envs.py 48.6 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import hashlib
5
import os
6
import sys
7
import tempfile
8
from typing import TYPE_CHECKING, Any, Callable, Optional
9
10
11

if TYPE_CHECKING:
    VLLM_HOST_IP: str = ""
12
    VLLM_PORT: Optional[int] = None
13
    VLLM_RPC_BASE_PATH: str = tempfile.gettempdir()
14
    VLLM_USE_MODELSCOPE: bool = False
15
    VLLM_RINGBUFFER_WARNING_INTERVAL: int = 60
16
17
    VLLM_NCCL_SO_PATH: Optional[str] = None
    LD_LIBRARY_PATH: Optional[str] = None
18
    VLLM_USE_TRITON_FLASH_ATTN: bool = True
19
    VLLM_V1_USE_PREFILL_DECODE_ATTENTION: bool = False
20
    VLLM_USE_AITER_UNIFIED_ATTENTION: bool = False
21
    VLLM_FLASH_ATTN_VERSION: Optional[int] = None
22
23
24
25
26
27
28
    LOCAL_RANK: int = 0
    CUDA_VISIBLE_DEVICES: Optional[str] = None
    VLLM_ENGINE_ITERATION_TIMEOUT_S: int = 60
    VLLM_API_KEY: Optional[str] = None
    S3_ACCESS_KEY_ID: Optional[str] = None
    S3_SECRET_ACCESS_KEY: Optional[str] = None
    S3_ENDPOINT_URL: Optional[str] = None
29
    VLLM_MODEL_REDIRECT_PATH: Optional[str] = None
30
31
    VLLM_CACHE_ROOT: str = os.path.expanduser("~/.cache/vllm")
    VLLM_CONFIG_ROOT: str = os.path.expanduser("~/.config/vllm")
32
33
34
35
36
    VLLM_USAGE_STATS_SERVER: str = "https://stats.vllm.ai"
    VLLM_NO_USAGE_STATS: bool = False
    VLLM_DO_NOT_TRACK: bool = False
    VLLM_USAGE_SOURCE: str = ""
    VLLM_CONFIGURE_LOGGING: int = 1
37
    VLLM_LOGGING_LEVEL: str = "INFO"
38
    VLLM_LOGGING_PREFIX: str = ""
39
    VLLM_LOGGING_CONFIG_PATH: Optional[str] = None
40
    VLLM_LOGITS_PROCESSOR_THREADS: Optional[int] = None
41
42
    VLLM_TRACE_FUNCTION: int = 0
    VLLM_ATTENTION_BACKEND: Optional[str] = None
43
    VLLM_USE_FLASHINFER_SAMPLER: Optional[bool] = None
44
    VLLM_FLASHINFER_FORCE_TENSOR_CORES: bool = False
45
    VLLM_PP_LAYER_PARTITION: Optional[str] = None
46
    VLLM_CPU_KVCACHE_SPACE: Optional[int] = 0
47
    VLLM_CPU_OMP_THREADS_BIND: str = ""
48
    VLLM_CPU_NUM_OF_RESERVED_CPU: Optional[int] = None
49
    VLLM_CPU_MOE_PREPACK: bool = True
50
    VLLM_CPU_SGL_KERNEL: bool = False
51
    VLLM_XLA_CACHE_PATH: str = os.path.join(VLLM_CACHE_ROOT, "xla_cache")
52
    VLLM_XLA_CHECK_RECOMPILATION: bool = False
53
    VLLM_FUSED_MOE_CHUNK_SIZE: int = 64 * 1024
54
    VLLM_ENABLE_FUSED_MOE_ACTIVATION_CHUNKING: bool = True
55
    VLLM_USE_RAY_SPMD_WORKER: bool = False
56
    VLLM_USE_RAY_COMPILED_DAG: bool = False
57
    VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE: str = "auto"
58
    VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM: bool = False
59
    VLLM_USE_RAY_WRAPPED_PP_COMM: bool = True
60
    VLLM_XLA_USE_SPMD: bool = False
61
    VLLM_WORKER_MULTIPROC_METHOD: str = "fork"
62
    VLLM_ASSETS_CACHE: str = os.path.join(VLLM_CACHE_ROOT, "assets")
63
    VLLM_IMAGE_FETCH_TIMEOUT: int = 5
64
    VLLM_VIDEO_FETCH_TIMEOUT: int = 30
65
    VLLM_AUDIO_FETCH_TIMEOUT: int = 10
66
    VLLM_MAX_AUDIO_CLIP_FILESIZE_MB: int = 25
67
    VLLM_VIDEO_LOADER_BACKEND: str = "opencv"
68
    VLLM_MM_INPUT_CACHE_GIB: int = 4
69
70
71
72
    VLLM_TARGET_DEVICE: str = "cuda"
    MAX_JOBS: Optional[str] = None
    NVCC_THREADS: Optional[str] = None
    VLLM_USE_PRECOMPILED: bool = False
73
    VLLM_DOCKER_BUILD_CONTEXT: bool = False
74
    VLLM_TEST_USE_PRECOMPILED_NIGHTLY_WHEEL: bool = False
75
    VLLM_KEEP_ALIVE_ON_ENGINE_DEATH: bool = False
76
77
    CMAKE_BUILD_TYPE: Optional[str] = None
    VERBOSE: bool = False
78
    VLLM_ALLOW_LONG_MAX_MODEL_LEN: bool = False
79
    VLLM_RPC_TIMEOUT: int = 10000  # ms
80
    VLLM_HTTP_TIMEOUT_KEEP_ALIVE: int = 5  # seconds
81
    VLLM_PLUGINS: Optional[list[str]] = None
82
    VLLM_LORA_RESOLVER_CACHE_DIR: Optional[str] = None
83
    VLLM_TORCH_PROFILER_DIR: Optional[str] = None
84
85
86
87
