envs.py 66.7 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import functools
5
import hashlib
6
import json
7
import os
8
import sys
9
import tempfile
10
11
from collections.abc import Callable
from typing import TYPE_CHECKING, Any, Literal
12
13
14

if TYPE_CHECKING:
    VLLM_HOST_IP: str = ""
15
    VLLM_PORT: int | None = None
16
    VLLM_RPC_BASE_PATH: str = tempfile.gettempdir()
17
    VLLM_USE_MODELSCOPE: bool = False
18
    VLLM_RINGBUFFER_WARNING_INTERVAL: int = 60
19
20
    VLLM_NCCL_SO_PATH: str | None = None
    LD_LIBRARY_PATH: str | None = None
21
    VLLM_USE_TRITON_FLASH_ATTN: bool = True
22
    VLLM_V1_USE_PREFILL_DECODE_ATTENTION: bool = False
23
    VLLM_FLASH_ATTN_VERSION: int | None = None
24
    LOCAL_RANK: int = 0
25
    CUDA_VISIBLE_DEVICES: str | None = None
26
    VLLM_ENGINE_ITERATION_TIMEOUT_S: int = 60
27
    VLLM_API_KEY: str | None = None
28
    VLLM_DEBUG_LOG_API_SERVER_RESPONSE: bool = False
29
30
31
32
    S3_ACCESS_KEY_ID: str | None = None
    S3_SECRET_ACCESS_KEY: str | None = None
    S3_ENDPOINT_URL: str | None = None
    VLLM_MODEL_REDIRECT_PATH: str | None = None
33
34
    VLLM_CACHE_ROOT: str = os.path.expanduser("~/.cache/vllm")
    VLLM_CONFIG_ROOT: str = os.path.expanduser("~/.config/vllm")
35
36
    VLLM_USAGE_STATS_SERVER: str = "https://stats.vllm.ai"
    VLLM_NO_USAGE_STATS: bool = False
37
    VLLM_DISABLE_FLASHINFER_PREFILL: bool = False
38
39
40
    VLLM_DO_NOT_TRACK: bool = False
    VLLM_USAGE_SOURCE: str = ""
    VLLM_CONFIGURE_LOGGING: int = 1
41
    VLLM_LOGGING_LEVEL: str = "INFO"
42
    VLLM_LOGGING_PREFIX: str = ""
43
    VLLM_LOGGING_STREAM: str = "ext://sys.stdout"
44
    VLLM_LOGGING_CONFIG_PATH: str | None = None
45
    VLLM_LOG_STATS_INTERVAL: float = 10.0
46
    VLLM_TRACE_FUNCTION: int = 0
47
48
49
50
    VLLM_ATTENTION_BACKEND: str | None = None
    VLLM_USE_FLASHINFER_SAMPLER: bool | None = None
    VLLM_PP_LAYER_PARTITION: str | None = None
    VLLM_CPU_KVCACHE_SPACE: int | None = 0
51
    VLLM_CPU_OMP_THREADS_BIND: str = ""
52
    VLLM_CPU_NUM_OF_RESERVED_CPU: int | None = None
53
    VLLM_CPU_MOE_PREPACK: bool = True
54
    VLLM_CPU_SGL_KERNEL: bool = False
55
    VLLM_XLA_CACHE_PATH: str = os.path.join(VLLM_CACHE_ROOT, "xla_cache")
56
    VLLM_XLA_CHECK_RECOMPILATION: bool = False
57
    VLLM_FUSED_MOE_CHUNK_SIZE: int = 64 * 1024
58
    VLLM_ENABLE_FUSED_MOE_ACTIVATION_CHUNKING: bool = True
59
    VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE: Literal["auto", "nccl", "shm"] = "auto"
60
    VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM: bool = False
61
    VLLM_USE_RAY_WRAPPED_PP_COMM: bool = True
62
    VLLM_XLA_USE_SPMD: bool = False
63
    VLLM_WORKER_MULTIPROC_METHOD: Literal["fork", "spawn"] = "fork"
64
    VLLM_ASSETS_CACHE: str = os.path.join(VLLM_CACHE_ROOT, "assets")
65
    VLLM_ASSETS_CACHE_MODEL_CLEAN: bool = False
66
    VLLM_IMAGE_FETCH_TIMEOUT: int = 5
67
    VLLM_VIDEO_FETCH_TIMEOUT: int = 30
68
    VLLM_AUDIO_FETCH_TIMEOUT: int = 10
69
    VLLM_MEDIA_URL_ALLOW_REDIRECTS: bool = True
70
    VLLM_MEDIA_LOADING_THREAD_COUNT: int = 8
71
    VLLM_MAX_AUDIO_CLIP_FILESIZE_MB: int = 25
72
    VLLM_VIDEO_LOADER_BACKEND: str = "opencv"
73
    VLLM_MM_INPUT_CACHE_GIB: int = 4
74
    VLLM_TARGET_DEVICE: str = "cuda"
75
    VLLM_MAIN_CUDA_VERSION: str = "12.8"
76
77
    MAX_JOBS: str | None = None
    NVCC_THREADS: str | None = None
78
    VLLM_USE_PRECOMPILED: bool = False
79
    VLLM_DOCKER_BUILD_CONTEXT: bool = False
80
    VLLM_TEST_USE_PRECOMPILED_NIGHTLY_WHEEL: bool = False
81
    VLLM_KEEP_ALIVE_ON_ENGINE_DEATH: bool = False
82
    CMAKE_BUILD_TYPE: Literal["Debug", "Release", "RelWithDebInfo"] | None = None
83
    VERBOSE: bool = False
84
    VLLM_ALLOW_LONG_MAX_MODEL_LEN: bool = False
85
    VLLM_RPC_TIMEOUT: int = 10000  # ms
86
    VLLM_HTTP_TIMEOUT_KEEP_ALIVE: int = 5  # seconds
87
88
89
    VLLM_PLUGINS: list[str] | None = None
    VLLM_LORA_RESOLVER_CACHE_DIR: str | None = None
    VLLM_TORCH_PROFILER_DIR: str | None = None
90
91
    VLLM_TORCH_PROFILER_RECORD_SHAPES: bool = False
    VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY: bool = False
92
93
    VLLM_USE_AOT_COMPILE: bool = False
    VLLM_FORCE_AOT_LOAD: bool = False
94
95
    VLLM_TORCH_PROFILER_WITH_STACK: bool = True
    VLLM_TORCH_PROFILER_WITH_FLOPS: bool = False
96
    VLLM_USE_TRITON_AWQ: bool = False
97
    VLLM_ALLOW_RUNTIME_LORA_UPDATING: bool = False
98
    VLLM_SKIP_P2P_CHECK: bool = False
99
    VLLM_DISABLED_KERNELS: list[str] = []
100
    VLLM_DISABLE_PYNCCL: bool = False
101
    VLLM_USE_V1: bool = True
102
    VLLM_ROCM_USE_AITER: bool = False
103
    VLLM_ROCM_USE_AITER_PAGED_ATTN: bool = False
104
    VLLM_ROCM_USE_AITER_LINEAR: bool = True
105
    VLLM_ROCM_USE_AITER_MOE: bool = True
106
    VLLM_ROCM_USE_AITER_RMSNORM: bool = True
107
    VLLM_ROCM_USE_AITER_MLA: bool = True
108
    VLLM_ROCM_USE_AITER_MHA: bool = True
109
110
    VLLM_ROCM_USE_AITER_FP4_ASM_GEMM: bool = False
    VLLM_ROCM_USE_TRITON_ROPE: bool = False
111
    VLLM_ROCM_USE_AITER_FP8BMM: bool = True
112
    VLLM_ROCM_USE_AITER_UNIFIED_ATTENTION: bool = False
113
    VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS: bool = True
114
    VLLM_ROCM_USE_SKINNY_GEMM: bool = True
115
    VLLM_ROCM_FP8_PADDING: bool = True
116
    VLLM_ROCM_MOE_PADDING: bool = True
117
    VLLM_ROCM_CUSTOM_PAGED_ATTN: bool = True
118
    VLLM_ENABLE_V1_MULTIPROCESSING: bool = True
119
    VLLM_LOG_BATCHSIZE_INTERVAL: float = -1
120
    VLLM_DISABLE_COMPILE_CACHE: bool = False
121
    Q_SCALE_CONSTANT: int = 200
122
123
    K_SCALE_CONSTANT: int = 200
    V_SCALE_CONSTANT: int = 100
124
    VLLM_SERVER_DEV_MODE: bool = False
125
    VLLM_V1_OUTPUT_PROC_CHUNK_SIZE: int = 128
126
    VLLM_MLA_DISABLE: bool = False
127
    VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH: int = 32
128
129
    VLLM_RAY_PER_WORKER_GPUS: float = 1.0
    VLLM_RAY_BUNDLE_INDICES: str = ""
130
    VLLM_CUDART_SO_PATH: str | None = None
131
    VLLM_DP_RANK: int = 0
132
    VLLM_DP_RANK_LOCAL: int = -1
133
    VLLM_DP_SIZE: int = 1
134
    VLLM_USE_STANDALONE_COMPILE: bool = False
135
136
    VLLM_DP_MASTER_IP: str = ""
    VLLM_DP_MASTER_PORT: int = 0
137
    VLLM_MOE_DP_CHUNK_SIZE: int = 256
138
    VLLM_RANDOMIZE_DP_DUMMY_INPUTS: bool = False
139
    VLLM_RAY_DP_PACK_STRATEGY: Literal["strict", "fill", "span"] = "strict"
140
    VLLM_MARLIN_USE_ATOMIC_ADD: bool = False
141
    VLLM_MXFP4_USE_MARLIN: bool | None = None
142
    VLLM_V1_USE_OUTLINES_CACHE: bool = False
143
    VLLM_TPU_BUCKET_PADDING_GAP: int = 0
144
    VLLM_TPU_MOST_MODEL_LEN: int | None = None
145
    VLLM_TPU_USING_PATHWAYS: bool = False
146
    VLLM_USE_DEEP_GEMM: bool = True
147
    VLLM_USE_DEEP_GEMM_E8M0: bool = True
148
149
150
151
152
    VLLM_DEEP_GEMM_WARMUP: Literal[
        "skip",
        "full",
        "relax",
    ] = "relax"
153
    VLLM_USE_FUSED_MOE_GROUPED_TOPK: bool = True
154
    VLLM_USE_FLASHINFER_MOE_FP16: bool = False
155
156
    VLLM_USE_FLASHINFER_MOE_FP8: bool = False
    VLLM_USE_FLASHINFER_MOE_FP4: bool = False
157
    VLLM_FLASHINFER_MOE_BACKEND: Literal["throughput", "latency"] = "throughput"
158
    VLLM_XGRAMMAR_CACHE_MB: int = 0
159
    VLLM_MSGPACK_ZERO_COPY_THRESHOLD: int = 256
160
    VLLM_ALLOW_INSECURE_SERIALIZATION: bool = False
Robert Shaw's avatar
Robert Shaw committed
161
    VLLM_NIXL_SIDE_CHANNEL_HOST: str = "localhost"
162
    VLLM_NIXL_SIDE_CHANNEL_PORT: int = 5600
163
164
165
166
167
168
169
170
    VLLM_ALL2ALL_BACKEND: Literal[
        "naive",
        "pplx",
        "deepep_high_throughput",
        "deepep_low_latency",
        "allgather_reducescatter",
        "flashinfer_all2allv",
    ] = "allgather_reducescatter"
171
    VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE: int = 163840
172
    VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS: int = 1
173
    VLLM_SLEEP_WHEN_IDLE: bool = False
174
    VLLM_MQ_MAX_CHUNK_BYTES_MB: int = 16
175
    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: int = 300
176
    VLLM_KV_CACHE_LAYOUT: Literal["NHD", "HND"] | None = None
177
    VLLM_COMPUTE_NANS_IN_LOGITS: bool = False
178
    VLLM_USE_NVFP4_CT_EMULATIONS: bool = False
179
180
181
    VLLM_ROCM_QUICK_REDUCE_QUANTIZATION: Literal[
        "FP", "INT8", "INT6", "INT4", "NONE"
    ] = "NONE"
182
    VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16: bool = True
183
    VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB: int | None = None
184
    VLLM_NIXL_ABORT_REQUEST_TIMEOUT: int = 480
185
    VLLM_USE_CUDNN_PREFILL: bool = False
186
    VLLM_ENABLE_CUDAGRAPH_GC: bool = False
187
    VLLM_LOOPBACK_IP: str = ""
188
    VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE: bool = False
189
    VLLM_ENABLE_RESPONSES_API_STORE: bool = False
190
    VLLM_USE_TRTLLM_ATTENTION: str | None = None
191
    VLLM_NVFP4_GEMM_BACKEND: str | None = None
192
    VLLM_FLASHINFER_DISABLE_Q_QUANTIZATION: bool = False
193
    VLLM_HAS_FLASHINFER_CUBIN: bool = False
194
195
    VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8: bool = False
    VLLM_USE_FLASHINFER_MOE_MXFP4_BF16: bool = False
xiao-llm's avatar
xiao-llm committed
196
    VLLM_ROCM_FP8_MFMA_PAGE_ATTN: bool = False
197
    VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS: bool = False
198
    VLLM_ALLREDUCE_USE_SYMM_MEM: bool = False
199
    VLLM_TUNED_CONFIG_FOLDER: str | None = None
200
    VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS: bool = False
201
    VLLM_CUSTOM_SCOPES_FOR_PROFILING: bool = False
202
    VLLM_NVTX_SCOPES_FOR_PROFILING: bool = False
203
    VLLM_KV_EVENTS_USE_INT_BLOCK_HASHES: bool = True
204
    VLLM_OBJECT_STORAGE_SHM_BUFFER_NAME: str = "VLLM_OBJECT_STORAGE_SHM_BUFFER"
205
206
    VLLM_DEEPEP_BUFFER_SIZE_MB: int = 1024
    VLLM_DBO_COMM_SMS: int = 20
207
    GPT_OSS_SYSTEM_TOOL_MCP_LABELS: list[str] = []
208
209
    VLLM_PATTERN_MATCH_DEBUG: str | None = None
    VLLM_DEBUG_DUMP_PATH: str | None = None
210
211
    VLLM_ENABLE_INDUCTOR_MAX_AUTOTUNE: bool = True
    VLLM_ENABLE_INDUCTOR_COORDINATE_DESCENT_TUNING: bool = True
212
    VLLM_USE_NCCL_SYMM_MEM: bool = False
213
    VLLM_NCCL_INCLUDE_PATH: str | None = None
214
    VLLM_USE_FBGEMM: bool = False
215
    VLLM_GC_DEBUG: str = ""
216

