tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-08-06 01:05:03 -04:00
Code Issues Packages Projects Releases Wiki Activity

context : always use non-causal attention for encoder graphs

Browse Source 
ggml-ci
This commit is contained in:
Georgi Gerganov
2025-03-18 11:14:48 +02:00
parent d9a14523bb
commit a0554c3cdc
1 changed files with 16 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
src/llama-context.cpp
View File

@@ -1627,7 +1627,16 @@ llm_graph_result_ptr llama_context::graph_build(
ggml_cgraph * gf,
const llama_ubatch & ubatch,
llm_graph_type gtype) {
return model.build_graph(
const auto causal_attn_org = cparams.causal_attn;
// always use non-causal attention for encoder graphs
// TODO: this is a tmp solution until we have a proper way to support enc-dec models
// ref: https://github.com/ggml-org/llama.cpp/pull/12181#issuecomment-2730451223
if (gtype == LLM_GRAPH_TYPE_ENCODER) {
cparams.causal_attn = false;
}
auto res = model.build_graph(
{
/*.ctx =*/ ctx,
/*.arch =*/ model.arch,
@@ -1643,6 +1652,12 @@ llm_graph_result_ptr llama_context::graph_build(
/*.n_outputs =*/ n_outputs,
/*.cb =*/ graph_get_cb(),
}, gf, gtype);
if (gtype == LLM_GRAPH_TYPE_ENCODER) {
cparams.causal_attn = causal_attn_org;
}
return res;
}
ggml_status llama_context::graph_compute(