envs.py 61.2 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 json
6
import os
7
import sys
8
import tempfile
9
from typing import TYPE_CHECKING, Any, Callable, Optional
10
11
12

if TYPE_CHECKING:
    VLLM_HOST_IP: str = ""
13
    VLLM_PORT: Optional[int] = None
14
    VLLM_RPC_BASE_PATH: str = tempfile.gettempdir()
15
    VLLM_USE_MODELSCOPE: bool = False
16
    VLLM_RINGBUFFER_WARNING_INTERVAL: int = 60
17
18
    VLLM_NCCL_SO_PATH: Optional[str] = None
    LD_LIBRARY_PATH: Optional[str] = None
19
    VLLM_USE_TRITON_FLASH_ATTN: bool = True
20
    VLLM_V1_USE_PREFILL_DECODE_ATTENTION: bool = False
21
    VLLM_USE_AITER_UNIFIED_ATTENTION: bool = False
22
    VLLM_FLASH_ATTN_VERSION: Optional[int] = None
23
24
25
26
27
28
29
    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
30
    VLLM_MODEL_REDIRECT_PATH: Optional[str] = None
31
32
    VLLM_CACHE_ROOT: str = os.path.expanduser("~/.cache/vllm")
    VLLM_CONFIG_ROOT: str = os.path.expanduser("~/.config/vllm")
33
34
35
36
37
    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
38
    VLLM_LOGGING_LEVEL: str = "INFO"
39
    VLLM_LOGGING_PREFIX: str = ""
40
    VLLM_LOGGING_STREAM: str = "ext://sys.stdout"
41
    VLLM_LOGGING_CONFIG_PATH: Optional[str] = None
42
    VLLM_LOGITS_PROCESSOR_THREADS: Optional[int] = None
43
    VLLM_LOG_STATS_INTERVAL: float = 10.
44
45
    VLLM_TRACE_FUNCTION: int = 0
    VLLM_ATTENTION_BACKEND: Optional[str] = None
46
    VLLM_USE_FLASHINFER_SAMPLER: Optional[bool] = None
47
    VLLM_PP_LAYER_PARTITION: Optional[str] = None
48
    VLLM_CPU_KVCACHE_SPACE: Optional[int] = 0
49
    VLLM_CPU_OMP_THREADS_BIND: str = ""
50
    VLLM_CPU_NUM_OF_RESERVED_CPU: Optional[int] = None
51
    VLLM_CPU_MOE_PREPACK: bool = True
52
    VLLM_CPU_SGL_KERNEL: bool = False
53
    VLLM_XLA_CACHE_PATH: str = os.path.join(VLLM_CACHE_ROOT, "xla_cache")
54
    VLLM_XLA_CHECK_RECOMPILATION: bool = False
55
    VLLM_FUSED_MOE_CHUNK_SIZE: int = 64 * 1024
56
    VLLM_ENABLE_FUSED_MOE_ACTIVATION_CHUNKING: bool = True
57
    VLLM_USE_RAY_SPMD_WORKER: bool = False
58
    VLLM_USE_RAY_COMPILED_DAG: bool = False
59
    VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE: str = "auto"
60
    VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM: bool = False
61
    VLLM_USE_RAY_WRAPPED_PP_COMM: bool = True
62
    VLLM_XLA_USE_SPMD: bool = False
63
    VLLM_WORKER_MULTIPROC_METHOD: str = "spawn"
64
    VLLM_ASSETS_CACHE: str = os.path.join(VLLM_CACHE_ROOT, "assets")
65
    VLLM_IMAGE_FETCH_TIMEOUT: int = 5
66
    VLLM_VIDEO_FETCH_TIMEOUT: int = 30
67
    VLLM_AUDIO_FETCH_TIMEOUT: int = 10
68
    VLLM_MEDIA_LOADING_THREAD_COUNT: int = 8
69
    VLLM_MAX_AUDIO_CLIP_FILESIZE_MB: int = 25
70
    VLLM_VIDEO_LOADER_BACKEND: str = "opencv"
71
    VLLM_MM_INPUT_CACHE_GIB: int = 4
72
73
74
75
    VLLM_TARGET_DEVICE: str = "cuda"
    MAX_JOBS: Optional[str] = None
    NVCC_THREADS: Optional[str] = None
    VLLM_USE_PRECOMPILED: bool = False
76
    VLLM_DOCKER_BUILD_CONTEXT: bool = False
77
    VLLM_TEST_USE_PRECOMPILED_NIGHTLY_WHEEL: bool = False
78
    VLLM_KEEP_ALIVE_ON_ENGINE_DEATH: bool = False
79
80
    CMAKE_BUILD_TYPE: Optional[str] = None
    VERBOSE: bool = False
81
    VLLM_ALLOW_LONG_MAX_MODEL_LEN: bool = False
82
    VLLM_RPC_TIMEOUT: int = 10000  # ms
83
    VLLM_HTTP_TIMEOUT_KEEP_ALIVE: int = 5  # seconds
84
    VLLM_PLUGINS: Optional[list[str]] = None
85
    VLLM_LORA_RESOLVER_CACHE_DIR: Optional[str] = None
86
    VLLM_TORCH_PROFILER_DIR: Optional[str] = None
87
88
89
90
    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
91
    VLLM_USE_TRITON_AWQ: bool = False
92
    VLLM_ALLOW_RUNTIME_LORA_UPDATING: bool = False
93
    VLLM_SKIP_P2P_CHECK: bool = False
94
    VLLM_DISABLED_KERNELS: list[str] = []
95
    VLLM_USE_V1: bool = True
96
    VLLM_ROCM_USE_AITER: bool = False
97
    VLLM_ROCM_USE_AITER_PAGED_ATTN: bool = False
98
    VLLM_ROCM_USE_AITER_LINEAR: bool = True
99