217
218
219
220
221
222
223
224
225
226
227
228
229
230
231

def get_default_cache_root():
    return os.getenv(
        "XDG_CACHE_HOME",
        os.path.join(os.path.expanduser("~"), ".cache"),
    )


def get_default_config_root():
    return os.getenv(
        "XDG_CONFIG_HOME",
        os.path.join(os.path.expanduser("~"), ".config"),
    )


232
def maybe_convert_int(value: str | None) -> int | None:
233
234
235
236
237
    if value is None:
        return None
    return int(value)


238
def maybe_convert_bool(value: str | None) -> bool | None:
239
240
241
242
243
    if value is None:
        return None
    return bool(int(value))


244
def use_aot_compile() -> bool:
245
    from vllm.utils.torch_utils import is_torch_equal_or_newer
246
247
248
249
250

    default_value = "1" if is_torch_equal_or_newer("2.10.0.dev") else "0"
    return os.environ.get("VLLM_USE_AOT_COMPILE", default_value) == "1"


251
def env_with_choices(
252
    env_name: str,
253
254
    default: str | None,
    choices: list[str] | Callable[[], list[str]],
255
    case_sensitive: bool = True,
256
) -> Callable[[], str | None]:
257
258
    """
    Create a lambda that validates environment variable against allowed choices
259

260
261
262
263
264
    Args:
        env_name: Name of the environment variable
        default: Default value if not set (can be None)
        choices: List of valid string options or callable that returns list
        case_sensitive: Whether validation should be case sensitive
265

266
267
268
269
    Returns:
        Lambda function for environment_variables dict
    """

270
    def _get_validated_env() -> str | None:
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
        value = os.getenv(env_name)
        if value is None:
            return default

        # Resolve choices if it's a callable (for lazy loading)
        actual_choices = choices() if callable(choices) else choices

        if not case_sensitive:
            check_value = value.lower()
            check_choices = [choice.lower() for choice in actual_choices]
        else:
            check_value = value
            check_choices = actual_choices

        if check_value not in check_choices:
286
287
288
289
            raise ValueError(
                f"Invalid value '{value}' for {env_name}. "
                f"Valid options: {actual_choices}."
            )
290
291
292
293
294
295

        return value

    return _get_validated_env


296
def env_list_with_choices(
297
298
    env_name: str,
    default: list[str],
299
    choices: list[str] | Callable[[], list[str]],
300
301
    case_sensitive: bool = True,
) -> Callable[[], list[str]]:
302
    """
303
    Create a lambda that validates environment variable
304
    containing comma-separated values against allowed choices
305

306
307
308
309
310
    Args:
        env_name: Name of the environment variable
        default: Default list of values if not set
        choices: List of valid string options or callable that returns list
        case_sensitive: Whether validation should be case sensitive
311

312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
    Returns:
        Lambda function for environment_variables
        dict that returns list of strings
    """

    def _get_validated_env_list() -> list[str]:
        value = os.getenv(env_name)
        if value is None:
            return default

        # Split comma-separated values and strip whitespace
        values = [v.strip() for v in value.split(",") if v.strip()]

        if not values:
            return default

        # Resolve choices if it's a callable (for lazy loading)
        actual_choices = choices() if callable(choices) else choices

        # Validate each value
        for val in values:
            if not case_sensitive:
                check_value = val.lower()
                check_choices = [choice.lower() for choice in actual_choices]
            else:
                check_value = val
                check_choices = actual_choices

            if check_value not in check_choices:
341
342
343
344
                raise ValueError(
                    f"Invalid value '{val}' in {env_name}. "
                    f"Valid options: {actual_choices}."
                )
345
346
347
348
349
350

        return values

    return _get_validated_env_list


351
def get_vllm_port() -> int | None:
352
    """Get the port from VLLM_PORT environment variable.
353

354
355
    Returns:
        The port number as an integer if VLLM_PORT is set, None otherwise.
356

357
358
359
    Raises:
        ValueError: If VLLM_PORT is a URI, suggest k8s service discovery issue.
    """
360
    if "VLLM_PORT" not in os.environ:
361
362
        return None

363
    port = os.getenv("VLLM_PORT", "0")
364
365
366
367
368

    try:
        return int(port)
    except ValueError as err:
        from urllib.parse import urlparse
369

370
371
372
373
374
375
376
        parsed = urlparse(port)
        if parsed.scheme:
            raise ValueError(
                f"VLLM_PORT '{port}' appears to be a URI. "
                "This may be caused by a Kubernetes service discovery issue,"
                "check the warning in: https://docs.vllm.ai/en/stable/serving/env_vars.html"
            ) from None
377
        raise ValueError(f"VLLM_PORT '{port}' must be a valid integer") from err
378
379


380
381
382
# The begin-* and end* here are used by the documentation generator
# to extract the used env vars.

