环境变量¶
vLLM 使用以下环境变量来配置系统:
Warning
请注意,VLLM_PORT 和 VLLM_HOST_IP 设置的是 vLLM 内部使用 的端口和 IP 地址,而非 API 服务器的端口和 IP 地址。如果您使用 --host $VLLM_HOST_IP 和 --port $VLLM_PORT 来启动 API 服务器,将无法正常工作。
vLLM 使用的所有环境变量均以 VLLM_ 作为前缀。Kubernetes 用户需特别注意:请勿将服务命名为 vllm,否则 Kubernetes 设置的环境变量可能会与 vLLM 的环境变量发生冲突,因为 Kubernetes 会为每个服务设置以该服务名称的大写形式作为前缀的环境变量。
logger = logging.getLogger(__name__)
def _deprecated_triton_attn_use_td() -> None:
"""Warn that VLLM_TRITON_ATTN_USE_TD was renamed to VLLM_TRITON_USE_TD.
The old name is ignored; VLLM_TRITON_USE_TD is the supported variable.
"""
if "VLLM_TRITON_ATTN_USE_TD" in os.environ:
logger.warning(
"VLLM_TRITON_ATTN_USE_TD is deprecated and will be removed in "
"v0.25. Use VLLM_TRITON_USE_TD instead."
)
return None
def _resolve_rust_cli_path() -> str | None:
"""Resolve the vllm-rs binary path.
Returns None unless VLLM_USE_RUST_FRONTEND or VLLM_USE_RUST_BENCH is enabled.
When enabled, resolves VLLM_RUST_FRONTEND_PATH ("auto" by default)
to the actual binary path.
"""
use_rust = bool(int(os.environ.get("VLLM_USE_RUST_FRONTEND", "0"))) or bool(
int(os.environ.get("VLLM_USE_RUST_BENCH", "0"))
)
raw = os.environ.get("VLLM_RUST_FRONTEND_PATH", "auto")
if not use_rust:
if os.environ.get("VLLM_RUST_FRONTEND_PATH") is not None:
logger.warning(
"VLLM_RUST_FRONTEND_PATH is set without enabling "
"VLLM_USE_RUST_FRONTEND or VLLM_USE_RUST_BENCH. "
"Set one of them to 1 to use the vllm-rs binary."
)
return None
if raw.lower() in ("auto", "1", "true"):
pkg_dir = os.path.dirname(os.path.abspath(__file__))
candidate = os.path.join(pkg_dir, "vllm-rs")
if os.path.isfile(candidate) and os.access(candidate, os.X_OK):
return candidate
raise FileNotFoundError(
"VLLM_RUST_FRONTEND_PATH=auto but the vllm-rs binary was "
f"not found at {candidate}. "
"Build with setuptools-rust or set the path explicitly."
)
return raw
environment_variables: dict[str, Callable[[], Any]] = {
# ================== Installation Time Env Vars ==================
# Target device of vLLM, supporting [cuda (by default),
# rocm, cpu]
"VLLM_TARGET_DEVICE": lambda: os.getenv("VLLM_TARGET_DEVICE", "cuda").lower(),
# Main CUDA version of vLLM. This follows PyTorch but can be overridden.
"VLLM_MAIN_CUDA_VERSION": lambda: (
os.getenv("VLLM_MAIN_CUDA_VERSION", "").lower() or "13.0"
),
# Controls PyTorch float32 matmul precision mode within vLLM workers.
# Valid options mirror torch.set_float32_matmul_precision
"VLLM_FLOAT32_MATMUL_PRECISION": env_with_choices(
"VLLM_FLOAT32_MATMUL_PRECISION",
"highest",
["highest", "high", "medium"],
case_sensitive=False,
),
# Enable batch-invariant mode: deterministic results regardless of
# batch composition. Requires NVIDIA GPU with compute capability >= 9.0.
"VLLM_BATCH_INVARIANT": lambda: bool(int(os.getenv("VLLM_BATCH_INVARIANT", "0"))),
# Use tensor descriptors for Q/K/V loads and output stores in the
# Triton unified-attention kernel. Enables HW 2D block reads on
# Intel XPU; the non-TD branch is dead-code-eliminated at Triton
# compile time so other platforms see no overhead. Tri-state override:
# unset (default) lets the `triton_attn` backend auto-select per
# platform (currently auto-enabled on XPU only); ``1`` forces TD on;
# ``0`` forces TD off. Useful for A/B benchmarking the TD path.
"VLLM_TRITON_USE_TD": lambda: {"1": True, "0": False}.get(
os.getenv("VLLM_TRITON_USE_TD", "").strip()
),
# If set, enable PyTorch's GPU<->CPU synchronization debug mode around
# the worker's `execute_model` and `sample_tokens` calls. Valid values
# are "warn" (print a warning on each sync) or "error" (raise on sync).
# Unset disables the check. See `torch.cuda.set_sync_debug_mode`.
"VLLM_GPU_SYNC_CHECK": env_with_choices(
"VLLM_GPU_SYNC_CHECK", None, ["warn", "error"]
),
# Deprecated: renamed to VLLM_TRITON_USE_TD. Kept registered so it does
# not trip the unknown-env-var check; warns on use and is otherwise
# ignored.
"VLLM_TRITON_ATTN_USE_TD": lambda: _deprecated_triton_attn_use_td(),
# Maximum number of compilation jobs to run in parallel.
# By default this is the number of CPUs
"MAX_JOBS": lambda: os.getenv("MAX_JOBS", None),
# Number of threads to use for nvcc
# By default this is 1.
# If set, `MAX_JOBS` will be reduced to avoid oversubscribing the CPU.
"NVCC_THREADS": lambda: os.getenv("NVCC_THREADS", None),
# If set, vllm will use precompiled native binaries (*.so and vllm-rs).
"VLLM_USE_PRECOMPILED": lambda: (
os.environ.get("VLLM_USE_PRECOMPILED", "").strip().lower() in ("1", "true")
or bool(os.environ.get("VLLM_PRECOMPILED_WHEEL_LOCATION"))
),
# If set, vllm will use the precompiled Rust frontend binary (vllm-rs).
"VLLM_USE_PRECOMPILED_RUST": lambda: (
os.environ.get("VLLM_USE_PRECOMPILED_RUST", "").strip().lower() in ("1", "true")
),
# If set, skip adding +precompiled suffix to version string
"VLLM_SKIP_PRECOMPILED_VERSION_SUFFIX": lambda: bool(
int(os.environ.get("VLLM_SKIP_PRECOMPILED_VERSION_SUFFIX", "0"))
),
# Used to mark that setup.py is running in a Docker build context,
# in order to force the use of precompiled binaries.
"VLLM_DOCKER_BUILD_CONTEXT": lambda: (
os.environ.get("VLLM_DOCKER_BUILD_CONTEXT", "").strip().lower() in ("1", "true")
),
# Build provenance metadata embedded in official vllm-openai images.
# Set via Docker ENV at image build time; informational only.
"VLLM_BUILD_COMMIT": lambda: os.environ.get("VLLM_BUILD_COMMIT", "unknown"),
"VLLM_BUILD_PIPELINE": lambda: os.environ.get("VLLM_BUILD_PIPELINE", "local"),
"VLLM_BUILD_URL": lambda: os.environ.get("VLLM_BUILD_URL", ""),
"VLLM_IMAGE_TAG": lambda: os.environ.get("VLLM_IMAGE_TAG", ""),
# CMake build type
# If not set, defaults to "Debug" or "RelWithDebInfo"
# Available options: "Debug", "Release", "RelWithDebInfo"
"CMAKE_BUILD_TYPE": env_with_choices(
"CMAKE_BUILD_TYPE", None, ["Debug", "Release", "RelWithDebInfo"]
),
# If set, vllm will print verbose logs during installation
"VERBOSE": lambda: bool(int(os.getenv("VERBOSE", "0"))),
# Root directory for vLLM configuration files
# Defaults to `~/.config/vllm` unless `XDG_CONFIG_HOME` is set
# Note that this not only affects how vllm finds its configuration files
# during runtime, but also affects how vllm installs its configuration
# files during **installation**.
"VLLM_CONFIG_ROOT": lambda: os.path.expanduser(
os.getenv(
"VLLM_CONFIG_ROOT",
os.path.join(get_default_config_root(), "vllm"),
)
),
# ================== Runtime Env Vars ==================
# Root directory for vLLM cache files
# Defaults to `~/.cache/vllm` unless `XDG_CACHE_HOME` is set
"VLLM_CACHE_ROOT": lambda: os.path.expanduser(
os.getenv(
"VLLM_CACHE_ROOT",
os.path.join(get_default_cache_root(), "vllm"),
)
),
# used in distributed environment to determine the ip address
# of the current node, when the node has multiple network interfaces.
# If you are using multi-node inference, you should set this differently
# on each node.
"VLLM_HOST_IP": lambda: os.getenv("VLLM_HOST_IP", ""),
# used in distributed environment to manually set the communication port
# Note: if VLLM_PORT is set, and some code asks for multiple ports, the
# VLLM_PORT will be used as the first port, and the rest will be generated
# by incrementing the VLLM_PORT value.
"VLLM_PORT": get_vllm_port,
# path used for ipc when the frontend api server is running in
# multi-processing mode to communicate with the backend engine process.
"VLLM_RPC_BASE_PATH": lambda: os.getenv(
"VLLM_RPC_BASE_PATH", tempfile.gettempdir()
),
# If true, will load models from ModelScope instead of Hugging Face Hub.
# note that the value is true or false, not numbers
"VLLM_USE_MODELSCOPE": lambda: (
os.environ.get("VLLM_USE_MODELSCOPE", "False").lower() == "true"
),
# If true, replace the Rust BPE backend that powers HF fast tokenizers
# with the `fastokens` (https://github.com/crusoecloud/fastokens) shim.
# Applies to any tokenizer mode that loads an HF fast tokenizer
# (`hf`, `deepseek_v32`, `deepseek_v4`, …). The `fastokens`
# Python package must be installed.
"VLLM_USE_FASTOKENS": lambda: bool(int(os.getenv("VLLM_USE_FASTOKENS", "0"))),
# Interval in seconds to log a warning message when the ring buffer is full
"VLLM_RINGBUFFER_WARNING_INTERVAL": lambda: int(
os.environ.get("VLLM_RINGBUFFER_WARNING_INTERVAL", "60")
),
# path to cudatoolkit home directory, under which should be bin, include,
# and lib directories.
"CUDA_HOME": lambda: os.environ.get("CUDA_HOME", None),
# Path to the NCCL library file. It is needed because nccl>=2.19 brought
# by PyTorch contains a bug: https://github.com/NVIDIA/nccl/issues/1234
"VLLM_NCCL_SO_PATH": lambda: os.environ.get("VLLM_NCCL_SO_PATH", None),
# when `VLLM_NCCL_SO_PATH` is not set, vllm will try to find the nccl
# library file in the locations specified by `LD_LIBRARY_PATH`
"LD_LIBRARY_PATH": lambda: os.environ.get("LD_LIBRARY_PATH", None),
# flag to control the chunk size (in MB) for sleeping memory allocations under ROCm
"VLLM_ROCM_SLEEP_MEM_CHUNK_SIZE": lambda: int(
os.environ.get("VLLM_ROCM_SLEEP_MEM_CHUNK_SIZE", "256")
