NUMA-aware KV cache buffer type (experimental)

Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com>

ggml/include/ggml-backend.h

@@ -109,6 +109,8 @@ extern "C" {
 
     GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void);
 
+    GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_numa_buffer_type(void);
+
 #ifdef GGML_USE_CPU_HBM
     GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
 #endif

ggml/src/ggml-backend.c

@@ -936,6 +936,90 @@ GGML_CALL ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, siz
     return ggml_backend_buffer_init(ggml_backend_cpu_buffer_type(), cpu_backend_buffer_i_from_ptr, ptr, size);
 }
 
+
+
+// NUMA buffer interface - similar to CPU, but with pages allocated accordingly to a NUMA first-touch policy
+
+#include <sys/mman.h>
+
+GGML_CALL static void ggml_backend_numa_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    if (munmap((char *) buffer->context, buffer->size)) {
+        GGML_LOG_WARN("warning: munmap failed: %s\n", strerror(errno));
+    }
+}
+
+GGML_CALL static void ggml_backend_numa_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    if (posix_madvise(buffer->context, buffer->size, POSIX_MADV_DONTNEED)) {
+        GGML_LOG_WARN("warning: posix_madvise(.., POSIX_MADV_DONTNEED) failed: %s\n",
+                strerror(errno));
+    }
+}
+
+static const struct ggml_backend_buffer_i ggml_backend_numa_buffer_i = {
+    /* .free_buffer     = */ ggml_backend_numa_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_cpu_buffer_get_base,
+    /* .init_tensor     = */ NULL, // no initialization required
+    /* .memset_tensor   = */ ggml_backend_cpu_buffer_memset_tensor,
+    /* .set_tensor      = */ ggml_backend_cpu_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_cpu_buffer_get_tensor,
+    /* .cpy_tensor      = */ ggml_backend_cpu_buffer_cpy_tensor,
+    /* .clear           = */ ggml_backend_numa_buffer_clear,
+    /* .reset           = */ NULL,
+};
+
+// NUMA buffer type - similar to CPU, but with pages allocated accordingly to a NUMA first-touch policy
+
+GGML_CALL static const char * ggml_backend_numa_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    return "NUMA";
+
+    GGML_UNUSED(buft);
+}
+
+GGML_CALL static ggml_backend_buffer_t ggml_backend_numa_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    int flags = MAP_SHARED | MAP_ANONYMOUS;
+    void * data = mmap(NULL, size, PROT_READ|PROT_WRITE, flags, -1, 0);
+    if (data == MAP_FAILED) {
+        GGML_LOG_ERROR("%s: failed to allocate buffer of size %zu\n", __func__, size);
+        return NULL;
+    }
+    if (posix_madvise(data, size, POSIX_MADV_RANDOM)) {
+        GGML_LOG_WARN("warning: posix_madvise(.., POSIX_MADV_RANDOM) failed: %s\n",
+                strerror(errno));
+    }
+
+    return ggml_backend_buffer_init(buft, ggml_backend_numa_buffer_i, data, size);
+}
+
+GGML_CALL static size_t ggml_backend_numa_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+    return TENSOR_ALIGNMENT;
+
+    GGML_UNUSED(buft);
+}
+
+GGML_CALL static bool ggml_backend_numa_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
+    return true;
+
+    GGML_UNUSED(buft);
+}
+
+GGML_CALL ggml_backend_buffer_type_t ggml_backend_numa_buffer_type(void) {
+    static struct ggml_backend_buffer_type ggml_backend_numa_buffer_type = {
+        /* .iface   = */ {
+            /* .get_name         = */ ggml_backend_numa_buffer_type_get_name,
+            /* .alloc_buffer     = */ ggml_backend_numa_buffer_type_alloc_buffer,
+            /* .get_alignment    = */ ggml_backend_numa_buffer_type_get_alignment,
+            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
+            /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
+            /* .is_host          = */ ggml_backend_numa_buffer_type_is_host,
+        },
+        /* .device  = */ NULL, // FIXME ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
+        /* .context = */ NULL,
+    };
+
+    return &ggml_backend_numa_buffer_type;
+}
+
+
 GGML_CALL static ggml_backend_t ggml_backend_reg_cpu_init(const char * params, void * user_data) {
     return ggml_backend_cpu_init();
 

src/llama.cpp

@@ -3249,6 +3249,15 @@ static bool llama_kv_cache_init(
 
     bool warn = true;
     int n_mla = 0;
+    auto * reg = ggml_backend_dev_backend_reg(ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU));
+    auto * is_numa_fn = (decltype(ggml_is_numa) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_is_numa");
+    bool is_numa = is_numa_fn();
+    if (!offload && is_numa) {
+        LLAMA_LOG_INFO("%s: NUMA usage detected, using NUMA-aware buffer for KV cache\n", __func__);
+    }
+
+
+
     for (int i = 0; i < (int) n_layer; i++) {
         const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i) + hparams.n_embd_k_s();
         const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(i) + hparams.n_embd_v_s();
@@ -3257,6 +3266,7 @@ static bool llama_kv_cache_init(
         const uint32_t n_embd_head_k= hparams.n_embd_head_k;
 
 
+<<<<<<< HEAD
         struct ggml_context * ctx = offload ? ctx_map.at(model.buft_layer[i].buft) : cache.ctxs.front();
         ggml_tensor * k;
         ggml_tensor * v;