Add config for Conv2D Direct for the VAE

Author: wbruna

expose.h

@@ -165,6 +165,7 @@ struct sd_load_model_inputs
     const int threads = 0;
     const int quant = 0;
     const bool flash_attention = false;
+    const bool vae_conv_direct = false;
     const bool taesd = false;
     const int tiled_vae_threshold = 0;
     const char * t5xxl_filename = nullptr;

koboldcpp.py

@@ -278,6 +278,7 @@ class sd_load_model_inputs(ctypes.Structure):
                 ("threads", ctypes.c_int),
                 ("quant", ctypes.c_int),
                 ("flash_attention", ctypes.c_bool),
+                ("vae_conv_direct", ctypes.c_bool),
                 ("taesd", ctypes.c_bool),
                 ("tiled_vae_threshold", ctypes.c_int),
                 ("t5xxl_filename", ctypes.c_char_p),
@@ -1646,6 +1647,7 @@ def sd_load_model(model_filename,vae_filename,lora_filename,t5xxl_filename,clipl
     inputs.threads = thds
     inputs.quant = quant
     inputs.flash_attention = args.sdflashattention
+    inputs.vae_conv_direct = args.sdvaeconvdir
     inputs.taesd = True if args.sdvaeauto else False
     inputs.tiled_vae_threshold = args.sdtiledvae
     inputs.vae_filename = vae_filename.encode("UTF-8")
@@ -4559,6 +4561,7 @@ def hide_tooltip(event):
     sd_flash_attention_var = ctk.IntVar(value=0)
     sd_vaeauto_var = ctk.IntVar(value=0)
     sd_tiled_vae_var = ctk.StringVar(value=str(default_vae_tile_threshold))
+    sd_vae_convdir_var = ctk.IntVar(value=0)
     sd_clamped_var = ctk.StringVar(value="0")
     sd_clamped_soft_var = ctk.StringVar(value="0")
     sd_threads_var = ctk.StringVar(value=str(default_threads))
@@ -5302,7 +5305,8 @@ def toggletaesd(a,b,c):
                 sdvaeitem1.grid()
                 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, "TAE SD (AutoFix Broken VAE)", sd_vaeauto_var, 42,command=toggletaesd,tooltiptxt="Replace VAE with TAESD. May fix bad VAE.")
+    makecheckbox(images_tab, "Conv2D Direct for VAE", sd_vae_convdir_var, 42, padx=220, tooltiptxt="Enable Conv2D Direct for VAE. Saves memory and improves performance.\nMight crash if not supported by the backend.")
     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.")
     makecheckbox(images_tab, "Flash Attention", sd_flash_attention_var, 48, tooltiptxt="Enable Flash Attention for diffusion. May save memory or improve performance.")
 
@@ -5549,6 +5553,8 @@ def export_vars():
             args.sdvae = ""
             if sd_vae_var.get() != "":
                 args.sdvae = sd_vae_var.get()
+        if sd_vae_convdir_var.get()==1:
+            args.sdvaeconvdir = True
         if sd_t5xxl_var.get() != "":
             args.sdt5xxl = sd_t5xxl_var.get()
         if sd_clipl_var.get() != "":
@@ -5766,6 +5772,7 @@ def import_vars(dict):
         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_tiled_vae_var.set(str(dict["sdtiledvae"]) if ("sdtiledvae" in dict and dict["sdtiledvae"]) else str(default_vae_tile_threshold))
+        sd_vae_convdir_var.set(1 if ("sdvaeconvdir" in dict and dict["sdvaeconvdir"]) else 0)
 
         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")
@@ -7564,6 +7571,7 @@ def range_checker(arg: str):
     sdparsergroupvae = sdparsergroup.add_mutually_exclusive_group()
     sdparsergroupvae.add_argument("--sdvae", metavar=('[filename]'), help="Specify an image generation safetensors VAE which replaces the one in the model.", default="")
     sdparsergroupvae.add_argument("--sdvaeauto", help="Uses a built-in VAE via TAE SD, which is very fast, and fixed bad VAEs.", action='store_true')
+    sdparsergroupvae.add_argument("--sdvaeconvdir", help="Enables Conv2D Direct for the image diffusion model. Should improve performance and reduce memory usage. Might crash if not supported by the backend.", action='store_true')
     sdparsergrouplora = sdparsergroup.add_mutually_exclusive_group()
     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="")

otherarch/sdcpp/sdtype_adapter.cpp

@@ -99,6 +99,7 @@ struct SDParams {
     bool clip_on_cpu              = false;
     bool vae_on_cpu               = false;
     bool diffusion_flash_attn     = false;
+    bool vae_conv_direct          = false;
     bool canny_preprocess         = false;
     bool color                    = false;
     int upscale_repeats           = 1;
@@ -211,6 +212,10 @@ bool sdtype_load_model(const sd_load_model_inputs inputs) {
     {
         printf("Flash Attention is enabled\n");
     }
+    if(inputs.vae_conv_direct)
+    {
+        printf("Conv2D Direct for VAE model is enabled\n");
+    }
     if(inputs.quant)
     {
         printf("Note: Loading a pre-quantized model is always faster than using compress weights!\n");
@@ -246,6 +251,7 @@ bool sdtype_load_model(const sd_load_model_inputs inputs) {
     sd_params->wtype = (inputs.quant==0?SD_TYPE_COUNT:SD_TYPE_Q4_0);
     sd_params->n_threads = inputs.threads; //if -1 use physical cores
     sd_params->diffusion_flash_attn = inputs.flash_attention;
+    sd_params->vae_conv_direct = inputs.vae_conv_direct;
     sd_params->input_path = ""; //unused
     sd_params->batch_count = 1;
     sd_params->vae_path = vaefilename;
@@ -312,6 +318,7 @@ bool sdtype_load_model(const sd_load_model_inputs inputs) {
     params.keep_control_net_on_cpu = sd_params->control_net_cpu;
     params.keep_vae_on_cpu = sd_params->vae_on_cpu;
     params.diffusion_flash_attn = sd_params->diffusion_flash_attn;
+    params.vae_conv_direct = sd_params->vae_conv_direct;
     params.chroma_use_dit_mask = sd_params->chroma_use_dit_mask;
     params.chroma_use_t5_mask = sd_params->chroma_use_t5_mask;
     params.chroma_t5_mask_pad = sd_params->chroma_t5_mask_pad;