vllm.envs ¶

CMAKE_BUILD_TYPE: (
    Literal["Debug", "Release", "RelWithDebInfo"] | None
) = None

CUDA_VISIBLE_DEVICES: str | None = None

GPT_OSS_SYSTEM_TOOL_MCP_LABELS: list[str] = []

K_SCALE_CONSTANT: int = 200

LD_LIBRARY_PATH: str | None = None

LOCAL_RANK: int = 0

MAX_JOBS: str | None = None

NVCC_THREADS: str | None = None

Q_SCALE_CONSTANT: int = 200

S3_ACCESS_KEY_ID: str | None = None

S3_ENDPOINT_URL: str | None = None

S3_SECRET_ACCESS_KEY: str | None = None

VERBOSE: bool = False

VLLM_ALL2ALL_BACKEND: Literal[
    "naive",
    "pplx",
    "deepep_high_throughput",
    "deepep_low_latency",
    "allgather_reducescatter",
    "flashinfer_all2allv",
] = "allgather_reducescatter"

VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE: bool = (
    False
)

VLLM_ALLOW_INSECURE_SERIALIZATION: bool = False

VLLM_ALLOW_LONG_MAX_MODEL_LEN: bool = False

VLLM_ALLOW_RUNTIME_LORA_UPDATING: bool = False

VLLM_ALLREDUCE_USE_SYMM_MEM: bool = True

VLLM_API_KEY: str | None = None

VLLM_ASSETS_CACHE: str = join(VLLM_CACHE_ROOT, 'assets')

VLLM_ASSETS_CACHE_MODEL_CLEAN: bool = False

VLLM_ATTENTION_BACKEND: str | None = None

VLLM_AUDIO_FETCH_TIMEOUT: int = 10

VLLM_CACHE_ROOT: str = expanduser('~/.cache/vllm')

VLLM_COMPUTE_NANS_IN_LOGITS: bool = False

VLLM_CONFIGURE_LOGGING: int = 1

VLLM_CONFIG_ROOT: str = expanduser('~/.config/vllm')

VLLM_CPU_KVCACHE_SPACE: int | None = 0

VLLM_CPU_MOE_PREPACK: bool = True

VLLM_CPU_NUM_OF_RESERVED_CPU: int | None = None

VLLM_CPU_OMP_THREADS_BIND: str = ''

VLLM_CPU_SGL_KERNEL: bool = False

VLLM_CUDART_SO_PATH: str | None = None

VLLM_CUSTOM_SCOPES_FOR_PROFILING: bool = False

VLLM_DBO_COMM_SMS: int = 20

VLLM_DEBUG_DUMP_PATH: str | None = None

VLLM_DEBUG_LOG_API_SERVER_RESPONSE: bool = False

VLLM_DEEPEP_BUFFER_SIZE_MB: int = 1024

VLLM_DEEP_GEMM_WARMUP: Literal["skip", "full", "relax"] = (
    "relax"
)

VLLM_DISABLED_KERNELS: list[str] = []

VLLM_DISABLE_COMPILE_CACHE: bool = False

VLLM_DISABLE_FLASHINFER_PREFILL: bool = False

VLLM_DISABLE_PYNCCL: bool = False

VLLM_DOCKER_BUILD_CONTEXT: bool = False

VLLM_DO_NOT_TRACK: bool = False

VLLM_DP_MASTER_IP: str = ''

VLLM_DP_MASTER_PORT: int = 0

VLLM_DP_RANK: int = 0

VLLM_DP_RANK_LOCAL: int = -1

VLLM_DP_SIZE: int = 1

VLLM_ENABLE_CUDAGRAPH_GC: bool = False

VLLM_ENABLE_FUSED_MOE_ACTIVATION_CHUNKING: bool = True

VLLM_ENABLE_INDUCTOR_COORDINATE_DESCENT_TUNING: bool = True

VLLM_ENABLE_INDUCTOR_MAX_AUTOTUNE: bool = True

VLLM_ENABLE_RESPONSES_API_STORE: bool = False

VLLM_ENABLE_V1_MULTIPROCESSING: bool = True

VLLM_ENGINE_ITERATION_TIMEOUT_S: int = 60

VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: int = 300

VLLM_FLASHINFER_DISABLE_Q_QUANTIZATION: bool = False

VLLM_FLASHINFER_MOE_BACKEND: Literal[
    "throughput", "latency"
] = "throughput"

VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH: int = 32

VLLM_FLASH_ATTN_VERSION: int | None = None

VLLM_FORCE_AOT_LOAD: bool = False

VLLM_FUSED_MOE_CHUNK_SIZE: int = 64 * 1024

VLLM_GC_DEBUG: str = ''

VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS: bool = False

VLLM_HAS_FLASHINFER_CUBIN: bool = False

VLLM_HOST_IP: str = ''

VLLM_HTTP_TIMEOUT_KEEP_ALIVE: int = 5

VLLM_IMAGE_FETCH_TIMEOUT: int = 5

VLLM_KEEP_ALIVE_ON_ENGINE_DEATH: bool = False

VLLM_KV_CACHE_LAYOUT: Literal['NHD', 'HND'] | None = None

VLLM_KV_EVENTS_USE_INT_BLOCK_HASHES: bool = True

VLLM_LOGGING_CONFIG_PATH: str | None = None

VLLM_LOGGING_LEVEL: str = 'INFO'

VLLM_LOGGING_PREFIX: str = ''

VLLM_LOGGING_STREAM: str = 'ext://sys.stdout'

VLLM_LOGITS_PROCESSOR_THREADS: int | None = None

VLLM_LOG_BATCHSIZE_INTERVAL: float = -1

VLLM_LOG_STATS_INTERVAL: float = 10.0

VLLM_LOOPBACK_IP: str = ''

