Model Serving limits and regions
This article summarizes the limitations and region availability for Mosaic AI Model Serving and supported endpoint types.
Resource and payload limits
Mosaic AI Model Serving imposes default limits to ensure reliable performance. If you have feedback on these limits, reach out to your Databricks account team.
The following table summarizes resource and payload limitations for model serving endpoints.
| Feature | Granularity | Limit |
|---|---|---|
| Payload size | Per request | 16 MB. For endpoints serving foundation models or external models the limit is 4 MB. |
| Queries per second (QPS) | Per workspace | 200, but can be increased to 25,000 or more by reaching out to your Databricks account team. |
| Model execution duration | Per request | 120 seconds |
| CPU endpoint model memory usage | Per endpoint | 4GB |
| GPU endpoint model memory usage | Per endpoint | Greater than or equal to assigned GPU memory, depends on the GPU workload size |
| Provisioned concurrency | Per model and per workspace | 200 concurrency. Can be increased by reaching out to your Databricks account team. |
| Overhead latency | Per request | Less than 50 milliseconds |
| Init scripts | Init scripts are not supported. | |
| Foundation Model APIs (pay-per-token) rate limits | Per workspace | If the following limits are insufficient for your use case, Databricks recommends using provisioned throughput. - Llama 3.1 70B Instruct has a limit of 2 queries per second and 1200 queries per hour. - Llama 3.1 405B Instruct has a limit of 1 query per second and 1200 queries per hour. - The DBRX Instruct model has a limit of 1 query per second. - Mixtral-8x 7B Instruct has a default rate limit of 2 queries per second. - GTE Large (En) has a rate limit of 150 queries per second - BGE Large (En) has a rate limit of 600 queries per second. |
| Foundation Model APIs (provisioned throughput) rate limits | Per workspace | 200 |
Additional limitations exist:
- If your workspace is deployed in a region that supports model serving but is served by a control plane in an unsupported region, the workspace does not support model serving. If you attempt to use model serving in such a workspace, you will see in an error message stating that your workspace is not supported. Reach out to your Azure Databricks account team for more information.
- Model Serving does not support init scripts.
- By default, Model Serving does not support PrivateLink to external endpoints. Support for this functionality is evaluated and implemented on a per region basis. Reach out to your Azure Databricks account team for more information.
- Model Serving does not provide security patches to existing model images because of the risk of destabilization to production deployments. A new model image created from a new model version will contain the latest patches. Reach out to your Databricks account team for more information.
| Feature | Granularity | Limit |
|---|---|---|
| Payload size | Per request | 16 MB. For endpoints serving foundation models or external models the limit is 4 MB. |
| Queries per second (QPS) | Per workspace | 200 QPS. Can be increased to 3000 or more by reaching out to your Databricks account team. |
| Model execution duration | Per request | 120 seconds |
| CPU endpoint model memory usage | Per endpoint | 4GB |
| GPU endpoint model memory usage | Per endpoint | Greater than or equal to assigned GPU memory, depends on the GPU workload size |
| Provisioned concurrency | Per model and per workspace | 200 concurrency. Can be increased by reaching out to your Databricks account. |
| Overhead latency | Per request | Less than 50 milliseconds |
| Foundation Model APIs (pay-per-token) rate limits | Per workspace | If the following limits are insufficient for your use case, Databricks recommends using provisioned throughput. - Llama 3.1 70B Instruct has a limit of 2 queries per second and 1200 queries per hour. - Llama 3.1 405B Instruct has a limit of 1 query per second and 1200 queries per hour. - The DBRX Instruct model has a limit of 1 query per second. - Mixtral-8x 7B Instruct has a default rate limit of 2 queries per second. - GTE Large (En) has a rate limit of 150 queries per second - BGE Large (En) has a rate limit of 600 queries per second. |
| Foundation Model APIs (provisioned throughput) rate limits | Per workspace | Same as Model Serving QPS limit listed above. |
Model Serving endpoints are protected by access control and respect networking-related ingress rules configured on the workspace, like IP allowlists and Private Link.
Azure Private Link is only supported for model serving endpoints that use provisioned throughput or endpoints that serve custom models.
Additional limitations exist as well:
- It is possible for a workspace to be deployed in a supported region, but be served by a control plane in a different region. These workspaces do not support Model Serving and result in an error message saying that your workspace is not supported. Reach out to your Azure Databricks account team for more information.
- Model Serving does not support init scripts.
