ggml-backend : sync remnant

ggml-backend.c

@@ -475,7 +475,7 @@ ggml_backend_buffer_t ggml_backend_reg_alloc_buffer(size_t i, size_t size) {
 
 // backend CPU
 
-static const size_t TENSOR_ALIGNMENT = 64; // should be enough for AVX 512
+static const size_t TENSOR_ALIGNMENT = 32; // required for mmap as gguf only guarantees 32-byte alignment
 
 GGML_CALL static const char * ggml_backend_cpu_buffer_name(ggml_backend_buffer_t buffer) {
     return "CPU";
@@ -818,6 +818,7 @@ void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_
 }
 
 GGML_CALL ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size) {
+    GGML_ASSERT((uintptr_t)ptr % TENSOR_ALIGNMENT == 0 && "buffer pointer must be aligned");
     return ggml_backend_buffer_init(ggml_backend_cpu_buffer_type(), cpu_backend_buffer_i_from_ptr, ptr, size);
 }
 
@@ -1411,6 +1412,8 @@ static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
             return false;
         }
     }
+
+    return true;
 }
 
 static bool ggml_backend_sched_compute_splits(ggml_backend_sched_t sched) {

ggml-metal.h

@@ -27,7 +27,6 @@
 
 // max memory buffers that can be mapped to the device
 #define GGML_METAL_MAX_BUFFERS 64
-#define GGML_METAL_MAX_COMMAND_BUFFERS 32
 
 struct ggml_tensor;
 struct ggml_cgraph;