),
# Feature flag to enable/disable Inductor standalone compile.
# In torch <= 2.7 we ignore this flag; in torch >= 2.9 this is
# enabled by default.
"VLLM_USE_STANDALONE_COMPILE": lambda: (
os.environ.get("VLLM_USE_STANDALONE_COMPILE", "1") == "1"
),
# Inductor's pre-grad passes don't do anything for vLLM.
# The pre-grad passes get run even on cache-hit and negatively impact
# vllm cold compile times by O(1s)
# Can remove this after the following issue gets fixed
# TODO(luka): maybe_inplace requires this
# https://github.com/pytorch/pytorch/issues/174502
"VLLM_ENABLE_PREGRAD_PASSES": lambda: (
os.environ.get("VLLM_ENABLE_PREGRAD_PASSES", "1") == "1"
),
# Experimental: breakable cudagraph does not rely on torch.compile
"VLLM_USE_BREAKABLE_CUDAGRAPH": lambda: (
os.environ.get("VLLM_USE_BREAKABLE_CUDAGRAPH", "0") == "1"
),
# Debug pattern matching inside custom passes.
# Should be set to the fx.Node name (e.g. 'getitem_34' or 'scaled_mm_3').
"VLLM_PATTERN_MATCH_DEBUG": lambda: os.environ.get(
"VLLM_PATTERN_MATCH_DEBUG", None
),
# Dump fx graphs to the given directory.
# It will override CompilationConfig.debug_dump_path if set.
"VLLM_DEBUG_DUMP_PATH": lambda: os.environ.get("VLLM_DEBUG_DUMP_PATH", None),
# Feature flag to enable/disable AOT compilation. This will ensure
# compilation is done in warmup phase and the compilation will be
# reused in subsequent calls.
"VLLM_USE_AOT_COMPILE": use_aot_compile,
# Feature flag to enable/disable bytecode in
# TorchCompileWithNoGuardsWrapper.
"VLLM_USE_BYTECODE_HOOK": lambda: bool(
int(os.environ.get("VLLM_USE_BYTECODE_HOOK", "1"))
),
# Force vllm to always load AOT compiled models from disk. Failure
# to load will result in a hard error when this is enabled.
# Will be ignored when VLLM_USE_AOT_COMPILE is disabled.
"VLLM_FORCE_AOT_LOAD": lambda: os.environ.get("VLLM_FORCE_AOT_LOAD", "0") == "1",
# Enable loading compiled models directly from cached standalone compile artifacts
# without re-splitting graph modules. This reduces overhead during model
# loading by using reconstruct_serializable_fn_from_mega_artifact.
"VLLM_USE_MEGA_AOT_ARTIFACT": use_mega_aot_artifact,
# local rank of the process in the distributed setting, used to determine
# the GPU device id
"LOCAL_RANK": lambda: int(os.environ.get("LOCAL_RANK", "0")),
# used to control the visible devices in the distributed setting
"CUDA_VISIBLE_DEVICES": lambda: os.environ.get("CUDA_VISIBLE_DEVICES", None),
# timeout for each iteration in the engine
"VLLM_ENGINE_ITERATION_TIMEOUT_S": lambda: int(
os.environ.get("VLLM_ENGINE_ITERATION_TIMEOUT_S", "60")
),
# Timeout in seconds for waiting for engine cores to become ready
# during startup. Default is 600 seconds (10 minutes).
"VLLM_ENGINE_READY_TIMEOUT_S": lambda: int(
os.environ.get("VLLM_ENGINE_READY_TIMEOUT_S", "600")
),
# API key for vLLM API server
"VLLM_API_KEY": lambda: os.environ.get("VLLM_API_KEY", None),
# Whether to log responses from API Server for debugging
"VLLM_DEBUG_LOG_API_SERVER_RESPONSE": lambda: (
os.environ.get("VLLM_DEBUG_LOG_API_SERVER_RESPONSE", "False").lower() == "true"
),
# S3 access information, used for tensorizer to load model from S3
"S3_ACCESS_KEY_ID": lambda: os.environ.get("S3_ACCESS_KEY_ID", None),
"S3_SECRET_ACCESS_KEY": lambda: os.environ.get("S3_SECRET_ACCESS_KEY", None),
"S3_ENDPOINT_URL": lambda: os.environ.get("S3_ENDPOINT_URL", None),
# Usage stats collection
"VLLM_USAGE_STATS_SERVER": lambda: os.environ.get(
"VLLM_USAGE_STATS_SERVER", "https://stats.vllm.ai"
),
"VLLM_NO_USAGE_STATS": lambda: os.environ.get("VLLM_NO_USAGE_STATS", "0") == "1",
"VLLM_DO_NOT_TRACK": lambda: (
(
os.environ.get("VLLM_DO_NOT_TRACK", None)
or os.environ.get("DO_NOT_TRACK", None)
or "0"
)
== "1"
),
"VLLM_USAGE_SOURCE": lambda: os.environ.get("VLLM_USAGE_SOURCE", "production"),
# Logging configuration
# If set to 0, vllm will not configure logging
# If set to 1, vllm will configure logging using the default configuration
# or the configuration file specified by VLLM_LOGGING_CONFIG_PATH
"VLLM_CONFIGURE_LOGGING": lambda: bool(
int(os.getenv("VLLM_CONFIGURE_LOGGING", "1"))
),
"VLLM_LOGGING_CONFIG_PATH": lambda: os.getenv("VLLM_LOGGING_CONFIG_PATH"),
# this is used for configuring the default logging level
"VLLM_LOGGING_LEVEL": lambda: os.getenv("VLLM_LOGGING_LEVEL", "INFO").upper(),
# this is used for configuring the default logging stream
"VLLM_LOGGING_STREAM": lambda: os.getenv("VLLM_LOGGING_STREAM", "ext://sys.stdout"),
# if set, VLLM_LOGGING_PREFIX will be prepended to all log messages
"VLLM_LOGGING_PREFIX": lambda: os.getenv("VLLM_LOGGING_PREFIX", ""),
# Controls colored logging output. Options: "auto" (default, colors when terminal),
# "1" (always use colors), "0" (never use colors)
"VLLM_LOGGING_COLOR": lambda: os.getenv("VLLM_LOGGING_COLOR", "auto"),
# Standard unix flag for disabling ANSI color codes
"NO_COLOR": lambda: os.getenv("NO_COLOR", "0") != "0",
# If set, vllm will log stats at this interval in seconds
# If not set, vllm will log stats every 10 seconds.
"VLLM_LOG_STATS_INTERVAL": lambda: (
val
if (val := float(os.getenv("VLLM_LOG_STATS_INTERVAL", "10."))) > 0.0
else 10.0
),
# Trace function calls
# If set to 1, vllm will trace function calls
# Useful for debugging
"VLLM_TRACE_FUNCTION": lambda: int(os.getenv("VLLM_TRACE_FUNCTION", "0")),
# Whether to use the FlashInfer top-k / top-p sampler on CUDA. Enabled
# by default when the hardware supports it — set to 0 to opt out
# explicitly, which forces the PyTorch-native (Triton for bs>=8) path.
"VLLM_USE_FLASHINFER_SAMPLER": lambda: (
bool(int(os.environ["VLLM_USE_FLASHINFER_SAMPLER"]))
if "VLLM_USE_FLASHINFER_SAMPLER" in os.environ
else True
),
# Pipeline stage partition strategy
"VLLM_PP_LAYER_PARTITION": lambda: os.getenv("VLLM_PP_LAYER_PARTITION", None),
# (CPU backend only) CPU key-value cache space.
# default is None and will be set as 4 GB
"VLLM_CPU_KVCACHE_SPACE": lambda: (
int(os.getenv("VLLM_CPU_KVCACHE_SPACE", "0"))
if "VLLM_CPU_KVCACHE_SPACE" in os.environ
else None
),
# (CPU backend only) CPU core ids bound by OpenMP threads, e.g., "0-31",
# "0,1,2", "0-31,33". CPU cores of different ranks are separated by '|'.
"VLLM_CPU_OMP_THREADS_BIND": lambda: os.getenv("VLLM_CPU_OMP_THREADS_BIND", "auto"),
# (CPU backend only) CPU cores not used by OMP threads .
# Those CPU cores will not be used by OMP threads of a rank.
"VLLM_CPU_NUM_OF_RESERVED_CPU": lambda: (
int(os.getenv("VLLM_CPU_NUM_OF_RESERVED_CPU", "0"))
if "VLLM_CPU_NUM_OF_RESERVED_CPU" in os.environ
else None
),
# (CPU backend only) whether to enable attention spilt KV.
"VLLM_CPU_ATTN_SPLIT_KV": lambda: bool(
int(os.getenv("VLLM_CPU_ATTN_SPLIT_KV", "1"))
),
# (Zen CPU backend) eagerly prepack weights into ZenDNN blocked layout
# at model load time. Eliminates per-inference layout conversion overhead.
"VLLM_ZENTORCH_WEIGHT_PREPACK": lambda: bool(
int(os.getenv("VLLM_ZENTORCH_WEIGHT_PREPACK", "1"))
),
# (CPU backend only) whether to use SGLang INT4 W4A8 kernels for AWQ.
"VLLM_CPU_INT4_W4A8": lambda: bool(int(os.getenv("VLLM_CPU_INT4_W4A8", "1"))),
# If the env var is set, Ray Compiled Graph uses the specified
# channel type to communicate between workers belonging to
# different pipeline-parallel stages.
# Available options:
# - "auto": use the default channel type
# - "nccl": use NCCL for communication
# - "shm": use shared memory and gRPC for communication
"VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE": env_with_choices(
"VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE", "auto", ["auto", "nccl", "shm"]
),
# If the env var is set, it enables GPU communication overlap
# (experimental feature) in Ray's Compiled Graph.
"VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM": lambda: bool(
int(os.getenv("VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM", "0"))
),
# If the env var is set, it uses a Ray Communicator wrapping
# vLLM's pipeline parallelism communicator to interact with Ray's
# Compiled Graph. Otherwise, it uses Ray's NCCL communicator.
"VLLM_USE_RAY_WRAPPED_PP_COMM": lambda: bool(
int(os.getenv("VLLM_USE_RAY_WRAPPED_PP_COMM", "1"))
),
# When True and distributed_executor_backend="ray", use RayExecutorV2
# (MQ-based) instead of RayDistributedExecutor (compiled-graph backend).
"VLLM_USE_RAY_V2_EXECUTOR_BACKEND": lambda: bool(
int(os.getenv("VLLM_USE_RAY_V2_EXECUTOR_BACKEND", "1"))
),
# When True, GroupCoordinator constructs its CPU/device subgroups via
# ``torch.distributed.split_group(backend=...)``
# and ``init_distributed_environment`` initializes the default PG with
# mixed ``cpu:gloo,cuda:nccl`` backend + eager ``device_id`` binding.
"VLLM_DISTRIBUTED_USE_SPLIT_GROUP": lambda: bool(
int(os.getenv("VLLM_DISTRIBUTED_USE_SPLIT_GROUP", "0"))
