upstream-archive/stable-diffusion.cpp
1
0
Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2026-09-29 09:28:14 -05:00
Code Issues Packages Projects Releases Wiki Activity

fix: reserve 128 MiB headroom when selecting monolithic execution (#2046)

Browse Source
This commit is contained in:
leejet
2026-09-25 01:53:11 +08:00
committed by GitHub
parent 740c7ae193
commit 1a2330de68
2 changed files with 16 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/core/ggml_runner.cpp
+10 -1
View File
@@ -837,11 +837,20 @@ std::optional<Tensor<float>> GGMLRunner::execute_graph(ggml_cgraph* graph, int n
last_compute_status_ = GGML_STATUS_ALLOC_FAILED;
return std::nullopt;
}
auto fits_monolithic = [&]() {
// Planning headroom absorbs allocation estimate drift; execution keeps the normal limits.
constexpr size_t planning_headroom = 128ULL * 1024ULL * 1024ULL;
auto requests = memory_requests(full_measurement.buffers, cache_.pending_bytes(graph));
for (auto& request : requests) {
request.pending_allocation_bytes = add_bytes(request.pending_allocation_bytes, planning_headroom);
}
return fits(requests, params);
};
auto manager = residency_manager.lock();
const bool segmented = !is_multi_device() && !sd_backend_is_cpu(runtime_backend) &&
manager != nullptr && manager->segmented_compute_enabled() &&
cached_plan.valid && cached_plan.has_cuts && cached_plan.segments.size() > 1 &&
!fits(memory_requests(full_measurement.buffers, cache_.pending_bytes(graph)), params);
!fits_monolithic();
ggml_graph_cut::Plan monolithic_plan;
if (!segmented) {
monolithic_plan.segments.emplace_back();