Gemma 4 on an Amazon SageMaker Endpoint: AWS CLI, NVIDIA L4, and an MCP Server

A developer published a step-by-step guide and open-source repository for deploying Google's Gemma 4 E2B model to an Amazon SageMaker real-time endpoint on a single NVIDIA L4 GPU using AWS's vLLM container. The project wraps every AWS CLI call in a Python MCP server so Claude Code can manage the deployment through annotated tools, with tests that run offline by faking the aws subprocess.

This article provides a step by step deployment guide for Gemma 4 E2B to an Amazon SageMaker hosted GPU enabled system. A suite of Python MCP tools is built to simplify management of the vLLM hosted deployment with Claude Code. https://github.com/xbill9/sagemaker-gemma https://github.com/xbill9/sagemaker-gemma This project serves Gemma 4 E2B from a SageMaker real-time endpoint on one NVIDIA L4 GPU, using the vLLM container AWS publishes for SageMaker. Every AWS call is a plain aws CLI command, so each step can be run by hand or by the MCP server. A SageMaker real-time endpoint is a managed HTTPS inference server. SageMaker places the container on a GPU instance, health-checks it, routes requests to it and writes its logs to CloudWatch. There is no instance to patch, no security group to open and no load balancer to build. The strategy for starting MCP development for model management is a incremental step by step approach. First, the basic development environment is setup with the required system variables and a working Claude Code configuration. Then, the Python MCP server is brought up over stdio and validated with Claude Code in the local environment. The deployment follows as eight steps, each shown as the raw aws command and the MCP tool that runs it. aws login ml.g6.xlarge , ml.g6.2xlarge or ml.g6.4xlarge mcp 2.x jq for reading JSON replies Clone the repository and install the one requirement into the system Python: git clone https://github.com/xbill9/sagemaker-gemma cd sagemaker-gemma python3 -m pip install -r requirements.txt cp .env.example .env .env is gitignored. It holds the settings every tool reads: cat .env Copy to .env gitignored and edit. sm.py reads it, so the MCP server, CLI and Makefile all see it. AWS REGION=us-east-2 MODEL ID=google/gemma-4-E2B-it INSTANCE TYPE=ml.g6.xlarge ENDPOINT NAME=gemma-4-e2b ROLE NAME=sagemaker-gemma-execution-role MAX MODEL LEN=8192 Leave empty to use the newest SageMaker vLLM image in the region. IMAGE URI= Fallback instance types, highest priority first same GPU keeps runs comparable . INSTANCE POOLS=ml.g6.xlarge,ml.g6.2xlarge,ml.g6.4xlarge Gemma 4 is Apache-2.0 and ungated on Hugging Face, so no Hugging Face token is needed. The simplest MCP transport is stdio: the client launches the server as a local process and talks to it over stdin and stdout. In this project Claude Code is the MCP client. The server is one file, server.py , on the MCP Python SDK 2.x: mcp = MCPServer RIG NAME READ ONLY = ToolAnnotations readOnlyHint=True, idempotentHint=True WRITE = ToolAnnotations destructiveHint=False DESTRUCTIVE = ToolAnnotations destructiveHint=True Every tool carries one of the three annotations, so a client can tell a status check from a deploy from a delete. The tools do no AWS work themselves. They call sm.py , which runs each request as an aws CLI subprocess: php def aws args: str, region: str | None = REGION, parse: bool = True - Any: """Run aws <args --output json and return the parsed result.""" cmd = "aws", args, "--output", "json" The CLI renews an aws login session on its own, so a server that stays up for hours keeps working credentials. sm.py also drops any AWS SESSION TOKEN the server inherits from its parent process, because a static token expires inside a long-running server and outranks the login session. The project can be linted: make lint All checks passed 16 files already formatted and tested: make test ---------------------------------------------------------------------- Ran 16 tests in 0.010s OK The tests replace the aws subprocess with a fake, so they run offline with no credentials. One of them compares the registered tool set and annotations against a fixed list: a tool that failed to register, or a delete that lost its destructive flag, fails the suite. ✅ A client speaks JSON-RPC over stdio. Hold stdin open with sleep , or the server sees end-of-input and exits before it answers: { printf '%s\n' \ '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"probe","version":"0"}}}' \ '{"jsonrpc":"2.0","method":"notifications/initialized"}' \ '{"jsonrpc":"2.0","id":2,"method":"tools/list","params":{}}'; sleep 3; } \ | python3 server.py 2 /dev/null Summarised: 1 sagemaker-gemma 2 'check quotas', 'delete endpoint', 'deploy endpoint', 'find vllm image', 'get deployment config', 'get endpoint logs', 'get endpoint status', 'get help', 'list endpoints', 'query model', 'verify model health' 🟢 The server answers the handshake and lists 11 tools. Claude Code reads .mcp.json in the project directory and