DarthAffe/stable-diffusion.cpp
1
0
Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2026-09-25 04:32:29 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix: reuse graph cut plans across CFG passes (#1943)

Browse Source
This commit is contained in:
leejet 2026-09-07 00:39:49 +08:00 committed by GitHub
parent 80bac2d5fc
commit d8fb10c029
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 25 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

31
src/core/ggml_graph_cut.cpp
View File

@ -537,10 +537,12 @@ namespace sd::ggml_graph_cut {
return signature; return signature;
} }
bool plan_matches_graph(ggml_cgraph* gf, const Plan& plan) { static bool plan_matches_graph(ggml_cgraph* gf,
const Plan& plan,
const std::vector<uint64_t>& layout) {
GGML_ASSERT(gf != nullptr); GGML_ASSERT(gf != nullptr);
if (plan.leaf_names.size() != static_cast<size_t>(gf->n_leafs) || if (plan.leaf_names.size() != static_cast<size_t>(gf->n_leafs) ||
plan.layout != graph_layout(gf, false)) { plan.layout != layout) {
return false; return false;
} }
for (int i = 0; i < gf->n_leafs; ++i) { for (int i = 0; i < gf->n_leafs; ++i) {
@ -558,6 +560,11 @@ namespace sd::ggml_graph_cut {
return cut_markers == plan.cut_markers; return cut_markers == plan.cut_markers;
} }
bool plan_matches_graph(ggml_cgraph* gf, const Plan& plan) {
GGML_ASSERT(gf != nullptr);
return plan_matches_graph(gf, plan, graph_layout(gf, false));
}
ggml_tensor* output_tensor(ggml_cgraph* gf, const Segment& segment, size_t output_index) { ggml_tensor* output_tensor(ggml_cgraph* gf, const Segment& segment, size_t output_index) {
GGML_ASSERT(gf != nullptr); GGML_ASSERT(gf != nullptr);
if (output_index >= segment.output_node_indices.size()) { if (output_index >= segment.output_node_indices.size()) {
@ -938,20 +945,26 @@ namespace sd::ggml_graph_cut {
GGML_ASSERT(gf != nullptr); GGML_ASSERT(gf != nullptr);
GGML_ASSERT(cache != nullptr); GGML_ASSERT(cache != nullptr);
if (cache->graph_cut_plan.available && const auto layout = graph_layout(gf, false);
plan_matches_graph(gf, cache->graph_cut_plan)) { auto& plans = cache->graph_cut_plans;
return cache->graph_cut_plan; for (auto it = plans.begin(); it != plans.end(); ++it) {
if (it->available && plan_matches_graph(gf, *it, layout)) {
plans.splice(plans.begin(), plans, it);
return plans.front();
}
} }
int64_t t_plan_begin = ggml_time_ms(); int64_t t_plan_begin = ggml_time_ms();
Plan plan = build_plan(backend, gf, params_tensor_set, log_desc); plans.push_front(build_plan(backend, gf, params_tensor_set, log_desc));
cache->graph_cut_plan = plan; if (plans.size() > PlanCache::MAX_PLANS) {
plans.pop_back();
}
if (log_desc != nullptr) { if (log_desc != nullptr) {
LOG_INFO("%s build cached graph cut plan done (taking %lld ms)", LOG_INFO("%s build cached graph cut plan done (taking %lld ms)",
log_desc, log_desc,
ggml_time_ms() - t_plan_begin); ggml_time_ms() - t_plan_begin);
} }
return plan; return plans.front();
} }
} // namespace sd::ggml_graph_cut } // namespace sd::ggml_graph_cut

4
src/core/ggml_graph_cut.h
View File

@ -3,6 +3,7 @@
#include <array> #include <array>
#include <cstdint> #include <cstdint>
#include <list>
#include <string> #include <string>
#include <unordered_map> #include <unordered_map>
#include <unordered_set> #include <unordered_set>
@ -48,7 +49,8 @@ namespace sd::ggml_graph_cut {
}; };
struct PlanCache { struct PlanCache {
Plan graph_cut_plan; static constexpr size_t MAX_PLANS = 4;
std::list<Plan> graph_cut_plans;
}; };
static constexpr const char* GGML_RUNNER_CUT_PREFIX = "ggml_runner_cut:"; static constexpr const char* GGML_RUNNER_CUT_PREFIX = "ggml_runner_cut:";