    VLLM_TORCH_PROFILER_RECORD_SHAPES: bool = False
    VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY: bool = False
    VLLM_TORCH_PROFILER_WITH_STACK: bool = True
    VLLM_TORCH_PROFILER_WITH_FLOPS: bool = False
88
    VLLM_USE_TRITON_AWQ: bool = False
89
    VLLM_ALLOW_RUNTIME_LORA_UPDATING: bool = False
90
    VLLM_SKIP_P2P_CHECK: bool = False
91
    VLLM_DISABLED_KERNELS: list[str] = []
92
    VLLM_USE_V1: bool = True
93
    VLLM_ROCM_USE_AITER: bool = False
94
    VLLM_ROCM_USE_AITER_PAGED_ATTN: bool = False
95
    VLLM_ROCM_USE_AITER_LINEAR: bool = True
96
    VLLM_ROCM_USE_AITER_MOE: bool = True
97
    VLLM_ROCM_USE_AITER_RMSNORM: bool = True
98
    VLLM_ROCM_USE_AITER_MLA: bool = True
99
    VLLM_ROCM_USE_AITER_MHA: bool = True
100
    VLLM_ROCM_USE_SKINNY_GEMM: bool = True
101
    VLLM_ROCM_FP8_PADDING: bool = True
102
    VLLM_ROCM_MOE_PADDING: bool = True
103
    VLLM_ROCM_CUSTOM_PAGED_ATTN: bool = True
104
    VLLM_ENABLE_V1_MULTIPROCESSING: bool = True
105
    VLLM_LOG_BATCHSIZE_INTERVAL: float = -1
106
    VLLM_DISABLE_COMPILE_CACHE: bool = False
107
    Q_SCALE_CONSTANT: int = 200
108
109
    K_SCALE_CONSTANT: int = 200
    V_SCALE_CONSTANT: int = 100
110
    VLLM_SERVER_DEV_MODE: bool = False
111
    VLLM_V1_OUTPUT_PROC_CHUNK_SIZE: int = 128
112
    VLLM_MLA_DISABLE: bool = False
113
114
    VLLM_RAY_PER_WORKER_GPUS: float = 1.0
    VLLM_RAY_BUNDLE_INDICES: str = ""
115
    VLLM_CUDART_SO_PATH: Optional[str] = None
116
    VLLM_DP_RANK: int = 0
117
    VLLM_DP_RANK_LOCAL: int = -1
118
119
120
    VLLM_DP_SIZE: int = 1
    VLLM_DP_MASTER_IP: str = ""
    VLLM_DP_MASTER_PORT: int = 0
121
    VLLM_MOE_DP_CHUNK_SIZE: int = 256
122
    VLLM_RANDOMIZE_DP_DUMMY_INPUTS: bool = False
123
    VLLM_MARLIN_USE_ATOMIC_ADD: bool = False
124
    VLLM_V0_USE_OUTLINES_CACHE: bool = False
125
    VLLM_V1_USE_OUTLINES_CACHE: bool = False
126
    VLLM_TPU_BUCKET_PADDING_GAP: int = 0
127
    VLLM_TPU_MOST_MODEL_LEN: Optional[int] = None
128
    VLLM_TPU_USING_PATHWAYS: bool = False
129
    VLLM_USE_DEEP_GEMM: bool = False
130
    VLLM_SKIP_DEEP_GEMM_WARMUP: bool = False
131
132
    VLLM_USE_FLASHINFER_MOE_FP8: bool = False
    VLLM_USE_FLASHINFER_MOE_FP4: bool = False
133
    VLLM_FLASHINFER_MOE_BACKEND: str = "throughput"
134
    VLLM_XGRAMMAR_CACHE_MB: int = 0
135
    VLLM_MSGPACK_ZERO_COPY_THRESHOLD: int = 256
136
    VLLM_ALLOW_INSECURE_SERIALIZATION: bool = False
Robert Shaw's avatar
Robert Shaw committed
137
138
    VLLM_NIXL_SIDE_CHANNEL_HOST: str = "localhost"
    VLLM_NIXL_SIDE_CHANNEL_PORT: int = 5557
139
    VLLM_ALL2ALL_BACKEND: str = "naive"
140
    VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE: int = 163840
141
    VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS: int = 1
142
    VLLM_SLEEP_WHEN_IDLE: bool = False
143
    VLLM_MQ_MAX_CHUNK_BYTES_MB: int = 16
144
    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: int = 300
145
    VLLM_KV_CACHE_LAYOUT: Optional[str] = None
146
    VLLM_COMPUTE_NANS_IN_LOGITS: bool = False
147
    VLLM_USE_NVFP4_CT_EMULATIONS: bool = False
148
149
150
    VLLM_ROCM_QUICK_REDUCE_QUANTIZATION: str = "NONE"
    VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16: bool = True
    VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB: Optional[int] = None
151
    VLLM_NIXL_ABORT_REQUEST_TIMEOUT: int = 120
152
    VLLM_USE_CUDNN_PREFILL: bool = False
153
    VLLM_ENABLE_CUDAGRAPH_GC: bool = False
154
    VLLM_LOOPBACK_IP: str = ""
155
    VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE: bool = False
156
    VLLM_ENABLE_RESPONSES_API_STORE: bool = False
157
    VLLM_USE_TRTLLM_ATTENTION: Optional[str] = None
158
159
    VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8: bool = False
    VLLM_USE_FLASHINFER_MOE_MXFP4_BF16: bool = False
160

