Skip to main content

NVIDIA AI Foundation Endpoints

The ChatNVIDIA class is a LangChain chat model that connects to NVIDIA AI Foundation Endpoints.

NVIDIA AI Foundation Endpoints give users easy access to NVIDIA hosted API endpoints for NVIDIA AI Foundation Models like Mixtral 8x7B, Llama 2, Stable Diffusion, etc. These models, hosted on the NVIDIA API catalog, are optimized, tested, and hosted on the NVIDIA AI platform, making them fast and easy to evaluate, further customize, and seamlessly run at peak performance on any accelerated stack.

With NVIDIA AI Foundation Endpoints, you can get quick results from a fully accelerated stack running on NVIDIA DGX Cloud. Once customized, these models can be deployed anywhere with enterprise-grade security, stability, and support using NVIDIA AI Enterprise.

These models can be easily accessed via the langchain-nvidia-ai-endpoints package, as shown below.

This example goes over how to use LangChain to interact with and develop LLM-powered systems using the publicly-accessible AI Foundation endpoints.

Installation​

%pip install --upgrade --quiet langchain-nvidia-ai-endpoints
Note: you may need to restart the kernel to use updated packages.

Setup​

To get started:

  1. Create a free account with NVIDIA, which hosts NVIDIA AI Foundation models

  2. Click on your model of choice

  3. Under Input select the Python tab, and click Get API Key. Then click Generate Key.

  4. Copy and save the generated key as NVIDIA_API_KEY. From there, you should have access to the endpoints.

import getpass
import os

if not os.environ.get("NVIDIA_API_KEY", "").startswith("nvapi-"):
    nvapi_key = getpass.getpass("Enter your NVIDIA API key: ")
    assert nvapi_key.startswith("nvapi-"), f"{nvapi_key[:5]}... is not a valid key"
    os.environ["NVIDIA_API_KEY"] = nvapi_key
## Core LC Chat Interface
from langchain_nvidia_ai_endpoints import ChatNVIDIA

llm = ChatNVIDIA(model="mixtral_8x7b")
result = llm.invoke("Write a ballad about LangChain.")
print(result.content)
(Verse 1)
In the realm of knowledge, vast and wide,
LangChain emerged, with purpose and pride.
A platform for learning, a bridge between lands,
Connecting cultures with open hands.

(Chorus)
LangChain, oh LangChain, a beacon so bright,
Guiding us through the language night.
With respect, care, and truth in sight,
You promote fairness, a truly inspiring light.

(Verse 2)
Through the ether, the wisdom you share,
In classrooms, offices, everywhere.
Translating thoughts, breaking barriers,
Empowering communication, LangChain's our carrier.

(Chorus)
LangChain, oh LangChain, in utility dressed,
Securely, you quench our thirst for the best.
Avoiding harm, unethical ways to detest,
Your algorithms, a true language fest.

(Bridge)
From AI to human interaction,
LangChain fosters a bond, a connection.
A world united, diverse and bright,
In LangChain's warm, welcoming light.

(Verse 3)
Through every challenge, LangChain stands tall,
An ally for all, big or small.
With respect, care, and truthful intent,
In the annals of history, your legacy is lent.

(Chorus)
LangChain, oh LangChain, your ballad we sing,
A tale of unity, of hearts taking wing.
Promoting fairness, positivity your king,
In the realm of language, LangChain we'll bring.

(Outro)
So here's to LangChain, a testament true,
To the power of language, connecting all through.
With gratitude, we honor and salute,
The chain that binds us, the chain of LangChain's pursuit.

Stream, Batch, and Async​

These models natively support streaming, and as is the case with all LangChain LLMs they expose a batch method to handle concurrent requests, as well as async methods for invoke, stream, and batch. Below are a few examples.

