fix upscale and add some compile options

CHANGELOG.md

@@ -43,13 +43,14 @@
   - Support `--ckpt none` to skip loading a model  
 - **XYZ grid**
   - Add refiner options to XYZ Grid  
-  - Add option to create only subimages in XYZ grid, thanks @midcoastal
+  - Add option to create only subgrids in XYZ grid, thanks @midcoastal
   - Allow custom font, background and text color in settings
 - **Fixes**  
   - Fix `params.txt` saved before actual image
   - Fix inpaint  
   - Fix manual grid image save  
   - Fix img2img init image save  
+  - Fix upscale in txt2img for batch counts when no hires is used  
   - More uniform models paths  
   - Safe scripts callback execution  
   - Improved extension compatibility  

installer.py

@@ -775,26 +775,26 @@ def install_requirements():
 # set environment variables controling the behavior of various libraries
 def set_environment():
     log.debug('Setting environment tuning')
-    os.environ.setdefault('USE_TORCH', '1')
-    os.environ.setdefault('TF_CPP_MIN_LOG_LEVEL', '2')
     os.environ.setdefault('ACCELERATE', 'True')
-    os.environ.setdefault('FORCE_CUDA', '1')
     os.environ.setdefault('ATTN_PRECISION', 'fp16')
-    os.environ.setdefault('PYTORCH_CUDA_ALLOC_CONF', 'garbage_collection_threshold:0.8,max_split_size_mb:512')
-    os.environ.setdefault('CUDA_LAUNCH_BLOCKING', '0')
-    os.environ.setdefault('CUDA_CACHE_DISABLE', '0')
     os.environ.setdefault('CUDA_AUTO_BOOST', '1')
-    os.environ.setdefault('CUDA_MODULE_LOADING', 'LAZY')
+    os.environ.setdefault('CUDA_CACHE_DISABLE', '0')
     os.environ.setdefault('CUDA_DEVICE_DEFAULT_PERSISTING_L2_CACHE_PERCENTAGE_LIMIT', '0')
+    os.environ.setdefault('CUDA_LAUNCH_BLOCKING', '0')
+    os.environ.setdefault('CUDA_MODULE_LOADING', 'LAZY')
+    os.environ.setdefault('FORCE_CUDA', '1')
     os.environ.setdefault('GRADIO_ANALYTICS_ENABLED', 'False')
-    os.environ.setdefault('SAFETENSORS_FAST_GPU', '1')
-    os.environ.setdefault('NUMEXPR_MAX_THREADS', '16')
-    os.environ.setdefault('PYTHONHTTPSVERIFY', '0')
-    os.environ.setdefault('HF_HUB_DISABLE_TELEMETRY', '1')
     os.environ.setdefault('HF_HUB_DISABLE_EXPERIMENTAL_WARNING', '1')
-    os.environ.setdefault('UVICORN_TIMEOUT_KEEP_ALIVE', '60')
+    os.environ.setdefault('HF_HUB_DISABLE_TELEMETRY', '1')
     os.environ.setdefault('K_DIFFUSION_USE_COMPILE', '0')
+    os.environ.setdefault('NUMEXPR_MAX_THREADS', '16')
+    os.environ.setdefault('PYTHONHTTPSVERIFY', '0')
+    os.environ.setdefault('PYTORCH_CUDA_ALLOC_CONF', 'garbage_collection_threshold:0.8,max_split_size_mb:512')
+    os.environ.setdefault('SAFETENSORS_FAST_GPU', '1')
+    os.environ.setdefault('TF_CPP_MIN_LOG_LEVEL', '2')
     os.environ.setdefault('TF_ENABLE_ONEDNN_OPTS', '0')
+    os.environ.setdefault('USE_TORCH', '1')
+    os.environ.setdefault('UVICORN_TIMEOUT_KEEP_ALIVE', '60')
     if sys.platform == 'darwin':
         os.environ.setdefault('PYTORCH_ENABLE_MPS_FALLBACK', '1')
 

modules/processing_diffusers.py

@@ -42,16 +42,14 @@ def hires_resize(latents): # input=latents output=pil
         if latent_upscaler is not None:
             latents = torch.nn.functional.interpolate(latents, size=(p.hr_upscale_to_y // 8, p.hr_upscale_to_x // 8), mode=latent_upscaler["mode"], antialias=latent_upscaler["antialias"])
         first_pass_images = vae_decode(latents=latents, model=shared.sd_model, full_quality=p.full_quality, output_type='pil')
-        p.init_images = []
+        resized_images = []
         for img in first_pass_images:
             if latent_upscaler is None:
-                init_image = images.resize_image(1, img, p.hr_upscale_to_x, p.hr_upscale_to_y, upscaler_name=p.hr_upscaler)
+                resized_image = images.resize_image(1, img, p.hr_upscale_to_x, p.hr_upscale_to_y, upscaler_name=p.hr_upscaler)
             else:
-                init_image = img
-            # if is_refiner_enabled:
-            #    init_image = vae_encode(init_image, model=shared.sd_model, full_quality=p.full_quality)
-            p.init_images.append(init_image)
-        return p.init_images
+                resized_image = img
+            resized_images.append(resized_image)
+        return resized_images
 
     def save_intermediate(latents, suffix):
         for i in range(len(latents)):
@@ -489,7 +487,7 @@ def calculate_refiner_steps():
         return results
 