VLLM_LORA_RESOLVER_CACHE_DIR: str | None = None

VLLM_MAIN_CUDA_VERSION: str = '12.8'

VLLM_MARLIN_USE_ATOMIC_ADD: bool = False

VLLM_MAX_AUDIO_CLIP_FILESIZE_MB: int = 25

VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE: int = 163840

VLLM_MEDIA_LOADING_THREAD_COUNT: int = 8

VLLM_MEDIA_URL_ALLOW_REDIRECTS: bool = True

VLLM_MLA_DISABLE: bool = False

VLLM_MM_INPUT_CACHE_GIB: int = 4

VLLM_MODEL_REDIRECT_PATH: str | None = None

VLLM_MOE_DP_CHUNK_SIZE: int = 256

VLLM_MQ_MAX_CHUNK_BYTES_MB: int = 16

VLLM_MSGPACK_ZERO_COPY_THRESHOLD: int = 256

VLLM_MXFP4_USE_MARLIN: bool | None = None

VLLM_NCCL_INCLUDE_PATH: str | None = None

VLLM_NCCL_SO_PATH: str | None = None

VLLM_NIXL_ABORT_REQUEST_TIMEOUT: int = 480

VLLM_NIXL_SIDE_CHANNEL_HOST: str = 'localhost'

VLLM_NIXL_SIDE_CHANNEL_PORT: int = 5600

VLLM_NO_USAGE_STATS: bool = False

VLLM_NVTX_SCOPES_FOR_PROFILING: bool = False

VLLM_OBJECT_STORAGE_SHM_BUFFER_NAME: str = (
    "VLLM_OBJECT_STORAGE_SHM_BUFFER"
)

VLLM_PATTERN_MATCH_DEBUG: str | None = None

VLLM_PLUGINS: list[str] | None = None

VLLM_PORT: int | None = None

VLLM_PP_LAYER_PARTITION: str | None = None

VLLM_RANDOMIZE_DP_DUMMY_INPUTS: bool = False

VLLM_RAY_BUNDLE_INDICES: str = ''

VLLM_RAY_DP_PACK_STRATEGY: str = 'strict'

VLLM_RAY_PER_WORKER_GPUS: float = 1.0

VLLM_RINGBUFFER_WARNING_INTERVAL: int = 60

VLLM_ROCM_CUSTOM_PAGED_ATTN: bool = True

VLLM_ROCM_FP8_MFMA_PAGE_ATTN: bool = False

VLLM_ROCM_FP8_PADDING: bool = True

VLLM_ROCM_MOE_PADDING: bool = True

VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16: bool = True

VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB: int | None = None

VLLM_ROCM_QUICK_REDUCE_QUANTIZATION: Literal[
    "FP", "INT8", "INT6", "INT4", "NONE"
] = "NONE"

VLLM_ROCM_USE_AITER: bool = False

VLLM_ROCM_USE_AITER_FP4_ASM_GEMM: bool = False

VLLM_ROCM_USE_AITER_FP8BMM: bool = True

VLLM_ROCM_USE_AITER_LINEAR: bool = True

VLLM_ROCM_USE_AITER_MHA: bool = True

VLLM_ROCM_USE_AITER_MLA: bool = True

VLLM_ROCM_USE_AITER_MOE: bool = True

VLLM_ROCM_USE_AITER_PAGED_ATTN: bool = False

VLLM_ROCM_USE_AITER_RMSNORM: bool = True

VLLM_ROCM_USE_AITER_UNIFIED_ATTENTION: bool = False

VLLM_ROCM_USE_SKINNY_GEMM: bool = True

VLLM_ROCM_USE_TRITON_ROPE: bool = False

VLLM_RPC_BASE_PATH: str = gettempdir()

VLLM_RPC_TIMEOUT: int = 10000

VLLM_SERVER_DEV_MODE: bool = False

VLLM_SKIP_P2P_CHECK: bool = False

VLLM_SLEEP_WHEN_IDLE: bool = False

VLLM_TARGET_DEVICE: str = 'cuda'

VLLM_TEST_USE_PRECOMPILED_NIGHTLY_WHEEL: bool = False

VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS: int = 1

VLLM_TORCH_PROFILER_DIR: str | None = None

VLLM_TORCH_PROFILER_RECORD_SHAPES: bool = False

VLLM_TORCH_PROFILER_WITH_FLOPS: bool = False

VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY: bool = False

VLLM_TORCH_PROFILER_WITH_STACK: bool = True

VLLM_TPU_BUCKET_PADDING_GAP: int = 0

VLLM_TPU_MOST_MODEL_LEN: int | None = None

VLLM_TPU_USING_PATHWAYS: bool = False

VLLM_TRACE_FUNCTION: int = 0

VLLM_TUNED_CONFIG_FOLDER: str | None = None

VLLM_USAGE_SOURCE: str = ''

VLLM_USAGE_STATS_SERVER: str = 'https://stats.vllm.ai'

VLLM_USE_AOT_COMPILE: bool = False

VLLM_USE_CUDNN_PREFILL: bool = False

VLLM_USE_DEEP_GEMM: bool = True

VLLM_USE_DEEP_GEMM_E8M0: bool = True

VLLM_USE_FBGEMM: bool = False

VLLM_USE_FLASHINFER_MOE_FP16: bool = False

VLLM_USE_FLASHINFER_MOE_FP4: bool = False

VLLM_USE_FLASHINFER_MOE_FP8: bool = False

VLLM_USE_FLASHINFER_MOE_MXFP4_BF16: bool = False

VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8: bool = False

VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS: bool = False

VLLM_USE_FLASHINFER_SAMPLER: bool | None = None

VLLM_USE_FUSED_MOE_GROUPED_TOPK: bool = True

VLLM_USE_MODELSCOPE: bool = False

VLLM_USE_NCCL_SYMM_MEM: bool = False

VLLM_USE_NVFP4_CT_EMULATIONS: bool = False

VLLM_USE_PRECOMPILED: bool = False

VLLM_USE_RAY_COMPILED_DAG: bool = False

VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE: Literal[
    "auto", "nccl", "shm"
] = "auto"

VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM: bool = False

VLLM_USE_RAY_SPMD_WORKER: bool = False

VLLM_USE_RAY_WRAPPED_PP_COMM: bool = True

VLLM_USE_STANDALONE_COMPILE: bool = False

VLLM_USE_TRITON_AWQ: bool = False

VLLM_USE_TRITON_FLASH_ATTN: bool = True

VLLM_USE_TRTLLM_ATTENTION: str | None = None

VLLM_USE_V1: bool = True

VLLM_V0_USE_OUTLINES_CACHE: bool = False

VLLM_V1_OUTPUT_PROC_CHUNK_SIZE: int = 128

VLLM_V1_USE_OUTLINES_CACHE: bool = False

VLLM_V1_USE_PREFILL_DECODE_ATTENTION: bool = False

VLLM_VIDEO_FETCH_TIMEOUT: int = 30

VLLM_VIDEO_LOADER_BACKEND: str = 'opencv'

VLLM_WORKER_MULTIPROC_METHOD: Literal["fork", "spawn"] = (
    "fork"
)

VLLM_XGRAMMAR_CACHE_MB: int = 0

VLLM_XLA_CACHE_PATH: str = join(
    VLLM_CACHE_ROOT, "xla_cache"
)

VLLM_XLA_CHECK_RECOMPILATION: bool = False

VLLM_XLA_USE_SPMD: bool = False

V_SCALE_CONSTANT: int = 100

environment_variables: dict[str, Callable[[], Any]] = {
    "VLLM_TARGET_DEVICE": lambda: lower(),
    "VLLM_MAIN_CUDA_VERSION": lambda: lower() or "12.8",
    "MAX_JOBS": lambda: getenv("MAX_JOBS", None),
    "NVCC_THREADS": lambda: getenv("NVCC_THREADS", None),
    "VLLM_USE_PRECOMPILED": lambda: lower() in ("1", "true")
    or bool(get("VLLM_PRECOMPILED_WHEEL_LOCATION")),
    "VLLM_DOCKER_BUILD_CONTEXT": lambda: lower()
    in ("1", "true"),
    "VLLM_TEST_USE_PRECOMPILED_NIGHTLY_WHEEL": lambda: bool(
        int(
            getenv(
                "VLLM_TEST_USE_PRECOMPILED_NIGHTLY_WHEEL",
                "0",
            )
        )
    ),
    "CMAKE_BUILD_TYPE": env_with_choices(
        "CMAKE_BUILD_TYPE",
        None,
        ["Debug", "Release", "RelWithDebInfo"],
    ),
    "VERBOSE": lambda: bool(int(getenv("VERBOSE", "0"))),
    "VLLM_CONFIG_ROOT": lambda: expanduser(
        getenv(
            "VLLM_CONFIG_ROOT",
            join(get_default_config_root(), "vllm"),
        )
    ),
    "VLLM_CACHE_ROOT": lambda: expanduser(
        getenv(
            "VLLM_CACHE_ROOT",
            join(get_default_cache_root(), "vllm"),
        )
    ),
    "VLLM_HOST_IP": lambda: getenv("VLLM_HOST_IP", ""),
    "VLLM_PORT": get_vllm_port,
    "VLLM_RPC_BASE_PATH": lambda: getenv(
        "VLLM_RPC_BASE_PATH", gettempdir()
    ),
    "VLLM_USE_MODELSCOPE": lambda: lower() == "true",
    "VLLM_RINGBUFFER_WARNING_INTERVAL": lambda: int(
        get("VLLM_RINGBUFFER_WARNING_INTERVAL", "60")
    ),
    "CUDA_HOME": lambda: get("CUDA_HOME", None),
    "VLLM_NCCL_SO_PATH": lambda: get(
        "VLLM_NCCL_SO_PATH", None
    ),
    "LD_LIBRARY_PATH": lambda: get("LD_LIBRARY_PATH", None),
    "VLLM_USE_TRITON_FLASH_ATTN": lambda: lower()
    in ("true", "1"),
    "VLLM_V1_USE_PREFILL_DECODE_ATTENTION": lambda: lower()
    in ("true", "1"),
    "VLLM_FLASH_ATTN_VERSION": lambda: maybe_convert_int(
        get("VLLM_FLASH_ATTN_VERSION", None)
    ),
    "VLLM_USE_STANDALONE_COMPILE": lambda: get(
        "VLLM_USE_STANDALONE_COMPILE", "0"
    )
    == "1",
    "VLLM_PATTERN_MATCH_DEBUG": lambda: get(
        "VLLM_PATTERN_MATCH_DEBUG", None
    ),
    "VLLM_DEBUG_DUMP_PATH": lambda: get(
        "VLLM_DEBUG_DUMP_PATH", None
    ),
    "VLLM_USE_AOT_COMPILE": use_aot_compile,
    "VLLM_FORCE_AOT_LOAD": lambda: get(
        "VLLM_FORCE_AOT_LOAD", "0"
    )
    == "1",
    "LOCAL_RANK": lambda: int(get("LOCAL_RANK", "0")),
    "CUDA_VISIBLE_DEVICES": lambda: get(
        "CUDA_VISIBLE_DEVICES", None
    ),
    "VLLM_ENGINE_ITERATION_TIMEOUT_S": lambda: int(
        get("VLLM_ENGINE_ITERATION_TIMEOUT_S", "60")
    ),
    "VLLM_API_KEY": lambda: get("VLLM_API_KEY", None),
    "VLLM_DEBUG_LOG_API_SERVER_RESPONSE": lambda: lower()
    == "true",
    "S3_ACCESS_KEY_ID": lambda: get(
        "S3_ACCESS_KEY_ID", None
    ),
    "S3_SECRET_ACCESS_KEY": lambda: get(
        "S3_SECRET_ACCESS_KEY", None
    ),
    "S3_ENDPOINT_URL": lambda: get("S3_ENDPOINT_URL", None),