383
# --8<-- [start:env-vars-definition]
384

385
environment_variables: dict[str, Callable[[], Any]] = {
386
    # ================== Installation Time Env Vars ==================
387
    # Target device of vLLM, supporting [cuda (by default),
388
    # rocm, cpu]
389
    "VLLM_TARGET_DEVICE": lambda: os.getenv("VLLM_TARGET_DEVICE", "cuda").lower(),
390
391
    # Main CUDA version of vLLM, supporting [12.6, 12.8, 12.9],
    # 12.8 is the default. This follows PyTorch but can be overridden.
392
393
    "VLLM_MAIN_CUDA_VERSION": lambda: os.getenv("VLLM_MAIN_CUDA_VERSION", "").lower()
    or "12.8",
394
395
    # Maximum number of compilation jobs to run in parallel.
    # By default this is the number of CPUs
396
    "MAX_JOBS": lambda: os.getenv("MAX_JOBS", None),
397
398
399
    # Number of threads to use for nvcc
    # By default this is 1.
    # If set, `MAX_JOBS` will be reduced to avoid oversubscribing the CPU.
400
    "NVCC_THREADS": lambda: os.getenv("NVCC_THREADS", None),
401
    # If set, vllm will use precompiled binaries (*.so)
402
403
404
405
406
    "VLLM_USE_PRECOMPILED": lambda: os.environ.get("VLLM_USE_PRECOMPILED", "")
    .strip()
    .lower()
    in ("1", "true")
    or bool(os.environ.get("VLLM_PRECOMPILED_WHEEL_LOCATION")),
407
408
    # Used to mark that setup.py is running in a Docker build context,
    # in order to force the use of precompiled binaries.
409
410
411
412
    "VLLM_DOCKER_BUILD_CONTEXT": lambda: os.environ.get("VLLM_DOCKER_BUILD_CONTEXT", "")
    .strip()
    .lower()
    in ("1", "true"),
413
414
    # Whether to force using nightly wheel in python build.
    # This is used for testing the nightly wheel in python build.
415
416
417
    "VLLM_TEST_USE_PRECOMPILED_NIGHTLY_WHEEL": lambda: bool(
        int(os.getenv("VLLM_TEST_USE_PRECOMPILED_NIGHTLY_WHEEL", "0"))
    ),
418
419
420
    # CMake build type
    # If not set, defaults to "Debug" or "RelWithDebInfo"
    # Available options: "Debug", "Release", "RelWithDebInfo"
421
422
423
    "CMAKE_BUILD_TYPE": env_with_choices(
        "CMAKE_BUILD_TYPE", None, ["Debug", "Release", "RelWithDebInfo"]
    ),
424
    # If set, vllm will print verbose logs during installation
425
    "VERBOSE": lambda: bool(int(os.getenv("VERBOSE", "0"))),
426
    # Root directory for vLLM configuration files
427
    # Defaults to `~/.config/vllm` unless `XDG_CONFIG_HOME` is set
428
429
430
    # 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**.
431
    "VLLM_CONFIG_ROOT": lambda: os.path.expanduser(
432
433
434
        os.getenv(
            "VLLM_CONFIG_ROOT",
            os.path.join(get_default_config_root(), "vllm"),
435
436
        )
    ),
437
    # ================== Runtime Env Vars ==================
438
    # Root directory for vLLM cache files
439
    # Defaults to `~/.cache/vllm` unless `XDG_CACHE_HOME` is set
440
    "VLLM_CACHE_ROOT": lambda: os.path.expanduser(
441
442
443
        os.getenv(
            "VLLM_CACHE_ROOT",
            os.path.join(get_default_cache_root(), "vllm"),
444
445
        )
    ),
446
447
448
449
    # 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.
450
    "VLLM_HOST_IP": lambda: os.getenv("VLLM_HOST_IP", ""),
451
    # used in distributed environment to manually set the communication port
452
453
454
    # 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.
455
    "VLLM_PORT": get_vllm_port,
456
457
    # path used for ipc when the frontend api server is running in
    # multi-processing mode to communicate with the backend engine process.
458
459
460
    "VLLM_RPC_BASE_PATH": lambda: os.getenv(
        "VLLM_RPC_BASE_PATH", tempfile.gettempdir()
    ),
461
462
    # If true, will load models from ModelScope instead of Hugging Face Hub.
    # note that the value is true or false, not numbers
463
464
465
466
    "VLLM_USE_MODELSCOPE": lambda: os.environ.get(
        "VLLM_USE_MODELSCOPE", "False"
    ).lower()
    == "true",
467
    # Interval in seconds to log a warning message when the ring buffer is full
468
469
470
    "VLLM_RINGBUFFER_WARNING_INTERVAL": lambda: int(
        os.environ.get("VLLM_RINGBUFFER_WARNING_INTERVAL", "60")
    ),
471
472
    # path to cudatoolkit home directory, under which should be bin, include,
    # and lib directories.
473
    "CUDA_HOME": lambda: os.environ.get("CUDA_HOME", None),
474
475
    # 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
476
    "VLLM_NCCL_SO_PATH": lambda: os.environ.get("VLLM_NCCL_SO_PATH", None),
477
478
    # 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`
479
    "LD_LIBRARY_PATH": lambda: os.environ.get("LD_LIBRARY_PATH", None),
480
    # flag to control if vllm should use triton flash attention
481
482
483
    "VLLM_USE_TRITON_FLASH_ATTN": lambda: (
        os.environ.get("VLLM_USE_TRITON_FLASH_ATTN", "True").lower() in ("true", "1")
    ),
484
485
    # Use separate prefill and decode kernels for V1 attention instead of
    # the unified triton kernel.
486
487
488
489
    "VLLM_V1_USE_PREFILL_DECODE_ATTENTION": lambda: (
        os.getenv("VLLM_V1_USE_PREFILL_DECODE_ATTENTION", "False").lower()
        in ("true", "1")
    ),
490
491
    # Force vllm to use a specific flash-attention version (2 or 3), only valid
    # when using the flash-attention backend.
492
493
494
    "VLLM_FLASH_ATTN_VERSION": lambda: maybe_convert_int(
        os.environ.get("VLLM_FLASH_ATTN_VERSION", None)
    ),
495
496
    # Feature flag to enable/disable Inductor standalone compile.
    # In torch <= 2.7 we ignore this flag; in torch >= 2.8 this is
497
    # disabled by default.
498
499
500
501
    "VLLM_USE_STANDALONE_COMPILE": lambda: os.environ.get(
        "VLLM_USE_STANDALONE_COMPILE", "0"
    )
    == "1",
502
503
    # Debug pattern matching inside custom passes.
    # Should be set to the fx.Node name (e.g. 'getitem_34' or 'scaled_mm_3').
504
505
506
    "VLLM_PATTERN_MATCH_DEBUG": lambda: os.environ.get(
        "VLLM_PATTERN_MATCH_DEBUG", None
    ),
507
508
    # Dump fx graphs to the given directory.
    # It will override CompilationConfig.debug_dump_path if set.
509
    "VLLM_DEBUG_DUMP_PATH": lambda: os.environ.get("VLLM_DEBUG_DUMP_PATH", None),
510
511
512
513
514
515
516
517
    # 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,
    # 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",
518
519
    # local rank of the process in the distributed setting, used to determine
    # the GPU device id
520
    "LOCAL_RANK": lambda: int(os.environ.get("LOCAL_RANK", "0")),
521
    # used to control the visible devices in the distributed setting
522
    "CUDA_VISIBLE_DEVICES": lambda: os.environ.get("CUDA_VISIBLE_DEVICES", None),
523
    # timeout for each iteration in the engine
524
525
526
    "VLLM_ENGINE_ITERATION_TIMEOUT_S": lambda: int(
        os.environ.get("VLLM_ENGINE_ITERATION_TIMEOUT_S", "60")
    ),
527
    # API key for vLLM API server
528
    "VLLM_API_KEY": lambda: os.environ.get("VLLM_API_KEY", None),
529
    # Whether to log responses from API Server for debugging
530
531
532
533
    "VLLM_DEBUG_LOG_API_SERVER_RESPONSE": lambda: os.environ.get(
        "VLLM_DEBUG_LOG_API_SERVER_RESPONSE", "False"
    ).lower()
    == "true",
534
    # S3 access information, used for tensorizer to load model from S3
535
536
537
    "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),
538
    # Usage stats collection
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
    "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_DISABLE_FLASHINFER_PREFILL": lambda: os.environ.get(
        "VLLM_DISABLE_FLASHINFER_PREFILL", "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"),
554
555
556
557
    # 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
558
559
    "VLLM_CONFIGURE_LOGGING": lambda: int(os.getenv("VLLM_CONFIGURE_LOGGING", "1")),
    "VLLM_LOGGING_CONFIG_PATH": lambda: os.getenv("VLLM_LOGGING_CONFIG_PATH"),
560
    # this is used for configuring the default logging level
561
    "VLLM_LOGGING_LEVEL": lambda: os.getenv("VLLM_LOGGING_LEVEL", "INFO").upper(),
562
    # this is used for configuring the default logging stream
563
    "VLLM_LOGGING_STREAM": lambda: os.getenv("VLLM_LOGGING_STREAM", "ext://sys.stdout"),
564
    # if set, VLLM_LOGGING_PREFIX will be prepended to all log messages
565
    "VLLM_LOGGING_PREFIX": lambda: os.getenv("VLLM_LOGGING_PREFIX", ""),
566
567
    # If set, vllm will log stats at this interval in seconds
    # If not set, vllm will log stats every 10 seconds.
568
569
570
    "VLLM_LOG_STATS_INTERVAL": lambda: val
    if (val := float(os.getenv("VLLM_LOG_STATS_INTERVAL", "10."))) > 0.0
    else 10.0,
571
572
573
    # Trace function calls
    # If set to 1, vllm will trace function calls
    # Useful for debugging
574
    "VLLM_TRACE_FUNCTION": lambda: int(os.getenv("VLLM_TRACE_FUNCTION", "0")),
575
    # Backend for attention computation
576
    # Example options:
577
578
579
    # - "TORCH_SDPA": use torch.nn.MultiheadAttention
    # - "FLASH_ATTN": use FlashAttention
    # - "XFORMERS": use XFormers