),
# Use dedicated multiprocess context for workers.
# Both spawn and fork work
"VLLM_WORKER_MULTIPROC_METHOD": env_with_choices(
"VLLM_WORKER_MULTIPROC_METHOD", "fork", ["spawn", "fork"]
),
# Path to the cache for storing downloaded assets
"VLLM_ASSETS_CACHE": lambda: os.path.expanduser(
os.getenv(
"VLLM_ASSETS_CACHE",
os.path.join(get_default_cache_root(), "vllm", "assets"),
)
),
# If the env var is set, we will clean model file in
# this path $VLLM_ASSETS_CACHE/model_streamer/$model_name
"VLLM_ASSETS_CACHE_MODEL_CLEAN": lambda: bool(
int(os.getenv("VLLM_ASSETS_CACHE_MODEL_CLEAN", "0"))
),
# Timeout for fetching images when serving multimodal models
# Default is 5 seconds
"VLLM_IMAGE_FETCH_TIMEOUT": lambda: int(os.getenv("VLLM_IMAGE_FETCH_TIMEOUT", "5")),
# Timeout for fetching videos when serving multimodal models
# Default is 30 seconds
"VLLM_VIDEO_FETCH_TIMEOUT": lambda: int(
os.getenv("VLLM_VIDEO_FETCH_TIMEOUT", "30")
),
# Timeout for fetching audio when serving multimodal models
# Default is 10 seconds
"VLLM_AUDIO_FETCH_TIMEOUT": lambda: int(
os.getenv("VLLM_AUDIO_FETCH_TIMEOUT", "10")
),
# Directory for caching media downloads (images, video, audio fetched
# from URLs during inference). Empty string disables caching.
"VLLM_MEDIA_CACHE": lambda: os.getenv("VLLM_MEDIA_CACHE", ""),
# Maximum cache size in MB. When exceeded, least-recently-used entries
# are evicted. Default is 5120 (5 GB).
"VLLM_MEDIA_CACHE_MAX_SIZE_MB": lambda: int(
os.getenv("VLLM_MEDIA_CACHE_MAX_SIZE_MB", "5120")
),
# Time-to-live in hours for cached media files. Entries older than this
# are evicted regardless of cache size. Default is 24 hours.
"VLLM_MEDIA_CACHE_TTL_HOURS": lambda: float(
os.getenv("VLLM_MEDIA_CACHE_TTL_HOURS", "24")
),
# Maximum number of retries for fetching media (images, audio, video)
# from URLs. Each retry quadruples the timeout. Default is 3.
"VLLM_MEDIA_FETCH_MAX_RETRIES": lambda: int(
os.getenv("VLLM_MEDIA_FETCH_MAX_RETRIES", "3")
),
# Whether to allow HTTP redirects when fetching from media URLs.
# Default to True
"VLLM_MEDIA_URL_ALLOW_REDIRECTS": lambda: bool(
int(os.getenv("VLLM_MEDIA_URL_ALLOW_REDIRECTS", "1"))
),
# Max number of workers for the thread pool handling
# media bytes loading. Set to 1 to disable parallel processing.
# Default is 8
"VLLM_MEDIA_LOADING_THREAD_COUNT": lambda: int(
os.getenv("VLLM_MEDIA_LOADING_THREAD_COUNT", "8")
),
# Maximum filesize in MB for a single audio file when processing
# speech-to-text requests. Files larger than this will be rejected.
# Default is 25 MB
"VLLM_MAX_AUDIO_CLIP_FILESIZE_MB": lambda: int(
os.getenv("VLLM_MAX_AUDIO_CLIP_FILESIZE_MB", "25")
),
# Maximum decoded audio duration in seconds. Compressed audio files
# (e.g. OPUS at very low bitrate) can expand into gigabytes of float32
# PCM. This limit is enforced *during* decoding so the memory is never
# allocated. Default is 600s (10 minutes).
"VLLM_MAX_AUDIO_DECODE_DURATION_S": lambda: int(
os.getenv("VLLM_MAX_AUDIO_DECODE_DURATION_S", "600")
),
# Maximum float32 PCM bytes that audio decoding may allocate.
# Guards against sample-rate forgery where an attacker inflates
# the header sample rate to bypass the duration guard while the
# actual frame count still causes a multi-GiB allocation.
# Default is 256 MiB (sufficient for 600s mono 48 kHz float32).
"VLLM_MAX_AUDIO_DECODE_BYTES": lambda: int(
os.getenv("VLLM_MAX_AUDIO_DECODE_BYTES", "268435456")
),
# Maximum number of worker threads used for STT preprocessing. The default
# intentionally caps at 2 because that performed best in profiling.
# https://github.com/vllm-project/vllm/pull/44612#issuecomment-4662757781
"VLLM_MAX_AUDIO_PREPROCESS_WORKERS": lambda: int(
os.getenv(
"VLLM_MAX_AUDIO_PREPROCESS_WORKERS",
str(max(1, min(os.cpu_count() or 1, 2))),
)
),
# Maximum decoded image size in pixels. Small compressed images can
# expand into gigabytes of raster memory. This limit is enforced before
# decoding so the memory is never allocated. Default matches PIL's
# built-in 2x decompression-bomb threshold (~179M pixels, ~680 MB RGB).
"VLLM_MAX_IMAGE_PIXELS": lambda: int(
os.getenv("VLLM_MAX_IMAGE_PIXELS", "178956970")
),
# Backend for Video IO — selects the frame-sampling algorithm.
# - "opencv": uniform sampling.
# - "opencv_dynamic": duration-aware dynamic sampling.
#
# Custom backend implementations can be registered
# via `@VIDEO_LOADER_REGISTRY.register("my_custom_video_loader")` and
# imported at runtime.
# If a non-existing backend is used, an AssertionError will be thrown.
"VLLM_VIDEO_LOADER_BACKEND": lambda: os.getenv(
"VLLM_VIDEO_LOADER_BACKEND", "opencv"
),
# Media connector implementation.
# - "http": Default connector that supports fetching media via HTTP.
#
# Custom implementations can be registered
# via `@MEDIA_CONNECTOR_REGISTRY.register("my_custom_media_connector")` and
# imported at runtime.
# If a non-existing backend is used, an AssertionError will be thrown.
"VLLM_MEDIA_CONNECTOR": lambda: os.getenv("VLLM_MEDIA_CONNECTOR", "http"),
# Hash algorithm for multimodal content hashing.
# - "blake3": Default, fast cryptographic hash (not FIPS 140-3 compliant)
# - "sha256": FIPS 140-3 compliant, widely supported
# - "sha512": FIPS 140-3 compliant, faster on 64-bit systems
# Use sha256 or sha512 for FIPS compliance in government/enterprise deployments
"VLLM_MM_HASHER_ALGORITHM": env_with_choices(
"VLLM_MM_HASHER_ALGORITHM",
"blake3",
["blake3", "sha256", "sha512"],
case_sensitive=False,
),
# Path to the XLA persistent cache directory.
# Only used for XLA devices such as TPUs.
"VLLM_XLA_CACHE_PATH": lambda: os.path.expanduser(
os.getenv(
"VLLM_XLA_CACHE_PATH",
os.path.join(get_default_cache_root(), "vllm", "xla_cache"),
)
),
# If set, assert on XLA recompilation after each execution step.
"VLLM_XLA_CHECK_RECOMPILATION": lambda: bool(
int(os.getenv("VLLM_XLA_CHECK_RECOMPILATION", "0"))
),
# Enable SPMD mode for TPU backend.
"VLLM_XLA_USE_SPMD": lambda: bool(int(os.getenv("VLLM_XLA_USE_SPMD", "0"))),
# Maximum size (in MB) for logits tensor in sparse MLA indexer prefill chunks.
# Bounds the [M, N] float32 logits tensor to prevent CUDA OOM.
# Default: 512 MB
"VLLM_SPARSE_INDEXER_MAX_LOGITS_MB": lambda: int(
os.getenv("VLLM_SPARSE_INDEXER_MAX_LOGITS_MB", "512")
),
# If set, the OpenAI API server will stay alive even after the underlying
# AsyncLLMEngine errors and stops serving requests
"VLLM_KEEP_ALIVE_ON_ENGINE_DEATH": lambda: bool(
int(os.getenv("VLLM_KEEP_ALIVE_ON_ENGINE_DEATH", "0"))
),
# If the env var VLLM_ALLOW_LONG_MAX_MODEL_LEN is set, it allows
# the user to specify a max sequence length greater than
# the max length derived from the model's config.json.
# To enable this, set VLLM_ALLOW_LONG_MAX_MODEL_LEN=1.
"VLLM_ALLOW_LONG_MAX_MODEL_LEN": lambda: (
os.environ.get("VLLM_ALLOW_LONG_MAX_MODEL_LEN", "0").strip().lower()
in ("1", "true")
),
# If set, forces FP8 Marlin to be used for FP8 quantization regardless
# of the hardware support for FP8 compute.
"VLLM_TEST_FORCE_FP8_MARLIN": lambda: (
os.environ.get("VLLM_TEST_FORCE_FP8_MARLIN", "0").strip().lower()
in ("1", "true")
),
"VLLM_TEST_FORCE_LOAD_FORMAT": lambda: os.getenv(
"VLLM_TEST_FORCE_LOAD_FORMAT", "dummy"
),
# Queue size for fastsafetensors ParallelLoader pipelined weight
# loading. Peak load-time VRAM is roughly
# model_weights + (1 + queue_size) * shard_size.
# Default 0 preserves the non-pipelined memory footprint so this
# change does not shrink the loadable-model envelope. Set to 1
# (or higher) to overlap producing the next shard's device buffer
# with the consumer copying the current shard into model params,
# at the cost of `queue_size` extra shard-sized buffers resident
# at peak during loading.
"VLLM_FASTSAFETENSORS_QUEUE_SIZE": lambda: int(
os.getenv("VLLM_FASTSAFETENSORS_QUEUE_SIZE", "0")
),
# Timeout in seconds for keeping HTTP connections alive in API server
"VLLM_HTTP_TIMEOUT_KEEP_ALIVE": lambda: int(
os.environ.get("VLLM_HTTP_TIMEOUT_KEEP_ALIVE", "5")
),
# Maximum allowed value for the `n` sampling parameter (number of output
# sequences per request). Limits resource consumption to prevent
# denial-of-service via excessively large fan-out. Default: 16384.
"VLLM_MAX_N_SEQUENCES": lambda: int(
os.environ.get("VLLM_MAX_N_SEQUENCES", "16384")
),
# Maximum number of prompts allowed in a single /v1/completions request
# when the prompt field is a list. Prevents unbounded fan-out of engine
# requests from a single API call. Default: 1024.
"VLLM_MAX_COMPLETION_PROMPTS": lambda: int(
os.environ.get("VLLM_MAX_COMPLETION_PROMPTS", "1024")
),
# Maximum number of stop strings allowed in a single request.
"VLLM_MAX_STOP_STRINGS": lambda: int(os.environ.get("VLLM_MAX_STOP_STRINGS", "4")),
# Maximum number of bad-word token sequences generated per request.
"VLLM_MAX_NUM_BAD_WORDS": lambda: int(
os.environ.get("VLLM_MAX_NUM_BAD_WORDS", "128")
),
# Maximum total number of bad-word tokens (summed across all bad words)
# allowed per request. Bounds the per-request GPU buffer width.
"VLLM_MAX_BAD_WORDS_TOTAL_TOKENS": lambda: int(
os.environ.get("VLLM_MAX_BAD_WORDS_TOTAL_TOKENS", "1024")
),
# a list of plugin names to load, separated by commas.
# if this is not set, it means all plugins will be loaded
# if this is set to an empty string, no plugins will be loaded
"VLLM_PLUGINS": lambda: (
None
if "VLLM_PLUGINS" not in os.environ
else os.environ["VLLM_PLUGINS"].split(",")
),
# Retain local sliding-window KV checkpoints for prefix caching.
# Unset (default) preserves the dense local checkpointing behavior. `0`
# retains only the latest completed prompt boundary. Positive values retain
# checkpoints at the specified interval boundaries (rounded up to the
# prefix-cache alignment).
# Applies to sliding-window attention for now but not yet Mamba/linear attention.
"VLLM_PREFIX_CACHE_RETENTION_INTERVAL": lambda: (
int(os.environ["VLLM_PREFIX_CACHE_RETENTION_INTERVAL"])
if "VLLM_PREFIX_CACHE_RETENTION_INTERVAL" in os.environ
else None
),
# a local directory to look in for unrecognized LoRA adapters.
# only works if plugins are enabled and
# VLLM_ALLOW_RUNTIME_LORA_UPDATING is enabled.
"VLLM_LORA_RESOLVER_CACHE_DIR": lambda: os.getenv(
"VLLM_LORA_RESOLVER_CACHE_DIR", None
),
# A remote HF repo(s) containing one or more LoRA adapters, which
# may be downloaded and leveraged as needed. Only works if plugins
# are enabled and VLLM_ALLOW_RUNTIME_LORA_UPDATING is enabled.
# Values should be comma separated.
"VLLM_LORA_RESOLVER_HF_REPO_LIST": lambda: os.getenv(
"VLLM_LORA_RESOLVER_HF_REPO_LIST", None
),
# If set, vLLM will use Triton implementations of AWQ.
"VLLM_USE_TRITON_AWQ": lambda: bool(int(os.getenv("VLLM_USE_TRITON_AWQ", "0"))),
# If set, monkey-patch triton.runtime.autotuner.Autotuner.run to skip
# benchmarking and select the first valid config (walking past invalid
# ones). Used to eliminate autotuning variability when measuring kernel
# performance and applied before running any kernel.
"VLLM_TRITON_FORCE_FIRST_CONFIG": lambda: (
os.environ.get("VLLM_TRITON_FORCE_FIRST_CONFIG", "0").strip().lower()
in ("1", "true")
),
# If set, allow loading or unloading lora adapters in runtime,
"VLLM_ALLOW_RUNTIME_LORA_UPDATING": lambda: (
os.environ.get("VLLM_ALLOW_RUNTIME_LORA_UPDATING", "0").strip().lower()
in ("1", "true")
),
# We assume drivers can report p2p status correctly.
# If the program hangs when using custom allreduce,
# potantially caused by a bug in the driver (535 series),
# if might be helpful to set VLLM_SKIP_P2P_CHECK=0
# so that vLLM can verify if p2p is actually working.
# See https://github.com/vllm-project/vllm/blob/a9b15c606fea67a072416ea0ea115261a2756058/vllm/distributed/device_communicators/custom_all_reduce_utils.py#L101-L108 for details. # noqa
"VLLM_SKIP_P2P_CHECK": lambda: os.getenv("VLLM_SKIP_P2P_CHECK", "1") == "1",
# List of quantization kernels that should be disabled, used for testing
# and performance comparisons. Currently only affects MPLinearKernel
# selection
# (kernels: MacheteLinearKernel, MarlinLinearKernel, ExllamaLinearKernel)
"VLLM_DISABLED_KERNELS": lambda: (
[]
if "VLLM_DISABLED_KERNELS" not in os.environ
else os.environ["VLLM_DISABLED_KERNELS"].split(",")
),
"VLLM_ENABLE_FLA_PACKED_RECURRENT_DECODE": lambda: bool(
int(os.getenv("VLLM_ENABLE_FLA_PACKED_RECURRENT_DECODE", "1"))
),
# Disable pynccl (using torch.distributed instead)
"VLLM_DISABLE_PYNCCL": lambda: (
os.getenv("VLLM_DISABLE_PYNCCL", "False").lower() in ("true", "1")
),
# Optional: enable external Oink custom ops (e.g., Blackwell RMSNorm).
# Disabled by default.
"VLLM_USE_OINK_OPS": lambda: (
os.getenv("VLLM_USE_OINK_OPS", "False").lower() in ("true", "1")
),
# Disable aiter ops unless specifically enabled.
# Acts as a parent switch to enable the rest of the other operations.
# On hardware without a native MXFP8 kernel (e.g. ROCm gfx942 / MI300), the
# MXFP8 emulation path dequantizes weights MXFP8->BF16 once at load time and
# runs as a BF16 checkpoint (no per-step dequant). Set to 0 to fall back to
# per-step dequant: keeps the 1-byte MXFP8 weights (~half the weight memory)
# at the cost of dequantizing every forward step (much slower). Default on.
"VLLM_MXFP8_EMULATION_DEQUANT_AT_LOAD": lambda: (
os.getenv("VLLM_MXFP8_EMULATION_DEQUANT_AT_LOAD", "True").lower()
in ("true", "1")
),
"VLLM_ROCM_USE_AITER": lambda: (
os.getenv("VLLM_ROCM_USE_AITER", "False").lower() in ("true", "1")
),
# Use AITER's CustomAllreduce as the custom-allreduce backend inside vLLM's
# CudaCommunicator on ROCm.
"VLLM_ROCM_USE_AITER_CUSTOM_AR": lambda: (
os.getenv("VLLM_ROCM_USE_AITER_CUSTOM_AR", "True").lower() in ("true", "1")
),
# use aiter linear op if aiter ops are enabled
# The following list of related ops
# - scaled_mm (per-tensor / rowwise)
# - use aiter tuned gemms for unquantized gemms
"VLLM_ROCM_USE_AITER_LINEAR": lambda: (
os.getenv("VLLM_ROCM_USE_AITER_LINEAR", "True").lower() in ("true", "1")
),
"VLLM_ROCM_USE_AITER_LINEAR_HIPBMM": lambda: (
os.getenv("VLLM_ROCM_USE_AITER_LINEAR_HIPBMM", "False").lower() in ("true", "1")
),
# Whether to use aiter moe ops.
# By default is enabled.
"VLLM_ROCM_USE_AITER_MOE": lambda: (
os.getenv("VLLM_ROCM_USE_AITER_MOE", "True").lower() in ("true", "1")
),
# Route K3 SiTU MXFP4 MoE through the a8w4 (fp8 activation) gate/up-
# interleaved flydsl kernels instead of the default a16w4 separated path.
# This is the only flag users need: vLLM picks the kernels by passing
# gate_mode to AITER and sets the AITER-side workaround env at init.
"VLLM_ROCM_USE_AITER_MOE_SITUV2_A8W4": lambda: (
os.getenv("VLLM_ROCM_USE_AITER_MOE_SITUV2_A8W4", "0").lower() in ("true", "1")
),
# MoE sorting dispatch policy for AITER fused MoE kernels.
# 0 = auto (default): single-pass for small batches, multi-pass
# for large batches
# 1 = always single-pass: one kernel launch, no workspace,
# may be preferred for low-concurrency decode workloads
# 2 = always multi-pass: can be faster for MoE-heavy models
# (e.g., +2-5% on Qwen3-Next, +1.5% on DeepSeek-V3 at TP4,
# see PR #39177 for benchmarks)
"VLLM_ROCM_AITER_MOE_DISPATCH_POLICY": lambda: int(
os.getenv("VLLM_ROCM_AITER_MOE_DISPATCH_POLICY", "0")
),
# use aiter rms norm op if aiter ops are enabled.
"VLLM_ROCM_USE_AITER_RMSNORM": lambda: (
os.getenv("VLLM_ROCM_USE_AITER_RMSNORM", "True").lower() in ("true", "1")
),
# Whether to use aiter mla ops.
# By default is enabled.
"VLLM_ROCM_USE_AITER_MLA": lambda: (
os.getenv("VLLM_ROCM_USE_AITER_MLA", "True").lower() in ("true", "1")
),
# Small-head (<16) AITER MLA decode kernel selection. Small head counts
# (e.g. Kimi-K3: 12 heads/rank at TP8, 6 at TP16) can decode either through
# the Gluon small-head kernel or through the padded persistent-scheduling
# (PS) ASM kernel. "auto" (default) keeps Gluon for head counts that divide
# 16 where a Gluon build exists (gfx950/CDNA4) and otherwise uses the padded
# PS ASM decode; "gluon" forces the Gluon path wherever a build exists;
# "asm" forces the padded PS ASM decode. On gfx942/CDNA3 there is no Gluon
# build, so the ASM path is always used regardless of this setting.
"VLLM_ROCM_AITER_MLA_ASM_PADDING": env_with_choices(
"VLLM_ROCM_AITER_MLA_ASM_PADDING",
"auto",
["auto", "gluon", "asm"],
case_sensitive=False,
),
# Whether to use aiter mha ops.
# By default is enabled.
"VLLM_ROCM_USE_AITER_MHA": lambda: (
os.getenv("VLLM_ROCM_USE_AITER_MHA", "True").lower() in ("true", "1")
),
# Whether to use aiter fp4 gemm asm.
# By default is disabled.
"VLLM_ROCM_USE_AITER_FP4_ASM_GEMM": lambda: (
os.getenv("VLLM_ROCM_USE_AITER_FP4_ASM_GEMM", "False").lower() in ("true", "1")
),
# Whether to use aiter rope.
# By default is disabled.
"VLLM_ROCM_USE_AITER_TRITON_ROPE": lambda: (
os.getenv("VLLM_ROCM_USE_AITER_TRITON_ROPE", "False").lower() in ("true", "1")
),
# Whether to use aiter triton fp8 bmm kernel
# By default is enabled.
"VLLM_ROCM_USE_AITER_FP8BMM": lambda: (
os.getenv("VLLM_ROCM_USE_AITER_FP8BMM", "True").lower() in ("true", "1")
),
# Whether to use aiter triton fp4 bmm kernel
# By default is enabled.
"VLLM_ROCM_USE_AITER_FP4BMM": lambda: (
os.getenv("VLLM_ROCM_USE_AITER_FP4BMM", "True").lower() in ("true", "1")
),
# Use AITER triton unified attention for V1 attention
"VLLM_ROCM_USE_AITER_UNIFIED_ATTENTION": lambda: (
os.getenv("VLLM_ROCM_USE_AITER_UNIFIED_ATTENTION", "False").lower()
in ("true", "1")
),
# Whether to use aiter fusion shared experts ops.
# By default is disabled.
"VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS": lambda: (
os.getenv("VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS", "False").lower()
in ("true", "1")
),
# Whether to use aiter triton kernels for gemm ops.
# By default is enabled.
"VLLM_ROCM_USE_AITER_TRITON_GEMM": lambda: (
os.getenv("VLLM_ROCM_USE_AITER_TRITON_GEMM", "True").lower() in ("true", "1")
),
# use rocm skinny gemms
"VLLM_ROCM_USE_SKINNY_GEMM": lambda: (
os.getenv("VLLM_ROCM_USE_SKINNY_GEMM", "True").lower() in ("true", "1")
),
# Pad the fp8 weights to 256 bytes for ROCm
"VLLM_ROCM_FP8_PADDING": lambda: bool(int(os.getenv("VLLM_ROCM_FP8_PADDING", "1"))),
# Pad the weights for the moe kernel
"VLLM_ROCM_MOE_PADDING": lambda: bool(int(os.getenv("VLLM_ROCM_MOE_PADDING", "1"))),
# Whether to use the shuffled kv cache layout
"VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT": lambda: (
os.getenv("VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT", "False").lower() in ("true", "1")
),
# Custom quick allreduce kernel for MI3* cards
# Choice of quantization level: FP, INT8, INT6, INT4, INT3 or NONE
# Recommended for large models to get allreduce
"VLLM_ROCM_QUICK_REDUCE_QUANTIZATION": env_with_choices(
"VLLM_ROCM_QUICK_REDUCE_QUANTIZATION",
"NONE",
["FP", "INT8", "INT6", "INT4", "INT3", "NONE"],
),
# Custom quick allreduce kernel for MI3* cards
# Due to the lack of the bfloat16 asm instruction, bfloat16
# kernels are slower than fp16,
# If environment variable is set to 1, the input is converted to fp16
"VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16": lambda: (
os.getenv("VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16", "True").lower()
in ("true", "1")
),
# Custom quick allreduce kernel for MI3* cards.
# Controls the maximum allowed number of data bytes(MB) for custom quick
# allreduce communication.
# Default: 2048 MB.
# Data exceeding this size will use either custom allreduce or RCCL
# communication.
"VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB": lambda: maybe_convert_int(
os.environ.get("VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB", None)
),
# Custom quick allreduce kernel for MI3* cards.
# Controls the minimum allowed number of data bytes(MB) required to use
# custom quick allreduce communication.
# If unset, use the built-in threshold table.
"VLLM_ROCM_QUICK_REDUCE_MIN_SIZE_BYTES_MB": lambda: maybe_convert_int(
os.environ.get("VLLM_ROCM_QUICK_REDUCE_MIN_SIZE_BYTES_MB", None)
),
# Controls the minimum tensor size (KB, where 1 KB = 1024 bytes) required
# to use the configured QuickReduce codec. Smaller tensors use FP
# QuickReduce. This does not affect QuickReduce eligibility.
"VLLM_ROCM_QUICK_REDUCE_QUANTIZATION_MIN_SIZE_KB": lambda: maybe_convert_int(
os.environ.get("VLLM_ROCM_QUICK_REDUCE_QUANTIZATION_MIN_SIZE_KB", None)
),
# If set, enable multiprocessing in LLM for the V1 code path.
"VLLM_ENABLE_V1_MULTIPROCESSING": lambda: bool(
int(os.getenv("VLLM_ENABLE_V1_MULTIPROCESSING", "1"))
),
"VLLM_LOG_BATCHSIZE_INTERVAL": lambda: float(
os.getenv("VLLM_LOG_BATCHSIZE_INTERVAL", "-1")
),
"VLLM_DISABLE_COMPILE_CACHE": disable_compile_cache,
# If set to "0", disable LayerName opaque type for layer_name
# parameters in custom ops. Defaults to enabled on torch >= 2.11.
"VLLM_USE_LAYERNAME": lambda: bool(int(os.getenv("VLLM_USE_LAYERNAME", "1"))),
# If set, use the Rust frontend binary instead of the Python API server
# process(es).
"VLLM_USE_RUST_FRONTEND": lambda: bool(
int(os.getenv("VLLM_USE_RUST_FRONTEND", "0"))
),
# If set, use the packaged Rust client for `vllm bench serve`.
"VLLM_USE_RUST_BENCH": lambda: bool(int(os.getenv("VLLM_USE_RUST_BENCH", "0"))),
# Path to the vllm-rs binary. Defaults to "auto" which discovers the
# binary installed with the vllm package. Used when VLLM_USE_RUST_FRONTEND=1
# or VLLM_USE_RUST_BENCH=1.
"VLLM_RUST_FRONTEND_PATH": lambda: _resolve_rust_cli_path(),
# If set, vllm will run in development mode, which will enable
# some additional endpoints for developing and debugging,
# e.g. `/reset_prefix_cache`
"VLLM_SERVER_DEV_MODE": lambda: bool(int(os.getenv("VLLM_SERVER_DEV_MODE", "0"))),
# Controls the maximum number of requests to handle in a
# single asyncio task when processing per-token outputs in the
# V1 AsyncLLM interface. It is applicable when handling a high
# concurrency of streaming requests.
# Setting this too high can result in a higher variance of
# inter-message latencies. Setting it too low can negatively impact
# TTFT and overall throughput.
"VLLM_V1_OUTPUT_PROC_CHUNK_SIZE": lambda: int(
os.getenv("VLLM_V1_OUTPUT_PROC_CHUNK_SIZE", "128")
),
# If set, vLLM will disable the MLA attention optimizations.
"VLLM_MLA_DISABLE": lambda: bool(int(os.getenv("VLLM_MLA_DISABLE", "0"))),
# If set, vLLM will pick up the provided Flash Attention MLA
# Number of GPUs per worker in Ray, if it is set to be a fraction,
# it allows ray to schedule multiple actors on a single GPU,
# so that users can colocate other actors on the same GPUs as vLLM.
"VLLM_RAY_PER_WORKER_GPUS": lambda: float(
os.getenv("VLLM_RAY_PER_WORKER_GPUS", "1.0")
),
# Bundle indices for Ray, if it is set, it can control precisely
# which indices are used for the Ray bundle, for every worker.
# Format: comma-separated list of integers, e.g. "0,1,2,3"
"VLLM_RAY_BUNDLE_INDICES": lambda: os.getenv("VLLM_RAY_BUNDLE_INDICES", ""),
# In some system, find_loaded_library() may not work. So we allow users to
# specify the path through environment variable VLLM_CUDART_SO_PATH.
"VLLM_CUDART_SO_PATH": lambda: os.getenv("VLLM_CUDART_SO_PATH", None),
# Rank of the process in the data parallel setting
"VLLM_DP_RANK": lambda: int(os.getenv("VLLM_DP_RANK", "0")),
# Rank of the process in the data parallel setting.
# Defaults to VLLM_DP_RANK when not set.
"VLLM_DP_RANK_LOCAL": lambda: int(
os.getenv("VLLM_DP_RANK_LOCAL", sys.modules[__name__].VLLM_DP_RANK)
),
# World size of the data parallel setting
"VLLM_DP_SIZE": lambda: int(os.getenv("VLLM_DP_SIZE", "1")),
# IP address of the master node in the data parallel setting
"VLLM_DP_MASTER_IP": lambda: os.getenv("VLLM_DP_MASTER_IP", "127.0.0.1"),
# Port of the master node in the data parallel setting
"VLLM_DP_MASTER_PORT": lambda: int(os.getenv("VLLM_DP_MASTER_PORT", "0")),
# Randomize inputs during dummy runs when using Data Parallel
"VLLM_RANDOMIZE_DP_DUMMY_INPUTS": lambda: (
os.environ.get("VLLM_RANDOMIZE_DP_DUMMY_INPUTS", "0") == "1"
),
# Strategy to pack the data parallel ranks for Ray.
# Available options:
# - "fill":
# for DP master node, allocate exactly data-parallel-size-local DP ranks,
# for non-master nodes, allocate as many DP ranks as can fit;
# - "strict":
# allocate exactly data-parallel-size-local DP ranks to each picked node;
# - "span":
# Should be used only when a single DP rank requires multiple nodes.
# allocate one DP rank over as many nodes as required for set world_size;
# This environment variable is ignored if data-parallel-backend is not Ray.
"VLLM_RAY_DP_PACK_STRATEGY": lambda: os.getenv(
"VLLM_RAY_DP_PACK_STRATEGY", "strict"
),
# Optional comma-separated list of node IPs that Ray data-parallel
# placement groups may use. When set, create_dp_placement_groups only
# considers these nodes (the DP master node is always included).
# This environment variable is ignored if data-parallel-backend is not Ray.
"VLLM_RAY_DP_PLACEMENT_NODE_IPS": lambda: os.getenv(
"VLLM_RAY_DP_PLACEMENT_NODE_IPS", ""
),
# Comma-separated *additional* prefixes of env vars to copy from the
# driver to Ray workers. These are merged with the built-in defaults
# defined in ``vllm.ray.ray_env`` (VLLM_, etc.). Example: "MYLIB_,OTHER_"
"VLLM_RAY_EXTRA_ENV_VAR_PREFIXES_TO_COPY": lambda: os.getenv(
"VLLM_RAY_EXTRA_ENV_VAR_PREFIXES_TO_COPY", ""
),
# Comma-separated *additional* individual env var names to copy from
# the driver to Ray workers. Merged with the built-in defaults
# defined in ``vllm.ray.ray_env`` (PYTHONHASHSEED).
# Example: "MY_SECRET,MY_FLAG"
"VLLM_RAY_EXTRA_ENV_VARS_TO_COPY": lambda: os.getenv(
"VLLM_RAY_EXTRA_ENV_VARS_TO_COPY", ""
),
# Use model_redirect to redirect the model name to a local folder.
# `model_redirect` can be a json file mapping the model between
# repo_id and local folder:
# {"meta-llama/Llama-3.2-1B": "/tmp/Llama-3.2-1B"}
# or a space separated values table file:
# meta-llama/Llama-3.2-1B /tmp/Llama-3.2-1B
"VLLM_MODEL_REDIRECT_PATH": lambda: os.environ.get(
"VLLM_MODEL_REDIRECT_PATH", None
),
# Whether to use atomicAdd reduce in gptq/awq marlin kernel.
"VLLM_MARLIN_USE_ATOMIC_ADD": lambda: (
os.environ.get("VLLM_MARLIN_USE_ATOMIC_ADD", "0") == "1"
),
# The activation dtype for marlin kernel
"VLLM_MARLIN_INPUT_DTYPE": env_with_choices(
"VLLM_MARLIN_INPUT_DTYPE", None, ["int8", "fp8"]
),
# The online quantization dtype for humming kernel
"VLLM_HUMMING_ONLINE_QUANT_CONFIG": lambda: maybe_convert_json_str_or_file(
os.environ.get("VLLM_HUMMING_ONLINE_QUANT_CONFIG", None)
),
# The activation dtype config for humming kernel
"VLLM_HUMMING_INPUT_QUANT_CONFIG": lambda: maybe_convert_json_str_or_file(
os.environ.get("VLLM_HUMMING_INPUT_QUANT_CONFIG", None)
),
# Whether to use fp16 accumulator mma
"VLLM_HUMMING_USE_F16_ACCUM": lambda: maybe_convert_bool(
os.environ.get("VLLM_HUMMING_USE_F16_ACCUM", "0")
),
# Whether to use indexed gemm for humming moe
# if 1, force use indexed gemm
# if 0, force use grouped gemm
# if None, choose better gemm type automatically
"VLLM_HUMMING_MOE_GEMM_TYPE": lambda: os.environ.get(
"VLLM_HUMMING_MOE_GEMM_TYPE", None
),
# Whether to use DeepEPLL kernels for NVFP4 quantization and dispatch method
# only supported on Blackwell GPUs and with
# https://github.com/deepseek-ai/DeepEP/pull/341
"VLLM_DEEPEPLL_NVFP4_DISPATCH": lambda: bool(
int(os.getenv("VLLM_DEEPEPLL_NVFP4_DISPATCH", "0"))
),
# Whether to turn on the outlines cache for V1
# This cache is unbounded and on disk, so it's not safe to use in
# an environment with potentially malicious users.
"VLLM_V1_USE_OUTLINES_CACHE": lambda: (
os.environ.get("VLLM_V1_USE_OUTLINES_CACHE", "0") == "1"
),
# Whether using Pathways
"VLLM_TPU_USING_PATHWAYS": lambda: bool(
"proxy" in os.getenv("JAX_PLATFORMS", "").lower()
