[CI/Build] Fix CI LoRA failure

tests/lora/conftest.py

@@ -256,3 +256,15 @@ def run_with_both_engines_lora(request, monkeypatch):
         monkeypatch.setenv('VLLM_USE_V1', '0')
 
     yield
+
+
+@pytest.fixture
+def reset_default_device():
+    """
+    Some tests, such as `test_punica_ops.py`, explicitly set the 
+    default device, which can affect subsequent tests. Adding this fixture 
+    helps avoid this problem.
+    """
+    original_device = torch.get_default_device()
+    yield
+    torch.set_default_device(original_device)

tests/lora/test_baichuan.py

@@ -73,7 +73,6 @@ def test_baichuan_tensor_parallel_equality(baichuan_lora_files,
                        max_num_seqs=16,
                        max_loras=4,
                        max_lora_rank=64,
-                       tensor_parallel_size=1,
                        trust_remote_code=True,
                        fully_sharded_loras=fully_sharded)
     output_tp1 = do_sample(llm_tp1, baichuan_lora_files, lora_id=1)

tests/lora/test_chatglm3_tp.py

@@ -61,7 +61,6 @@ def test_chatglm3_lora(chatglm3_lora_files):
                    enable_lora=True,
                    max_loras=4,
                    max_lora_rank=64,
-                   tensor_parallel_size=1,
                    trust_remote_code=True,
                    enable_chunked_prefill=True)
 

tests/lora/test_layers.py

@@ -65,7 +65,7 @@
 
 
 @pytest.fixture(autouse=True)
-def clean_cache():
+def clean_cache_reset_device(reset_default_device):
     # Release any memory we might be holding on to. CI runs OOMs otherwise.
     from vllm.lora.ops.triton_ops.utils import (_LORA_A_PTR_DICT,
                                                 _LORA_B_PTR_DICT)

tests/lora/test_llama_tp.py

@@ -88,7 +88,6 @@ def test_llama_lora(sql_lora_files):
         # also test odd max_num_seqs
         max_num_seqs=13,
         max_loras=4,
-        tensor_parallel_size=1,
         enable_chunked_prefill=True)
     generate_and_test(llm, sql_lora_files)
 

tests/lora/test_punica_ops.py

@@ -13,6 +13,11 @@
 from .utils import PunicaTensors, assert_close, generate_data_for_nslices
 
 
+@pytest.fixture(autouse=True)
+def reset_device(reset_default_device):
+    pass
+
+
 # Utility shrink and expand operations used as reference implementations.
 def sgmv_shrink_for_nslices(
         nslices: int, inputs_tensor: torch.Tensor,

tests/lora/test_quant_model.py

@@ -78,20 +78,14 @@ def format_prompt_tuples(prompt):
 
 
 @pytest.mark.parametrize("model", MODELS)
-@pytest.mark.parametrize("tp_size", [1])
-def test_quant_model_lora(tinyllama_lora_files, num_gpus_available, model,
-                          tp_size):
-    if num_gpus_available < tp_size and \
-        tp_size > 1 and current_platform.is_cuda_alike():
-        pytest.skip(f"Not enough GPUs for tensor parallelism {tp_size}")
+def test_quant_model_lora(tinyllama_lora_files, model):
 
     llm = vllm.LLM(
         model=model.model_path,
         enable_lora=True,
         max_num_seqs=16,
         max_loras=4,
         max_model_len=400,
-        tensor_parallel_size=tp_size,
         gpu_memory_utilization=0.2,  #avoid OOM
         quantization=model.quantization,
         trust_remote_code=True,
@@ -185,7 +179,6 @@ def test_quant_model_tp_equality(tinyllama_lora_files, num_gpus_available,
         enable_lora=True,
         max_num_seqs=16,
         max_loras=4,
-        tensor_parallel_size=1,
         gpu_memory_utilization=0.2,  #avoid OOM
         quantization=model.quantization,
         trust_remote_code=True,

tests/lora/test_transfomers_model.py

@@ -53,7 +53,6 @@ def test_ilama_lora(ilama_lora_files):
                    enable_lora=True,
                    max_loras=4,
                    max_lora_rank=16,
-                   tensor_parallel_size=1,
                    trust_remote_code=True,
                    enable_chunked_prefill=True)