    "VLLM_USAGE_STATS_SERVER": lambda: get(
        "VLLM_USAGE_STATS_SERVER", "https://stats.vllm.ai"
    ),
    "VLLM_NO_USAGE_STATS": lambda: get(
        "VLLM_NO_USAGE_STATS", "0"
    )
    == "1",
    "VLLM_DISABLE_FLASHINFER_PREFILL": lambda: get(
        "VLLM_DISABLE_FLASHINFER_PREFILL", "0"
    )
    == "1",
    "VLLM_DO_NOT_TRACK": lambda: (
        get("VLLM_DO_NOT_TRACK", None)
        or get("DO_NOT_TRACK", None)
        or "0"
    )
    == "1",
    "VLLM_USAGE_SOURCE": lambda: get(
        "VLLM_USAGE_SOURCE", "production"
    ),
    "VLLM_CONFIGURE_LOGGING": lambda: int(
        getenv("VLLM_CONFIGURE_LOGGING", "1")
    ),
    "VLLM_LOGGING_CONFIG_PATH": lambda: getenv(
        "VLLM_LOGGING_CONFIG_PATH"
    ),
    "VLLM_LOGGING_LEVEL": lambda: upper(),
    "VLLM_LOGGING_STREAM": lambda: getenv(
        "VLLM_LOGGING_STREAM", "ext://sys.stdout"
    ),
    "VLLM_LOGGING_PREFIX": lambda: getenv(
        "VLLM_LOGGING_PREFIX", ""
    ),
    "VLLM_LOGITS_PROCESSOR_THREADS": lambda: int(
        getenv("VLLM_LOGITS_PROCESSOR_THREADS", "0")
    )
    if "VLLM_LOGITS_PROCESSOR_THREADS" in environ
    else None,
    "VLLM_LOG_STATS_INTERVAL": lambda: val
    if (
        val := (
            float(getenv("VLLM_LOG_STATS_INTERVAL", "10."))
        )
    )
    > 0.0
    else 10.0,
    "VLLM_TRACE_FUNCTION": lambda: int(
        getenv("VLLM_TRACE_FUNCTION", "0")
    ),
    "VLLM_ATTENTION_BACKEND": env_with_choices(
        "VLLM_ATTENTION_BACKEND", None, lambda: list(keys())
    ),
    "VLLM_USE_FLASHINFER_SAMPLER": lambda: bool(
        int(environ["VLLM_USE_FLASHINFER_SAMPLER"])
    )
    if "VLLM_USE_FLASHINFER_SAMPLER" in environ
    else None,
    "VLLM_PP_LAYER_PARTITION": lambda: getenv(
        "VLLM_PP_LAYER_PARTITION", None
    ),
    "VLLM_CPU_KVCACHE_SPACE": lambda: int(
        getenv("VLLM_CPU_KVCACHE_SPACE", "0")
    )
    if "VLLM_CPU_KVCACHE_SPACE" in environ
    else None,
    "VLLM_CPU_OMP_THREADS_BIND": lambda: getenv(
        "VLLM_CPU_OMP_THREADS_BIND", "auto"
    ),
    "VLLM_CPU_NUM_OF_RESERVED_CPU": lambda: int(
        getenv("VLLM_CPU_NUM_OF_RESERVED_CPU", "0")
    )
    if "VLLM_CPU_NUM_OF_RESERVED_CPU" in environ
    else None,
    "VLLM_CPU_MOE_PREPACK": lambda: bool(
        int(getenv("VLLM_CPU_MOE_PREPACK", "1"))
    ),
    "VLLM_CPU_SGL_KERNEL": lambda: bool(
        int(getenv("VLLM_CPU_SGL_KERNEL", "0"))
    ),
    "VLLM_USE_RAY_SPMD_WORKER": lambda: bool(
        int(getenv("VLLM_USE_RAY_SPMD_WORKER", "0"))
    ),
    "VLLM_USE_RAY_COMPILED_DAG": lambda: bool(
        int(getenv("VLLM_USE_RAY_COMPILED_DAG", "0"))
    ),
    "VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE": env_with_choices(
        "VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE",
        "auto",
        ["auto", "nccl", "shm"],
    ),
    "VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM": lambda: bool(
        int(
            getenv(
                "VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM",
                "0",
            )
        )
    ),
    "VLLM_USE_RAY_WRAPPED_PP_COMM": lambda: bool(
        int(getenv("VLLM_USE_RAY_WRAPPED_PP_COMM", "1"))
    ),
    "VLLM_WORKER_MULTIPROC_METHOD": env_with_choices(
        "VLLM_WORKER_MULTIPROC_METHOD",
        "fork",
        ["spawn", "fork"],
    ),
    "VLLM_ASSETS_CACHE": lambda: expanduser(
        getenv(
            "VLLM_ASSETS_CACHE",
            join(
                get_default_cache_root(), "vllm", "assets"
            ),
        )
    ),
    "VLLM_ASSETS_CACHE_MODEL_CLEAN": lambda: bool(
        int(getenv("VLLM_ASSETS_CACHE_MODEL_CLEAN", "0"))
    ),
    "VLLM_IMAGE_FETCH_TIMEOUT": lambda: int(
        getenv("VLLM_IMAGE_FETCH_TIMEOUT", "5")
    ),
    "VLLM_VIDEO_FETCH_TIMEOUT": lambda: int(
        getenv("VLLM_VIDEO_FETCH_TIMEOUT", "30")
    ),
    "VLLM_AUDIO_FETCH_TIMEOUT": lambda: int(
