envs.py 35.5 KB
Newer Older
1
2
# SPDX-License-Identifier: Apache-2.0

3
import hashlib
4
import os
5
import sys
6
import tempfile
7
from typing import TYPE_CHECKING, Any, Callable, Optional
8
9
10

if TYPE_CHECKING:
    VLLM_HOST_IP: str = ""
11
    VLLM_PORT: Optional[int] = None
12
    VLLM_RPC_BASE_PATH: str = tempfile.gettempdir()
13
    VLLM_USE_MODELSCOPE: bool = False
14
    VLLM_RINGBUFFER_WARNING_INTERVAL: int = 60
15
16
17
    VLLM_NCCL_SO_PATH: Optional[str] = None
    LD_LIBRARY_PATH: Optional[str] = None
    VLLM_USE_TRITON_FLASH_ATTN: bool = False
18
    VLLM_V1_USE_PREFILL_DECODE_ATTENTION: bool = False
19
    VLLM_FLASH_ATTN_VERSION: Optional[int] = None
20
21
22
23
24
25
26
    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
27
    VLLM_MODEL_REDIRECT_PATH: Optional[str] = None
28
29
    VLLM_CACHE_ROOT: str = os.path.expanduser("~/.cache/vllm")
    VLLM_CONFIG_ROOT: str = os.path.expanduser("~/.config/vllm")
30
31
32
33
34
    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
35
    VLLM_LOGGING_LEVEL: str = "INFO"
36
    VLLM_LOGGING_PREFIX: str = ""
37
    VLLM_LOGGING_CONFIG_PATH: Optional[str] = None
38
    VLLM_LOGITS_PROCESSOR_THREADS: Optional[int] = None
39
40
    VLLM_TRACE_FUNCTION: int = 0
    VLLM_ATTENTION_BACKEND: Optional[str] = None
41
    VLLM_USE_FLASHINFER_SAMPLER: Optional[bool] = None
42
    VLLM_FLASHINFER_FORCE_TENSOR_CORES: bool = False
43
    VLLM_PP_LAYER_PARTITION: Optional[str] = None
44
    VLLM_CPU_KVCACHE_SPACE: int = 0
45
    VLLM_CPU_OMP_THREADS_BIND: str = ""
46
    VLLM_CPU_MOE_PREPACK: bool = True
47
    VLLM_XLA_CACHE_PATH: str = os.path.join(VLLM_CACHE_ROOT, "xla_cache")
48
    VLLM_XLA_CHECK_RECOMPILATION: bool = False
49
    VLLM_FUSED_MOE_CHUNK_SIZE: int = 64 * 1024
50
    VLLM_USE_RAY_SPMD_WORKER: bool = False
51
    VLLM_USE_RAY_COMPILED_DAG: bool = False
52
    VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE: str = "auto"
53
    VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM: bool = False
54
    VLLM_WORKER_MULTIPROC_METHOD: str = "fork"
55
    VLLM_ASSETS_CACHE: str = os.path.join(VLLM_CACHE_ROOT, "assets")
56
    VLLM_IMAGE_FETCH_TIMEOUT: int = 5
57
    VLLM_VIDEO_FETCH_TIMEOUT: int = 30
58
    VLLM_AUDIO_FETCH_TIMEOUT: int = 10
59
    VLLM_VIDEO_LOADER_BACKEND: str = "opencv"
60
    VLLM_MM_INPUT_CACHE_GIB: int = 8
61
62
63
64
    VLLM_TARGET_DEVICE: str = "cuda"
    MAX_JOBS: Optional[str] = None
    NVCC_THREADS: Optional[str] = None
    VLLM_USE_PRECOMPILED: bool = False
65
    VLLM_TEST_USE_PRECOMPILED_NIGHTLY_WHEEL: bool = False
66
    VLLM_NO_DEPRECATION_WARNING: bool = False
67
    VLLM_KEEP_ALIVE_ON_ENGINE_DEATH: bool = False
68
69
    CMAKE_BUILD_TYPE: Optional[str] = None
    VERBOSE: bool = False
70
    VLLM_ALLOW_LONG_MAX_MODEL_LEN: bool = False
71
    VLLM_RPC_TIMEOUT: int = 10000  # ms
72
    VLLM_PLUGINS: Optional[list[str]] = None
73
    VLLM_LORA_RESOLVER_CACHE_DIR: Optional[str] = None
74
    VLLM_TORCH_PROFILER_DIR: Optional[str] = None
75
    VLLM_USE_TRITON_AWQ: bool = False
76
    VLLM_ALLOW_RUNTIME_LORA_UPDATING: bool = False
77
    VLLM_SKIP_P2P_CHECK: bool = False
78
    VLLM_DISABLED_KERNELS: list[str] = []
79
    VLLM_USE_V1: bool = True
80
    VLLM_ROCM_USE_AITER: bool = False
81
    VLLM_ROCM_USE_AITER_PAGED_ATTN: bool = False
82
    VLLM_ROCM_USE_AITER_LINEAR: bool = True
83
    VLLM_ROCM_USE_AITER_MOE: bool = True
84
    VLLM_ROCM_USE_AITER_RMSNORM: bool = True
85
    VLLM_ROCM_USE_AITER_MLA: bool = True
86
    VLLM_ROCM_USE_SKINNY_GEMM: bool = True
87
    VLLM_ROCM_FP8_PADDING: bool = True
88
    VLLM_ROCM_MOE_PADDING: bool = True
89
    VLLM_ROCM_CUSTOM_PAGED_ATTN: bool = True
90
    VLLM_QUARK_EMU_MEM_OPT: bool = False
91
    VLLM_ENABLE_V1_MULTIPROCESSING: bool = True
92
    VLLM_LOG_BATCHSIZE_INTERVAL: float = -1
93
    VLLM_DISABLE_COMPILE_CACHE: bool = False
94
    Q_SCALE_CONSTANT: int = 200
95
96
    K_SCALE_CONSTANT: int = 200
    V_SCALE_CONSTANT: int = 100
97
    VLLM_SERVER_DEV_MODE: bool = False
98
    VLLM_V1_OUTPUT_PROC_CHUNK_SIZE: int = 128
99
    VLLM_MLA_DISABLE: bool = False
100
    VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON: bool = False
101
102
    VLLM_RAY_PER_WORKER_GPUS: float = 1.0
    VLLM_RAY_BUNDLE_INDICES: str = ""
103
    VLLM_CUDART_SO_PATH: Optional[str] = None
104
    VLLM_USE_HPU_CONTIGUOUS_CACHE_FETCH: bool = True
105
    VLLM_HPU_USE_DELAYED_SAMPLING: bool = False
106
    VLLM_DP_RANK: int = 0
107
    VLLM_DP_RANK_LOCAL: int = -1
108
109
110
    VLLM_DP_SIZE: int = 1
    VLLM_DP_MASTER_IP: str = ""
    VLLM_DP_MASTER_PORT: int = 0
111
    VLLM_MARLIN_USE_ATOMIC_ADD: bool = False
112
    VLLM_V0_USE_OUTLINES_CACHE: bool = False
113
    VLLM_TPU_BUCKET_PADDING_GAP: int = 0
114
    VLLM_USE_DEEP_GEMM: bool = False
115
    VLLM_XGRAMMAR_CACHE_MB: int = 0
116
    VLLM_MSGPACK_ZERO_COPY_THRESHOLD: int = 256
117
    VLLM_ALLOW_INSECURE_SERIALIZATION: bool = False
Robert Shaw's avatar
Robert Shaw committed
118
119
    VLLM_NIXL_SIDE_CHANNEL_HOST: str = "localhost"
    VLLM_NIXL_SIDE_CHANNEL_PORT: int = 5557
120
    VLLM_ALL2ALL_BACKEND: str = "naive"
121
    VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE: int = 163840
122

123
124
125
126
127
128
129
130
131
132
133
134
135
136
137

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"),
    )


138
139
140
141
142
143
def maybe_convert_int(value: Optional[str]) -> Optional[int]:
    if value is None:
        return None
    return int(value)


144
145
def get_vllm_port() -> Optional[int]:
    """Get the port from VLLM_PORT environment variable.
146

147
148
    Returns:
        The port number as an integer if VLLM_PORT is set, None otherwise.
149

150
151
152
153
154
155
156
157
158
159
160
161
    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
162
163
164
165
166
167
168
        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
169
170
171
172
        raise ValueError(
            f"VLLM_PORT '{port}' must be a valid integer") from err


173
174
175
# The begin-* and end* here are used by the documentation generator
# to extract the used env vars.

176
# --8<-- [start:env-vars-definition]
177

178
environment_variables: dict[str, Callable[[], Any]] = {
179
180
181

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

182
    # Target device of vLLM, supporting [cuda (by default),
183
    # rocm, neuron, cpu]
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
    "VLLM_TARGET_DEVICE":
    lambda: os.getenv("VLLM_TARGET_DEVICE", "cuda"),

    # 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":
200
201
    lambda: bool(os.environ.get("VLLM_USE_PRECOMPILED")) or bool(
        os.environ.get("VLLM_PRECOMPILED_WHEEL_LOCATION")),
202

203
204
205
206
207
208
    # 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"))
                 ),

209
210
211
212
213
214
215
216
217
218
    # 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'))),

219
    # Root directory for vLLM configuration files
220
    # Defaults to `~/.config/vllm` unless `XDG_CONFIG_HOME` is set
221
222
223
224
    # 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":
225
226
227
228
229
    lambda: os.path.expanduser(
        os.getenv(
            "VLLM_CONFIG_ROOT",
            os.path.join(get_default_config_root(), "vllm"),
        )),
230
231
232

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

233
    # Root directory for vLLM cache files
234
235
236
237
238
239
240
241
    # 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"),
        )),

242
243
244
245
    # 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.
246
    'VLLM_HOST_IP':
247
    lambda: os.getenv('VLLM_HOST_IP', ""),
248

249
    # used in distributed environment to manually set the communication port
250
251
252
    # 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.
253
    'VLLM_PORT':
254
    get_vllm_port,
255

256
257
258
259
    # 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()),
260

261
262
263
264
265
    # 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",

266
267
268
269
    # 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")),

270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
    # 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")),

290
291
292
293
294
295
296
    # 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")),

297
298
299
300
301
    # 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)),

302
303
304
305
    # Internal flag to enable Dynamo fullgraph capture
    "VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE":
    lambda: bool(
        os.environ.get("VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE", "1") != "0"),
306

307
308
309
310
311
    # 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",
312

313
314
315
316
317
318
319
320
321
322
323
324
325
    # 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")),

326
    # API key for vLLM API server
327
328
329
    "VLLM_API_KEY":
    lambda: os.environ.get("VLLM_API_KEY", None),

330
331
    # Whether to log responses from API Server for debugging
    "VLLM_DEBUG_LOG_API_SERVER_RESPONSE":
332
333
    lambda: os.environ.get("VLLM_DEBUG_LOG_API_SERVER_RESPONSE", "False"
                           ).lower() == "true",
334

335
336
    # S3 access information, used for tensorizer to load model from S3
    "S3_ACCESS_KEY_ID":
337
    lambda: os.environ.get("S3_ACCESS_KEY_ID", None),
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
    "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"),

363
364
    # this is used for configuring the default logging level
    "VLLM_LOGGING_LEVEL":
365
    lambda: os.getenv("VLLM_LOGGING_LEVEL", "INFO").upper(),
366

367
368
369
370
    # if set, VLLM_LOGGING_PREFIX will be prepended to all log messages
    "VLLM_LOGGING_PREFIX":
    lambda: os.getenv("VLLM_LOGGING_PREFIX", ""),

371
372
373
374
375
376
377
378
    # 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,

379
380
381
382
383
384
385
386
387
388
389
390
    # 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
391
    # - "FLASHINFER": use flashinfer
392
    # - "FLASHMLA": use FlashMLA
393
394
395
    "VLLM_ATTENTION_BACKEND":
    lambda: os.getenv("VLLM_ATTENTION_BACKEND", None),

