actllminfer 0.2.0

Integrated LLM inference engine with a LangChain-Core-style interface across Kimi, GLM, MiniMax, DeepSeek, OpenAI, Anthropic, Hugging Face, and NVIDIA NIM.

ActLLMInfer

Integrated LLM inference engine with a LangChain-Core-style interface. One small package, one consistent API across the major Chinese and US chat-LLM providers — designed to be the inference layer for the actdecor package.

Supported providers

Provider	Class	Default model	API key env var
Moonshot (Kimi)	ChatMoonshot / ChatKimi	moonshot-v1-8k	MOONSHOT_API_KEY
ZhipuAI (GLM)	ChatZhipuAI / ChatGLM	glm-4-plus	ZHIPUAI_API_KEY
MiniMax	ChatMiniMax	abab6.5s-chat	MINIMAX_API_KEY
DeepSeek	ChatDeepSeek	deepseek-chat	DEEPSEEK_API_KEY
OpenAI	ChatOpenAI	gpt-4o-mini	OPENAI_API_KEY
Anthropic	ChatAnthropic	claude-sonnet-4-6	ANTHROPIC_API_KEY
Hugging Face	ChatHuggingFace	Hythcliff/canadian-address-checker-on	HF_TOKEN (or HUGGINGFACEHUB_API_TOKEN)
NVIDIA NIM	ChatNVIDIA	meta/llama-3.3-70b-instruct	NVIDIA_API_KEY (or NGC_API_KEY)

Install

pip install -e .

The only required dependency is requests. httpx is optional (for users who want async transports later).

Quick start

Direct invocation

from actllminfer import ChatKimi, HumanMessage, SystemMessage

llm = ChatKimi(model="moonshot-v1-8k", temperature=0.2)
reply = llm.invoke([
    SystemMessage(content="You are a concise assistant."),
    HumanMessage(content="Summarize the theory of relativity in one sentence."),
])
print(reply.content)

Unified completion() endpoint (LiteLLM-style)

For callers that just want a single function with an OpenAI-shaped response, completion() dispatches across every supported provider via a provider/model string. Providers can also be addressed with the legacy provider:model separator, and known prefixes (glm-, deepseek-, claude-, nemotron, …) are inferred automatically.

from actllminfer import completion

resp = completion(
    model="openai/gpt-4o-mini",
    messages=[{"role": "user", "content": "Summarize relativity in one line."}],
    temperature=0.2,
)
print(resp.choices[0].message.content)   # attribute access
print(resp["choices"][0]["message"]["content"])  # dict access
print(resp.usage["total_tokens"])

stream=True returns an iterator of OpenAI-shaped chat.completion.chunk objects:

for chunk in completion(model="kimi/moonshot-v1-8k", messages="Tell me a joke", stream=True):
    delta = chunk.choices[0].delta
    if delta.get("content"):
        print(delta.content, end="", flush=True)

Per-call generation knobs (temperature, max_tokens, tools, response_format, seed, stop, …) are forwarded to the provider. Constructor-time arguments (api_key, base_url, organization, request_timeout, default_headers, extra_body, …) build (and cache) the underlying client.

An embedding() counterpart and async variants (acompletion, aembedding) are provided too:

from actllminfer import embedding, acompletion

vecs = embedding(model="openai/text-embedding-3-small", input=["hi", "there"])
print(vecs.data[0].embedding[:4])

resp = await acompletion(model="anthropic/claude-sonnet-4-6", messages="Hi")

A primary-with-fallbacks Router retries on transient errors so a rate-limited primary transparently falls through to the next provider:

from actllminfer import Router

router = Router([
    "openai/gpt-4o-mini",
    "kimi/moonshot-v1-8k",
    "deepseek/deepseek-chat",
])
resp = router.completion(messages=[{"role": "user", "content": "Hi"}])

String spec via the factory

from actllminfer import init_chat_model

llm = init_chat_model("kimi:moonshot-v1-8k", temperature=0)
llm = init_chat_model("glm-4-plus")                 # provider inferred
llm = init_chat_model("deepseek-reasoner")          # provider inferred
llm = init_chat_model("anthropic:claude-sonnet-4-6")
llm = init_chat_model("abab6.5s-chat")              # MiniMax inferred
llm = init_chat_model("hf:meta-llama/Llama-3.3-70B-Instruct")
llm = init_chat_model("nvidia:meta/llama-3.3-70b-instruct")

Hugging Face

ChatHuggingFace defaults to the HF Inference Router (https://router.huggingface.co/v1/chat/completions), which is OpenAI-compatible and dispatches to whichever provider currently serves the model id you pass.

from actllminfer import ChatHuggingFace

llm = ChatHuggingFace(model="Qwen/Qwen2.5-72B-Instruct")
print(llm.invoke("Summarize the theory of relativity in one sentence.").content)

For a dedicated Inference Endpoint, a self-hosted TGI server, or any other OpenAI-compatible deployment, just point base_url at the /v1 root:

llm = ChatHuggingFace(
    model="tgi",  # placeholder; the endpoint already targets a single model
    base_url="https://my-endpoint.example.com/v1",
)

The class accepts HF_TOKEN or the older HUGGINGFACEHUB_API_TOKEN env var.

