开始使用 RunPod LLM。

概述

本指南介绍如何使用 LangChain RunPod LLM 类与托管在 RunPod Serverless.

设置

1. **安装包：**

   pip install -qU langchain-runpod
   

1. **部署 LLM 端点：** 按照 RunPod Provider Guide 中的设置步骤在 RunPod Serverless 上部署兼容的文本生成端点并获取其端点 ID。
2. **设置环境变量：** 确保 RUNPOD_API_KEY 和 RUNPOD_ENDPOINT_ID 已设置。

# Make sure environment variables are set (or pass them directly to RunPod)
if "RUNPOD_API_KEY" not in os.environ:
    os.environ["RUNPOD_API_KEY"] = getpass.getpass("Enter your RunPod API Key: ")
if "RUNPOD_ENDPOINT_ID" not in os.environ:
    os.environ["RUNPOD_ENDPOINT_ID"] = input("Enter your RunPod Endpoint ID: ")

实例化

初始化 RunPod 类。您可以通过 model_kwargs 传递特定于模型的参数并配置轮询行为。

from langchain_runpod import RunPod

llm = RunPod(
    # runpod_endpoint_id can be passed here if not set in env
    model_kwargs={
        "max_new_tokens": 256,
        "temperature": 0.6,
        "top_k": 50,
        # Add other parameters supported by your endpoint handler
    },
    # Optional: Adjust polling
    # poll_interval=0.3,
    # max_polling_attempts=100
)

调用

使用标准 LangChain .invoke() 和 .ainvoke() 方法调用模型。流式处理也支持通过 .stream() 和 .astream() （通过轮询 RunPod /stream 端点模拟）。

prompt = "Write a tagline for an ice cream shop on the moon."

# Invoke (Sync)
try:
    response = llm.invoke(prompt)
    print("--- Sync Invoke Response ---")
    print(response)
except Exception as e:
    print(
        f"Error invoking LLM: {e}. Ensure endpoint ID/API key are correct and endpoint is active/compatible."
    )

# Stream (Sync, simulated via polling /stream)
print("\n--- Sync Stream Response ---")
try:
    stream = llm.stream_events(prompt, version="v3")
    for token in stream.text:
        print(token, end="", flush=True)
    print()  # Newline
except Exception as e:
    print(
        f"\nError streaming LLM: {e}. Ensure endpoint handler supports streaming output format."
    )

异步用法

# AInvoke (Async)
try:
    async_response = await llm.ainvoke(prompt)
    print("--- Async Invoke Response ---")
    print(async_response)
except Exception as e:
    print(f"Error invoking LLM asynchronously: {e}.")

# AStream (Async)
print("\n--- Async Stream Response ---")
try:
    stream = await llm.astream_events(prompt, version="v3")
    async for token in stream.text:
        print(token, end="", flush=True)
    print()  # Newline
except Exception as e:
    print(
        f"\nError streaming LLM asynchronously: {e}. Ensure endpoint handler supports streaming output format."
    )

链式调用

LLM 与 LangChain 表达式语言 (LCEL) 链无缝集成。

from langchain_core.output_parsers import StrOutputParser
from langchain_core.prompts import PromptTemplate

# Assumes 'llm' variable is instantiated from the 'Instantiation' cell
prompt_template = PromptTemplate.from_template("Tell me a joke about {topic}")
parser = StrOutputParser()

chain = prompt_template | llm | parser

try:
    chain_response = chain.invoke({"topic": "bears"})
    print("--- Chain Response ---")
    print(chain_response)
except Exception as e:
    print(f"Error running chain: {e}")

# Async chain
try:
    async_chain_response = await chain.ainvoke({"topic": "robots"})
    print("--- Async Chain Response ---")
    print(async_chain_response)
except Exception as e:
    print(f"Error running async chain: {e}")

端点注意事项

• Input: 端点处理器应在 {"input": {"prompt": "...", ...}}.
• Output: 中预期提示字符串。处理器应在最终状态响应的 "output" 键中返回生成的文本（例如， {"output": "Generated text..."} or {"output": {"text": "..."}}).
• Streaming: 对于通过 /stream 端点的模拟流式处理，处理器必须在状态响应中填充 "stream" 键，使用分块字典列表，如 [{"output": "token1"}, {"output": "token2"}].

API 参考

有关 RunPod LLM 类、参数和方法的详细文档，请参阅源代码或生成的 API 参考（如果有）。

源代码链接： https://github.com/runpod/langchain-runpod/blob/main/langchain_runpod/llms.py