[Feature Request] Add vLLM CPU inference image for SageMaker

[timelfrink]

What

Requesting a CPU-only vLLM inference container for SageMaker, similar to the existing GPU-based vLLM images.

Why

This would enable running vLLM on CPU instances for cost-effective workloads that don't require GPU acceleration, such as:

• Reranking
• Scoring
• Embeddings
• Small generative models

Reference Implementation

I previously submitted PR #5670 which contains a working implementation (Dockerfile + buildspec) that was validated on EC2 (c5.4xlarge):

• Successful image build (~3.5 GB)
• Health endpoint returning 200
• /v1/completions working with facebook/opt-125m
• Reranker endpoint working with Alibaba-NLP model

Key Implementation Details from PR

• vLLM v0.15.1 with CPU target build
• Python 3.12, Ubuntu 22.04
• tcmalloc + Intel OpenMP for performance
• Reuses existing SageMaker entrypoint script
• Tag format: 0.15.1-cpu-py312-ubuntu22.04-sagemaker

I understand external contributions are not accepted, so filing this as a feature request for the team to consider. The PR can serve as a reference for the implementation.