Streaming LLM output

Large Language Models produce output token by token, and each token can take some time to generate. If you are waiting for the model to finish the entire response before outputting the result, these times add up quickly!

An alternative to this is streaming: Similar to video streaming, where you don’t wait for the entire video to be downloaded before playing it, you can stream the output of an LLM. In this recipe, we will explore how to do that with langchain_dartmouth!

Note

Many LLMs in the LangChain ecosystem support streaming, not just the ones in langchain_dartmouth! You could replace the model in this notebook with, e.g., ChatOpenAI from langchain_openai, and it would work exactly the same!

Importing and instantiating a model

Just as we saw in the previous recipe, we will import a chat model and then instantiate it. We will use the streaming parameter, however, to tell the model that we want it to stream its output!

from langchain_dartmouth.llms import ChatDartmouth
from dotenv import find_dotenv, load_dotenv

load_dotenv(find_dotenv())

True

llm = ChatDartmouth(model_name="meta.llama-3.2-11b-vision-instruct", streaming=True)

Streaming the output

We could use the invoke method as we did before, but that would still require us to wait for the entire response to be generated before returning. Since we have set our model’s streaming parameter to True, we can instead call the stream method. This will return a generator object. We can then iterate through this generator and print each returned chunk as it is generated:

for chunk in llm.stream("Write a haiku about Dartmouth College"):
    print(chunk.content)

Snow
y
 Han
over


D
art
mouth
's
 iv
y
 wisdom


Winter
's
 quiet
 h
ush

We can see that the chunks came in one by one, because the print function breaks the line after every chunk. We can use the end parameter to avoid that:

for chunk in llm.stream("Write a haiku about Dartmouth College"):
    print(chunk.content, end="")

Hanover's quiet
River flowing, green hills rise
Dartmouth's noble

That looks better! Let’s try a longer response to show the benefit of streaming:

for chunk in llm.stream("Write five haiku about Dartmouth College"):
    print(chunk.content, end="")

Snowy Hanover
Ancient trees stand tall, serene
Dart

mouth's quiet heart

River's gentle lap
Connecticut's soothing song

Peaceful summer eve

Baker-Berry's halls
Knowledge and wisdom's

 halls
Dartmouth's proud heart beats

Keddy's verd

ant slope
Mountain trails await, steep
Dartmouth's challenge

Winter

's snowy night
Tuckerman's slopes shine bright
Dartmouth's

 wild dream

Summary

This recipe showed how to stream output from an LLM using the stream method. Streaming long responses makes for a better user experience and a more efficient use of time when working with an LLM interactively.