Skip to main content
This feature is in beta. You can share feedback on the Together AI Discord or by contacting support.
If you already run a LoRA-enabled dedicated endpoint for an adapter’s base model, you can attach the adapter to that endpoint instead of deploying it on its own hardware. One endpoint can serve multiple LoRA adapters that share the same base model, and inference requests select among them by model name. Use this when you want to:
  • Serve several adapters trained against the same base model from one endpoint.
  • Avoid paying for separate hardware per adapter.
  • Change which adapter handles traffic without redeploying.

​
Requirements

The target endpoint must be a private dedicated endpoint with LoRA enabled, running a base model that’s compatible with the adapter. The adapter’s base model must match the endpoint’s model, and both the adapter and the endpoint must be owned by the same account. If you don’t have a LoRA-enabled endpoint for the base model yet, create one first.

​
Attach an adapter from the dashboard

1

Open the adapter

Go to My Models and open the adapter you want to attach.
2

Select Deploy adapter to endpoint

From the actions menu on the adapter, select Deploy adapter to endpoint. A dialog lists running endpoints that are LoRA-enabled and match the adapter’s base model.
3

Select an endpoint and deploy

Select a compatible endpoint, then select Deploy adapter. If no endpoints appear, create a LoRA-enabled dedicated endpoint for the base model first.

​
Attach an adapter from the API

The API uses two different identifiers for the base endpoint:
  • <ENDPOINT_ID> is the endpoint’s id, a system-generated handle like endpoint-e6c6b82f-90f7-45b7-af39-3ca3b51d08c1. It goes in the URL path and as the first argument to the SDK and CLI calls.
  • <ENDPOINT_NAME> is the endpoint’s name, an auto-generated namespaced string like tester/Qwen/Qwen3.5-9B-FP8-bb04c904. It’s the prefix of the combined model_id. This isn’t the display name you set at creation; read it from the endpoint’s name field.
Both are returned when you create an endpoint and by the list and get operations. Retrieve them for an existing endpoint with the same calls used to manage endpoints: 
together endpoints list
Each adapter is identified by a combined model_id in the form endpoint_name:adapter_model_name, where adapter_model_name is the model_name returned when you uploaded the adapter. The endpoint_name prefix must match the endpoint resolved from <ENDPOINT_ID>. 
from together import Together

client = Together()

result = client.endpoints.adapters.add(
    "<ENDPOINT_ID>",
    model_id="<ENDPOINT_NAME>:<ADAPTER_MODEL_NAME>",
)
print(result.api_model_id)
The response echoes the bound model_id: 
{ "model_id": "<ENDPOINT_NAME>:<ADAPTER_MODEL_NAME>" }

​
List attached adapters

adapters = client.endpoints.adapters.list("<ENDPOINT_ID>")
for adapter in adapters.data or []:
    print(adapter.api_model_id, adapter.adapter_name, adapter.endpoint_name)
The list alias ls is also accepted by the CLI. Add --json to print the raw API response: 
{
  "object": "list",
  "data": [
    {
      "model_id": "<ENDPOINT_NAME>:<ADAPTER_MODEL_NAME>",
      "adapter_name": "<ADAPTER_MODEL_NAME>",
      "endpoint_name": "<ENDPOINT_NAME>"
    }
  ]
}
In the Python SDK, the combined identifier is exposed on the response object as api_model_id, while the JSON field returned by the API is model_id. The TypeScript SDK and the raw API both use model_id.

​
Remove an adapter

Removing an adapter detaches it from the endpoint. The uploaded adapter stays in your account and can be re-attached or deployed elsewhere. 
client.endpoints.adapters.remove(
    "<ENDPOINT_ID>",
    model_id="<ENDPOINT_NAME>:<ADAPTER_MODEL_NAME>",
)
The CLI accepts delete and rm as aliases for remove.

​
Run inference

Once the adapter is attached, send inference requests using the adapter model name as the model parameter. Requests are routed to the endpoint automatically. You can also pass the full endpoint_name:adapter_model_name form. 
response = client.chat.completions.create(
    model="<ADAPTER_MODEL_NAME>",
    messages=[{"role": "user", "content": "Hello!"}],
    max_tokens=128,
)
print(response.choices[0].message.content)

​
Serve multiple adapters

A LoRA-enabled endpoint can hold several adapters at once, as long as they all share the endpoint’s base model. Attach each one with a separate add call, then route requests to whichever adapter you need by passing its model name. All attached adapters share the endpoint’s hardware and replicas, so concurrent traffic across adapters draws on the same capacity rather than scaling independently. If load grows, raise the endpoint’s replica count through autoscaling.

​
Troubleshooting

  • “model_id must be in format ‘endpoint_model_name:adapter_name’”: The model_id field must contain exactly one : separator with non-empty parts.
  • “endpoint name in model_id does not match endpoint”: The endpoint name prefix in model_id doesn’t match the endpoint resolved from the URL path.
  • “Could not find model”: The adapter model name after the : doesn’t exist or isn’t owned by your account.
  • “No endpoint with this id exists”: The endpoint ID in the URL doesn’t exist or isn’t owned by your account.
  • “endpoint does not have LoRA enabled”: The target endpoint wasn’t created with LoRA support enabled.
  • “adapter base model is not compatible with endpoint model”: The adapter’s base model doesn’t match the model running on the endpoint.
  • “adapter is already bound to endpoint”: The adapter is already attached to a different endpoint. Remove it first before attaching it to a new one. Re-attaching to the same endpoint is allowed.
  • “adapter is not bound to endpoint”: When removing, the adapter isn’t currently routed to this endpoint.
  • “new_lora_model must be a dedicated endpoint”: Only dedicated endpoints, not serverless, can be used as adapter targets.
  • “new_lora_model cannot be a public dedicated endpoint”: Only private endpoints can be used as adapter targets.