[Attention] Update tests to remove deprecated env vars (#30563)

Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>

.buildkite/scripts/hardware_ci/run-xpu-test.sh

@@ -39,7 +39,7 @@ docker run \
     python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend ray
     python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend mp
     python3 examples/offline_inference/basic/generate.py --model Intel/Qwen2.5-0.5B-W4A16-G128-AutoRound-LLMC-TEST-ONLY --enforce-eager
-    VLLM_ATTENTION_BACKEND=TRITON_ATTN python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager
+    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager --attention-backend=TRITON_ATTN
     cd tests
     pytest -v -s v1/core
     pytest -v -s v1/engine

tests/basic_correctness/test_basic_correctness.py

@@ -67,7 +67,6 @@ def _fix_prompt_embed_outputs(
 @pytest.mark.parametrize("model_executor", ["uni", "mp"])
 @pytest.mark.parametrize("enable_prompt_embeds", [True, False])
 def test_models(
-    monkeypatch: pytest.MonkeyPatch,
     hf_runner,
     model: str,
     backend: str,
@@ -77,48 +76,46 @@ def test_models(
     model_executor: str,
     enable_prompt_embeds: bool,
 ) -> None:
-    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]
-
-        with hf_runner(model) as hf_model:
-            hf_outputs = hf_model.generate_greedy(example_prompts, max_tokens)
-            if enable_prompt_embeds:
-                with torch.no_grad():
-                    prompt_embeds = hf_model.get_prompt_embeddings(example_prompts)
-
-        with VllmRunner(
-            model,
-            max_model_len=8192,
-            enforce_eager=enforce_eager,
-            enable_prompt_embeds=enable_prompt_embeds,
-            gpu_memory_utilization=0.7,
-            async_scheduling=async_scheduling,
-            distributed_executor_backend=model_executor,
-        ) as vllm_model:
-            if enable_prompt_embeds:
-                vllm_outputs = vllm_model.generate_greedy(prompt_embeds, max_tokens)
-                vllm_outputs = _fix_prompt_embed_outputs(
-                    vllm_outputs, hf_model, example_prompts
-                )
-            else:
-                vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
+    # 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) as hf_model:
+        hf_outputs = hf_model.generate_greedy(example_prompts, max_tokens)
+        if enable_prompt_embeds:
+            with torch.no_grad():
+                prompt_embeds = hf_model.get_prompt_embeddings(example_prompts)
+
+    with VllmRunner(
+        model,
+        max_model_len=8192,
+        enforce_eager=enforce_eager,
+        enable_prompt_embeds=enable_prompt_embeds,
+        gpu_memory_utilization=0.7,
+        async_scheduling=async_scheduling,
+        distributed_executor_backend=model_executor,
+        attention_config={"backend": backend},
+    ) as vllm_model:
+        if enable_prompt_embeds:
+            vllm_outputs = vllm_model.generate_greedy(prompt_embeds, max_tokens)
+            vllm_outputs = _fix_prompt_embed_outputs(
+                vllm_outputs, hf_model, example_prompts
+            )
+        else:
+            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)
@@ -161,12 +158,6 @@ def test_models_distributed(
         ):  # noqa
             pytest.skip("enable_prompt_embeds does not work with ray compiled dag.")
 
-        if attention_backend:
-            monkeypatch_context.setenv(
-                "VLLM_ATTENTION_BACKEND",
-                attention_backend,
-            )
-
         for k, v in extra_env.items():
             monkeypatch_context.setenv(k, v)
 
@@ -178,6 +169,7 @@ def test_models_distributed(
         # if we run HF first, the cuda initialization will be done and it
         # will hurt multiprocessing backend with fork method
         # (the default method).
+        attention_config = {"backend": attention_backend} if attention_backend else None
         with vllm_runner(
             model,
             dtype=dtype,
@@ -185,6 +177,7 @@ def test_models_distributed(
             distributed_executor_backend=distributed_executor_backend,
             enable_prompt_embeds=enable_prompt_embeds,
             gpu_memory_utilization=0.7,
+            attention_config=attention_config,
         ) as vllm_model:
             if enable_prompt_embeds:
                 with hf_runner(model, dtype=dtype) as hf_model:

tests/compile/distributed/test_fusions_e2e.py

@@ -208,7 +208,8 @@ def test_attn_quant(
     # To capture subprocess logs, we need to know whether spawn or fork is used.
     # Force spawn as it is more general.
     monkeypatch.setenv("VLLM_WORKER_MULTIPROC_METHOD", "spawn")
-    monkeypatch.setenv("VLLM_ATTENTION_BACKEND", backend.name)
+
+    model_kwargs["attention_config"] = {"backend": backend.name}
 
     compilation_config = CompilationConfig(
         # Testing properties
@@ -297,7 +298,8 @@ def test_tp2_attn_quant_allreduce_rmsnorm(
     # To capture subprocess logs, we need to know whether spawn or fork is used.
     # Force spawn as it is more general.
     monkeypatch.setenv("VLLM_WORKER_MULTIPROC_METHOD", "spawn")
-    monkeypatch.setenv("VLLM_ATTENTION_BACKEND", backend.name)
+
+    model_kwargs["attention_config"] = {"backend": backend.name}
 
     compilation_config = CompilationConfig(
         # Testing properties
@@ -409,7 +411,8 @@ def test_tp2_attn_quant_async_tp(
     # To capture subprocess logs, we need to know whether spawn or fork is used.
     # Force spawn as it is more general.
     monkeypatch.setenv("VLLM_WORKER_MULTIPROC_METHOD", "spawn")
-    monkeypatch.setenv("VLLM_ATTENTION_BACKEND", backend.name)
+
+    model_kwargs["attention_config"] = {"backend": backend.name}
 
     compilation_config = CompilationConfig(
         # Testing properties

tests/compile/fullgraph/test_basic_correctness.py

@@ -89,7 +89,6 @@ class TestSetting:
     ],
 )
 def test_compile_correctness(
-    monkeypatch: pytest.MonkeyPatch,
     test_setting: TestSetting,
 ):
     # this test is run under multiple suits, with different GPUs.
@@ -107,49 +106,48 @@ def test_compile_correctness(
             f"{cuda_device_count_stateless()}"
         )
 