396
397
    # If set, vllm will use flashinfer sampler
    "VLLM_USE_FLASHINFER_SAMPLER":
398
399
    lambda: bool(int(os.environ["VLLM_USE_FLASHINFER_SAMPLER"]))
    if "VLLM_USE_FLASHINFER_SAMPLER" in os.environ else None,
400

401
402
403
404
405
    # 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"))),

406
407
408
409
    # Pipeline stage partition strategy
    "VLLM_PP_LAYER_PARTITION":
    lambda: os.getenv("VLLM_PP_LAYER_PARTITION", None),

410
    # (CPU backend only) CPU key-value cache space.
411
    # default is 4 GiB
412
413
414
    "VLLM_CPU_KVCACHE_SPACE":
    lambda: int(os.getenv("VLLM_CPU_KVCACHE_SPACE", "0")),

415
416
417
418
419
    # (CPU backend only) CPU core ids bound by OpenMP threads, e.g., "0-31",
    # "0,1,2", "0-31,33". CPU cores of different ranks are separated by '|'.
    "VLLM_CPU_OMP_THREADS_BIND":
    lambda: os.getenv("VLLM_CPU_OMP_THREADS_BIND", "all"),

420
421
422
423
424
425
    # (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"))),

426
427
428
429
430
    # 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":
431
    lambda: bool(int(os.getenv("VLLM_USE_RAY_SPMD_WORKER", "0"))),
432

433
434
435
    # If the env var is set, it uses the Ray's Compiled Graph
    # (previously known as ADAG) API which optimizes the
    # control plane overhead.
436
    # Run vLLM with VLLM_USE_RAY_COMPILED_DAG=1 to enable it.
437
438
    # Note that this variable is set to 1 in V1 by default
    # when ray distributed executor is used.
439
    "VLLM_USE_RAY_COMPILED_DAG":
440
441
    lambda: bool(int(os.getenv("VLLM_USE_RAY_COMPILED_DAG", "0"))),

442
443
444
445
446
447
448
449
450
451
    # 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"),
452

453
    # If the env var is set, it enables GPU communication overlap
454
    # (experimental feature) in Ray's Compiled Graph. This flag is ignored if
455
456
    # VLLM_USE_RAY_COMPILED_DAG is not set.
    "VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM":
457
    lambda: bool(int(os.getenv("VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM", "0"))
458
459
                 ),

460
461
462
    # Use dedicated multiprocess context for workers.
    # Both spawn and fork work
    "VLLM_WORKER_MULTIPROC_METHOD":
463
    lambda: os.getenv("VLLM_WORKER_MULTIPROC_METHOD", "fork"),
464

465
466
467
468
469
470
471
472
    # 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"),
        )),

473
474
475
476
    # 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")),
477

478
    # Timeout for fetching videos when serving multimodal models
479
    # Default is 30 seconds
480
    "VLLM_VIDEO_FETCH_TIMEOUT":
481
    lambda: int(os.getenv("VLLM_VIDEO_FETCH_TIMEOUT", "30")),
482

483
    # Timeout for fetching audio when serving multimodal models
484
    # Default is 10 seconds
485
    "VLLM_AUDIO_FETCH_TIMEOUT":
486
    lambda: int(os.getenv("VLLM_AUDIO_FETCH_TIMEOUT", "10")),
487

488
489
490
491
492
493
494
495
496
497
    # 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"),

498
    # Cache size (in GiB) for multimodal input cache
499
    # Default is 4 GiB
500
    "VLLM_MM_INPUT_CACHE_GIB":
501
    lambda: int(os.getenv("VLLM_MM_INPUT_CACHE_GIB", "4")),
502

503
504
505
    # Path to the XLA persistent cache directory.
    # Only used for XLA devices such as TPUs.
    "VLLM_XLA_CACHE_PATH":
506
507
    lambda: os.path.expanduser(
        os.getenv(
508
            "VLLM_XLA_CACHE_PATH",
509
510
            os.path.join(get_default_cache_root(), "vllm", "xla_cache"),
        )),
511
512
513
514

    # If set, assert on XLA recompilation after each execution step.
    "VLLM_XLA_CHECK_RECOMPILATION":
    lambda: bool(int(os.getenv("VLLM_XLA_CHECK_RECOMPILATION", "0"))),
515
    "VLLM_FUSED_MOE_CHUNK_SIZE":
516
    lambda: int(os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", "32768")),
517
518
519
520

    # If set, vllm will skip the deprecation warnings.
    "VLLM_NO_DEPRECATION_WARNING":
    lambda: bool(int(os.getenv("VLLM_NO_DEPRECATION_WARNING", "0"))),
521

522
523
524
525
526
    # 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)),

527
528
529
530
531
532
533
534
    # 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")),
535
536
537
538
539
540
541

    # 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")),
542
543
    "VLLM_TEST_FORCE_LOAD_FORMAT":
    lambda: os.getenv("VLLM_TEST_FORCE_LOAD_FORMAT", "dummy"),
544

545
546
    # Time in ms for the zmq client to wait for a response from the backend
    # server for simple data operations
547
548
    "VLLM_RPC_TIMEOUT":
    lambda: int(os.getenv("VLLM_RPC_TIMEOUT", "10000")),
549

550
551
552
553
554
555
    # 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(","),
556

557
558
559
560
561
562
    # 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),

563
564
565
566
567
    # 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", "."))),
568
569
570
571

    # If set, vLLM will use Triton implementations of AWQ.
    "VLLM_USE_TRITON_AWQ":
    lambda: bool(int(os.getenv("VLLM_USE_TRITON_AWQ", "0"))),
572
573
574
575
576
577

    # 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")),
578
579
580
581
582
583
584

    # By default, vLLM will check the peer-to-peer capability itself,
    # in case of broken drivers. See https://github.com/vllm-project/vllm/blob/a9b15c606fea67a072416ea0ea115261a2756058/vllm/distributed/device_communicators/custom_all_reduce_utils.py#L101-L108 for details. # noqa
    # If this env var is set to 1, vLLM will skip the peer-to-peer check,
    # and trust the driver's peer-to-peer capability report.
    "VLLM_SKIP_P2P_CHECK":
    lambda: os.getenv("VLLM_SKIP_P2P_CHECK", "0") == "1",
585

586
587
588
589
590
591
592
    # 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(","),
593
594
595

    # If set, use the V1 code path.
    "VLLM_USE_V1":
596
    lambda: bool(int(os.getenv("VLLM_USE_V1", "1"))),
597

598
599
600
601
602
603
    # 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")),

604
605
606
607
608
609
    # 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")),

610
611
612
613
614
615
616
    # 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")),

617
618
619
620
621
622
    # 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")),

623
624
625
626
627
    # 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")),

628
629
630
631
632
    # 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")),
633
634
635
636
637
    # use rocm skinny gemms
    "VLLM_ROCM_USE_SKINNY_GEMM":
    lambda: (os.getenv("VLLM_ROCM_USE_SKINNY_GEMM", "True").lower() in
             ("true", "1")),

638
639
640
    # Pad the fp8 weights to 256 bytes for ROCm
    "VLLM_ROCM_FP8_PADDING":
    lambda: bool(int(os.getenv("VLLM_ROCM_FP8_PADDING", "1"))),
641

642
643
644
645
    # Pad the weights for the moe kernel
    "VLLM_ROCM_MOE_PADDING":
    lambda: bool(int(os.getenv("VLLM_ROCM_MOE_PADDING", "1"))),

646
647
648
649
650
    # 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")),

651
652
653
654
655
656
657
658
    # If set, when running in Quark emulation mode, do not dequantize the
    # weights at load time. Instead, dequantize weights on-the-fly during
    # kernel execution.
    # This allows running larger models at the cost of slower inference.
    # This flag has no effect when not running in Quark emulation mode.
    "VLLM_QUARK_EMU_MEM_OPT":
    lambda: bool(int(os.getenv("VLLM_QUARK_EMU_MEM_OPT", "0"))),

659
660
661
    # Divisor for dynamic query scale factor calculation for FP8 KV Cache
    "Q_SCALE_CONSTANT":
    lambda: int(os.getenv("Q_SCALE_CONSTANT", "200")),
662
663
664
665
666
667
    # 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")),
668

