[Lambda] Fix missing private ip

[bend-works]

Handle the setup failure when single-node A6000 VMs on Lambda Cloud do not return a private IP: in this case, cluster info now returns 127.0.0.1.

Tested: sky launch -c private-ip-bug --cloud lambda --gpus A6000 is successful.

[bend-works]

fixes #4634

[Michaelvll]

Thanks @bend-works for submitting the fix! Please find a quick comment below. : )

[Michaelvll]

How about we just use a argument called single_node for the function, so we don't have to call the cloud API multiple times for initializing the cluster_info for better performance? We can set single_node=(len(running_instances) > 1)

Suggested change

	def _get_private_ip(cluster_name_on_cloud: str,
	instance_info: Dict[str, Any]) -> str:
	if len(_filter_instances(cluster_name_on_cloud, None)) == 1:
	return instance_info.get('private_ip', '127.0.0.1')
	return instance_info['private_ip']
	def _get_private_ip(instance_info: Dict[str, Any], single_node: bool) -> str:
	private_ip = instance_info.get('private_ip')
	if private_ip is None:
	if single_node:
	# The Lambda cloud API may return an instance info without
	# private IP. It does not align with their docs, but we still
	# allow single-node cluster to proceed with the provisioning, by using
	# 127.0.0.1, as private IP is not critical for single-node case.
	return '127.0.0.1'
	msg = f'Failed to retrieve private IP for instance {instance_info}.'
	logger.error(msg)
	raise RuntimeError(msg)
	return private_ip

[bend-works]

I like it and it’s easier to unit test! I introduced the extra API call because I wasn’t sure about the assumption that all nodes of a cluster would be running at the same time but you know better.

[Michaelvll]

Thanks @bend-works! This is great! Merging now.