        getenv("VLLM_AUDIO_FETCH_TIMEOUT", "10")
    ),
    "VLLM_MEDIA_URL_ALLOW_REDIRECTS": lambda: bool(
        int(getenv("VLLM_MEDIA_URL_ALLOW_REDIRECTS", "1"))
    ),
    "VLLM_MEDIA_LOADING_THREAD_COUNT": lambda: int(
        getenv("VLLM_MEDIA_LOADING_THREAD_COUNT", "8")
    ),
    "VLLM_MAX_AUDIO_CLIP_FILESIZE_MB": lambda: int(
        getenv("VLLM_MAX_AUDIO_CLIP_FILESIZE_MB", "25")
    ),
    "VLLM_VIDEO_LOADER_BACKEND": lambda: getenv(
        "VLLM_VIDEO_LOADER_BACKEND", "opencv"
    ),
    "VLLM_MM_INPUT_CACHE_GIB": lambda: int(
        getenv("VLLM_MM_INPUT_CACHE_GIB", "4")
    ),
    "VLLM_XLA_CACHE_PATH": lambda: expanduser(
        getenv(
            "VLLM_XLA_CACHE_PATH",
            join(
                get_default_cache_root(),
                "vllm",
                "xla_cache",
            ),
        )
    ),
    "VLLM_XLA_CHECK_RECOMPILATION": lambda: bool(
        int(getenv("VLLM_XLA_CHECK_RECOMPILATION", "0"))
    ),
    "VLLM_XLA_USE_SPMD": lambda: bool(
        int(getenv("VLLM_XLA_USE_SPMD", "0"))
    ),
    "VLLM_FUSED_MOE_CHUNK_SIZE": lambda: int(
        getenv("VLLM_FUSED_MOE_CHUNK_SIZE", "32768")
    ),
    "VLLM_ENABLE_FUSED_MOE_ACTIVATION_CHUNKING": lambda: bool(
        int(
            getenv(
                "VLLM_ENABLE_FUSED_MOE_ACTIVATION_CHUNKING",
                "1",
            )
        )
    ),
    "VLLM_KEEP_ALIVE_ON_ENGINE_DEATH": lambda: bool(
        getenv("VLLM_KEEP_ALIVE_ON_ENGINE_DEATH", 0)
    ),
    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": lambda: lower()
    in ("1", "true"),
    "VLLM_TEST_FORCE_FP8_MARLIN": lambda: lower()
    in ("1", "true"),
    "VLLM_TEST_FORCE_LOAD_FORMAT": lambda: getenv(
        "VLLM_TEST_FORCE_LOAD_FORMAT", "dummy"
    ),
    "VLLM_RPC_TIMEOUT": lambda: int(
        getenv("VLLM_RPC_TIMEOUT", "10000")
    ),
    "VLLM_HTTP_TIMEOUT_KEEP_ALIVE": lambda: int(
        get("VLLM_HTTP_TIMEOUT_KEEP_ALIVE", "5")
    ),
    "VLLM_PLUGINS": lambda: None
    if "VLLM_PLUGINS" not in environ
    else split(","),
    "VLLM_LORA_RESOLVER_CACHE_DIR": lambda: getenv(
        "VLLM_LORA_RESOLVER_CACHE_DIR", None
    ),
    "VLLM_TORCH_PROFILER_DIR": lambda: None
    if getenv("VLLM_TORCH_PROFILER_DIR", None) is None
    else abspath(
        expanduser(getenv("VLLM_TORCH_PROFILER_DIR", "."))
    ),
    "VLLM_TORCH_PROFILER_RECORD_SHAPES": lambda: bool(
        getenv("VLLM_TORCH_PROFILER_RECORD_SHAPES", "0")
        != "0"
    ),
    "VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY": lambda: bool(
        getenv(
            "VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY", "0"
        )
        != "0"
    ),
    "VLLM_TORCH_PROFILER_WITH_STACK": lambda: bool(
        getenv("VLLM_TORCH_PROFILER_WITH_STACK", "1") != "0"
    ),
    "VLLM_TORCH_PROFILER_WITH_FLOPS": lambda: bool(
        getenv("VLLM_TORCH_PROFILER_WITH_FLOPS", "0") != "0"
    ),
    "VLLM_USE_TRITON_AWQ": lambda: bool(
        int(getenv("VLLM_USE_TRITON_AWQ", "0"))
    ),
    "VLLM_ALLOW_RUNTIME_LORA_UPDATING": lambda: lower()
    in ("1", "true"),
    "VLLM_SKIP_P2P_CHECK": lambda: getenv(
        "VLLM_SKIP_P2P_CHECK", "1"
    )
    == "1",
    "VLLM_DISABLED_KERNELS": lambda: []
    if "VLLM_DISABLED_KERNELS" not in environ
    else split(","),
    "VLLM_DISABLE_PYNCCL": lambda: lower() in ("true", "1"),
    "VLLM_USE_V1": lambda: bool(
        int(getenv("VLLM_USE_V1", "1"))
    ),
    "VLLM_ROCM_USE_AITER": lambda: lower() in ("true", "1"),
    "VLLM_ROCM_USE_AITER_PAGED_ATTN": lambda: lower()
    in ("true", "1"),
    "VLLM_ROCM_USE_AITER_LINEAR": lambda: lower()
    in ("true", "1"),
    "VLLM_ROCM_USE_AITER_MOE": lambda: lower()
    in ("true", "1"),
    "VLLM_ROCM_USE_AITER_RMSNORM": lambda: lower()
    in ("true", "1"),
    "VLLM_ROCM_USE_AITER_MLA": lambda: lower()
    in ("true", "1"),
    "VLLM_ROCM_USE_AITER_MHA": lambda: lower()
