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

llama : separate compute buffer reserve from fattn check (#15696)

Browse Source 
Exposes ggml_backend_sched_split_graph() to allow splitting the graph without allocating compute buffers and uses it to split the graph for the automatic Flash Attention check.
This commit is contained in:
Diego Devesa
2025-08-31 06:49:03 -07:00
committed by GitHub
parent 7d3c9f2b21
commit 9777032dcc
4 changed files with 64 additions and 58 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
ggml/src/ggml-backend.cpp
View File

@@ -902,7 +902,7 @@ static void ggml_backend_sched_set_if_supported(ggml_backend_sched_t sched, stru
}
// assigns backends to ops and splits the graph into subgraphs that can be computed on the same backend
static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
// reset splits
sched->n_splits = 0;
sched->n_graph_inputs = 0;
@@ -1687,6 +1687,8 @@ bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph *
GGML_ASSERT(sched);
GGML_ASSERT((int)sched->hash_set.size >= measure_graph->n_nodes + measure_graph->n_leafs);
ggml_backend_sched_reset(sched);
ggml_backend_sched_synchronize(sched);
ggml_backend_sched_split_graph(sched, measure_graph);