-    with monkeypatch.context() as m:
-        m.setenv("VLLM_ATTENTION_BACKEND", attn_backend)
-        final_args = [
-            *model_args,
-            "-pp",
-            str(pp_size),
-            "-tp",
-            str(tp_size),
-            "-cc.cudagraph_mode=none",
-        ]
+    final_args = [
+        *model_args,
+        "-pp",
+        str(pp_size),
+        "-tp",
+        str(tp_size),
+        "-cc.cudagraph_mode=none",
+        f"--attention-backend={attn_backend}",
+    ]
 
-        all_args: list[list[str]] = []
-        all_envs: list[dict[str, str] | None] = []
+    all_args: list[list[str]] = []
+    all_envs: list[dict[str, str] | None] = []
 
-        for comp_mode in [
-            CompilationMode.STOCK_TORCH_COMPILE,
-            CompilationMode.DYNAMO_TRACE_ONCE,
-            CompilationMode.VLLM_COMPILE,
-        ]:
-            for mode in [CompilationMode.NONE, comp_mode]:
-                all_args.append(
-                    final_args + [f"-cc.mode={mode.name}", "-cc.backend=inductor"]
-                )
-
-            # inductor will change the output, so we only compare if the output
-            # is close, not exactly the same.
-            compare_all_settings(
-                model,
-                all_args,
-                all_envs,
-                method=method if method != "generate" else "generate_close",
+    for comp_mode in [
+        CompilationMode.STOCK_TORCH_COMPILE,
+        CompilationMode.DYNAMO_TRACE_ONCE,
+        CompilationMode.VLLM_COMPILE,
+    ]:
+        for mode in [CompilationMode.NONE, comp_mode]:
+            all_args.append(
+                final_args + [f"-cc.mode={mode.name}", "-cc.backend=inductor"]
             )
-            all_envs.clear()
-            all_args.clear()
 
-        for mode in [
-            CompilationMode.NONE,
-            CompilationMode.STOCK_TORCH_COMPILE,
-            CompilationMode.DYNAMO_TRACE_ONCE,
-            CompilationMode.VLLM_COMPILE,
-        ]:
-            all_args.append(final_args + [f"-cc.mode={mode.name}", "-cc.backend=eager"])
-            all_envs.append({})
-            all_envs.append({})
+        # inductor will change the output, so we only compare if the output
+        # is close, not exactly the same.
+        compare_all_settings(
+            model,
+            all_args,
+            all_envs,
+            method=method if method != "generate" else "generate_close",
+        )
+        all_envs.clear()
+        all_args.clear()
+
+    for mode in [
+        CompilationMode.NONE,
+        CompilationMode.STOCK_TORCH_COMPILE,
+        CompilationMode.DYNAMO_TRACE_ONCE,
+        CompilationMode.VLLM_COMPILE,
+    ]:
+        all_args.append(final_args + [f"-cc.mode={mode.name}", "-cc.backend=eager"])
+        all_envs.append({})
+        all_envs.append({})
 
-        compare_all_settings(model, all_args * 3, all_envs, method=method)
+    compare_all_settings(model, all_args * 3, all_envs, method=method)

tests/compile/fullgraph/test_full_cudagraph.py

@@ -74,7 +74,6 @@ def llm_pair(request):
         # Force native sampler to avoid potential nondeterminism in FlashInfer
         # when per-request generators are not used in V1.
         "VLLM_USE_FLASHINFER_SAMPLER": "0",
-        **backend_config.env_vars,
     }
     with temporary_environ(env_vars):
         full = LLM(
@@ -170,16 +169,10 @@ class TestFullCUDAGraph:
 
 @pytest.mark.skipif(not current_platform.is_cuda(), reason="Skip if not cuda")
 def test_full_cudagraph_with_invalid_backend():
-    with (
-        temporary_environ(
-            {
-                "VLLM_ATTENTION_BACKEND": "FLEX_ATTENTION",
-                # Flex_Attention is not supported with full cuda graph
-            }
-        ),
-        pytest.raises(RuntimeError),
-    ):
+    # Flex_Attention is not supported with full cuda graph
+    with pytest.raises(RuntimeError):
         LLM(
             model="Qwen/Qwen2-1.5B-Instruct",
             compilation_config=CompilationConfig(cudagraph_mode="FULL"),
+            attention_config={"backend": "FLEX_ATTENTION"},
         )

tests/compile/fullgraph/test_full_graph.py

@@ -197,20 +197,19 @@ def test_custom_compile_config(
     ],
 )
 def test_fp8_kv_scale_compile(
-    monkeypatch: pytest.MonkeyPatch,
     compilation_mode: int,
     model: str,
     backend: AttentionBackendEnum | None,
 ):
-    if backend:
-        monkeypatch.setenv("VLLM_ATTENTION_BACKEND", backend.name)
-
     model_kwargs = {
         "quantization": "fp8",
         "kv_cache_dtype": "fp8_e4m3",
         "calculate_kv_scales": True,
         "max_model_len": 512,
     }
+    if backend:
+        model_kwargs["attention_config"] = {"backend": backend.name}
+
     run_model(compilation_mode, model, **model_kwargs)
 
 

tests/distributed/test_context_parallel.py

@@ -219,14 +219,12 @@ def _test_cp_gsm8k(
         ]
     )
 
