Responses vLLM backend passes TP environment value as a string

Summary

The Responses API vLLM backend defines tensor parallelism with:

TP = os.environ.get("TP", 2)

When TP is unset, the fallback is integer 2. When a user configures the environment variable, os.environ.get() returns a string, and that string is passed directly to:

LLM(..., tensor_parallel_size=TP, ...)

tensor_parallel_size is an integer configuration value, so the explicit environment override changes the value's type instead of only changing its size.

Impact

The documented/configurable TP path can fail during vLLM engine setup when users set values such as TP=4, even though the default path works.

Proposed resolution

Parse the environment variable as an integer:

TP = int(os.environ.get("TP", 2))

Add a regression that imports the backend with TP=4 and verifies load_model() passes integer 4 to LLM.