|
import os |
|
from fastapi import FastAPI, HTTPException |
|
from pydantic import BaseModel |
|
from huggingface_hub import InferenceClient |
|
import uvicorn |
|
|
|
app = FastAPI() |
|
|
|
|
|
primary = "mistralai/Mixtral-8x7B-Instruct-v0.1" |
|
fallbacks = ["mistralai/Mistral-7B-Instruct-v0.2", "mistralai/Mistral-7B-Instruct-v0.1"] |
|
|
|
|
|
class Item(BaseModel): |
|
input: str = None |
|
system_prompt: str = None |
|
system_output: str = None |
|
history: list = None |
|
templates: list = None |
|
temperature: float = 0.0 |
|
max_new_tokens: int = 1048 |
|
top_p: float = 0.15 |
|
repetition_penalty: float = 1.0 |
|
key: str = None |
|
|
|
|
|
def generate_response_json(item, output, tokens, model_name): |
|
return { |
|
"settings": { |
|
"input": item.input if item.input is not None else "", |
|
"system prompt": item.system_prompt if item.system_prompt is not None else "", |
|
"system output": item.system_output if item.system_output is not None else "", |
|
"temperature": f"{item.temperature}" if item.temperature is not None else "", |
|
"max new tokens": f"{item.max_new_tokens}" if item.max_new_tokens is not None else "", |
|
"top p": f"{item.top_p}" if item.top_p is not None else "", |
|
"repetition penalty": f"{item.repetition_penalty}" if item.repetition_penalty is not None else "", |
|
"do sample": "True", |
|
"seed": "42" |
|
}, |
|
"response": { |
|
"output": output.strip().lstrip('\n').rstrip('\n').lstrip('<s>').rstrip('</s>').strip(), |
|
"unstripped": output, |
|
"tokens": tokens, |
|
"model": "primary" if model_name == primary else "fallback", |
|
"name": model_name |
|
} |
|
} |
|
|
|
|
|
@app.post("/") |
|
async def generate_text(item: Item = None): |
|
try: |
|
if item is None: |
|
raise HTTPException(status_code=400, detail="JSON body is required.") |
|
|
|
if item.input is None and item.system_prompt is None or item.input == "" and item.system_prompt == "": |
|
raise HTTPException(status_code=400, detail="Parameter `input` or `system prompt` is required.") |
|
|
|
input_ = "" |
|
if item.system_prompt != None and item.system_output != None: |
|
input_ = f"<s>[INST] {item.system_prompt} [/INST] {item.system_output}</s>" |
|
elif item.system_prompt != None: |
|
input_ = f"<s>[INST] {item.system_prompt} [/INST]</s>" |
|
elif item.system_output != None: |
|
input_ = f"<s>{item.system_output}</s>" |
|
|
|
if item.templates != None: |
|
for num, template in enumerate(item.templates, start=1): |
|
input_ += f"\n<s>[INST] Beginning of archived conversation {num} [/INST]</s>" |
|
for i in range(0, len(template), 2): |
|
input_ += f"\n<s>[INST] {template[i]} [/INST]" |
|
input_ += f"\n{template[i + 1]}</s>" |
|
input_ += f"\n<s>[INST] End of archived conversation {num} [/INST]</s>" |
|
|
|
input_ += f"\n<s>[INST] Beginning of active conversation [/INST]</s>" |
|
if item.history != None: |
|
for input_, output_ in item.history: |
|
input_ += f"\n<s>[INST] {input_} [/INST]" |
|
input_ += f"\n{output_}" |
|
input_ += f"\n<s>[INST] {item.input} [/INST]" |
|
|
|
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, |
|
) |
|
|
|
tokens = 0 |
|
client = InferenceClient(primary) |
|
stream = client.text_generation(input_, **generate_kwargs, stream=True, details=True, return_full_text=True) |
|
output = "" |
|
for response in stream: |
|
tokens += 1 |
|
output += response.token.text |
|
return generate_response_json(item, output, tokens, primary) |
|
|
|
except HTTPException as http_error: |
|
raise http_error |
|
|
|
except Exception as e: |
|
tokens = 0 |
|
error = "" |
|
|
|
for model in fallbacks: |
|
try: |
|
client = InferenceClient(model) |
|
stream = client.text_generation(input_, **generate_kwargs, stream=True, details=True, return_full_text=True) |
|
output = "" |
|
for response in stream: |
|
tokens += 1 |
|
output += response.token.text |
|
return generate_response_json(item, output, tokens, model) |
|
|
|
except Exception as e: |
|
error = f"All models failed. {e}" if e else "All models failed." |
|
continue |
|
|
|
raise HTTPException(status_code=500, detail=error) |
|
|
|
if "KEY" in os.environ: |
|
if item.key != os.environ["KEY"]: |
|
raise HTTPException(status_code=401, detail="Valid key is required.") |
|
|
|
if __name__ == "__main__": |
|
uvicorn.run(app, host="0.0.0.0", port=8000) |