    VLLM_ROCM_USE_AITER_MOE: bool = True
100
    VLLM_ROCM_USE_AITER_RMSNORM: bool = True
101
    VLLM_ROCM_USE_AITER_MLA: bool = True
102
    VLLM_ROCM_USE_AITER_MHA: bool = True
103
    VLLM_ROCM_USE_AITER_FP8BMM: bool = True
104
    VLLM_ROCM_USE_SKINNY_GEMM: bool = True
105
    VLLM_ROCM_FP8_PADDING: bool = True
106
    VLLM_ROCM_MOE_PADDING: bool = True
107
    VLLM_ROCM_CUSTOM_PAGED_ATTN: bool = True
108
    VLLM_ENABLE_V1_MULTIPROCESSING: bool = True
109
    VLLM_LOG_BATCHSIZE_INTERVAL: float = -1
110
    VLLM_DISABLE_COMPILE_CACHE: bool = False
111
    Q_SCALE_CONSTANT: int = 200
112
113
    K_SCALE_CONSTANT: int = 200
    V_SCALE_CONSTANT: int = 100
114
    VLLM_SERVER_DEV_MODE: bool = False
115
    VLLM_V1_OUTPUT_PROC_CHUNK_SIZE: int = 128
116
    VLLM_MLA_DISABLE: bool = False
117
118
    VLLM_RAY_PER_WORKER_GPUS: float = 1.0
    VLLM_RAY_BUNDLE_INDICES: str = ""
119
    VLLM_CUDART_SO_PATH: Optional[str] = None
120
    VLLM_DP_RANK: int = 0
121
    VLLM_DP_RANK_LOCAL: int = -1
122
123
124
    VLLM_DP_SIZE: int = 1
    VLLM_DP_MASTER_IP: str = ""
    VLLM_DP_MASTER_PORT: int = 0
125
    VLLM_MOE_DP_CHUNK_SIZE: int = 256
126
    VLLM_RANDOMIZE_DP_DUMMY_INPUTS: bool = False
127
    VLLM_MARLIN_USE_ATOMIC_ADD: bool = False
128
    VLLM_MXFP4_USE_MARLIN: Optional[bool] = None
129
    VLLM_V0_USE_OUTLINES_CACHE: bool = False
130
    VLLM_V1_USE_OUTLINES_CACHE: bool = False
131
    VLLM_TPU_BUCKET_PADDING_GAP: int = 0
132
    VLLM_TPU_MOST_MODEL_LEN: Optional[int] = None
133
    VLLM_TPU_USING_PATHWAYS: bool = False
134
    VLLM_USE_DEEP_GEMM: bool = False
135
    VLLM_USE_DEEP_GEMM_E8M0: bool = True
136
    VLLM_USE_DEEP_GEMM_E8M0_HOPPER: bool = False
137
    VLLM_SKIP_DEEP_GEMM_WARMUP: bool = False
138
    VLLM_USE_FUSED_MOE_GROUPED_TOPK: bool = True
139
140
    VLLM_USE_FLASHINFER_MOE_FP8: bool = False
    VLLM_USE_FLASHINFER_MOE_FP4: bool = False
141
    VLLM_FLASHINFER_MOE_BACKEND: str = "throughput"
142
    VLLM_XGRAMMAR_CACHE_MB: int = 0
143
    VLLM_MSGPACK_ZERO_COPY_THRESHOLD: int = 256
144
    VLLM_ALLOW_INSECURE_SERIALIZATION: bool = False
Robert Shaw's avatar
Robert Shaw committed
145
146
    VLLM_NIXL_SIDE_CHANNEL_HOST: str = "localhost"
    VLLM_NIXL_SIDE_CHANNEL_PORT: int = 5557
147
    VLLM_ALL2ALL_BACKEND: str = "naive"
148
    VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE: int = 163840
149
    VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS: int = 1
150
    VLLM_SLEEP_WHEN_IDLE: bool = False
151
    VLLM_MQ_MAX_CHUNK_BYTES_MB: int = 16
152
    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: int = 300
153
    VLLM_KV_CACHE_LAYOUT: Optional[str] = None
154
    VLLM_COMPUTE_NANS_IN_LOGITS: bool = False
155
    VLLM_USE_NVFP4_CT_EMULATIONS: bool = False
156
157
158
    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
159
    VLLM_NIXL_ABORT_REQUEST_TIMEOUT: int = 120
160
    VLLM_USE_CUDNN_PREFILL: bool = False
161
    VLLM_ENABLE_CUDAGRAPH_GC: bool = False
162
    VLLM_LOOPBACK_IP: str = ""
163
    VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE: bool = False
164
    VLLM_ENABLE_RESPONSES_API_STORE: bool = False
165
    VLLM_USE_TRTLLM_ATTENTION: Optional[str] = None
166
    VLLM_FLASHINFER_DISABLE_Q_QUANTIZATION: bool = False
167
    VLLM_HAS_FLASHINFER_CUBIN: bool = False
168
169
    VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8: bool = False
    VLLM_USE_FLASHINFER_MOE_MXFP4_BF16: bool = False
170
    VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS: bool = False
171
    VLLM_ALLREDUCE_USE_SYMM_MEM: bool = False
172
    VLLM_TUNED_CONFIG_FOLDER: Optional[str] = None
173
    VLLM_DISABLE_PAD_FOR_CUDAGRAPH: bool = False
174
175
    VLLM_GPT_OSS_USE_CONTAINER_TOOL: bool = False
    VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS: bool = False
176
    VLLM_CUSTOM_SCOPES_FOR_PROFILING: bool = False
177
    VLLM_KV_EVENTS_USE_INT_BLOCK_HASHES: bool = True
zhuwenwen's avatar
zhuwenwen committed
178
    