-    server_env = {}
     if attn_backend:
-        server_env["VLLM_ATTENTION_BACKEND"] = attn_backend
+        server_args.append(f"--attention-backend={attn_backend}")
 
     with RemoteOpenAIServer(
         model_id,
         server_args,
-        env_dict=server_env,
         max_wait_seconds=720,
     ) as remote_server:
         host = f"http://{remote_server.host}"

tests/distributed/test_pp_cudagraph.py

@@ -20,23 +20,21 @@ from ..utils import compare_two_settings, create_new_process_for_each_test
 )
 @create_new_process_for_each_test()
 def test_pp_cudagraph(
-    monkeypatch: pytest.MonkeyPatch,
     PP_SIZE: int,
     MODEL_NAME: str,
     ATTN_BACKEND: LiteralString,
 ):
-    with monkeypatch.context() as m:
-        cudagraph_args = [
-            # use half precision for speed and memory savings in CI environment
-            "--dtype",
-            "float16",
-            "--pipeline-parallel-size",
-            str(PP_SIZE),
-            "--distributed-executor-backend",
-            "mp",
-        ]
-        m.setenv("VLLM_ATTENTION_BACKEND", ATTN_BACKEND)
+    cudagraph_args = [
+        # use half precision for speed and memory savings in CI environment
+        "--dtype",
+        "float16",
+        "--pipeline-parallel-size",
+        str(PP_SIZE),
+        "--distributed-executor-backend",
+        "mp",
+        f"--attention-backend={ATTN_BACKEND}",
+    ]
 
-        eager_args = cudagraph_args + ["--enforce-eager"]
+    eager_args = cudagraph_args + ["--enforce-eager"]
 
-        compare_two_settings(MODEL_NAME, eager_args, cudagraph_args)
+    compare_two_settings(MODEL_NAME, eager_args, cudagraph_args)

tests/engine/test_arg_utils.py

@@ -9,7 +9,7 @@ from typing import Annotated, Literal
 
 import pytest
 
-from vllm.config import CompilationConfig, config
+from vllm.config import AttentionConfig, CompilationConfig, config
 from vllm.engine.arg_utils import (
     EngineArgs,
     contains_type,
@@ -298,6 +298,139 @@ def test_compilation_config():
     )
 
 
+def test_attention_config():
+    from vllm.attention.backends.registry import AttentionBackendEnum
+
+    parser = EngineArgs.add_cli_args(FlexibleArgumentParser())
+
+    # default value
+    args = parser.parse_args([])
+    assert args is not None
+    engine_args = EngineArgs.from_cli_args(args)
+    assert engine_args.attention_config == AttentionConfig()
+
+    # set backend via dot notation
+    args = parser.parse_args(["--attention-config.backend", "FLASH_ATTN"])
+    assert args is not None
+    engine_args = EngineArgs.from_cli_args(args)
+    assert engine_args.attention_config.backend is not None
+    assert engine_args.attention_config.backend.name == "FLASH_ATTN"
+
+    # set backend via --attention-backend shorthand
+    args = parser.parse_args(["--attention-backend", "FLASHINFER"])
+    assert args is not None
+    engine_args = EngineArgs.from_cli_args(args)
+    assert engine_args.attention_backend is not None
+    assert engine_args.attention_backend == "FLASHINFER"
+
+    # set all fields via dot notation
+    args = parser.parse_args(
+        [
+            "--attention-config.backend",
+            "FLASH_ATTN",
+            "--attention-config.flash_attn_version",
+            "3",
+            "--attention-config.use_prefill_decode_attention",
+            "true",
+            "--attention-config.flash_attn_max_num_splits_for_cuda_graph",
+            "16",
+            "--attention-config.use_cudnn_prefill",
+            "true",
+            "--attention-config.use_trtllm_ragged_deepseek_prefill",
+            "true",
+            "--attention-config.use_trtllm_attention",
+            "true",
+            "--attention-config.disable_flashinfer_prefill",
+            "true",
+            "--attention-config.disable_flashinfer_q_quantization",
+            "true",
+        ]
+    )
+    assert args is not None
+    engine_args = EngineArgs.from_cli_args(args)
+    assert engine_args.attention_config.backend is not None
+    assert engine_args.attention_config.backend.name == "FLASH_ATTN"
+    assert engine_args.attention_config.flash_attn_version == 3
+    assert engine_args.attention_config.use_prefill_decode_attention is True
+    assert engine_args.attention_config.flash_attn_max_num_splits_for_cuda_graph == 16
+    assert engine_args.attention_config.use_cudnn_prefill is True
+    assert engine_args.attention_config.use_trtllm_ragged_deepseek_prefill is True
+    assert engine_args.attention_config.use_trtllm_attention is True
+    assert engine_args.attention_config.disable_flashinfer_prefill is True
+    assert engine_args.attention_config.disable_flashinfer_q_quantization is True
+
+    # set to string form of a dict with all fields
+    args = parser.parse_args(
+        [
+            "--attention-config="
+            '{"backend": "FLASHINFER", "flash_attn_version": 2, '
+            '"use_prefill_decode_attention": false, '
+            '"flash_attn_max_num_splits_for_cuda_graph": 8, '
+            '"use_cudnn_prefill": false, '
+            '"use_trtllm_ragged_deepseek_prefill": false, '
+            '"use_trtllm_attention": false, '
+            '"disable_flashinfer_prefill": false, '
+            '"disable_flashinfer_q_quantization": false}',
+        ]
+    )
+    assert args is not None
+    engine_args = EngineArgs.from_cli_args(args)
+    assert engine_args.attention_config.backend is not None
+    assert engine_args.attention_config.backend.name == "FLASHINFER"
+    assert engine_args.attention_config.flash_attn_version == 2
+    assert engine_args.attention_config.use_prefill_decode_attention is False
+    assert engine_args.attention_config.flash_attn_max_num_splits_for_cuda_graph == 8
+    assert engine_args.attention_config.use_cudnn_prefill is False
+    assert engine_args.attention_config.use_trtllm_ragged_deepseek_prefill is False
+    assert engine_args.attention_config.use_trtllm_attention is False
+    assert engine_args.attention_config.disable_flashinfer_prefill is False
+    assert engine_args.attention_config.disable_flashinfer_q_quantization is False
+
+    # test --attention-backend flows into VllmConfig.attention_config
+    args = parser.parse_args(
+        [
+            "--model",
+            "facebook/opt-125m",
+            "--attention-backend",
+            "FLASH_ATTN",
+        ]
+    )
+    assert args is not None
+    engine_args = EngineArgs.from_cli_args(args)
+    vllm_config = engine_args.create_engine_config()
+    assert vllm_config.attention_config.backend == AttentionBackendEnum.FLASH_ATTN
+
+    # test --attention-config.backend flows into VllmConfig.attention_config
+    args = parser.parse_args(
+        [
+            "--model",
+            "facebook/opt-125m",
+            "--attention-config.backend",
+            "FLASHINFER",
+        ]
+    )
+    assert args is not None
+    engine_args = EngineArgs.from_cli_args(args)
+    vllm_config = engine_args.create_engine_config()
+    assert vllm_config.attention_config.backend == AttentionBackendEnum.FLASHINFER
+
+    # test --attention-backend and --attention-config.backend are mutually exclusive
+    args = parser.parse_args(
+        [
+            "--model",
+            "facebook/opt-125m",
+            "--attention-backend",
+            "FLASH_ATTN",
+            "--attention-config.backend",
+            "FLASHINFER",
+        ]
+    )
+    assert args is not None
+    engine_args = EngineArgs.from_cli_args(args)
+    with pytest.raises(ValueError, match="mutually exclusive"):
+        engine_args.create_engine_config()
+
+
 def test_prefix_cache_default():
     parser = EngineArgs.add_cli_args(FlexibleArgumentParser())
     args = parser.parse_args([])

