Spaces:
Sleeping
Sleeping
Commit
•
ed12453
1
Parent(s):
52f79c7
Upload app.py
Browse files
app.py
ADDED
@@ -0,0 +1,72 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import time
|
2 |
+
import gradio as gr
|
3 |
+
from transformers import AutoTokenizer
|
4 |
+
import json
|
5 |
+
import requests
|
6 |
+
|
7 |
+
|
8 |
+
models = [
|
9 |
+
"TinyLLama 1b 4_K_M 2048",
|
10 |
+
"TinyLLama 1b OpenOrca 4_K_M 2048",
|
11 |
+
"OpenLLama 3b 4_K_M 196k",
|
12 |
+
"Phi-2 2.7b 4_K_M 2048",
|
13 |
+
"Stable Zephyr 3b 4_K_M 4096"
|
14 |
+
]
|
15 |
+
|
16 |
+
|
17 |
+
def make_request_to_llm(llm, prompt, max_new_tokens, nctx):
|
18 |
+
headers = {
|
19 |
+
"Authorization": f"Bearer {o.getenv('HF_TOKEN')}",
|
20 |
+
"Content-Type": "application/json"
|
21 |
+
}
|
22 |
+
|
23 |
+
body = {'prompt': prompt, 'max_new_tokens': max_new_tokens, "llm": llm, "nctx":nctx}
|
24 |
+
response = requests.post('https://daniellefranca96-cpu-inf.hf.space/llm_on_cpu', headers=headers, json=body)
|
25 |
+
return response.text
|
26 |
+
|
27 |
+
|
28 |
+
def change(llm):
|
29 |
+
return int(models[llm]['nctx']/1000)
|
30 |
+
|
31 |
+
def update(prompt, llm, nctx, max_tokens):
|
32 |
+
|
33 |
+
answer = {}
|
34 |
+
|
35 |
+
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-chat-hf")
|
36 |
+
|
37 |
+
# Tokenize the input text
|
38 |
+
tokenized_input = tokenizer.encode(prompt, return_tensors="pt")
|
39 |
+
|
40 |
+
# Measure processing time
|
41 |
+
start_time = time.time()
|
42 |
+
result = make_request_to_llm(llm, prompt, max_tokens, int(nctx)*1000)
|
43 |
+
print(result)
|
44 |
+
end_time = time.time()
|
45 |
+
|
46 |
+
# Calculate tokens per second
|
47 |
+
duration = end_time - start_time
|
48 |
+
answer['Duration'] = duration
|
49 |
+
print("Duration: "+str(duration))
|
50 |
+
tokens_per_second = len(tokenized_input) / duration
|
51 |
+
answer['Tokens Per Second'] = tokens_per_second
|
52 |
+
answer['answer'] = result
|
53 |
+
return json.dumps(answer)
|
54 |
+
|
55 |
+
with gr.Blocks() as demo:
|
56 |
+
gr.Markdown("Test LM inferences speeds on CPU.")
|
57 |
+
|
58 |
+
with gr.Row():
|
59 |
+
select = gr.Dropdown(models, label="LLM", value=models[0])
|
60 |
+
with gr.Row():
|
61 |
+
nctx = gr.Slider(minimum=1, maximum=100, label='Context (consider as 1000* the value chosen)', value="1")
|
62 |
+
with gr.Row():
|
63 |
+
max_tokens = gr.Slider(minimum=512, maximum=4096, label='Max Tokens Generated')
|
64 |
+
with gr.Row():
|
65 |
+
inp = gr.Textbox(placeholder="What is your prompt?", label="Prompt")
|
66 |
+
with gr.Row():
|
67 |
+
out = gr.Textbox(label="Output", lines=20)
|
68 |
+
btn = gr.Button("Run")
|
69 |
+
btn.click(fn=update, inputs=[inp, select, nctx, max_tokens], outputs=out)
|
70 |
+
select.change(fn=change, inputs=[select], outputs=nctx)
|
71 |
+
|
72 |
+
demo.launch()
|