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

backend : add eval callback (#4935)

Browse source 
* backend : add eval callback

ggml-ci

* backend : group nodes in a single compute when user don't need them

* backend : clean-up the implementation

ggml-ci

* simple : do not perform tensor data copy if not needed

* simple : fix

* simple : no need for ggml_is_contiguous + fix bool parse

* llama : fix callback placement in llama_context_params

* backend : avoid double-ask callback calls

* simple : restore examples, imatrix will serve as a demo
This commit is contained in:
Georgi Gerganov 2024-01-17 18:39:41 +02:00 • committed by GitHub
parent c918fe8dca
commit 44a1a4a41a
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 64 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

42
ggml-backend.c
View file

@ -802,6 +802,9 @@ struct ggml_backend_sched {
__attribute__((aligned(GGML_MEM_ALIGN)))
#endif
char context_buffer[GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS*sizeof(struct ggml_tensor) + sizeof(struct ggml_cgraph)];
ggml_backend_sched_eval_callback callback_eval;
void * callback_eval_user_data;
};
#define hash_id(node) ggml_hash_find_or_insert(sched->hash_set, node)
@ -1324,9 +1327,38 @@ static void sched_compute_splits(ggml_backend_sched_t sched) {
ggml_graph_dump_dot(split->graph, NULL, split_filename);
#endif
uint64_t compute_start_us = ggml_time_us();
ggml_backend_graph_compute(split_backend, &split->graph);
//ggml_backend_synchronize(split_backend); // necessary to measure compute time
if (!sched->callback_eval) {
ggml_backend_graph_compute(split_backend, &split->graph);
//ggml_backend_synchronize(split_backend); // necessary to measure compute time
} else {
// similar to ggml_backend_compare_graph_backend
for (int j0 = 0; j0 < split->graph.n_nodes; j0++) {
struct ggml_tensor * t = split->graph.nodes[j0];
// check if the user needs data from this node
bool need = sched->callback_eval(t, true, sched->callback_eval_user_data);
int j1 = j0;
// determine the range [j0, j1] of nodes that can be computed together
while (!need && j1 < split->graph.n_nodes - 1) {
t = split->graph.nodes[++j1];
need = sched->callback_eval(t, true, sched->callback_eval_user_data);
}
struct ggml_cgraph gv = ggml_graph_view(&split->graph, j0, j1 + 1);
ggml_backend_graph_compute(split_backend, &gv);
if (need && !sched->callback_eval(t, false, sched->callback_eval_user_data)) {
break;
}
j0 = j1;
}
}
uint64_t compute_end_us = ggml_time_us();
compute_us[split_backend_id] += compute_end_us - compute_start_us;
}
@ -1431,6 +1463,12 @@ void ggml_backend_sched_reset(ggml_backend_sched_t sched) {
sched_reset(sched);
}
void ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data) {
sched->callback_eval = callback;
sched->callback_eval_user_data = user_data;
}
int ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched) {
return sched->n_splits;
}