Azure ML HuggingFace custom model

Question

Azure ML HuggingFace custom model

Adrien O'Hana 15

Hello,

I see now in preview we can select models from the HuggingFace Hub and deploy them on an Azure ML endpoint.

What if my model is a custom model ? I see two options

A) Ask Huggingface to add it to their model list (if they start doing that for everyone the catalog will become a huge mess)
B) Deploy on Azure ML via the HuggingFace Hub (only CPU and East US for now, apparent separate subscription)

What I'm looking for is a tutorial/method to deploy a huggingface hub custom model on an azure endpoint, selecting my own compute instance (not via the huggingface hub).

Ramr-msft 17,826 Reputation points

2023-11-20T11:02:09.2133333+00:00

Thanks for the question, You’re correct that you can deploy models from the HuggingFace Hub to an Azure ML endpoint. However, if you have a custom model that’s not in the HuggingFace Hub, you can still deploy it to an Azure ML endpoint. Here is the Azure Machine Learning documentation for more detailed instructions.
Adrien O'Hana 15 Reputation points

2023-11-21T17:04:30.0266667+00:00

That link is specifically the problem, I don't want to deploy a model that's part of the proposed list, I want to deploy my own model that is already on huggingface, and I don't want to do this via the huggingface endpoints hosted on azure but rather via the Azure Endpoints directly. Is there any tutorial or example, or is something being developed to facilitate such a deployment ?

In my case it's a transformer based classifier that I have fined-tuned and pushed on the huggingface hub.
Ramr-msft 17,826 Reputation points

2023-11-27T03:39:24.2566667+00:00

Thanks for the details. I have forwaded to the team, will be supported in the near future.
Adrien O'Hana 15 Reputation points

2024-04-01T12:33:38.6866667+00:00

any update ?
Arthuro 20 Reputation points

2024-04-25T14:50:36.74+00:00

Someone managed to get a custom model running on Azure?
Vincent Nguyen 0 Reputation points

2024-06-06T10:51:33.41+00:00

Is there any way to do it? The model catalog is a huge list, but unfortunately, models that we are interested in are not on the list yet.

1 answer

Your answer

Ramr-msft 17,826 Reputation points

2023-11-20T11:02:09.2133333+00:00

Thanks for the question, You’re correct that you can deploy models from the HuggingFace Hub to an Azure ML endpoint. However, if you have a custom model that’s not in the HuggingFace Hub, you can still deploy it to an Azure ML endpoint. Here is the Azure Machine Learning documentation for more detailed instructions.
Adrien O'Hana 15 Reputation points

2023-11-21T17:04:30.0266667+00:00

That link is specifically the problem, I don't want to deploy a model that's part of the proposed list, I want to deploy my own model that is already on huggingface, and I don't want to do this via the huggingface endpoints hosted on azure but rather via the Azure Endpoints directly. Is there any tutorial or example, or is something being developed to facilitate such a deployment ?

In my case it's a transformer based classifier that I have fined-tuned and pushed on the huggingface hub.
Ramr-msft 17,826 Reputation points

2023-11-27T03:39:24.2566667+00:00

Thanks for the details. I have forwaded to the team, will be supported in the near future.
Adrien O'Hana 15 Reputation points

2024-04-01T12:33:38.6866667+00:00

any update ?
Arthuro 20 Reputation points

2024-04-25T14:50:36.74+00:00

Someone managed to get a custom model running on Azure?
Vincent Nguyen 0 Reputation points

2024-06-06T10:51:33.41+00:00

Is there any way to do it? The model catalog is a huge list, but unfortunately, models that we are interested in are not on the list yet.

Answer 1

Arthuro 20

I managed to deploy a custom model from hugging face but it will work with basically every other model as well. You can follow these steps: Notebook

Share via

Azure ML HuggingFace custom model

1 answer

Your answer