161
162
163
164
165
166
167
168
169
170
171
172
173
174
175

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"),
    )


176
177
178
179
180
181
def maybe_convert_int(value: Optional[str]) -> Optional[int]:
    if value is None:
        return None
    return int(value)


182
183
def get_vllm_port() -> Optional[int]:
    """Get the port from VLLM_PORT environment variable.
184

185
186
    Returns:
        The port number as an integer if VLLM_PORT is set, None otherwise.
187

188
189
190
191
192
193
194
195
196
197
198
199
    Raises:
        ValueError: If VLLM_PORT is a URI, suggest k8s service discovery issue.
    """
    if 'VLLM_PORT' not in os.environ:
        return None

    port = os.getenv('VLLM_PORT', '0')

    try:
        return int(port)
    except ValueError as err:
        from urllib.parse import urlparse
200
201
202
203
204
205
206
        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
207
208
209
210
        raise ValueError(
            f"VLLM_PORT '{port}' must be a valid integer") from err


211
212
213
# The begin-* and end* here are used by the documentation generator
# to extract the used env vars.

214
# --8<-- [start:env-vars-definition]
215

216
environment_variables: dict[str, Callable[[], Any]] = {
217
218
219

    # ================== Installation Time Env Vars ==================

220
    # Target device of vLLM, supporting [cuda (by default),
221
    # rocm, neuron, cpu]
222
    "VLLM_TARGET_DEVICE":
223
    lambda: os.getenv("VLLM_TARGET_DEVICE", "cuda").lower(),
224
225
226
227
228
229
230
231
232
233
234
235
236
237

    # 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 binaries (*.so)
    "VLLM_USE_PRECOMPILED":
238
239
240
241
242
243
244
245
    lambda: os.environ.get("VLLM_USE_PRECOMPILED", "").strip().lower() in
    ("1", "true") or bool(os.environ.get("VLLM_PRECOMPILED_WHEEL_LOCATION")),

    # 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"),
246

247
248
249
250
251
252
    # Whether to force using nightly wheel in python build.
    # This is used for testing the nightly wheel in python build.
    "VLLM_TEST_USE_PRECOMPILED_NIGHTLY_WHEEL":
    lambda: bool(int(os.getenv("VLLM_TEST_USE_PRECOMPILED_NIGHTLY_WHEEL", "0"))
                 ),

253
254
255
256
257
258
259
260
261
262
    # CMake build type
    # If not set, defaults to "Debug" or "RelWithDebInfo"
    # Available options: "Debug", "Release", "RelWithDebInfo"
    "CMAKE_BUILD_TYPE":
    lambda: os.getenv("CMAKE_BUILD_TYPE"),

    # If set, vllm will print verbose logs during installation
    "VERBOSE":
    lambda: bool(int(os.getenv('VERBOSE', '0'))),

263
    # Root directory for vLLM configuration files
264
    # Defaults to `~/.config/vllm` unless `XDG_CONFIG_HOME` is set
265
266
267
268
    # 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":
269
270
271
272
273
    lambda: os.path.expanduser(
        os.getenv(
            "VLLM_CONFIG_ROOT",
            os.path.join(get_default_config_root(), "vllm"),
        )),
274
275
276

    # ================== Runtime Env Vars ==================

277
    # Root directory for vLLM cache files
278
279
280
281
282
283
284
285
    # 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"),
        )),

286
287
288
289
    # 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.
290
    'VLLM_HOST_IP':
291
    lambda: os.getenv('VLLM_HOST_IP', ""),
292

293
    # used in distributed environment to manually set the communication port
294
295
296
    # 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.
297
    'VLLM_PORT':
298
    get_vllm_port,
299

300
301
302
303
    # 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()),
304

305
306
307
308
309
    # 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",

310
311
312
313
    # 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")),

314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
    # 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 if vllm should use triton flash attention
    "VLLM_USE_TRITON_FLASH_ATTN":
    lambda: (os.environ.get("VLLM_USE_TRITON_FLASH_ATTN", "True").lower() in
             ("true", "1")),

334
335
336
337
338
339
340
    # Use separate prefill and decode kernels for V1 attention instead of
    # the unified triton kernel.
    "VLLM_V1_USE_PREFILL_DECODE_ATTENTION":
    lambda:
    (os.getenv("VLLM_V1_USE_PREFILL_DECODE_ATTENTION", "False").lower() in
     ("true", "1")),

341
342
343
344
345
346
    # Use AITER triton unified attention for V1 attention
    "VLLM_USE_AITER_UNIFIED_ATTENTION":
    lambda:
    (os.getenv("VLLM_USE_AITER_UNIFIED_ATTENTION", "False").lower() in
     ("true", "1")),

347
348
349
350
351
    # Force vllm to use a specific flash-attention version (2 or 3), only valid
    # when using the flash-attention backend.
    "VLLM_FLASH_ATTN_VERSION":
    lambda: maybe_convert_int(os.environ.get("VLLM_FLASH_ATTN_VERSION", None)),

