Download, launch, and integrate AI models curated by Anaconda.

Anaconda provides quantization files for a curated collection of large-language-models (LLMs). This package provides programmatic access and an SDK to access the curated models, download them, and start servers.

Below you will find documentation for

• How to install
• Command line interface to list, download, run API servers for models
• Anaconda AI SDK
• Integration with LLM CLI
• Langchain
• LlamaIndex
• LiteLLM
• DSPy
• Panel ChatInterface

conda install -c anaconda-cloud anaconda-ai

The backend for anaconda-ai is Anaconda AI Navigator. This package package utilizes the backend API to list and download models and manage running servers. All activities performed by the CLI, SDK, and integrations here are visible within Anaconda AI Navigator.

Anaconda AI supports configuration management in the ~/.anaconda/config.toml file. The following parameters are supported under the table [plugin.ai] or by setting ANACONDA_AI_<parameter>=<value> environment variables.

In the CLI, SDK, and integrations below individual model quantizations are are referenced according the following scheme.

[<author>/]<model_name></ or _><quantization>[.<format>]

Fields surrounded by [] are optional. The essential elements are the model name and quantization method separated by either / or _. The supported quantization methods are

• Q4_K_M
• Q5_K_M
• Q6_K
• Q8_0

The CLI subcommands within anaconda ai provide full access to list and download model files, start and stop servers through the backend.

See the --help for each command for more details.

The SDK actions are initiated by creating a client connection to the backend.

from anaconda_ai import get_default_client

client = get_default_client()

The client provides two top-level accessors .models and .servers.

The .models attribute provides actions to list available models and download specific quantization files.

The ModelSummary class holds metadata for each available model

The ModelMetadata holds

Each ModelQuantization object provides

There are two methods to download a quantization file:

1. Calling .download() from a ModelQuantization object
   • For example: client.models.get('<model>').get_quantization('<method>').download()
2. client.models.download('quantized-file-name')
   • the .models.download() method accepts two types of input: string name of the model with quantization or a ModelQuantization object

If the model file has already been downloaded this function returns immediately. Otherwise a progress bar is shown showing the download progress.

The .servers accessor provides methods to list running servers, start new servers, and stop servers.

The .create method will create a new server configuration. If there is already a running server with the same model file and API parameters the matched server configuration is returned rather than creating and starting a new server.

The .create function has the following inputs

The three server parameters Pydantic classes are shown here. If the value None is used for any parameter the server will utilize the backend default value.

class APIParams(BaseModel, extra="forbid"):
    host: str = "127.0.0.1"
    port: int = 0            # 0 means find a random unused port
    api_key: str | None = None
    log_disable: bool | None = None
    mmproj: str | None = None
    timeout: int | None = None
    verbose: bool | None = None
    n_gpu_layers: int | None = None
    main_gpu: int | None = None
    metrics: bool | None = None


class LoadParams(BaseModel, extra="forbid"):
    batch_size: int | None = None
    cont_batching: bool | None = None
    ctx_size: int | None = None
    main_gpu: int | None = None
    memory_f32: bool | None = None
    mlock: bool | None = None
    n_gpu_layers: int | None = None
    rope_freq_base: int | None = None
    rope_freq_scale: int | None = None
    seed: int | None = None
    tensor_split: list[int] | None = None
    use_mmap: bool | None = None
    embedding: bool | None = None


class InferParams(BaseModel, extra="forbid"):
    threads: int | None = None
    n_predict: int | None = None
    top_k: int | None = None
    top_p: float | None = None
    min_p: float | None = None
    repeat_last: int | None = None
    repeat_penalty: float | None = None
    temp: float | None = None
    parallel: int | None = None

For example to create a server with the OpenHermes model with default values

from anaconda_ai import get_default_client

client = get_default_client()
server = client.servers.create(
  'OpenHermes-2.5-Mistral-7B/Q4_K_M',
)

By default creating a server configuration will

• download the model file if needed
• run the server API on a random unused port

The optional server parameters listed above can be passed as dictionaries as well as avoiding automatic model downloads. For example

server = client.servers.create(
  'OpenHermes-2.5-Mistral-7B/Q4_K_M',
  api_params={"main_gpu": 1, "port": 9999},
  load_params={"ctx_size": 512, "n_gpu_layers": 10},
  infer_params={"temp": 0.1},
  download_if_needed=False
)

When a server is created it is not automatically started. A server can be started and stopped in a number of ways

From the server object

server.start()
server.stop()

From the .servers accessor

client.servers.start(server)
client.servers.stop(server)

Alternatively you can use .create as a context manager, which will automatically stop the server on exit of the indented block.

with client.servers.create('OpenHermes-2.5-Mistral-7B/Q4_K_M') as server:
    openai_client = server.openai_client()
    # make requests to the server

• .url: is the full url to the running server
• .openai_url: is the url with /v1 appended to utilize the OpenAI compatibility endpoints
• .openai_client(): creates a pre-configured OpenAI client for this url
• .openai_async_client(): creates a pre-configured Async OpenAI client for this url

Each of .openai_client() and opeanai_async_client() allow extra keyword parameters to pass to the client initialization.

