Oracle Cloud Infrastructure Documentation

Model Limitations in Generative AI

Review the following model requirements for the OCI Generative AI custom and base models to get the most out of your models.

Note

For key features of the pretrained base models, see Pretrained Foundational Models in Generative AI.

Matching Clusters to Base Models

Expand the following sections to review the dedicated AI cluster unit size and units that match each foundational model.

Chat
Base Model Fine-tuning Cluster Hosting Cluster
meta.llama-3-70b-instruct
  • Unit Size: Large Generic
  • Required Units: 2
  • Unit Size: Large Generic
  • Required Units: 1
cohere.command-r-plus Not available for fine-tuning
  • Unit Size: Large Cohere V2
  • Required Units: 1
cohere.command-r-16k
  • Unit Size: Small Cohere V2
  • Required Units: 8
  • Unit Size: Small Cohere V2
  • Required Units: 1
Text Generation
Base Model Fine-tuning Cluster Hosting Cluster
cohere.command
  • Unit Size: Large Cohere
  • Required Units: 2
  • Unit Size: Large Cohere
  • Required Units: 1
cohere.command-light
  • Unit Size: Small Cohere
  • Required Units: 2
  • Unit Size: Small Cohere
  • Required Units: 1
meta.llama-2-70b-chat Not available for fine-tuning
  • Unit Size: Llama2 70
  • Required Units: 1
Summarization
Base Model Fine-tuning Cluster Hosting Cluster
cohere.command Not available for fine-tuning
  • Unit Size: Large Cohere
  • Required Units: 1
Embedding
Base Model Fine-tuning Cluster Hosting Cluster
cohere.embed Not available for fine-tuning
  • Unit Size: Embed Cohere
  • Required Units: 1
Units for Fine-Tuning Clusters
Creating a fine-tuning dedicated AI cluster automatically provisions a fixed number of units based on the base model: 8 units for cohere.command-r-16k and 2 units for other models. This cannot be changed, but you can use the same cluster to fine-tune multiple models.
Units for Hosting Clusters
  • When creating a cluster, by default, one unit is created for the selected base model.
  • You can increase throughput or requests per minute (RPM) by adding model replicas. For example, 2 replicas require 2 units. You can add model replicas when creating or editing a hosting cluster.
  • Host up to 50 models on the same cluster, with the following restrictions:
    • Host up to 50 of the the same version of a fine-tuned or a pretrained model on the same cluster.
    • Host different versions of the same base model, only if using T-FEW fine-tuning method for cohere.command and cohere.command-light base models.
Note

Instead of committing to dedicated AI clusters, you can pay as you go for on-demand inferencing. With on-demand inferencing you reach the foundational models either through the Console, in the playground or through the API. For on-demand features, see Calculating Cost in Generative AI.

Adding Endpoints to Hosting Clusters

Important

To host a model for inference on a hosting dedicated AI cluster, you must create an endpoint for that model. Then, you can add either a custom model or a pretrained foundational model to that endpoint. A hosting dedicated AI cluster can have up to 50 endpoints. Use these endpoints for the following use cases:

Creating Endpoint Aliases

Create aliases with many endpoints. These 50 endpoints must either point to the same base model or the same version of a custom model. Creating many endpoints that point to the same model makes it easier to manage the endpoints, because you can use the endpoints for different users or different purposes.

Stack Serving

Host several versions of a cusom model on one cluster. This applies to cohere.command and cohere.command-light models that are fine-tuned with the T-Few training method. Hosting various versions of a fine-tuned model can help you to assess the custom models for different use cases.

Tip

You can increase the instance count to increase the call volume supported by a hosting cluster.

Expand the following sections to review the requirements for hosting models on the same cluster.

Chat
For hosting the pretrained or fine-tuned chat models on a hosting dedicated AI cluster, use the following cluster unit size and endpoint rules that match your base model.
Hosting Cluster Unit Size Matching Rules
Large Generic

To host the same pretrained base model through several endpoints on the same cluster:

  • Create as many endpoints as needed for the meta.llama-3-70b-instruct model on the same hosting cluster.