352
353
354
355
    # Internal flag to enable Dynamo fullgraph capture
    "VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE":
    lambda: bool(
        os.environ.get("VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE", "1") != "0"),
356

357
358
359
360
361
    # Feature flag to enable/disable Inductor standalone compile.
    # In torch <= 2.7 we ignore this flag; in torch >= 2.8 this is
    # enabled by default.
    "VLLM_USE_STANDALONE_COMPILE":
    lambda: os.environ.get("VLLM_USE_STANDALONE_COMPILE", "1") == "1",
362

363
364
365
366
367
368
369
370
371
372
373
374
375
    # 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")),

376
    # API key for vLLM API server
377
378
379
    "VLLM_API_KEY":
    lambda: os.environ.get("VLLM_API_KEY", None),

380
381
    # Whether to log responses from API Server for debugging
    "VLLM_DEBUG_LOG_API_SERVER_RESPONSE":
382
383
    lambda: os.environ.get("VLLM_DEBUG_LOG_API_SERVER_RESPONSE", "False"
                           ).lower() == "true",
384

385
386
    # S3 access information, used for tensorizer to load model from S3
    "S3_ACCESS_KEY_ID":
387
    lambda: os.environ.get("S3_ACCESS_KEY_ID", None),
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
    "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: int(os.getenv("VLLM_CONFIGURE_LOGGING", "1")),
    "VLLM_LOGGING_CONFIG_PATH":
    lambda: os.getenv("VLLM_LOGGING_CONFIG_PATH"),

413
414
    # this is used for configuring the default logging level
    "VLLM_LOGGING_LEVEL":
415
    lambda: os.getenv("VLLM_LOGGING_LEVEL", "INFO").upper(),
416

417
418
419
420
    # if set, VLLM_LOGGING_PREFIX will be prepended to all log messages
    "VLLM_LOGGING_PREFIX":
    lambda: os.getenv("VLLM_LOGGING_PREFIX", ""),

421
422
423
424
425
426
427
428
    # if set, vllm will call logits processors in a thread pool with this many
    # threads. This is useful when using custom logits processors that either
    # (a) launch additional CUDA kernels or (b) do significant CPU-bound work
    # while not holding the python GIL, or both.
    "VLLM_LOGITS_PROCESSOR_THREADS":
    lambda: int(os.getenv("VLLM_LOGITS_PROCESSOR_THREADS", "0"))
    if "VLLM_LOGITS_PROCESSOR_THREADS" in os.environ else None,

429
430
431
432
433
434
435
436
437
438
439
440
    # 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")),

    # Backend for attention computation
    # Available options:
    # - "TORCH_SDPA": use torch.nn.MultiheadAttention
    # - "FLASH_ATTN": use FlashAttention
    # - "XFORMERS": use XFormers
    # - "ROCM_FLASH": use ROCmFlashAttention
441
    # - "FLASHINFER": use flashinfer
442
    # - "FLASHMLA": use FlashMLA
443
444
445
    "VLLM_ATTENTION_BACKEND":
    lambda: os.getenv("VLLM_ATTENTION_BACKEND", None),

446
447
    # If set, vllm will use flashinfer sampler
    "VLLM_USE_FLASHINFER_SAMPLER":
448
449
    lambda: bool(int(os.environ["VLLM_USE_FLASHINFER_SAMPLER"]))
    if "VLLM_USE_FLASHINFER_SAMPLER" in os.environ else None,
450

451
452
453
454
455
    # If set, vllm will force flashinfer to use tensor cores;
    # otherwise will use heuristic based on model architecture.
    "VLLM_FLASHINFER_FORCE_TENSOR_CORES":
    lambda: bool(int(os.getenv("VLLM_FLASHINFER_FORCE_TENSOR_CORES", "0"))),

456
457
458
459
    # Pipeline stage partition strategy
    "VLLM_PP_LAYER_PARTITION":
    lambda: os.getenv("VLLM_PP_LAYER_PARTITION", None),

460
    # (CPU backend only) CPU key-value cache space.
461
    # default is None and will be set as 4 GB
462
    "VLLM_CPU_KVCACHE_SPACE":
463
464
    lambda: int(os.getenv("VLLM_CPU_KVCACHE_SPACE", "0"))
    if "VLLM_CPU_KVCACHE_SPACE" in os.environ else None,
465

466
467
468
    # (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":
469
470
471
472
473
    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":
474
475
    lambda: int(os.getenv("VLLM_CPU_NUM_OF_RESERVED_CPU", "0"))
    if "VLLM_CPU_NUM_OF_RESERVED_CPU" in os.environ else None,
476

477
478
479
480
481
482
    # (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).
    "VLLM_CPU_MOE_PREPACK":
    lambda: bool(int(os.getenv("VLLM_CPU_MOE_PREPACK", "1"))),

483
484
485
486
    # (CPU backend only) whether to use SGL kernels, optimized for small batch.
    "VLLM_CPU_SGL_KERNEL":
    lambda: bool(int(os.getenv("VLLM_CPU_SGL_KERNEL", "0"))),

487
488
489
490
491
    # If the env var is set, then all workers will execute as separate
    # processes from the engine, and we use the same mechanism to trigger
    # execution on all workers.
    # Run vLLM with VLLM_USE_RAY_SPMD_WORKER=1 to enable it.
    "VLLM_USE_RAY_SPMD_WORKER":
492
    lambda: bool(int(os.getenv("VLLM_USE_RAY_SPMD_WORKER", "0"))),
493