Networking and security limitations
- Model Serving endpoints are protected by access control and respect networking-related ingress rules configured on the workspace, like IP allowlists and Private Link.
- Private connectivity (such as Azure Private Link) is only supported for model serving endpoints that use provisioned throughput or endpoints that serve custom models.
- By default, Model Serving does not support Private Link to external endpoints (like, Azure OpenAI). Support for this functionality is evaluated and implemented on a per-region basis. Reach out to your Azure Databricks account team for more information.
- Model Serving does not provide security patches to existing model images because of the risk of destabilization to production deployments. A new model image created from a new model version will contain the latest patches. Reach out to your Databricks account team for more information.
Foundation Model APIs limits
Note
As part of providing the Foundation Model APIs, Databricks might process your data outside of the region where your data originated, but not outside of the relevant geographical location.
For both pay-per-token and provisioned throughput workloads:
- Only workspace admins can change the governance settings, such as rate limits for Foundation Model APIs endpoints. To change rate limits use the following steps:
- Open the Serving UI in your workspace to see your serving endpoints.
- From the kebab menu on the Foundation Model APIs endpoint you want to edit, select View details.
- From the kebab menu on the upper-right side of the endpoints details page, select Change rate limit.
- The GTE Large (En) embedding models do not generate normalized embeddings.
Pay-per-token limits
The following are limits relevant to Foundation Model APIs pay-per-token workloads:
- Pay-per-token workloads are not HIPAA or compliance security profile compliant.
- GTE Large (En) and Meta Llama 3.1 70B Instruct models are available in pay-per-token EU and US supported regions.
- The following pay-per-token models are supported only in the Foundation Model APIs pay-per-token supported US regions:
- Meta Llama 3.1 405B Instruct
- DBRX Instruct
- Mixtral-8x7B Instruct
- BGE Large (En)
- If your workspace is in a Model Serving region but not a U.S. or EU region, your workspace must be enabled for cross-Geo data processing. When enabled, your pay-per-token workload is routed to the U.S. Databricks Geo. To see which geographic regions process pay-per-token workloads, see Databricks Designated Services.
Provisioned throughput limits
The following are limits relevant to Foundation Model APIs provisioned throughput workloads:
- Provisioned throughput supports the HIPAA compliance profile and is recommended for workloads that require compliance certifications.
- To use the DBRX model architecture for a provisioned throughput workload, your serving endpoint must be in one of the following regions:
eastuseastus2westuscentraluswesteuropenortheuropeaustraliaeastcanadacentralbrazilsouth
- The following table shows the region availability of the supported Meta Llama 3.1 and 3.2 models. See Deploy fine-tuned foundation models for guidance on how to deploy fine-tuned models.
| Meta Llama model variant | Regions |
|---|---|
| meta-llama/Llama-3.1-8B | - centralus- eastus- eastus2- northcentralus- westus- westus2 |
| meta-llama/Llama-3.1-8B-Instruct | - centralus- eastus- eastus2- northcentralus- westus- westus2 |
| meta-llama/Llama-3.1-70B | - centralus- eastus- eastus2- northcentralus- westus- westus2 |
| meta-llama/Llama-3.1-70B-Instruct | - centralus- eastus- eastus2- northcentralus- westus- westus2 |
| meta-llama/Llama-3.1-405B | - centralus- eastus- eastus2- northcentralus- westus- westus2 |
| meta-llama/Llama-3.1-405B-Instruct | - centralus- eastus- eastus2- northcentralus- westus- westus2 |
| meta-llama/Llama-3.2-1B | - centralus- eastus- eastus2- northcentralus- westus- westus2 |
| meta-llama/Llama-3.2-1B-Instruct | - centralus- eastus- eastus2- northcentralus- westus- westus2 |
| meta-llama/Llama-3.2-3B | - centralus- eastus- eastus2- northcentralus- westus- westus2 |
| meta-llama/Llama-3.2-3B-Instruct | - centralus- eastus- eastus2- northcentralus- westus- westus2 |
Region availability
Note
If you require an endpoint in an unsupported region, reach out to your Azure Databricks account team.
If your workspace is deployed in a region that supports model serving but is served by a control plane in an unsupported region, the workspace does not support model serving. If you attempt to use model serving in such a workspace, you will see in an error message stating that your workspace is not supported. Reach out to your Azure Databricks account team for more information.
For more information on regional availability of features, see Model serving regional availability.