669
670
    # If set, enable multiprocessing in LLM for the V1 code path.
    "VLLM_ENABLE_V1_MULTIPROCESSING":
671
    lambda: bool(int(os.getenv("VLLM_ENABLE_V1_MULTIPROCESSING", "1"))),
672
673
    "VLLM_LOG_BATCHSIZE_INTERVAL":
    lambda: float(os.getenv("VLLM_LOG_BATCHSIZE_INTERVAL", "-1")),
674
675
    "VLLM_DISABLE_COMPILE_CACHE":
    lambda: bool(int(os.getenv("VLLM_DISABLE_COMPILE_CACHE", "0"))),
676
677
678
679
680
681

    # 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"))),
682
683
684
685
686
687
688
689
690
691

    # 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")),
692
693
694
695
696

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

697
698
699
700
701
    # If set, vLLM will use the Triton implementation of moe_align_block_size,
    # i.e. moe_align_block_size_triton in fused_moe.py.
    "VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON":
    lambda: bool(int(os.getenv("VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON", "0"))
                 ),
702

703
704
705
706
707
708
709
710
711
712
713
714
    # 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", ""),

715
716
717
718
    # 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),
719
720
721
722
723
724
725

    # Contiguous cache fetching to avoid using costly gather operation on
    # Gaudi3. This is only applicable to HPU contiguous cache. If set to true,
    # contiguous cache fetch will be used.
    "VLLM_USE_HPU_CONTIGUOUS_CACHE_FETCH":
    lambda: os.environ.get("VLLM_CONTIGUOUS_PA", "true").lower() in
    ("1", "true"),
726

727
728
729
730
731
732
    # Use delayed sampling for HPU to reduce host cpu overhead
    # between each step.
    "VLLM_HPU_USE_DELAYED_SAMPLING":
    lambda: os.environ.get("VLLM_DELAYED_SAMPLING", "false").lower() in
    ("1", "true"),

733
734
735
736
    # Rank of the process in the data parallel setting
    "VLLM_DP_RANK":
    lambda: int(os.getenv("VLLM_DP_RANK", "0")),

737
738
739
740
741
742
    # 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)),

743
744
745
746
747
748
749
750
751
752
753
    # 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")),
754
755
756
757

    # 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",
758

759
    # Use model_redirect to redirect the model name to a local folder.
760
761
762
763
764
    # `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
765
766
767
    "VLLM_MODEL_REDIRECT_PATH":
    lambda: os.environ.get("VLLM_MODEL_REDIRECT_PATH", None),

768
769
770
    # 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",
771
772
773
774
775
776

    # 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",
777

778
779
780
781
    # 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"])
782
    if "VLLM_TPU_BUCKET_PADDING_GAP" in os.environ else 0,
783
784
785
786

    # Allow use of DeepGemm kernels for fused moe ops.
    "VLLM_USE_DEEP_GEMM":
    lambda: bool(int(os.getenv("VLLM_USE_DEEP_GEMM", "0"))),
787
788
789
790
791
792

    # 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")),
793
794
795
796
797
798
799
800
801
802

    # 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")),
803
804
805
806
807
808

    # 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
809
810
811
812
813
814
815
816

    # 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")),
817
818

    # all2all backend for vllm's expert parallel communication
819
820
821
    # Available options:
    # - "naive": naive all2all implementation using all-reduce
    # - "pplx": use pplx kernels
822
823
    "VLLM_ALL2ALL_BACKEND":
    lambda: os.getenv("VLLM_ALL2ALL_BACKEND", "naive"),
824
825
826
827
828
829
830

    # 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")),
831
832
}

833
# --8<-- [end:env-vars-definition]
834

835

836
def __getattr__(name: str):
837
838
839
840
841
842
843
844
    # 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())
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860


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"
861
862
863
864
865
866
867
868


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
869
    graphs, so it is included in the factors list. The env vars that
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
    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",
893
        "VLLM_USE_STANDALONE_COMPILE",
894
895
896
897
898
    ]
    for key in environment_variables_to_hash:
        if key in environment_variables:
            factorize(key)

899
900
    hash_str = hashlib.md5(str(factors).encode(),
                           usedforsecurity=False).hexdigest()
901
902

    return hash_str