ggml : add NUMA-aware buffer type that allocates pages accordingly to the first-touch policy

llama : use NUMA-aware buffer type for KV cache

Author: sszymczy

ggml/include/ggml-backend.h

@@ -348,6 +348,7 @@ extern "C" {
     // CPU buffer types are always available
     GGML_API ggml_backend_buffer_t      ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size);
     GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void);
+    GGML_API ggml_backend_buffer_type_t ggml_backend_numa_buffer_type(void);
 
 #ifdef  __cplusplus
 }

ggml/src/ggml-backend.cpp

@@ -2000,3 +2000,84 @@ ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size)
     GGML_ASSERT((uintptr_t)ptr % TENSOR_ALIGNMENT == 0 && "buffer pointer must be aligned");
     return ggml_backend_buffer_init(ggml_backend_cpu_buffer_from_ptr_type(), ggml_backend_cpu_buffer_from_ptr_i, ptr, size);
 }
+
+// NUMA buffer interface - similar to CPU, but with pages allocated accordingly to a NUMA first-touch policy
+
+#include <sys/mman.h>
+
+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));
+    }
+}
+
+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
+
+static const char * ggml_backend_numa_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    return "NUMA";
+
+    GGML_UNUSED(buft);
+}
+
+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);
+}
+
+static size_t ggml_backend_numa_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+    return TENSOR_ALIGNMENT;
+
+    GGML_UNUSED(buft);
+}
+
+static bool ggml_backend_numa_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
+    return true;
+
+    GGML_UNUSED(buft);
+}
+
+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;
+}

src/llama-kv-cache.cpp

@@ -71,6 +71,13 @@ bool llama_kv_cache_init(
     cache.k_l.reserve(n_layer);
     cache.v_l.reserve(n_layer);
 
+    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 < 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();
@@ -82,7 +89,11 @@ bool llama_kv_cache_init(
             auto * dev = model.dev_layer(i);
             buft = ggml_backend_dev_buffer_type(dev);
         } else {
-            buft = ggml_backend_cpu_buffer_type();
+            if (is_numa) {
+                buft = ggml_backend_numa_buffer_type();
+            } else {
+                buft = ggml_backend_cpu_buffer_type();
+            }
         }
         ggml_context * ctx = ctx_for_buft(buft);