A worked example using Hythcliff/canadian-address-checker-on to validate a batch of Canadian addresses and parse a structured JSON response is in examples/canadian_address_checker.py.

NVIDIA NIM (free serverless inference)

ChatNVIDIA targets NVIDIA's free OpenAI-compatible NIM endpoint at https://integrate.api.nvidia.com/v1/chat/completions. Grab a free nvapi-... key from build.nvidia.com and set NVIDIA_API_KEY (the legacy NGC_API_KEY is also accepted).

from actllminfer import ChatNVIDIA

llm = ChatNVIDIA(model="meta/llama-3.3-70b-instruct", temperature=0.2)
print(llm.invoke("Summarize the theory of relativity in one sentence.").content)

The same key fans out to dozens of hosted models — Llama 3.x, Mixtral, Nemotron, Qwen, DeepSeek, Phi, Gemma, etc. To target a self-hosted NIM microservice instead, point base_url at any OpenAI-compatible /v1 root:

llm = ChatNVIDIA(
    model="meta/llama-3.1-8b-instruct",
    base_url="https://my-nim.example.com/v1",
)

Composable chains (LCEL-style)

from actllminfer import ChatPromptTemplate, StrOutputParser, init_chat_model

llm = init_chat_model("kimi:moonshot-v1-8k")
prompt = ChatPromptTemplate.from_messages([
    ("system", "You are a helpful translator."),
    ("user", "Translate to {language}: {text}"),
])
chain = prompt | llm | StrOutputParser()

print(chain.invoke({"language": "French", "text": "Good morning"}))

Streaming

for chunk in llm.stream("Tell me a short story about a robot."):
    print(chunk.text, end="", flush=True)

Tool / function calling (OpenAI-shaped)

tools = [{
    "type": "function",
    "function": {
        "name": "get_weather",
        "description": "Get current weather for a city.",
        "parameters": {
            "type": "object",
            "properties": {"city": {"type": "string"}},
            "required": ["city"],
        },
    },
}]

llm_with_tools = init_chat_model("kimi:moonshot-v1-8k").with_tools(tools)
ai_msg = llm_with_tools.invoke("What's the weather in Beijing?")
for call in ai_msg.tool_calls:
    print(call.name, call.args)

The same tools=[{"type": "function", ...}] spec is automatically translated to Anthropic's input_schema shape for ChatAnthropic.

JSON output

from actllminfer import JsonOutputParser

chain = prompt | llm | JsonOutputParser()
data = chain.invoke({"language": "JSON", "text": "Return {\"ok\": true} only."})

Architecture

actllminfer/
├── messages.py         # BaseMessage, SystemMessage, HumanMessage, AIMessage, ToolMessage, ...
├── outputs.py          # ChatGeneration, ChatResult, ChatGenerationChunk
├── prompts.py          # PromptTemplate, ChatPromptTemplate, MessagesPlaceholder
├── output_parsers.py   # StrOutputParser, JsonOutputParser, CommaSeparatedListOutputParser
├── runnables.py        # Runnable, RunnableLambda, RunnablePassthrough, RunnableSequence
├── callbacks.py        # BaseCallbackHandler, CallbackManager, StdOutCallbackHandler
├── language_models/
│   └── base.py         # BaseChatModel
├── chat_models/
│   ├── _openai_compat.py   # shared OpenAI /v1/chat/completions backend
│   ├── openai.py
│   ├── moonshot.py     # Kimi
│   ├── deepseek.py
│   ├── zhipuai.py      # GLM
│   ├── minimax.py
|   ├── anthropic.py    # Claude (different transport)
│   ├── huggingface.py  # HF Inference Router / TGI / dedicated endpoints
│   └── nvidia.py       # NVIDIA NIM serverless / self-hosted
├── factory.py          # init_chat_model("kimi:moonshot-v1-8k")
└── exceptions.py

Every provider implements the same BaseChatModel contract: invoke, batch, stream, generate, with_tools, bind. That means the actdecor package can keep one code path and switch providers via configuration.

License

Apache 2.0.