    in ("true", "1"),
    "VLLM_ROCM_USE_AITER_FP4_ASM_GEMM": lambda: lower()
    in ("true", "1"),
    "VLLM_ROCM_USE_TRITON_ROPE": lambda: lower()
    in ("true", "1"),
    "VLLM_ROCM_USE_AITER_FP8BMM": lambda: lower()
    in ("true", "1"),
    "VLLM_ROCM_USE_AITER_UNIFIED_ATTENTION": lambda: lower()
    in ("true", "1"),
    "VLLM_ROCM_USE_SKINNY_GEMM": lambda: lower()
    in ("true", "1"),
    "VLLM_ROCM_FP8_PADDING": lambda: bool(
        int(getenv("VLLM_ROCM_FP8_PADDING", "1"))
    ),
    "VLLM_ROCM_MOE_PADDING": lambda: bool(
        int(getenv("VLLM_ROCM_MOE_PADDING", "1"))
    ),
    "VLLM_ROCM_CUSTOM_PAGED_ATTN": lambda: lower()
    in ("true", "1"),
    "VLLM_ROCM_QUICK_REDUCE_QUANTIZATION": env_with_choices(
        "VLLM_ROCM_QUICK_REDUCE_QUANTIZATION",
        "NONE",
        ["FP", "INT8", "INT6", "INT4", "NONE"],
    ),
    "VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16": lambda: lower()
    in ("true", "1"),
    "VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB": lambda: maybe_convert_int(
        get(
            "VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB", None
        )
    ),
    "Q_SCALE_CONSTANT": lambda: int(
        getenv("Q_SCALE_CONSTANT", "200")
    ),
    "K_SCALE_CONSTANT": lambda: int(
        getenv("K_SCALE_CONSTANT", "200")
    ),
    "V_SCALE_CONSTANT": lambda: int(
        getenv("V_SCALE_CONSTANT", "100")
    ),
    "VLLM_ENABLE_V1_MULTIPROCESSING": lambda: bool(
        int(getenv("VLLM_ENABLE_V1_MULTIPROCESSING", "1"))
    ),
    "VLLM_LOG_BATCHSIZE_INTERVAL": lambda: float(
        getenv("VLLM_LOG_BATCHSIZE_INTERVAL", "-1")
    ),
    "VLLM_DISABLE_COMPILE_CACHE": lambda: bool(
        int(getenv("VLLM_DISABLE_COMPILE_CACHE", "0"))
    ),
    "VLLM_SERVER_DEV_MODE": lambda: bool(
        int(getenv("VLLM_SERVER_DEV_MODE", "0"))
    ),
    "VLLM_V1_OUTPUT_PROC_CHUNK_SIZE": lambda: int(
        getenv("VLLM_V1_OUTPUT_PROC_CHUNK_SIZE", "128")
    ),
    "VLLM_MLA_DISABLE": lambda: bool(
        int(getenv("VLLM_MLA_DISABLE", "0"))
    ),
    "VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH": lambda: int(
        getenv(
            "VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH",
            "32",
        )
    ),
    "VLLM_RAY_PER_WORKER_GPUS": lambda: float(
        getenv("VLLM_RAY_PER_WORKER_GPUS", "1.0")
    ),
    "VLLM_RAY_BUNDLE_INDICES": lambda: getenv(
        "VLLM_RAY_BUNDLE_INDICES", ""
    ),
    "VLLM_CUDART_SO_PATH": lambda: getenv(
        "VLLM_CUDART_SO_PATH", None
    ),
    "VLLM_DP_RANK": lambda: int(
        getenv("VLLM_DP_RANK", "0")
    ),
    "VLLM_DP_RANK_LOCAL": lambda: int(
        getenv("VLLM_DP_RANK_LOCAL", VLLM_DP_RANK)
    ),
    "VLLM_DP_SIZE": lambda: int(
        getenv("VLLM_DP_SIZE", "1")
    ),
    "VLLM_DP_MASTER_IP": lambda: getenv(
        "VLLM_DP_MASTER_IP", "127.0.0.1"
    ),
    "VLLM_DP_MASTER_PORT": lambda: int(
        getenv("VLLM_DP_MASTER_PORT", "0")
    ),
    "VLLM_MOE_DP_CHUNK_SIZE": lambda: int(
        getenv("VLLM_MOE_DP_CHUNK_SIZE", "256")
    ),
    "VLLM_RANDOMIZE_DP_DUMMY_INPUTS": lambda: get(
        "VLLM_RANDOMIZE_DP_DUMMY_INPUTS", "0"
    )
    == "1",
    "VLLM_RAY_DP_PACK_STRATEGY": lambda: getenv(
        "VLLM_RAY_DP_PACK_STRATEGY", "strict"
    ),
    "VLLM_CI_USE_S3": lambda: get("VLLM_CI_USE_S3", "0")
    == "1",
    "VLLM_MODEL_REDIRECT_PATH": lambda: get(
        "VLLM_MODEL_REDIRECT_PATH", None
    ),
    "VLLM_MARLIN_USE_ATOMIC_ADD": lambda: get(
        "VLLM_MARLIN_USE_ATOMIC_ADD", "0"
    )
    == "1",
    "VLLM_MXFP4_USE_MARLIN": lambda: maybe_convert_bool(
        get("VLLM_MXFP4_USE_MARLIN", None)
    ),
    "VLLM_V0_USE_OUTLINES_CACHE": lambda: get(
        "VLLM_V0_USE_OUTLINES_CACHE", "0"
    )
    == "1",
    "VLLM_V1_USE_OUTLINES_CACHE": lambda: get(
        "VLLM_V1_USE_OUTLINES_CACHE", "0"
    )
    == "1",
    "VLLM_TPU_BUCKET_PADDING_GAP": lambda: int(
