Skip to content

Configurable VAE threshold limit #1601

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
merged 5 commits into from
Jun 21, 2025
Merged

Configurable VAE threshold limit #1601

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
5919014
add backend support for changing the VAE tiling threshold
wbruna Jun 11, 2025
496e5f2
trigger VAE tiling by image area instead of dimensions
wbruna Jun 11, 2025
18b24ae
replace the notile option with a configurable threshold
wbruna Jun 20, 2025
3adb93d
Merge branch 'concedo_experimental' into koboldcpp_vae_threshold
LostRuins Jun 21, 2025
830ae98
simplify tiling checks, 768 default visible in launcher
LostRuins Jun 21, 2025
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
2 changes: 1 addition & 1 deletion expose.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -162,7 +162,7 @@ struct sd_load_model_inputs
const int threads = 0;
const int quant = 0;
const bool taesd = false;
const bool notile = false;
const int tiled_vae_threshold = 0;
const char * t5xxl_filename = nullptr;
const char * clipl_filename = nullptr;
const char * clipg_filename = nullptr;
Expand Down
20 changes: 12 additions & 8 deletions koboldcpp.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -52,6 +52,7 @@
default_visionmaxres = 1024
net_save_slots = 10
savestate_limit = 3 #3 savestate slots
default_vae_tile_threshold = 768

# abuse prevention
stop_token_max = 256
Expand Down Expand Up @@ -272,7 +273,7 @@ class sd_load_model_inputs(ctypes.Structure):
("threads", ctypes.c_int),
("quant", ctypes.c_int),
("taesd", ctypes.c_bool),
("notile", ctypes.c_bool),
("tiled_vae_threshold", ctypes.c_int),
("t5xxl_filename", ctypes.c_char_p),
("clipl_filename", ctypes.c_char_p),
("clipg_filename", ctypes.c_char_p),
Expand Down Expand Up @@ -1549,7 +1550,7 @@ def sd_load_model(model_filename,vae_filename,lora_filename,t5xxl_filename,clipl
inputs.threads = thds
inputs.quant = quant
inputs.taesd = True if args.sdvaeauto else False
inputs.notile = True if args.sdnotile else False
inputs.tiled_vae_threshold = args.sdtiledvae
inputs.vae_filename = vae_filename.encode("UTF-8")
inputs.lora_filename = lora_filename.encode("UTF-8")
inputs.lora_multiplier = args.sdloramult
Expand Down Expand Up @@ -4303,7 +4304,7 @@ def hide_tooltip(event):
sd_clipg_var = ctk.StringVar()
sd_photomaker_var = ctk.StringVar()
sd_vaeauto_var = ctk.IntVar(value=0)
sd_notile_var = ctk.IntVar(value=0)
sd_tiled_vae_var = ctk.StringVar(value=str(default_vae_tile_threshold))
sd_clamped_var = ctk.StringVar(value="0")
sd_clamped_soft_var = ctk.StringVar(value="0")
sd_threads_var = ctk.StringVar(value=str(default_threads))
Expand Down Expand Up @@ -5033,7 +5034,7 @@ def toggletaesd(a,b,c):
sdvaeitem2.grid()
sdvaeitem3.grid()
makecheckbox(images_tab, "Use TAE SD (AutoFix Broken VAE)", sd_vaeauto_var, 42,command=toggletaesd,tooltiptxt="Replace VAE with TAESD. May fix bad VAE.")
makecheckbox(images_tab, "No VAE Tiling", sd_notile_var, 44,tooltiptxt="Disables VAE tiling, may not work for large images.")
makelabelentry(images_tab, "VAE Tiling Threshold:", sd_tiled_vae_var, 44, 50, padx=144,singleline=True,tooltip="Enable VAE Tiling for images above this size, to save memory.\nSet to 0 to disable VAE tiling.")

# audio tab
audio_tab = tabcontent["Audio"]
Expand Down Expand Up @@ -5266,7 +5267,7 @@ def export_vars():
args.sdthreads = (0 if sd_threads_var.get()=="" else int(sd_threads_var.get()))
args.sdclamped = (0 if int(sd_clamped_var.get())<=0 else int(sd_clamped_var.get()))
args.sdclampedsoft = (0 if int(sd_clamped_soft_var.get())<=0 else int(sd_clamped_soft_var.get()))
args.sdnotile = (True if sd_notile_var.get()==1 else False)
args.sdtiledvae = (default_vae_tile_threshold if sd_tiled_vae_var.get()=="" else int(sd_tiled_vae_var.get()))
if sd_vaeauto_var.get()==1:
args.sdvaeauto = True
args.sdvae = ""
Expand Down Expand Up @@ -5488,7 +5489,8 @@ def import_vars(dict):
sd_clipg_var.set(dict["sdclipg"] if ("sdclipg" in dict and dict["sdclipg"]) else "")
sd_photomaker_var.set(dict["sdphotomaker"] if ("sdphotomaker" in dict and dict["sdphotomaker"]) else "")
sd_vaeauto_var.set(1 if ("sdvaeauto" in dict and dict["sdvaeauto"]) else 0)
sd_notile_var.set(1 if ("sdnotile" in dict and dict["sdnotile"]) else 0)
sd_tiled_vae_var.set(str(dict["sdtiledvae"]) if ("sdtiledvae" in dict and dict["sdtiledvae"]) else str(default_vae_tile_threshold))

sd_lora_var.set(dict["sdlora"] if ("sdlora" in dict and dict["sdlora"]) else "")
sd_loramult_var.set(str(dict["sdloramult"]) if ("sdloramult" in dict and dict["sdloramult"]) else "1.0")