179
180
181
182
183
184
185
186
187
    # add envs
    VLLM_OPTEST_URLS_PORT: Optional[int] = None
    VLLM_OPTEST_MODELS_PATH: str = ""
    VLLM_USE_TRITON_PREFIX_FLASH_ATTN: bool = False
    VLLM_USE_TRITON_OPT_MLA: bool = False
    VLLM_USE_FLASH_MLA: bool = False
    VLLM_USE_OPT_OP: bool = False
    VLLM_USE_TC_PAGED_ATTN: bool = False
    VLLM_USE_PA_PRINT_PARAM: bool = False 
zhuwenwen's avatar
zhuwenwen committed
188
    VLLM_TREE_DECODING: bool = False
189
190
    VLLM_SPEC_DECODE_EAGER: bool = False
    VLLM_PCIE_USE_CUSTOM_ALLREDUCE: bool = False
zhuwenwen's avatar
zhuwenwen committed
191
    VLLM_CUSTOM_ALLREDUCE_SUPPORTED_WORLDSIZE_MAX: int = 16
192
193
    VLLM_ENFORCE_EAGER_BS_THRESHOLD: Optional[int] = None
    VLLM_HAS_CONTEXT_DEFAULT: bool = False
194
    VLLM_USE_NN: bool = False
195
    VLLM_ENABLE_TBO: bool = False
196
197
    VLLM_TBO_REQ_DELAY_MS: int = 0
    VLLM_TBO_DECODE_BS: int = 0
zhuwenwen's avatar
zhuwenwen committed
198
    VLLM_TBO_MIN_TOKENS: int = 200
199
    VLLM_ZERO_OVERHEAD: bool = False
200
    VLLM_ENABLE_MOE_FUSED_GATE: bool = False
201
    VLLM_USE_FLASH_ATTN_PA: bool = False
zhuwenwen's avatar
zhuwenwen committed
202
    VLLM_USE_APEX_RN: bool = False
203
    VLLM_USE_GLOBAL_CACHE13: bool = False
204
    VLLM_USE_LIGHT_OP: bool = False
205
    VLLM_USE_TRITON_CAT: bool = False
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221