tests/entrypoints/openai/test_serving_chat.py

@@ -76,15 +76,10 @@ def default_server_args(with_tool_parser: bool):
 
 
 @pytest.fixture(scope="module")
-def gptoss_server(
-    monkeypatch_module: pytest.MonkeyPatch, default_server_args: list[str]
-):
-    with monkeypatch_module.context() as m:
-        m.setenv("VLLM_ATTENTION_BACKEND", "TRITON_ATTN")
-        with RemoteOpenAIServer(
-            GPT_OSS_MODEL_NAME, default_server_args
-        ) as remote_server:
-            yield remote_server
+def gptoss_server(default_server_args: list[str]):
+    server_args = default_server_args + ["--attention-backend=TRITON_ATTN"]
+    with RemoteOpenAIServer(GPT_OSS_MODEL_NAME, server_args) as remote_server:
+        yield remote_server
 
 
 @pytest_asyncio.fixture

tests/kernels/attention/test_attention_selector.py

@@ -6,7 +6,9 @@ from unittest.mock import patch
 import pytest
 import torch
 
+from vllm.attention.backends.registry import AttentionBackendEnum
 from vllm.attention.selector import _cached_get_attn_backend, get_attn_backend
+from vllm.config import AttentionConfig, VllmConfig, set_current_vllm_config
 from vllm.platforms import current_platform
 from vllm.platforms.cpu import CpuPlatform
 from vllm.platforms.cuda import CudaPlatform
@@ -73,18 +75,18 @@ def generate_params():
 
 
 @pytest.mark.parametrize("device, name, use_mla, block_size", generate_params())
-def test_env(
+def test_backend_selection(
     device: str,
     name: str,
     use_mla: bool,
     block_size: int,
-    monkeypatch: pytest.MonkeyPatch,
 ):
     """Test attention backend selection with valid device-backend pairs."""
-    with monkeypatch.context() as m:
-        m.setenv("VLLM_ATTENTION_BACKEND", name)
-        m.setenv("VLLM_MLA_DISABLE", "1" if use_mla else "0")
+    # Create AttentionConfig with the specified backend
+    attention_config = AttentionConfig(backend=AttentionBackendEnum[name])
+    vllm_config = VllmConfig(attention_config=attention_config)
 
+    with set_current_vllm_config(vllm_config):
         if device == "cpu":
             with patch("vllm.platforms.current_platform", CpuPlatform()):
                 backend = get_attn_backend(16, torch.float16, None, block_size)
@@ -217,27 +219,32 @@ def test_env(
 @pytest.mark.parametrize("device", ["cpu", "cuda"])
 def test_fp32_fallback(device: str):
     """Test attention backend selection with fp32."""
-    if device == "cpu":
-        with patch("vllm.platforms.current_platform", CpuPlatform()):
-            backend = get_attn_backend(16, torch.float32, None, 16)
-        assert backend.get_name() == "CPU_ATTN"
+    # Use default config (no backend specified)
+    vllm_config = VllmConfig()
 
-    elif device == "cuda":
-        with patch("vllm.platforms.current_platform", CudaPlatform()):
-            backend = get_attn_backend(16, torch.float32, None, 16)
-        assert backend.get_name() == "FLEX_ATTENTION"
+    with set_current_vllm_config(vllm_config):
+        if device == "cpu":
+            with patch("vllm.platforms.current_platform", CpuPlatform()):
+                backend = get_attn_backend(16, torch.float32, None, 16)
+            assert backend.get_name() == "CPU_ATTN"
+
+        elif device == "cuda":
+            with patch("vllm.platforms.current_platform", CudaPlatform()):
+                backend = get_attn_backend(16, torch.float32, None, 16)
+            assert backend.get_name() == "FLEX_ATTENTION"
 
 
 def test_flash_attn(monkeypatch: pytest.MonkeyPatch):
     """Test FlashAttn validation."""
     pytest.skip(
         "Skipping as current backend selector does not "
-        "handle fallbacks when a backend is set via env var."
+        "handle fallbacks when a backend is explicitly set."
     )
 
-    with monkeypatch.context() as m:
-        m.setenv("VLLM_ATTENTION_BACKEND", "FLASH_ATTN")
+    attention_config = AttentionConfig(backend=AttentionBackendEnum.FLASH_ATTN)
+    vllm_config = VllmConfig(attention_config=attention_config)
 
+    with set_current_vllm_config(vllm_config):
         # Unsupported CUDA arch
         monkeypatch.setattr(torch.cuda, "get_device_capability", lambda _=None: (7, 5))
         backend = get_attn_backend(16, torch.float16, None, 16)
@@ -277,15 +284,10 @@ def test_flash_attn(monkeypatch: pytest.MonkeyPatch):
         assert backend.get_name() != "FLASH_ATTN"
 
