CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-31 08:51:55 +00:00
Code Issues Packages Projects Releases Wiki Activity

ggml : reduce hash table reset cost (#8698)

Browse Source 
* ggml : reduce hash table reset cost

* fix unreachable code warnings after GGML_ASSERT(false)

* GGML_ASSERT(false) -> GGML_ABORT("fatal error")

* GGML_ABORT use format string
This commit is contained in:
slaren
2024-07-27 04:41:55 +02:00
committed by GitHub
parent 01245f5b16
commit 2b1f616b20
46 changed files with 851 additions and 754 deletions
Download Patch File Download Diff File Expand all files Collapse all files

42
ggml/src/ggml-alloc.c
View File

@@ -91,8 +91,7 @@ void ggml_tallocr_alloc(struct ggml_tallocr * talloc, struct ggml_tensor * tenso
if (talloc->offset + size > ggml_backend_buffer_get_size(talloc->buffer)) {
fprintf(stderr, "%s: not enough space in the buffer to allocate %s (needed %zu, available %zu)\n",
__func__, tensor->name, size, ggml_backend_buffer_get_size(talloc->buffer) - talloc->offset);
GGML_ASSERT(!"not enough space in the buffer");
return;
GGML_ABORT("not enough space in the buffer");
}
void * addr = (char *)ggml_backend_buffer_get_base(talloc->buffer) + talloc->offset;
@@ -133,7 +132,7 @@ static void add_allocated_tensor(struct ggml_dyn_tallocr * alloc, size_t offset,
return;
}
}
GGML_ASSERT(!"out of allocated_tensors");
GGML_ABORT("out of allocated_tensors");
}
static void remove_allocated_tensor(struct ggml_dyn_tallocr * alloc, size_t offset, const struct ggml_tensor * tensor) {
for (int i = 0; i < 1024; i++) {
@@ -142,8 +141,7 @@ static void remove_allocated_tensor(struct ggml_dyn_tallocr * alloc, size_t offs
return;
}
}
fprintf(stderr, "tried to free tensor %s not found\n", tensor->name);
GGML_ASSERT(!"tensor not found");
GGML_ABORT("tried to free tensor %s not found\n", tensor->name);
}
#endif
@@ -176,8 +174,7 @@ static size_t ggml_dyn_tallocr_alloc(struct ggml_dyn_tallocr * alloc, size_t siz
// this should never happen
fprintf(stderr, "%s: not enough space in the buffer to allocate %zu bytes, largest block available %zu bytes\n",
__func__, size, max_avail);
GGML_ASSERT(!"not enough space in the buffer");
GGML_UNREACHABLE();
GGML_ABORT("not enough space in the buffer");
}
}
@@ -443,7 +440,7 @@ void ggml_gallocr_free(ggml_gallocr_t galloc) {
}
}
free(galloc->hash_set.keys);
ggml_hash_set_free(&galloc->hash_set);
free(galloc->hash_values);
free(galloc->bufts);
free(galloc->buffers);
@@ -456,7 +453,7 @@ void ggml_gallocr_free(ggml_gallocr_t galloc) {
typedef struct ggml_gallocr * ggml_gallocr_t;
static struct hash_node * ggml_gallocr_hash_get(ggml_gallocr_t galloc, struct ggml_tensor * t) {
size_t i = ggml_hash_find_or_insert(galloc->hash_set, t);
size_t i = ggml_hash_find_or_insert(&galloc->hash_set, t);
return &galloc->hash_values[i];
}
@@ -565,8 +562,8 @@ static int get_node_buffer_id(const int * node_buffer_ids, int i) {
static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids, const int * leaf_buffer_ids) {
// clear hash tables
memset(galloc->hash_set.keys, 0, galloc->hash_set.size * sizeof(struct ggml_tensor *));
memset(galloc->hash_values, 0, galloc->hash_set.size * sizeof(struct hash_node));
ggml_hash_set_reset(&galloc->hash_set);
memset(galloc->hash_values, 0, sizeof(struct hash_node) * galloc->hash_set.size);
// allocate leafs
// these may be tensors that the application is not using in the graph, but may still want to allocate for other purposes
@@ -671,21 +668,19 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
}
bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids, const int * leaf_buffer_ids) {
size_t hash_size = graph->visited_hash_table.size;
size_t min_hash_size = graph->n_nodes + graph->n_leafs;
// add 25% margin to avoid hash collisions
min_hash_size += min_hash_size / 4;
// initialize hash table
if (galloc->hash_set.size < hash_size) {
free(galloc->hash_set.keys);
free(galloc->hash_values);
galloc->hash_set.size = hash_size;
galloc->hash_set.keys = calloc(hash_size, sizeof(struct ggml_tensor *));
galloc->hash_values = calloc(hash_size, sizeof(struct hash_node));
if (galloc->hash_set.size < min_hash_size) {
ggml_hash_set_free(&galloc->hash_set);
galloc->hash_set = ggml_hash_set_new(min_hash_size);
GGML_ASSERT(galloc->hash_set.keys != NULL);
free(galloc->hash_values);
galloc->hash_values = malloc(sizeof(struct hash_node) * galloc->hash_set.size);
GGML_ASSERT(galloc->hash_values != NULL);
} else {
// reset hash table
memset(galloc->hash_set.keys, 0, sizeof(struct ggml_tensor *) * galloc->hash_set.size);
memset(galloc->hash_values, 0, sizeof(struct hash_node) * galloc->hash_set.size);
}
// reset allocators
@@ -817,8 +812,7 @@ static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor *
}
static bool ggml_gallocr_node_needs_realloc(ggml_gallocr_t galloc, struct ggml_tensor * node, struct tensor_alloc * talloc) {
ggml_backend_buffer_type_t buft = talloc->buffer_id != -1 ? galloc->bufts[talloc->buffer_id] : NULL;
size_t node_size = (node->data || node->view_src) ? 0 : ggml_backend_buft_get_alloc_size(buft, node);
size_t node_size = (node->data || node->view_src) ? 0 : ggml_backend_buft_get_alloc_size(galloc->bufts[talloc->buffer_id], node);
return talloc->size_max >= node_size;
}