Several improvements (#117)

CMD_FLAGS.txt

@@ -0,0 +1 @@
+--chat

INSTRUCTIONS.TXT

@@ -1,39 +0,0 @@
-Thank you for downloading oobabooga/text-generation-webui.
-
-# Installation
-
-Run the "start" script. If all goes right, it should take care of
-everything for you.
-
-To launch the web UI in the future after it is already installed, run
-the same "start" script.
-
-# Updating the web UI
-
-Run the "update" script. This will only install the updates, so it should
-be much faster than the initial installation.
-
-# Adding flags like --chat, --notebook, etc
-
-Edit the "webui.py" script using a text editor and add the desired flags
-to the CMD_FLAGS variable at the top. It should look like this:
-
-CMD_FLAGS = '--chat'
-
-For instance, to add the --api flag, change it to
-
-CMD_FLAGS = '--chat --api'
-
-# Running an interactive shell
-
-To run an interactive shell in the miniconda environment, run the "cmd"
-script. This is useful for installing additional requirements manually.
-
-# Using an AMD GPU in Linux
-
-Requires ROCm SDK 5.4.2 or 5.4.3 to be installed.
-Some systems may also need: sudo apt-get install libstdc++-12-dev
-
-Edit the "webui.py" script using a text editor and un-comment and modify the
-lines near the top of the script according to your setup. In particular, modify
-the os.environ["ROCM_PATH"] = '/opt/rocm' line to point to your ROCm installation.

INSTRUCTIONS.txt

@@ -0,0 +1,38 @@
+Thank you for downloading oobabooga/text-generation-webui!
+
+# Installation
+
+Run the "start" script. It will install the web UI and all of its
+dependencies inside this folder.
+
+To launch the web UI in the future after it is already installed, run the
+"start" script again.
+
+# Updating the web UI
+
+Run the "update" script. It will install the updates only, so it should
+be much faster than the initial installation.
+
+# Adding flags like --chat, --notebook, etc
+
+Open the "CMD_FLAGS.txt" file with a text editor, add your flags, and
+save the file. For instance, to add the --api flag, change the file
+contents to
+
+--chat --api
+
+# Running an interactive shell
+
+Sometimes you may need to install some additional Python package. To do
+that, run the "cmd" script and type your commands inside the terminal
+window that will appear.
+
+# Using an AMD GPU in Linux
+
+Requires ROCm SDK 5.4.2 or 5.4.3 to be installed. Some systems may also
+need: sudo apt-get install libstdc++-12-dev
+
+Edit the "webui.py" script using a text editor and un-comment and
+modify the lines near the top of the script according to your setup. In
+particular, modify the os.environ["ROCM_PATH"] = '/opt/rocm' line to
+point to your ROCm installation.

generate_zips.sh

@@ -1,6 +1,6 @@
 mkdir oobabooga_{windows,linux,macos,wsl}
 for p in windows macos linux wsl; do
-  if [ "$p" == "wsl" ]; then cp {*$p*\.*,webui.py,INSTRUCTIONS-WSL.TXT} oobabooga_$p;
-  else cp {*$p*\.*,webui.py,INSTRUCTIONS.TXT} oobabooga_$p; fi
+  if [ "$p" == "wsl" ]; then cp {*$p*\.*,webui.py,INSTRUCTIONS-WSL.txt} oobabooga_$p;
+  else cp {*$p*\.*,webui.py,INSTRUCTIONS.txt,CMD_FLAGS.txt} oobabooga_$p; fi
   zip -r oobabooga_$p.zip oobabooga_$p;
 done

webui.py

@@ -1,7 +1,7 @@
 import argparse
 import glob
-import re
 import os
+import re
 import site
 import subprocess
 import sys
@@ -9,19 +9,19 @@ import sys
 script_dir = os.getcwd()
 conda_env_path = os.path.join(script_dir, "installer_files", "env")
 
-# Use this to set your command-line flags. For the full list, see:
-# https://github.com/oobabooga/text-generation-webui/#starting-the-web-ui
-# Example: CMD_FLAGS = '--chat --listen'
-CMD_FLAGS = '--chat'
-
-
-# Allows users to set flags in "OOBABOOGA_FLAGS" environment variable
+# Command-line flags
 if "OOBABOOGA_FLAGS" in os.environ:
     CMD_FLAGS = os.environ["OOBABOOGA_FLAGS"]
     print("The following flags have been taken from the environment variable 'OOBABOOGA_FLAGS':")
     print(CMD_FLAGS)
     print("To use the CMD_FLAGS Inside webui.py, unset 'OOBABOOGA_FLAGS'.\n")
-    
+else:
+    cmd_flags_path = os.path.join(script_dir, "CMD_FLAGS.txt")
+    if os.path.exists(cmd_flags_path):
+        CMD_FLAGS = open(cmd_flags_path, 'r').read().strip()
+    else:
+        CMD_FLAGS = '--chat'
+
 
 # Remove the '# ' from the following lines as needed for your AMD GPU on Linux
 # os.environ["ROCM_PATH"] = '/opt/rocm'
@@ -74,6 +74,11 @@ def check_env():
         sys.exit()
 
