reorder startup

Signed-off-by: Vladimir Mandic <mandic00@live.com>

CHANGELOG.md

@@ -1,8 +1,8 @@
 # Change Log for SD.Next
 
-## Update for 2026-09-04
+## Update for 2026-09-05
 
-### Highlights for 2026-09-04
+### Highlights for 2026-09-05
 
 All-about-optimizations: 
 - improved LoRA performance and quality, especially with quantized models
@@ -11,8 +11,11 @@ All-about-optimizations:
 - support for different caching stacks
 - compute updates across the board
 
-### Details for 2026-09-04
+### Details for 2026-09-05
 
+- **Models**
+  - [Anima 2.9B Preview v1](https://huggingface.co/yeoj34760/Anima-2.9B)  
+    expanded version of Anima 2B  
 - **LoRA**
   - *TODO*: see [LoRA docs](https://vladmandic.github.io/sdnext-docs/LoRA) for all of the improvements and usage instructions  
     *note*: lora now has its own settings section in *settings -> lora*
@@ -48,9 +51,11 @@ All-about-optimizations:
 - **Compute**
   - cuda: update `torch==2.14.0` with `cuda==13.2`
   - openvino: update `openvino==2026.3.1` with `torch==2.13.0`
-  - add `SD_SKIP_AUTOTUNE` env variable to skip triton autotune and use default config for all triton kernels  
+  - option to skip triton autotune and use default config for all triton kernels  
+    in *settings -> compute settings*  
     *note*: this may improve initial generate time, but may also reduce performance on some models  
-  - new optional transformer hooks: *settings -> compute add-ons*  
+  - new optional transformer hooks  
+    in *settings -> compute add-ons*  
     *PAG: Perturbed attention guidance, PAB: Pyramid attention broadcast, FBC: First Block Cache, FC: Faster Cache, LS: Layer Skip, MC: Mag Cache, TS: TaylorSeer*  
     *note*: compatibility of different methods varies across different models  
   - update `numpy` and `scipy` frozen requirements as required by new compute drivers  

modules/attention/dispatcher.py

@@ -111,9 +111,9 @@ def set_attention_dispatcher(pipe):
         current = a._AttentionBackendRegistry.get_active_backend() # pylint: disable=protected-access
         log.debug(f'Attention dispatcher: target={attn} previous={prev[0].value} active={current[0]} list={backends}')
     elif len(attn) > 0:
-        log.warning(f'Attention dispatcher: active={prev[0].value} list={backends} target={attn} not found')
+        log.warning(f'Attention dispatcher: active={prev[0].value} kernels={backends} target={attn} not found')
     else:
-        log.debug(f'Attention dispatcher: active={prev[0].value} list={backends}')
+        log.debug(f'Attention dispatcher: active={prev[0].value} kernels={backends}')
 
 
 def list_dispatcher_backends() -> list:

modules/attention/registry.py

@@ -69,6 +69,9 @@ class AttentionBackend:
     def available_on(self, platform: Platform) -> bool:
         return self.platforms is None or platform.backend in self.platforms
 
+    def __repr__(self) -> str:
+        return f'AttentionBackend(name="{self.name}" label="{self.label}" priority={self.priority} terminal={self.terminal} platforms={list(self.platforms) if self.platforms is not None else []} options={self.options} caps={list(self.caps)})'
+
 
 class Registry:
     def __init__(self):
@@ -98,5 +101,8 @@ class Registry:
     def with_cap(self, cap: str) -> list[AttentionBackend]:
         return [backend for backend in self.ordered() if cap in backend.caps]
 
+    def __repr__(self) -> str:
+        return f'Registry(backends={list(self.backends.keys())})'
+
 
 registry = Registry()

modules/loader.py

@@ -133,13 +133,6 @@ if ".dev" in torch.__version__ or "+git" in torch.__version__:
     torch.__version__ = re.search(r'[\d.]+[\d]', torch.__version__).group(0)
 timer.startup.record("torch")
 