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"),
    )


222
223
224
225
226
227
def maybe_convert_int(value: Optional[str]) -> Optional[int]:
    if value is None:
        return None
    return int(value)


228
229
230
231
232
233
def maybe_convert_bool(value: Optional[str]) -> Optional[bool]:
    if value is None:
        return None
    return bool(int(value))


234
235
def get_vllm_port() -> Optional[int]:
    """Get the port from VLLM_PORT environment variable.
236

237
238
    Returns:
        The port number as an integer if VLLM_PORT is set, None otherwise.
239

240
241
242
243
244
245
246
247
248
249
250
251
    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
252
253
254
255
256
257
258
        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
259
260
261
262
        raise ValueError(
            f"VLLM_PORT '{port}' must be a valid integer") from err


263
264
265
# The begin-* and end* here are used by the documentation generator
# to extract the used env vars.

266
# --8<-- [start:env-vars-definition]
267

268
environment_variables: dict[str, Callable[[], Any]] = {
269
270
271

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

272
    # Target device of vLLM, supporting [cuda (by default),
273
    # rocm, cpu]
274
    "VLLM_TARGET_DEVICE":
275
    lambda: os.getenv("VLLM_TARGET_DEVICE", "cuda").lower(),
276
277
278
279
280
281
282
283
284
285
286
287
288
289

    # 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":
290
291
292
293
294
295
296
297
    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"),
298

299
300
301
302
303
304
    # 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"))
                 ),

305
306
307
308
309
310
311
312
313
314
    # 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'))),

315
    # Root directory for vLLM configuration files
316
    # Defaults to `~/.config/vllm` unless `XDG_CONFIG_HOME` is set
317
318
319
320
    # 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":
321
322
323
324
325
    lambda: os.path.expanduser(
        os.getenv(
            "VLLM_CONFIG_ROOT",
            os.path.join(get_default_config_root(), "vllm"),
        )),
326
327
328

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

329
    # Root directory for vLLM cache files
330
331
332
333
334
335
336
337
    # 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"),
        )),

338
339
340
341
    # 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.
342
    'VLLM_HOST_IP':
343
    lambda: os.getenv('VLLM_HOST_IP', ""),
344

345
    # used in distributed environment to manually set the communication port
346
347
348
    # 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.
349
    'VLLM_PORT':
350
    get_vllm_port,
351

352
353
354
355
    # 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()),
356

357
358
359
360
361
    # 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",

362
363
364
365
    # 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")),

366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
    # 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":
383
    lambda: (os.environ.get("VLLM_USE_TRITON_FLASH_ATTN", "False").lower() in
384
385
             ("true", "1")),

386
387
388
389
390
391
392
    # 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")),

393
394
395
396
397
398
    # 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")),