494
495
496
    # If the env var is set, it uses the Ray's Compiled Graph
    # (previously known as ADAG) API which optimizes the
    # control plane overhead.
497
    # Run vLLM with VLLM_USE_RAY_COMPILED_DAG=1 to enable it.
498
499
    # Note that this variable is set to 1 in V1 by default
    # when ray distributed executor is used.
500
    "VLLM_USE_RAY_COMPILED_DAG":
501
502
    lambda: bool(int(os.getenv("VLLM_USE_RAY_COMPILED_DAG", "0"))),

503
504
505
506
507
508
509
510
511
512
    # 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
    # This flag is ignored if VLLM_USE_RAY_COMPILED_DAG is not set.
    "VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE":
    lambda: os.getenv("VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE", "auto"),
513

514
    # If the env var is set, it enables GPU communication overlap
515
    # (experimental feature) in Ray's Compiled Graph. This flag is ignored if
516
517
    # VLLM_USE_RAY_COMPILED_DAG is not set.
    "VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM":
518
    lambda: bool(int(os.getenv("VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM", "0"))
519
520
                 ),

521
522
523
524
525
526
527
    # 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.
    # This flag is ignored if VLLM_USE_RAY_COMPILED_DAG is not set.
    "VLLM_USE_RAY_WRAPPED_PP_COMM":
    lambda: bool(int(os.getenv("VLLM_USE_RAY_WRAPPED_PP_COMM", "1"))),

528
529
530
    # Use dedicated multiprocess context for workers.
    # Both spawn and fork work
    "VLLM_WORKER_MULTIPROC_METHOD":
531
    lambda: os.getenv("VLLM_WORKER_MULTIPROC_METHOD", "fork"),
532

533
534
535
536
537
538
539
540
    # 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"),
        )),

541
542
543
544
    # 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")),
545

546
    # Timeout for fetching videos when serving multimodal models
547
    # Default is 30 seconds
548
    "VLLM_VIDEO_FETCH_TIMEOUT":
549
    lambda: int(os.getenv("VLLM_VIDEO_FETCH_TIMEOUT", "30")),
550

551
    # Timeout for fetching audio when serving multimodal models
552
    # Default is 10 seconds
553
    "VLLM_AUDIO_FETCH_TIMEOUT":
554
    lambda: int(os.getenv("VLLM_AUDIO_FETCH_TIMEOUT", "10")),
555

556
557
558
559
560
561
    # 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")),

562
563
564
565
566
567
568
569
570
571
    # 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.
    "VLLM_VIDEO_LOADER_BACKEND":
    lambda: os.getenv("VLLM_VIDEO_LOADER_BACKEND", "opencv"),

572
    # [DEPRECATED] Cache size (in GiB per process) for multimodal input cache
573
    # Default is 4 GiB per API process + 4 GiB per engine core process
574
    "VLLM_MM_INPUT_CACHE_GIB":
575
    lambda: int(os.getenv("VLLM_MM_INPUT_CACHE_GIB", "4")),
576

577
578
579
    # Path to the XLA persistent cache directory.
    # Only used for XLA devices such as TPUs.
    "VLLM_XLA_CACHE_PATH":
580
581
    lambda: os.path.expanduser(
        os.getenv(
582
            "VLLM_XLA_CACHE_PATH",
583
584
            os.path.join(get_default_cache_root(), "vllm", "xla_cache"),
        )),
585
586
587
588

    # If set, assert on XLA recompilation after each execution step.
    "VLLM_XLA_CHECK_RECOMPILATION":
    lambda: bool(int(os.getenv("VLLM_XLA_CHECK_RECOMPILATION", "0"))),
589
590
591
592

    # Enable SPMD mode for TPU backend.
    "VLLM_XLA_USE_SPMD":
    lambda: bool(int(os.getenv("VLLM_XLA_USE_SPMD", "0"))),
593
    "VLLM_FUSED_MOE_CHUNK_SIZE":
594
    lambda: int(os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", "32768")),
595
596
597
598
599
600
    # 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.
    "VLLM_ENABLE_FUSED_MOE_ACTIVATION_CHUNKING":
    lambda: bool(
        int(os.getenv("VLLM_ENABLE_FUSED_MOE_ACTIVATION_CHUNKING", "1"))),
601

602
603
604
605
606
    # 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(os.getenv("VLLM_KEEP_ALIVE_ON_ENGINE_DEATH", 0)),

607
608
609
610
611
612
613
614
    # 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")),
615
616
617
618
619
620
621

    # 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")),
622
623
    "VLLM_TEST_FORCE_LOAD_FORMAT":
    lambda: os.getenv("VLLM_TEST_FORCE_LOAD_FORMAT", "dummy"),
624

625
626
    # Time in ms for the zmq client to wait for a response from the backend
    # server for simple data operations
627
628
    "VLLM_RPC_TIMEOUT":
    lambda: int(os.getenv("VLLM_RPC_TIMEOUT", "10000")),
629

630
631
632
633
    # 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")),

634
635
636
637
638
639
    # 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(","),
640

641
642
643
644
645
646
    # 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),

647
648
649
650
651
    # Enables torch profiler if set. Path to the directory where torch profiler
    # traces are saved. Note that it must be an absolute path.
    "VLLM_TORCH_PROFILER_DIR":
    lambda: (None if os.getenv("VLLM_TORCH_PROFILER_DIR", None) is None else os
             .path.expanduser(os.getenv("VLLM_TORCH_PROFILER_DIR", "."))),
652

