feat: add configurable Qwen cache types and early cache scheduling (#2045)

This commit is contained in:
leejet authored and GitHub committed 2026-09-25 01:51:01 +08:00
1 parent 4dfe8f5d45
commit 740c7ae193
11 files changed
+185 -61

No files matched your search

+2 -3
View File
@@ -376,7 +376,7 @@ struct ControlNet : public GGMLRunner {
hint = make_input(hint_tensor);
}
auto runner_ctx = get_context();
auto runner_ctx = get_context(gf);
auto outs = control_net.forward(&runner_ctx,
x,
@@ -389,8 +389,7 @@ struct ControlNet : public GGMLRunner {
if (guided_hint_input == nullptr && !outs.empty()) {
guided_hint_output_ggml = outs[0];
ggml_set_output(guided_hint_output_ggml);
cache(guided_hint_cache_name(), guided_hint_output_ggml);
ggml_build_forward_expand(gf, guided_hint_output_ggml);
runner_ctx.persist_cache_tensor(guided_hint_cache_name(), guided_hint_output_ggml);
}
control_outputs_ggml.reserve(outs.size() > 0 ? outs.size() - 1 : 0);