A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 8, 02:30 UTC
[BUG] Fixing Weight Sync unit test (#34841)
Signed-off-by: ahao-anyscale <ahao@anyscale.com>
Aaron Hao committed
Feb 18, 2026 at 14:20 UTC
e99ba957ec3953abd65332c26880efa489effc3a
1 file changed
+6
tests/entrypoints/weight_transfer/test_weight_transfer_llm.py
+6
@@ -124,6 +124,8 @@ def test_init_weight_transfer_engine_calls_engine():
124
if torch.cuda.device_count() < 1:
125
pytest.skip("Need at least 1 GPU for this test")
126
127
+ # Run in-process so mock.patch works (spawn won't inherit the mock)
128
+ os.environ["VLLM_ENABLE_V1_MULTIPROCESSING"] = "0"
129
# Enable insecure serialization to allow pickling functions for collective_rpc
130
os.environ["VLLM_ALLOW_INSECURE_SERIALIZATION"] = "1"
131
@@ -171,6 +173,8 @@ def test_update_weights_calls_engine():
173
if torch.cuda.device_count() < 1:
174
pytest.skip("Need at least 1 GPU for this test")
175
176
+ # Run in-process so mock.patch works (spawn won't inherit the mock)
177
+ os.environ["VLLM_ENABLE_V1_MULTIPROCESSING"] = "0"
178
# Enable insecure serialization to allow pickling functions for collective_rpc
179
os.environ["VLLM_ALLOW_INSECURE_SERIALIZATION"] = "1"
180
@@ -228,6 +232,8 @@ def test_full_weight_transfer_flow():
232
if torch.cuda.device_count() < 1:
233
pytest.skip("Need at least 1 GPU for this test")
234
235
+ # Run in-process so mock.patch works (spawn won't inherit the mock)
236
+ os.environ["VLLM_ENABLE_V1_MULTIPROCESSING"] = "0"
237
# Enable insecure serialization to allow pickling functions for collective_rpc
238
os.environ["VLLM_ALLOW_INSECURE_SERIALIZATION"] = "1"
239