Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-01-01 00:39:00 +01:00
Code Issues Packages Projects Releases Wiki Activity

allocator cleanup

Browse Source
This commit is contained in:
slaren 2023-07-22 13:29:44 +02:00
parent 7de7882537
commit e2b9575951
3 changed files with 12 additions and 18 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
ggml-backend.c
View File

@ -264,7 +264,6 @@ void ggml_allocator_default_free_tensor(struct ggml_backend_buffer * alloc, stru
size_t size = ggml_backend_buffer_get_alloc_size(alloc, tensor);
size = aligned_offset(NULL, size, allocator_ctx->alignment);
AT_PRINTF("%s: freeing %s (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, size, allocator_ctx->n_free_blocks);
tensor->freed = true;
#ifdef GGML_ALLOCATOR_DEBUG
remove_allocated_tensor(allocator_ctx, tensor);
@ -858,13 +857,12 @@ static void ggml_graph_allocate_tensors_n(
struct ggml_tensor * node = gf->nodes[i];
node->n_children = 0;
node->n_views = 0;
//node->freed = false;
}
for (int i = 0; i < gf->n_leafs; i++) {
struct ggml_tensor * leaf = gf->leafs[i];
leaf->n_children = 0;
leaf->n_views = 0;
//leaf->freed = false;
}
}
@ -912,7 +910,6 @@ static void ggml_graph_allocate_tensors_n(
if (parent == NULL) {
break;
}
GGML_ASSERT(!parent->freed && "tensor used after free");
allocate_node(buffer, parent);
}

14
ggml.c
View File

@ -4530,16 +4530,15 @@ struct ggml_tensor * ggml_new_tensor_impl(
/*.is_param =*/ false,
/*.grad =*/ NULL,
/*.src =*/ { NULL },
/*.node_id =*/ -1,
/*.visited =*/ false,
/*.n_children =*/ 0,
/*.n_views =*/ 0,
/*.freed =*/ false,
/*.perf_runs =*/ 0,
/*.perf_cycles =*/ 0,
/*.perf_time_us =*/ 0,
/*.data =*/ data,
/*.name =*/ { 0 },
/*.extra =*/ NULL,
/*.name =*/ { 0 },
/*.padding =*/ { 0 },
};
@ -15818,9 +15817,10 @@ static void ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor *
}
// check if already visited
if (node->node_id != -1) {
if (node->visited) {
return;
}
node->visited = true;
for (int i = 0; i < GGML_MAX_SRC; ++i) {
if (node->src[i]) {
@ -15839,7 +15839,6 @@ static void ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor *
ggml_format_name(node, "leaf_%d", cgraph->n_leafs);
}
node->node_id = cgraph->n_leafs;
cgraph->leafs[cgraph->n_leafs] = node;
cgraph->n_leafs++;
} else {
@ -15849,7 +15848,6 @@ static void ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor *
ggml_format_name(node, "node_%d", cgraph->n_nodes);
}
node->node_id = cgraph->n_nodes;
cgraph->nodes[cgraph->n_nodes] = node;
cgraph->grads[cgraph->n_nodes] = node->grad;
cgraph->n_nodes++;
@ -15883,10 +15881,10 @@ void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor *
// TODO: this can be removed when ggml_build_forward_expand is removed
void ggml_graph_close(struct ggml_cgraph * cgraph) {
for (int i = 0; i < cgraph->n_nodes; ++i) {
cgraph->nodes[i]->node_id = -1;
cgraph->nodes[i]->visited = false;
}
for (int i = 0; i < cgraph->n_leafs; ++i) {
cgraph->leafs[i]->node_id = -1;
cgraph->leafs[i]->visited = false;
}
cgraph->closed = true;
}

11
ggml.h
View File

@ -422,10 +422,9 @@ extern "C" {
struct ggml_tensor * grad;
struct ggml_tensor * src[GGML_MAX_SRC];
int node_id; // used to build graphs
int n_children;
bool visited; // used to build graphs
int n_children; // used by the allocator
int n_views;
bool freed; // debug
// performance
int perf_runs;
@ -434,11 +433,11 @@ extern "C" {
void * data;
char name[GGML_MAX_NAME];
void * extra; // extra things e.g. for ggml-cuda.cu
char padding[8];
char name[GGML_MAX_NAME];
char padding[12];
};
static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor);