fix(azure): use azure.mgmt.network.SubResource for IPConfiguration subnet (was compute.SubResource) — unbreaks VM creation

[murraju]

Problem

sky launch --cloud azure fails at VM creation with Object of type SubResource is not JSON serializable when resolved against current/new-generation azure-mgmt-network / azure-mgmt-compute. This blocks every Azure launch end-to-end.

Root cause

sky/provision/azure/instance.py::_create_network_interface builds a network IPConfiguration whose subnet is a compute.SubResource (a model from the azure-mgmt-compute package):

network = azure.azure_mgmt_models('network')
compute = azure.azure_mgmt_models('compute')
ip_config = network.IPConfiguration(
    name=...,
    subnet=compute.SubResource(id=provider_config['subnet']),   # cross-package model
    ...)
network_client.network_interfaces.begin_create_or_update(
    parameters=network.NetworkInterface(..., ip_configurations=[ip_config], ...))

The network SDK serializes the NetworkInterface body. With new-gen azure-mgmt-network (>=~25.x; reproducible with 31.0.0b1) serialization goes through stdlib json, which has no encoder for a compute.SubResource nested in a network model → TypeError. It was latent under the older msrest-era SDKs. azure-mgmt-compute==33.0.0 + azure-mgmt-network==31.0.0b1 (what a fresh pip install skypilot[azure] resolves today) expose it. azure-cli version is irrelevant (confirmed: pinning azure-cli<2.70 → 2.69.0 does not fix it).

Fix

Use the network SDK's own SubResource — same package as IPConfiguration, serializes natively. network is already bound in the function, so this is a 2-line, drop-in change (both the internal-IP and public-IP branches).

network.SubResource exists in every azure-mgmt-network release sky supports, is constructed identically (id=...), and round-trips through the serializer (as_dict() → {\id\: \...\}).

Reproduction

uv tool install "skypilot[azure]>=0.12,<0.13" --prerelease=allow   # resolves azure-mgmt-network 31.0.0b1
az login
sky launch --cloud azure --gpus T4:1 -y -- echo ok
# -> Failed to create instances: [TypeError] Object of type SubResource is not JSON serializable

Reproduces across every Azure region. Fails pre-VM-creation (no compute spend). After the patch, the VM is created and the cluster comes UP.

Verification

Live: with the patch applied, sky launch --cloud azure --gpus T4:1 provisions a Standard_NC4as_T4_v3; the NIC (the exact object that failed) is created; cluster reaches UP. (Azure eastus, 2026-06-28.)

Notes

• master carries the same compute.SubResource usage (unfixed as of 2026-06-28).
• Independent of azure-cli; skypilot[azure] pulls azure-mgmt-* with floors but no ceilings, so a fresh resolve gets the new-gen SDKs that expose this.
• A downstream tool (mpl) carries the equivalent transform as a post-install patch today; converging upstream lets it retire that divergence.

[gemini-code-assist]

Code Review

This pull request updates sky/provision/azure/instance.py to use network.SubResource instead of compute.SubResource when defining the subnet for both private and public IP configurations. I have no feedback to provide as there are no review comments to evaluate.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.