Add support for presets

README.md

@@ -63,6 +63,10 @@ If I get enough ⭐s on this repository, I will make the process of loading mode
 
 Then browse to `http://localhost:7860/?__theme=dark`
 
+## Presets
+
+Inference settings presets can be created under `presets/` as text files. These files are detected automatically at startup.
+
 ## Contributing
 
 Pull requests are welcome.

presets/Default.txt

@@ -0,0 +1,5 @@
+do_sample=True,
+max_new_tokens=max_length,
+top_p=1,
+typical_p=0.3,
+temperature=temperature, 

presets/Verbose.txt

@@ -0,0 +1,10 @@
+num_beams=10,
+min_length=max_length,
+max_new_tokens=max_length,
+length_penalty =1.4,
+no_repeat_ngram_size=2,
+early_stopping=True,
+temperature=0.7,
+top_k=150,
+top_p=0.92,
+repetition_penalty=4.5,

server.py

@@ -1,5 +1,6 @@
-import time
 import re
+import time
+import glob
 import torch
 import gradio as gr
 import transformers
@@ -16,6 +17,8 @@ model_name = 'galactica-6.7b'
 #model_name = 'flan-t5'
 #model_name = 'OPT-13B-Erebus'
 
+settings_name = "Default"
+
 def load_model(model_name):
     print(f"Loading {model_name}")
 
@@ -48,7 +51,7 @@ def fix_gpt4chan(s):
     return s
 
 def fn(question, temperature, max_length, inference_settings, selected_model):
-    global model, tokenizer, model_name
+    global model, tokenizer, model_name, settings_name
 
     if selected_model != model_name:
         model_name = selected_model
@@ -56,35 +59,17 @@ def fn(question, temperature, max_length, inference_settings, selected_model):
         tokenier = None
         torch.cuda.empty_cache()
         model, tokenizer = load_model(model_name)
+    if inference_settings != settings_name:
+        with open(f'presets/{inference_settings}.txt', 'r') as infile:
+            preset = infile.read()
+        settings_name = inference_settings
 
     torch.cuda.empty_cache()
     input_text = question
     input_ids = tokenizer.encode(str(input_text), return_tensors='pt').cuda()
 
-    if inference_settings == 'Default':
-        output = model.generate(
-            input_ids,
-            do_sample=True,
-            max_new_tokens=max_length,
-            #max_length=max_length+len(input_ids[0]),
-            top_p=1,
-            typical_p=0.3,
-            temperature=temperature, 
-        ).cuda()
-    elif inference_settings == 'Verbose':
-        output = model.generate(
-            input_ids,
-            num_beams=10,
-            min_length=max_length,
-            max_new_tokens=max_length,
-            length_penalty =1.4,
-            no_repeat_ngram_size=2,
-            early_stopping=True,
-            temperature=0.7,
-            top_k=150,
-            top_p=0.92,
-            repetition_penalty=4.5,
-        ).cuda()
+
+    output = eval(f"model.generate(input_ids, {preset}).cuda()")
 
     reply = tokenizer.decode(output[0], skip_special_tokens=True)
     if model_name.startswith('gpt4chan'):
@@ -104,7 +89,7 @@ interface = gr.Interface(
         gr.Textbox(value=default_text, lines=15),
         gr.Slider(minimum=0.0, maximum=1.0, step=0.01, label='Temperature', value=0.7),
         gr.Slider(minimum=1, maximum=2000, step=1, label='max_length', value=200),
-        gr.Dropdown(choices=["Default", "Verbose"], value="Default"),
+        gr.Dropdown(choices=list(map(lambda x : x.split('/')[-1].split('.')[0], glob.glob("presets/*.txt"))), value="Default"),
         gr.Dropdown(choices=["gpt4chan_model_float16", "galactica-6.7b", "opt-6.7b",  "opt-13b", "gpt-neox-20b", "gpt-j-6B-float16", "flan-t5", "bloomz-7b1-p3", "OPT-13B-Erebus"], value=model_name),
     ],
     outputs=[