File size: 2,327 Bytes
f84e083
caa64e7
f84e083
 
 
215f4a9
f84e083
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
215f4a9
f84e083
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
215f4a9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f84e083
 
 
215f4a9
 
 
f84e083
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
61
62
63
64
65
66
67
68
69
70
71
72
73
from fastapi import FastAPI
from fastapi.responses import StreamingResponse
from pydantic import BaseModel
from huggingface_hub import InferenceClient
import uvicorn
import json  # Make sure to import json


app = FastAPI()

client = InferenceClient("mistralai/Mixtral-8x7B-Instruct-v0.1")

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):
    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


def generate(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}, {item.prompt}", item.history)
    stream = client.text_generation(formatted_prompt, **generate_kwargs, stream=True, details=True, return_full_text=False)

    # Initialize a variable to track whether this is the last item
    is_last = False

    # Since we're yielding JSON, each chunk must be a complete JSON object.
    # We'll iterate over the stream and yield each response as a JSON string.
    for i, response in enumerate(stream):
        # Check if this is the last item by attempting to peek ahead
        is_last = True  # Assume it's the last unless proven otherwise in the next iteration

        # Construct the chunk of data to include the text and completion status
        chunk_data = {
            "text": response.token.text,
            "complete": is_last
        }

        # Yield this chunk as a JSON-encoded string followed by a newline to separate chunks
        yield json.dumps(chunk_data) + "\n"

@app.post("/generate/")
async def generate_text(item: Item):
    # Note the change to media_type to indicate we're streaming JSON
    return StreamingResponse(generate(item), media_type="application/x-ndjson")