Import customized fashions in Amazon Bedrock (preview)

With Amazon Bedrock, you’ve entry to a alternative of high-performing basis fashions (FMs) from main synthetic intelligence (AI) corporations that make it simpler to construct and scale generative AI functions. A few of these fashions present publicly out there weights that may be fine-tuned and customised for particular use instances. Nonetheless, deploying custom-made FMs in a safe and scalable method isn’t a straightforward activity.

Beginning at this time, Amazon Bedrock provides in preview the aptitude to import customized weights for supported mannequin architectures (similar to Meta Llama 2, Llama 3, and Mistral) and serve the customized mannequin utilizing On-Demand mode. You possibly can import fashions with weights in Hugging Face safetensors format from Amazon SageMaker and Amazon Easy Storage Service (Amazon S3).

On this method, you should use Amazon Bedrock with present custom-made fashions similar to Code Llama, a code-specialized model of Llama 2 that was created by additional coaching Llama 2 on code-specific datasets, or use your knowledge to fine-tune fashions to your personal distinctive enterprise case and import the ensuing mannequin in Amazon Bedrock.

Let’s see how this works in follow.

Bringing a customized mannequin to Amazon Bedrock
Within the Amazon Bedrock console, I select Imported fashions from the Basis fashions part of the navigation pane. Now, I can create a customized mannequin by importing mannequin weights from an Amazon Easy Storage Service (Amazon S3) bucket or from an Amazon SageMaker mannequin.

I select to import mannequin weights from an S3 bucket. In one other browser tab, I obtain the MistralLite mannequin from the Hugging Face web site utilizing this pull request (PR) that gives weights in safetensors format. The pull request is at the moment Able to merge, so it is perhaps a part of the primary department if you learn this. MistralLite is a fine-tuned Mistral-7B-v0.1 language mannequin with enhanced capabilities of processing lengthy context as much as 32K tokens.

When the obtain is full, I add the recordsdata to an S3 bucket in the identical AWS Area the place I’ll import the mannequin. Listed here are the MistralLite mannequin recordsdata within the Amazon S3 console:

Again on the Amazon Bedrock console, I enter a reputation for the mannequin and hold the proposed import job identify.

I choose Mannequin weights within the Mannequin import settings and browse S3 to decide on the situation the place I uploaded the mannequin weights.

To authorize Amazon Bedrock to entry the recordsdata on the S3 bucket, I choose the choice to create and use a brand new AWS Id and Entry Administration (IAM) service function. I take advantage of the View permissions particulars hyperlink to verify what shall be within the function. Then, I submit the job.

About ten minutes later, the import job is accomplished.

Now, I see the imported mannequin within the console. The listing additionally reveals the mannequin Amazon Useful resource Identify (ARN) and the creation date.

I select the mannequin to get extra data, such because the S3 location of the mannequin recordsdata.

Within the mannequin element web page, I select Open in playground to check the mannequin within the console. Within the textual content playground, I kind a query utilizing the immediate template of the mannequin:

<|prompter|>What are the primary challenges to assist an extended context for LLM?</s><|assistant|>

The MistralLite imported mannequin is fast to answer and describe a few of these challenges.

Within the playground, I can tune responses for my use case utilizing configurations similar to temperature and most size or add cease sequences particular to the imported mannequin.

To see the syntax of the API request, I select the three small vertical dots on the prime proper of the playground.

I select View API syntax and run the command utilizing the AWS Command Line Interface (AWS CLI):

aws bedrock-runtime invoke-model 
--model-id arn:aws:bedrock:us-east-1:123412341234:imported-model/a82bkefgp20f 
--body "prompter" 
--cli-binary-format raw-in-base64-out 
--region us-east-1 
invoke-model-output.txt

The output is just like what I bought within the playground. As you possibly can see, for imported fashions, the mannequin ID is the ARN of the imported mannequin. I can use the mannequin ID to invoke the imported mannequin with the AWS CLI and AWS SDKs.

Issues to know
You possibly can convey your individual weights for supported mannequin architectures to Amazon Bedrock within the US East (N. Virginia) AWS Area. The mannequin import functionality is at the moment out there in preview.

When utilizing customized weights, Amazon Bedrock serves the mannequin with On-Demand mode, and also you solely pay for what you employ with no time-based time period commitments. For detailed data, see Amazon Bedrock pricing.

The flexibility to import fashions is managed utilizing AWS Id and Entry Administration (IAM), and you’ll permit this functionality solely to the roles in your group that have to have it.

With this launch, it’s now simpler to construct and scale generative AI functions utilizing customized fashions with safety and privateness in-built.

To study extra:

— Danilo