580
    # - "FLASHINFER": use flashinfer
581
    # - "FLASHMLA": use FlashMLA
582
    # - "FLASH_ATTN_MLA": use FlashAttention for MLA
583
584
    # - "FLASHINFER_MLA": use FlashInfer for MLA
    # - "CUTLASS_MLA": use CUTLASS for MLA
585
    # All possible options loaded dynamically from _Backend enum
586
587
588
589
590
591
592
593
594
    "VLLM_ATTENTION_BACKEND": env_with_choices(
        "VLLM_ATTENTION_BACKEND",
        None,
        lambda: list(
            __import__(
                "vllm.attention.backends.registry", fromlist=["_Backend"]
            )._Backend.__members__.keys()
        ),
    ),
595
    # If set, vllm will use flashinfer sampler
596
597
598
599
600
    "VLLM_USE_FLASHINFER_SAMPLER": lambda: bool(
        int(os.environ["VLLM_USE_FLASHINFER_SAMPLER"])
    )
    if "VLLM_USE_FLASHINFER_SAMPLER" in os.environ
    else None,
601
    # Pipeline stage partition strategy
602
    "VLLM_PP_LAYER_PARTITION": lambda: os.getenv("VLLM_PP_LAYER_PARTITION", None),
603
    # (CPU backend only) CPU key-value cache space.
604
    # default is None and will be set as 4 GB
605
606
607
    "VLLM_CPU_KVCACHE_SPACE": lambda: int(os.getenv("VLLM_CPU_KVCACHE_SPACE", "0"))
    if "VLLM_CPU_KVCACHE_SPACE" in os.environ
    else None,
608
609
    # (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 '|'.
610
    "VLLM_CPU_OMP_THREADS_BIND": lambda: os.getenv("VLLM_CPU_OMP_THREADS_BIND", "auto"),
611
612
    # (CPU backend only) CPU cores not used by OMP threads .
    # Those CPU cores will not be used by OMP threads of a rank.
613
614
615
616
617
    "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,
618
619
620
    # (CPU backend only) whether to use prepack for MoE layer. This will be
    # passed to ipex.llm.modules.GatedMLPMOE. On unsupported CPUs, you might
    # need to set this to "0" (False).
621
    "VLLM_CPU_MOE_PREPACK": lambda: bool(int(os.getenv("VLLM_CPU_MOE_PREPACK", "1"))),
622
    # (CPU backend only) whether to use SGL kernels, optimized for small batch.
623
    "VLLM_CPU_SGL_KERNEL": lambda: bool(int(os.getenv("VLLM_CPU_SGL_KERNEL", "0"))),
624
625
626
627
628
629
630
    # 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
631
632
633
    "VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE": env_with_choices(
        "VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE", "auto", ["auto", "nccl", "shm"]
    ),
634
    # If the env var is set, it enables GPU communication overlap
635
    # (experimental feature) in Ray's Compiled Graph.
636
637
638
    "VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM": lambda: bool(
        int(os.getenv("VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM", "0"))
    ),
639
640
641
    # 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.
642
643
644
    "VLLM_USE_RAY_WRAPPED_PP_COMM": lambda: bool(
        int(os.getenv("VLLM_USE_RAY_WRAPPED_PP_COMM", "1"))
    ),
645
646
    # Use dedicated multiprocess context for workers.
    # Both spawn and fork work
647
648
649
    "VLLM_WORKER_MULTIPROC_METHOD": env_with_choices(
        "VLLM_WORKER_MULTIPROC_METHOD", "fork", ["spawn", "fork"]
    ),
650
    # Path to the cache for storing downloaded assets
651
    "VLLM_ASSETS_CACHE": lambda: os.path.expanduser(
652
653
654
        os.getenv(
            "VLLM_ASSETS_CACHE",
            os.path.join(get_default_cache_root(), "vllm", "assets"),
655
656
        )
    ),
657
658
    # If the env var is set, we will clean model file in
    # this path $VLLM_ASSETS_CACHE/model_streamer/$model_name
659
660
661
    "VLLM_ASSETS_CACHE_MODEL_CLEAN": lambda: bool(
        int(os.getenv("VLLM_ASSETS_CACHE_MODEL_CLEAN", "0"))
    ),
662
663
    # Timeout for fetching images when serving multimodal models
    # Default is 5 seconds
664
    "VLLM_IMAGE_FETCH_TIMEOUT": lambda: int(os.getenv("VLLM_IMAGE_FETCH_TIMEOUT", "5")),
665
    # Timeout for fetching videos when serving multimodal models
666
    # Default is 30 seconds
667
668
669
    "VLLM_VIDEO_FETCH_TIMEOUT": lambda: int(
        os.getenv("VLLM_VIDEO_FETCH_TIMEOUT", "30")
    ),
670
    # Timeout for fetching audio when serving multimodal models
671
    # Default is 10 seconds
672
673
674
    "VLLM_AUDIO_FETCH_TIMEOUT": lambda: int(
        os.getenv("VLLM_AUDIO_FETCH_TIMEOUT", "10")
    ),
675
676
    # Whether to allow HTTP redirects when fetching from media URLs.
    # Default to True
677
678
679
    "VLLM_MEDIA_URL_ALLOW_REDIRECTS": lambda: bool(
        int(os.getenv("VLLM_MEDIA_URL_ALLOW_REDIRECTS", "1"))
    ),
680
681
682
    # Max number of workers for the thread pool handling
    # media bytes loading. Set to 1 to disable parallel processing.
    # Default is 8
683
684
685
    "VLLM_MEDIA_LOADING_THREAD_COUNT": lambda: int(
        os.getenv("VLLM_MEDIA_LOADING_THREAD_COUNT", "8")
    ),
686
687
688
    # 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
689
690
691
    "VLLM_MAX_AUDIO_CLIP_FILESIZE_MB": lambda: int(
        os.getenv("VLLM_MAX_AUDIO_CLIP_FILESIZE_MB", "25")
    ),
692
693
694
695
696
697
698
    # Backend for Video IO
    # - "opencv": Default backend that uses OpenCV stream buffered backend.
    #
    # 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.
699
700
701
    "VLLM_VIDEO_LOADER_BACKEND": lambda: os.getenv(
        "VLLM_VIDEO_LOADER_BACKEND", "opencv"
    ),
702
    # [DEPRECATED] Cache size (in GiB per process) for multimodal input cache
703
    # Default is 4 GiB per API process + 4 GiB per engine core process
704
    "VLLM_MM_INPUT_CACHE_GIB": lambda: int(os.getenv("VLLM_MM_INPUT_CACHE_GIB", "4")),
705
706
    # Path to the XLA persistent cache directory.
    # Only used for XLA devices such as TPUs.
707
    "VLLM_XLA_CACHE_PATH": lambda: os.path.expanduser(
708
        os.getenv(
709
            "VLLM_XLA_CACHE_PATH",
710
            os.path.join(get_default_cache_root(), "vllm", "xla_cache"),
711
712
        )
    ),
713
    # If set, assert on XLA recompilation after each execution step.
714
715
716
    "VLLM_XLA_CHECK_RECOMPILATION": lambda: bool(
        int(os.getenv("VLLM_XLA_CHECK_RECOMPILATION", "0"))
    ),
717
    # Enable SPMD mode for TPU backend.
718
719
720
721
    "VLLM_XLA_USE_SPMD": lambda: bool(int(os.getenv("VLLM_XLA_USE_SPMD", "0"))),
    "VLLM_FUSED_MOE_CHUNK_SIZE": lambda: int(
        os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", "32768")
    ),
722
723
724
    # Control whether to use fused MoE activation chunking. Current chunking
    # logic is incompatible with torch.compile and causes IMA. See issue
    # https://github.com/vllm-project/vllm/issues/19631.
725
726
727
    "VLLM_ENABLE_FUSED_MOE_ACTIVATION_CHUNKING": lambda: bool(
        int(os.getenv("VLLM_ENABLE_FUSED_MOE_ACTIVATION_CHUNKING", "1"))
    ),
728
729
    # If set, the OpenAI API server will stay alive even after the underlying
    # AsyncLLMEngine errors and stops serving requests
730
731
732
    "VLLM_KEEP_ALIVE_ON_ENGINE_DEATH": lambda: bool(
        os.getenv("VLLM_KEEP_ALIVE_ON_ENGINE_DEATH", 0)
    ),
733
734
735
736
    # 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.
737
738
739
740
    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": lambda: (
        os.environ.get("VLLM_ALLOW_LONG_MAX_MODEL_LEN", "0").strip().lower()
        in ("1", "true")
    ),
741
742
    # If set, forces FP8 Marlin to be used for FP8 quantization regardless
    # of the hardware support for FP8 compute.
743
744
745
746
747
748
749
    "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"
    ),
750
751
    # Time in ms for the zmq client to wait for a response from the backend
    # server for simple data operations
752
    "VLLM_RPC_TIMEOUT": lambda: int(os.getenv("VLLM_RPC_TIMEOUT", "10000")),
753
    # Timeout in seconds for keeping HTTP connections alive in API server
754
755
756
    "VLLM_HTTP_TIMEOUT_KEEP_ALIVE": lambda: int(
        os.environ.get("VLLM_HTTP_TIMEOUT_KEEP_ALIVE", "5")
    ),
757
758
759
    # 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
760
761
762
    "VLLM_PLUGINS": lambda: None
    if "VLLM_PLUGINS" not in os.environ
    else os.environ["VLLM_PLUGINS"].split(","),
763
764
765
    # a local directory to look in for unrecognized LoRA adapters.
    # only works if plugins are enabled and
    # VLLM_ALLOW_RUNTIME_LORA_UPDATING is enabled.
766
767
768
    "VLLM_LORA_RESOLVER_CACHE_DIR": lambda: os.getenv(
        "VLLM_LORA_RESOLVER_CACHE_DIR", None
    ),
769
770
771
772
    # Enables torch profiler if set.
    # Both AsyncLLM's CPU traces as well as workers'
    # traces (CPU & GPU) will be saved under this directory.
    # Note that it must be an absolute path.
773
774
775
776
777
778
779
    "VLLM_TORCH_PROFILER_DIR": lambda: (
        None
        if os.getenv("VLLM_TORCH_PROFILER_DIR", None) is None
        else os.path.abspath(
            os.path.expanduser(os.getenv("VLLM_TORCH_PROFILER_DIR", "."))
        )
    ),
780
781
782
    # Enable torch profiler to record shapes if set
