update

README.md

@@ -13,8 +13,13 @@ All code changes are merged upstream whenever possible
 
 Fork does differ in few things:
 
+- Different start script  
+  > ./automatic.sh  
+- Drops compatibility with `python` **3.7** and requires **3.10**  
+- If you're using **PyTorch 2.0** models will be auto-compiled and optimized on load  
+  Using `max-tune` 
 - Updated **Python** libraries to latest known compatible versions  
-  e.g. `accelerate`, `transformers`, `numpy`, etc.    
+  e.g. `accelerate`, `transformers`, `numpy`, etc.  
 - Includes opinionated **System** and **Options** configuration  
   e.g. `samplers`, `upscalers`, etc.  
 - Includes reskinned **UI**  
@@ -23,16 +28,17 @@ Fork does differ in few things:
 - Ships with additional **extensions**  
   e.g. `System Info`  
 - Uses simplified folder structure  
-  e.g. `/train`, `outputs`  
+  e.g. `/train`, `/outputs/*`  
 - Modified training templates  
 
 Only Python library which is not auto-updated is `PyTorch` itself as that is very system specific  
-I'm currently using **PyTorch 2.0-nightly** compiled with **CUDA 11.8**:
+I'm currently using **PyTorch 2.0-nightly** compiled with **CUDA 11.8** and with **Triton** optimizations:
 
-> pip3 install --pre torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/nightly/cu118  
+> pip3 install --pre torch torchvision torchaudio torchtriton --extra-index-url https://download.pytorch.org/whl/nightly/cu118  
 > pip show torch  
 > 2.0.0.dev20230111+cu118  
 
+
 <br>
 
 ## Docs

automatic.sh

@@ -1,5 +1,5 @@
 #/bin/env bash
 
 export PYTORCH_CUDA_ALLOC_CONF=garbage_collection_threshold:0.9,max_split_size_mb:512
-python launch.py --api --disable-console-progressbars
+python launch.py --api --disable-console-progressbars 
-# python launch.py --api --xformers --disable-console-progressbars
+# python launch.py --api --xformers --disable-console-progressbars --opt-channelslast

config.json

@@ -48,7 +48,7 @@
     "memmon_poll_rate": 1,
     "samples_log_stdout": false,
     "multiple_tqdm": true,
-    "unload_models_when_training": true,
+    "unload_models_when_training": false,
     "pin_memory": true,
     "save_optimizer_state": false,
     "dataset_filename_word_regex": "",

modules/sd_models.py

@@ -358,6 +358,17 @@ def load_model(checkpoint_info=None):
     sd_hijack.model_hijack.hijack(sd_model)
 
     sd_model.eval()
+
+    """
+    try:
+        t0 = time.time()
+        sd_model = torch.compile(sd_model, mode="max-autotune", fullgraph=True)
+        t1 = time.time()
+        print(f"Model compiled in {round(t1 - t0, 2)} sec")
+    except Exception as err:
+        print(f"Model compile not supported: {err}")
+    """
+
     shared.sd_model = sd_model
 
     sd_hijack.model_hijack.embedding_db.load_textual_inversion_embeddings(force_reload=True)  # Reload embeddings after model load as they may or may not fit the model

requirements_versions.txt

@@ -1,17 +1,15 @@
 accelerate==0.15.0
 basicsr==1.4.2
-blendmodes==2022
-transformers==4.25.1
 basicsr==1.4.2
+blendmodes==2022
 gfpgan==1.3.8
 GitPython==3.1.27
 gradio==3.15.0
 numpy==1.23.5
-Pillow==9.4.0
-realesrgan==0.3.0
-torch
 omegaconf==2.2.3
+Pillow==9.4.0
 pytorch_lightning==1.7.7
+realesrgan==0.3.0
 scikit-image==0.19.2
 timm==0.6.7
 torchdiffeq==0.2.3