Limit max CUDA pinned host buffer size to 4 GB (#30449)

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -1348,6 +1348,12 @@ static ggml_backend_buffer_t ggml_backend_cuda_host_buffer_type_alloc_buffer(ggm
     return buffer;
 }
 
+static size_t ggml_backend_cuda_host_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
+    GGML_UNUSED(buft);
+    // very large pinned allocations can fail (e.g. on Windows) and fall back to pageable memory, smaller ones succeed
+    return 4ull*1024*1024*1024;
+}
+
 ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type() {
     static struct ggml_backend_buffer_type ggml_backend_cuda_buffer_type_host = {
         /* .iface    = */ {
@@ -1355,7 +1361,7 @@ ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type() {
             /* .alloc_buffer        = */ ggml_backend_cuda_host_buffer_type_alloc_buffer,
             /* .alloc_buffer_n      = */ NULL,
             /* .get_alignment       = */ ggml_backend_cpu_buffer_type()->iface.get_alignment,
-            /* .get_max_size        = */ NULL, // defaults to SIZE_MAX
+            /* .get_max_size        = */ ggml_backend_cuda_host_buffer_type_get_max_size,
             /* .get_alloc_size      = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
             /* .get_alloc_size_n    = */ NULL,
             /* .is_host             = */ ggml_backend_cpu_buffer_type()->iface.is_host,

src/llama-model.cpp

@@ -1950,7 +1950,8 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
             if (buf == nullptr) {
                 throw std::runtime_error(format("unable to allocate %s buffer", ggml_backend_buft_name(buft)));
             }
-            if (use_mlock && ggml_backend_buffer_is_host(buf)) {
+            // a multi_buffer has no base to lock, its parts are usually pinned memory (host buffer types with a max size)
+            if (use_mlock && ggml_backend_buffer_is_host(buf) && ggml_backend_buffer_get_base(buf) != nullptr) {
                 pimpl->mlock_bufs.emplace_back(new llama_mlock);
                 auto & mlock_buf = pimpl->mlock_bufs.back();
                 mlock_buf->init   (ggml_backend_buffer_get_base(buf));