llama model interleave

llama-model-interleave.diff

diff --git a/ggml/CMakeLists.txt b/ggml/CMakeLists.txt
index 9a4ee49..137e3ea 100644
--- a/ggml/CMakeLists.txt
+++ b/ggml/CMakeLists.txt
@@ -337,6 +337,11 @@ set(GGML_INCLUDE_INSTALL_DIR ${CMAKE_INSTALL_INCLUDEDIR} CACHE PATH "Location of
 set(GGML_LIB_INSTALL_DIR     ${CMAKE_INSTALL_LIBDIR}     CACHE PATH "Location of library files")
 set(GGML_BIN_INSTALL_DIR     ${CMAKE_INSTALL_BINDIR}     CACHE PATH "Location of binary  files")
 
+foreach(lib "ggml" "ggml-base")
+    target_link_libraries(${lib} PUBLIC numa)
+endforeach()
+
+
 configure_package_config_file(
         ${CMAKE_CURRENT_SOURCE_DIR}/cmake/ggml-config.cmake.in
         ${CMAKE_CURRENT_BINARY_DIR}/ggml-config.cmake
diff --git a/src/llama-mmap.cpp b/src/llama-mmap.cpp
index 3970b74..ae51025 100644
--- a/src/llama-mmap.cpp
+++ b/src/llama-mmap.cpp
@@ -269,6 +269,9 @@ void llama_file::write_u32(uint32_t val) const { pimpl->write_u32(val); }
 
 // llama_mmap
 
+#include <numa.h>
+#include <numaif.h>
+
 struct llama_mmap::impl {
 #ifdef _POSIX_MAPPED_FILES
     std::vector<std::pair<size_t, size_t>> mapped_fragments;
@@ -285,6 +288,41 @@ struct llama_mmap::impl {
         }
         if (prefetch) { flags |= MAP_POPULATE; }
 #endif
+#if 1
+        // mmap the file as read-only
+        void *mapped = mmap(NULL, size, PROT_READ, MAP_SHARED, fd, 0);
+        if (mapped == MAP_FAILED) {
+            perror("mmap");
+            exit(1);
+        }
+
+        // Backup the current NUMA policy
+        struct bitmask *old_policy = numa_get_interleave_mask();
+        if (!old_policy) {
+            fprintf(stderr, "Failed to get current NUMA interleave mask.\n");
+            exit(1);
+        }
+
+        // Set new interleave policy to nodes 0 and 1
+        struct bitmask *new_policy = numa_allocate_nodemask();
+        numa_bitmask_clearall(new_policy);
+        numa_bitmask_setbit(new_policy, 0);
+        numa_bitmask_setbit(new_policy, 1);
+        numa_set_interleave_mask(new_policy);
+
+        // Explicitly populate pages by reading them once
+        size_t pagesize = sysconf(_SC_PAGESIZE);
+        volatile char sum = 0;  // volatile to avoid optimization
+        for (off_t offset = 0; offset < size; offset += pagesize) {
+            sum += *((volatile char *)mapped + offset); // Read a byte to trigger page fault
+        }
+
+        printf("File pages populated with interleave policy on nodes 0 and 1.\n");
+
+        // Restore original NUMA policy
+        numa_set_interleave_mask(old_policy);
+        addr = mapped;
+#else
         addr = mmap(NULL, file->size(), PROT_READ, flags, fd, 0);
         if (addr == MAP_FAILED) {
             throw std::runtime_error(format("mmap failed: %s", strerror(errno)));
@@ -304,6 +342,7 @@ struct llama_mmap::impl {
         }
 
         mapped_fragments.emplace_back(0, file->size());
+#endif
     }
 
     static void align_range(size_t * first, size_t * last, size_t page_size) {
@@ -319,6 +358,7 @@ struct llama_mmap::impl {
     }
 
     void unmap_fragment(size_t first, size_t last) {
+#if 0
         int page_size = sysconf(_SC_PAGESIZE);
         align_range(&first, &last, page_size);
         size_t len = last - first;
@@ -352,6 +392,7 @@ struct llama_mmap::impl {
             }
         }
         mapped_fragments = std::move(new_mapped_fragments);
+#endif
     }
 
     ~impl() {