653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
    # Enable torch profiler to record shapes if set
    # VLLM_TORCH_PROFILER_RECORD_SHAPES=1. If not set, torch profiler will
    # not record shapes.
    "VLLM_TORCH_PROFILER_RECORD_SHAPES":
    lambda: bool(os.getenv("VLLM_TORCH_PROFILER_RECORD_SHAPES", "0") != "0"),

    # Enable torch profiler to profile memory if set
    # VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY=1. If not set, torch profiler
    # will not profile memory.
    "VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY":
    lambda: bool(
        os.getenv("VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY", "0") != "0"),

    # Enable torch profiler to profile stack if set
    # VLLM_TORCH_PROFILER_WITH_STACK=1. If not set, torch profiler WILL
    # profile stack by default.
    "VLLM_TORCH_PROFILER_WITH_STACK":
    lambda: bool(os.getenv("VLLM_TORCH_PROFILER_WITH_STACK", "1") != "0"),

    # Enable torch profiler to profile flops if set
    # VLLM_TORCH_PROFILER_WITH_FLOPS=1. If not set, torch profiler will
    # not profile flops.
    "VLLM_TORCH_PROFILER_WITH_FLOPS":
    lambda: bool(os.getenv("VLLM_TORCH_PROFILER_WITH_FLOPS", "0") != "0"),

678
679
680
    # If set, vLLM will use Triton implementations of AWQ.
    "VLLM_USE_TRITON_AWQ":
    lambda: bool(int(os.getenv("VLLM_USE_TRITON_AWQ", "0"))),
681
682
683
684
685
686

    # 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")),
687

688
689
690
691
692
693
    # 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
694
    "VLLM_SKIP_P2P_CHECK":
695
    lambda: os.getenv("VLLM_SKIP_P2P_CHECK", "1") == "1",
696

697
698
699
700
701
702
703
    # 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(","),
704
705
706

    # If set, use the V1 code path.
    "VLLM_USE_V1":
707
    lambda: bool(int(os.getenv("VLLM_USE_V1", "1"))),
708

709
710
711
712
713
714
    # Disable aiter ops unless specifically enabled.
    # Acts as a parent switch to enable the rest of the other operations.
    "VLLM_ROCM_USE_AITER":
    lambda: (os.getenv("VLLM_ROCM_USE_AITER", "False").lower() in
             ("true", "1")),

715
716
717
718
719
720
    # Whether to use aiter paged attention.
    # By default is disabled.
    "VLLM_ROCM_USE_AITER_PAGED_ATTN":
    lambda: (os.getenv("VLLM_ROCM_USE_AITER_PAGED_ATTN", "False").lower() in
             ("true", "1")),

721
722
723
724
725
726
727
    # use aiter linear op if aiter ops are enabled
    # The following list of related ops
    # - scaled_mm (per-tensor / rowwise)
    "VLLM_ROCM_USE_AITER_LINEAR":
    lambda: (os.getenv("VLLM_ROCM_USE_AITER_LINEAR", "True").lower() in
             ("true", "1")),

728
729
730
731
732
733
    # 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")),

734
735
736
737
738
    # 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")),

739
740
741
742
743
    # 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")),
744
745
746
747
748
749
750

    # 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")),

751
752
753
754
755
    # use rocm skinny gemms
    "VLLM_ROCM_USE_SKINNY_GEMM":
    lambda: (os.getenv("VLLM_ROCM_USE_SKINNY_GEMM", "True").lower() in
             ("true", "1")),

756
757
758
    # Pad the fp8 weights to 256 bytes for ROCm
    "VLLM_ROCM_FP8_PADDING":
    lambda: bool(int(os.getenv("VLLM_ROCM_FP8_PADDING", "1"))),
759

760
761
762
763
    # Pad the weights for the moe kernel
    "VLLM_ROCM_MOE_PADDING":
    lambda: bool(int(os.getenv("VLLM_ROCM_MOE_PADDING", "1"))),

764
765
766
767
768
    # custom paged attention kernel for MI3* cards
    "VLLM_ROCM_CUSTOM_PAGED_ATTN":
    lambda: (os.getenv("VLLM_ROCM_CUSTOM_PAGED_ATTN", "True").lower() in
             ("true", "1")),

769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
    # Custom quick allreduce kernel for MI3* cards
    # Choice of quantization level: FP, INT8, INT6, INT4 or NONE
    # Recommended for large models to get allreduce
    "VLLM_ROCM_QUICK_REDUCE_QUANTIZATION":
    lambda: os.getenv("VLLM_ROCM_QUICK_REDUCE_QUANTIZATION", "NONE").upper(),

    # 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)),

794
795
796
    # Divisor for dynamic query scale factor calculation for FP8 KV Cache
    "Q_SCALE_CONSTANT":
    lambda: int(os.getenv("Q_SCALE_CONSTANT", "200")),
797
798
799
800
801
802
    # Divisor for dynamic key scale factor calculation for FP8 KV Cache
    "K_SCALE_CONSTANT":
    lambda: int(os.getenv("K_SCALE_CONSTANT", "200")),
    # Divisor for dynamic value scale factor calculation for FP8 KV Cache
    "V_SCALE_CONSTANT":
    lambda: int(os.getenv("V_SCALE_CONSTANT", "100")),
803