),
# Allow use of DeepGemm kernels for fused moe ops.
"VLLM_USE_DEEP_GEMM": lambda: bool(int(os.getenv("VLLM_USE_DEEP_GEMM", "1"))),
# Allow use of DeepGemm specifically for MoE fused ops (overrides only MoE).
"VLLM_MOE_USE_DEEP_GEMM": lambda: bool(
int(os.getenv("VLLM_MOE_USE_DEEP_GEMM", "1"))
),
# Whether to use E8M0 scaling when DeepGEMM is used on Blackwell GPUs.
"VLLM_USE_DEEP_GEMM_E8M0": lambda: bool(
int(os.getenv("VLLM_USE_DEEP_GEMM_E8M0", "1"))
),
# Whether to create TMA-aligned scale tensor when DeepGEMM is used.
"VLLM_USE_DEEP_GEMM_TMA_ALIGNED_SCALES": lambda: bool(
int(os.getenv("VLLM_USE_DEEP_GEMM_TMA_ALIGNED_SCALES", "1"))
),
# Opt-in MLA DCP query replication: skip the decode query all-gather.
"VLLM_DCP_Q_REPLICATE": lambda: bool(int(os.getenv("VLLM_DCP_Q_REPLICATE", "0"))),
# DeepGemm JITs the kernels on-demand. The warmup attempts to make DeepGemm
# JIT all the required kernels before model execution so there is no
# JIT'ing in the hot-path. However, this warmup increases the engine
# startup time by a couple of minutes.
# Available options:
# - "skip" : Skip warmup.
# - "full" : Warmup deepgemm by running all possible gemm shapes the
# engine could encounter.
# - "relax" : Select gemm shapes to run based on some heuristics. The
# heuristic aims to have the same effect as running all possible gemm
# shapes, but provides no guarantees.
"VLLM_DEEP_GEMM_WARMUP": env_with_choices(
"VLLM_DEEP_GEMM_WARMUP",
"relax",
[
"skip",
"full",
"relax",
],
),
# Whether to use fused grouped_topk used for MoE expert selection.
"VLLM_USE_FUSED_MOE_GROUPED_TOPK": lambda: bool(
int(os.getenv("VLLM_USE_FUSED_MOE_GROUPED_TOPK", "1"))
),
# Skip cudagraph/DP padding tokens in the MoE path by forcing their expert
# ids to -1 so the dispatch and experts drop them. Requires a MoE kernel that
# treats topk_id == -1 as a skip sentinel
"VLLM_MOE_SKIP_PADDING": lambda: bool(int(os.getenv("VLLM_MOE_SKIP_PADDING", "1"))),
# Kimi-K3 only. Under sequence-parallel MoE the dense and shared-expert MLPs
# are replicated on every rank, so each rank streams the whole weight to
# serve its own token shard. Shard them across TP instead: the MLP then
# all-gathers the full token set, computes this rank's partial, and
# reduce-scatters (which both sums across TP and restores the sequence
# sharding). Trades weight bandwidth and resident memory for two collectives
# per layer, so it only wins at low token counts: intended for decode
# instances in a P/D disaggregated deployment, not for prefill or unified
# serving.
"VLLM_KIMI_K3_SHARD_SP_SHARED_EXPERT": lambda: bool(
int(os.getenv("VLLM_KIMI_K3_SHARD_SP_SHARED_EXPERT", "0"))
),
# Allow use of FlashInfer FP8 block-scale GEMM for linear layers.
# This uses TensorRT-LLM kernels and requires SM90+ (Hopper).
"VLLM_BLOCKSCALE_FP8_GEMM_FLASHINFER": lambda: bool(
int(os.getenv("VLLM_BLOCKSCALE_FP8_GEMM_FLASHINFER", "1"))
),
# Allow use of FlashInfer MxInt4 MoE kernels for fused moe ops.
"VLLM_USE_FLASHINFER_MOE_INT4": lambda: bool(
int(os.getenv("VLLM_USE_FLASHINFER_MOE_INT4", "0"))
),
# Control the cache sized used by the xgrammar compiler. The default
# of 512 MB should be enough for roughly 1000 JSON schemas.
# It can be changed with this variable if needed for some reason.
"VLLM_XGRAMMAR_CACHE_MB": lambda: int(os.getenv("VLLM_XGRAMMAR_CACHE_MB", "512")),
# Maximum time in seconds allowed for regex compilation in structured
# output backends (xgrammar, outlines). Prevents ReDoS attacks where
# adversarial patterns cause exponential DFA state-space explosion.
# Set to 0 to disable the timeout (not recommended in production).
"VLLM_REGEX_COMPILATION_TIMEOUT_S": lambda: int(
os.getenv("VLLM_REGEX_COMPILATION_TIMEOUT_S", "5")
),
# Control the threshold for msgspec to use 'zero copy' for
# serialization/deserialization of tensors. Tensors below
# this limit will be encoded into the msgpack buffer, and
# tensors above will instead be sent via a separate message.
# While the sending side still actually copies the tensor
# in all cases, on the receiving side, tensors above this
# limit will actually be zero-copy decoded. The unit is bytes.
"VLLM_MSGPACK_ZERO_COPY_THRESHOLD": lambda: int(
os.getenv("VLLM_MSGPACK_ZERO_COPY_THRESHOLD", "256")
),
# If set, allow insecure serialization using pickle.
# This is useful for environments where it is deemed safe to use the
# insecure method and it is needed for some reason.
"VLLM_ALLOW_INSECURE_SERIALIZATION": lambda: bool(
int(os.getenv("VLLM_ALLOW_INSECURE_SERIALIZATION", "0"))
),
# Temporary: skip adding random suffix to internal request IDs. May be
# needed for KV connectors that match request IDs across instances.
"VLLM_DISABLE_REQUEST_ID_RANDOMIZATION": lambda: bool(
int(os.getenv("VLLM_DISABLE_REQUEST_ID_RANDOMIZATION", "0"))
),
# IP address used for NIXL handshake between remote agents.
"VLLM_NIXL_SIDE_CHANNEL_HOST": lambda: os.getenv(
"VLLM_NIXL_SIDE_CHANNEL_HOST", "localhost"
),
# Port used for NIXL handshake between remote agents.
"VLLM_NIXL_SIDE_CHANNEL_PORT": lambda: int(
os.getenv("VLLM_NIXL_SIDE_CHANNEL_PORT", "5600")
),
# Address the P2P KV-offload control socket binds to. Defaults to
# ``localhost`` (loopback only); must be set to the node IP for
# cross-host P2P so remote peers can reach the socket.
"VLLM_P2P_SIDE_CHANNEL_HOST": lambda: os.getenv(
"VLLM_P2P_SIDE_CHANNEL_HOST", "localhost"
),
# Port the P2P KV-offload control socket binds to.
"VLLM_P2P_SIDE_CHANNEL_PORT": lambda: int(
os.getenv("VLLM_P2P_SIDE_CHANNEL_PORT", "5710")
),
# IP address used for the EC connector's ZMQ side channel
# (producer ROUTER bind, consumer DEALER dial).
"VLLM_EC_SIDE_CHANNEL_HOST": lambda: os.getenv(
"VLLM_EC_SIDE_CHANNEL_HOST", "localhost"
),
# Port for the EC connector's ZMQ side channel; advertised to peers
# via `ec_transfer_params.peer_port` on the producer's response.
"VLLM_EC_SIDE_CHANNEL_PORT": lambda: int(
os.getenv("VLLM_EC_SIDE_CHANNEL_PORT", "5601")
),
# Port used for Mooncake handshake between remote agents.
"VLLM_MOONCAKE_BOOTSTRAP_PORT": lambda: int(
os.getenv("VLLM_MOONCAKE_BOOTSTRAP_PORT", "8998")
),
# Log per-batch memory/disk tier breakdown on external GETs.
"VLLM_MOONCAKE_STORE_TIER_LOG": lambda: (
os.getenv("VLLM_MOONCAKE_STORE_TIER_LOG", "False").lower() in ("true", "1")
),
# Number of parallel KV-load receive threads per worker rank. Lets the
# per-request control overhead (Python prep + master key lookup) of one
# request overlap with the RDMA transfer of another, keeping the transfer
# engine's queue pairs busy. Helps when that overhead is significant or
# per-request batches are too small to saturate the link on their own.
"VLLM_MOONCAKE_LOAD_RECV_THREADS": lambda: int(
os.getenv("VLLM_MOONCAKE_LOAD_RECV_THREADS", "1")
),
# Fraction of the owner's DirectIO staging buffer to fill per GET batch.
"VLLM_MOONCAKE_DISK_STAGING_USABLE_RATIO": lambda: float(
os.getenv("VLLM_MOONCAKE_DISK_STAGING_USABLE_RATIO", "0.9")
),
# Pin this rank to a specific owner segment ("host:port").
"MOONCAKE_PREFERRED_SEGMENT": lambda: os.getenv("MOONCAKE_PREFERRED_SEGMENT"),
# Override the hostname the rank registers as a Mooncake requester.
"MOONCAKE_REQUESTER_LOCAL_HOSTNAME": lambda: os.getenv(
"MOONCAKE_REQUESTER_LOCAL_HOSTNAME"
),
# Override the directory for the FlashInfer autotune config cache.
"VLLM_FLASHINFER_AUTOTUNE_CACHE_DIR": lambda: os.getenv(
"VLLM_FLASHINFER_AUTOTUNE_CACHE_DIR", None
),
# Comma-separated FlashInfer op names to exclude from autotuning, using
# the heuristic fallback tactic instead. Unset: skip "fp4_gemm" when the
# CuTe-DSL NVFP4 linear kernel is selected. Empty: skip nothing.
"VLLM_FLASHINFER_AUTOTUNE_SKIP_OPS": lambda: (
None
if "VLLM_FLASHINFER_AUTOTUNE_SKIP_OPS" not in os.environ
else [
v.strip()
for v in os.environ["VLLM_FLASHINFER_AUTOTUNE_SKIP_OPS"].split(",")
if v.strip()
]
),
# Flashinfer fused allreduce backend.
"VLLM_FLASHINFER_ALLREDUCE_BACKEND": env_with_choices(
"VLLM_FLASHINFER_ALLREDUCE_BACKEND",
"auto",
["auto", "trtllm", "mnnvl"],
),
# Control the workspace buffer size for the FlashInfer backend.
"VLLM_FLASHINFER_WORKSPACE_BUFFER_SIZE": lambda: int(
os.getenv("VLLM_FLASHINFER_WORKSPACE_BUFFER_SIZE", str(394 * 1024 * 1024))
),
# Control the maximum number of tokens per expert supported by the
# NVFP4 MoE CUTLASS Kernel. This value is used to create a buffer for
# the blockscale tensor of activations NVFP4 Quantization.
# This is used to prevent the kernel from running out of memory.
"VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE": lambda: int(
os.getenv("VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE", "163840")
),
# MoE routing strategy selector.
# See `RoutingSimulator.get_available_strategies()` # for available
# strategies.
# Custom routing strategies can be registered by
# RoutingSimulator.register_strategy()
# Note: custom strategies may not produce correct model outputs
"VLLM_MOE_ROUTING_SIMULATION_STRATEGY": lambda: os.environ.get(
"VLLM_MOE_ROUTING_SIMULATION_STRATEGY", ""
).lower(),
# Regex timeout for use by the vLLM tool parsing plugins.
"VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS": lambda: int(
os.getenv("VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS", "1")