Expand Down Expand Up @@ -5856,6 +5858,8 @@ def convert_invalid_args(args):
dict["model_param"] = model_value
elif isinstance(model_value, list) and model_value: # Non-empty list
dict["model_param"] = model_value[0] # Take the first file in the list
if "sdnotile" in dict and "sdtiledvae" not in dict:
dict["sdtiledvae"] = (0 if (dict["sdnotile"]) else default_vae_tile_threshold) # convert legacy option
return args

def setuptunnel(global_memory, has_sd):
Expand Down Expand Up @@ -7269,8 +7273,7 @@ def range_checker(arg: str):
sdparsergrouplora.add_argument("--sdquant", help="If specified, loads the model quantized to save memory.", action='store_true')
sdparsergrouplora.add_argument("--sdlora", metavar=('[filename]'), help="Specify an image generation LORA safetensors model to be applied.", default="")
sdparsergroup.add_argument("--sdloramult", metavar=('[amount]'), help="Multiplier for the image LORA model to be applied.", type=float, default=1.0)
sdparsergroup.add_argument("--sdnotile", help="Disables VAE tiling, may not work for large images.", action='store_true')

sdparsergroup.add_argument("--sdtiledvae", metavar=('[maxres]'), help="Adjust the automatic VAE tiling trigger for images above this size. 0 disables vae tiling.", type=int, default=default_vae_tile_threshold)
whisperparsergroup = parser.add_argument_group('Whisper Transcription Commands')
whisperparsergroup.add_argument("--whispermodel", metavar=('[filename]'), help="Specify a Whisper .bin model to enable Speech-To-Text transcription.", default="")

Expand All @@ -7296,5 +7299,6 @@ def range_checker(arg: str):
deprecatedgroup.add_argument("--sdconfig", help=argparse.SUPPRESS, nargs='+')
compatgroup.add_argument("--noblas", help=argparse.SUPPRESS, action='store_true')
compatgroup3.add_argument("--nommap", help=argparse.SUPPRESS, action='store_true')
deprecatedgroup.add_argument("--sdnotile", help=argparse.SUPPRESS, action='store_true') # legacy option, see sdtiledvae

main(launch_args=parser.parse_args(),default_args=parser.parse_args([]))
9 changes: 6 additions & 3 deletions otherarch/sdcpp/sdtype_adapter.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -119,7 +119,7 @@ static uint8_t * input_mask_buffer = NULL;
static uint8_t * input_photomaker_buffer = NULL;

static std::string sdplatformenv, sddeviceenv, sdvulkandeviceenv;
static bool notiling = false;
static int cfg_tiled_vae_threshold = 0;
static int cfg_square_limit = 0;
static int cfg_side_limit = 0;
static bool sd_is_quiet = false;
Expand All @@ -136,7 +136,9 @@ bool sdtype_load_model(const sd_load_model_inputs inputs) {
std::string clipl_filename = inputs.clipl_filename;
std::string clipg_filename = inputs.clipg_filename;
std::string photomaker_filename = inputs.photomaker_filename;
notiling = inputs.notile;
cfg_tiled_vae_threshold = inputs.tiled_vae_threshold;
cfg_tiled_vae_threshold = (cfg_tiled_vae_threshold > 8192 ? 8192 : cfg_tiled_vae_threshold);
cfg_tiled_vae_threshold = (cfg_tiled_vae_threshold <= 0 ? 8192 : cfg_tiled_vae_threshold); //if negative dont tile
cfg_side_limit = inputs.img_hard_limit;
cfg_square_limit = inputs.img_soft_limit;
printf("\nImageGen Init - Load Model: %s\n",inputs.model_filename);
Expand Down Expand Up @@ -482,7 +484,8 @@ sd_generation_outputs sdtype_generate(const sd_generation_inputs inputs)
printf("\nKCPP SD: Requested dimensions %dx%d changed to %dx%d\n", inputs.width, inputs.height, sd_params->width, sd_params->height);
}

bool dotile = (sd_params->width>768 || sd_params->height>768) && !notiling;
// trigger tiling by image area, the memory used for the VAE buffer is 6656 bytes per image pixel, default 768x768
bool dotile = (sd_params->width*sd_params->height > cfg_tiled_vae_threshold*cfg_tiled_vae_threshold);
set_sd_vae_tiling(sd_ctx,dotile); //changes vae tiling, prevents memory related crash/oom

if (sd_params->clip_skip <= 0) {
Expand Down