Skip to content
Python

added setting sdpa override, blocked sage with rtx30xx+zimage base

DeepBeepMeep31438d5fe923
diff --git a/README.md b/README.md
index be22842..4395ca5 100644
--- a/README.md
+++ b/README.md
@@ -34,7 +34,7 @@ WanGP supports the Wan (and derived models), Hunyuan Video, Flux, Qwen, Z-Image,
## 🔥 Latest Updates :
-### January 27th 2026: WanGP v10.54, Music for your Hearts
+### January 27th 2026: WanGP v10.55, Music for your Hearts
WanGP Special *TTS* (Text To Speech) Release:
@@ -48,7 +48,9 @@ WanGP Special *TTS* (Text To Speech) Release:
- **Custom Output folder for Audio Gens**: you can now choose a different folder for the *Audio Outputs*
- **Default Memory Profile for Audio Models**: TTS models can get very slow if you use profile 4 (being autoregressive models, they will need to load all the layers one per one to generate one single audio token then rinse & repeat). On the other hand, they dont't need as much VRAM, so you can now define a more agressive profile (3+ for instance)
-- **Z Image Base**: try it if you are into the *Z Image* hype but it will be probably useless for you unless you are a researcher and / or want to build a finetune out of it. This model requires from 35 to 50 steps (4x to 6x slower than *Z Image turbo*) and cfg > 1 (an additional 2x slower) and there is no *Reinforcement Learning* so Output Images wont be as good. The plus side is a higher diversity and *Native Negative Prompt* (versus Z Image virtual Negative Prompt using *NAG*)
+- **Z Image Base**: try it if you are into the *Z Image* hype but it will be probably useless for you unless you are a researcher and / or want to build a finetune out of it. This model requires from 35 to 50 steps (4x to 6x slower than *Z Image turbo*) and cfg > 1 (an additional 2x slower) and there is no *Reinforcement Learning* so Output Images wont be as good. The plus side is a higher diversity and *Native Negative Prompt* (versus Z Image virtual Negative Prompt using *NAG*).
+
+Note that Z Image Base is very sensitive to the *Attention Mode*: it is not compatible with *Sage 1* as it produces black frames. So I have disabled Sage for RTX 30xx. Also there are reports it produces some vertical banding artifacts with *Sage 2*
- **Various Improvements**:
- Video /Audio Galleries now support deletions of gens done outside WanGP
@@ -56,11 +58,14 @@ WanGP Special *TTS* (Text To Speech) Release:
- *Check for Updates* button for *Plugins* to see in a glance if any of your plugin can be updated
- *Prompt Enhancer* generates a different enhanced prompt each timee you click on it. You can define in the config tab its gen parameters (top k, temperature)
- New *Root Loras* folder can be defined in the config Tab. Useful if you have multiple WanGP instances or want to store easily all your loras in a different hard drive
+ - added new setting *Attention Mode Override" in the *Misc* tab
+ - Experimental: allowed changing *Configuration* during a *Generation*
*update 10.51*: new Heart Mula Finetune better at following instructions, Extra settings (cfg, top k) for TTS models, Rife v4\
*update 10.52*: updated plugin list and added version tracking\
*update 10.53*: video/audio galleries now support deletions\
-*update 10.54*: added Z Image Omnibase, prompt enhancers improvements, configurable loras root folder
+*update 10.54*: added Z Image Base, prompt enhancers improvements, configurable loras root folder\
+*update 10.55*: blocked Sage with Z Image on RTX30xx and added override attention mode settings, allowed changing config during generation
### January 20th 2026: WanGP v10.43, The Cost Saver
*GPUs are expensive, RAM is expensive, SSD are expensive, sadly we live now in a GPU & RAM poor.*
diff --git a/defaults/z_image_base.json b/defaults/z_image_base.json
index 7c2e00f..69375ef 100644
--- a/defaults/z_image_base.json
+++ b/defaults/z_image_base.json
@@ -6,7 +6,10 @@
"URLs": [
"https://huggingface.co/DeepBeepMeep/Z-Image/resolve/main/ZImageBase_bf16.safetensors",
"https://huggingface.co/DeepBeepMeep/Z-Image/resolve/main/ZImageBase_quanto_bf16_int8.safetensors"
- ]
+ ],
+ "attention": {
+ "<89": "sdpa"
+ }
},
"resolution": "1024x1024",
"batch_size": 1,
diff --git a/models/_settings.json b/models/_settings.json
index 7798da8..ceb2a3f 100644
--- a/models/_settings.json