        environ["VLLM_TPU_BUCKET_PADDING_GAP"]
    )
    if "VLLM_TPU_BUCKET_PADDING_GAP" in environ
    else 0,
    "VLLM_TPU_MOST_MODEL_LEN": lambda: maybe_convert_int(
        get("VLLM_TPU_MOST_MODEL_LEN", None)
    ),
    "VLLM_TPU_USING_PATHWAYS": lambda: bool(
        "proxy" in lower()
    ),
    "VLLM_USE_DEEP_GEMM": lambda: bool(
        int(getenv("VLLM_USE_DEEP_GEMM", "1"))
    ),
    "VLLM_USE_DEEP_GEMM_E8M0": lambda: bool(
        int(getenv("VLLM_USE_DEEP_GEMM_E8M0", "1"))
    ),
    "VLLM_DEEP_GEMM_WARMUP": env_with_choices(
        "VLLM_DEEP_GEMM_WARMUP",
        "relax",
        ["skip", "full", "relax"],
    ),
    "VLLM_USE_FUSED_MOE_GROUPED_TOPK": lambda: bool(
        int(getenv("VLLM_USE_FUSED_MOE_GROUPED_TOPK", "1"))
    ),
    "VLLM_USE_FLASHINFER_MOE_FP16": lambda: bool(
        int(getenv("VLLM_USE_FLASHINFER_MOE_FP16", "0"))
    ),
    "VLLM_USE_FLASHINFER_MOE_FP8": lambda: bool(
        int(getenv("VLLM_USE_FLASHINFER_MOE_FP8", "0"))
    ),
    "VLLM_USE_FLASHINFER_MOE_FP4": lambda: bool(
        int(getenv("VLLM_USE_FLASHINFER_MOE_FP4", "0"))
    ),
    "VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8": lambda: bool(
        int(
            getenv(
                "VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8", "0"
            )
        )
    ),
    "VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS": lambda: bool(
        int(
            getenv(
                "VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS",
                "0",
            )
        )
    ),
    "VLLM_USE_FLASHINFER_MOE_MXFP4_BF16": lambda: bool(
        int(
            getenv(
                "VLLM_USE_FLASHINFER_MOE_MXFP4_BF16", "0"
            )
        )
    ),
    "VLLM_XGRAMMAR_CACHE_MB": lambda: int(
        getenv("VLLM_XGRAMMAR_CACHE_MB", "512")
    ),
    "VLLM_MSGPACK_ZERO_COPY_THRESHOLD": lambda: int(
        getenv("VLLM_MSGPACK_ZERO_COPY_THRESHOLD", "256")
    ),
    "VLLM_ALLOW_INSECURE_SERIALIZATION": lambda: bool(
        int(
            getenv("VLLM_ALLOW_INSECURE_SERIALIZATION", "0")
        )
    ),
    "VLLM_NIXL_SIDE_CHANNEL_HOST": lambda: getenv(
        "VLLM_NIXL_SIDE_CHANNEL_HOST", "localhost"
    ),
    "VLLM_NIXL_SIDE_CHANNEL_PORT": lambda: int(
        getenv("VLLM_NIXL_SIDE_CHANNEL_PORT", "5600")
    ),
    "VLLM_ALL2ALL_BACKEND": env_with_choices(
        "VLLM_ALL2ALL_BACKEND",
        "allgather_reducescatter",
        [
            "naive",
            "pplx",
            "deepep_high_throughput",
            "deepep_low_latency",
            "allgather_reducescatter",
            "flashinfer_all2allv",
        ],
    ),
    "VLLM_FLASHINFER_MOE_BACKEND": env_with_choices(
        "VLLM_FLASHINFER_MOE_BACKEND",
        "throughput",
        ["throughput", "latency"],
    ),
    "VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE": lambda: int(
        getenv(
            "VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE", "163840"
        )
    ),
    "VLLM_FLASHINFER_ALLREDUCE_FUSION_THRESHOLDS_MB": lambda: loads(
        getenv(
            "VLLM_FLASHINFER_ALLREDUCE_FUSION_THRESHOLDS_MB",
            "{}",
        )
    ),
    "VLLM_MOE_ROUTING_SIMULATION_STRATEGY": lambda: lower(),
    "VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS": lambda: int(
        getenv("VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS", "1")
    ),
    "VLLM_SLEEP_WHEN_IDLE": lambda: bool(
        int(getenv("VLLM_SLEEP_WHEN_IDLE", "0"))
    ),
    "VLLM_MQ_MAX_CHUNK_BYTES_MB": lambda: int(
        getenv("VLLM_MQ_MAX_CHUNK_BYTES_MB", "16")
    ),
    "VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS": lambda: int(
        getenv("VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS", "300")
    ),
    "VLLM_KV_CACHE_LAYOUT": env_with_choices(
        "VLLM_KV_CACHE_LAYOUT", None, ["NHD", "HND"]
    ),