    # VLLM_TORCH_PROFILER_RECORD_SHAPES=1. If not set, torch profiler will
    # not record shapes.
783
784
785
    "VLLM_TORCH_PROFILER_RECORD_SHAPES": lambda: bool(
        os.getenv("VLLM_TORCH_PROFILER_RECORD_SHAPES", "0") != "0"
    ),
786
787
788
    # Enable torch profiler to profile memory if set
    # VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY=1. If not set, torch profiler
    # will not profile memory.
789
790
791
    "VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY": lambda: bool(
        os.getenv("VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY", "0") != "0"
    ),
792
793
794
    # Enable torch profiler to profile stack if set
    # VLLM_TORCH_PROFILER_WITH_STACK=1. If not set, torch profiler WILL
    # profile stack by default.
795
796
797
    "VLLM_TORCH_PROFILER_WITH_STACK": lambda: bool(
        os.getenv("VLLM_TORCH_PROFILER_WITH_STACK", "1") != "0"
    ),
798
799
800
    # Enable torch profiler to profile flops if set
    # VLLM_TORCH_PROFILER_WITH_FLOPS=1. If not set, torch profiler will
    # not profile flops.
801
802
803
    "VLLM_TORCH_PROFILER_WITH_FLOPS": lambda: bool(
        os.getenv("VLLM_TORCH_PROFILER_WITH_FLOPS", "0") != "0"
    ),
804
    # If set, vLLM will use Triton implementations of AWQ.
805
    "VLLM_USE_TRITON_AWQ": lambda: bool(int(os.getenv("VLLM_USE_TRITON_AWQ", "0"))),
806
    # If set, allow loading or unloading lora adapters in runtime,
807
808
809
810
    "VLLM_ALLOW_RUNTIME_LORA_UPDATING": lambda: (
        os.environ.get("VLLM_ALLOW_RUNTIME_LORA_UPDATING", "0").strip().lower()
        in ("1", "true")
    ),
811
812
813
814
815
816
    # 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
817
    "VLLM_SKIP_P2P_CHECK": lambda: os.getenv("VLLM_SKIP_P2P_CHECK", "1") == "1",
818
819
820
821
    # List of quantization kernels that should be disabled, used for testing
    # and performance comparisons. Currently only affects MPLinearKernel
    # selection
    # (kernels: MacheteLinearKernel, MarlinLinearKernel, ExllamaLinearKernel)
822
823
824
    "VLLM_DISABLED_KERNELS": lambda: []
    if "VLLM_DISABLED_KERNELS" not in os.environ
    else os.environ["VLLM_DISABLED_KERNELS"].split(","),
825
    # Disable pynccl (using torch.distributed instead)
826
827
828
    "VLLM_DISABLE_PYNCCL": lambda: (
        os.getenv("VLLM_DISABLE_PYNCCL", "False").lower() in ("true", "1")
    ),
829
    # If set, use the V1 code path.
830
    "VLLM_USE_V1": lambda: bool(int(os.getenv("VLLM_USE_V1", "1"))),
831
832
    # Disable aiter ops unless specifically enabled.
    # Acts as a parent switch to enable the rest of the other operations.
833
834
835
    "VLLM_ROCM_USE_AITER": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER", "False").lower() in ("true", "1")
    ),
836
837
    # Whether to use aiter paged attention.
    # By default is disabled.
838
839
840
    "VLLM_ROCM_USE_AITER_PAGED_ATTN": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_PAGED_ATTN", "False").lower() in ("true", "1")
    ),
841
842
843
    # use aiter linear op if aiter ops are enabled
    # The following list of related ops
    # - scaled_mm (per-tensor / rowwise)
844
845
846
    "VLLM_ROCM_USE_AITER_LINEAR": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_LINEAR", "True").lower() in ("true", "1")
    ),
847
848
    # Whether to use aiter moe ops.
    # By default is enabled.
849
850
851
    "VLLM_ROCM_USE_AITER_MOE": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_MOE", "True").lower() in ("true", "1")
    ),
852
    # use aiter rms norm op if aiter ops are enabled.
853
854
855
    "VLLM_ROCM_USE_AITER_RMSNORM": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_RMSNORM", "True").lower() in ("true", "1")
    ),
856
857
    # Whether to use aiter mla ops.
    # By default is enabled.
858
859
860
    "VLLM_ROCM_USE_AITER_MLA": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_MLA", "True").lower() in ("true", "1")
    ),
861
862
    # Whether to use aiter mha ops.
    # By default is enabled.
863
864
865
    "VLLM_ROCM_USE_AITER_MHA": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_MHA", "True").lower() in ("true", "1")
    ),
866
867
    # Whether to use aiter fp4 gemm asm.
    # By default is disabled.
868
869
870
    "VLLM_ROCM_USE_AITER_FP4_ASM_GEMM": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_FP4_ASM_GEMM", "False").lower() in ("true", "1")
    ),
871
872
    # Whether to use aiter rope.
    # By default is disabled.
873
874
875
    "VLLM_ROCM_USE_TRITON_ROPE": lambda: (
        os.getenv("VLLM_ROCM_USE_TRITON_ROPE", "False").lower() in ("true", "1")
    ),
876
877
    # Whether to use aiter triton fp8 bmm kernel
    # By default is enabled.
878
879
880
    "VLLM_ROCM_USE_AITER_FP8BMM": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_FP8BMM", "True").lower() in ("true", "1")
    ),
881
882
883
884
885
    # 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")
    ),
886
887
888
889
890
891
    # Whether to use aiter fusion shared experts ops.
    # By default is enabled.
    "VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS", "True").lower()
        in ("true", "1")
    ),
892
    # use rocm skinny gemms
893
894
895
    "VLLM_ROCM_USE_SKINNY_GEMM": lambda: (
        os.getenv("VLLM_ROCM_USE_SKINNY_GEMM", "True").lower() in ("true", "1")
    ),
896
    # Pad the fp8 weights to 256 bytes for ROCm
897
    "VLLM_ROCM_FP8_PADDING": lambda: bool(int(os.getenv("VLLM_ROCM_FP8_PADDING", "1"))),
898
    # Pad the weights for the moe kernel
899
    "VLLM_ROCM_MOE_PADDING": lambda: bool(int(os.getenv("VLLM_ROCM_MOE_PADDING", "1"))),
900
    # custom paged attention kernel for MI3* cards
901
902
903
    "VLLM_ROCM_CUSTOM_PAGED_ATTN": lambda: (
        os.getenv("VLLM_ROCM_CUSTOM_PAGED_ATTN", "True").lower() in ("true", "1")
    ),
904
905
906
    # Custom quick allreduce kernel for MI3* cards
    # Choice of quantization level: FP, INT8, INT6, INT4 or NONE
    # Recommended for large models to get allreduce
907
908
909
910
911
    "VLLM_ROCM_QUICK_REDUCE_QUANTIZATION": env_with_choices(
        "VLLM_ROCM_QUICK_REDUCE_QUANTIZATION",
        "NONE",
        ["FP", "INT8", "INT6", "INT4", "NONE"],
    ),
912
913
914
915
    # 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
916
917
918
919
    "VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16": lambda: (
        os.getenv("VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16", "True").lower()
        in ("true", "1")
    ),
920
921
922
923
924
925
    # 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.
926
927
928
    "VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB": lambda: maybe_convert_int(
        os.environ.get("VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB", None)
    ),
929
    # Divisor for dynamic query scale factor calculation for FP8 KV Cache
930
    "Q_SCALE_CONSTANT": lambda: int(os.getenv("Q_SCALE_CONSTANT", "200")),
931
    # Divisor for dynamic key scale factor calculation for FP8 KV Cache
932
    "K_SCALE_CONSTANT": lambda: int(os.getenv("K_SCALE_CONSTANT", "200")),
933
    # Divisor for dynamic value scale factor calculation for FP8 KV Cache
934
    "V_SCALE_CONSTANT": lambda: int(os.getenv("V_SCALE_CONSTANT", "100")),
935
    # If set, enable multiprocessing in LLM for the V1 code path.
936
937
938
939
940
941
942
943
944
    "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": lambda: bool(
        int(os.getenv("VLLM_DISABLE_COMPILE_CACHE", "0"))
    ),
945
946
947
    # If set, vllm will run in development mode, which will enable
    # some additional endpoints for developing and debugging,
    # e.g. `/reset_prefix_cache`
948
    "VLLM_SERVER_DEV_MODE": lambda: bool(int(os.getenv("VLLM_SERVER_DEV_MODE", "0"))),
949
950
951
952
953
954
955
    # 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.
956
957
958
    "VLLM_V1_OUTPUT_PROC_CHUNK_SIZE": lambda: int(
        os.getenv("VLLM_V1_OUTPUT_PROC_CHUNK_SIZE", "128")
    ),
959
    # If set, vLLM will disable the MLA attention optimizations.
960
    "VLLM_MLA_DISABLE": lambda: bool(int(os.getenv("VLLM_MLA_DISABLE", "0"))),
961
962
    # If set, vLLM will pick up the provided Flash Attention MLA
    # max number splits for cuda graph decode
963
964
965
    "VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH": lambda: int(
        os.getenv("VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH", "32")
    ),
966
967
968
    # 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.
969
970
971
    "VLLM_RAY_PER_WORKER_GPUS": lambda: float(
        os.getenv("VLLM_RAY_PER_WORKER_GPUS", "1.0")
    ),
972
973
974
    # 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"
975
    "VLLM_RAY_BUNDLE_INDICES": lambda: os.getenv("VLLM_RAY_BUNDLE_INDICES", ""),
976
977
    # In some system, find_loaded_library() may not work. So we allow users to
    # specify the path through environment variable VLLM_CUDART_SO_PATH.
978
    "VLLM_CUDART_SO_PATH": lambda: os.getenv("VLLM_CUDART_SO_PATH", None),
979
