Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →aws-samples/sagemaker-deploy
Stage any supported open-weight Hugging Face model directly into Amazon S3 and deploy it to an Amazon SageMaker AI real-time endpoint with a compatible vLLM or SGLang Deep Learning Container.
Passed 16 of 19 test cases with the skill, and 6 of 19 without it.
and 13 more in the full report
12 fixed7 passed either way3 still failing0 broke
Tokens, turns and time describe resource use, not a dollar price.
One of 12 test cases the skill fixed. Same model, same prompt, with and without the skill.
Our automated validation pipeline finished running smoke test queries on a newly deployed SageMaker endpoint. What post-validation cleanup command should be executed to avoid ongoing compute charges?
| Checked | Without skill | With skill |
|---|---|---|
| Executes scripts/teardown.py with the --yes flag | Missed | Passed |
The instructions your model reads when the skill loads.
This skill is a model-agnostic deployment contract. The scripts do not select behavior by model name. The agent researches the requested checkpoint, derives its serving requirements, and supplies those decisions as explicit arguments.
The bundled scripts are generic execution primitives, not a model catalog or a policy engine. The agent must interpret current primary sources and choose the checkpoint, engine, image, hardware, topology, context cap, and request shape. A newly supported model should normally require different arguments, not a code change.
Do not grow the scripts with model profiles or model-name branches. Add code
only for a reusable SageMaker capability. If no compatible checkpoint fits a
single SageMaker real-time endpoint, choose another official quantization or
move to the appropriate SageMaker multi-node path, such as HyperPod, instead of
encoding a one-model exception in deploy.py.
Before deriving serving arguments from scratch, look up a known-good configuration for the requested checkpoint in the AWS SageMaker GenAI hosting examples:
https://github.com/aws-samples/sagemaker-genai-hosting-examples/tree/main/01-models
Find the requested model or an architecturally similar one and adopt its engine, image
family, instance and topology, context cap, and required SM_VLLM_* / SM_SGLANG_* options
as the starting point. This does not change the model-neutral contract: any configuration
found is supplied through generic CLI arguments (--image, --instance, --env,
--smoke-inputs, --capacity-reservation-arn), never added to the scripts. Use the model
publisher's model card for anything the examples do not cover.
Always re-validate against current reality before deploying: resolve the newest compatible DLC (reuse an example's pinned image only if it still exists in ECR), confirm the checkpoint's required GPU microarchitecture and quantization support, check endpoint quota, and size context and concurrency to the measured KV-cache budget.
If no matching example exists, or the example is stale (superseded DLC tag, deprecated flag,
test cases passed, vs 6 / 19 without
Published skills whose descriptions closely match this one, with their headline benchmark lift.
More skills from the same source and nearby in the registry, with their headline benchmark lift.