venv uvicorn python 虚拟服务器外网无法访问

python -m venv .venv
source ./.venv/bin/activate
pip install -r requirements.txt
./run.sh

source ./.venv/bin/activate
uvicorn main:app --reload

虚拟web服务器外网访问控制台启动命令用以下代码启动
uvicorn main:app --host 0.0.0.0 --port 8501 --reload
启动到后台

nohup uvicorn main:app --host 0.0.0.0 --port 8501 --reload &

main.py

import xmltodict
import anthropic
from fastapi import FastAPI, Request
from fastapi.responses import JSONResponse, StreamingResponse

from prompt_constructors import *

from claude import ClaudeLlm  # claude.py から ClaudeLlm クラスをインポート
import os
import base64

app = FastAPI()
api_key = "your key"
client = anthropic.Anthropic(api_key=api_key)

# @app.middleware("http")
# async def log_body(request: Request, call_next):
#     body = await request.body()
#     print("HTTP REQUEST BODY: ", body)
#     return await call_next(request)


# The anthropic API does not have a method to list models, so we are hard coding the models here
@app.get("/models")
async def list_models() -> JSONResponse:
    return JSONResponse(content={"data": [
        {"id": "claude-3-sonnet-20240229", "name": "Anthropic Claude 3 Sonnet"},
        {"id": "anthropic.claude-3-sonnet-20240229-v1:0", "name": "AWS Bedrock Anthropic Claude 3 Sonnet"},
        {"id": "claude-3-opus-20240229", "name": "Anthropic Claude 3 Opus"},
    ]
    })


def map_req(req: dict) -> dict:
    messages = req["messages"]
    mapped_req = {
        "messages": messages,
    }
    return mapped_req


@app.post("/chat/completions")
async def completions(request: Request) -> StreamingResponse:
    data = await request.body()
    req = map_req(json.loads(data))
    messages=req["messages"]
    claude = ClaudeLlm(client, messages)
    resp = claude.generate_responses("claude-3-opus-20240229")
    return StreamingResponse(resp, media_type="application/x-ndjson")


def map_resp(response) -> str:
    data = json.loads(response)
    finish_reason = None
    parsed_tool_calls = []

    for message in data["content"]:
        if 'text' in message.keys() and message["text"].startswith("<function_calls>"):
            xml_tool_calls = message["text"] + "</function_calls>"
            tool_calls = xmltodict.parse(xml_tool_calls)
            if tool_calls["function_calls"]["invoke"] is list:
                for key, value in tool_calls["function_calls"]["invoke"].items():
                    parsed_tool_calls.append({
                        "index": 0,
                        "id": value['tool_name'],
                        "type": "function",
                        "function": {
                            "name": value["tool_name"],
                            "arguments": str(value["parameters"]),
                        },
                    })
            else:
                parsed_tool_calls.append({
                    "index": 0,
                    "id": tool_calls["function_calls"]["invoke"]["tool_name"],
                    "type": "function",
                    "function": {
                        "name": tool_calls["function_calls"]["invoke"]["tool_name"],
                        "arguments": json.dumps(tool_calls["function_calls"]["invoke"]["parameters"]),
                    },
                })

            message.pop("text", None)
            message.pop("type", None)
            message["tool_calls"] = parsed_tool_calls
            message["content"] = None
            message["role"] = "assistant"

        if 'text' in message.keys():
            message["content"] = message["text"]

    if "stop_reason" in data.keys() and data["stop_reason"] == "stop_sequence":
        finish_reason = "tool_calls"

    if "stop_reason" in data.keys() and data["stop_reason"] == "end_turn":
        finish_reason = "stop"

    translated = {
        "id": data["id"],
        "object": "chat.completion.chunk",
        "created": 0,
        "model": data["model"],
        "system_fingerprint": "TEMP",
        "choices": [
            {
                "index": 0,
                "delta": data["content"][0],
            },
        ],
        "finish_reason": finish_reason,
    }

    return json.dumps(translated)

claude.py

import pandas as pd
from forex_python.converter import CurrencyRates
import time
import anthropic
from anthropic.types.message_stream_event import MessageStartEvent, MessageDeltaEvent, ContentBlockDeltaEvent


class ClaudeLlm:
    def __init__(self, client, user_input):
        self.client = client
        self.user_input = user_input
        self.cost_df = pd.DataFrame(columns=["Model", "Input Tokens", "Output Tokens", "Input Cost", "Output Cost", "Total Cost", "総計_円換算", "処理時間"])

    def convert_usd_to_jpy(self, usd_amount):
        c = CurrencyRates()
        try:
            rate = c.get_rate('USD', 'JPY')
            jpy_rate = (f"為替レート: {rate:.2f}円/ドル")
            return usd_amount * rate, jpy_rate
        except Exception as e:
            rate = 150  # フォールバックとして使用する為替レート
            jpy_rate = (f"為替レート: {rate:.2f}円/ドル想定")
            return usd_amount * rate, jpy_rate

    def calculate_cost(self, model, input_tokens, output_tokens):
        token_costs = {
            "claude-3-opus-20240229": {"input": 0.000015, "output": 0.000075},
            "claude-3-sonnet-20240229": {"input": 0.000003, "output": 0.000015},
        }
        model_costs = token_costs[model]
        input_cost = input_tokens * model_costs["input"]
        output_cost = output_tokens * model_costs["output"]
        total_cost = input_cost + output_cost
        return input_cost, output_cost, total_cost

    def generate_responses(self, model_name):
        start_time = time.time()
        input_tokens = 0
        output_tokens = 0
        try:
            with self.client.messages.stream(
                model=model_name,
                max_tokens=1024,
                messages=[{"role": "user", "content": self.user_input}],
            ) as stream:
                for event in stream:
                    if isinstance(event, MessageStartEvent):
                        usage_info = event.message.usage
                        input_tokens = usage_info.input_tokens
                    elif isinstance(event, MessageDeltaEvent):
                        output_tokens = event.usage.output_tokens
                    elif isinstance(event, ContentBlockDeltaEvent):
                        return_text = event.delta.text
                        yield return_text
        except anthropic.APIStatusError as e:
            error_response = e.response.json()
            if 'error' in error_response and error_response['error'].get('type') == 'overloaded_error':
                return "APIが過負荷状態です。しばらくしてから再試行してください。"

        input_cost, output_cost, total_cost = self.calculate_cost(model_name, input_tokens, output_tokens)
        jpy_total_cost, _ = self.convert_usd_to_jpy(total_cost)
        end_time = time.time()
        response_time = end_time - start_time
        new_row = {
            "Model": model_name,
            "Input Tokens": input_tokens,
            "Output Tokens": output_tokens,
            "Input Cost": f"${input_cost:.6f}",
            "Output Cost": f"${output_cost:.6f}",
            "Total Cost": f"${total_cost:.6f}",
            "総計_円換算": f"¥{jpy_total_cost:.3f}",
            "処理時間": f"{response_time:.2f}秒"
        }
        new_row_df = pd.DataFrame([new_row])
        self.cost_df = pd.concat([self.cost_df, new_row_df], ignore_index=True)
        return self.cost_df

curl -H “Content-Type: application/json” -X POST -d ‘{“user_id”: “123”, “coin”:100, “success”:1, “msg”:“OK!” }’ “http://192.168.0.1:8001/test”

netstat -ntulp
lsof -i:8000