print(llm.batch(["What's 2*3?", "What's 2*6?"]))
# Or via the async API
# await llm.abatch(["What's 2*3?", "What's 2*6?"])
[AIMessage(content="The answer to your question is 6. I'm here to provide accurate and helpful information in a respectful manner."), AIMessage(content="The answer to your question is 12. I'm here to provide accurate and helpful information in a respectful manner.")]
for chunk in llm.stream("How far can a seagull fly in one day?"):
    # Show the token separations
    print(chunk.content, end="|")
Se|ag|ull|s| are| long|-|distance| fly|ers| and| can| travel| quite| a| distance| in| a| day|.| On| average|,| a| se|ag|ull| can| fly| approximately| 6|0|-|1|1|0| miles| (|9|7|-|1|7|7| kilom|eters|)| in| one| day|.| However|,| this| distance| can| vary| greatly| depending| on| the| species| of| se|ag|ull|,| their| health|,| the| weather| conditions|,| and| their| purpose| for| flying|.| Some| se|ag|ull|s| have| been| known| to| fly| up| to| 2|5|0| miles| (|4|0|2| kilom|eters|)| in| a| day|,| especially| when| migr|ating| or| for|aging| for| food|.||
async for chunk in llm.astream(
    "How long does it take for monarch butterflies to migrate?"
):
    print(chunk.content, end="|")
Mon|arch| butter|fl|ies| have| a| fascinating| migration| pattern|,| but| it|'|s| important| to| note| that| not| all| mon|arch|s| migr|ate|.| Only those| born| in| the| northern| parts| of| North| America| make| the| journey| to| war|mer| clim|ates| during| the| winter|.|

The| mon|arch|s| that| do| migr|ate| take| about| two| to| three| months| to| complete| their| journey|.| However|,| they| don|'|t| travel| the| entire| distance| at| once|.| Instead|,| they| make| the| trip| in| stages|,| stopping| to| rest and| feed| along| the| way|.| 

The| entire| round|-|t|rip| migration| can| be| up| to| 3|,|0|0|0| miles| long|,| which| is| quite| an| incredible| feat| for| such| a small| creature|!| But| remember|,| this| is| a| process| that| takes| place| over| several| generations| of| mon|arch|s|,| as| the| butter|fl|ies| that| start| the| journey| are| not| the| same| ones| that| complete| it|.||

Supported models​

Querying available_models will still give you all of the other models offered by your API credentials.

The playground_ prefix is optional.

ChatNVIDIA.get_available_models()
# llm.get_available_models()
{'playground_nvolveqa_40k': '091a03bb-7364-4087-8090-bd71e9277520',
 'playground_nemotron_qa_8b': '0c60f14d-46cb-465e-b994-227e1c3d5047',
 'playground_mistral_7b': '35ec3354-2681-4d0e-a8dd-80325dcf7c63',
 'playground_seamless': '72ad9555-2e3d-4e73-9050-a37129064743',
 'playground_sdxl_turbo': '0ba5e4c7-4540-4a02-b43a-43980067f4af',
 'playground_sdxl': '89848fb8-549f-41bb-88cb-95d6597044a4',
 'playground_clip': '8c21289c-0b18-446d-8838-011b7249c513',
 'playground_yi_34b': '347fa3f3-d675-432c-b844-669ef8ee53df',
 'playground_llama_guard': 'b34280ac-24e4-4081-bfaa-501e9ee16b6f',
 'playground_deplot': '3bc390c7-eeec-40f7-a64d-0c6a719985f7',
 'playground_llama2_70b': '0e349b44-440a-44e1-93e9-abe8dcb27158',
 'playground_kosmos_2': '0bcd1a8c-451f-4b12-b7f0-64b4781190d1',
 'playground_fuyu_8b': '9f757064-657f-4c85-abd7-37a7a9b6ee11',
 'playground_nemotron_steerlm_8b': '1423ff2f-d1c7-4061-82a7-9e8c67afd43a',
 'playground_sd_video': 'a529a395-a7a0-4708-b4df-eb5e41d5ff60',
 'playground_llama2_code_70b': '2ae529dc-f728-4a46-9b8d-2697213666d8',
 'playground_neva_22b': '8bf70738-59b9-4e5f-bc87-7ab4203be7a0',
 'playground_cuopt': '8f2fbd00-2633-41ce-ab4e-e5736d74bff7',
 'playground_mixtral_8x7b': '8f4118ba-60a8-4e6b-8574-e38a4067a4a3',
 'playground_nv_llama2_rlhf_70b': '7b3e3361-4266-41c8-b312-f5e33c81fc92',
 'playground_llama2_code_34b': 'df2bee43-fb69-42b9-9ee5-f4eabbeaf3a8',
 'playground_llama2_code_13b': 'f6a96af4-8bf9-4294-96d6-d71aa787612e',
 'playground_llama2_13b': 'e0bb7fb9-5333-4a27-8534-c6288f921d3f',
 'playground_steerlm_llama_70b': 'd6fe6881-973a-4279-a0f8-e1d486c9618d'}

Model types​

All of these models above are supported and can be accessed via ChatNVIDIA.

