6 @@ bool ggml_allocr_is_measure(struct ggml_allocr * alloc) {
 
 //////////// compute graph allocator
 
-static bool ggml_is_view(struct ggml_tensor * t) {
-    return t->view_src != NULL;
-}
-
 static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
     if (a->type != b->type) {
         return false;