initial multi-lora support (#1103)

---------

Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>

modules/LoRA.py

@@ -4,19 +4,31 @@ import torch
 from peft import PeftModel
 
 import modules.shared as shared
-from modules.models import reload_model
 
 
-def add_lora_to_model(lora_name):
+def add_lora_to_model(lora_names):
+    prior_set = set(shared.lora_names)
+    added_set = set(lora_names) - prior_set
+    removed_set = prior_set - set(lora_names)
+    shared.lora_names = list(lora_names)
 
-    # If a LoRA had been previously loaded, or if we want
-    # to unload a LoRA, reload the model
-    if shared.lora_name not in ['None', ''] or lora_name in ['None', '']:
-        reload_model()
-    shared.lora_name = lora_name
+    # Nothing to do = skip.
+    if len(added_set) == 0 and len(removed_set) == 0:
+        return
 
-    if lora_name not in ['None', '']:
-        print(f"Adding the LoRA {lora_name} to the model...")
+    # Only adding, and already peft? Do it the easy way.
+    if len(removed_set) == 0 and len(prior_set) > 0:
+        print(f"Adding the LoRA(s) named {added_set} to the model...")
+        for lora in added_set:
+            shared.model.load_adapter(Path(f"{shared.args.lora_dir}/{lora}"), lora)
+        return
+
+    # If removing anything, disable all and re-add.
+    if len(removed_set) > 0:
+        shared.model.disable_adapter()
+
+    if len(lora_names) > 0:
+        print("Applying the following LoRAs to {}: {}".format(shared.model_name, ', '.join(lora_names)))
         params = {}
         if not shared.args.cpu:
             params['dtype'] = shared.model.dtype
@@ -25,7 +37,11 @@ def add_lora_to_model(lora_name):
             elif shared.args.load_in_8bit:
                 params['device_map'] = {'': 0}
 
-        shared.model = PeftModel.from_pretrained(shared.model, Path(f"{shared.args.lora_dir}/{lora_name}"), **params)
+        shared.model = PeftModel.from_pretrained(shared.model, Path(f"{shared.args.lora_dir}/{lora_names[0]}"), **params)
+
+        for lora in lora_names[1:]:
+            shared.model.load_adapter(Path(f"{shared.args.lora_dir}/{lora}"), lora)
+
         if not shared.args.load_in_8bit and not shared.args.cpu:
             shared.model.half()
             if not hasattr(shared.model, "hf_device_map"):

modules/shared.py

@@ -6,7 +6,7 @@ import yaml
 model = None
 tokenizer = None
 model_name = "None"
-lora_name = "None"
+lora_names = []
 soft_prompt_tensor = None
 soft_prompt = False
 is_RWKV = False
@@ -68,7 +68,7 @@ settings = {
     },
     'lora_prompts': {
         'default': 'QA',
-        '.*(alpaca-lora-7b|alpaca-lora-13b|alpaca-lora-30b)': "Alpaca",
+        '.*alpaca': "Alpaca",
     }
 }
 

requirements.txt

@@ -5,13 +5,13 @@ gradio==3.24.1
 markdown
 numpy
 Pillow>=9.5.0
-peft==0.2.0
 requests
 rwkv==0.7.3
 safetensors==0.3.0
 sentencepiece
 pyyaml
 tqdm
+git+https://github.com/huggingface/peft
 transformers==4.28.0
 bitsandbytes==0.38.1; platform_system != "Windows"
 llama-cpp-python==0.1.33; platform_system != "Windows"

server.py

@@ -88,9 +88,10 @@ def load_model_wrapper(selected_model):
         yield traceback.format_exc()
 
 
-def load_lora_wrapper(selected_lora):
-    add_lora_to_model(selected_lora)
-    return selected_lora
+def load_lora_wrapper(selected_loras):
+    yield ("Applying the following LoRAs to {}:\n\n{}".format(shared.model_name, '\n'.join(selected_loras)))
+    add_lora_to_model(selected_loras)
+    yield ("Successfuly applied the LoRAs")
 
 
 def load_preset_values(preset_menu, state, return_dict=False):
@@ -275,10 +276,12 @@ def create_model_menus():
 
                 with gr.Column():
                     with gr.Row():
-                        shared.gradio['lora_menu'] = gr.Dropdown(choices=get_available_loras(), value=shared.lora_name, label='LoRA')
-                        ui.create_refresh_button(shared.gradio['lora_menu'], lambda: None, lambda: {'choices': get_available_loras()}, 'refresh-button')
+                        shared.gradio['lora_menu'] = gr.Dropdown(multiselect=True, choices=get_available_loras(), value=shared.lora_names, label='LoRA(s)')
+                        ui.create_refresh_button(shared.gradio['lora_menu'], lambda: None, lambda: {'choices': get_available_loras(), 'value': shared.lora_names}, 'refresh-button')
 
         with gr.Column():
+            shared.gradio['lora_menu_apply'] = gr.Button(value='Apply the selected LoRAs')
+            with gr.Row():
                 unload = gr.Button("Unload the model")
                 reload = gr.Button("Reload the model")
 
@@ -338,7 +341,7 @@ def create_model_menus():
         update_model_parameters, shared.gradio['interface_state'], None).then(
         load_model_wrapper, shared.gradio['model_menu'], shared.gradio['model_status'], show_progress=True)
 
-    shared.gradio['lora_menu'].change(load_lora_wrapper, shared.gradio['lora_menu'], shared.gradio['lora_menu'], show_progress=True)
+    shared.gradio['lora_menu_apply'].click(load_lora_wrapper, shared.gradio['lora_menu'], shared.gradio['model_status'], show_progress=False)
     shared.gradio['download_button'].click(download_model_wrapper, shared.gradio['custom_model_menu'], shared.gradio['model_status'], show_progress=False)
 
 
@@ -428,8 +431,8 @@ def create_interface():
     # Defining some variables
     gen_events = []
     default_preset = shared.settings['presets'][next((k for k in shared.settings['presets'] if re.match(k.lower(), shared.model_name.lower())), 'default')]
-    if shared.lora_name != "None":
-        default_text = load_prompt(shared.settings['lora_prompts'][next((k for k in shared.settings['lora_prompts'] if re.match(k.lower(), shared.lora_name.lower())), 'default')])
+    if len(shared.lora_names) == 1:
+        default_text = load_prompt(shared.settings['lora_prompts'][next((k for k in shared.settings['lora_prompts'] if re.match(k.lower(), shared.lora_names[0].lower())), 'default')])
     else:
         default_text = load_prompt(shared.settings['prompts'][next((k for k in shared.settings['prompts'] if re.match(k.lower(), shared.model_name.lower())), 'default')])
     title = 'Text generation web UI'
@@ -861,7 +864,7 @@ if __name__ == "__main__":
         # Load the model
         shared.model, shared.tokenizer = load_model(shared.model_name)
         if shared.args.lora:
-            add_lora_to_model(shared.args.lora)
+            add_lora_to_model([shared.args.lora])
 
     # Launch the web UI
     create_interface()