+++ b/models/_settings.json
@@ -81,6 +81,7 @@
"RIFLEx_setting": 0,
"force_fps": "",
"override_profile": -1,
+ "override_attention": "",
"output_filename": "",
"activated_loras": [],
"loras_multipliers": "",
diff --git a/plugins/wan2gp-configuration/plugin.py b/plugins/wan2gp-configuration/plugin.py
index 1421bba..c493a13 100644
--- a/plugins/wan2gp-configuration/plugin.py
+++ b/plugins/wan2gp-configuration/plugin.py
@@ -29,6 +29,7 @@ class ConfigTabPlugin(WAN2GPPlugin):
self.request_global("attention_modes_supported")
self.request_global("displayed_model_types")
self.request_global("memory_profile_choices")
+ self.request_global("attention_modes_choices")
self.request_global("save_path")
self.request_global("image_save_path")
self.request_global("audio_save_path")
@@ -87,22 +88,8 @@ class ConfigTabPlugin(WAN2GPPlugin):
interactive=not self.args.lock_config
)
- def check_attn(mode):
- if mode not in self.attention_modes_installed: return " (NOT INSTALLED)"
- if mode not in self.attention_modes_supported: return " (NOT SUPPORTED)"
- return ""
-
self.attention_choice = gr.Dropdown(
- choices=[
- ("Auto: Best available (sage2 > sage > sdpa)", "auto"),
- ("sdpa: Default, always available", "sdpa"),
- (f'flash{check_attn("flash")}: High quality, requires manual install', "flash"),
- (f'xformers{check_attn("xformers")}: Good quality, less VRAM, requires manual install', "xformers"),
- (f'sage{check_attn("sage")}: ~30% faster, requires manual install', "sage"),
- (f'sage2/sage2++{check_attn("sage2")}: ~40% faster, requires manual install', "sage2"),
- ] + ([(f'radial{check_attn("radial")}: Experimental, may be faster, requires manual install', "radial")] if self.args.betatest else []) + [
- (f'sage3{check_attn("sage3")}: >50% faster, may have quality trade-offs, requires manual install', "sage3"),
- ],
+ choices=self.attention_modes_choices,
value=self.attention_mode, label="Attention Type", interactive=not self.args.lock_config
)
self.preload_model_policy_choice = gr.CheckboxGroup(
@@ -323,8 +310,8 @@ class ConfigTabPlugin(WAN2GPPlugin):
return [self.release_RAM_btn]
def _save_changes(self, state, *args):
- if self.is_generation_in_progress():
- return "<div style='color:red; text-align:center;'>Unable to change config when a generation is in progress.</div>", *[gr.update()]*5
+ gen_in_progress = self.is_generation_in_progress()
+ # return "<div style='color:red; text-align:center;'>Unable to change config when a generation is in progress.</div>", *[gr.update()]*5
if self.args.lock_config:
return "<div style='color:red; text-align:center;'>Configuration is locked by command-line arguments.</div>", *[gr.update()]*5
@@ -459,9 +446,14 @@ class ConfigTabPlugin(WAN2GPPlugin):
model_family_update, model_base_type_update, model_choice_update = self.generate_dropdown_model_list(model_type)
header_update = self.generate_header(model_type, compile=new_server_config["compile"], attention_mode=new_server_config["attention_mode"])
-
- return (
- "<div style='color:green; text-align:center;'>The new configuration has been succesfully applied.</div>",
+
+ if gen_in_progress:
+ msg = "<div style='color:green; text-align:center;'>The new configuration has been succesfully applied. Some of the Settings will be only effective when you will start another Generation</div>"
+ else:
+ msg = "<div style='color:green; text-align:center;'>The new configuration has been succesfully applied.</div>"
+
+ return (msg
+ ,
header_update,
model_family_update,
model_base_type_update,
diff --git a/wgp.py b/wgp.py
index 7a2d91f..568bf1f 100644
--- a/wgp.py
+++ b/wgp.py
@@ -85,7 +85,7 @@ AUTOSAVE_TEMPLATE_PATH = AUTOSAVE_FILENAME
CONFIG_FILENAME = "wgp_config.json"
PROMPT_VARS_MAX = 10
target_mmgp_version = "3.7.2"
-WanGP_version = "10.54"
+WanGP_version = "10.55"
settings_version = 2.44
max_source_video_frames = 3000
prompt_enhancer_image_caption_model, prompt_enhancer_image_caption_processor, prompt_enhancer_llm_model, prompt_enhancer_llm_tokenizer = None, None, None, None
@@ -5481,6 +5481,7 @@ def generate_video(
prompt_enhancer,
min_frames_if_references,
override_profile,
+ override_attention,
pace,
exaggeration,
temperature,
@@ -5573,7 +5574,7 @@ def generate_video(
if args.test:
send_cmd("info", "Test mode: model loaded, skipping generation.")
