[V0 Deprecation] Remove VLLM_USE_V1 from tests (#26341)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>

tests/compile/piecewise/test_full_cudagraph.py

@@ -66,7 +66,6 @@ def llm_pair(request):
             pytest.skip("Only Blackwell GPUs support Cutlass MLA")
 
     env_vars = {
-        "VLLM_USE_V1": "1",
         # Force native sampler to avoid potential nondeterminism in FlashInfer
         # when per-request generators are not used in V1.
         "VLLM_USE_FLASHINFER_SAMPLER": "0",
@@ -161,7 +160,6 @@ def test_full_cudagraph_with_invalid_backend():
     with (
         temporary_environ(
             {
-                "VLLM_USE_V1": "1",
                 "VLLM_ATTENTION_BACKEND": "FLEX_ATTENTION",
                 # Flex_Attention is not supported with full cuda graph
             }

tests/compile/piecewise/test_simple.py

@@ -18,7 +18,6 @@ from vllm.config import (
     VllmConfig,
     set_current_vllm_config,
 )
-from vllm.envs import VLLM_USE_V1
 from vllm.forward_context import BatchDescriptor, set_forward_context
 from vllm.utils import is_torch_equal_or_newer
 
@@ -127,7 +126,6 @@ def _run_simple_model(
 @pytest.mark.parametrize("use_inductor", [True, False])
 @torch.inference_mode()
 def test_simple_piecewise_compile(use_inductor):
-    assert VLLM_USE_V1
     _run_simple_model(
         splitting_ops=["silly.attention"],
         use_inductor_graph_partition=False,
@@ -146,7 +144,6 @@ def test_simple_piecewise_compile(use_inductor):
 @torch.inference_mode()
 @pytest.mark.parametrize("splitting_ops", [["silly.attention"], []])
 def test_simple_inductor_graph_partition(splitting_ops):
-    assert VLLM_USE_V1
     if not is_torch_equal_or_newer("2.9.0.dev"):
         pytest.skip("inductor graph partition is only available in PyTorch 2.9+")
 

tests/compile/test_async_tp.py

@@ -388,10 +388,6 @@ def test_async_tp_pass_correctness(
         "pass_config": {"enable_async_tp": async_tp_enabled},
     }
 
-    async_tp_env = tp_env = {
-        "VLLM_USE_V1": "1",
-    }
-
     async_tp_args = [
         *common_args,
         "--tensor-parallel-size",
@@ -410,6 +406,4 @@ def test_async_tp_pass_correctness(
         "mp",
     ]
 
-    compare_two_settings(
-        model_id, async_tp_args, tp_args, async_tp_env, tp_env, method="generate"
-    )
+    compare_two_settings(model_id, async_tp_args, tp_args, method="generate")

tests/compile/test_config.py

@@ -2,7 +2,6 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import pytest
 
-import vllm
 from vllm.compilation.counter import compilation_counter
 from vllm.config import CompilationConfig, CUDAGraphMode, VllmConfig
 from vllm.utils import _is_torch_equal_or_newer
@@ -16,15 +15,10 @@ def test_version():
     assert not _is_torch_equal_or_newer("2.7.1", "2.8.0.dev")
 
 
-def test_use_cudagraphs_dynamic(monkeypatch):
-    assert vllm.envs.VLLM_USE_V1
+def test_use_cudagraphs_dynamic():
     vllm_config = VllmConfig()
     assert vllm_config.compilation_config.use_cudagraph
 
-    monkeypatch.setenv("VLLM_USE_V1", "0")
-    vllm_config = VllmConfig()
-    assert not vllm_config.compilation_config.use_cudagraph
-
 
 def test_custom_op():
     # proper syntax
@@ -41,8 +35,6 @@ def test_custom_op():
 # may be influenced by other tests.
 @pytest.mark.parametrize("val", ["1"])
 def test_VLLM_DISABLE_COMPILE_CACHE(vllm_runner, monkeypatch, val):
-    assert vllm.envs.VLLM_USE_V1
-
     # Disable multiprocessing so that the counter is in the same process
     monkeypatch.setenv("VLLM_ENABLE_V1_MULTIPROCESSING", "0")
     monkeypatch.setenv("VLLM_DISABLE_COMPILE_CACHE", val)
@@ -68,8 +60,6 @@ def test_VLLM_DISABLE_COMPILE_CACHE(vllm_runner, monkeypatch, val):
 @pytest.mark.forked
 @pytest.mark.parametrize("enabled", [True, False])
 def test_use_cudagraphs(vllm_runner, monkeypatch, enabled):
-    assert vllm.envs.VLLM_USE_V1
-
     # Disable multiprocessing so that the counter is in the same process
     monkeypatch.setenv("VLLM_ENABLE_V1_MULTIPROCESSING", "0")
 

tests/compile/test_fusion_attn.py

@@ -303,7 +303,6 @@ def test_attention_quant_pattern(
     model_class: type[AttentionQuantPatternModel],
     backend: _Backend,
     use_inductor_graph_partition: bool,
-    monkeypatch,
     dist_init,
     caplog_vllm,
 ):
@@ -312,8 +311,6 @@ def test_attention_quant_pattern(
     if use_inductor_graph_partition and not is_torch_equal_or_newer("2.9.0.dev"):
         pytest.skip("inductor graph partition is only available in PyTorch 2.9+")
 
-    monkeypatch.setenv("VLLM_USE_V1", "1")
-
     device = torch.device("cuda:0")
     torch.manual_seed(42)