     # optional hires pass
-    if p.enable_hr and p.hr_upscaler != 'None' and p.denoising_strength > 0 and len(getattr(p, 'init_images', [])) == 0:
+    if p.enable_hr and getattr(p, 'hr_upscaler', 'None') != 'None' and len(getattr(p, 'init_images', [])) == 0:
         p.is_hr_pass = True
     latent_scale_mode = shared.latent_upscale_modes.get(p.hr_upscaler, None) if (hasattr(p, "hr_upscaler") and p.hr_upscaler is not None) else shared.latent_upscale_modes.get(shared.latent_upscale_default_mode, "None")
     if p.is_hr_pass:
@@ -501,7 +499,7 @@ def calculate_refiner_steps():
                 save_intermediate(latents=output.images, suffix="-before-hires")
             shared.state.job = 'upscale'
             output.images = hires_resize(latents=output.images)
-            if latent_scale_mode is not None or p.hr_force:
+            if (latent_scale_mode is not None or p.hr_force) and p.denoising_strength > 0:
                 p.ops.append('hires')
                 shared.sd_model = sd_models.set_diffuser_pipe(shared.sd_model, sd_models.DiffusersTaskType.IMAGE_2_IMAGE)
                 recompile_model(hires=True)
@@ -518,7 +516,7 @@ def calculate_refiner_steps():
                     guidance_rescale=p.diffusers_guidance_rescale,
                     output_type='latent' if hasattr(shared.sd_model, 'vae') else 'np',
                     clip_skip=p.clip_skip,
-                    image=p.init_images,
+                    image=output.images,
                     strength=p.denoising_strength,
                     desc='Hires',
                 )

modules/sd_models_compile.py

@@ -76,6 +76,9 @@ def compile_stablefast(sd_model):
     warnings.filterwarnings("ignore", category=torch.jit.TracerWarning)
     config.enable_cuda_graph = shared.opts.cuda_compile_fullgraph
     config.enable_jit_freeze = shared.opts.diffusers_eval
+    config.memory_format = torch.channels_last if shared.opts.opt_channelslast else torch.contiguous_format
+    # config.enable_cnn_optimization
+    # config.prefer_lowp_gemm
     try:
         t0 = time.time()
         sd_model = sf.compile(sd_model, config)

modules/shared.py

@@ -307,7 +307,7 @@ def default(obj):
     "cuda_compile_vae": OptionInfo(True if cmd_opts.use_openvino else False, "Compile VAE"),
     "cuda_compile_upscaler": OptionInfo(True if cmd_opts.use_openvino else False, "Compile upscaler"),
     "cuda_compile_backend": OptionInfo("openvino_fx" if cmd_opts.use_openvino else "none", "Model compile backend", gr.Radio, {"choices": ['none', 'inductor', 'cudagraphs', 'aot_ts_nvfuser', 'hidet', 'ipex', 'openvino_fx', 'stable-fast']}),
-    "cuda_compile_mode": OptionInfo("default", "Model compile mode", gr.Radio, {"choices": ['default', 'reduce-overhead', 'max-autotune']}),
+    "cuda_compile_mode": OptionInfo("default", "Model compile mode", gr.Radio, {"choices": ['default', 'reduce-overhead', 'max-autotune', 'max-autotune-no-cudagraphs']}),
     "cuda_compile_fullgraph": OptionInfo(False, "Model compile fullgraph"),
     "cuda_compile_precompile": OptionInfo(False if cmd_opts.use_openvino else True, "Model compile precompile"),
     "cuda_compile_verbose": OptionInfo(False, "Model compile verbose mode"),

modules/upscaler.py

@@ -228,7 +228,7 @@ def compile_upscaler(model, name=""):
 
             if modules.shared.opts.cuda_compile_backend == "openvino_fx":
                 from modules.intel.openvino import openvino_fx # pylint: disable=unused-import
-                from modules.sd_models_compile import CompiledModelState
+                from modules.sd_models_compile import CompiledModelState # pylint: disable=unused-import
                 torch._dynamo.eval_frame.check_if_dynamo_supported = lambda: True # pylint: disable=protected-access
 
             log_level = logging.WARNING if modules.shared.opts.cuda_compile_verbose else logging.CRITICAL # pylint: disable=protected-access