return
- overridden_attention = get_overridden_attention(model_type)
+ overridden_attention = override_attention if len(override_attention) else get_overridden_attention(model_type)
# if overridden_attention is not None and overridden_attention != attention_mode: print(f"Attention mode has been overriden to {overridden_attention} for model type '{model_type}'")
attn = overridden_attention if overridden_attention is not None else attention_mode
if attn == "auto":
@@ -8357,6 +8358,7 @@ def save_inputs(
prompt_enhancer,
min_frames_if_references,
override_profile,
+ override_attention,
pace,
exaggeration,
temperature,
@@ -9024,6 +9026,22 @@ memory_profile_choices= [ ("Profile 1, HighRAM_HighVRAM: at least 64 GB of RAM
("Profile 4+, LowRAM_LowVRAM+: at least 32 GB of RAM and 12 GB of VRAM, variant of Profile 4, slightly slower but needs less VRAM",4.5),
("Profile 5, VerylowRAM_LowVRAM (Fail safe): at least 24 GB of RAM and 10 GB of VRAM, if you don't have much it won't be fast but maybe it will work",5)]
+def check_attn(mode):
+ if mode not in attention_modes_installed: return " (NOT INSTALLED)"
+ if mode not in attention_modes_supported: return " (NOT SUPPORTED)"
+ return ""
+
+attention_modes_choices= [
+ ("Auto: Best available (sage2 > sage > sdpa)", "auto"),
+ ("sdpa: Default, always available", "sdpa"),
+ (f'flash{check_attn("flash")}: High quality, requires manual install', "flash"),
+ (f'xformers{check_attn("xformers")}: Good quality, less VRAM, requires manual install', "xformers"),
+ (f'sage{check_attn("sage")}: ~30% faster, requires manual install', "sage"),
+ (f'sage2/sage2++{check_attn("sage2")}: ~40% faster, requires manual install', "sage2"),
+] + ([(f'radial{check_attn("radial")}: Experimental, may be faster, requires manual install', "radial")] if args.betatest else []) + [
+ (f'sage3{check_attn("sage3")}: >50% faster, may have quality trade-offs, requires manual install', "sage3"),
+]
+
def detect_auto_save_form(state, evt:gr.SelectData):
last_tab_id = state.get("last_tab_id", 0)
state["last_tab_id"] = new_tab_id = evt.index
@@ -10147,6 +10165,13 @@ def generate_video_tab(update_form = False, state_dict = None, ui_defaults = Non
label=f"Override Memory Profile"
)
+ gr.Markdown("<B>You can set a different Attention Mode to improve the quality / compatibility<B>")
+ override_attention = gr.Dropdown(
+ choices=[("Default Attention Mode", "")] + attention_modes_choices,
+ value=ui_get("override_attention"),
+ label=f"Override Attention Mode"
+ )
+
with gr.Column():
gr.Markdown('<B>Customize the Output Filename using Settings Values (<I>date, seed, resolution, num_inference_steps, prompt, flow_shift, video_length, guidance_scale</I>). For Instance:<BR>"<I>{date(YYYY-MM-DD_HH-mm-ss)}_{seed}_{prompt(50)}, {num_inference_steps}</I>"</B>')
output_filename = gr.Text( label= " Output Filename ( Leave Blank for Auto Naming)", value= ui_get("output_filename"))