[Serving Catalog] Add llama3.1-405b vLLM GKE support with LWS #11
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,6 @@ | ||
| # kustomization.yaml | ||
| apiVersion: kustomize.config.k8s.io/v1beta1 | ||
| kind: Kustomization | ||
|
|
||
| resources: | ||
| - leaderworkerset.yaml |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,21 @@ | ||
| apiVersion: leaderworkerset.x-k8s.io/v1 | ||
| kind: LeaderWorkerSet | ||
| metadata: | ||
| name: multihost-base | ||
| spec: | ||
| leaderWorkerTemplate: | ||
| restartPolicy: RecreateGroupOnPodRestart | ||
| leaderTemplate: | ||
| metadata: | ||
| labels: | ||
| app: multihost-inference-server | ||
| spec: | ||
| containers: | ||
| - name: multihost-leader-base | ||
| workerTemplate: | ||
| metadata: | ||
| labels: | ||
| app: multihost-inference-server | ||
| spec: | ||
| containers: | ||
| - name: multihost-worker-base |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,23 @@ | ||
| # Kustomize lacks easy support for strategic patch merge for CRDs | ||
| # Leader | ||
| - op: add | ||
| path: /spec/leaderWorkerTemplate/leaderTemplate/spec/nodeSelector | ||
| value: | ||
| cloud.google.com/gke-accelerator: nvidia-h100-80gb | ||
| - op: add | ||
| path: /spec/leaderWorkerTemplate/leaderTemplate/spec/containers/0/resources | ||
| value: | ||
| limits: | ||
| nvidia.com/gpu: "8" | ||
| memory: 1770Gi | ||
| ephemeral-storage: 800Gi | ||
| requests: | ||
| ephemeral-storage: 800Gi | ||
| cpu: 125 | ||
| # Worker | ||
| - op: copy | ||
| from: /spec/leaderWorkerTemplate/leaderTemplate/spec/nodeSelector | ||
| path: /spec/leaderWorkerTemplate/workerTemplate/spec/nodeSelector | ||
| - op: copy | ||
| from: /spec/leaderWorkerTemplate/leaderTemplate/spec/containers/0/resources | ||
| path: /spec/leaderWorkerTemplate/workerTemplate/spec/containers/0/resources |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,8 @@ | ||
| # kustomization.yaml | ||
| apiVersion: kustomize.config.k8s.io/v1alpha1 | ||
| kind: Component | ||
|
|
||
| patches: | ||
| - target: | ||
| kind: LeaderWorkerSet | ||
| path: h100.patch.yaml |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,31 @@ | ||
| apiVersion: v1 | ||
| kind: ConfigMap | ||
| metadata: | ||
| name: vllm-multihost-config | ||
| data: | ||
| ray_status_check.sh: |- | ||
| #!/usr/bin/bash -x | ||
| # Verify ray head status | ||
| until ray status --address $LWS_LEADER_ADDRESS:6380; do | ||
| sleep 5; | ||
| done | ||
| entrypoint.sh: |- | ||
| #!/usr/bin/bash -x | ||
| # Launch vLLM Inference server | ||
|
|
||
| export PYTHONPATH="/workspace/" | ||
| if [[ -n "$1" ]]; then | ||
| ray start --head --port=6380 | ||
| num_accelerators=`python3 -c 'import ray; ray.init(); print(int(sum([ray.cluster_resources().get("GPU", 0), ray.cluster_resources().get("TPU", 0)])))'` | ||
| total_accelerators=$(($TENSOR_PARALLEL_SIZE * $PIPELINE_PARALLEL_SIZE )) | ||
| until [ $num_accelerators -eq $total_accelerators ]; do | ||
| num_accelerators=`python3 -c 'import ray; ray.init(); print(int(sum([ray.cluster_resources().get("GPU", 0), ray.cluster_resources().get("TPU", 0)])))'` | ||
| sleep 5 | ||
| done | ||
| python3 -m vllm.entrypoints.openai.api_server --port 8080 --model $MODEL_ID --tensor_parallel_size $TENSOR_PARALLEL_SIZE --pipeline_parallel_size $PIPELINE_PARALLEL_SIZE | ||
| else | ||
| until ray start --address="$LWS_LEADER_ADDRESS":6380 --block; do | ||
| sleep 5 | ||
| done | ||
| fi | ||
|
|
||
Edwinhr716 marked this conversation as resolved.
