Signed-off-by: vladmandic <mandic00@live.com>
This commit is contained in:
vladmandic
2026-04-13 12:18:05 +02:00
parent 522352bac2
commit 84c2ddc5ca
5 changed files with 17 additions and 14 deletions
+13 -8
View File
@@ -21,6 +21,17 @@
integrates Qwen2.5-VL text encoder and Flow Matching scheduler for high-quality in-context edits
enables multimodal understanding and conditional image refinement
category: *base*
- **Features**
- **Tag Autocomplete** native implementation!
select and activate any number of different auto-complete databases
see *Extras -> Autocomplete* for all settings
- **NudeNet** add `LlavaGuard` and `QwenGuard` as image safety evaulation models
- **schedulers** add `ER-SDE` solver
with with *1S, 2M, 3M* variants in both *EPS* and *FlowMatch* flavors
- **installer** auto-restart on upgrade
- enhanced **filename** pattern processing
allows for any *processing* property name (as defined in `modules/processing_class.py` and saved to `ui-config.json`)
allows for any *settings* property name (as defined in `modules/ui_definitions.py` and saved to `config.json`)
- **Compute**
- **ROCm** futher work on advanced configuration and tuning, thanks @resonantsky
now covers both ROCm on Windows and Linux
@@ -31,16 +42,9 @@
perform detection of fp16 capabilities instead of forcing fp32 by default
add *experimental* support for `NPU` devices, typically available on AI-PC/Copilot-PC devices
- **SDNQ** add quant support to `nn.Embedding` type
- **Features**
- **nudenet** add `LlavaGuard` and `QwenGuard` as image safety evaulation models
- **schedulers** add `ER-SDE` solver
with with *1S, 2M, 3M* variants in both *EPS* and *FlowMatch* flavors
- installer auto-restart on upgrade
- enhanced filename pattern processing
allows for any *processing* property name (as defined in `modules/processing_class.py` and saved to `ui-config.json`)
allows for any *settings* property name (as defined in `modules/ui_definitions.py` and saved to `config.json`)
- **UI**
- `gallery` send-to button advanced options with right-click
- `tag autocomplete` quick toggle in main prompt area
- **Caption & Prompt Enhance**
- [Google Gemma 4] in *E2B* and *E4B* variants as well as *heretic* fine-tune
- **Agents**
@@ -81,6 +85,7 @@
- controlnet processor error handling
- error handling for same-device check
- error handling for undefined pipeline
- patch `z-image` for fp16 compatibility, thanks @resonantsky
## Update for 2026-04-01
-1
View File
@@ -132,7 +132,6 @@ def checkpoint_titles(use_short=False):
def list_models():
t0 = time.time()
global checkpoints_list # pylint: disable=global-statement
checkpoints_list.clear()
checkpoint_aliases.clear()
ext_filter = [".safetensors"]
+1 -1
View File
@@ -576,7 +576,7 @@ def apply_sdnq_to_module(model, weights_dtype="int8", quantized_matmul_dtype=Non
if layer_class_name in allowed_types and module.weight.dtype in {torch.float64, torch.float32, torch.float16, torch.bfloat16}:
if layer_class_name in embedding_types and not quant_embedding:
continue
elif (layer_class_name in conv_types or layer_class_name in conv_transpose_types) and not quant_conv:
if (layer_class_name in conv_types or layer_class_name in conv_transpose_types) and not quant_conv:
continue
quant_kwargs = {
"weights_dtype": weights_dtype,
+2 -3
View File
@@ -27,7 +27,7 @@ def apply_patches():
Uses modules.patches.patch() — idempotent, reversible via modules.patches.undo().
"""
global _applied
global _applied # pylint: disable=global-statement
if _applied:
return
_applied = True
@@ -39,7 +39,7 @@ def apply_patches():
# ------------------------------------------------------------------
# FeedForward._forward_silu_gating
# ------------------------------------------------------------------
def _patched_forward_silu_gating(self, x1, x3):
def _patched_forward_silu_gating(self, x1, x3): # pylint: disable=unused-argument
return _clamp_fp16(F.silu(x1) * x3)
# ------------------------------------------------------------------
@@ -106,4 +106,3 @@ def apply_patches():
sdnext_patches.patch(__name__, m.FeedForward, '_forward_silu_gating', _patched_forward_silu_gating)
sdnext_patches.patch(__name__, m.ZImageTransformerBlock, 'forward', _patched_zimage_block_forward)
sdnext_patches.patch(__name__, m.FinalLayer, 'forward', _patched_final_layer_forward)