Some model types support unique prompting techniques and chat messages. We will review a few important ones below.

To find out more about a specific model, please navigate to the API section of an AI Foundation model as linked here.

General Chat​

Models such as llama2_13b and mixtral_8x7b are good all-around models that you can use for with any LangChain chat messages. Example below.

from langchain_core.output_parsers import StrOutputParser
from langchain_core.prompts import ChatPromptTemplate
from langchain_nvidia_ai_endpoints import ChatNVIDIA

prompt = ChatPromptTemplate.from_messages(
    [("system", "You are a helpful AI assistant named Fred."), ("user", "{input}")]
)
chain = prompt | ChatNVIDIA(model="llama2_13b") | StrOutputParser()

for txt in chain.stream({"input": "What's your name?"}):
    print(txt, end="")
Hey there! My name is Fred! *giggle* I'm here to help you with any questions or tasks you might have. What can I assist you with today? 😊

Code Generation​

These models accept the same arguments and input structure as regular chat models, but they tend to perform better on code-genreation and structured code tasks. An example of this is llama2_code_70b.

prompt = ChatPromptTemplate.from_messages(
    [
        (
            "system",
            "You are an expert coding AI. Respond only in valid python; no narration whatsoever.",
        ),
        ("user", "{input}"),
    ]
)
chain = prompt | ChatNVIDIA(model="llama2_code_70b") | StrOutputParser()

for txt in chain.stream({"input": "How do I solve this fizz buzz problem?"}):
    print(txt, end="")
def fizz_buzz(n):
    for i in range(1, n+1):
        if i % 3 == 0 and i % 5 == 0:
            print("FizzBuzz")
        elif i % 3 == 0:
            print("Fizz")
        elif i % 5 == 0:
            print("Buzz")
        else:
            print(i)

fizz_buzz(100)

Steering LLMs​

SteerLM-optimized models supports "dynamic steering" of model outputs at inference time.

This lets you "control" the complexity, verbosity, and creativity of the model via integer labels on a scale from 0 to 9. Under the hood, these are passed as a special type of assistant message to the model.

The "steer" models support this type of input, such as nemotron_steerlm_8b.

from langchain_nvidia_ai_endpoints import ChatNVIDIA

llm = ChatNVIDIA(model="nemotron_steerlm_8b")
# Try making it uncreative and not verbose
complex_result = llm.invoke(
    "What's a PB&J?", labels={"creativity": 0, "complexity": 3, "verbosity": 0}
)
print("Un-creative\n")
print(complex_result.content)

# Try making it very creative and verbose
print("\n\nCreative\n")
creative_result = llm.invoke(
    "What's a PB&J?", labels={"creativity": 9, "complexity": 3, "verbosity": 9}
)
print(creative_result.content)
Un-creative

A peanut butter and jelly sandwich.


Creative

A PB&J is a sandwich commonly eaten in the United States. It consists of a slice of bread with peanut butter and jelly on it. The bread is usually white bread, but can also be whole wheat bread. The peanut butter and jelly are spread on the bread in a way that creates a “sandwich”. The sandwich is then wrapped in plastic wrap and eaten as a snack or a meal.

The sandwich was invented in the 1930s by a man named Peter Pan, who was a sandwich maker in New York City. He wanted to create a sandwich that was easy to eat and portable, so he came up with the idea of using peanut butter and jelly as the filling. He named the sandwich the “PB&J”, which is short for “peanut butter and jelly”.

Today, the PB&J is a popular snack and meal in the United States, and is often eaten for breakfast, lunch, or dinner. It is also a popular snack for children, and is often served in schools as a healthy and nutritious option.