To host several custom models on the same cluster:

  • Fine-tune one model with the LoRA training method.
  • Use the meta.llama-3-70b-instruct model as the base.
  • Create as many endpoints as needed for the custom model (same version).
Large Cohere V2

To host the same pretrained base model through several endpoints on the same cluster:

  • Create as many endpoints as needed for the cohere.command-r-plus model on the same hosting cluster.
Small Cohere V2

To host the same pretrained base model through several endpoints on the same cluster:

  • Create as many endpoints as needed for the cohere.command-r-16k model on the same hosting cluster.

To host several custom models on the same cluster:

  • Fine-tune one model with the T-Few training method.
  • Use the cohere.command-r-16k model as the base.
  • Create as many endpoints as needed for the custom model (same version).

You can't host different versions of a custom model trained on the cohere.command-r-16k base model on the same cluster, as stack serving is not supported.
Text Generation

You can use the chat models for text generation. If you want to use the text generation models instead, review the following table for rules that match the hosting clusters to the base models. To decide which models to use, review the pretrained models. For hosting text generation models on a hosting dedicated AI cluster, use the following cluster unit size and endpoint rules that match your base model.

Hosting Cluster Unit Size Matching Rules
Small Cohere
To host the same pretrained base model through several endpoints on the same cluster:
  • Create as many endpoints as needed for the cohere.command-light model on the same hosting cluster.

To host different custom models on the same cluster:

  • Fine-tune all the models with the T-Few training method.
  • Use the cohere.command-light model as the base.
  • Ensure that all base models have the same version.
  • Create an endpoint for each model on the same hosting cluster.
Large Cohere

To host the same pretrained base model through several endpoints on the same cluster:

  • Create as many endpoints as needed for the cohere.command model with the same version on the same hosting cluster.

To host different custom models on the same cluster:

  • Fine-tune all the models with the T-Few training method.
  • Use the cohere.command model as the base.
  • Ensure that all base models have the same version.
  • Add an endpoint to the hosting cluster for each model.
Llama2 70 To host the same pretrained base model through several endpoints on the same cluster:
  • Create as many endpoints as needed for the meta.llama-2-70b-chat model on the same hosting cluster.
Summarization

You can use the chat models for summarization. If you want to use the cohere.command model instead, review the following table for rules that match the hosting clusters to the base models. To decide which model to use, review the pretrained models. For hosting the pretrained cohere.command summarization model on a hosting dedicated AI cluster, use the following cluster unit size and endpoint rules.

Hosting Cluster Unit Size Matching Rules
Large Cohere

To host the same pretrained base model through several endpoints on the same cluster:

  • Create as many endpoints as needed for the cohere.command model with the same version on the same hosting cluster.

To host different custom models on the same cluster:

  • Fine-tune all the models with the T-Few training method.
  • Use the cohere.command model as the base.
  • Ensure that all base models have the same version.
  • Add an endpoint to the hosting cluster for each model.
Embedding

For hosting the embedding models on a hosting dedicated AI cluster, use the following cluster unit size and endpoint rules.

Hosting Cluster Unit Size Matching Rules
Embed Cohere To host the same pretrained base model through several endpoints on the same cluster:
  • Create as many endpoints as needed for the cohere.embed-<models> model with the exact same model name and version on the same hosting cluster.

Training Data

Datasets for training custom models have the following requirements:

  • A maximum of one fine-tuning dataset is allowed per custom model. This dataset is randomly split to a 80:20 ratio for training and validating.
  • Each file must have at least 32 prompt/completion pair examples.
  • The file format is JSONL.
  • Each line in the JSONL file has the following format:

    {"prompt": "<a prompt>", "completion": "<expected response given the prompt>"}\n

  • The file must be stored in an OCI Object Storage bucket.

Learn about Training Data Requirements in Generative AI.

Input Data for Text Embeddings

Input data for creating text embeddings has the following requirements:

  • You can add sentences, phrases, or paragraphs for embeddings either one phrase at a time, or by uploading a file.
  • Only files with a .txt extension are allowed.
  • If you use an input file, each input sentence, phrase, or paragraph in the file must be separated with a newline character.
  • A maximum of 96 inputs are allowed for each run.
  • Each input must be less than 512 tokens. If an input is too long, select whether to cut off the start or the end of the text to fit within the token limit by setting the Truncate parameter to Start or End. If an input exceeds the 512 token limit and the Truncate parameter is set to None, you get an error message.

Learn about Creating text embeddings in OCI Generative AI.