 
+def clear_cache():
+    run_cmd("conda clean -a -y", environment=True)
+    run_cmd("python -m pip cache purge", environment=True)
+
+
 def install_dependencies():
     # Select your GPU or, choose to run in CPU mode
     print("What is your GPU")
@@ -81,12 +86,12 @@ def install_dependencies():
     print("A) NVIDIA")
     print("B) AMD (Linux/MacOS only. Requires ROCm SDK 5.4.2/5.4.3 on Linux)")
     print("C) Apple M Series")
-    print("D) None (I want to run in CPU mode)")
+    print("D) None (I want to run models in CPU mode)")
     print()
     gpuchoice = input("Input> ").lower()
 
     if gpuchoice == "d":
-        print_big_message("Once the installation ends, make sure to open webui.py with a text editor\nand add the --cpu flag to CMD_FLAGS.")
+        print_big_message("Once the installation ends, make sure to open CMD_FLAGS.txt with\na text editor and add the --cpu flag.")
 
     # Install the version of PyTorch needed
     if gpuchoice == "a":
@@ -113,10 +118,10 @@ def install_dependencies():
     run_cmd("git clone https://github.com/oobabooga/text-generation-webui.git", assert_success=True, environment=True)
 
     # Install the webui dependencies
-    update_dependencies()
+    update_dependencies(initial_installation=True)
 
 
-def update_dependencies():
+def update_dependencies(initial_installation=False):
     os.chdir("text-generation-webui")
     run_cmd("git pull", assert_success=True, environment=True)
 
@@ -135,16 +140,19 @@ def update_dependencies():
         run_cmd("python -m pip uninstall -y " + package_name, environment=True)
         print(f"Uninstalled {package_name}")
 
-    # Installs/Updates dependencies from all requirements.txt
+    # Installs/Updates the project dependencies
     run_cmd("python -m pip install -r requirements.txt --upgrade", assert_success=True, environment=True)
-    extensions = next(os.walk("extensions"))[1]
-    for extension in extensions:
-        if extension in ['superbooga']:  # No wheels available for dependencies
-            continue
 
-        extension_req_path = os.path.join("extensions", extension, "requirements.txt")
-        if os.path.exists(extension_req_path):
-            run_cmd("python -m pip install -r " + extension_req_path + " --upgrade", assert_success=True, environment=True)
+    # Installs the extensions dependencies (only on the first install)
+    if initial_installation:
+        extensions = next(os.walk("extensions"))[1]
+        for extension in extensions:
+            if extension in ['superbooga']:  # No wheels available for dependencies
+                continue
+
+            extension_req_path = os.path.join("extensions", extension, "requirements.txt")
+            if os.path.exists(extension_req_path):
+                run_cmd("python -m pip install -r " + extension_req_path + " --upgrade", assert_success=True, environment=True)
 
     # The following dependencies are for CUDA, not CPU
     # Parse output of 'pip show torch' to determine torch version
@@ -153,6 +161,7 @@ def update_dependencies():
 
     # Check for '+cu' or '+rocm' in version string to determine if torch uses CUDA or ROCm   check for pytorch-cuda as well for backwards compatibility
     if '+cu' not in torver and '+rocm' not in torver and run_cmd("conda list -f pytorch-cuda | grep pytorch-cuda", environment=True, capture_output=True).returncode == 1:
+        clear_cache()
         return
 
     # Get GPU CUDA/compute support
@@ -178,7 +187,7 @@ def update_dependencies():
         os.chdir("exllama")
         run_cmd("git pull", environment=True)
         os.chdir("..")
-        
+
     # Pre-installed exllama module does not support AMD GPU
     if '+rocm' in torver:
         run_cmd("python -m pip uninstall -y exllama", environment=True)
@@ -222,7 +231,7 @@ def update_dependencies():
     # Install GPTQ-for-LLaMa dependencies
     os.chdir("GPTQ-for-LLaMa")
     run_cmd("git pull", environment=True)
-    
+
     # Finds the path to your dependencies
     for sitedir in site.getsitepackages():
         if "site-packages" in sitedir:
@@ -272,6 +281,8 @@ def update_dependencies():
 
         print("Continuing with install..")
 
+    clear_cache()
+
 
 def download_model():
     os.chdir("text-generation-webui")
@@ -301,7 +312,7 @@ if __name__ == "__main__":
 
         # Check if a model has been downloaded yet
         if len([item for item in glob.glob('text-generation-webui/models/*') if not item.endswith(('.txt', '.yaml'))]) == 0:
-            print_big_message("WARNING: You haven't downloaded any model yet.\nOnce the web UI launches, head over to the bottom of the \"Model\" tab and download one.")
+            print_big_message("WARNING: You haven't downloaded any model yet.\nOnce the web UI launches, head over to the \"Model\" tab and download one.")
 
         # Workaround for llama-cpp-python loading paths in CUDA env vars even if they do not exist
         conda_path_bin = os.path.join(conda_env_path, "bin")