[Bug]: model_hosting_container_standards changes a shared root/vllm handler to ERROR and suppresses vLLM startup logs #53655

Your current environment

vLLM: 0.18.1.dev44+gcfae1bf2d
Python: 3.11.14
model-hosting-container-standards: 0.1.16
Platform: reproduced with vllm-ascend/NPU, but the root cause is platform-independent

🐛 Describe the bug

Importing vllm.entrypoints.openai.api_server imports the SageMaker integration and, transitively, model_hosting_container_standards.logging_config.

At module import time, model_hosting_container_standards.logging_config.configure_root_logger() changes the levels of existing root handlers to ERROR.

When a custom VLLM_LOGGING_CONFIG_PATH assigns the same handler object to both the root logger and the vllm logger, this also changes the effective vLLM handler level to ERROR. As a result, vLLM startup INFO and WARNING records are silently discarded.

For example, the following scheduler message is not printed even though the corresponding branch executes:

Chunked prefill is enabled with max_num_batched_tokens=16384.

This is not caused by chunked prefill, data parallelism, DPCoordinator, or vllm-ascend. Instrumentation confirms that SchedulerConfig.__post_init__() executes with:

enable_chunked_prefill=True
max_num_batched_tokens=16384

The record is discarded only because the shared handler has already been changed to ERROR.

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.