update numpy/scripy and add SD_SKIP_AUTOTUNE

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

CHANGELOG.md

@@ -1,6 +1,6 @@
 # Change Log for SD.Next
 
-## Update for 2026-09-02
+## Update for 2026-09-04
 
 - **LoRA**
   - *TODO*: see [LoRA docs](https://vladmandic.github.io/sdnext-docs/LoRA) for all of the improvements and usage instructions  
@@ -36,11 +36,15 @@
   - on-demand convert standard model on-demand 
 - **Compute**
   - openvino: update `openvino==2026.3.1` with `torch==2.13.0`
-- **Other**
-  - Video Preview: TAESD support for **MiniMax**
+  - add `SD_SKIP_AUTOTUNE` env variable to skip triton autotune and use default config for all triton kernels  
+    *note*: this may improve initial generate time, but may also reduce performance on some models  
   - new optional transformer hooks: *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  
+    *note*: this may break compatibility with some legacy packages, so report any finidings  
+- **Other**
+  - Video Preview: TAESD support for **MiniMax**
   - remove `/redocs` as `/docs` are primary api docs  
 - **Wiki/Docs**:
   - [MiniMax](https://vladmandic.github.io/sdnext-docs/MiniMax) updated with LoRA, Turbo and other improvements

constraints.txt

@@ -1,3 +1,3 @@
 fastapi==0.124.4
-numpy==2.1.2
+numpy==2.5.2
 Pillow==12.2.0

installer.py

@@ -1277,10 +1277,8 @@ def install_pydantic():
 
 
 def install_scipy():
-    if args.new or (sys.version_info >= (3, 14)):
-        install('scipy==1.17.1', ignore=True, quiet=True)
-    else:
-        install('scipy==1.14.1', ignore=True, quiet=True)
+    # install('scipy==1.14.1', ignore=True, quiet=True)
+    install('scipy==1.18.1', ignore=True, quiet=True)
 
 
 def install_opencv():
@@ -1311,7 +1309,6 @@ def install_insightface():
 def install_optional():
     t_start = time.time()
     log.info('Installing optional requirements...')
-    install('pillow-heif')
     install('addict')
     install('yapf')
     install('--no-build-isolation git+https://github.com/Disty0/BasicSR@23c1fb6f5c559ef5ce7ad657f2fa56e41b121754', 'basicsr', ignore=True, quiet=True)
@@ -1322,11 +1319,14 @@ def install_optional():
     install('Cython', ignore=True, quiet=True)
     install('gguf', ignore=True, quiet=True)
     install('hf_transfer', ignore=True, quiet=True)
-    install('hf_xet', ignore=True, quiet=True)
     install('nvidia-ml-py', ignore=True, quiet=True)
+    install('pillow-heif')
     install('pillow-jxl-plugin==1.3.7', ignore=True, quiet=True)
     install('ultralytics==8.4.67', ignore=True, quiet=True)
     install('open-clip-torch', no_deps=True, quiet=True)
+    install('runai_model_streamer', ignore=True, quiet=True)
+    install('facexlib', ignore=True, quiet=True)
+    install('omegaconf', ignore=True, quiet=True)
     install('git+https://github.com/tencent-ailab/IP-Adapter.git', 'ip_adapter', ignore=True, quiet=True)
     # install('git+https://github.com/openai/CLIP.git', 'clip', quiet=True, no_build_isolation=True)
     ts('optional', t_start)

modules/sd_hijack_triton.py

@@ -171,6 +171,12 @@ def run_hook(orig):
                 if hasattr(arg, 'dtype'):
                     key += (str(arg.dtype),)
             needs_benchmark = len(self.configs) > 1 and key not in self.cache
+            import os
+            skip_autotune = os.environ.get('SD_SKIP_AUTOTUNE', None) is not None
+            if needs_benchmark and skip_autotune:
+                log.trace('Autotune: skip')
+                self.cache[key] = self.configs[0] # pre-seed the cache so orig() takes its cache-hit path and skips the sweep
+                needs_benchmark = False
             if needs_benchmark:
                 try:
                     total = len(self.prune_configs(kwargs))

requirements.txt

@@ -32,15 +32,15 @@ requests==2.34.2
 tqdm==4.70.0
 accelerate==1.14.0
 einops==0.8.2
-huggingface_hub==1.28.0
+huggingface_hub==1.30.0
 hf_xet==1.6.0
-numpy==2.1.2
+numpy==2.5.2
 pandas==2.3.1
-protobuf==7.36.0
+protobuf==7.36.1
 pytorch_lightning==2.6.5
 urllib3==1.26.19
 Pillow==12.2.0
-timm==1.0.28
+timm==1.0.29
 pyparsing==3.3.2
 typing-extensions==4.16.0
 sentencepiece==0.2.1