    # Rank of the process in the data parallel setting
980
    "VLLM_DP_RANK": lambda: int(os.getenv("VLLM_DP_RANK", "0")),
981
982
    # Rank of the process in the data parallel setting.
    # Defaults to VLLM_DP_RANK when not set.
983
984
985
    "VLLM_DP_RANK_LOCAL": lambda: int(
        os.getenv("VLLM_DP_RANK_LOCAL", sys.modules[__name__].VLLM_DP_RANK)
    ),
986
    # World size of the data parallel setting
987
    "VLLM_DP_SIZE": lambda: int(os.getenv("VLLM_DP_SIZE", "1")),
988
    # IP address of the master node in the data parallel setting
989
    "VLLM_DP_MASTER_IP": lambda: os.getenv("VLLM_DP_MASTER_IP", "127.0.0.1"),
990
    # Port of the master node in the data parallel setting
991
    "VLLM_DP_MASTER_PORT": lambda: int(os.getenv("VLLM_DP_MASTER_PORT", "0")),
992
993
994
995
996
    # In the context of executing MoE models with Data-Parallel, Expert-Parallel
    # and Batched All-to-All dispatch/combine kernels, VLLM_MOE_DP_CHUNK_SIZE
    # dictates the quantum of tokens that can be dispatched from a DP
    # rank. All DP ranks process the activations in VLLM_MOE_DP_CHUNK_SIZE
    # units.
997
    "VLLM_MOE_DP_CHUNK_SIZE": lambda: int(os.getenv("VLLM_MOE_DP_CHUNK_SIZE", "256")),
998
    # Randomize inputs during dummy runs when using Data Parallel
999
1000
1001
1002
    "VLLM_RANDOMIZE_DP_DUMMY_INPUTS": lambda: os.environ.get(
        "VLLM_RANDOMIZE_DP_DUMMY_INPUTS", "0"
    )
    == "1",
1003
1004
1005
1006
1007
1008
1009
    # 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;
1010
1011
1012
    # - "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;
1013
1014
1015
1016
    # 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"
    ),
1017
    # Whether to use S3 path for model loading in CI via RunAI Streamer
1018
    "VLLM_CI_USE_S3": lambda: os.environ.get("VLLM_CI_USE_S3", "0") == "1",
1019
    # Use model_redirect to redirect the model name to a local folder.
1020
1021
1022
1023
1024
    # `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
1025
1026
1027
    "VLLM_MODEL_REDIRECT_PATH": lambda: os.environ.get(
        "VLLM_MODEL_REDIRECT_PATH", None
    ),
1028
    # Whether to use atomicAdd reduce in gptq/awq marlin kernel.
1029
1030
1031
1032
    "VLLM_MARLIN_USE_ATOMIC_ADD": lambda: os.environ.get(
        "VLLM_MARLIN_USE_ATOMIC_ADD", "0"
    )
    == "1",
1033
    # Whether to use marlin kernel in mxfp4 quantization method
1034
1035
1036
    "VLLM_MXFP4_USE_MARLIN": lambda: maybe_convert_bool(
        os.environ.get("VLLM_MXFP4_USE_MARLIN", None)
    ),
1037
1038
1039
    # 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.
1040
1041
1042
1043
    "VLLM_V1_USE_OUTLINES_CACHE": lambda: os.environ.get(
        "VLLM_V1_USE_OUTLINES_CACHE", "0"
    )
    == "1",
1044
1045
    # Gap between padding buckets for the forward pass. So we have
    # 8, we will run forward pass with [16, 24, 32, ...].
1046
1047
1048
1049
1050
1051
1052
1053
    "VLLM_TPU_BUCKET_PADDING_GAP": lambda: int(
        os.environ["VLLM_TPU_BUCKET_PADDING_GAP"]
    )
    if "VLLM_TPU_BUCKET_PADDING_GAP" in os.environ
    else 0,
    "VLLM_TPU_MOST_MODEL_LEN": lambda: maybe_convert_int(
        os.environ.get("VLLM_TPU_MOST_MODEL_LEN", None)
    ),
1054
    # Whether using Pathways
1055
1056
1057
    "VLLM_TPU_USING_PATHWAYS": lambda: bool(
        "proxy" in os.getenv("JAX_PLATFORMS", "").lower()
    ),
1058
    # Allow use of DeepGemm kernels for fused moe ops.
1059
    "VLLM_USE_DEEP_GEMM": lambda: bool(int(os.getenv("VLLM_USE_DEEP_GEMM", "1"))),
1060
    # Whether to use E8M0 scaling when DeepGEMM is used on Blackwell GPUs.
1061
1062
1063
    "VLLM_USE_DEEP_GEMM_E8M0": lambda: bool(
        int(os.getenv("VLLM_USE_DEEP_GEMM_E8M0", "1"))
    ),
1064
1065
1066
1067
    # 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.
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
    # 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",
        ],
1083
    ),
1084
    # Whether to use fused grouped_topk used for MoE expert selection.
1085
1086
1087
    "VLLM_USE_FUSED_MOE_GROUPED_TOPK": lambda: bool(
        int(os.getenv("VLLM_USE_FUSED_MOE_GROUPED_TOPK", "1"))
    ),
1088
    # Allow use of FlashInfer MoE kernels for fused moe ops.
1089
1090
1091
    "VLLM_USE_FLASHINFER_MOE_FP16": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_FP16", "0"))
    ),
1092
    # Allow use of FlashInfer MoE kernels for fused moe ops.
1093
1094
1095
    "VLLM_USE_FLASHINFER_MOE_FP8": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_FP8", "0"))
    ),
1096
    # Allow use of FlashInfer CUTLASS kernels for fused moe ops.
1097
1098
1099
    "VLLM_USE_FLASHINFER_MOE_FP4": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_FP4", "0"))
    ),
1100
1101
    # If set to 1, use the FlashInfer
    # MXFP8 (activation) x MXFP4 (weight) MoE backend.
1102
1103
1104
    "VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8", "0"))
    ),
1105
1106
1107
1108
    # If set to 1, use the FlashInfer CUTLASS backend for
    # MXFP8 (activation) x MXFP4 (weight) MoE.
    # This is separate from the TRTLLMGEN path controlled by
    # VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8.
1109
1110
1111
    "VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS", "0"))
    ),
1112
1113
    # If set to 1, use the FlashInfer
    # BF16 (activation) x MXFP4 (weight) MoE backend.
1114
1115
1116
    "VLLM_USE_FLASHINFER_MOE_MXFP4_BF16": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_MXFP4_BF16", "0"))
    ),
1117
1118
1119
    # 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.
1120
    "VLLM_XGRAMMAR_CACHE_MB": lambda: int(os.getenv("VLLM_XGRAMMAR_CACHE_MB", "512")),
1121
1122
1123
1124
1125
1126
1127
    # 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.
1128
1129
1130
    "VLLM_MSGPACK_ZERO_COPY_THRESHOLD": lambda: int(
        os.getenv("VLLM_MSGPACK_ZERO_COPY_THRESHOLD", "256")
    ),
1131
1132
1133
    # 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.
1134
1135
1136
    "VLLM_ALLOW_INSECURE_SERIALIZATION": lambda: bool(
        int(os.getenv("VLLM_ALLOW_INSECURE_SERIALIZATION", "0"))
    ),
Robert Shaw's avatar
Robert Shaw committed
1137
    # IP address used for NIXL handshake between remote agents.
1138
1139
1140
    "VLLM_NIXL_SIDE_CHANNEL_HOST": lambda: os.getenv(
        "VLLM_NIXL_SIDE_CHANNEL_HOST", "localhost"
    ),
Robert Shaw's avatar
Robert Shaw committed
1141
    # Port used for NIXL handshake between remote agents.
1142
1143
1144
    "VLLM_NIXL_SIDE_CHANNEL_PORT": lambda: int(
        os.getenv("VLLM_NIXL_SIDE_CHANNEL_PORT", "5600")
    ),
1145
    # all2all backend for vllm's expert parallel communication
1146
    # Available options:
1147
1148
1149
    # - "naive": naive all2all implementation using broadcasts
    # - "allgather_reducescatter": all2all implementation based on allgather and
    #  reducescatter
1150
    # - "pplx": use pplx kernels
1151
1152
    # - "deepep_high_throughput", use deepep high-throughput kernels
    # - "deepep_low_latency", use deepep low-latency kernels
1153
    # - "flashinfer_all2allv", use flashinfer alltoallv kernels for mnnvl
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
    "VLLM_ALL2ALL_BACKEND": env_with_choices(
        "VLLM_ALL2ALL_BACKEND",
        "allgather_reducescatter",
        [
            "naive",
            "pplx",
            "deepep_high_throughput",
            "deepep_low_latency",
            "allgather_reducescatter",
            "flashinfer_all2allv",
        ],
    ),
1166
1167
    # Flashinfer MoE backend for vLLM's fused Mixture-of-Experts support.
    # Both require compute capability 10.0 or above.
1168
1169
1170
1171
1172
    # Available options:
    # - "throughput":  [default]
    #     Uses CUTLASS kernels optimized for high-throughput batch inference.
    # - "latency":
    #     Uses TensorRT-LLM kernels optimized for low-latency inference.
1173
1174
1175
    "VLLM_FLASHINFER_MOE_BACKEND": env_with_choices(
        "VLLM_FLASHINFER_MOE_BACKEND", "throughput", ["throughput", "latency"]
    ),
1176
1177
1178
1179
    # 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.
1180
1181
1182
    "VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE": lambda: int(
        os.getenv("VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE", "163840")
    ),
1183
1184
1185
1186
    # Specifies the thresholds of the communicated tensor sizes under which
    # vllm should use flashinfer fused allreduce. The variable should be a
    # JSON with the following format:
    #     { <world size>: <max size in mb> }
1187
    # Unspecified world sizes will fall back to
1188
    #     { 2: 64, 4: 1, <everything else>: 0.5 }
1189
1190
1191
    "VLLM_FLASHINFER_ALLREDUCE_FUSION_THRESHOLDS_MB": lambda: json.loads(
        os.getenv("VLLM_FLASHINFER_ALLREDUCE_FUSION_THRESHOLDS_MB", "{}")
    ),
1192
1193
1194
1195
1196
1197
    # MoE routing strategy selector.
    # See `RoutingSimulator.get_available_strategies()` # for available
    # strategies.
