diff --git a/src/llama-mmap.cpp b/src/llama-mmap.cpp
index 0261e4c72c9..4b464ead2cb 100644
--- a/src/llama-mmap.cpp
+++ b/src/llama-mmap.cpp
@@ -448,6 +448,22 @@ struct llama_mmap::impl {
         mapped_fragments.emplace_back(0, file->size());
     }
 
+    void prefetch(size_t offset, size_t len) const {
+
+        int page_size = sysconf(_SC_PAGESIZE);
+        size_t last = offset + len;
+        align_range(&offset, &last, page_size);
+        size_t aligned_len = last - offset;
+
+        int err = posix_madvise((void*)((uint8_t *)addr + offset),
+                                aligned_len,
+                                POSIX_MADV_WILLNEED);
+        if (err != 0) {
+            LLAMA_LOG_WARN("warning: posix_madvise(.., POSIX_MADV_WILLNEED) failed: %s\n",
+                           strerror(err));
+        }
+    }
+
     static void align_range(size_t * first, size_t * last, size_t page_size) {
         size_t offset_in_page = *first & (page_size - 1);
         size_t offset_to_page = offset_in_page == 0 ? 0 : page_size - offset_in_page;
@@ -587,6 +603,7 @@ size_t llama_mmap::size() const { return pimpl->size; }
 void * llama_mmap::addr() const { return pimpl->addr; }
 
 void llama_mmap::unmap_fragment(size_t first, size_t last) { pimpl->unmap_fragment(first, last); }
+void llama_mmap::prefetch(size_t offset, size_t len) const { pimpl->prefetch(offset, len); }
 
 #if defined(_POSIX_MEMLOCK_RANGE) || defined(_WIN32)
 const bool llama_mmap::SUPPORTED  = true;
diff --git a/src/llama-mmap.h b/src/llama-mmap.h
index 29ce4d24685..c5fbc70b11d 100644
--- a/src/llama-mmap.h
+++ b/src/llama-mmap.h
@@ -48,6 +48,7 @@ struct llama_mmap {
     void * addr() const;
 
     void unmap_fragment(size_t first, size_t last);
+    void prefetch(size_t offset, size_t len) const;
 
     static const bool SUPPORTED;
 
diff --git a/src/llama-model-loader.cpp b/src/llama-model-loader.cpp
index 383b8dc7618..4139673ef1a 100644
--- a/src/llama-model-loader.cpp
+++ b/src/llama-model-loader.cpp
@@ -536,21 +536,15 @@ llama_model_loader::llama_model_loader(
     get_key(llm_kv(LLM_KV_GENERAL_ARCHITECTURE), arch_name, false);
     llm_kv = LLM_KV(llm_arch_from_string(arch_name));
 
-    files.emplace_back(new llama_file(fname.c_str(), "rb", use_direct_io));
     contexts.emplace_back(ctx);
 
-    if (use_mmap && use_direct_io) {
-        if (files.back()->has_direct_io()) {
-            // Disable mmap, as DirectIO is available
-            use_mmap = false;
-            LLAMA_LOG_WARN("%s: direct I/O is enabled, disabling mmap\n", __func__);
-        } else {
-            // Disable DirectIO and reopen file using std::fopen for mmap
-            use_direct_io = false;
-            files.pop_back();
-            files.emplace_back(new llama_file(fname.c_str(), "rb", false));
-            LLAMA_LOG_WARN("%s: direct I/O is not available, using mmap\n", __func__);
-        }
+    files.emplace_back(new llama_file(fname.c_str(), "rb", use_direct_io));
+
+    if (use_direct_io && !files.back()->has_direct_io()) {
+        use_direct_io = false;
+        LLAMA_LOG_WARN("%s: direct I/O is not available, disabling\n", __func__);
+        files.pop_back();
+        files.emplace_back(new llama_file(fname.c_str(), "rb", false));
     }
 
     // Save tensors data offset of the main file.
@@ -997,7 +991,7 @@ bool llama_model_loader::load_all_data(
     std::vector<void *> host_ptrs;
     size_t buffer_idx = 0; // buffer to use for async loads
     ggml_backend_t upload_backend = [&](const char * func) -> ggml_backend_t {
-        if (use_mmap || check_tensors) {
+        if (!use_direct_io || check_tensors) {
             return nullptr;
         }
         // When not using mmaped io use async uploads from pinned memory to GPU memory.
@@ -1092,7 +1086,7 @@ bool llama_model_loader::load_all_data(
 
         size_t n_size = ggml_nbytes(cur);
 
-        if (use_mmap) {
+        if (use_mmap && (!use_direct_io || !upload_backend || ggml_backend_buffer_is_host(cur->buffer))) {
             const auto & mapping = mappings.at(weight->idx);
             ggml_backend_buffer_t buf_mmap = nullptr;
             if (bufs.count(weight->idx)) {
@@ -1108,6 +1102,8 @@ bool llama_model_loader::load_all_data(
 
             GGML_ASSERT(buf_mmap || cur->data); // either we have a buffer to allocate the tensor in, or it is already allocated
             if (buf_mmap && cur->data == nullptr) {
+                mapping->prefetch(weight->offs, n_size);
+
                 ggml_backend_tensor_alloc(buf_mmap, cur, data);
                 if (lmlocks) {
                     const auto & lmlock = lmlocks->at(weight->idx);
diff --git a/src/llama-model.cpp b/src/llama-model.cpp
index b58b35a4268..dfdfd8101e3 100644
--- a/src/llama-model.cpp
+++ b/src/llama-model.cpp
@@ -6985,7 +6985,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
 
     ml.done_getting_tensors();
 
-    ml.init_mappings(true, use_mlock ? &pimpl->mlock_mmaps : nullptr);
+    ml.init_mappings(false, use_mlock ? &pimpl->mlock_mmaps : nullptr);
     pimpl->mappings.reserve(ml.mappings.size());
 
     // create the backend buffers