The canonical schema for the WorkflowRun CRD. Source of truth: plnt/operators/crds/workflowrun.yaml.
Which runtime chart to deploy.
| Value | Chart |
|---|
vllm | plnt/charts/vllm-runtime |
tgi | plnt/charts/tgi-runtime |
trt-llm | plnt/charts/trt-llm-runtime |
sglang | plnt/charts/sglang-runtime |
| Field | Type | Required | Notes |
|---|
name | string | ✅ | Canonical model name (e.g. meta-llama/Llama-3.1-70B-Instruct). |
storageUri | string | ✅ | Where weights live: s3://, gs://, oci://, hf://. |
quantization | string | | fp16 | fp8 | awq-4bit | gptq-4bit. Default fp16. |
| Field | Type | Required | Notes |
|---|
gpu | integer ≥ 1 | ✅ | GPU count per replica. |
gpuClass | string | ✅ | Kubernetes GPU resource class (nvidia.com/h100, nvidia.com/a100, nvidia.com/a10). |
memoryGiB | integer ≥ 1 | | Container memory request. |
| Field | Type | Default | Notes |
|---|
min | integer ≥ 0 | 1 | HPA lower bound. |
max | integer ≥ 1 | 3 | HPA upper bound. |
| Field | Type | Notes |
|---|
targetTokensPerSecond | integer ≥ 1 | Custom-metric HPA target. Requires prometheus-adapter mapped to plnt_tokens_per_sec_per_gpu. |
| Field | Type | Default | Notes |
|---|
trafficPercent | integer 1..100 | 5 | Envoy VirtualService weight for the canary release. |
smokeTest.prompts | integer ≥ 1 | 10 | How many test prompts to send. |
smokeTest.ttftBudgetMs | integer ≥ 1 | | Max TTFT p50 for the canary to pass. |
smokeTest.tpotBudgetMs | integer ≥ 1 | | Max TPOT p50 for the canary to pass. |
| Field | Type | Values |
|---|
phase | string | Pending | Validating | PullingWeights | Deploying | SmokeTesting | Promoting | Ready | RolledBack | Failed |
endpoint | string | In-cluster URL once Ready. |
workflowRunId | string | Temporal workflow run id. |
lastError | string | Message from the most recent failed activity. |
observedGeneration | integer | Matches metadata.generation when reconciled. |
kubectl get workflowrun shows:
NAME RUNTIME GPU PHASE ENDPOINT AGE
llama-3.1-70b-instruct vllm nvidia.com/h100 Ready ... 3h
kubectl apply -f plnt/operators/crds/workflowrun.yaml
kubectl apply -f examples/llama-3.1-70b-instruct.yaml
