DarthAffe/stable-diffusion.cpp
1
0
Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2026-06-23 14:46:39 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix: avoid writable mmap for read-only weights

Browse Source
This commit is contained in:
leejet 2026-06-23 00:26:20 +08:00
parent 41f7acbfb0
commit 2666e2a658
6 changed files with 24 additions and 22 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
src/model_manager.cpp
View File

@ -480,7 +480,7 @@ bool ModelManager::mmap_params(const std::vector<TensorState*>& states,
return true;
}
auto mmap_store = model_loader_.mmap_tensors(mmap_candidates, {}, true);
auto mmap_store = model_loader_.mmap_tensors(mmap_candidates, {}, writable_mmap_);
if (mmap_store.empty()) {
return true;
}

2
src/model_manager.h
View File

@ -69,6 +69,7 @@ private:
uint64_t current_lora_epoch_ = 0;
int n_threads_ = 0;
bool enable_mmap_ = false;
bool writable_mmap_ = false;
void finish_compute_backend_usage(const std::vector<TensorState*>& states);
void release_all();
@ -110,6 +111,7 @@ public:
model_loader_.set_n_threads(n_threads);
}
void set_enable_mmap(bool enable_mmap) { enable_mmap_ = enable_mmap; }
void set_writable_mmap(bool writable_mmap) { writable_mmap_ = writable_mmap; }
void set_common_ignore_tensors(std::set<std::string> ignore_tensors);
void set_loras(std::vector<LoraSpec> loras, SDVersion version);

2
src/runtime/guidance.cpp
View File

@ -3,9 +3,9 @@
#include <algorithm>
#include <cmath>
#include <cstdlib>
#include <optional>
#include <string>
#include <utility>
#include <optional>
#include "core/util.h"

2
src/runtime/guidance.h
View File

@ -3,8 +3,8 @@
#include <cstddef>
#include <functional>
#include <vector>
#include <optional>
#include <vector>
#include "core/tensor.hpp"

8
src/stable-diffusion.cpp
View File

@ -532,7 +532,6 @@ public:
if (wtype != GGML_TYPE_COUNT || tensor_type_rules.size() > 0) {
model_loader.set_wtype_override(wtype, tensor_type_rules);
}
model_loader.process_model_files(enable_mmap, true);
std::map<ggml_type, uint32_t> wtype_stat = model_loader.get_wtype_stat();
std::map<ggml_type, uint32_t> conditioner_wtype_stat = model_loader.get_conditioner_wtype_stat();
@ -586,9 +585,12 @@ public:
apply_lora_immediately = false;
}
bool needs_writable_mmap = enable_mmap && apply_lora_immediately;
model_manager->set_writable_mmap(needs_writable_mmap);
if (enable_mmap && apply_lora_immediately) {
LOG_WARN("in mode 'immediately', LoRAs will cause extra memory usage with mmap");
}
model_loader.process_model_files(enable_mmap, needs_writable_mmap);
load_alphas_cumprod(model_loader);
size_t text_encoder_params_mem_size = 0;
@ -2208,9 +2210,7 @@ public:
guidance_input.pred_uncond = uncond_out.empty() ? nullptr : &uncond_out;
guidance_input.pred_img_uncond = img_uncond_out.empty() ? nullptr : &img_uncond_out;
sd::guidance::GuiderOutput guided = guidance_schedule.empty()?
primary_guidance.forward(guidance_input, {}):
primary_guidance.forward(guidance_input, {}, guidance_schedule[guidance_schedule.size() - 1 - step]);
sd::guidance::GuiderOutput guided = guidance_schedule.empty() ? primary_guidance.forward(guidance_input, {}) : primary_guidance.forward(guidance_input, {}, guidance_schedule[guidance_schedule.size() - 1 - step]);
if (guided.pred.empty()) {
return {};
}