CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-27 08:21:30 +00:00
Code Issues Packages Projects Releases Wiki Activity

CANN: Disable acl_graph for prefill stage (#15933)

Browse Source 
Since the prefill length is not fixed, graphs constructed for the
prefill stage cannot be reused. For this reason, ACL graph
execution is disabled by default during prefill.
This commit is contained in:
hipudding
2025-09-11 15:59:37 +08:00
committed by GitHub
parent 00681dfc16
commit c0389dba43
2 changed files with 19 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
ggml/src/ggml-cann/ggml-cann.cpp
View File

@@ -2360,6 +2360,21 @@ static enum ggml_status ggml_backend_cann_graph_compute(
bool use_cann_graph = true;
bool cann_graph_update_required = false;
static bool prefill_use_graph = parse_bool(get_env("GGML_CANN_PREFILL_USE_GRAPH").value_or(""));
if (!prefill_use_graph) {
// Do not use acl_graph for prefill.
for (int i = 0; i < cgraph->n_nodes; i++) {
ggml_tensor * node = cgraph->nodes[i];
// TODO: Optimize here. Currently, we can only
// get seq_len by FA's input.
if (node->op == GGML_OP_FLASH_ATTN_EXT) {
// Q -> src[0], shape: [B, S, N, D]
use_cann_graph = (node->src[0]->ne[1] == 1);
break;
}
}
}
if (!cann_ctx->acl_graph_mode) {
use_cann_graph = false;
}