399
400
401
402
403
    # 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)),

404
405
406
407
408
    # Internal flag to enable Dynamo fullgraph capture
    "VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE":
    lambda: bool(
        os.environ.get("VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE", "1") != "0"),

409
410
411
412
413
    # 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",
414

415
416
417
418
419
420
421
422
423
424
425
    # 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":
426
    lambda: int(os.environ.get("VLLM_ENGINE_ITERATION_TIMEOUT_S", "120")),
427

428
    # API key for vLLM API server
429
430
431
    "VLLM_API_KEY":
    lambda: os.environ.get("VLLM_API_KEY", None),

432
433
    # Whether to log responses from API Server for debugging
    "VLLM_DEBUG_LOG_API_SERVER_RESPONSE":
434
435
    lambda: os.environ.get("VLLM_DEBUG_LOG_API_SERVER_RESPONSE", "False"
                           ).lower() == "true",
436

437
438
    # S3 access information, used for tensorizer to load model from S3
    "S3_ACCESS_KEY_ID":
439
    lambda: os.environ.get("S3_ACCESS_KEY_ID", None),
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
    "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"),

465
466
    # this is used for configuring the default logging level
    "VLLM_LOGGING_LEVEL":
467
    lambda: os.getenv("VLLM_LOGGING_LEVEL", "INFO").upper(),
468

469
470
471
472
    # this is used for configuring the default logging stream
    "VLLM_LOGGING_STREAM":
    lambda: os.getenv("VLLM_LOGGING_STREAM", "ext://sys.stdout"),

473
474
475
476
    # if set, VLLM_LOGGING_PREFIX will be prepended to all log messages
    "VLLM_LOGGING_PREFIX":
    lambda: os.getenv("VLLM_LOGGING_PREFIX", ""),

477
478
479
480
481
482
483
484
    # 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,

485
486
487
488
489
490
    # If set, vllm will log stats at this interval in seconds
    # If not set, vllm will log stats every 10 seconds.
    "VLLM_LOG_STATS_INTERVAL":
    lambda: val if (val := float(os.getenv("VLLM_LOG_STATS_INTERVAL", "10.")))
        > 0. else 10.,

491
492
493
494
495
496
497
498
499
500
501
502
    # 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
503
    # - "FLASHINFER": use flashinfer
504
    # - "FLASHMLA": use FlashMLA
505
    # - "FLASH_ATTN_MLA": use FlashAttention for MLA
506
507
508
    "VLLM_ATTENTION_BACKEND":
    lambda: os.getenv("VLLM_ATTENTION_BACKEND", None),

509
510
    # If set, vllm will use flashinfer sampler
    "VLLM_USE_FLASHINFER_SAMPLER":
511
512
    lambda: bool(int(os.environ["VLLM_USE_FLASHINFER_SAMPLER"]))
    if "VLLM_USE_FLASHINFER_SAMPLER" in os.environ else None,
513

514
515
516
517
    # Pipeline stage partition strategy
    "VLLM_PP_LAYER_PARTITION":
    lambda: os.getenv("VLLM_PP_LAYER_PARTITION", None),

518
    # (CPU backend only) CPU key-value cache space.
519
    # default is None and will be set as 4 GB
520
    "VLLM_CPU_KVCACHE_SPACE":
521
522
    lambda: int(os.getenv("VLLM_CPU_KVCACHE_SPACE", "0"))
    if "VLLM_CPU_KVCACHE_SPACE" in os.environ else None,
523

524
525
526
    # (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":
527
528
529
530
531
    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":
532
533
    lambda: int(os.getenv("VLLM_CPU_NUM_OF_RESERVED_CPU", "0"))
    if "VLLM_CPU_NUM_OF_RESERVED_CPU" in os.environ else None,
534

535
536
537
538
539
    # (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"))),
540

541
542
543
544
    # (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"))),

545
546
547
548
549
    # 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":
550
    lambda: bool(int(os.getenv("VLLM_USE_RAY_SPMD_WORKER", "0"))),
551

