addressed comments

Signed-off-by: Kinjal Patel <kinjalpravin@nvidia.com>

Author: kinjalpatel27

.github/workflows/gpu_tests.yml

@@ -79,6 +79,9 @@ jobs:
         run: |
           python3 -m pip install nox && nox -s ${{ matrix.example }}
       - name: Upload GPU coverage to Codecov
+        # vLLM container has no ``git``, which codecov-action needs; gpu_vllm
+        # also runs without ``--cov`` so there's no coverage.xml to upload.
+        if: matrix.example != 'gpu_vllm'
         uses: codecov/codecov-action@v5
         with:
           token: ${{ secrets.CODECOV_TOKEN }}

tests/_test_utils/torch/transformers_models.py

@@ -153,8 +153,14 @@ def get_tiny_deepseek_v3(**config_kwargs) -> PreTrainedModel:
     return AutoModelForCausalLM.from_config(cfg)
 
 
-def create_tiny_deepseek_v3_dir(tmp_path: Path | str, **config_kwargs) -> Path:
+def create_tiny_deepseek_v3_dir(
+    tmp_path: Path | str, with_tokenizer: bool = False, **config_kwargs
+) -> Path:
     deepseek_dir = Path(tmp_path) / "tiny_deepseek_v3"
+    if with_tokenizer:
+        tokenizer = get_tiny_tokenizer()
+        tokenizer.save_pretrained(deepseek_dir)
+        config_kwargs["vocab_size"] = tokenizer.vocab_size
     get_tiny_deepseek_v3(**config_kwargs).save_pretrained(deepseek_dir)
     return deepseek_dir
 

tests/gpu_vllm/conftest.py

@@ -13,18 +13,12 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-"""Shared setup for vLLM tests.
-
-vLLM handles its own distributed init, current-vllm-config context, and
-parallel-state setup when ``LLM(...)`` is constructed, so this conftest only
-opts into ``VLLM_ALLOW_INSECURE_SERIALIZATION=1`` *before* importing vLLM so
-``LLM.collective_rpc(callable)`` can ship our worker callables over the engine
-IPC channel via pickle. Without this, the default msgpack encoder rejects raw
-functions and the call raises ``TypeError``. Only safe in a controlled test
-environment.
+"""Set ``VLLM_ALLOW_INSECURE_SERIALIZATION=1`` before vLLM is imported so
+``LLM.collective_rpc(callable)`` can pickle worker callables. pytest loads
+conftests before sibling test modules, so this beats the top-level
+``from vllm import LLM`` in ``test_*.py``.
 """
 
 import os
 
-# Must precede any ``import vllm``: the env is read at module-import time.
 os.environ.setdefault("VLLM_ALLOW_INSECURE_SERIALIZATION", "1")