langchain-crynux

Drop-in replacement for langchain-openai ChatOpenAI that lets existing OpenAI-compatible LLM code run on the Crynux network without changes.

Installation

pip install langchain-crynux

Dependencies:

• langchain-openai>=1.0.1

Usage

import os
from langchain_crynux import ChatCrynux

# Option 1: environment variable (same as langchain-openai)
os.environ["OPENAI_API_KEY"] = "your-api-key"

chat = ChatCrynux(
    base_url="crynux_base_url",
    model="Qwen/Qwen2.5-7B-Instruct",
    vram_limit=24,
    timeout=60,
    # Option 2: pass api_key directly
    # api_key="your-api-key",
)

response = chat.invoke("Hello from Crynux.")
print(response.content)

• base_url is the Crynux API endpoint. There is no default.
• vram_limit is the minimum GPU VRAM (in GB) required for the inference run. Default is 24.

ChatCrynux is a single ChatModel. Call it with invoke / ainvoke / with_structured_output the same way as ChatOpenAI. LangGraph graphs that call ainvoke do not need to change; set the switches on the ChatCrynux constructor.

Responses API and background polling

Background example:

from langchain_crynux import ChatCrynux

chat = ChatCrynux(
    base_url="crynux_base_url",
    model="Qwen/Qwen2.5-7B-Instruct",
    use_responses_api=True,
    background=True,
    timeout=600,
    http_timeout=60,
    poll_interval=2.0,
)

response = chat.invoke("Hello from Crynux.")
print(response.content)

When background=True, invoke / ainvoke still return a finished AIMessage. ChatCrynux creates the Responses job, polls until completed or failed, and then returns. A failed job or a polling timeout raises an exception.

Synchronous Responses without internal polling:

chat = ChatCrynux(
    base_url="crynux_base_url",
    model="Qwen/Qwen2.5-7B-Instruct",
    use_responses_api=True,
    background=False,
)

Structured Output

You can use the with_structured_output method to get structured output from the model.

from typing import Optional
from langchain_crynux import ChatCrynux
from pydantic import BaseModel, Field

class Weather(BaseModel):
    """The weather in a specific location."""
    city: str = Field(description="The city to get the weather for")
    temperature: float = Field(description="The temperature in celsius")
    condition: Optional[str] = Field(description="The weather condition (e.g., sunny, rainy)")

chat = ChatCrynux(
    base_url="crynux_base_url",
    model="Qwen/Qwen2.5-7B-Instruct",
    temperature=0
)

structured_llm = chat.with_structured_output(Weather)

response = structured_llm.invoke("The weather in Tokyo is 25.5 degrees celsius and sunny.")
print(response)
# Weather(city='Tokyo', temperature=25.5, condition='sunny')

The same with_structured_output call works with background=True.

ChatCrynux defaults with_structured_output to method="function_calling". ChatOpenAI defaults to method="json_schema". Many models on Crynux do not support OpenAI json mode, so ChatCrynux uses function calling instead.