launches the server with the system python3 : { "mcpServers": { "sagemaker-gemma": { "command": "python3", "args": "/home/xbill/sagemaker-gemma/server.py" } } } Gemini CLI reads the same entry from .gemini/settings.json . claude mcp get sagemaker-gemma sagemaker-gemma: Scope: Project config shared via .mcp.json Status: ✔ Connected From inside Claude Code, get help returns the resolved settings and the order of work: sagemaker-gemma Serving google/gemma-4-E2B-it with vLLM on a SageMaker real-time endpoint , managed entirely through the aws CLI. | Setting | Value | | --- | --- | | Region | us-east-2 | | Endpoint | gemma-4-e2b | | Instance types priority order | ml.g6.xlarge , ml.g6.2xlarge , ml.g6.4xlarge | | Max model length | 8192 | | Image | newest SageMaker vLLM image find vllm image | | Execution role | sagemaker-gemma-execution-role | Order of work: check quotas → deploy endpoint → get endpoint status until InService about 10 minutes once an instance is placed → verify model health → query model → delete endpoint. The steps below follow that order. Each one shows the raw CLI command; deploy endpoint runs Steps 2 to 6 in one call. export AWS REGION=us-east-2 AWS PAGER= export NAME=gemma-4-e2b export MODEL ID=google/gemma-4-E2B-it SageMaker endpoint quotas are per instance type, per region, and count instances. aws service-quotas list-service-quotas --service-code sagemaker \ --query "Quotas ?QuotaName=='ml.g6.xlarge for endpoint usage' .Value" 1.0 The check quotas tool reads the same quota for each fallback type across the US regions: | Instance | us-east-1 | us-east-2 | us-west-1 | us-west-2 | | --- | ---: | ---: | ---: | ---: | | ml.g6.xlarge | 1 | 1 | - | 1 | | ml.g6.2xlarge | 1 | 1 | - | 1 | | ml.g6.4xlarge | 1 | 1 | - | 1 | Regions with quota for at least one of these types: us-east-1, us-east-2, us-west-2. A - means the type is not offered in that region. A 0 means a quota increase request first. list-service-quotas pages through every SageMaker quota in the region, and the Service Quotas API is rate limited per account. Twelve calls at once, one per type per region, came back as: An error occurred TooManyRequestsException when calling the ListServiceQuotas operation reached max retries: 2 check quotas makes one call per region, one region at a time, and filters the types from that one result. sm.py also sets AWS RETRY MODE=adaptive so the CLI backs off and retries. The AWS vLLM repository holds pinned release tags such as 0.30.0-gpu-py312-cu130-ubuntu24.04-sagemaker-v1.1 , floating aliases such as 0.30-gpu-py312 , and -soci index tags. Older release lines receive patch rebuilds, so the most recent push can carry an older vLLM. Sort the pinned tags by version: TAG=$ aws ecr describe-images --registry-id 763104351884 --repository-name vllm \ --query "imageDetails .imageTags " --output text | tr '\t' '\n' \ | grep -E -- '^ 0-9 +\. 0-9 +\. 0-9 +-. -sagemaker-v 0-9 +\. 0-9 +$' | sort -V | tail -1 export IMAGE=763104351884.dkr.ecr.$AWS REGION.amazonaws.com/vllm:$TAG echo $IMAGE 763104351884.dkr.ecr.us-east-2.amazonaws.com/vllm:0.30.0-gpu-py312-cu130-ubuntu24.04-sagemaker-v1.1 The find vllm image tool applies the same rule. The image runs vLLM 0.30.0. SageMaker assumes this role to pull the image and write logs. It is created once per account. aws iam create-role --role-name sagemaker-gemma-execution-role \ --assume-role-policy-document '{"Version":"2012-10-17","Statement": {"Effect":"Allow","Principal":{"Service":"sagemaker.amazonaws.com"},"Action":"sts:AssumeRole"} }' aws iam attach-role-policy --role-name sagemaker-gemma-execution-role \ --policy-arn arn:aws:iam::aws:policy/AmazonSageMakerFullAccess export ROLE=$ aws iam get-role --role-name sagemaker-gemma-execution-role --query Role.Arn --output text On a second run create-role reports EntityAlreadyExists ; get-role still sets ROLE . deploy endpoint creates the role only when it is missing. A SageMaker model pairs an image with its settings. The container turns each SM VLLM variable into the matching vLLM flag: SM VLLM MAX MODEL LEN becomes --max-model-len . aws sagemaker create-model --model-name $NAME --execution-role-arn $ROLE \ --primary-container "{\"Image\":\"$IMAGE\",\"Environment\":{ \"SM VLLM MODEL\":\"$MODEL ID\", \"SM VLLM MAX MODEL LEN\":\"8192\", \"SM VLLM GPU MEMORY UTILIZATION\":\"0.9\"}}" { "ModelArn": "arn:aws:sagemaker:us-east-2:<account-id :model/gemma-4-e2b" } The endpoint config says where the model runs. InstancePools lists up to five instance types in priority order, and SageMaker places the first one with a free instance. All three types below carry one L4, so the model sees the same GPU whichever is placed. aws sagemaker create-endpoint-config --endpoint-config-name $NAME \ --production-variants " {\"VariantName\":\"AllTraffic\",\"ModelName\":\"$NAME\", \"InitialInstanceCount\":1, \"InstancePools\": {\"InstanceType\":\"ml.g6.xlarge\",\"Priority\":1}, {\"InstanceType\":\"ml.g6.2xlarge\",\"Priority\":2}, {\"InstanceType\":\"ml.g6.4xlarge\",\"Priority\":3} , \"ContainerStartupHealthCheckTimeoutInSeconds\":1800, \"ModelDataDownloadTimeoutInSeconds\":1800} " { "EndpointConfigArn": "arn:aws:sagemaker:us-east-2:<account-id :endpoint-config/gemma-4-e2b" } The two 1800-second timeouts give the container time to download the weights and compile before SageMaker's health check gives up. With one Gemma endpoint running on ml.g6.xlarge from the three-type list above, a second endpoint asking for ml.g6.2xlarge in the same region was refused: ResourceLimitExceeded: The account-level service limit 'ml.g6.2xlarge for endpoint usage' is 1 Instances, with current utilization of 1 Instances and a request delta of 1 Instances. With a quota of 1 per type, a second endpoint goes in another region, or uses types outside the first endpoint's list. aws sagemaker create-endpoint --endpoint-name $NAME --endpoint-config-name $NAME { "EndpointArn": "arn:aws:sagemaker:us-east-2:<account-id :endpoint/gemma-4-e2b" } Billing starts when an instance is placed. aws sagemaker wait endpoint-in-service --endpoint-name $NAME blocks until it is ready; get endpoint status reports the same state and the instance type that was placed: ✅ gemma-4-e2b in us-east-2 : InService - Instance: ml.g6.xlarge The container log is in CloudWatch, and get endpoint logs tails it: aws logs tail /aws/sagemaker/Endpoints/$NAME --follow Start-up on the L4, in minutes after create-endpoint : | Phase | Minutes | |---|---| | Weights loaded, 9.75 GiB in 82.75 s | 7.4 | | KV cache sized, 723,484 tokens | 9.2 | | InService | 9.9 | create-endpoint cannot be undone until the endpoint settles. A delete-endpoint sent while it is Creating is refused: aws: ERROR : An error occurred ValidationException when calling the DeleteEndpoint operation: Cannot update in-progress endpoint "arn:aws:sagemaker:us-east-2:<account-id :endpoint/gemma-4-e2b". The endpoint finishes starting, bills from the moment its instance is placed, and can be deleted once it reaches InService or Failed . Check the model ID and instance types before Step 6. A quota of 1 lets you request one instance; the region still has to have one free. In us-east-1 , two requests for L4 instances each stayed in Creating for about 30 minutes and then failed: Unable to provision requested ML compute capacity due to InsufficientInstanceCapacity error. Please retry using a different ML instance type or after some time. The same request in us-east-2 placed an instance at once. While SageMaker waits for capacity, no container starts and the CloudWatch log group never appears, which is how get endpoint logs tells a capacity wait from a slow model load. No charge accrues during the wait. When it fails, repeat Steps 4 to 6 in another region where Step 1 shows the same quota. The vLLM container accepts an OpenAI chat body on invoke-endpoint : echo '{"messages": {"role":"user","content":"Why is the sky blue?"} ,"max tokens":256}' req.json aws sagemaker-runtime invoke-endpoint --endpoint-name $NAME \ --content-type application/json --body fileb://req.json out.json jq -r '.choices 0 .message.content' out.json | head -1 jq -c '{model,usage}' out.json { "ContentType": "application/json", "InvokedProductionVariant": "AllTraffic" } The sky is blue due to a phenomenon called Rayleigh scattering . This process is caused by how sunlight interacts with the Earth's atmosphere. {"model":"google/gemma-4-E2B-it","usage":{"prompt tokens":15,"total tokens":271,"completion tokens":256,"prompt tokens details":null,"completion tokens details":null}} verify model health sends one short request and checks for a reply: ✅ model= google/gemma-4-E2B-it tokens=2 wall=0.749s reply='ok' The endpoint bills by the hour until it is deleted. Deleting the config and the model as well leaves nothing behind. delete endpoint runs all three: aws sagemaker delete-endpoint --endpoint-name $NAME aws sagemaker delete-endpoint-config --endpoint-config-name $NAME aws sagemaker delete-model --model-name $NAME Each command prints nothing and exits 0. list endpoints confirms the account is clear: 0 endpoint s matching gemma The goal of this article was to deploy Gemma 4 E2B to an Amazon SageMaker real-time endpoint with the AWS CLI and manage it from an MCP server. The key to the solution was the AWS vLLM SageMaker container, which turns the deployment into three create- calls and makes the endpoint answer OpenAI-style chat requests. The deployment results were: InService 9.9 minutes after create-endpoint , with the weights using 9.75 GiB of the L4's 24 GB InstancePools gives one endpoint config several fallback instance types us-east-1 refused twice, about 30 minutes each, while us-east-2 placed an instance at once Scope: one account, ml.g6.xlarge with one NVIDIA L4, vLLM 0.30.0 from the AWS container, google/gemma-4-E2B-it at full precision, deployed in us-east-2 on 2026-09-25. Start-up times are from a single deployment. The strategy for using MCP for SageMaker deployment was validated with an incremental step by step approach.