diff --git a/ggml/include/ggml-alloc.h b/ggml/include/ggml-alloc.h
index 434c13b34a..91194f2816 100644
--- a/ggml/include/ggml-alloc.h
+++ b/ggml/include/ggml-alloc.h
@@ -71,6 +71,9 @@ GGML_API size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_i
GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft);
GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, ggml_backend_t backend);
+// Export tensor allocations in a graph to a file that can be plotted
+GGML_API void ggml_gallocr_export_allocs(const char * filename, struct ggml_cgraph * graph);
+
#ifdef __cplusplus
}
#endif
diff --git a/ggml/include/ggml-backend.h b/ggml/include/ggml-backend.h
index 5f3f1e2869..b5c3f7cb8f 100644
--- a/ggml/include/ggml-backend.h
+++ b/ggml/include/ggml-backend.h
@@ -208,6 +208,9 @@ extern "C" {
// Set a callback to be called for each resulting node during graph compute
GGML_API void ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data);
+ // internal
+ GGML_API struct ggml_cgraph * ggml_backend_sched_get_graph_copy(ggml_backend_sched_t sched);
+
//
// Utils
//
diff --git a/ggml/src/ggml-alloc.c b/ggml/src/ggml-alloc.c
index e485326abc..65710cfd2e 100644
--- a/ggml/src/ggml-alloc.c
+++ b/ggml/src/ggml-alloc.c
@@ -1034,3 +1034,30 @@ ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_conte
ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, ggml_backend_t backend) {
return ggml_backend_alloc_ctx_tensors_from_buft(ctx, ggml_backend_get_default_buffer_type(backend));
}
+
+
+static void export_tensor(FILE * f, struct ggml_tensor * t) {
+ size_t offset = (uintptr_t)t->data - (uintptr_t)ggml_backend_buffer_get_base(t->buffer);
+ // [tensor_id] [tensor_view_src_id] [tensor_view_offs] [tensor_name] [buffer_id] [buffer_name] [offset] [size]
+ fprintf(f, "%p,%p,%zu,\"%s\",%p,\"%s\",%zu,%zu\n",
+ (void *)t, (void *)t->view_src, t->view_offs, t->name,
+ (void *)t->buffer, ggml_backend_buffer_name(t->buffer),
+ offset, ggml_backend_buft_get_alloc_size(t->buffer->buft, t));
+
+}
+
+void ggml_gallocr_export_allocs(const char * filename, struct ggml_cgraph * graph) {
+ FILE * f = fopen(filename, "wb");
+
+ fprintf(f, "tensor_id,tensor_view_src_id,tensor_view_offs,tensor_name,buffer_id,buffer_name,offset,size\n");
+
+ for (int i = 0; i < graph->n_leafs; i++) {
+ export_tensor(f, graph->leafs[i]);
+ }
+
+ for (int i = 0; i < graph->n_nodes; i++) {
+ export_tensor(f, graph->nodes[i]);
+ }
+
+ fclose(f);
+}
diff --git a/ggml/src/ggml-backend.c b/ggml/src/ggml-backend.c
index 954ab20725..a2600edba8 100644
--- a/ggml/src/ggml-backend.c
+++ b/ggml/src/ggml-backend.c
@@ -2028,6 +2028,10 @@ ggml_backend_t ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched,
return sched->backends[backend_index];
}
+GGML_API struct ggml_cgraph * ggml_backend_sched_get_graph_copy(ggml_backend_sched_t sched) {
+ return &sched->graph;
+}
+
// utils
void ggml_backend_view_init(struct ggml_tensor * tensor) {
diff --git a/plot-allocs.ipynb b/plot-allocs.ipynb
new file mode 100644
index 0000000000..2397735732
--- /dev/null
+++ b/plot-allocs.ipynb
@@ -0,0 +1,292 @@
+{
+ "cells": [
+ {
+ "cell_type": "code",
+ "execution_count": 77,
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "import pandas as pd\n",
+ "\n",
+ "df = pd.read_csv('allocs7.csv')\n",
+ "# remove all views\n",
+ "df = df[df['tensor_view_src_id'] == \"(nil)\"]\n",
+ "df = df.drop_duplicates(subset=['tensor_id'])\n",
+ "df = df.sort_values(by='size', ascending=False, kind='stable')"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 78,
+ "metadata": {},
+ "outputs": [
+ {
+ "data": {
+ "text/html": [
+ "\n",
+ "\n",
+ "
\n",
+ " \n",
+ " \n",
+ " | \n",
+ " buffer_id | \n",
+ " buffer_name | \n",
+ " offset+size | \n",
+ "
\n",
+ " \n",
+ " \n",
+ " \n",
+ " | 7 | \n",
+ " 0x56fa9b3b02f0 | \n",
+ " CPU | \n",
+ " 430940160 | \n",
+ "
\n",
+ " \n",
+ " | 1 | \n",
+ " 0x56fa95d0a890 | \n",
+ " CUDA0 | \n",
+ " 402653184 | \n",
+ "
\n",
+ " \n",
+ " | 5 | \n",
+ " 0x56fa9af97b00 | \n",
+ " CUDA0 | \n",
+ " 4294705152 | \n",
+ "
\n",
+ " \n",
+ " | 6 | \n",
+ " 0x56fa9afff860 | \n",
+ " CUDA0 | \n",
+ " 155197440 | \n",
+ "
\n",
+ " \n",
+ " | 0 | \n",
+ " 0x56fa95cd17d0 | \n",
+ " CUDA1 | \n",
+ " 134217728 | \n",
+ "
\n",
+ " \n",
+ " | 3 | \n",
+ " 0x56fa9ae68ec0 | \n",
+ " CUDA1 | \n",
+ " 321404928 | \n",
+ "
\n",
+ " \n",
+ " | 4 | \n",
+ " 0x56fa9af5b340 | \n",
+ " CUDA1 | \n",
+ " 1862524928 | \n",
+ "
\n",
+ " \n",
+ " | 2 | \n",
+ " 0x56fa9ae63710 | \n",
+ " CUDA_Host | \n",
+ " 41967616 | \n",
+ "
\n",
+ " \n",
+ "
\n",
+ "