 
-def test_invalid_env(monkeypatch: pytest.MonkeyPatch):
+def test_invalid_backend():
     """Test that invalid attention backend names raise ValueError."""
     with (
-        monkeypatch.context() as m,
-        patch("vllm.platforms.current_platform", CudaPlatform()),
+        pytest.raises(ValueError),
     ):
-        m.setenv("VLLM_ATTENTION_BACKEND", "INVALID")
-
-        # Should raise ValueError for invalid backend
-        with pytest.raises(ValueError) as exc_info:
-            get_attn_backend(32, torch.float16, None, 16)
-        assert "Invalid value 'INVALID'" in str(exc_info.value)
+        # Invalid backend name should raise ValueError when creating enum
+        AttentionConfig(backend=AttentionBackendEnum["INVALID"])

tests/kernels/attention/test_rocm_attention_selector.py

@@ -4,7 +4,9 @@
 import pytest
 import torch
 
+from vllm.attention.backends.registry import AttentionBackendEnum
 from vllm.attention.selector import _cached_get_attn_backend, get_attn_backend
+from vllm.config import AttentionConfig, VllmConfig, set_current_vllm_config
 from vllm.platforms.rocm import RocmPlatform
 
 
@@ -16,40 +18,56 @@ def clear_cache():
 
 @pytest.mark.skip(reason="Skipped for now. Should be revisited.")
 def test_selector(monkeypatch: pytest.MonkeyPatch):
-    with monkeypatch.context() as m:
-        m.setenv("VLLM_ATTENTION_BACKEND", "ROCM_ATTN")
+    # Set the current platform to ROCm using monkeypatch
+    monkeypatch.setattr("vllm.attention.selector.current_platform", RocmPlatform())
 
-        # Set the current platform to ROCm using monkeypatch
-        monkeypatch.setattr("vllm.attention.selector.current_platform", RocmPlatform())
+    # Test standard ROCm attention
+    attention_config = AttentionConfig(backend=AttentionBackendEnum.ROCM_ATTN)
+    vllm_config = VllmConfig(attention_config=attention_config)
 
-        # Test standard ROCm attention
+    with set_current_vllm_config(vllm_config):
         backend = get_attn_backend(16, torch.float16, torch.float16, 16, False)
         assert backend.get_name() == "ROCM_FLASH" or backend.get_name() == "TRITON_ATTN"
 
-        # MLA test for deepseek related
+    # MLA test for deepseek related
+    # Change the attention backend to triton MLA
+    attention_config = AttentionConfig(backend=AttentionBackendEnum.TRITON_MLA)
+    vllm_config = VllmConfig(attention_config=attention_config)
 
-        # change the attention backend to triton MLA
-        m.setenv("VLLM_ATTENTION_BACKEND", "TRITON_MLA")
+    with set_current_vllm_config(vllm_config):
         backend = get_attn_backend(576, torch.bfloat16, "auto", 16, False, use_mla=True)
         assert backend.get_name() == "TRITON_MLA"
 
-        # If attention backend is None
-        # If use_mla is true
-        # The selected backend is triton MLA
-        m.setenv("VLLM_ATTENTION_BACKEND", "")
+    # If attention backend is None
+    # If use_mla is true
+    # The selected backend is triton MLA
+    attention_config = AttentionConfig(backend=None)
+    vllm_config = VllmConfig(attention_config=attention_config)
+
+    with set_current_vllm_config(vllm_config):
         backend = get_attn_backend(576, torch.bfloat16, "auto", 16, False, use_mla=True)
         assert backend.get_name() == "TRITON_MLA"
 
-        # change the attention backend to AITER MLA
-        m.setenv("VLLM_ATTENTION_BACKEND", "ROCM_AITER_MLA")
+    # Change the attention backend to AITER MLA
+    attention_config = AttentionConfig(backend=AttentionBackendEnum.ROCM_AITER_MLA)
+    vllm_config = VllmConfig(attention_config=attention_config)
+
+    with set_current_vllm_config(vllm_config):
         backend = get_attn_backend(576, torch.bfloat16, "auto", 1, False, use_mla=True)
         assert backend.get_name() == "ROCM_AITER_MLA"
 
-        # If attention backend is None
-        # If use_mla is true
-        # If VLLM_ROCM_USE_AITER is enabled
-        # The selected backend is ROCM_AITER_MLA
-        m.setenv("VLLM_ATTENTION_BACKEND", "")
+    # If attention backend is None
+    # If use_mla is true
+    # If VLLM_ROCM_USE_AITER is enabled
+    # The selected backend is ROCM_AITER_MLA
+    with monkeypatch.context() as m:
         m.setenv("VLLM_ROCM_USE_AITER", "1")
-        backend = get_attn_backend(576, torch.bfloat16, "auto", 1, False, use_mla=True)
-        assert backend.get_name() == "ROCM_AITER_MLA"
+
+        attention_config = AttentionConfig(backend=None)
+        vllm_config = VllmConfig(attention_config=attention_config)
+
+        with set_current_vllm_config(vllm_config):
+            backend = get_attn_backend(
+                576, torch.bfloat16, "auto", 1, False, use_mla=True
+            )
+            assert backend.get_name() == "ROCM_AITER_MLA"

tests/kernels/test_flex_attention.py

@@ -37,7 +37,7 @@ def set_seed(seed):
     not torch.cuda.is_available() or TORCH_VERSION < MINIMUM_TORCH_VERSION,
     reason="CUDA not available or PyTorch version < 2.7",
 )
-def test_flex_attention_vs_default_backend(vllm_runner, monkeypatch):
+def test_flex_attention_vs_default_backend(vllm_runner):
     """Test that FlexAttention produces the same outputs as the default backend.
 
