[py_connector] make connector unit tests runnable without torch

The open-source CI image has no torch/triton/orjson/zmq/requests, so the
three pure-logic py_tests failed at import time. vllm_stubs now registers
stand-ins for missing third-party deps (real modules always win), the
connector uses typing.TYPE_CHECKING instead of typing_extensions, and the
GPU-only Triton kernel test is tagged manual like the other GPU suites.

Author: lpdink

kv_cache_manager/py_connector/test/kernel/BUILD

@@ -8,9 +8,10 @@ py_test(
     ],
     main = "test_strided_gather_scatter.py",
     tags = [
-        "no-remote-exec",
-        "gpu",  # requires 1 GPU
         "exclusive",  # GPU tests run serially to avoid CUDA contention
+        "gpu",  # requires 1 GPU
+        "manual",  # needs torch/triton + GPU; not runnable in open-source CI
+        "no-remote-exec",
     ],
     deps = [
         "//kv_cache_manager/py_connector/kernel",

kv_cache_manager/py_connector/test/vllm_stubs.py

@@ -1,12 +1,16 @@
 """Shared test stubs: make ``v1_connector`` importable without vLLM/CUDA/pybind.
 
-``v1_connector`` imports vLLM and the compiled ``kvcm_py_client`` at module
+``v1_connector`` imports vLLM, the compiled ``kvcm_py_client`` and several
+third-party runtime deps (torch, triton, orjson, zmq, requests) at module
 level. For pure-logic unit tests we register lightweight stand-ins in
 ``sys.modules`` *before* the first import, then build connector instances via
 ``__new__`` with only the attributes the code under test reads. No production
-module is modified.
+module is modified; real modules are preferred whenever they are importable
+(e.g. on a dev machine with a full vLLM venv).
 """
 
+import importlib.util
+import json
 import sys
 import types
 from typing import Optional
@@ -21,7 +25,35 @@ def _module(name: str) -> types.ModuleType:
     return mod
 
 
+def _importable(name: str) -> bool:
+    try:
+        return importlib.util.find_spec(name) is not None
+    except (ImportError, ValueError):
+        return False
+
+
+def _stub_third_party():
+    """Register stand-ins for third-party deps missing from the environment
+    (the open-source CI runs these tests without torch/triton/orjson/zmq/
+    requests installed). Real modules always win."""
+    # Pure-attribute deps: a MagicMock module is enough because the pure-logic
+    # tests never execute tensor/socket/http work at module import time.
+    for name in ("torch", "triton", "triton.language", "zmq", "requests"):
+        if name not in sys.modules and not _importable(name):
+            sys.modules[name] = MagicMock(__name__=name)
+
+    # orjson is used functionally (CoordinateMsgSerializer round trips), so
+    # the stand-in must actually (de)serialize; stdlib json handles the
+    # dataclass payloads via __dict__.
+    if "orjson" not in sys.modules and not _importable("orjson"):
+        orjson = _module("orjson")
+        orjson.dumps = lambda obj: json.dumps(
+            obj, default=lambda o: o.__dict__).encode()
+        orjson.loads = json.loads
+
+
 def _install_stubs():
+    _stub_third_party()
     existing = sys.modules.get("vllm")
     if existing is not None:
         # Either our stub is already in place or the real vLLM is importable;

kv_cache_manager/py_connector/vllm/v1_connector.py

@@ -32,7 +32,6 @@
 from kv_cache_manager.client.pybind import kvcm_py_client
 
 import torch
-import typing_extensions
 from vllm.config import VllmConfig
 from vllm.distributed import get_tensor_model_parallel_rank
 from vllm.distributed.kv_transfer.kv_connector.v1.base import (
@@ -67,7 +66,7 @@
 from kv_cache_manager.py_connector.vllm.location_query_manager import LocationQueryManager
 from kv_cache_manager.py_connector.vllm.data_transfer import MultiResult, DataTransferManager, _get_device_module
 
-if typing_extensions.TYPE_CHECKING:
+if typing.TYPE_CHECKING:
     from vllm.forward_context import ForwardContext
     from vllm.attention import AttentionMetadata
     from vllm.v1.request import Request