    "VLLM_COMPUTE_NANS_IN_LOGITS": lambda: bool(
        int(getenv("VLLM_COMPUTE_NANS_IN_LOGITS", "0"))
    ),
    "VLLM_USE_NVFP4_CT_EMULATIONS": lambda: bool(
        int(getenv("VLLM_USE_NVFP4_CT_EMULATIONS", "0"))
    ),
    "VLLM_NIXL_ABORT_REQUEST_TIMEOUT": lambda: int(
        getenv("VLLM_NIXL_ABORT_REQUEST_TIMEOUT", "480")
    ),
    "VLLM_USE_CUDNN_PREFILL": lambda: bool(
        int(getenv("VLLM_USE_CUDNN_PREFILL", "0"))
    ),
    "VLLM_USE_TRTLLM_ATTENTION": lambda: None
    if "VLLM_USE_TRTLLM_ATTENTION" not in environ
    else lower() in ("1", "true"),
    "VLLM_FLASHINFER_DISABLE_Q_QUANTIZATION": lambda: bool(
        int(
            getenv(
                "VLLM_FLASHINFER_DISABLE_Q_QUANTIZATION",
                "0",
            )
        )
    ),
    "VLLM_HAS_FLASHINFER_CUBIN": lambda: getenv(
        "VLLM_HAS_FLASHINFER_CUBIN", False
    ),
    "VLLM_USE_TRTLLM_FP4_GEMM": lambda: bool(
        int(getenv("VLLM_USE_TRTLLM_FP4_GEMM", "0"))
    ),
    "VLLM_ENABLE_CUDAGRAPH_GC": lambda: bool(
        int(getenv("VLLM_ENABLE_CUDAGRAPH_GC", "0"))
    ),
    "VLLM_LOOPBACK_IP": lambda: getenv(
        "VLLM_LOOPBACK_IP", ""
    ),
    "VLLM_PROCESS_NAME_PREFIX": lambda: getenv(
        "VLLM_PROCESS_NAME_PREFIX", "VLLM"
    ),
    "VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE": lambda: bool(
        int(
            getenv(
                "VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE",
                "0",
            )
        )
    ),
    "VLLM_ENABLE_RESPONSES_API_STORE": lambda: bool(
        int(getenv("VLLM_ENABLE_RESPONSES_API_STORE", "0"))
    ),
    "VLLM_ROCM_FP8_MFMA_PAGE_ATTN": lambda: bool(
        int(getenv("VLLM_ROCM_FP8_MFMA_PAGE_ATTN", "0"))
    ),
    "VLLM_ALLREDUCE_USE_SYMM_MEM": lambda: bool(
        int(getenv("VLLM_ALLREDUCE_USE_SYMM_MEM", "1"))
    ),
    "VLLM_TUNED_CONFIG_FOLDER": lambda: getenv(
        "VLLM_TUNED_CONFIG_FOLDER", None
    ),
    "VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS": lambda: bool(
        int(
            getenv(
                "VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS",
                "0",
            )
        )
    ),
    "VLLM_CUSTOM_SCOPES_FOR_PROFILING": lambda: bool(
        int(getenv("VLLM_CUSTOM_SCOPES_FOR_PROFILING", "0"))
    ),
    "VLLM_NVTX_SCOPES_FOR_PROFILING": lambda: bool(
        int(getenv("VLLM_NVTX_SCOPES_FOR_PROFILING", "0"))
    ),
    "VLLM_KV_EVENTS_USE_INT_BLOCK_HASHES": lambda: bool(
        int(
            getenv(
                "VLLM_KV_EVENTS_USE_INT_BLOCK_HASHES", "1"
            )
        )
    ),
    "VLLM_OBJECT_STORAGE_SHM_BUFFER_NAME": lambda: getenv(
        "VLLM_OBJECT_STORAGE_SHM_BUFFER_NAME",
        "VLLM_OBJECT_STORAGE_SHM_BUFFER",
    ),
    "VLLM_DEEPEP_BUFFER_SIZE_MB": lambda: int(
        getenv("VLLM_DEEPEP_BUFFER_SIZE_MB", "1024")
    ),
    "VLLM_DBO_COMM_SMS": lambda: int(
        getenv("VLLM_DBO_COMM_SMS", "20")
    ),
    "GPT_OSS_SYSTEM_TOOL_MCP_LABELS": env_list_with_choices(
        "GPT_OSS_SYSTEM_TOOL_MCP_LABELS",
        [],
        [
            "container",
            "code_interpreter",
            "web_search_preview",
        ],
    ),
    "VLLM_ENABLE_INDUCTOR_MAX_AUTOTUNE": lambda: bool(
        int(
            getenv("VLLM_ENABLE_INDUCTOR_MAX_AUTOTUNE", "1")
        )
    ),
    "VLLM_ENABLE_INDUCTOR_COORDINATE_DESCENT_TUNING": lambda: bool(
        int(
            getenv(
                "VLLM_ENABLE_INDUCTOR_COORDINATE_DESCENT_TUNING",
                "1",
            )
        )
    ),
    "VLLM_USE_NCCL_SYMM_MEM": lambda: bool(
        int(getenv("VLLM_USE_NCCL_SYMM_MEM", "0"))
    ),
    "VLLM_NCCL_INCLUDE_PATH": lambda: get(
        "VLLM_NCCL_INCLUDE_PATH", None
    ),
    "VLLM_USE_FBGEMM": lambda: bool(
        int(getenv("VLLM_USE_FBGEMM", "0"))
    ),
    "VLLM_GC_DEBUG": lambda: getenv("VLLM_GC_DEBUG", ""),
}