     This test compares the outputs from the FlexAttention backend with
@@ -54,35 +54,32 @@ def test_flex_attention_vs_default_backend(vllm_runner, monkeypatch):
     ]
 
     # Run with flex attention
-    with monkeypatch.context() as m:
-        m.setenv("VLLM_ATTENTION_BACKEND", "FLEX_ATTENTION")
-
-        set_seed(seed)
-        with vllm_runner(
-            model_name,
-            runner="generate",
-            tensor_parallel_size=1,
-            num_gpu_blocks_override=128,
-            enforce_eager=True,
-        ) as llm_flex:
-            output_flex = llm_flex.generate_greedy_logprobs(
-                prompts, max_tokens, num_logprobs
-            )
+    set_seed(seed)
+    with vllm_runner(
+        model_name,
+        runner="generate",
+        tensor_parallel_size=1,
+        num_gpu_blocks_override=128,
+        enforce_eager=True,
+        attention_config={"backend": "FLEX_ATTENTION"},
+    ) as llm_flex:
+        output_flex = llm_flex.generate_greedy_logprobs(
+            prompts, max_tokens, num_logprobs
+        )
 
     # Run with default backend
-    with monkeypatch.context() as m:
-        set_seed(seed)
-        with vllm_runner(
-            model_name,
-            runner="generate",
-            tensor_parallel_size=1,
-            num_gpu_blocks_override=128,
-            enforce_eager=True,
-            gpu_memory_utilization=0.85,
-        ) as llm_default:
-            output_default = llm_default.generate_greedy_logprobs(
-                prompts, max_tokens, num_logprobs
-            )
+    set_seed(seed)
+    with vllm_runner(
+        model_name,
+        runner="generate",
+        tensor_parallel_size=1,
+        num_gpu_blocks_override=128,
+        enforce_eager=True,
+        gpu_memory_utilization=0.85,
+    ) as llm_default:
+        output_default = llm_default.generate_greedy_logprobs(
+            prompts, max_tokens, num_logprobs
+        )
 
     check_logprobs_close(
         outputs_0_lst=output_flex,
@@ -96,7 +93,7 @@ def test_flex_attention_vs_default_backend(vllm_runner, monkeypatch):
     not torch.cuda.is_available() or TORCH_VERSION < MINIMUM_TORCH_VERSION,
     reason="CUDA not available or PyTorch version < 2.7",
 )
-def test_encoder_flex_attention_vs_default_backend(vllm_runner, monkeypatch):
+def test_encoder_flex_attention_vs_default_backend(vllm_runner):
     """Test that FlexAttention produces the same outputs as the default backend.
 
     This test compares the outputs from the FlexAttention backend with
@@ -110,30 +107,26 @@ def test_encoder_flex_attention_vs_default_backend(vllm_runner, monkeypatch):
     ]
 
     # Run with flex attention
-    with monkeypatch.context() as m:
-        m.setenv("VLLM_ATTENTION_BACKEND", "FLEX_ATTENTION")
-        with vllm_runner(
-            model_name,
-            runner="pooling",
-            dtype=torch.bfloat16,
-            tensor_parallel_size=1,
-            max_model_len=100,
-            enforce_eager=True,
-        ) as llm_flex:
-            flex_outputs = llm_flex.embed(prompts)
+    with vllm_runner(
+        model_name,
+        runner="pooling",
+        dtype=torch.bfloat16,
+        tensor_parallel_size=1,
+        max_model_len=100,
+        enforce_eager=True,
+        attention_config={"backend": "FLEX_ATTENTION"},
+    ) as llm_flex:
+        flex_outputs = llm_flex.embed(prompts)
 
     # Run with default backend
-    with (
-        monkeypatch.context() as m,
-        vllm_runner(
-            model_name,
-            runner="pooling",
-            dtype=torch.bfloat16,
-            tensor_parallel_size=1,
-            max_model_len=100,
-            enforce_eager=True,
-        ) as llm_default,
-    ):
+    with vllm_runner(
+        model_name,
+        runner="pooling",
+        dtype=torch.bfloat16,
+        tensor_parallel_size=1,
+        max_model_len=100,
+        enforce_eager=True,
+    ) as llm_default:
         default_outputs = llm_default.embed(prompts)
 
     check_embeddings_close(

tests/models/multimodal/generation/test_granite_speech.py

@@ -35,10 +35,12 @@ audio_lora_path = MODEL_NAME
 models = [MODEL_NAME]
 
 
-@pytest.fixture(autouse=True)
-def set_attention_backend_for_rocm(monkeypatch):
+@pytest.fixture
+def granite_speech_attention_config():
+    """Return attention config for Granite Speech tests on ROCm."""
     if current_platform.is_rocm():
-        monkeypatch.setenv("VLLM_ATTENTION_BACKEND", "TRITON_ATTN")
+        return {"backend": "TRITON_ATTN"}
+    return None
 
 
 def run_test(
@@ -53,6 +55,7 @@ def run_test(
     num_logprobs: int,
     tensor_parallel_size: int,
     distributed_executor_backend: str | None = None,
+    attention_config: dict | None = None,
 ):
     """Inference result should be the same between hf and vllm.
 
@@ -80,6 +83,7 @@ def run_test(
         enable_lora=True,
         max_lora_rank=64,
         enforce_eager=True,
+        attention_config=attention_config,
     ) as vllm_model:
         lora_request = LoRARequest("audio", 1, audio_lora_path)
         vllm_outputs_per_case = [
@@ -131,6 +135,7 @@ def test_models(
     vllm_runner,
     model: str,
     audio_assets: AudioTestAssets,
+    granite_speech_attention_config,
     dtype: str,
     max_model_len: int,
     max_tokens: int,
@@ -157,4 +162,5 @@ def test_models(
         max_tokens=max_tokens,
         num_logprobs=num_logprobs,
         tensor_parallel_size=1,
+        attention_config=granite_speech_attention_config,
     )

tests/models/multimodal/pooling/conftest.py

@@ -2,23 +2,17 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """Pytest configuration for vLLM pooling tests."""
 
