Skip to content

fix: avoid writable mmap for read-only weights #1698

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
leejet merged 1 commit into from
Jun 22, 2026
Merged

fix: avoid writable mmap for read-only weights #1698

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 3 additions & 3 deletions examples/common/common.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -261,15 +261,15 @@ bool parse_options(int argc, const char** argv, const std::vector<ArgOptions>& o
invalid_arg = true;
return;
}
if(option.concat && !option.target->empty()){
if(option.concat > 0 && option.concat <= 0xff){
if (option.concat && !option.target->empty()) {
if (option.concat > 0 && option.concat <= 0xff) {
*option.target += static_cast<char>(option.concat);
}
*option.target += argv_to_utf8(i, argv);
} else {
*option.target = argv_to_utf8(i, argv);
}
found_arg = true;
found_arg = true;
}))
break;

Expand Down
2 changes: 1 addition & 1 deletion src/model_manager.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -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;
}
Expand Down
2 changes: 2 additions & 0 deletions src/model_manager.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -69,6 +69,7 @@ class ModelManager : public RunnerWeightManager {
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();
Expand Down Expand Up @@ -110,6 +111,7 @@ class ModelManager : public RunnerWeightManager {
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);

Expand Down
10 changes: 5 additions & 5 deletions src/runtime/guidance.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3,9 +3,9 @@
#include <algorithm>
#include <cmath>
#include <cstdlib>
#include <optional>
#include <string>
#include <utility>
#include <optional>

#include "core/util.h"

Expand Down Expand Up @@ -127,7 +127,7 @@ namespace sd::guidance {
std::vector<float> parse_guidance_schedule(const char* extra_sample_args) {
std::vector<float> guidance_schedule;
std::string guidance_schedule_str = "";
for (const auto& [key, value] : parse_key_value_args(extra_sample_args, "extra sample arg")) {
for (const auto& [key, value] : parse_key_value_args(extra_sample_args, "extra sample arg")) {
float parsed = 0.0f;
if (key == "guidance_schedule") {
guidance_schedule_str = value;
Expand All @@ -148,9 +148,9 @@ namespace sd::guidance {

GuiderOutput ClassifierFreeGuidance::forward(const GuidanceInput& input,
GuiderOutput previous,
std::optional<float> scale_override) const {
std::optional<float> scale_override) const {
(void)previous;
float guidance_scale = scale_override.value_or(guidance_scale_);
float guidance_scale = scale_override.value_or(guidance_scale_);

GuiderOutput output;
if (!has_tensor(input.pred_cond)) {
Expand Down Expand Up @@ -210,7 +210,7 @@ namespace sd::guidance {
GuiderOutput previous,
std::optional<float> scale_override) const {
(void)previous;
float guidance_scale = scale_override.value_or(guidance_scale_);
float guidance_scale = scale_override.value_or(guidance_scale_);

GuiderOutput output;
if (!has_tensor(input.pred_cond)) {
Expand Down
4 changes: 2 additions & 2 deletions src/runtime/guidance.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3,8 +3,8 @@

#include <cstddef>
#include <functional>
#include <vector>
#include <optional>
#include <vector>

#include "core/tensor.hpp"

Expand Down Expand Up @@ -42,7 +42,7 @@ namespace sd::guidance {

class BaseGuidance {
public:
virtual ~BaseGuidance() = default;
virtual ~BaseGuidance() = default;
virtual GuiderOutput forward(const GuidanceInput& input,
GuiderOutput previous,
std::optional<float> scale_override = std::nullopt) const = 0;
Expand Down
22 changes: 11 additions & 11 deletions src/stable-diffusion.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -532,7 +532,6 @@ class StableDiffusionGGML {
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();
Expand Down Expand Up @@ -586,9 +585,12 @@ class StableDiffusionGGML {
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;
Expand Down Expand Up @@ -1941,26 +1943,26 @@ class StableDiffusionGGML {
float img_cfg_scale = guidance.img_cfg;
float slg_scale = guidance.slg.scale;
bool slg_uncond = sd::guidance::parse_skip_layer_guidance_uncond_arg(extra_sample_args);

std::vector<float> guidance_schedule = sd::guidance::parse_guidance_schedule(extra_sample_args);
if(!guidance_schedule.empty() && guidance_schedule.size() != sigmas.size() - 1) {
if(guidance_schedule.size() > sigmas.size()) {
if (!guidance_schedule.empty() && guidance_schedule.size() != sigmas.size() - 1) {
if (guidance_schedule.size() > sigmas.size()) {
LOG_WARN("guidance_schedule length (%zu) is greater than number of steps (%zu)", guidance_schedule.size(), sigmas.size() - 1);
LOG_WARN("truncating guidance_schedule to match step count");
guidance_schedule.resize(sigmas.size() - 1);
} else {
LOG_INFO("padding guidance_schedule with cfg_scale");
while(guidance_schedule.size() < sigmas.size() - 1) {
while (guidance_schedule.size() < sigmas.size() - 1) {
guidance_schedule.push_back(cfg_scale);
}
}
}

if(!guidance_schedule.empty()) {
if (!guidance_schedule.empty()) {
std::string schedule_str = "[";
for(size_t i = 0; i < guidance_schedule.size(); ++i) {
for (size_t i = 0; i < guidance_schedule.size(); ++i) {
schedule_str += std::to_string(guidance_schedule[i]);
if(i < guidance_schedule.size() - 1) {
if (i < guidance_schedule.size() - 1) {
schedule_str += ", ";
}
}
Expand Down Expand Up @@ -2208,9 +2210,7 @@ class StableDiffusionGGML {
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 {};
}
Expand Down
Loading