Creates a postgres vector db and returns the connection information.

anaconda ai launch-vectordb

To use the llm integration you will need to also install llm package

conda install -c conda-forge llm

then you can list downloaded model quantizations

llm models

or to show only the Anaconda AI models

llm models list -q anaconda

When utilizing a model it will first ensure that the model has been downloaded and start the server though the backend. Standard OpenAI parameters are supported.

llm -m 'anaconda:meta-llama/llama-2-7b-chat-hf_Q4_K_M.gguf' -o temperature 0.1 'what is pi?'

Standard OpenAI and the above server options are available for Anaconda AI models, to see the parameter names run

llm models list -q anaconda --options

The LangChain integration provides Chat and Embedding classes that automatically manage downloading and starting servers. You will need the langchain-openai package.

from langchain.prompts import ChatPromptTemplate
from anaconda_ai.integrations.langchain import AnacondaQuantizedModelChat, AnacondaQuantizedModelEmbeddings

prompt = ChatPromptTemplate.from_template("tell me a joke about {topic}")
model = AnacondaQuantizedModelChat(model_name='meta-llama/llama-2-7b-chat-hf_Q4_K_M.gguf')

chain = prompt | model

message = chain.invoke({'topic': 'python'})

The following keyword arguments are supported:

• api_params: Dict or APIParams class above
• load_params: Dict or LoadParams class above
• infer_params: Dict or InferParams class above (excluding AnacondaQuantizedEmbedding)

You will need at least the llama-index-llms-openai package installed to use the integration.

from anaconda_ai.integrations.llama_index import AnacondaModel

llm = AnacondaModel(
    model='OpenHermes-2.5-Mistral-7B_q4_k_m'
)

The AnacondaModel class supports the following arguments

• model: Name of the model using the pattern defined above
• system_prompt: Optional system prompt to apply to completions and chats
• temperature: Optional temperature to apply to all completions and chats (default is 0.1)
• max_tokens: Optional Max tokens to predict (default is to let the model decide when to finish)
• api_params: Optional dict or APIParams object
• load_params: Optional dict or LoadParams object
• infer_params: Optional dict or InferParams object

This provides a CustomLLM provider for use with litellm. But, since litellm does not currently support entrypoints to register the provider, the user must import the module first.

import litellm
import anaconda_ai.integrations.litellm

response = litellm.completion(
    'anaconda/openhermes-2.5-mistral-7b/q4_k_m',
    messages=[{'role': 'user', 'content': 'what is pi?'}]
)

Supported usage:

• completion (with and without stream=True)
• acompletion (with and without stream=True)
• Most OpenAI inference parameters
  • n: number of completions is not supported
• Server parameters (api_params, load_params, infer_params) can be passed as dictionaries to the optional_params keyword argument
  • optional_params={"load_params": {"ctx_size": 512}}

Since DSPy uses LiteLLM, Anaconda models can be used with dspy. Streaming and async are supported for raw LLM calls and for modules like Predict or ChainofThought .

import dspy
import anaconda_ai.integrations.litellm

lm = dspy.LM('anaconda/openhermes-2.5-mistral-7b/q4_k_m')
dspy.configure(lm=lm)

chain = dspy.ChainOfThought("question -> answer")
chain(question="Who are you?")

dspy.LM supports optional_params= keyword argument as explained in the previous section.

A callback is available to work with Panel's ChatInterface

To use it you will need to have panel, httpx, and numpy installed.

Here's an example application that can be written in Python script or Jupyter Notebook

import panel as pn
from anaconda_ai.integrations.panel import AnacondaModelHandler

pn.extension('echarts', 'tabulator', 'terminal')

llm = AnacondaModelHandler('TinyLlama/TinyLlama-1.1B-Chat-v1.0_Q4_K_M.gguf', display_throughput=True)

chat = pn.chat.ChatInterface(
    callback=llm.callback,
    show_button_name=False)

chat.send(
    "I am your assistant. How can I help you?",
    user=llm.model_id, avatar=llm.avatar, respond=False
)
chat.servable()

the AnacondaModelHandler supports the following keyword arguments

• display_throughput: Show a speed dial next to the response. Default is False
• system_message: Default system message applied to all responses
• client_options: Optional dict passed as kwargs to chat.completions.create
• api_params: Optional dict or APIParams object
• load_params: Optional dict or LoadParams object
• infer_params: Optional dict or InferParams object

Ensure you have conda installed. Then run:

make setup

make test

make tox