    # Cutstom routing strategies can be registered by
    # RoutingSimulator.register_strategy()
    # Note: custom strategies may not produce correct model outputs
1198
1199
1200
    "VLLM_MOE_ROUTING_SIMULATION_STRATEGY": lambda: os.environ.get(
        "VLLM_MOE_ROUTING_SIMULATION_STRATEGY", ""
    ).lower(),
1201
    # Regex timeout for use by the vLLM tool parsing plugins.
1202
1203
1204
    "VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS": lambda: int(
        os.getenv("VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS", "1")
    ),
1205
1206
    # Reduce CPU usage when vLLM is idle. Enabling this will incur small
    # latency penalty when a request eventually comes.
1207
    "VLLM_SLEEP_WHEN_IDLE": lambda: bool(int(os.getenv("VLLM_SLEEP_WHEN_IDLE", "0"))),
1208
1209
1210
    # 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.
1211
1212
1213
    "VLLM_MQ_MAX_CHUNK_BYTES_MB": lambda: int(
        os.getenv("VLLM_MQ_MAX_CHUNK_BYTES_MB", "16")
    ),
1214
1215
    # Timeout in seconds for execute_model RPC calls in multiprocessing
    # executor (only applies when TP > 1).
1216
1217
1218
    "VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS": lambda: int(
        os.getenv("VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS", "300")
    ),
1219
1220
1221
1222
1223
1224
1225
    # 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.
1226
1227
1228
    "VLLM_KV_CACHE_LAYOUT": env_with_choices(
        "VLLM_KV_CACHE_LAYOUT", None, ["NHD", "HND"]
    ),
1229
1230
1231
    # 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.
1232
1233
1234
    "VLLM_COMPUTE_NANS_IN_LOGITS": lambda: bool(
        int(os.getenv("VLLM_COMPUTE_NANS_IN_LOGITS", "0"))
    ),
1235
1236
1237
    # Controls whether or not emulations are used for NVFP4
    # generations on machines < 100 for compressed-tensors
    # models
1238
1239
1240
    "VLLM_USE_NVFP4_CT_EMULATIONS": lambda: bool(
        int(os.getenv("VLLM_USE_NVFP4_CT_EMULATIONS", "0"))
    ),
1241
1242
1243
1244
    # Time (in seconds) after which the KV cache on the producer side is
    # automatically cleared if no READ notification is received from the
    # consumer. This is only applicable when using NixlConnector in a
    # disaggregated decode-prefill setup.
1245
1246
1247
    "VLLM_NIXL_ABORT_REQUEST_TIMEOUT": lambda: int(
        os.getenv("VLLM_NIXL_ABORT_REQUEST_TIMEOUT", "480")
    ),
1248
    # Controls whether or not to use cudnn prefill
1249
1250
1251
    "VLLM_USE_CUDNN_PREFILL": lambda: bool(
        int(os.getenv("VLLM_USE_CUDNN_PREFILL", "0"))
    ),
1252
1253
1254
    # If set to 1/True, use the TRTLLM attention backend in flashinfer.
    # If set to 0/False, use the default attention backend in flashinfer.
    # If not set, auto-detect the attention backend in flashinfer.
1255
1256
1257
1258
1259
    "VLLM_USE_TRTLLM_ATTENTION": lambda: (
        None
        if "VLLM_USE_TRTLLM_ATTENTION" not in os.environ
        else os.environ["VLLM_USE_TRTLLM_ATTENTION"].lower() in ("1", "true")
    ),
1260
    # If set to 1, when we use fp8 kv, we do not quantize Q to fp8
1261
1262
1263
    "VLLM_FLASHINFER_DISABLE_Q_QUANTIZATION": lambda: bool(
        int(os.getenv("VLLM_FLASHINFER_DISABLE_Q_QUANTIZATION", "0"))
    ),
1264
1265
    # If set, it means we pre-downloaded cubin files and flashinfer will
    # read the cubin files directly.
1266
    "VLLM_HAS_FLASHINFER_CUBIN": lambda: os.getenv("VLLM_HAS_FLASHINFER_CUBIN", False),
1267
1268
1269
1270
1271
1272
1273
1274
1275
    # Supported options:
    # - "flashinfer-cudnn": use flashinfer cudnn GEMM backend
    # - "flashinfer-trtllm": use flashinfer trtllm GEMM backend
    # - "flashinfer-cutlass": use flashinfer cutlass GEMM backend
    # - <none>: automatically pick an available backend
    "VLLM_NVFP4_GEMM_BACKEND": env_with_choices(
        "VLLM_NVFP4_GEMM_BACKEND",
        None,
        ["flashinfer-cudnn", "flashinfer-trtllm", "flashinfer-cutlass"],
1276
    ),
1277
1278
1279
    # 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.
1280
1281
1282
    "VLLM_ENABLE_CUDAGRAPH_GC": lambda: bool(
        int(os.getenv("VLLM_ENABLE_CUDAGRAPH_GC", "0"))
    ),
1283
    # Used to force set up loopback IP
1284
    "VLLM_LOOPBACK_IP": lambda: os.getenv("VLLM_LOOPBACK_IP", ""),
1285
1286
1287
    # Used to set the process name prefix for vLLM processes.
    # This is useful for debugging and monitoring purposes.
    # The default value is "VLLM".
1288
    "VLLM_PROCESS_NAME_PREFIX": lambda: os.getenv("VLLM_PROCESS_NAME_PREFIX", "VLLM"),
1289
1290
1291
1292
1293
1294
1295
    # 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.
1296
1297
1298
    "VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE": lambda: bool(
        int(os.getenv("VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE", "0"))
    ),
1299
1300
    # 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
1301
1302
    # messages for those requests in memory. By default, this is disabled (0),
    # and the "store" option is ignored.
1303
1304
1305
1306
1307
    # 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.
1308
1309
1310
    "VLLM_ENABLE_RESPONSES_API_STORE": lambda: bool(
        int(os.getenv("VLLM_ENABLE_RESPONSES_API_STORE", "0"))
    ),
xiao-llm's avatar
xiao-llm committed
1311
    # If set, use the fp8 mfma in rocm paged attention.
1312
1313
1314
    "VLLM_ROCM_FP8_MFMA_PAGE_ATTN": lambda: bool(
        int(os.getenv("VLLM_ROCM_FP8_MFMA_PAGE_ATTN", "0"))
    ),
1315
    # Whether to use pytorch symmetric memory for allreduce
1316
    "VLLM_ALLREDUCE_USE_SYMM_MEM": lambda: bool(
1317
        int(os.getenv("VLLM_ALLREDUCE_USE_SYMM_MEM", "0"))
1318
    ),
1319
    # Allows vllm to find tuned config under customized folder
1320
    "VLLM_TUNED_CONFIG_FOLDER": lambda: os.getenv("VLLM_TUNED_CONFIG_FOLDER", None),
1321
    # Allows harmony instructions to be injected on system messages
1322
1323
1324
    "VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS": lambda: bool(
        int(os.getenv("VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS", "0"))
    ),
1325
    # Add optional custom scopes for profiling, disable to avoid overheads
1326
1327
1328
    "VLLM_CUSTOM_SCOPES_FOR_PROFILING": lambda: bool(
        int(os.getenv("VLLM_CUSTOM_SCOPES_FOR_PROFILING", "0"))
    ),
1329
    # Add optional nvtx scopes for profiling, disable to avoid overheads
1330
1331
1332
    "VLLM_NVTX_SCOPES_FOR_PROFILING": lambda: bool(
        int(os.getenv("VLLM_NVTX_SCOPES_FOR_PROFILING", "0"))
    ),
1333
1334
    # Represent block hashes in KV cache events as 64-bit integers instead of
    # raw bytes. Defaults to True for backward compatibility.
1335
1336
1337
    "VLLM_KV_EVENTS_USE_INT_BLOCK_HASHES": lambda: bool(
        int(os.getenv("VLLM_KV_EVENTS_USE_INT_BLOCK_HASHES", "1"))
    ),
1338
1339
    # Name of the shared memory buffer used for object storage.
    # Only effective when mm_config.mm_processor_cache_type == "shm".
1340
1341
1342
    "VLLM_OBJECT_STORAGE_SHM_BUFFER_NAME": lambda: os.getenv(
        "VLLM_OBJECT_STORAGE_SHM_BUFFER_NAME", "VLLM_OBJECT_STORAGE_SHM_BUFFER"
    ),
1343
    # The size in MB of the buffers (NVL and RDMA) used by DeepEP
1344
1345
1346
    "VLLM_DEEPEP_BUFFER_SIZE_MB": lambda: int(
        os.getenv("VLLM_DEEPEP_BUFFER_SIZE_MB", "1024")
    ),
1347
1348
    # The number of SMs to allocate for communication kernels when running DBO
    # the rest of the SMs on the device will be allocated to compute
1349
    "VLLM_DBO_COMM_SMS": lambda: int(os.getenv("VLLM_DBO_COMM_SMS", "20")),
1350
1351
    # Valid values are container,code_interpreter,web_search_preview
    # ex GPT_OSS_SYSTEM_TOOL_MCP_LABELS=container,code_interpreter
1352
1353
1354
1355
1356
    "GPT_OSS_SYSTEM_TOOL_MCP_LABELS": env_list_with_choices(
        "GPT_OSS_SYSTEM_TOOL_MCP_LABELS",
        [],
        ["container", "code_interpreter", "web_search_preview"],
    ),
1357
1358
1359
    # 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)
1360
1361
1362
    "VLLM_ENABLE_INDUCTOR_MAX_AUTOTUNE": lambda: bool(
        int(os.getenv("VLLM_ENABLE_INDUCTOR_MAX_AUTOTUNE", "1"))
    ),
1363
1364
    # If set to 1, enable coordinate_descent_tuning;
    # By default, this is enabled (1)
1365
1366
1367
    "VLLM_ENABLE_INDUCTOR_COORDINATE_DESCENT_TUNING": lambda: bool(
        int(os.getenv("VLLM_ENABLE_INDUCTOR_COORDINATE_DESCENT_TUNING", "1"))
    ),
1368
    # Flag to enable NCCL symmetric memory allocation and registration
1369
1370
1371
    "VLLM_USE_NCCL_SYMM_MEM": lambda: bool(
        int(os.getenv("VLLM_USE_NCCL_SYMM_MEM", "0"))
    ),
1372
    # NCCL header path
1373
    "VLLM_NCCL_INCLUDE_PATH": lambda: os.environ.get("VLLM_NCCL_INCLUDE_PATH", None),
1374
1375
    # Flag to enable FBGemm kernels on model execution
    "VLLM_USE_FBGEMM": lambda: bool(int(os.getenv("VLLM_USE_FBGEMM", "0"))),
1376
1377
1378
1379
1380
1381
    # 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", ""),
1382
1383
}