-import os
-import warnings
+import pytest
 
 from vllm.platforms import current_platform
 
 
-def pytest_collection_modifyitems(config, items):
-    """Set FLEX_ATTENTION backend for SigLIP tests on ROCm."""
-    if not current_platform.is_rocm():
-        return
+@pytest.fixture
+def siglip_attention_config():
+    """Return attention config for SigLIP tests on ROCm.
 
-    siglip_tests = [item for item in items if "test_siglip" in item.nodeid]
-
-    if siglip_tests:
-        os.environ["VLLM_ATTENTION_BACKEND"] = "FLEX_ATTENTION"
-        warnings.warn(
-            "ROCm: Set VLLM_ATTENTION_BACKEND=FLEX_ATTENTION for SigLIP tests",
-            UserWarning,
-            stacklevel=1,
-        )
+    On ROCm, SigLIP tests require FLEX_ATTENTION backend.
+    """
+    if current_platform.is_rocm():
+        return {"backend": "FLEX_ATTENTION"}
+    return None

tests/models/multimodal/pooling/test_siglip.py

@@ -38,6 +38,7 @@ def _run_test(
     *,
     dtype: str,
     tokenization_kwargs: dict[str, Any] | None = None,
+    attention_config: dict[str, Any] | None = None,
 ) -> None:
     if tokenization_kwargs is None:
         tokenization_kwargs = {}
@@ -49,6 +50,7 @@ def _run_test(
         enforce_eager=True,
         max_model_len=64,
         gpu_memory_utilization=0.7,
+        attention_config=attention_config,
     ) as vllm_model:
         vllm_outputs = vllm_model.embed(
             input_texts, images=input_images, tokenization_kwargs=tokenization_kwargs
@@ -90,6 +92,7 @@ def test_models_text(
     hf_runner,
     vllm_runner,
     image_assets,
+    siglip_attention_config,
     model: str,
     dtype: str,
 ) -> None:
@@ -108,6 +111,7 @@ def test_models_text(
             "padding": "max_length",
             "max_length": 64,
         },  # siglip2 was trained with this padding setting.
+        attention_config=siglip_attention_config,
     )
 
 
@@ -117,6 +121,7 @@ def test_models_image(
     hf_runner,
     vllm_runner,
     image_assets,
+    siglip_attention_config,
     model: str,
     dtype: str,
 ) -> None:
@@ -133,6 +138,7 @@ def test_models_image(
         input_images,
         model,
         dtype=dtype,
+        attention_config=siglip_attention_config,
     )
 
 
@@ -141,6 +147,7 @@ def test_models_image(
 def test_models_text_image_no_crash(
     vllm_runner,
     image_assets,
+    siglip_attention_config,
     model: str,
     dtype: str,
 ) -> None:
@@ -154,6 +161,7 @@ def test_models_text_image_no_crash(
         enforce_eager=True,
         max_model_len=64,
         gpu_memory_utilization=0.7,
+        attention_config=siglip_attention_config,
     ) as vllm_model:
         with pytest.raises(ValueError, match="not both"):
             vllm_model.embed(texts, images=images)

tests/models/quantization/test_fp8.py

@@ -75,7 +75,6 @@ def test_models(
 
     with monkeypatch.context() as m:
         m.setenv("TOKENIZERS_PARALLELISM", "true")
-        m.setenv("VLLM_ATTENTION_BACKEND", backend)
 
         MAX_MODEL_LEN = 1024
         NUM_LOG_PROBS = 8
@@ -86,6 +85,7 @@ def test_models(
             tensor_parallel_size=tensor_parallel_size,
             enforce_eager=enforce_eager,
             kv_cache_dtype="auto",
+            attention_config={"backend": backend},
         ) as vllm_model:
             baseline_outputs = vllm_model.generate_greedy_logprobs(
                 example_prompts, max_tokens, NUM_LOG_PROBS
@@ -97,6 +97,7 @@ def test_models(
             tensor_parallel_size=tensor_parallel_size,
             enforce_eager=enforce_eager,
             kv_cache_dtype=kv_cache_dtype,
+            attention_config={"backend": backend},
         ) as vllm_model:
             test_outputs = vllm_model.generate_greedy_logprobs(
                 example_prompts, max_tokens, NUM_LOG_PROBS

tests/models/test_initialization.py

@@ -108,11 +108,12 @@ def can_initialize(
         patch.object(V1EngineCore, "_initialize_kv_caches", _initialize_kv_caches_v1),
         monkeypatch.context() as m,
     ):
-        if model_arch == "GptOssForCausalLM":
-            # FIXME: A hack to bypass FA3 assertion because our CI's L4 GPU
-            # has cc==8.9 which hasn't supported FA3 yet. Remove this hack when
-            # L4 supports FA3.
-            m.setenv("VLLM_ATTENTION_BACKEND", "TRITON_ATTN")
+        # FIXME: A hack to bypass FA3 assertion because our CI's L4 GPU
+        # has cc==8.9 which hasn't supported FA3 yet. Remove this hack when
+        # L4 supports FA3.
+        attention_config = (
+            {"backend": "TRITON_ATTN"} if model_arch == "GptOssForCausalLM" else None
+        )
         if model_arch == "WhisperForConditionalGeneration":
             m.setenv("VLLM_WORKER_MULTIPROC_METHOD", "spawn")
 
@@ -143,6 +144,7 @@ def can_initialize(
             else "vllm",
             hf_overrides=hf_overrides_fn,
             max_num_seqs=model_info.max_num_seqs,
+            attention_config=attention_config,
         )
 
