Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-27 14:49:25 +01:00
Code Issues Packages Projects Releases Wiki Activity

ggml : dynamic ggml_sched_max_splits based on graph_size (#9047)

Browse Source 
* ggml : Dynamic ggml_sched_max_splits based on graph_size

* Fixed and readded debug code for causes
This commit is contained in:
Nico Bosshard 2024-08-16 04:22:55 +02:00 committed by GitHub
parent 4b9afbbe90
commit e3f6fd56b1
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 5 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
ggml/src/ggml-backend.c
View File

@ -1018,10 +1018,6 @@ static bool ggml_is_view_op(enum ggml_op op) {
#define GGML_SCHED_MAX_BACKENDS 16 #define GGML_SCHED_MAX_BACKENDS 16
#endif #endif
#ifndef GGML_SCHED_MAX_SPLITS
#define GGML_SCHED_MAX_SPLITS 2048
#endif
#ifndef GGML_SCHED_MAX_SPLIT_INPUTS #ifndef GGML_SCHED_MAX_SPLIT_INPUTS
#define GGML_SCHED_MAX_SPLIT_INPUTS GGML_MAX_SRC #define GGML_SCHED_MAX_SPLIT_INPUTS GGML_MAX_SRC
#endif #endif
@ -1125,7 +1121,8 @@ static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, co
} }
#if 0 #if 0
static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS][128]; // debug only #define GGML_SCHED_MAX_SPLITS_DEBUG 4096
static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_SCHED_MAX_SPLITS_DEBUG*GGML_SCHED_MAX_SPLIT_INPUTS][128]; // debug only
#define SET_CAUSE(node, ...) sprintf(causes[hash_id(node)], __VA_ARGS__) #define SET_CAUSE(node, ...) sprintf(causes[hash_id(node)], __VA_ARGS__)
#define GET_CAUSE(node) causes[hash_id(node)] #define GET_CAUSE(node) causes[hash_id(node)]
#else #else
@ -1549,7 +1546,6 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
sched->splits = realloc(sched->splits, sched->splits_capacity * sizeof(struct ggml_backend_sched_split)); sched->splits = realloc(sched->splits, sched->splits_capacity * sizeof(struct ggml_backend_sched_split));
GGML_ASSERT(sched->splits != NULL); GGML_ASSERT(sched->splits != NULL);
} }
GGML_ASSERT(i_split < GGML_SCHED_MAX_SPLITS);
split = &sched->splits[i_split]; split = &sched->splits[i_split];
split->backend_id = node_backend_id; split->backend_id = node_backend_id;
split->i_start = i; split->i_start = i;
@ -1865,13 +1861,14 @@ ggml_backend_sched_t ggml_backend_sched_new(
sched->hv_tensor_backend_ids = malloc(sched->hash_set.size * sizeof(sched->hv_tensor_backend_ids[0])); sched->hv_tensor_backend_ids = malloc(sched->hash_set.size * sizeof(sched->hv_tensor_backend_ids[0]));
sched->hv_tensor_copies = malloc(sched->hash_set.size * sched->n_backends * sched->n_copies * sizeof(struct ggml_tensor *)); sched->hv_tensor_copies = malloc(sched->hash_set.size * sched->n_backends * sched->n_copies * sizeof(struct ggml_tensor *));
const size_t nodes_size = graph_size + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS*2; const size_t ggml_sched_max_splits = graph_size; // at most there is one split for each node in the graph
const size_t nodes_size = graph_size + ggml_sched_max_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2;
sched->node_backend_ids = calloc(nodes_size, sizeof(sched->node_backend_ids[0])); sched->node_backend_ids = calloc(nodes_size, sizeof(sched->node_backend_ids[0]));
sched->leaf_backend_ids = calloc(nodes_size, sizeof(sched->leaf_backend_ids[0])); sched->leaf_backend_ids = calloc(nodes_size, sizeof(sched->leaf_backend_ids[0]));
sched->prev_node_backend_ids = calloc(nodes_size, sizeof(sched->prev_node_backend_ids[0])); sched->prev_node_backend_ids = calloc(nodes_size, sizeof(sched->prev_node_backend_ids[0]));
sched->prev_leaf_backend_ids = calloc(nodes_size, sizeof(sched->prev_leaf_backend_ids[0])); sched->prev_leaf_backend_ids = calloc(nodes_size, sizeof(sched->prev_leaf_backend_ids[0]));
sched->context_buffer_size = GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS*2*sizeof(struct ggml_tensor) + ggml_graph_overhead_custom(graph_size, false); sched->context_buffer_size = ggml_sched_max_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2*sizeof(struct ggml_tensor) + ggml_graph_overhead_custom(graph_size, false);
sched->context_buffer = malloc(sched->context_buffer_size); sched->context_buffer = malloc(sched->context_buffer_size);
const int initial_splits_capacity = 16; const int initial_splits_capacity = 16;