-try:
-    from modules.sd_hijack_triton import install as install_autotune_report # pylint: disable=ungrouped-imports
-    install_autotune_report()
-except Exception as e:
-    log.warning(f'Triton logging: {e}')
-timer.startup.record("triton")
-
 try:
     import bitsandbytes # pylint: disable=unused-import
     _bnb = True
@@ -201,8 +194,6 @@ except Exception as e:
     _onnx = False
 timer.startup.record("onnx")
 
-timer.startup.record("fastapi")
-
 import gradio # pylint: disable=W0611,C0411
 timer.startup.record("gradio")
 errors.install([gradio])
@@ -237,6 +228,13 @@ diffusers.utils.import_utils._sdnq_available = True # pylint: disable=protected-
 diffusers.utils.import_utils._sdnq_version = sdnq.__version__ # pylint: disable=protected-access
 timer.startup.record("sdnq")
 
+try:
+    from modules.sd_hijack_triton import install as install_autotune_report # pylint: disable=ungrouped-imports
+    install_autotune_report()
+except Exception as e:
+    log.warning(f'Triton logging: {e}')
+timer.startup.record("triton")
+
 try:
     import pillow_jxl # pylint: disable=W0611,C0411
 except Exception:

modules/sd_hijack_triton.py

@@ -156,6 +156,7 @@ def run_hook(orig):
     """Report the register use of the config a sweep just chose. n_regs and n_spills are filled in
     when the driver loads the binary, so they exist only once the kernel has run, not at compile."""
     def wrapped(self, *args, **kwargs):
+        from modules import shared
         session = status['session']
         if session is not None:
             stop_progress(session)
@@ -172,7 +173,7 @@ def run_hook(orig):
                 if hasattr(arg, 'dtype'):
                     key += (str(arg.dtype),)
             needs_benchmark = len(self.configs) > 1 and key not in self.cache
-            skip_autotune = os.environ.get('SD_SKIP_AUTOTUNE', None) is not None
+            skip_autotune = (os.environ.get('SD_SKIP_AUTOTUNE', None) is not None) or shared.opts.triton_skip_autotune
             if needs_benchmark and skip_autotune:
                 self.cache[key] = self.configs[0] # pre-seed the cache so orig() takes its cache-hit path and skips the sweep
                 needs_benchmark = False

modules/shared.py

@@ -10,8 +10,8 @@ from typing import TYPE_CHECKING
 import gradio as gr
 from installer import print_dict # pylint: disable=unused-import
 from modules.logger import log
-log.debug('Initializing: shared module')
 
+log.debug('Initializing: shared module')
 import modules.memmon
 import modules.paths as paths
 from modules.json_helpers import readfile # pylint: disable=W0611
@@ -19,6 +19,9 @@ from modules.shared_helpers import listdir, req # pylint: disable=W0611
 from modules import errors, devices, shared_state, cmd_args, theme, history, files_cache # pylint: disable=unused-import
 from modules.memstats import memory_stats # pylint: disable=unused-import
 
+# main entry point that triggers package imports
+from modules import loader # pylint: disable=unused-import
+
 log.debug('Initializing: pipelines')
 from modules import shared_items # pylint: disable=unused-import
 from modules.caption.openclip import get_clip_models, refresh_clip_models # pylint: disable=unused-import

modules/ui_definitions.py

@@ -287,6 +287,7 @@ def create_settings(cmd_opts):
         "diffusers_fuse_projections": OptionInfo(False, "Fused projections"),
         "torch_expandable_segments": OptionInfo(False, "Expandable segments"),
         "torch_sync": OptionInfo(True, "Force synchronize"),
+        "triton_skip_autotune": OptionInfo(False, "Skip Triton autotune", gr.Checkbox),
         "cudnn_enabled": OptionInfo("default", "cuDNN enabled", gr.Radio, {"choices": ["default", "true", "false"]}),
         "cudnn_benchmark": OptionInfo(devices.backend != "rocm", "cuDNN full-depth benchmark"),
         "cudnn_benchmark_limit": OptionInfo(10, "cuDNN benchmark limit", gr.Slider, {"minimum": 0, "maximum": 100, "step": 1}),