[Bugfix][Frontend] Parse VLLM_USE_MODELSCOPE via vllm.envs

[yashkgp]

Purpose

Fixes #5427.

omni_snapshot_download() checked VLLM_USE_MODELSCOPE for truthiness instead of
parsing it, so every non-empty value enabled ModelScope, including 0, False,
false, no and off.

vLLM only treats the literal string "true" (case-insensitive) as enabled, see
vllm/envs.py @ v0.25.0:

"VLLM_USE_MODELSCOPE": lambda: (
    os.environ.get("VLLM_USE_MODELSCOPE", "False").lower() == "true"
),

omni_snapshot_download() is the first call in OmniBase.__init__, so with
VLLM_USE_MODELSCOPE=0 both offline Omni(...) and vllm-omni serve took the
ModelScope branch. modelscope is not a vLLM-Omni dependency (it is only
installed in docker/Dockerfile.xpu), so a default install failed at startup with
ModuleNotFoundError: No module named 'modelscope'. Where it is installed,
weights were fetched from ModelScope despite the opt-out.

vllm_omni/model_executor/model_loader/weight_utils.py#L9 already reads the flag
through vllm.envs, so the two paths disagreed inside one process. This routes
the call site through vllm.envs as well. VLLM_USE_MODELSCOPE=true behaves as
before.

Test Plan

Added tests/entrypoints/test_omni_snapshot_download.py, marked core_model and
cpu so the existing L1 Simple · Other Test lane in
.buildkite/cuda/test-ready.yml collects it, no CI wiring changes needed. It
stubs the ModelScope module into sys.modules, otherwise the wrong branch raises
ModuleNotFoundError instead of being observable, then asserts which backend
omni_snapshot_download() selects for each flag value.

pytest -sv tests/entrypoints/test_omni_snapshot_download.py -m 'core_model and cpu'

Repro for the original bug:

VLLM_USE_MODELSCOPE=0 python -c "
from vllm_omni.entrypoints.omni_base import omni_snapshot_download
omni_snapshot_download('Qwen/Qwen3-Omni-30B-A3B-Instruct')
"

vLLM Version: v0.25.0

vLLM-Omni Commit: 2bb6189

Test Result

On main, 6 of the 10 cases fail:

FAILED tests/entrypoints/test_omni_snapshot_download.py::test_non_true_values_do_not_enable_modelscope[0]
FAILED tests/entrypoints/test_omni_snapshot_download.py::test_non_true_values_do_not_enable_modelscope[1]
FAILED tests/entrypoints/test_omni_snapshot_download.py::test_non_true_values_do_not_enable_modelscope[False]
FAILED tests/entrypoints/test_omni_snapshot_download.py::test_non_true_values_do_not_enable_modelscope[false]
FAILED tests/entrypoints/test_omni_snapshot_download.py::test_non_true_values_do_not_enable_modelscope[no]
FAILED tests/entrypoints/test_omni_snapshot_download.py::test_non_true_values_do_not_enable_modelscope[off]
6 failed, 4 passed in 0.02s

E         - huggingface
E         + modelscope

With the fix:

10 passed in 0.01s

ruff check, ruff format --check and tools/pre_commit/check_test_marks.py are
clean on both changed files.

I do not have a CUDA box, so that pytest run was done in a CPU sandbox that pairs
the real vllm/envs.py from v0.25.0 with the source of omni_snapshot_download()
taken from this branch, and from main for the "before" numbers. The test file is
the same one in this PR. The full tests/entrypoints/ suite needs CI.

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Credits must be used to enable repository wide code reviews.

[Gaohan123]

LGTM. Thanks