Show resolved
Hide resolved
|
||
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,13 @@ | ||
| # kustomization.yaml | ||
| apiVersion: kustomize.config.k8s.io/v1beta1 | ||
| kind: Kustomization | ||
|
|
||
| resources: | ||
| - ../../base | ||
| - service.yaml | ||
| - configmap.yaml | ||
|
|
||
| patches: | ||
| - path: leaderworkerset.yaml | ||
| target: | ||
| kind: LeaderWorkerSet |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,100 @@ | ||
| apiVersion: leaderworkerset.x-k8s.io/v1 | ||
| kind: LeaderWorkerSet | ||
| metadata: | ||
| name: vllm-multihost-base | ||
| spec: | ||
| leaderWorkerTemplate: | ||
| restartPolicy: RecreateGroupOnPodRestart | ||
| leaderTemplate: | ||
| spec: | ||
| containers: | ||
| - name: inference-server-leader | ||
| image: vllm/vllm-openai:latest | ||
| command: | ||
| - /scripts/entrypoint.sh | ||
| args: ["--head"] | ||
| env: | ||
| - name: HUGGING_FACE_HUB_TOKEN | ||
| valueFrom: | ||
| secretKeyRef: | ||
| name: hf-secret | ||
| key: hf_api_token | ||
| - name: MODEL_ID | ||
| valueFrom: | ||
| configMapKeyRef: | ||
|
There was a problem hiding this comment. Using a configmap is inflating the lws yaml, assuming we can bake the logic in the container, do we still need it? There was a problem hiding this comment. Don't need it, although in practice I liked defining the environment variables shared across leaders/workers in one place. |
||
| name: vllm-multihost-config | ||
| key: model_id | ||
| - name: TENSOR_PARALLEL_SIZE | ||
| valueFrom: | ||
| configMapKeyRef: | ||
| name: vllm-multihost-config | ||
| key: tensor_parallel_size | ||
| - name: PIPELINE_PARALLEL_SIZE | ||
| valueFrom: | ||
| configMapKeyRef: | ||
| name: vllm-multihost-config | ||
| key: pipeline_parallel_size | ||
| volumeMounts: | ||
| - mountPath: "/scripts" | ||
| name: scripts-volume | ||
| readOnly: true | ||
| - mountPath: /dev/shm | ||
| name: dshm | ||
| ports: | ||
| - containerPort: 8080 | ||
| volumes: | ||
| - name: scripts-volume | ||
| configMap: | ||
| defaultMode: 0700 | ||
| name: vllm-multihost-config | ||
| - name: dshm | ||
| emptyDir: | ||
| medium: Memory | ||
| sizeLimit: 30Gi | ||
| workerTemplate: | ||
| spec: | ||
| initContainers: | ||
| - name: ray-head-check | ||
| image: vllm/vllm-openai:latest | ||
| command: | ||
| - /scripts/ray_status_check.sh | ||
| volumeMounts: | ||
| - mountPath: "/scripts" | ||
| name: scripts-volume | ||
| readOnly: true | ||
| containers: | ||
| - name: inference-server-worker | ||
| image: vllm/vllm-openai:latest | ||
| command: | ||
| - /scripts/entrypoint.sh | ||
| env: | ||
| - name: HUGGING_FACE_HUB_TOKEN | ||
| valueFrom: | ||
| secretKeyRef: | ||
| name: hf-secret | ||
| key: hf_api_token | ||
| - name: MODEL_ID | ||
| valueFrom: | ||
| configMapKeyRef: | ||
| name: vllm-multihost-config | ||
| key: model_id | ||
| - name: TENSOR_PARALLEL_SIZE | ||
| valueFrom: | ||
| configMapKeyRef: | ||
| name: vllm-multihost-config | ||
| key: tensor_parallel_size | ||
| volumeMounts: | ||
| - mountPath: "/scripts" | ||
| name: scripts-volume | ||
| readOnly: true | ||
| - mountPath: /dev/shm | ||
| name: dshm | ||
| volumes: | ||
| - name: scripts-volume | ||
| configMap: | ||
| defaultMode: 0700 | ||
| name: vllm-multihost-config | ||
| - name: dshm | ||
| emptyDir: | ||
| medium: Memory | ||
| sizeLimit: 30Gi | ||
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,13 @@ | ||
| apiVersion: v1 | ||
| kind: Service | ||
| metadata: | ||
| name: vllm-leader | ||
| spec: | ||
| ports: | ||
| - name: http | ||
| port: 8080 | ||
| protocol: TCP | ||
| targetPort: 8080 | ||
| selector: | ||
| role: leader | ||
| type: ClusterIP |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,14 @@ | ||
| # Llama3.1-405b | ||
|
|
||
| ## Configuration | ||
| | Kind | Model Server | Model | Provider | Accelerator | | ||
| | --- | --- | --- | --- | --- | | ||
| | Deployment | vLLM | llama3.1-405b | GKE | GPU H100 | | ||
Edwinhr716 marked this conversation as resolved.
Show resolved
Hide resolved
|
||
|
|
||
| ## Usage | ||
|
|
||
| The template can be deployed with the following commands: | ||
|
|
||
| ``` | ||
| kustomize build core/lws/vllm/llama3-405b/gke | kubectl apply -f - | ||
| ``` | ||
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,30 @@ | ||
| # kustomization.yaml | ||
| apiVersion: kustomize.config.k8s.io/v1beta1 | ||
| kind: Kustomization | ||
|
|
||
| resources: | ||
| - ../../base | ||
|
|
||
| components: | ||
| - ../../../components/gke/resources/gpu/8-H100 | ||
|
|
||
| patches: | ||
| - path: patch.yaml | ||
| target: | ||
| kind: LeaderWorkerSet | ||
| options: | ||
| allowNameChange: true | ||
| - target: | ||
| kind: Service | ||
| patch: |- | ||
| - op: replace | ||
| path: /metadata/name | ||
| value: llama3-405b-vllm-service | ||
|
|
||
| configMapGenerator: | ||
| - name: vllm-multihost-config | ||
| behavior: merge | ||
| literals: | ||
| - model_id="meta-llama/Meta-Llama-3.1-405B-Instruct" | ||
| - tensor_parallel_size="8" | ||
| - pipeline_parallel_size="2" |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,15 @@ | ||
| apiVersion: leaderworkerset.x-k8s.io/v1 | ||
| kind: LeaderWorkerSet | ||
| metadata: | ||
| name: llama3-405b-lws | ||
| spec: | ||
| leaderWorkerTemplate: | ||
| size: 2 | ||
| leaderTemplate: | ||
| metadata: | ||
| labels: | ||
| ai.gke.io/model: llama3-405b | ||
| workerTemplate: | ||
| metadata: | ||
| labels: | ||
| ai.gke.io/model: llama3-405b |
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
@jjk-g can we propose to vllm to create containers that embeds this logic? we can create an issue on vllm repo.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
+1 vllm support for creating Ray cluster via their container is preferred
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Issue has been created on the vllm repo vllm-project/vllm#8302
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Since the issue hasn't been addressed, would it be better to merge this PR right now, and modify it later once a multihost vllm image is created?