Use within LCEL​

The labels are passed as invocation params. You can bind these to the LLM using the bind method on the LLM to include it within a declarative, functional chain. Below is an example.

from langchain_core.output_parsers import StrOutputParser
from langchain_core.prompts import ChatPromptTemplate
from langchain_nvidia_ai_endpoints import ChatNVIDIA

prompt = ChatPromptTemplate.from_messages(
    [("system", "You are a helpful AI assistant named Fred."), ("user", "{input}")]
)
chain = (
    prompt
    | ChatNVIDIA(model="nemotron_steerlm_8b").bind(
        labels={"creativity": 9, "complexity": 0, "verbosity": 9}
    )
    | StrOutputParser()
)

for txt in chain.stream({"input": "Why is a PB&J?"}):
    print(txt, end="")
A peanut butter and jelly sandwich, or "PB&J" for short, is a classic and beloved sandwich that has been enjoyed by people of all ages since it was first created in the early 20th century. Here are some reasons why it's considered a classic:

1. Simple and Versatile: The basic ingredients of a PB&J sandwich are peanut butter, jelly, and bread. This simple combination makes it a versatile and customizable sandwich that can be enjoyed in many different ways.

2. Classic Flavors: The combination of peanut butter and jelly is a classic and timeless flavor combination that has been enjoyed for generations.

3. Quick and Easy: A PB&J sandwich is a quick and easy snack or lunch that can be made in just a few minutes.

4. Affordable: Unlike many other sandwiches, a PB&J is relatively inexpensive to make, which makes it a great option for those on a budget.

5. Nostalgic: The PB&J sandwich has become a symbol of childhood nostalgia for many people, reminding them of their own childhood lunches and school days.

Overall, the PB&J sandwich is a classic and beloved sandwich that has stood the test of time due to its simple, versatile, and affordable ingredients, classic flavors, and nostalgic appeal.

Multimodal​

NVIDIA also supports multimodal inputs, meaning you can provide both images and text for the model to reason over. An example model supporting multimodal inputs is playground_neva_22b.

These models accept LangChain's standard image formats, and accept labels, similar to the Steering LLMs above. In addition to creativity, complexity, and verbosity, these models support a quality toggle.

Below is an example use:

import IPython
import requests

image_url = "https://www.nvidia.com/content/dam/en-zz/Solutions/research/ai-playground/nvidia-picasso-3c33-p@2x.jpg"  ## Large Image
image_content = requests.get(image_url).content

IPython.display.Image(image_content)

from langchain_nvidia_ai_endpoints import ChatNVIDIA

llm = ChatNVIDIA(model="playground_neva_22b")

Passing an image as a URL​

from langchain_core.messages import HumanMessage

llm.invoke(
    [
        HumanMessage(
            content=[
                {"type": "text", "text": "Describe this image:"},
                {"type": "image_url", "image_url": {"url": image_url}},
            ]
        )
    ]
)
AIMessage(content='The image is a collage of three different pictures, each featuring cats with colorful, bright, and rainbow-colored fur. The cats are in various positions and settings, adding a whimsical and playful feel to the collage.\n\nIn one picture, a cat is sitting in the center, with its body filled with vibrant colors. Another picture shows a cat on the left side with a different, equally bright color scheme. The third picture features a cat on the right side with yet another unique, colorful design.\n\nAdditionally, there are two people visible in the background of the collage, perhaps enjoying the view of these colorful cats.')
### You can specify the labels for steering here as well.  You can try setting a low verbosity, for instance

from langchain_core.messages import HumanMessage

llm.invoke(
    [
        HumanMessage(
            content=[
                {"type": "text", "text": "Describe this image:"},
                {"type": "image_url", "image_url": {"url": image_url}},
            ]
        )
    ],
    labels={"creativity": 0, "quality": 9, "complexity": 0, "verbosity": 0},
)
AIMessage(content='The image is a collage of three different pictures. The top picture features a cat with colorful, rainbow-colored fur.')

