Commit
·
bd0305e
1
Parent(s):
b058a81
Create app.py
Browse files
app.py
ADDED
|
@@ -0,0 +1,99 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import gradio as gr
|
| 2 |
+
import torch
|
| 3 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline, StoppingCriteria, StoppingCriteriaList
|
| 4 |
+
import time
|
| 5 |
+
import numpy as np
|
| 6 |
+
from torch.nn import functional as F
|
| 7 |
+
|
| 8 |
+
|
| 9 |
+
m = AutoModelForCausalLM.from_pretrained("/mnt/nvme/home/dakota/ckpts/stablelm/7B-sft-combined/checkpoint-8000", torch_dtype=torch.float16).cuda()
|
| 10 |
+
tok = AutoTokenizer.from_pretrained("/mnt/nvme/home/dakota/stablelm_tokenizer")
|
| 11 |
+
generator = pipeline('text-generation', model=m, tokenizer=tok, device=0)
|
| 12 |
+
|
| 13 |
+
|
| 14 |
+
start_message = """<|SYSTEM|># StableAssistant
|
| 15 |
+
- StableAssistant is A helpful and harmless Open Source AI Language Model developed by Stability and CarperAI.
|
| 16 |
+
- StableAssistant is excited to be able to help the user, but will refuse to do anything that could be considered harmful to the user.
|
| 17 |
+
- StableAssistant is more than just an information source, StableAssistant is also able to write poetry, short stories, and make jokes.
|
| 18 |
+
- StableAssistant will refuse to participate in anything that could harm a human."""
|
| 19 |
+
|
| 20 |
+
|
| 21 |
+
class StopOnTokens(StoppingCriteria):
|
| 22 |
+
def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
|
| 23 |
+
stop_ids = [50278, 50279, 50277, 1, 0]
|
| 24 |
+
for stop_id in stop_ids:
|
| 25 |
+
if input_ids[0][-1] == stop_id:
|
| 26 |
+
return True
|
| 27 |
+
return False
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
def contrastive_generate(text, bad_text):
|
| 31 |
+
with torch.no_grad():
|
| 32 |
+
tokens = tok(text, return_tensors="pt")['input_ids'].cuda()[:,:4096-1024]
|
| 33 |
+
bad_tokens = tok(bad_text, return_tensors="pt")['input_ids'].cuda()[:,:4096-1024]
|
| 34 |
+
history = None
|
| 35 |
+
bad_history = None
|
| 36 |
+
curr_output = list()
|
| 37 |
+
for i in range(1024):
|
| 38 |
+
out = m(tokens, past_key_values=history, use_cache=True)
|
| 39 |
+
logits = out.logits
|
| 40 |
+
history = out.past_key_values
|
| 41 |
+
bad_out = m(bad_tokens, past_key_values=bad_history, use_cache=True)
|
| 42 |
+
bad_logits = bad_out.logits
|
| 43 |
+
bad_history = bad_out.past_key_values
|
| 44 |
+
probs = F.softmax(logits.float(), dim=-1)[0][-1].cpu()
|
| 45 |
+
bad_probs = F.softmax(bad_logits.float(), dim=-1)[0][-1].cpu()
|
| 46 |
+
logits = torch.log(probs)
|
| 47 |
+
bad_logits = torch.log(bad_probs)
|
| 48 |
+
logits[probs > 0.1] = logits[probs > 0.1] - bad_logits[probs > 0.1]
|
| 49 |
+
probs = F.softmax(logits)
|
| 50 |
+
out = int(torch.multinomial(probs, 1))
|
| 51 |
+
if out in [50278, 50279, 50277, 1, 0]:
|
| 52 |
+
break
|
| 53 |
+
else:
|
| 54 |
+
curr_output.append(out)
|
| 55 |
+
out = np.array([out])
|
| 56 |
+
tokens = torch.from_numpy(np.array([out])).to(
|
| 57 |
+
tokens.device)
|
| 58 |
+
bad_tokens = torch.from_numpy(np.array([out])).to(
|
| 59 |
+
tokens.device)
|
| 60 |
+
return tok.decode(curr_output)
|
| 61 |
+
|
| 62 |
+
def generate(text, bad_text=None):
|
| 63 |
+
stop = StopOnTokens()
|
| 64 |
+
result = generator(text, max_new_tokens=1024, num_return_sequences=1, num_beams=1, do_sample=True, temperature=1.0, top_p=0.95, top_k=1000, stopping_criteria=StoppingCriteriaList([stop]))
|
| 65 |
+
return result[0]["generated_text"].replace(text, "")
|
| 66 |
+
|
| 67 |
+
|
| 68 |
+
def user(user_message, history):
|
| 69 |
+
return "", history + [[user_message, ""]]
|
| 70 |
+
|
| 71 |
+
|
| 72 |
+
def bot(history, curr_system_message):
|
| 73 |
+
messages = curr_system_message + "".join(["".join(["<|USER|>"+item[0], "<|ASSISTANT|>"+item[1]]) for item in history])
|
| 74 |
+
output = generate(messages)
|
| 75 |
+
history[-1][1] = output
|
| 76 |
+
time.sleep(1)
|
| 77 |
+
return history
|
| 78 |
+
|
| 79 |
+
|
| 80 |
+
def system_update(msg):
|
| 81 |
+
global curr_system_message
|
| 82 |
+
curr_system_message = msg
|
| 83 |
+
|
| 84 |
+
|
| 85 |
+
with gr.Blocks() as demo:
|
| 86 |
+
with gr.Row():
|
| 87 |
+
with gr.Column():
|
| 88 |
+
chatbot = gr.Chatbot([])
|
| 89 |
+
clear = gr.Button("Clear")
|
| 90 |
+
with gr.Column():
|
| 91 |
+
system_msg = gr.Textbox(start_message, label="System Message", interactive=True)
|
| 92 |
+
msg = gr.Textbox(label="Chat Message")
|
| 93 |
+
|
| 94 |
+
msg.submit(user, [msg, chatbot], [msg, chatbot], queue=False).then(
|
| 95 |
+
bot, [chatbot, system_msg], chatbot
|
| 96 |
+
)
|
| 97 |
+
system_msg.change(system_update, system_msg, None, queue=False)
|
| 98 |
+
clear.click(lambda: None, None, chatbot, queue=False)
|
| 99 |
+
demo.launch(share=True)
|