804
805
    # If set, enable multiprocessing in LLM for the V1 code path.
    "VLLM_ENABLE_V1_MULTIPROCESSING":
806
    lambda: bool(int(os.getenv("VLLM_ENABLE_V1_MULTIPROCESSING", "1"))),
807
808
    "VLLM_LOG_BATCHSIZE_INTERVAL":
    lambda: float(os.getenv("VLLM_LOG_BATCHSIZE_INTERVAL", "-1")),
809
810
    "VLLM_DISABLE_COMPILE_CACHE":
    lambda: bool(int(os.getenv("VLLM_DISABLE_COMPILE_CACHE", "0"))),
811
812
813
814
815
816

    # 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"))),
817
818
819
820
821
822
823
824
825
826

    # 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")),
827
828
829
830
831

    # If set, vLLM will disable the MLA attention optimizations.
    "VLLM_MLA_DISABLE":
    lambda: bool(int(os.getenv("VLLM_MLA_DISABLE", "0"))),

832
833
834
835
836
837
838
839
840
841
842
843
    # 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", ""),

844
845
846
847
    # 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),
848

849
850
851
852
    # Rank of the process in the data parallel setting
    "VLLM_DP_RANK":
    lambda: int(os.getenv("VLLM_DP_RANK", "0")),

853
854
855
856
857
858
    # 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)),

859
860
861
862
863
864
865
866
867
868
869
    # 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")),
870

871
872
873
874
875
876
877
878
    # 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.
    "VLLM_MOE_DP_CHUNK_SIZE":
    lambda: int(os.getenv("VLLM_MOE_DP_CHUNK_SIZE", "256")),

879
880
881
882
    # 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",

883
884
885
    # Whether to use S3 path for model loading in CI via RunAI Streamer
    "VLLM_CI_USE_S3":
    lambda: os.environ.get("VLLM_CI_USE_S3", "0") == "1",
886

887
    # Use model_redirect to redirect the model name to a local folder.
888
889
890
891
892
    # `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
893
894
895
    "VLLM_MODEL_REDIRECT_PATH":
    lambda: os.environ.get("VLLM_MODEL_REDIRECT_PATH", None),

896
897
898
    # 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",
899
900
901
902
903
904

    # Whether to turn on the outlines cache for V0
    # This cache is unbounded and on disk, so it's not safe to use in
    # an environment with potentially malicious users.
    "VLLM_V0_USE_OUTLINES_CACHE":
    lambda: os.environ.get("VLLM_V0_USE_OUTLINES_CACHE", "0") == "1",
905

906
907
908
909
910
911
    # 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",

912
913
914
915
    # Gap between padding buckets for the forward pass. So we have
    # 8, we will run forward pass with [16, 24, 32, ...].
    "VLLM_TPU_BUCKET_PADDING_GAP":
    lambda: int(os.environ["VLLM_TPU_BUCKET_PADDING_GAP"])
916
    if "VLLM_TPU_BUCKET_PADDING_GAP" in os.environ else 0,
917
918
    "VLLM_TPU_MOST_MODEL_LEN":
    lambda: maybe_convert_int(os.environ.get("VLLM_TPU_MOST_MODEL_LEN", None)),
919

920
921
922
923
    # Whether using Pathways
    "VLLM_TPU_USING_PATHWAYS":
    lambda: bool("proxy" in os.getenv("JAX_PLATFORMS", "").lower()),

924
925
926
    # Allow use of DeepGemm kernels for fused moe ops.
    "VLLM_USE_DEEP_GEMM":
    lambda: bool(int(os.getenv("VLLM_USE_DEEP_GEMM", "0"))),
927

928
929
930
931
932
933
934
935
    # 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.
    # Set `VLLM_SKIP_DEEP_GEMM_WARMUP` to disable the warmup.
    "VLLM_SKIP_DEEP_GEMM_WARMUP":
    lambda: bool(int(os.getenv("VLLM_SKIP_DEEP_GEMM_WARMUP", "0"))),

936
937
938
939
    # Allow use of FlashInfer MoE kernels for fused moe ops.
    "VLLM_USE_FLASHINFER_MOE_FP8":
    lambda: bool(int(os.getenv("VLLM_USE_FLASHINFER_MOE_FP8", "0"))),

940
    # Allow use of FlashInfer CUTLASS kernels for fused moe ops.
941
942
    "VLLM_USE_FLASHINFER_MOE_FP4":
    lambda: bool(int(os.getenv("VLLM_USE_FLASHINFER_MOE_FP4", "0"))),
943

944
945
946
947
948
949
950
951
952
953
    # If set to 1, use the FlashInfer
    # MXFP8 (activation) x MXFP4 (weight) MoE backend.
    "VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8":
    lambda: bool(int(os.getenv("VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8", "0"))),

    # If set to 1, use the FlashInfer
    # BF16 (activation) x MXFP4 (weight) MoE backend.
    "VLLM_USE_FLASHINFER_MOE_MXFP4_BF16":
    lambda: bool(int(os.getenv("VLLM_USE_FLASHINFER_MOE_MXFP4_BF16", "0"))),

954
955
956
957
958
    # 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")),
959
960
961
962
963
964
965
966
967
968

    # 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.
    "VLLM_MSGPACK_ZERO_COPY_THRESHOLD":
    lambda: int(os.getenv("VLLM_MSGPACK_ZERO_COPY_THRESHOLD", "256")),
969
970
971
972
973
974

    # 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"))),
Robert Shaw's avatar
Robert Shaw committed
975
976
977
978
979
980
981
982