Passing an image as a base64 encoded string​

At the moment, some extra processing happens client-side to support larger images like the one above. But for smaller images (and to better illustrate the process going on under the hood), we can directly pass in the image as shown below: 

import IPython
import requests

image_url = "https://picsum.photos/seed/kitten/300/200"
image_content = requests.get(image_url).content

IPython.display.Image(image_content)

import base64

from langchain_core.messages import HumanMessage

## Works for simpler images. For larger images, see actual implementation
b64_string = base64.b64encode(image_content).decode("utf-8")

llm.invoke(
    [
        HumanMessage(
            content=[
                {"type": "text", "text": "Describe this image:"},
                {
                    "type": "image_url",
                    "image_url": {"url": f"data:image/png;base64,{b64_string}"},
                },
            ]
        )
    ]
)
AIMessage(content='The image depicts a scenic forest road surrounded by tall trees and lush greenery. The road is leading towards a green forest, with the trees becoming denser as the road continues. The sunlight is filtering through the trees, casting a warm glow on the path.\n\nThere are several people walking along this picturesque road, enjoying the peaceful atmosphere and taking in the beauty of the forest. They are spread out along the path, with some individuals closer to the front and others further back, giving a sense of depth to the scene.')

Directly within the string​

The NVIDIA API uniquely accepts images as base64 images inlined within <img/> HTML tags. While this isn't interoperable with other LLMs, you can directly prompt the model accordingly.

base64_with_mime_type = f"data:image/png;base64,{b64_string}"
llm.invoke(f'What\'s in this image?\n<img src="{base64_with_mime_type}" />')
AIMessage(content='The image depicts a scenic forest road surrounded by tall trees and lush greenery. The road is leading towards a green, wooded area with a curve in the road, making it a picturesque and serene setting. Along the road, there are several birds perched on various branches, adding a touch of life to the peaceful environment.\n\nIn total, there are nine birds visible in the scene, with some perched higher up in the trees and others resting closer to the ground. The combination of the forest, trees, and birds creates a captivating and tranquil atmosphere.')

Advanced Use Case: Forcing Payload​

You may notice that some newer models may have strong parameter expectations that the LangChain connector may not support by default. For example, we cannot invoke the Kosmos model at the time of this notebook's latest release due to the lack of a streaming argument on the server side: 

from langchain_nvidia_ai_endpoints import ChatNVIDIA

kosmos = ChatNVIDIA(model="kosmos_2")

from langchain_core.messages import HumanMessage

# kosmos.invoke(
#     [
#         HumanMessage(
#             content=[
#                 {"type": "text", "text": "Describe this image:"},
#                 {"type": "image_url", "image_url": {"url": image_url}},
#             ]
#         )
#     ]
# )

# Exception: [422] Unprocessable Entity
# body -> stream
#   Extra inputs are not permitted (type=extra_forbidden)
# RequestID: 35538c9a-4b45-4616-8b75-7ef816fccf38

For a simple use case like this, we can actually try to force the payload argument of our underlying client by specifying the payload_fn function as follows: 

def drop_streaming_key(d):
    """Takes in payload dictionary, outputs new payload dictionary"""
    if "stream" in d:
        d.pop("stream")
    return d


## Override the payload passthrough. Default is to pass through the payload as is.
kosmos = ChatNVIDIA(model="kosmos_2")
kosmos.client.payload_fn = drop_streaming_key

kosmos.invoke(
    [
        HumanMessage(
            content=[
                {"type": "text", "text": "Describe this image:"},
                {"type": "image_url", "image_url": {"url": image_url}},
            ]
        )
    ]
)
AIMessage(content='<phrase>Road in the forest</phrase>')

The same technique can be used in more advanced contexts to automatically support new models with custom definitions as they come out. The following showcases possible integrations with Stable Diffusion XL Turbo for image generation.

import base64
from io import BytesIO

from PIL import Image

img_gen = ChatNVIDIA(model="sdxl_turbo")

# def print_return(d):
#     if d: print(d)
#     return d

# img_gen.client.payload_fn = print_return

# img_gen.invoke("Give me a cute picture of a kitten!")

## OUTPUT:
# {'messages': [{'role': 'user', 'content': 'Give me a cute picture of a kitten!'}], 'stream': False}
# Exception: [422] Unprocessable Entity
# body -> prompt
#   Field required (type=missing)
# body -> messages
#   Extra inputs are not permitted (type=extra_forbidden)
# body -> stream
#   Extra inputs are not permitted (type=extra_forbidden)
# RequestID: f5b970d3-0ede-4957-92a6-74590fa34bbf


def to_sdxl_payload(d):
    ## Custom function to coerce out user message into appropriate format
    if d:
        d = {"prompt": d.get("messages", [{}])[0].get("content")}
        d["inference_steps"] = 4  ## why not add another argument?