 

tests/v1/attention/test_rocm_attention_backends_selection.py

@@ -94,26 +94,20 @@ def mock_on_gfx9():
             None,
             AttentionBackendEnum.ROCM_AITER_UNIFIED_ATTN.get_path(),
         ),
-        # Test Case 9: VLLM_V1_USE_PREFILL_DECODE_ATTENTION=1
-        (
-            {"VLLM_V1_USE_PREFILL_DECODE_ATTENTION": "1"},
-            None,
-            AttentionBackendEnum.ROCM_ATTN.get_path(),
-        ),
-        # Test Case 10: VLLM_ROCM_USE_AITER=1 + explicit TRITON_ATTN
+        # Test Case 9: VLLM_ROCM_USE_AITER=1 + explicit TRITON_ATTN
         (
             {"VLLM_ROCM_USE_AITER": "1"},
             "TRITON_ATTN",
             AttentionBackendEnum.TRITON_ATTN.get_path(),
         ),
-        # Test Case 11: VLLM_ROCM_USE_AITER=1 + VLLM_ROCM_USE_AITER_MHA=0
+        # Test Case 10: VLLM_ROCM_USE_AITER=1 + VLLM_ROCM_USE_AITER_MHA=0
         # (explicitly disabled)
         (
             {"VLLM_ROCM_USE_AITER": "1", "VLLM_ROCM_USE_AITER_MHA": "0"},
             None,
             AttentionBackendEnum.TRITON_ATTN.get_path(),
         ),
-        # Test Case 12: VLLM_ROCM_USE_AITER=1 + explicit ROCM_ATTN
+        # Test Case 11: VLLM_ROCM_USE_AITER=1 + explicit ROCM_ATTN
         (
             {"VLLM_ROCM_USE_AITER": "1"},
             "ROCM_ATTN",

tests/v1/attention/utils.py

@@ -249,8 +249,8 @@ def create_dummy_kv_cache(
 @dataclass
 class BackendConfig:
     name: str
-    env_vars: dict
-    comp_config: dict  # compilation config
+    attention_config: dict
+    comp_config: dict
     specific_gpu_arch: tuple | None = None
 
 
@@ -259,10 +259,10 @@ full_cg_backend_configs = {
     # FA3 on Hopper
     "FA3": BackendConfig(
         name="FA3",
-        env_vars={
-            "VLLM_ATTENTION_BACKEND": "FLASH_ATTN",
-            "VLLM_FLASH_ATTN_VERSION": "3",
-            "VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH": "16",
+        attention_config={
+            "backend": "FLASH_ATTN",
+            "flash_attn_version": 3,
+            "flash_attn_max_num_splits_for_cuda_graph": 16,
         },
         comp_config={
             "cudagraph_mode": "FULL",
@@ -272,9 +272,7 @@ full_cg_backend_configs = {
     # FlashMLA on Hopper
     "FlashMLA": BackendConfig(
         name="FlashMLA",
-        env_vars={
-            "VLLM_ATTENTION_BACKEND": "FLASHMLA",
-        },
+        attention_config={"backend": "FLASHMLA"},
         comp_config={
             "cudagraph_mode": "FULL_AND_PIECEWISE",
         },
@@ -283,9 +281,7 @@ full_cg_backend_configs = {
     # Cutlass MLA on Blackwell
     "CutlassMLA": BackendConfig(
         name="CutlassMLA",
-        env_vars={
-            "VLLM_ATTENTION_BACKEND": "CUTLASS_MLA",
-        },
+        attention_config={"backend": "CUTLASS_MLA"},
         comp_config={
             "cudagraph_mode": "FULL_AND_PIECEWISE",
         },
@@ -294,9 +290,7 @@ full_cg_backend_configs = {
     # FlashInfer MLA on Blackwell
     "FlashInferMLA": BackendConfig(
         name="FlashInferMLA",
-        env_vars={
-            "VLLM_ATTENTION_BACKEND": "FLASHINFER_MLA",
-        },
+        attention_config={"backend": "FLASHINFER_MLA"},
         comp_config={
             "cudagraph_mode": "FULL_AND_PIECEWISE",
         },
@@ -305,9 +299,9 @@ full_cg_backend_configs = {
     # FlashAttention MLA on Hopper
     "FlashAttentionMLA": BackendConfig(
         name="FlashAttentionMLA",
-        env_vars={
-            "VLLM_ATTENTION_BACKEND": "FLASH_ATTN_MLA",
-            "VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH": "16",
+        attention_config={
+            "backend": "FLASH_ATTN_MLA",
+            "flash_attn_max_num_splits_for_cuda_graph": 16,
         },
         comp_config={
             "cudagraph_mode": "FULL_DECODE_ONLY",
@@ -317,10 +311,10 @@ full_cg_backend_configs = {
     # FA2
     "FA2": BackendConfig(
         name="FA2",
-        env_vars={
-            "VLLM_ATTENTION_BACKEND": "FLASH_ATTN",
-            "VLLM_FLASH_ATTN_VERSION": "2",
-            "VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH": "16",
+        attention_config={
+            "backend": "FLASH_ATTN",
+            "flash_attn_version": 2,
+            "flash_attn_max_num_splits_for_cuda_graph": 16,
         },
         comp_config={
             "cudagraph_mode": "FULL_AND_PIECEWISE",
@@ -329,7 +323,7 @@ full_cg_backend_configs = {
     # Triton Attention
     "TritonAttn": BackendConfig(
         name="TritonAttn",
-        env_vars={"VLLM_ATTENTION_BACKEND": "TRITON_ATTN"},
+        attention_config={"backend": "TRITON_ATTN"},
         comp_config={
             "cudagraph_mode": "FULL_AND_PIECEWISE",
         },
@@ -337,14 +331,17 @@ full_cg_backend_configs = {
     # FlashInfer
     "FlashInfer": BackendConfig(
         name="FlashInfer",
-        env_vars={"VLLM_ATTENTION_BACKEND": "FLASHINFER"},
+        attention_config={"backend": "FLASHINFER"},
         comp_config={
             "cudagraph_mode": "FULL_AND_PIECEWISE",
         },
     ),
     "RocmAttn": BackendConfig(
         name="RocmAttn",
-        env_vars={"VLLM_V1_USE_PREFILL_DECODE_ATTENTION": "1"},
+        attention_config={
+            "backend": "ROCM_ATTN",
+            "use_prefill_decode_attention": True,
+        },
         comp_config={
             "cudagraph_mode": "FULL",
         },