File size: 1,735 Bytes
5fa76ab
 
 
63b4fe7
5fa76ab
 
 
 
d1a0a0d
5fa76ab
 
 
 
 
79c2343
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5fa76ab
 
63b4fe7
79c2343
 
 
 
 
 
 
 
 
 
 
 
5fa76ab
63b4fe7
79c2343
 
 
 
 
 
63b4fe7
79c2343
5fa76ab
 
 
79c2343
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
from fastapi import FastAPI
from pydantic import BaseModel
from huggingface_hub import InferenceClient
from fastapi.responses import StreamingResponse
import uvicorn

app = FastAPI()

client = InferenceClient("HuggingFaceH4/zephyr-7b-beta")

class Item(BaseModel):
    prompt: str
    history: list
    system_prompt: str
    temperature: float = 0.0
    max_new_tokens: int = 1048
    top_p: float = 0.15
    repetition_penalty: float = 1.0

def format_prompt(message, history):
    print("````")
    print(message)
    print("++++")
    print(history)
    print("````")
    prompt = "<s>"
    for user_prompt, bot_response in history:
        prompt += f"[INST] {user_prompt} [/INST]"
        prompt += f" {bot_response}</s> "
    prompt += f"[INST] {message} [/INST]"
    return prompt

async def generate_stream(item: Item):
    temperature = float(item.temperature)
    if temperature < 1e-2:
        temperature = 1e-2
    top_p = float(item.top_p)

    generate_kwargs = dict(
        temperature=temperature,
        max_new_tokens=item.max_new_tokens,
        top_p=top_p,
        repetition_penalty=item.repetition_penalty,
        do_sample=True,
        seed=42,
    )

    formatted_prompt = format_prompt(f"{item.system_prompt} [/INST] Ok..! </s> [INST] {item.prompt}", item.history)
    print(formatted_prompt)
    print("=======")
    print(item.history)
    stream = client.text_generation(formatted_prompt, **generate_kwargs, stream=True, details=True, return_full_text=False)

    for response in stream:
        yield response.token.text  # Stream each token as it's received

@app.post("/generate/")
async def generate_text(item: Item):
    return StreamingResponse(generate_stream(item), media_type="text/plain")