Machine Learning: Deployment Failures with InternalServerError at 0% (Marketplace model)

Problem

A managed online deployment of the Marketplace model Cohere embed-v-4-0 (registry azureml-cohere) suddenly stopped serving. It had been running without problems before. Every redeploy attempt now fails at 0% with ResourceOperationFailure / InternalServerError. There is no inner error, and get-logs returns BadRequest, so no container ever seems to start.

What I tried

• Standard_NV36ads_A10_v5 in Sweden Central: failed
• Standard_NV36ads_A10_v5 in West Europe (new hub and project): failed
• Standard_NC24ads_A100_v4: failed
• Other models in the same workspace (CPU catalog model, custom container on A100): succeeded

Already ruled out

• Quota is available
• No resource locks
• Required resource providers are registered (including Microsoft.SaaS and Microsoft.Marketplace)
• No managed network isolation
• The Marketplace subscription for the offer shows Subscribed