[Misc] Replace os environ to monkeypatch in test suite (#14516)

Signed-off-by: sibi <[email protected]>
Signed-off-by: Aaron Pham <[email protected]>
Co-authored-by: Cyrus Leung <[email protected]>
Co-authored-by: Aaron Pham <[email protected]>

Author: 3 people

.buildkite/test-pipeline.yaml

@@ -522,7 +522,7 @@ steps:
   # TODO: investigate and fix
   # - pytest -v -s spec_decode/e2e/test_integration_dist_tp2.py
   - VLLM_USE_V1=0 CUDA_VISIBLE_DEVICES=0,1 pytest -v -s test_sharded_state_loader.py
-  - VLLM_USE_V1=0 CUDA_VISIBLE_DEVICES=0,1 pytest -v -s kv_transfer/disagg_test.py
+  - VLLM_USE_V1=0 CUDA_VISIBLE_DEVICES=0,1 pytest -v -s kv_transfer/test_disagg.py
 
 - label: Plugin Tests (2 GPUs) # 40min
   working_dir: "/vllm-workspace/tests"

tests/basic_correctness/test_basic_correctness.py

@@ -47,6 +47,7 @@ def test_vllm_gc_ed():
 @pytest.mark.parametrize("max_tokens", [5])
 @pytest.mark.parametrize("enforce_eager", [False])
 def test_models(
+    monkeypatch: pytest.MonkeyPatch,
     hf_runner,
     model: str,
     backend: str,
@@ -63,31 +64,33 @@ def test_models(
         pytest.skip(
             f"{backend} does not support gemma2 with full context length.")
 
-    os.environ["VLLM_ATTENTION_BACKEND"] = backend
+    with monkeypatch.context() as m:
+        m.setenv("VLLM_ATTENTION_BACKEND", backend)
 
-    # 5042 tokens for gemma2
-    # gemma2 has alternating sliding window size of 4096
-    # we need a prompt with more than 4096 tokens to test the sliding window
-    prompt = "The following numbers of the sequence " + ", ".join(
-        str(i) for i in range(1024)) + " are:"
-    example_prompts = [prompt]
+        # 5042 tokens for gemma2
+        # gemma2 has alternating sliding window size of 4096
+        # we need a prompt with more than 4096 tokens to test the sliding window
+        prompt = "The following numbers of the sequence " + ", ".join(
+            str(i) for i in range(1024)) + " are:"
+        example_prompts = [prompt]
 
-    with hf_runner(model, dtype=dtype) as hf_model:
-        hf_outputs = hf_model.generate_greedy(example_prompts, max_tokens)
+        with hf_runner(model, dtype=dtype) as hf_model:
+            hf_outputs = hf_model.generate_greedy(example_prompts, max_tokens)
 
-    with VllmRunner(model,
-                    max_model_len=8192,
-                    dtype=dtype,
-                    enforce_eager=enforce_eager,
-                    gpu_memory_utilization=0.7) as vllm_model:
-        vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
+        with VllmRunner(model,
+                        max_model_len=8192,
+                        dtype=dtype,
+                        enforce_eager=enforce_eager,
+                        gpu_memory_utilization=0.7) as vllm_model:
+            vllm_outputs = vllm_model.generate_greedy(example_prompts,
+                                                      max_tokens)
 
-    check_outputs_equal(
-        outputs_0_lst=hf_outputs,
-        outputs_1_lst=vllm_outputs,
-        name_0="hf",
-        name_1="vllm",
-    )
+        check_outputs_equal(
+            outputs_0_lst=hf_outputs,
+            outputs_1_lst=vllm_outputs,
+            name_0="hf",
+            name_1="vllm",
+        )
 
 
 @multi_gpu_test(num_gpus=2)
@@ -104,6 +107,7 @@ def test_models(
         ("meta-llama/Meta-Llama-3-8B", "ray", "FLASHINFER", "A100"),
     ])
 def test_models_distributed(
+    monkeypatch: pytest.MonkeyPatch,
     hf_runner,
     vllm_runner,
     example_prompts,
@@ -116,34 +120,41 @@ def test_models_distributed(
     if test_suite != TARGET_TEST_SUITE:
         pytest.skip(f"Skip test for {test_suite}")
 
-    if model == "meta-llama/Llama-3.2-1B-Instruct" and distributed_executor_backend == "ray" and attention_backend == "" and test_suite == "L4":  # noqa
-        # test Ray Compiled Graph
-        os.environ['VLLM_USE_RAY_SPMD_WORKER'] = "1"
-        os.environ['VLLM_USE_RAY_COMPILED_DAG'] = "1"
-
-    if attention_backend:
-        os.environ["VLLM_ATTENTION_BACKEND"] = attention_backend
-
-    dtype = "half"
-    max_tokens = 5
-
-    # NOTE: take care of the order. run vLLM first, and then run HF.
-    # vLLM needs a fresh new process without cuda initialization.
-    # if we run HF first, the cuda initialization will be done and it
-    # will hurt multiprocessing backend with fork method (the default method).
-    with vllm_runner(model,
-                     dtype=dtype,
-                     tensor_parallel_size=2,
-                     distributed_executor_backend=distributed_executor_backend
-                     ) as vllm_model:
-        vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
-
-    with hf_runner(model, dtype=dtype) as hf_model:
-        hf_outputs = hf_model.generate_greedy(example_prompts, max_tokens)
-
-    check_outputs_equal(
-        outputs_0_lst=hf_outputs,
-        outputs_1_lst=vllm_outputs,
-        name_0="hf",
-        name_1="vllm",
-    )
+    with monkeypatch.context() as monkeypatch_context:
+        if model == "meta-llama/Llama-3.2-1B-Instruct" and distributed_executor_backend == "ray" and attention_backend == "" and test_suite == "L4":  # noqa
+            # test Ray Compiled Graph
+            monkeypatch_context.setenv("VLLM_USE_RAY_SPMD_WORKER", "1")
+            monkeypatch_context.setenv("VLLM_USE_RAY_COMPILED_DAG", "1")
+
+        if attention_backend:
+            monkeypatch_context.setenv(
+                "VLLM_ATTENTION_BACKEND",
+                attention_backend,
+            )
+
+        dtype = "half"
+        max_tokens = 5
+
+        # NOTE: take care of the order. run vLLM first, and then run HF.
+        # vLLM needs a fresh new process without cuda initialization.
+        # if we run HF first, the cuda initialization will be done and it
+        # will hurt multiprocessing backend with fork method
+        # (the default method).
+        with vllm_runner(
+                model,
+                dtype=dtype,
+                tensor_parallel_size=2,
+                distributed_executor_backend=distributed_executor_backend,
+        ) as vllm_model:
+            vllm_outputs = vllm_model.generate_greedy(example_prompts,
+                                                      max_tokens)
+
+        with hf_runner(model, dtype=dtype) as hf_model:
+            hf_outputs = hf_model.generate_greedy(example_prompts, max_tokens)
+
+        check_outputs_equal(
+            outputs_0_lst=hf_outputs,
+            outputs_1_lst=vllm_outputs,
+            name_0="hf",
+            name_1="vllm",
+        )