fix(ci): unblock main build (stop_reason refs + omni-test CPU collection)

components/src/dynamo/vllm/tests/conftest.py

@@ -6,11 +6,35 @@
 framework is not installed in the current container.
 """
 
+import importlib
 import importlib.util
 import sys
 
 import pytest
 
+# Cached result of attempting to import the omni handler module.
+# `None` = not yet attempted, `True` = succeeded, `False` = raised.
+_omni_importable: bool | None = None
+
+
+def _can_import_omni() -> bool:
+    """Try to import dynamo.vllm.omni.base_handler once and cache the result.
+
+    Catches any exception, not just ImportError — vllm_omni's import chain
+    can raise NotImplementedError (and other types) when vllm._C / libcuda
+    aren't available on a CPU-only runner. importlib.util.find_spec is
+    insufficient because it only resolves the top-level package, not the
+    transitive imports that actually fail.
+    """
+    global _omni_importable
+    if _omni_importable is None:
+        try:
+            importlib.import_module("dynamo.vllm.omni.base_handler")
+            _omni_importable = True
+        except Exception:
+            _omni_importable = False
+    return _omni_importable
+
 
 def pytest_ignore_collect(collection_path, config):
     """Skip collecting vllm test files if vllm module isn't installed.
@@ -20,6 +44,16 @@ def pytest_ignore_collect(collection_path, config):
     if filename.startswith("test_vllm_"):
         if importlib.util.find_spec("vllm") is None:
             return True  # vllm not available, skip this file
+    # Omni tests import dynamo.vllm.omni.* which transitively imports
+    # vllm_omni at module load. On CPU-only sample-runtime runners the
+    # import chain reaches vllm._C and raises (NotImplementedError when
+    # libcuda.so.1 is missing). Each file's local try/except ImportError
+    # doesn't catch this, so skip collection up-front if the canonical
+    # omni module isn't importable.
+    parts = collection_path.parts
+    if "omni" in parts and filename.startswith("test_"):
+        if not _can_import_omni():
+            return True
     return None
 
 

lib/llm/src/preprocessor.rs

@@ -1578,7 +1578,6 @@ impl OpenAIPreprocessor {
                             choice.delta.refusal = None;
                             choice.delta.reasoning_content = None;
                             choice.finish_reason = None;
-                            choice.stop_reason = None;
                             choice.logprobs = None;
                             true
                         } else {

lib/llm/tests/postprocessor_parsing_stream.rs

@@ -329,7 +329,6 @@ fn mock_multi_choice_content_chunk(
                 reasoning_content: None,
             },
             finish_reason: None,
-            stop_reason: None,
             logprobs: None,
         })
         .collect();

tests/serve/test_vllm_omni.py

@@ -10,7 +10,10 @@
 
 try:
     from dynamo.vllm.omni.args import OmniConfig  # noqa: F401
-except ImportError:
+except Exception:
+    # vllm_omni's import chain can raise NotImplementedError (and other
+    # non-ImportError types) on platforms it doesn't support — e.g. a
+    # CPU-only runner where vllm._C can't load libcuda.so.1.
     pytest.skip("vLLM omni dependencies not available", allow_module_level=True)
 
 from tests.serve.common import (