upstream-archive/stable-diffusion.cpp
1
0
Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2026-10-03 11:27:43 -05:00
Code Issues Packages Projects Releases Wiki Activity

feat: add max-vram based segmented param offload (#1476)

Browse Source
This commit is contained in:
leejet
2026-05-06 21:56:02 +08:00
committed by GitHub
parent 586b6f1481
commit 90e87bc846
26 changed files with 1576 additions and 66 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/cli/README.md
+2
View File
@@ -54,6 +54,8 @@ Context Options:
-t, --threads <int> number of threads to use during computation (default: -1). If threads <= 0,
then threads will be set to the number of CPU physical cores
--chroma-t5-mask-pad <int> t5 mask pad size of chroma
--max-vram <float> maximum VRAM budget in GiB for graph-cut segmented execution. 0 disables
graph splitting
--force-sdxl-vae-conv-scale force use of conv scale on sdxl vae
--offload-to-cpu place the weights in RAM to save VRAM, and automatically load them into VRAM
when needed
examples/common/common.cpp
+8 -1
View File
@@ -394,7 +394,12 @@ ArgOptions SDContextParams::get_options() {
&chroma_t5_mask_pad},
};
options.float_options = {};
options.float_options = {
{"",
"--max-vram",
"maximum VRAM budget in GiB for graph-cut segmented execution. 0 disables graph splitting",
&max_vram},
};
options.bool_options = {
{"",
@@ -670,6 +675,7 @@ std::string SDContextParams::to_string() const {
<< " rng_type: " << sd_rng_type_name(rng_type) << ",\n"
<< " sampler_rng_type: " << sd_rng_type_name(sampler_rng_type) << ",\n"
<< " offload_params_to_cpu: " << (offload_params_to_cpu ? "true" : "false") << ",\n"
<< " max_vram: " << max_vram << ",\n"
<< " enable_mmap: " << (enable_mmap ? "true" : "false") << ",\n"
<< " control_net_cpu: " << (control_net_cpu ? "true" : "false") << ",\n"
<< " clip_on_cpu: " << (clip_on_cpu ? "true" : "false") << ",\n"
@@ -744,6 +750,7 @@ sd_ctx_params_t SDContextParams::to_sd_ctx_params_t(bool vae_decode_only, bool f
chroma_use_t5_mask,
chroma_t5_mask_pad,
qwen_image_zero_cond_t,
max_vram,
};
return sd_ctx_params;
}
examples/common/common.h
+1
View File
@@ -109,6 +109,7 @@ struct SDContextParams {
rng_type_t rng_type = CUDA_RNG;
rng_type_t sampler_rng_type = RNG_TYPE_COUNT;
bool offload_params_to_cpu = false;
float max_vram = 0.f;
bool enable_mmap = false;
bool control_net_cpu = false;
bool clip_on_cpu = false;
examples/server/README.md
+2
View File
@@ -156,6 +156,8 @@ Context Options:
-t, --threads <int> number of threads to use during computation (default: -1). If threads <= 0,
then threads will be set to the number of CPU physical cores
--chroma-t5-mask-pad <int> t5 mask pad size of chroma
--max-vram <float> maximum VRAM budget in GiB for graph-cut segmented execution. 0 disables
graph splitting
--force-sdxl-vae-conv-scale force use of conv scale on sdxl vae
--offload-to-cpu place the weights in RAM to save VRAM, and automatically load them into VRAM
when needed