vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

test-backend-ops : check buffer allocation failures

Browse source
This commit is contained in:
slaren 2024-01-07 13:50:02 +01:00
parent f77c72f371
commit 7c16cf106d
3 changed files with 53 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

37
ggml-backend.c
View file

@ -1503,6 +1503,21 @@ struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, s
struct ggml_context * ctx_allocated = ggml_init(params);
struct ggml_context * ctx_unallocated = ggml_init(params);
if (ctx_allocated == NULL || ctx_unallocated == NULL) {
fprintf(stderr, "failed to allocate context for graph copy\n");
free(hash_set.keys);
free(node_copies);
free(node_init);
ggml_free(ctx_allocated);
ggml_free(ctx_unallocated);
return (struct ggml_backend_graph_copy) {
/* .buffer = */ NULL,
/* .ctx_allocated = */ NULL,
/* .ctx_unallocated = */ NULL,
/* .graph = */ NULL,
};
}
// dup nodes
for (int i = 0; i < graph->n_nodes; i++) {
struct ggml_tensor * node = graph->nodes[i];
@ -1511,6 +1526,20 @@ struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, s
// allocate nodes
ggml_backend_buffer_t buffer = ggml_backend_alloc_ctx_tensors(ctx_allocated, backend);
if (buffer == NULL) {
fprintf(stderr, "failed to allocate buffer for graph copy\n");
free(hash_set.keys);
free(node_copies);
free(node_init);
ggml_free(ctx_allocated);
ggml_free(ctx_unallocated);
return (struct ggml_backend_graph_copy) {
/* .buffer = */ NULL,
/* .ctx_allocated = */ NULL,
/* .ctx_unallocated = */ NULL,
/* .graph = */ NULL,
};
}
//printf("copy buffer size: %zu MB\n", ggml_backend_buffer_get_size(buffer) / 1024 / 1024);
@ -1547,8 +1576,12 @@ void ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy) {
ggml_free(copy.ctx_unallocated);
}
void ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data) {
bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data) {
struct ggml_backend_graph_copy copy = ggml_backend_graph_copy(backend2, graph);
if (copy.buffer == NULL) {
return false;
}
struct ggml_cgraph * g1 = graph;
struct ggml_cgraph * g2 = copy.graph;
@ -1578,4 +1611,6 @@ void ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t
}
ggml_backend_graph_copy_free(copy);
return true;
}

2
ggml-backend.h
View file

@ -189,7 +189,7 @@ extern "C" {
typedef bool (*ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);
// Compare the output of two backends
GGML_API void ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);
GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);
// Tensor initialization
GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);

18
tests/test-backend-ops.cpp
View file

@ -376,6 +376,11 @@ struct test_case {
// allocate
ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors(ctx, backend1);
if (buf == NULL) {
printf("failed to allocate tensors [%s] ", ggml_backend_name(backend1));
ggml_free(ctx);
return false;
}
// build graph
ggml_build_forward_expand(gf, out);
@ -463,9 +468,13 @@ struct test_case {
GGML_UNUSED(index);
};
ggml_backend_compare_graph_backend(backend1, backend2, gf, callback, &ud);
bool cmp_ok = ggml_backend_compare_graph_backend(backend1, backend2, gf, callback, &ud);
if (ud.ok) {
if (!cmp_ok) {
printf("compare failed ");
}
if (ud.ok && cmp_ok) {
printf("\033[1;32mOK\033[0m\n");
} else {
printf("\033[1;31mFAIL\033[0m\n");
@ -519,6 +528,11 @@ struct test_case {
// allocate
ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors(ctx, backend);
if (buf == NULL) {
printf("failed to allocate tensors\n");
ggml_free(ctx);
return false;
}
// randomize tensors
initialize_tensors(ctx);