CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-28 08:31:25 +00:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: sort graph to allow more parallel execution (#15850)

Browse Source 
* vulkan: sort graph to allow more parallel execution

Add a backend proc to allow the backend to modify the graph. The
vulkan implementation looks at which nodes depend on each other
and greedily reorders them to group together nodes that don't
depend on each other. It only reorders the nodes, doesn't change
the contents of any of them.

With #15489, this reduces the number of synchronizations needed.

* call optimize_graph per-split
This commit is contained in:
Jeff Bolz
2025-09-08 13:10:07 -05:00
committed by GitHub
parent 0a16bf52e6
commit e68aa10d8f
13 changed files with 154 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
ggml/src/ggml-backend.cpp
View File

@@ -463,6 +463,13 @@ void ggml_backend_event_wait(ggml_backend_t backend, ggml_backend_event_t event)
backend->iface.event_wait(backend, event);
}
static void ggml_backend_optimize_graph(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
GGML_ASSERT(backend);
if (backend->iface.optimize_graph != NULL) {
backend->iface.optimize_graph(backend, cgraph);
}
}
// Backend device
const char * ggml_backend_dev_name(ggml_backend_dev_t device) {
@@ -1298,6 +1305,10 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
struct ggml_backend_sched_split * split = &sched->splits[i];
split->graph = ggml_graph_view(graph, split->i_start, split->i_end);
// Optimize this split of the graph. This needs to happen before we make graph_copy,
// so they are in sync.
ggml_backend_optimize_graph(sched->backends[split->backend_id], &split->graph);
// add inputs to the graph copy so that they are allocated by ggml-alloc at the start of the split
for (int j = 0; j < split->n_inputs; j++) {
assert(graph_copy->size > (graph_copy->n_nodes + 1));