Support for additional formats for azurerm_cognitive_deployment

[DanHyland]

Is there an existing issue for this?

• I have searched the existing issues

Community Note

• Please vote on this issue by adding a 👍 reaction to the original issue to help the community and maintainers prioritize this request
• Please do not leave comments along the lines of "+1", "me too" or "any updates", they generate extra noise for issue followers and do not help prioritize the request
• If you are interested in working on this issue or have submitted a pull request, please leave a comment and review the contribution guide to help.

Description

Today, only format of type "OpenAI" is supported for azurerm_cognitive_deployments (https://registry.terraform.io/providers/hashicorp/azurerm/latest/docs/resources/cognitive_deployment#format-1). This makes attempting to manage non OpenAI model deployments impossible. My goal is to manage a Azure AI model catalog MaaS deployment for DeepSeek (and other models) via the azurerm_cognitive_deployment. Before anyone already mentions that this can be done via azapi, I do not view azapi as a true TF provider in that true lifecycle support does not exist. We have seen large inconsistencies in formats of API calls, and outputs in azapi. We limit its use in our environment as much as possible.

When attempting to do an import, I get the proper diff, but the limitation in the format input makes it impossible to manage.

Diff and plan:

`-/+ resource "azurerm_cognitive_deployment" "foobar" {
cognitive_account_id = "/subscriptions/xxxxxxxxxxxxxxxx/resourceGroups/xxxxxxxxx/providers/Microsoft.CognitiveServices/accounts/xxxxxxxxx"
- dynamic_throttling_enabled = false -> null
~ id = "/subscriptions/xxxxxxxxx/resourceGroups/xxxxxxxxxx/providers/Microsoft.CognitiveServices/accounts/xxxxxxxxxx/deployments/DeepSeek-R1" -> (known after apply)
~ name = "DeepSeek-R1" -> "testing" # forces replacement
version_upgrade_option = "OnceNewDefaultVersionAvailable"

  ~ model {
      ~ format  = "DeepSeek" -> "OpenAI" # forces replacement
      ~ name    = "DeepSeek-R1" -> "test" # forces replacement
      ~ version = "1" -> "4"
    }

  ~ sku {
      ~ capacity = 1 -> 50
        name     = "GlobalStandard"
    }
}

Plan: 1 to import, 2 to add, 0 to change, 1 to destroy.

────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────

Note: You didn't use the -out option to save this plan, so Terraform can't guarantee to take exactly these actions if you run "terraform apply" now.
------New TF Plan with updated format to DeepSeek)
╷
│ Error: expected model.0.format to be one of ["OpenAI"], got DeepSeek
│
│ with azurerm_cognitive_deployment.foobar,
│ on main.tf line 83, in resource "azurerm_cognitive_deployment" "foobar":
│ 83: format = "DeepSeek"
│
╵`

New or Affected Resource(s)/Data Source(s)

azurerm_cognitive_deployment

Potential Terraform Configuration

resource "azurerm_cognitive_deployment" "foobar" {
  name = "DeepSeek-R1"
  cognitive_account_id = azurerm_ai_services.account.id
  model {
    format = "DeepSeek"
    name = "DeepSeek-R1"
    version = "1"
  }
  sku {
    name = "GlobalStandard"
    capacity = "1"

  }
}

References

No response

[sinbai]

Hi @DanHyland thanks for opening this issue. I would like to explain that Terraform manages Azure Resources through Azure Rest API. After verifying the behavior of the API, it seems that the DeepSeek format is not supported by this API. Therefore, Terraform could not support it until the API supports it. Thank you for your understanding

[DanHyland]

@sinbai does that mean models from the Ai Model Catalog should have a separate distinct resource instead of the azurerm_cognitive_deployment?

[sinbai]

As far as I know, the only resource in the Terraform AzureRM provider that supports this API is azurerm_cognitive_deployment. Since I am not an expert on the API, it is recommended to contact Microsoft Support for more information about the feature you mentioned.

[josh-barker]

Hi @rcskosir @sinbai , I removed the validation on format, recompiled the provider locally, and was able to deploy the Deep Seek and grok model.

Code Snippet from internal/services/cognitive/cognitive_deployment_resource.go:

"model": {
		Type:     pluginsdk.TypeList,
		Required: true,
		ForceNew: true,
		MaxItems: 1,
		Elem: &pluginsdk.Resource{
			Schema: map[string]*pluginsdk.Schema{
				"format": {
					Type:     pluginsdk.TypeString,
					Required: true,
					ForceNew: true,
					// ValidateFunc: validation.StringInSlice([]string{
					//	"OpenAI",
					//	"Cohere",
					//}, false),
				},
...

resource "azurerm_ai_services" "example" {
  name                = "example"
  location            = "australiaeast"
  resource_group_name = "example-rg"
  sku_name            = "S0"

  custom_subdomain_name = "example"
}

# azureml://registries/azureml-deepseek/models/DeepSeek-R1/versions/1
resource "azurerm_cognitive_deployment" "deepseek" {
  name                   = "deepseek"
  version_upgrade_option = "NoAutoUpgrade"

  cognitive_account_id = azurerm_ai_services.example.id

  sku {
    name = "GlobalStandard"
  }
  model {
    name             = "DeepSeek-R1"
    version          = "1"
    format           = "DeepSeek"
  }
}

# azureml://registries/azureml-xai/models/grok-3-mini/versions/1
resource "azurerm_cognitive_deployment" "grok" {
  name                   = "grok"
  version_upgrade_option = "NoAutoUpgrade"

  cognitive_account_id = azurerm_ai_services.example.id

  sku {
    name = "GlobalStandard"
  }
  model {
    name             = "grok-3-mini"
    version          = "1"
    format           = "xAI" # case is sensitive
  }
}

I'm happy to submit a PR to remove the validation?

[sinbai]

If you could add the test case at the same time, and the test passes. I assume adding DeepSeek to ValidateFunc instead of removing the validation. Thank you for your contribution!

[DanHyland]

@sinbai, curious to why the ValidateFunc even exists when historically Terraform isn't really used for API input validation. Making every model provider need a Terraform change for each model format seems wrong.

[sreallymatt]

Validation was removed in #30276

[github-actions]

I'm going to lock this issue because it has been closed for 30 days ⏳. This helps our maintainers find and focus on the active issues.
If you have found a problem that seems similar to this, please open a new issue and complete the issue template so we can capture all the details necessary to investigate further.