fixes

Author: slaren

ggml/include/ggml-backend.h

@@ -127,6 +127,8 @@ extern "C" {
         bool async;
         // pinned host buffer
         bool host_buffer;
+        // creating buffers from host ptr
+        bool buffer_from_host_ptr;
         // event synchronization
         bool events;
     };

ggml/src/ggml-backend.cpp

@@ -463,6 +463,7 @@ enum ggml_backend_dev_type ggml_backend_dev_type(ggml_backend_dev_t device) {
 }
 
 void ggml_backend_dev_get_props(ggml_backend_dev_t device, struct ggml_backend_dev_props * props) {
+    memset(props, 0, sizeof(*props));
     device->iface.get_props(device, props);
 }
 
@@ -1129,9 +1130,10 @@ static void ggml_backend_cpu_device_get_props(ggml_backend_dev_t dev, struct ggm
     props->type        = ggml_backend_cpu_device_get_type(dev);
     ggml_backend_cpu_device_get_memory(dev, &props->memory_free, &props->memory_total);
     props->caps = {
-        /* async       */ false,
-        /* host_buffer */ false,
-        /* events      */ false,
+        /* .async                 = */ false,
+        /* .host_buffer           = */ false,
+        /* .buffer_from_host_ptr  = */ true,
+        /* .events                = */ false,
     };
 }
 

ggml/src/ggml-cuda.cu

@@ -2920,9 +2920,10 @@ static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_back
 #endif
 
     props->caps = {
-        /* async       */ true,
-        /* host_buffer */ host_buffer,
-        /* events      */ events,
+        /* .async                 = */ true,
+        /* .host_buffer           = */ host_buffer,
+        /* .buffer_from_host_ptr  = */ false,
+        /* .events                = */ events,
     };
 }
 

ggml/src/ggml-metal.m

@@ -3567,12 +3567,14 @@ void ggml_backend_metal_capture_next_compute(ggml_backend_t backend) {
 }
 
 static void ggml_backend_metal_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
-    // TODO
-    *free = 0;
-    *total = 0;
-
     if (@available(macOS 10.12, iOS 16.0, *)) {
-        *total = g_state.mtl_device.recommendedMaxWorkingSetSize;
+        id<MTLDevice> device = ggml_backend_metal_get_device();
+        *total = device.recommendedMaxWorkingSetSize;
+        *free  = *total - device.currentAllocatedSize;
+        ggml_backend_metal_free_device();
+    } else {
+        *free = 1;
+        *total = 1;
     }
 
     GGML_UNUSED(dev);
@@ -3590,9 +3592,10 @@ static void ggml_backend_metal_device_get_props(ggml_backend_dev_t dev, struct g
     props->type        = ggml_backend_metal_device_get_type(dev);
     ggml_backend_metal_device_get_memory(dev, &props->memory_free, &props->memory_total);
     props->caps = (struct ggml_backend_dev_caps) {
-        /* async       */ false,
-        /* host_buffer */ false,
-        /* events      */ false,
+        /* .async                 = */ false,
+        /* .host_buffer           = */ false,
+        /* .buffer_from_host_ptr  = */ true,
+        /* .events                = */ false,
     };
 }
 

src/llama.cpp

@@ -8907,20 +8907,30 @@ static bool llm_load_tensors(
         llama_buf_map bufs;
         bufs.reserve(n_max_backend_buffer);
 
-        // only the mmap region containing the tensors in the model is mapped to the backend buffer
-        // this is important for metal with apple silicon: if the entire model could be mapped to a metal buffer, then we could just use metal for all layers
-        // this allows using partial offloading when the model size exceeds the metal buffer size, but not the RAM size
-        if (ml.use_mmap && use_mmap_buffer && buft == llama_default_buffer_type_cpu(model, true)) {
+        // check if this backend device supports buffer_from_host_ptr
+        ggml_backend_dev_t dev = ggml_backend_buft_get_device(buft);
+        bool buffer_from_host_ptr_supported = false;
+        if (dev) {
+            ggml_backend_dev_props props;
+            ggml_backend_dev_get_props(dev, &props);
+            buffer_from_host_ptr_supported = props.caps.buffer_from_host_ptr;
+        }
+
+        if (ml.use_mmap && use_mmap_buffer && buffer_from_host_ptr_supported) {
             for (uint32_t idx = 0; idx < ml.files.size(); idx++) {
+                // only the mmap region containing the tensors in the model is mapped to the backend buffer
+                // this is important for metal with apple silicon: if the entire model could be mapped to a metal buffer, then we could just use metal for all layers
+                // this allows using partial offloading when the model size exceeds the metal buffer size, but not the RAM size
                 void * addr = nullptr;
-                size_t first, last;
+                size_t first, last; // NOLINT
                 ml.get_mapping_range(&first, &last, &addr, idx, ctx);
                 if (first >= last) {
                     continue;
                 }
-                ggml_backend_buffer_t buf = ggml_backend_cpu_buffer_from_ptr((char *) addr + first, last - first);
+                const size_t max_size = ggml_get_max_tensor_size(ctx);
+                ggml_backend_buffer_t buf = ggml_backend_dev_buffer_from_host_ptr(dev, (char *) addr + first, last - first, max_size);
                 if (buf == nullptr) {
-                    throw std::runtime_error("unable to allocate backend CPU buffer");
+                    throw std::runtime_error(format("unable to allocate %s buffer", ggml_backend_buft_name(buft)));
                 }
                 model.bufs.push_back(buf);
                 bufs.emplace(idx, buf);
@@ -8929,7 +8939,7 @@ static bool llm_load_tensors(
         else {
             ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
             if (buf == nullptr) {
-                throw std::runtime_error("unable to allocate backend buffer");
+                throw std::runtime_error(format("unable to allocate %s buffer", ggml_backend_buft_name(buft)));
             }
             model.bufs.push_back(buf);
             if (use_mlock && ggml_backend_buffer_is_host(buf)) {