[ROCm][CI] Fix TP size issue for test_gpt_oss

[micah-wil]

Quantized Models Test is allocated to a 1 GPU agent pool in CI, but tries to run multi-GPU tests (example: https://buildkite.com/vllm/amd-ci/builds/5699/steps/canvas?sid=019cb28b-7107-44a7-adde-1af22fb4f7b7&tab=output#019cb28b-71fb-4bda-bc58-43ef57384abc/L1654)

This PR skips the multi-GPU test cases if there are not enough GPUs available.

[gemini-code-assist]

Code Review

The pull request effectively addresses the issue of multi-GPU tests failing in single-GPU CI environments by conditionally skipping these tests when insufficient GPUs are available. This is a practical fix that improves CI stability and efficiency. The implementation correctly uses cuda_device_count_stateless to determine available resources.

[gemini-code-assist]

According to PEP 8, imports should generally be grouped in the following order: standard library imports, third-party imports, and then local application/library specific imports. The vllm.utils.torch_utils import is a local application import and should be placed after packaging.version to maintain consistency with common Python style guidelines.

import pytest
from packaging import version

from vllm.utils.torch_utils import cuda_device_count_stateless

References

1. Imports should be grouped in the following order: standard library, third-party, and local application/library specific imports. Each group should be separated by a blank line. ^(link)