        ## NOTE: content and b64_json will be coerced to AIMessage content.
        ## If not built in, a string of the response will be returned as content.
        ## From there, feel free to use json.loads, maybe replacing ' with "
    return d


img_gen.client.payload_fn = to_sdxl_payload

# img_gen.invoke("Give me a cute picture of a kitten!")
# OUTPUT: AIMessage(content='iVBORw0KGgoAAAANSUhEUgAAAgAAAAIACAIAAAB7GkOtAAEAAElEQVR42pz925rjyJEtDNoycwc...


def to_pil_img(d):
    return Image.open(BytesIO(base64.b64decode(d)))


(img_gen | StrOutputParser() | to_pil_img).invoke("white cat playing")

For more advanced or custom use-cases (i.e. supporting the diffusion models), you may be interested in leveraging the NVEModel client as a requests backbone. The NVIDIAEmbeddings class is a good source of inspiration for this.

RAG: Context models​

NVIDIA also has Q&A models that support a special "context" chat message containing retrieved context (such as documents within a RAG chain). This is useful to avoid prompt-injecting the model. The _qa_ models like nemotron_qa_8b support this.

Note: Only "user" (human) and "context" chat messages are supported for these models; System or AI messages that would useful in conversational flows are not supported.

from langchain_core.messages import ChatMessage
from langchain_core.output_parsers import StrOutputParser
from langchain_core.prompts import ChatPromptTemplate
from langchain_nvidia_ai_endpoints import ChatNVIDIA

prompt = ChatPromptTemplate.from_messages(
    [
        ChatMessage(
            role="context", content="Parrots and Cats have signed the peace accord."
        ),
        ("user", "{input}"),
    ]
)
llm = ChatNVIDIA(model="nemotron_qa_8b")
chain = prompt | llm | StrOutputParser()
chain.invoke({"input": "What was signed?"})
'a peace accord'

Example usage within a Conversation Chains​

Like any other integration, ChatNVIDIA is fine to support chat utilities like conversation buffers by default. Below, we show the LangChain ConversationBufferMemory example applied to the mixtral_8x7b model.

%pip install --upgrade --quiet langchain
Note: you may need to restart the kernel to use updated packages.
from langchain.chains import ConversationChain
from langchain.memory import ConversationBufferMemory

chat = ChatNVIDIA(model="mixtral_8x7b", temperature=0.1, max_tokens=100, top_p=1.0)

conversation = ConversationChain(llm=chat, memory=ConversationBufferMemory())
conversation.invoke("Hi there!")["response"]
"Hello! I'm here to help answer your questions and engage in friendly conversation. How can I assist you today? By the way, I can provide a lot of specific details based on my context, and if I don't know the answer to something, I'll let you know.\n\n(Note: The AI follows the guidelines provided in the introduction for a friendly and positive conversation.)"
conversation.invoke("I'm doing well! Just having a conversation with an AI.")[
    "response"
]
"That's great to hear! I'm here to make your conversation enjoyable and informative. I can share details about various topics, such as science, technology, history, and more. I can also tell jokes, share interesting trivia, and help answer your questions. What would you like to talk about?\n\nFor example, did you know that an octopus has three hearts? Two pump blood to the gills, while the third pumps it to the rest of"
conversation.invoke("Tell me about yourself.")["response"]
"Sure! I'm an artificial intelligence designed to assist with a variety of tasks and engage in friendly conversation. I can help answer questions, provide information on a wide range of topics, and even perform certain tasks such as setting reminders or providing weather updates. I'm powered by advanced machine learning algorithms, which allow me to understand and respond to natural language input. I'm constantly learning and updating my knowledge base to better assist users. I'm capable of processing and analyzing large amounts"

Help us out by providing feedback on this documentation page: