Update app.py

app.py

@@ -17,48 +17,60 @@ InferenceClient(models[2]),
 InferenceClient(models[3]),
 ]
 
+VERBOSE=True
 
 def format_prompt(message, history):
     prompt = ""
     if history:
         #<start_of_turn>userHow does the brain work?<end_of_turn><start_of_turn>model
         for user_prompt, bot_response in history:
-            prompt += f"<start_of_turn>user{user_prompt}<end_of_turn>"
-            prompt += f"<start_of_turn>model{bot_response}"
+            prompt += f"{user_prompt}\n"
+            #print(prompt)
+            prompt += f"{bot_response}\n"
+            #print(prompt)
     prompt += f"<start_of_turn>user{message}<end_of_turn><start_of_turn>model"
     return prompt
 
-
-
-def chat_inf(system_prompt,prompt,history,client_choice,seed,temp,tokens,top_p,rep_p):
+def chat_inf(system_prompt,prompt,history,memory,client_choice,seed,temp,tokens,top_p,rep_p,chat_mem):
     #token max=8192
+    hist_len=0
     client=clients[int(client_choice)-1]
     if not history:
         history = []
         hist_len=0
-    if history:
-        hist_len=len(history)
-        print(hist_len)
-        
-    #seed = random.randint(1,1111111111111111)
-    generate_kwargs = dict(
-        temperature=temp,
-        max_new_tokens=tokens,
-        top_p=top_p,
-        repetition_penalty=rep_p,
-        do_sample=True,
-        seed=seed,
-    )
-    #formatted_prompt=prompt   
-    formatted_prompt = format_prompt(f"{system_prompt}, {prompt}", history)
-    stream = client.text_generation(formatted_prompt, **generate_kwargs, stream=True, details=True, return_full_text=False)
-    output = ""
-        
-    for response in stream:
-        output += response.token.text
-        yield [(prompt,output)]
-    history.append((prompt,output))
-    yield history
+    if not memory:
+        memory = []
+        mem_len=0        
+    if memory:
+        for ea in memory[0-chat_mem:]:
+            hist_len+=len(str(ea))
+    in_len=len(system_prompt+prompt)+hist_len
+
+    if (in_len+tokens) > 8000:
+        history.append((prompt,"Wait, that's too many tokens, please reduce the 'Chat Memory' value, or reduce the 'Max new tokens' value"))
+        yield history,memory
+    else:
+        generate_kwargs = dict(
+            temperature=temp,
+            max_new_tokens=tokens,
+            top_p=top_p,
+            repetition_penalty=rep_p,
+            do_sample=True,
+            seed=seed,
+        )
+        formatted_prompt = format_prompt(f"{system_prompt}, {prompt}", memory[0-chat_mem:])
+        stream = client.text_generation(formatted_prompt, **generate_kwargs, stream=True, details=True, return_full_text=False)
+        output = ""
+        for response in stream:
+            output += response.token.text
+            yield [(prompt,output)],memory
+        history.append((prompt,output))
+        memory.append((prompt,output))
+        yield history,memory
+    if VERBOSE==True:
+        print("\n######### HIST "+str(in_len))
+        print("\n######### TOKENS "+str(tokens))        
+        print("\n######### PROMPT "+str(len(formatted_prompt)))
 
 def get_screenshot(chat: list,height=5000,width=600,chatblock=[],theme="light",wait=3000,header=True):
     print(chatblock)
@@ -70,20 +82,18 @@ def get_screenshot(chat: list,height=5000,width=600,chatblock=[],theme="light",w
     print(out)
     return out
 
-
-
 def clear_fn():
-    return None
+    return None,None,None
 rand_val=random.randint(1,1111111111111111)
+
 def check_rand(inp,val):
     if inp==True:
         return gr.Slider(label="Seed", minimum=1, maximum=1111111111111111, value=random.randint(1,1111111111111111))
     else:
         return gr.Slider(label="Seed", minimum=1, maximum=1111111111111111, value=int(val))
-
-
     
 with gr.Blocks() as app:
+    memory=gr.State()
     gr.HTML("""<center><h1 style='font-size:xx-large;'>Google Gemma Models</h1><br><h3>running on Huggingface Inference Client</h3><br><h7>EXPERIMENTAL""")
     chat_b = gr.Chatbot(height=500)
     with gr.Group():
@@ -99,15 +109,15 @@ with gr.Blocks() as app:
                             stop_btn=gr.Button("Stop")
                             clear_btn=gr.Button("Clear")                
                 client_choice=gr.Dropdown(label="Models",type='index',choices=[c for c in models],value=models[0],interactive=True)
-
             with gr.Column(scale=1):
                 with gr.Group():
                     rand = gr.Checkbox(label="Random Seed", value=True)
                     seed=gr.Slider(label="Seed", minimum=1, maximum=1111111111111111,step=1, value=rand_val)
-                    tokens = gr.Slider(label="Max new tokens",value=6400,minimum=0,maximum=8000,step=64,interactive=True, visible=True,info="The maximum number of tokens")
+                    tokens = gr.Slider(label="Max new tokens",value=1600,minimum=0,maximum=8000,step=64,interactive=True, visible=True,info="The maximum number of tokens")
                     temp=gr.Slider(label="Temperature",step=0.01, minimum=0.01, maximum=1.0, value=0.9)
                     top_p=gr.Slider(label="Top-P",step=0.01, minimum=0.01, maximum=1.0, value=0.9)
                     rep_p=gr.Slider(label="Repetition Penalty",step=0.1, minimum=0.1, maximum=2.0, value=1.0)
+                    chat_mem=gr.Number(label="Chat Memory", info="Number of previous chats to retain",value=4)
         with gr.Accordion(label="Screenshot",open=False):
             with gr.Row():
                 with gr.Column(scale=3):
@@ -120,10 +130,10 @@ with gr.Blocks() as app:
                     wait_time=gr.Number(label="Wait Time",value=3000)
                     theme=gr.Radio(label="Theme", choices=["light","dark"],value="light")
                     chatblock=gr.Dropdown(label="Chatblocks",info="Choose specific blocks of chat",choices=[c for c in range(1,40)],multiselect=True)
-            
+
     im_go=im_btn.click(get_screenshot,[chat_b,im_height,im_width,chatblock,theme,wait_time],img)
-    chat_sub=inp.submit(check_rand,[rand,seed],seed).then(chat_inf,[sys_inp,inp,chat_b,client_choice,seed,temp,tokens,top_p,rep_p],chat_b)
-    go=btn.click(check_rand,[rand,seed],seed).then(chat_inf,[sys_inp,inp,chat_b,client_choice,seed,temp,tokens,top_p,rep_p],chat_b)
+    chat_sub=inp.submit(check_rand,[rand,seed],seed).then(chat_inf,[sys_inp,inp,chat_b,memory,client_choice,seed,temp,tokens,top_p,rep_p,chat_mem],[chat_b,memory])
+    go=btn.click(check_rand,[rand,seed],seed).then(chat_inf,[sys_inp,inp,chat_b,memory,client_choice,seed,temp,tokens,top_p,rep_p,chat_mem],[chat_b,memory])
     stop_btn.click(None,None,None,cancels=[go,im_go,chat_sub])
-    clear_btn.click(clear_fn,None,[chat_b])
+    clear_btn.click(clear_fn,None,[inp,sys_inp,chat_b,memory])
 app.queue(default_concurrency_limit=10).launch()