),
# Enforce function parameter schemas in structural-tag based tool calling.
"VLLM_ENFORCE_STRICT_TOOL_CALLING": lambda: (
os.getenv("VLLM_ENFORCE_STRICT_TOOL_CALLING", "True").lower() in ("true", "1")
),
# Control the max chunk bytes (in MB) for the rpc message queue.
# Object larger than this threshold will be broadcast to worker
# processes via zmq.
"VLLM_MQ_MAX_CHUNK_BYTES_MB": lambda: int(
os.getenv("VLLM_MQ_MAX_CHUNK_BYTES_MB", "16")
),
# Timeout in seconds for execute_model RPC calls in multiprocessing
# executor (only applies when TP > 1).
"VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS": lambda: int(
os.getenv("VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS", "300")
),
# Timeout in seconds for engine and worker process shutdown
"VLLM_WORKER_SHUTDOWN_TIMEOUT_SECONDS": lambda: int(
os.getenv("VLLM_WORKER_SHUTDOWN_TIMEOUT_SECONDS", "5")
),
# KV Cache layout used throughout vllm.
# Some common values are:
# - NHD
# - HND
# Where N=num_blocks, H=num_heads and D=head_size. The default value will
# leave the layout choice to the backend. Mind that backends may only
# implement and support a subset of all possible layouts.
"VLLM_KV_CACHE_LAYOUT": env_with_choices(
"VLLM_KV_CACHE_LAYOUT", None, ["NHD", "HND"]
),
# SSM conv state layout used for Mamba models.
# - SD: (state_len, dim) — dim contiguous (default)
# - DS: (dim, state_len) — TP-sharded dim on dim1,
# consistent with SSM temporal state and HND KV cache layout.
"VLLM_SSM_CONV_STATE_LAYOUT": env_with_choices(
"VLLM_SSM_CONV_STATE_LAYOUT", None, ["SD", "DS"]
),
# Enable checking whether the generated logits contain NaNs,
# indicating corrupted output. Useful for debugging low level bugs
# or bad hardware but it may add compute overhead.
"VLLM_COMPUTE_NANS_IN_LOGITS": lambda: bool(
int(os.getenv("VLLM_COMPUTE_NANS_IN_LOGITS", "0"))
or int(os.getenv("VLLM_RAISE_ON_LOGIT_NANS", "0"))
),
# Raise an exception when generated logits contain NaNs. Enabling this
# also enables the NaN computation required to detect them.
"VLLM_RAISE_ON_LOGIT_NANS": lambda: bool(
int(os.getenv("VLLM_RAISE_ON_LOGIT_NANS", "0"))
),
# Timeout (in seconds) for MooncakeConnector in PD disaggregated setup.
"VLLM_MOONCAKE_ABORT_REQUEST_TIMEOUT": lambda: int(
os.getenv("VLLM_MOONCAKE_ABORT_REQUEST_TIMEOUT", "480")
),
# If set, it means we pre-downloaded cubin files and flashinfer will
# read the cubin files directly.
"VLLM_HAS_FLASHINFER_CUBIN": lambda: bool(
int(os.getenv("VLLM_HAS_FLASHINFER_CUBIN", "0"))
),
# Controls garbage collection during CUDA graph capture.
# If set to 0 (default), enables GC freezing to speed up capture time.
# If set to 1, allows GC to run during capture.
"VLLM_ENABLE_CUDAGRAPH_GC": lambda: bool(
int(os.getenv("VLLM_ENABLE_CUDAGRAPH_GC", "0"))
),
# Used to force set up loopback IP
"VLLM_LOOPBACK_IP": lambda: os.getenv("VLLM_LOOPBACK_IP", ""),
# Used to set the process name prefix for vLLM processes.
# This is useful for debugging and monitoring purposes.
# The default value is "VLLM".
"VLLM_PROCESS_NAME_PREFIX": lambda: os.getenv("VLLM_PROCESS_NAME_PREFIX", "VLLM"),
# Allow chunked local attention with hybrid kv cache manager.
# Currently using the Hybrid KV cache manager with chunked local attention
# in the Llama4 models (the only models currently using chunked local attn)
# causes a latency regression. For this reason, we disable it by default.
# This flag is used to allow users to enable it if they want to (to save on
# kv-cache memory usage and enable longer contexts)
# TODO(lucas): Remove this flag once latency regression is resolved.
"VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE": lambda: bool(
int(os.getenv("VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE", "1"))
),
# Enables support for the "store" option in the OpenAI Responses API.
# When set to 1, vLLM's OpenAI server will retain the input and output
# messages for those requests in memory. By default, this is disabled (0),
# and the "store" option is ignored.
# NOTE/WARNING:
# 1. Messages are kept in memory only (not persisted to disk) and will be
# lost when the vLLM server shuts down.
# 2. Enabling this option will cause a memory leak, as stored messages are
# never removed from memory until the server terminates.
"VLLM_ENABLE_RESPONSES_API_STORE": lambda: bool(
int(os.getenv("VLLM_ENABLE_RESPONSES_API_STORE", "0"))
),
# If set to 1, expose the Cohere Chat v2 API at ``POST /cohere/v2/chat``.
# Default off
"VLLM_ENABLE_COHERE_API": lambda: bool(
int(os.getenv("VLLM_ENABLE_COHERE_API", "0"))
),
# If set, use the fp8 mfma in rocm paged attention.
"VLLM_ROCM_FP8_MFMA_PAGE_ATTN": lambda: bool(
int(os.getenv("VLLM_ROCM_FP8_MFMA_PAGE_ATTN", "0"))
),
# Whether to use pytorch symmetric memory for allreduce
"VLLM_ALLREDUCE_USE_SYMM_MEM": lambda: bool(
int(os.getenv("VLLM_ALLREDUCE_USE_SYMM_MEM", "1"))
),
# Whether to use FlashInfer allreduce
"VLLM_ALLREDUCE_USE_FLASHINFER": lambda: bool(
int(os.getenv("VLLM_ALLREDUCE_USE_FLASHINFER", "0"))
),
# Experimental: use this to enable MCP tool calling for non harmony models
"VLLM_USE_EXPERIMENTAL_PARSER_CONTEXT": lambda: bool(
int(os.getenv("VLLM_USE_EXPERIMENTAL_PARSER_CONTEXT", "0"))
),
# User override folder for tuned Triton-kernel configs. Shared by MoE,
# Mamba SSU, and LoRA. Filenames are distinct so one folder can hold all.
# Each component first checks this folder, then the configs shipped with
# vLLM (if any). If no JSON matches, it uses a hard-coded heuristic.
"VLLM_TUNED_CONFIG_FOLDER": lambda: os.getenv("VLLM_TUNED_CONFIG_FOLDER", None),
# Opt-in persistence of the startup plan. When enabled, each worker
# saves the memory-profiling result (the suggested --kv-cache-memory value
# and the free-memory baseline) under VLLM_CACHE_ROOT/startup_plan/,
# keyed by a hardware+config fingerprint, and later boots auto-apply it
# -- skipping memory profiling -- when the fingerprint matches and
# current free memory >= the recorded baseline.
# See vllm/v1/worker/startup_plan.py.
"VLLM_ENABLE_STARTUP_PLAN": lambda: bool(
int(os.getenv("VLLM_ENABLE_STARTUP_PLAN", "0"))
),
# Valid values are container,code_interpreter,web_search_preview
# ex VLLM_GPT_OSS_SYSTEM_TOOL_MCP_LABELS=container,code_interpreter
# If the server_label of your mcp tool is not in this list it will
# be completely ignored.
"VLLM_GPT_OSS_SYSTEM_TOOL_MCP_LABELS": env_set_with_choices(
"VLLM_GPT_OSS_SYSTEM_TOOL_MCP_LABELS",
default=[],
choices=["container", "code_interpreter", "web_search_preview"],
),
# Allows harmony instructions to be injected on system messages
"VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS": lambda: bool(
int(os.getenv("VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS", "0"))
),
# Pin the conversation start date injected into the Harmony system
# message. When unset the current date is used, which introduces
# non-determinism (different tokens -> different model behaviour at
# temperature=0). Set to an ISO date string, e.g. "2023-09-12",
# for reproducible inference or testing.
"VLLM_SYSTEM_START_DATE": lambda: os.getenv("VLLM_SYSTEM_START_DATE", None),
# Enable automatic retry when tool call JSON parsing fails
# If enabled, returns an error message to the model to retry
# If disabled (default), raises an exception and fails the request
"VLLM_TOOL_JSON_ERROR_AUTOMATIC_RETRY": lambda: bool(
int(os.getenv("VLLM_TOOL_JSON_ERROR_AUTOMATIC_RETRY", "0"))
),
# Add optional custom scopes for profiling, disable to avoid overheads
"VLLM_CUSTOM_SCOPES_FOR_PROFILING": lambda: bool(
int(os.getenv("VLLM_CUSTOM_SCOPES_FOR_PROFILING", "0"))
),
# Add optional nvtx scopes for profiling, disable to avoid overheads
"VLLM_NVTX_SCOPES_FOR_PROFILING": lambda: bool(
int(os.getenv("VLLM_NVTX_SCOPES_FOR_PROFILING", "0"))
),
# Represent block hashes in KV cache events as 64-bit integers instead of
# raw bytes. Defaults to True for backward compatibility.
"VLLM_KV_EVENTS_USE_INT_BLOCK_HASHES": lambda: bool(
int(os.getenv("VLLM_KV_EVENTS_USE_INT_BLOCK_HASHES", "1"))
),
# Name of the shared memory buffer used for object storage.
# Only effective when mm_config.mm_processor_cache_type == "shm".
# Automatically generates a unique UUID-based name per process tree
# if not explicitly set.
"VLLM_OBJECT_STORAGE_SHM_BUFFER_NAME": get_env_or_set_default(
"VLLM_OBJECT_STORAGE_SHM_BUFFER_NAME",
lambda: f"VLLM_OBJECT_STORAGE_SHM_BUFFER_{uuid.uuid4().hex}",
),
# The size in MB of the buffers (NVL and RDMA) used by DeepEP
"VLLM_DEEPEP_BUFFER_SIZE_MB": lambda: int(
os.getenv("VLLM_DEEPEP_BUFFER_SIZE_MB", "1024")
),
# Force DeepEP to use intranode kernel for inter-node communication in
# high throughput mode. This is useful archive higher prefill throughput
# on system supports multi-node nvlink (e.g GB200).
"VLLM_DEEPEP_HIGH_THROUGHPUT_FORCE_INTRA_NODE": lambda: bool(
int(os.getenv("VLLM_DEEPEP_HIGH_THROUGHPUT_FORCE_INTRA_NODE", "0"))
),
# Allow DeepEP to use MNNVL (multi-node nvlink) for internode_ll kernel,
# turn this for better latency on GB200 like system
"VLLM_DEEPEP_LOW_LATENCY_USE_MNNVL": lambda: bool(
int(os.getenv("VLLM_DEEPEP_LOW_LATENCY_USE_MNNVL", "0"))
),
# DeepEP v2: enable two-tier NVLink+RDMA hybrid mode
"VLLM_DEEPEP_V2_ALLOW_HYBRID_MODE": lambda: bool(
int(os.getenv("VLLM_DEEPEP_V2_ALLOW_HYBRID_MODE", "0"))
),
# DeepEP v2: use fewer SMs at slight throughput cost
"VLLM_DEEPEP_V2_PREFER_OVERLAP": lambda: bool(
int(os.getenv("VLLM_DEEPEP_V2_PREFER_OVERLAP", "0"))
),
# DeepEP v2: trade precision for transfer size in combine
"VLLM_DEEPEP_V2_ALLOW_MULTIPLE_REDUCTION": lambda: bool(
int(os.getenv("VLLM_DEEPEP_V2_ALLOW_MULTIPLE_REDUCTION", "0"))