1384
# --8<-- [end:env-vars-definition]
1385

1386

1387
def __getattr__(name: str):
1388
1389
1390
1391
1392
1393
    """
    Gets environment variables lazily.

    NOTE: After enable_envs_cache() invocation (which triggered after service
    initialization), all environment variables will be cached.
    """
1394
1395
1396
1397
1398
    if name in environment_variables:
        return environment_variables[name]()
    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")


1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
def enable_envs_cache() -> None:
    """
    Enables caching of environment variables. This is useful for performance
    reasons, as it avoids the need to re-evaluate environment variables on
    every call.

    NOTE: Currently, it's invoked after service initialization to reduce
    runtime overhead. This also means that environment variables should NOT
    be updated after the service is initialized.
    """
    # Tag __getattr__ with functools.cache
    global __getattr__
    __getattr__ = functools.cache(__getattr__)

    # Cache all environment variables
    for key in environment_variables:
        __getattr__(key)


1418
1419
def __dir__():
    return list(environment_variables.keys())
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433


def is_set(name: str):
    """Check if an environment variable is explicitly set."""
    if name in environment_variables:
        return name in os.environ
    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")


def set_vllm_use_v1(use_v1: bool):
    if is_set("VLLM_USE_V1"):
        raise ValueError(
            "Should not call set_vllm_use_v1() if VLLM_USE_V1 is set "
            "explicitly by the user. Please raise this as a Github "
1434
1435
            "Issue and explicitly set VLLM_USE_V1=0 or 1."
        )
1436
    os.environ["VLLM_USE_V1"] = "1" if use_v1 else "0"
1437
1438
1439
1440
1441
1442
1443
1444


def compute_hash() -> str:
    """
    WARNING: Whenever a new key is added to this environment
    variables, ensure that it is included in the factors list if
    it affects the computation graph. For example, different values
    of VLLM_PP_LAYER_PARTITION will generate different computation
1445
    graphs, so it is included in the factors list. The env vars that
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
    affect the choice of different kernels or attention backends should
    also be included in the factors list.
    """

    # The values of envs may affects the computation graph.
    # TODO(DefTruth): hash all environment variables?
    # for key in environment_variables:
    #     factorize(key)
    environment_variables_to_hash = [
        "VLLM_PP_LAYER_PARTITION",
        "VLLM_MLA_DISABLE",
1457
        "VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH",
1458
1459
1460
1461
        "VLLM_USE_TRITON_FLASH_ATTN",
        "VLLM_USE_TRITON_AWQ",
        "VLLM_DP_RANK",
        "VLLM_DP_SIZE",
1462
        "VLLM_USE_STANDALONE_COMPILE",
1463
        "VLLM_FUSED_MOE_CHUNK_SIZE",
1464
1465
1466
1467
1468
1469
        "VLLM_FLASHINFER_MOE_BACKEND",
        "VLLM_V1_USE_PREFILL_DECODE_ATTENTION",
        "VLLM_ATTENTION_BACKEND",
        "VLLM_USE_FLASHINFER_SAMPLER",
        "VLLM_DISABLED_KERNELS",
        "VLLM_USE_DEEP_GEMM",
1470
        "VLLM_USE_DEEP_GEMM_E8M0",
1471
        "VLLM_USE_FUSED_MOE_GROUPED_TOPK",
1472
        "VLLM_USE_FLASHINFER_MOE_FP16",
1473
1474
1475
        "VLLM_USE_FLASHINFER_MOE_FP8",
        "VLLM_USE_FLASHINFER_MOE_FP4",
        "VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8",
1476
        "VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS",
1477
1478
1479
        "VLLM_USE_FLASHINFER_MOE_MXFP4_BF16",
        "VLLM_USE_CUDNN_PREFILL",
        "VLLM_USE_TRTLLM_ATTENTION",
1480
        "VLLM_FLASHINFER_DISABLE_Q_QUANTIZATION",
1481
1482
1483
1484
1485
1486
1487
        "VLLM_ROCM_USE_AITER",
        "VLLM_ROCM_USE_AITER_PAGED_ATTN",
        "VLLM_ROCM_USE_AITER_LINEAR",
        "VLLM_ROCM_USE_AITER_MOE",
        "VLLM_ROCM_USE_AITER_RMSNORM",
        "VLLM_ROCM_USE_AITER_MLA",
        "VLLM_ROCM_USE_AITER_MHA",
1488
1489
        "VLLM_ROCM_USE_AITER_FP4_ASM_GEMM",
        "VLLM_ROCM_USE_TRITON_ROPE",
1490
        "VLLM_ROCM_USE_AITER_FP8BMM",
1491
        "VLLM_ROCM_USE_AITER_UNIFIED_ATTENTION",
1492
1493
1494
1495
1496
1497
1498
        "VLLM_ROCM_USE_SKINNY_GEMM",
        "VLLM_ROCM_FP8_PADDING",
        "VLLM_ROCM_MOE_PADDING",
        "VLLM_ROCM_CUSTOM_PAGED_ATTN",
        "VLLM_ROCM_QUICK_REDUCE_QUANTIZATION",
        "VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16",
        "VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB",
xiao-llm's avatar
xiao-llm committed
1499
        "VLLM_ROCM_FP8_MFMA_PAGE_ATTN",
1500
1501
        "VLLM_ENABLE_INDUCTOR_MAX_AUTOTUNE",
        "VLLM_ENABLE_INDUCTOR_COORDINATE_DESCENT_TUNING",
1502
        "VLLM_NVFP4_GEMM_BACKEND",
1503
        "VLLM_USE_FBGEMM",
1504
1505
    ]
    for key in environment_variables_to_hash:
1506
1507
        # if this goes out of sync with environment_variables,
        # it's not a user error, it's a bug
1508
        assert key in environment_variables, (
1509
            "Please update environment_variables_to_hash in envs.py"
1510
        )
1511

1512
    factors = [environment_variables[key]() for key in environment_variables_to_hash]
1513

1514
    hash_str = hashlib.md5(str(factors).encode(), usedforsecurity=False).hexdigest()
1515
1516

    return hash_str