handle xformers

CHANGELOG.md

@@ -135,9 +135,14 @@ or even free speedups and quality improvements (regardless of which workflows yo
   - to enable search, make sure all models have set hash values  
     *Models -> Valida -> Calculate hashes*  
 - **LoRA**
-  - for *backend:original*, lyco handler has been removed and replaced with new  
-    unified lora/lyco handler that supports all variants of loras  
+  - new unified LoRA handler for all LoRA types (lora, lyco, loha, lokr, locon, etc.)  
+    applies to both original and diffusers backend  
+    thanks @AI-Casanova for diffusers port  
+  - for *backend:original*, separate lyco handler has been removed  
 - **Compute**  
+  - **CUDA**:  
+    - default updated to `torch` *2.1.0* with cuda *12.1*  
+    - testing moved to `torch` *2.2.0-dev/cu122*  
   - **Intel Arc/IPEX**:  
     - tons of optimizations, built-in binary wheels for Windows  
       i have to say, intel arc/ipex is getting to be quite a player, especially with openvino  
@@ -145,8 +150,10 @@ or even free speedups and quality improvements (regardless of which workflows yo
   - **AMD ROCm**:  
     - updated installer to support detect `ROCm` *5.4/5.5/5.6/5.7*  
     - support for `torch-rocm-5.7`
-  - **CUDA**:  
-    - testing moved to `torch` *2.2.0-dev/cu121*  
+  - **xFormers**:
+    - default updated to *0.0.22*  
+    - note that latest xformers are still not compatible with standard torch 2.1.0 with cuda 12.1  
+      either downgrade torch to 2.0.1 with cuda 11.8 or build xformers manually  
   - **GC**:  
     - custom garbage collect threshold to reduce vram memory usage, thanks @Disty0  
       see *settings -> compute -> gc*  

installer.py

@@ -341,8 +341,6 @@ def check_python():
 
 # check torch version
 def check_torch():
-    if args.quick:
-        return
     if args.skip_torch:
         log.info('Skipping Torch tests')
         return
@@ -362,8 +360,8 @@ def check_torch():
         pass
     elif allow_cuda and (shutil.which('nvidia-smi') is not None or os.path.exists(os.path.join(os.environ.get('SystemRoot') or r'C:\Windows', 'System32', 'nvidia-smi.exe'))):
         log.info('nVidia CUDA toolkit detected')
-        torch_command = os.environ.get('TORCH_COMMAND', 'torch torchvision --index-url https://download.pytorch.org/whl/cu118')
-        xformers_package = os.environ.get('XFORMERS_PACKAGE', 'xformers==0.0.20' if opts.get('cross_attention_optimization', '') == 'xFormers' else 'none')
+        torch_command = os.environ.get('TORCH_COMMAND', 'torch torchvision --index-url https://download.pytorch.org/whl/cu121')
+        xformers_package = os.environ.get('XFORMERS_PACKAGE', 'xformers==0.0.22' if opts.get('cross_attention_optimization', '') == 'xFormers' else 'none')
     elif allow_rocm and (shutil.which('rocminfo') is not None or os.path.exists('/opt/rocm/bin/rocminfo') or os.path.exists('/dev/kfd')):
         log.info('AMD ROCm toolkit detected')
         os.environ.setdefault('PYTORCH_HIP_ALLOC_CONF', 'garbage_collection_threshold:0.8,max_split_size_mb:512')
@@ -483,6 +481,10 @@ def check_torch():
     try:
         if 'xformers' in xformers_package:
             install(f'--no-deps {xformers_package}', ignore=True)
+            import torch
+            import xformers
+            if torch.__version__ != '2.0.1+cu118' and xformers.__version__ in ['0.0.22', '0.0.21', '0.0.20']:
+                log.warning(f'Likely incompatible torch with: xformers=={xformers.__version__} installed: torch=={torch.__version__} required: torch==2.0.1+cu118 - build xformers manually or downgrade torch')
         elif not args.experimental:
             x = pkg_resources.working_set.by_key.get('xformers', None)
             if x is not None:

modules/processing.py

@@ -643,7 +643,7 @@ def process_images(p: StableDiffusionProcessing) -> Processed:
 
         if not shared.opts.cuda_compile:
             modules.sd_models.apply_token_merging(p.sd_model, p.get_token_merging_ratio())
-            modules.sd_hijack_freeu.apply_freeu(p.sd_model, shared.backend == shared.Backend.ORIGINAL)
+            modules.sd_hijack_freeu.apply_freeu(p, shared.backend == shared.Backend.ORIGINAL)
 
         if shared.cmd_opts.profile:
             """

modules/sd_hijack_freeu.py

@@ -126,11 +126,12 @@ def ratio_to_region(width: float, offset: float, n: int):
     return round(start), round(end), inverted
 
 
-def apply_freeu(model, backend_original):
+def apply_freeu(p, backend_original):
     global state_enabled # pylint: disable=global-statement
     global cat_original # pylint: disable=global-statement
     if backend_original:
         if opts.freeu_enabled:
+            p.extra_generation_params['FreeU'] = f'b1={opts.freeu_b1} b2={opts.freeu_b2} s1={opts.freeu_s1} s2={opts.freeu_s2}'
             if not state_enabled: # otherwise already patched
                 cat_original = th.cat
                 th.cat = functools.partial(free_u_cat_hijack, original_function=th.cat)
@@ -139,12 +140,13 @@ def apply_freeu(model, backend_original):
             if cat_original is not None:
                 th.cat = cat_original
                 state_enabled = False
-    elif hasattr(model, 'enable_freeu'):
+    elif hasattr(p.sd_model, 'enable_freeu'):
         if opts.freeu_enabled:
-            model.enable_freeu(s1=opts.freeu_s1, s2=opts.freeu_s2, b1=opts.freeu_b1, b2=opts.freeu_b2)
+            p.extra_generation_params['FreeU'] = f'b1={opts.freeu_b1} b2={opts.freeu_b2} s1={opts.freeu_s1} s2={opts.freeu_s2}'
+            p.sd_model.enable_freeu(s1=opts.freeu_s1, s2=opts.freeu_s2, b1=opts.freeu_b1, b2=opts.freeu_b2)
             state_enabled = True
         elif state_enabled:
-            model.disable_freeu()
+            p.sd_model.disable_freeu()
             state_enabled = False
     if opts.freeu_enabled:
         log.info(f'Applying free-u: b1={opts.freeu_b1} b2={opts.freeu_b2} s1={opts.freeu_s1} s2={opts.freeu_s2}')

modules/sd_hijack_hypertile.py

@@ -142,6 +142,7 @@ def context_hypertile_vae(p):
         return nullcontext()
     else:
         shared.log.info(f'Applying hypertile: vae={shared.opts.hypertile_vae_tile}')
+        p.extra_generation_params['Hypertile VAE'] = shared.opts.hypertile_vae_tile
         return split_attention(vae, tile_size=shared.opts.hypertile_vae_tile, min_tile_size=128, swap_size=1)
 
 
@@ -161,6 +162,7 @@ def context_hypertile_unet(p):
         return nullcontext()
     else:
         shared.log.info(f'Applying hypertile: unet={shared.opts.hypertile_unet_tile}')
+        p.extra_generation_params['Hypertile UNet'] = shared.opts.hypertile_unet_tile
         return split_attention(unet, tile_size=shared.opts.hypertile_unet_tile, min_tile_size=128, swap_size=1)