),
# The number of SMs/CUs to allocate for communication kernels when
# running DBO; the rest will be allocated to compute.
# Default: 20 on CUDA (SMs), 64 on ROCm (CUs).
"VLLM_DBO_COMM_SMS": lambda: int(
os.getenv(
"VLLM_DBO_COMM_SMS",
"64"
if hasattr(__import__("torch").version, "hip")
and __import__("torch").version.hip is not None
else "20",
)
),
# Enable max_autotune & coordinate_descent_tuning in inductor_config
# to compile static shapes passed from compile_sizes in compilation_config
# If set to 1, enable max_autotune; By default, this is enabled (1)
"VLLM_ENABLE_INDUCTOR_MAX_AUTOTUNE": lambda: bool(
int(os.getenv("VLLM_ENABLE_INDUCTOR_MAX_AUTOTUNE", "1"))
),
# If set to 1, enable coordinate_descent_tuning;
# By default, this is enabled (1)
"VLLM_ENABLE_INDUCTOR_COORDINATE_DESCENT_TUNING": lambda: bool(
int(os.getenv("VLLM_ENABLE_INDUCTOR_COORDINATE_DESCENT_TUNING", "1"))
),
# Flag to enable NCCL symmetric memory allocation and registration
"VLLM_USE_NCCL_SYMM_MEM": lambda: bool(
int(os.getenv("VLLM_USE_NCCL_SYMM_MEM", "0"))
),
# NCCL header path
"VLLM_NCCL_INCLUDE_PATH": lambda: os.environ.get("VLLM_NCCL_INCLUDE_PATH", None),
# GC debug config
# - VLLM_GC_DEBUG=0: disable GC debugger
# - VLLM_GC_DEBUG=1: enable GC debugger with gc.collect elpased times
# - VLLM_GC_DEBUG='{"top_objects":5}': enable GC debugger with
# top 5 collected objects
"VLLM_GC_DEBUG": lambda: os.getenv("VLLM_GC_DEBUG", ""),
# Debug workspace allocations.
# logging of workspace resize operations.
"VLLM_DEBUG_WORKSPACE": lambda: bool(int(os.getenv("VLLM_DEBUG_WORKSPACE", "0"))),
# Disables parallel execution of shared_experts via separate cuda stream
"VLLM_DISABLE_SHARED_EXPERTS_STREAM": lambda: bool(
int(os.getenv("VLLM_DISABLE_SHARED_EXPERTS_STREAM", "0"))
),
# Limits when we run shared_experts in a separate stream.
# We found out that for large batch sizes, the separate stream
# execution is not beneficial (most likely because of the input clone)
# TODO(alexm-redhat): Tune to be more dynamic based on GPU type
"VLLM_SHARED_EXPERTS_STREAM_TOKEN_THRESHOLD": lambda: int(
int(os.getenv("VLLM_SHARED_EXPERTS_STREAM_TOKEN_THRESHOLD", 256))
),
# Token-count cutoff for multi-stream overlap of the attention input
# GEMM with auxiliary GEMMs (e.g. fused_wqa_wkv overlapped with indexer
# weights / kv-score projections in DeepSeek-V4). At or below this many
# tokens the FP8 main GEMM has idle SMs to share with the bf16 aux GEMMs
# and overlap is a 5-45% win; above it the FP8 GEMM saturates the device
# and the cross-stream sync becomes pure overhead. Set to 0 to disable
# the multi-stream path entirely. See #PR 41526 for the empirical result
# for the default value of 1024 tokens.
"VLLM_MULTI_STREAM_GEMM_TOKEN_THRESHOLD": lambda: int(
os.getenv("VLLM_MULTI_STREAM_GEMM_TOKEN_THRESHOLD", "1024")
),
# Format for saving torch.compile cache artifacts
# - "binary": saves as binary file
# Safe for multiple vllm serve processes accessing the same torch compile cache.
# - "unpacked": saves as directory structure (for inspection/debugging)
# NOT multiprocess safe - race conditions may occur with multiple processes.
# Allows viewing and setting breakpoints in Inductor's code output files.
"VLLM_COMPILE_CACHE_SAVE_FORMAT": env_with_choices(
"VLLM_COMPILE_CACHE_SAVE_FORMAT", "binary", ["binary", "unpacked"]
),
# Flag to control the v2 model runner. If unset, use config defaults.
"VLLM_USE_V2_MODEL_RUNNER": lambda: maybe_convert_bool(
os.getenv("VLLM_USE_V2_MODEL_RUNNER", None)
),
# Log model inspection after loading.
# If enabled, logs a transformers-style hierarchical view of the model
# with quantization methods and attention backends.
"VLLM_LOG_MODEL_INSPECTION": lambda: bool(
int(os.getenv("VLLM_LOG_MODEL_INSPECTION", "0"))
),
# Debug logging for --enable-mfu-metrics
"VLLM_DEBUG_MFU_METRICS": lambda: bool(
int(os.getenv("VLLM_DEBUG_MFU_METRICS", "0"))
),
# Disable using pytorch's pin memory for CPU offloading.
"VLLM_WEIGHT_OFFLOADING_DISABLE_PIN_MEMORY": lambda: bool(
int(os.getenv("VLLM_WEIGHT_OFFLOADING_DISABLE_PIN_MEMORY", "0"))
),
# Disable using UVA (Unified Virtual Addressing) for CPU offloading.
"VLLM_WEIGHT_OFFLOADING_DISABLE_UVA": lambda: bool(
int(os.getenv("VLLM_WEIGHT_OFFLOADING_DISABLE_UVA", "0"))
),
# On WSL2 with a compatible kernel (>= 4.19.121), pinned memory is
# supported but disabled by default due to a small performance regression.
# Set to 1 when pinned memory or UVA is required (e.g. CPU offloading
# or v2 model runner).
"VLLM_WSL2_ENABLE_PIN_MEMORY": lambda: bool(
int(os.getenv("VLLM_WSL2_ENABLE_PIN_MEMORY", "0"))
),
# Disable logging of vLLM logo at server startup time.
"VLLM_DISABLE_LOG_LOGO": lambda: bool(int(os.getenv("VLLM_DISABLE_LOG_LOGO", "0"))),
# Disable PDL for LoRA, as enabling PDL with LoRA on SM100 causes
# Triton compilation to fail.
"VLLM_LORA_DISABLE_PDL": lambda: bool(int(os.getenv("VLLM_LORA_DISABLE_PDL", "0"))),
# Enable CUDA compatibility mode for datacenter GPUs with older
# driver versions than the CUDA toolkit major version of vLLM.
"VLLM_ENABLE_CUDA_COMPATIBILITY": lambda: (
os.environ.get("VLLM_ENABLE_CUDA_COMPATIBILITY", "0").strip().lower()
in ("1", "true")
),
# Path to the CUDA compatibility libraries when CUDA compatibility is enabled.
"VLLM_CUDA_COMPATIBILITY_PATH": lambda: os.environ.get(
"VLLM_CUDA_COMPATIBILITY_PATH", None
),
# Skip model name validation in OpenAI API requests.
# When set to 1, any model name will be accepted in the 'model' field
# of API requests. This is useful for proxy/gateway scenarios where
# the actual model is served but different names may be used in requests.
"VLLM_SKIP_MODEL_NAME_VALIDATION": lambda: (
os.getenv("VLLM_SKIP_MODEL_NAME_VALIDATION", "0").strip().lower()
in ("1", "true")
),
# Whether it is a scale up launch engine for elastic EP,
# Should only be set by EngineCoreClient.
"VLLM_ELASTIC_EP_SCALE_UP_LAUNCH": lambda: bool(
int(os.getenv("VLLM_ELASTIC_EP_SCALE_UP_LAUNCH", "0"))
),
# Whether to wait for all requests to drain before sending the
# scaling command in elastic EP.
"VLLM_ELASTIC_EP_DRAIN_REQUESTS": lambda: bool(
int(os.getenv("VLLM_ELASTIC_EP_DRAIN_REQUESTS", "0"))
),
# If set to 1, enable CUDA graph memory estimation during memory profiling.
# This profiles CUDA graph memory usage to provide more accurate KV cache
# memory allocation. Enabled by default as of v0.21.0
"VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS": lambda: bool(
int(os.getenv("VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS", "1"))
),
# NIXL EP environment variables
"VLLM_NIXL_EP_MAX_NUM_RANKS": lambda: int(
os.getenv("VLLM_NIXL_EP_MAX_NUM_RANKS", "32")
),
# Whether enable XPU graph on Intel GPU
"VLLM_XPU_ENABLE_XPU_GRAPH": lambda: bool(
int(os.getenv("VLLM_XPU_ENABLE_XPU_GRAPH", "0"))
),
# whether use xpu specific sample kernel
"VLLM_XPU_USE_SAMPLER_KERNEL": lambda: bool(
int(os.getenv("VLLM_XPU_USE_SAMPLER_KERNEL", "1"))
),
# Enable simple KV offload.
"VLLM_USE_SIMPLE_KV_OFFLOAD": lambda: bool(
int(os.getenv("VLLM_USE_SIMPLE_KV_OFFLOAD", "0"))
),
# Direct DCP ops default on when applicable; set to 1 to enforce or 0 to disable.
"VLLM_USE_DIRECT_DCP_A2A": lambda: maybe_convert_bool(
os.getenv("VLLM_USE_DIRECT_DCP_A2A")
),
"VLLM_USE_DIRECT_DCP_Q_GATHER": lambda: maybe_convert_bool(
os.getenv("VLLM_USE_DIRECT_DCP_Q_GATHER")
),
"VLLM_USE_DIRECT_DCP_KV_GATHER": lambda: maybe_convert_bool(
os.getenv("VLLM_USE_DIRECT_DCP_KV_GATHER")
),
# Whether to enable dual cuda streams for LoRA computation
# (used by both BaseLinearLayerWithLoRA and FusedMoEWithLoRA to
# overlap the base layer compute with the LoRA fast path).
"VLLM_LORA_ENABLE_DUAL_STREAM": lambda: bool(
int(os.getenv("VLLM_LORA_ENABLE_DUAL_STREAM", "0"))
),
# If set to 1, use Python spinloop extension to poll in a more efficient
# way when using the mp backend.
"VLLM_USE_SPINLOOP_EXT": lambda: bool(int(os.getenv("VLLM_USE_SPINLOOP_EXT", "0"))),
# Comma-separated GPU_BDF=NIC_BDF pairs for RDMA NIC selection.
# Must be set together with VLLM_NIC_SELECTION_VARS.
"VLLM_GPU_NIC_PCIE_MAPPING": lambda: os.getenv("VLLM_GPU_NIC_PCIE_MAPPING", ""),
# Comma-separated list of env vars to set from the GPU-NIC mapping.
# Each entry is VAR_NAME or VAR_NAME:<suffix> (suffix appended to
# RDMA device name). Must be set together with VLLM_GPU_NIC_PCIE_MAPPING.
"VLLM_NIC_SELECTION_VARS": lambda: os.getenv("VLLM_NIC_SELECTION_VARS", ""),
}