vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-05-22 11:16:08 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

sd: add backend support for max_vram (#2221)

Browse source
This commit is contained in:
Wagner Bruna 2026-05-21 00:51:00 -03:00 • committed by GitHub
parent 095bf63b58
commit f85a747dc0
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
3 changed files with 6 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

1
expose.h
View file

@ -207,6 +207,7 @@ struct sd_load_model_inputs
const char * upscaler_filename = nullptr;
const int img_hard_limit = 0;
const int img_soft_limit = 0;
float max_vram = 0.f;
const char * devices_override = nullptr;
const bool quiet = false;
const int debugmode = 0;

1
koboldcpp.py
View file

@ -389,6 +389,7 @@ class sd_load_model_inputs(ctypes.Structure):
("upscaler_filename", ctypes.c_char_p),
("img_hard_limit", ctypes.c_int),
("img_soft_limit", ctypes.c_int),
("max_vram", ctypes.c_float),
("devices_override", ctypes.c_char_p),
("quiet", ctypes.c_bool),
("debugmode", ctypes.c_int)]

4
otherarch/sdcpp/sdtype_adapter.cpp
View file

@ -397,6 +397,9 @@ bool sdtype_load_model(const sd_load_model_inputs inputs) {
} else if (inputs.use_mmap) {
printf("Using mmap for I/O\n");
}
if(inputs.max_vram != 0.f) {
printf("Using max VRAM = %0.2f\n", inputs.max_vram);
}
if(inputs.quant > 0)
{
printf("Note: Loading a pre-quantized model is always faster than using compress weights!\n");
@ -460,6 +463,7 @@ bool sdtype_load_model(const sd_load_model_inputs inputs) {
params.diffusion_conv_direct = sd_params->diffusion_conv_direct;
params.vae_conv_direct = sd_params->vae_conv_direct;
params.chroma_use_dit_mask = sd_params->chroma_use_dit_mask;
params.max_vram = inputs.max_vram;
params.enable_mmap = inputs.use_mmap;
// the _cpu flags are only used if the backend string is empty, but
// we always set both for consistency