DarthAffe/stable-diffusion.cpp
1
0 Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2026-11-12 11:56:36 +00:00
Code Issues Packages Projects Releases Wiki Activity

perf: optimize H3 inference and segmented weight prefetch (#2240)

Browse Source
This commit is contained in:
leejet authored and GitHub committed 2026-10-30 00:38:14 +08:00
1 parent e7b7ec6d6f
commit 37bc88b5e9
17 files changed
+367 -210
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

src/core/runner_cache.cpp
+45 -9
View File
@@ -24,15 +24,7 @@ namespace sd {
ggml_free(context);
}
std::unique_ptr<CachedTensor> CachedTensor::copy(ggml_backend_t backend,
const std::string& name,
ggml_tensor* source,
ggml_status& status) {
status = GGML_STATUS_FAILED;
if (ggml_graph_cut::tensor_buffer(source) == nullptr) {
return nullptr;
}
status = GGML_STATUS_ALLOC_FAILED;
std::unique_ptr<CachedTensor> CachedTensor::allocate(ggml_backend_t backend, const std::string& name, ggml_tensor* source) {
auto entry = std::make_unique<CachedTensor>();
entry->context = ggml_init({2 * ggml_tensor_overhead(), nullptr, true});
if (entry->context == nullptr) {
@@ -46,6 +38,22 @@ namespace sd {
if (entry->buffer == nullptr) {
return nullptr;
}
return entry;
}
std::unique_ptr<CachedTensor> CachedTensor::copy(ggml_backend_t backend,
const std::string& name,
ggml_tensor* source,
ggml_status& status) {
status = GGML_STATUS_FAILED;
if (ggml_graph_cut::tensor_buffer(source) == nullptr) {
return nullptr;
}
status = GGML_STATUS_ALLOC_FAILED;
auto entry = allocate(backend, name, source);
if (entry == nullptr) {
return nullptr;
}
if (source->view_src != nullptr || !ggml_is_contiguous(source) || source->buffer == nullptr) {
std::vector<uint8_t> data(ggml_nbytes(source));
ggml_backend_tensor_get(source, data.data(), 0, data.size());
@@ -155,6 +163,26 @@ namespace sd {
return sd::resident_bytes(tensors_, device);
}
bool GraphCutTensorCache::can_bind_output(ggml_backend_t backend, const ggml_tensor* tensor) {
return tensor != nullptr && tensor->view_src == nullptr && tensor->op != GGML_OP_NONE &&
ggml_is_contiguous(tensor) && ggml_backend_supports_op(backend, tensor);
}
bool GraphCutTensorCache::bind_outputs(const std::unordered_set<const ggml_tensor*>& outputs) {
for (const ggml_tensor* output : outputs) {
auto tensor = const_cast<ggml_tensor*>(output);
auto entry = CachedTensor::allocate(backend_, tensor->name, tensor);
if (entry == nullptr) {
return false;
}
tensor->buffer = entry->tensor->buffer;
tensor->data = entry->tensor->data;
tensor->extra = entry->tensor->extra;
tensors_[tensor->name] = std::move(entry);
}
return true;
}
size_t GraphCutTensorCache::estimate_output_bytes(
ggml_cgraph* graph,
const ggml_graph_cut::Segment& segment) const {
@@ -171,6 +199,10 @@ namespace sd {
!segment.future_cut_names.count(output->name)) {
continue;
}
const ggml_tensor* cached = get(output->name);
if (cached != nullptr && output->buffer == cached->buffer && output->data == cached->data) {
continue;
}
ggml_tensor* source = ggml_graph_cut::cache_source_tensor(output);
const size_t tensor_size = GGML_PAD(
ggml_backend_buft_get_alloc_size(buffer_type, source), alignment);
@@ -196,6 +228,10 @@ namespace sd {
!segment.future_cut_names.count(output->name)) {
continue;
}
const ggml_tensor* cached = get(output->name);
if (cached != nullptr && output->buffer == cached->buffer && output->data == cached->data) {
continue;
}
ggml_status status;
auto entry = CachedTensor::copy(backend_, output->name, ggml_graph_cut::cache_source_tensor(output), status);
if (entry == nullptr) {