Add blobfuse support for AKS

[vofish]

Add Azure Blob Storage (blobfuse2) support to the Kubernetes AI inference benchmark

Enables the kubernetes_ai_inference benchmark to load model weights from Azure Blob Storage via the Azure Blob CSI driver (blobfuse2), as an Azure equivalent of the GCS Fuse.

• AksAutomaticCluster._Create: unconditionally pass --enable-blob-driver at cluster creation, matching GKE Autopilot (where the GCS Fuse CSI driver is on by default).
• _ApplyBlobFusePVC: renders and applies blobfuse_pv_pvc.yaml.j2 with the Secret, PV, and PVC for the blob CSI driver.

wg_serving_inference_server.py:

• Add three new flags: --k8s_inference_server_blobstorage_bucket, --k8s_inference_server_blobstorage_account, --k8s_inference_server_blobstorage_resource_group.
• _GetStorageType: return 'blobfuse' when catalog_components contains blobfuse.
• _Create: branch to _ApplyBlobFusePVC() when blobfuse is in catalog_components.
• _ResolveBlobStorageAccount: resolves credentials

Command used to run:

./pkb.py --benchmarks=kubernetes_ai_inference --cloud=Azure --config_override=kubernetes_ai_inference.container_cluster.type='Auto' \
--config_override=kubernetes_ai_inference.container_cluster.inference_server.model_server='vllm' \
--config_override=kubernetes_ai_inference.container_cluster.inference_server.model_name='llama3-8b' \
--config_override=kubernetes_ai_inference.container_cluster.inference_server.extra_deployment_args.pvc-name='gcs-fuse-csi-static-pvc' \
--config_override=kubernetes_ai_inference.container_cluster.inference_server.catalog_components='1-H100,blobfuse,xla-cache' \
--config_override=kubernetes_ai_inference.container_cluster.inference_server.deployment_timeout=2500 \
--config_override=kubernetes_ai_inference.container_cluster.inference_server.extra_deployment_args.model-path='vllm_models/llama3-8b-hf' \
--k8s_ai_inference_hf_token="${HF_TOKEN}" --k8s_ai_inference_request_rate=5 \
--k8s_inference_server_image_repo=us-docker.pkg.dev/p3rf-gke/public \
--metadata=cloud:Azure --zone=centralus --k8s_inference_server_blobstorage_bucket='gke-vllm-test' \
--wg_serving_repo_url='https://github.com/vofish/wg-serving' --wg_serving_repo_branch='azure-test'

It requires gke-vllm-test container storage to be created and the llama3-8b model to be uploaded to the vllm_models/llama3-8b-hf path prior to the run.

[hubatish]

Oof this looks quite ugly. We might finally be hitting some of the limit of the current wg_serving_inference_server.py if cloud implementation. It might be somewhat necessary currently (because idk what happens if you put this in root, whether that causes circular dependencies or errors with LoadProvider when running on eg GCP) but is not correct. We should instead have separate files in each provider with cloud specific code. I'll also forward this feedback to Xia internally.

[vofish]

Refactored and moved functionality into a new file perfkitbenchmarker/providers/azure/azure_blob_csi_mount.py

[hubatish]

approving but as mentioned I think this may fail pytype checks

[hubatish]

seems like odd gcs-fuse specific metadata?

[vofish]

This was updated as follows: name: {{pvc_name}}

[hubatish]

Yes this is basically what I was asking for. It is kubernetes_ai_inference specific atm but it is reasonable to think we'd use this in other benchmarks (none of the setup actually seems that AI specific).

Some additional changes to line up with this:

• Consider also moving the yaml to data/container/azure rather than data/container/kubernetes/ai_infernece.
• Add an empty version of this function to container_cluster.py (surprised the code as isn't failing pytype checks). Call it ApplyFusePVC rather than ApplyBlobFusePVC so it's more cloud agnostic.

[vofish]

Done:

• Jinja template was moved to data/container/azure
• The function was renamed and moved to container_cluster.py