    # 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", "5557")),
983
984

    # all2all backend for vllm's expert parallel communication
985
986
987
    # Available options:
    # - "naive": naive all2all implementation using all-reduce
    # - "pplx": use pplx kernels
988
989
    # - "deepep_high_throughput", use deepep high-throughput kernels
    # - "deepep_low_latency", use deepep low-latency kernels
990
991
    "VLLM_ALL2ALL_BACKEND":
    lambda: os.getenv("VLLM_ALL2ALL_BACKEND", "naive"),
992

993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
    # Flashinfer MoE backend for vLLM's fused Mixture-of-Experts support. Both
    # require compute capability 10.0 or above.
    # Available options:
    # - "throughput":  [default]
    #     Uses CUTLASS kernels optimized for high-throughput batch inference.
    # - "latency":
    #     Uses TensorRT-LLM kernels optimized for low-latency inference.
    # To set this backend, define the environment variable:
    #     export VLLM_FLASHINFER_MOE_BACKEND=latency.
    # If not set, defaults to "throughput".
    "VLLM_FLASHINFER_MOE_BACKEND": lambda: os.getenv(
    "VLLM_FLASHINFER_MOE_BACKEND", "throughput"
    ),

1007
1008
1009
1010
1011
1012
    # 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")),
1013

1014
1015
1016
1017
1018
1019
1020
1021
1022
    # 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
    "VLLM_MOE_ROUTING_SIMULATION_STRATEGY":
    lambda: os.environ.get("VLLM_MOE_ROUTING_SIMULATION_STRATEGY", "").lower(),

1023
1024
1025
    # 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")),
1026
1027
1028
1029
1030

    # Reduce CPU usage when vLLM is idle. Enabling this will incur small
    # latency penalty when a request eventually comes.
    "VLLM_SLEEP_WHEN_IDLE":
    lambda: bool(int(os.getenv("VLLM_SLEEP_WHEN_IDLE", "0"))),
1031
1032
1033
1034
1035
1036

    # 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")),
1037

1038
1039
1040
1041
1042
    # 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")),

1043
1044
1045
1046
1047
1048
1049
1050
    # 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":
1051
1052
1053
1054
1055
1056
1057
    lambda: os.getenv("VLLM_KV_CACHE_LAYOUT", None),

    # 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"))),
1058
1059
1060
1061
1062

    # Controls whether or not emulations are used for NVFP4
    # generations on machines < 100 for compressed-tensors
    # models
    "VLLM_USE_NVFP4_CT_EMULATIONS":
1063
1064
1065
1066
1067
1068
1069
    lambda: bool(int(os.getenv("VLLM_USE_NVFP4_CT_EMULATIONS", "0"))),

    # 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.
    "VLLM_NIXL_ABORT_REQUEST_TIMEOUT":
1070
1071
    lambda: int(os.getenv("VLLM_NIXL_ABORT_REQUEST_TIMEOUT", "120")),

1072
1073
1074
1075
    # Controls whether or not to use cudnn prefill
    "VLLM_USE_CUDNN_PREFILL":
    lambda: bool(int(os.getenv("VLLM_USE_CUDNN_PREFILL", "0"))),

1076
1077
1078
    # If set to 1, use the TRTLLM attention backend in flashinfer.
    "VLLM_USE_TRTLLM_ATTENTION":
    lambda: os.getenv("VLLM_USE_TRTLLM_ATTENTION", None),
1079

1080
1081
1082
1083
1084
1085
    # 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"))),

1086
1087
1088
    # Used to force set up loopback IP
    "VLLM_LOOPBACK_IP":
    lambda: os.getenv("VLLM_LOOPBACK_IP", ""),
1089
1090
1091
1092
1093
1094

    # 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"),
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105

    # 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", "0"))),
1106
1107
1108

    # 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
1109
1110
    # messages for those requests in memory. By default, this is disabled (0),
    # and the "store" option is ignored.
1111
1112
1113
1114
1115
1116
1117
    # 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"))),
1118
1119
}

1120
# --8<-- [end:env-vars-definition]
1121

1122

1123
def __getattr__(name: str):
1124
1125
1126
1127
1128
1129
1130
1131
    # lazy evaluation of environment variables
    if name in environment_variables:
        return environment_variables[name]()
    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")


def __dir__():
    return list(environment_variables.keys())
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147


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 "
            "Issue and explicitly set VLLM_USE_V1=0 or 1.")
    os.environ["VLLM_USE_V1"] = "1" if use_v1 else "0"
1148
1149
1150
1151
1152
1153
1154
1155


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
1156
    graphs, so it is included in the factors list. The env vars that
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
    affect the choice of different kernels or attention backends should
    also be included in the factors list.
    """
    factors: list[Any] = []

    # summarize environment variables
    def factorize(name: str):
        if __getattr__(name):
            factors.append(__getattr__(name))
        else:
            factors.append("None")

    # 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",
        "VLLM_USE_TRITON_FLASH_ATTN",
        "VLLM_USE_TRITON_AWQ",
        "VLLM_DP_RANK",
        "VLLM_DP_SIZE",
1180
        "VLLM_USE_STANDALONE_COMPILE",
1181
        "VLLM_FUSED_MOE_CHUNK_SIZE",
1182
1183
1184
1185
1186
    ]
    for key in environment_variables_to_hash:
        if key in environment_variables:
            factorize(key)

1187
1188
    hash_str = hashlib.md5(str(factors).encode(),
                           usedforsecurity=False).hexdigest()
1189
1190

    return hash_str