552
553
554
    # If the env var is set, it uses the Ray's Compiled Graph
    # (previously known as ADAG) API which optimizes the
    # control plane overhead.
555
    # Run vLLM with VLLM_USE_RAY_COMPILED_DAG=1 to enable it.
556
557
    # Note that this variable is set to 1 in V1 by default
    # when ray distributed executor is used.
558
    "VLLM_USE_RAY_COMPILED_DAG":
559
560
    lambda: bool(int(os.getenv("VLLM_USE_RAY_COMPILED_DAG", "0"))),

561
562
563
564
565
566
567
568
569
570
    # 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"),
571

572
    # If the env var is set, it enables GPU communication overlap
573
    # (experimental feature) in Ray's Compiled Graph. This flag is ignored if
574
575
    # VLLM_USE_RAY_COMPILED_DAG is not set.
    "VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM":
576
    lambda: bool(int(os.getenv("VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM", "0"))
577
578
                 ),

579
580
581
582
583
584
585
    # 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"))),

586
587
588
    # Use dedicated multiprocess context for workers.
    # Both spawn and fork work
    "VLLM_WORKER_MULTIPROC_METHOD":
589
    lambda: os.getenv("VLLM_WORKER_MULTIPROC_METHOD", "spawn"),
590

591
592
593
594
595
596
597
598
    # 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"),
        )),

599
600
601
602
    # 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")),
603

604
    # Timeout for fetching videos when serving multimodal models
605
    # Default is 30 seconds
606
    "VLLM_VIDEO_FETCH_TIMEOUT":
607
    lambda: int(os.getenv("VLLM_VIDEO_FETCH_TIMEOUT", "30")),
608

609
    # Timeout for fetching audio when serving multimodal models
610
    # Default is 10 seconds
611
    "VLLM_AUDIO_FETCH_TIMEOUT":
612
    lambda: int(os.getenv("VLLM_AUDIO_FETCH_TIMEOUT", "10")),
613

614
615
616
617
618
619
    # Max number of workers for the thread pool handling
    # media bytes loading. Set to 1 to disable parallel processing.
    # Default is 8
    "VLLM_MEDIA_LOADING_THREAD_COUNT":
    lambda: int(os.getenv("VLLM_MEDIA_LOADING_THREAD_COUNT", "8")),

620
621
622
623
624
625
    # 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")),

626
627
628
629
630
631
632
633
634
635
    # 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"),

636
    # [DEPRECATED] Cache size (in GiB per process) for multimodal input cache
637
    # Default is 4 GiB per API process + 4 GiB per engine core process
638
    "VLLM_MM_INPUT_CACHE_GIB":
639
    lambda: int(os.getenv("VLLM_MM_INPUT_CACHE_GIB", "4")),
640

641
642
643
    # Path to the XLA persistent cache directory.
    # Only used for XLA devices such as TPUs.
    "VLLM_XLA_CACHE_PATH":
644
645
    lambda: os.path.expanduser(
        os.getenv(
646
            "VLLM_XLA_CACHE_PATH",
647
648
            os.path.join(get_default_cache_root(), "vllm", "xla_cache"),
        )),
649
650
651
652

    # If set, assert on XLA recompilation after each execution step.
    "VLLM_XLA_CHECK_RECOMPILATION":
    lambda: bool(int(os.getenv("VLLM_XLA_CHECK_RECOMPILATION", "0"))),
653
654
655
656

    # Enable SPMD mode for TPU backend.
    "VLLM_XLA_USE_SPMD":
    lambda: bool(int(os.getenv("VLLM_XLA_USE_SPMD", "0"))),
657
    "VLLM_FUSED_MOE_CHUNK_SIZE":
658
    lambda: int(os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", "32768")),
659
660
661
662
663
664
    # 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"))),
665

666
667
668
669
670
    # 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)),

671
672
673
674
675
676
677
678
    # 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")),
679
680
681
682
683
684
685

    # 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")),
686
687
    "VLLM_TEST_FORCE_LOAD_FORMAT":
    lambda: os.getenv("VLLM_TEST_FORCE_LOAD_FORMAT", "dummy"),
688

689
690
    # Time in ms for the zmq client to wait for a response from the backend
    # server for simple data operations
691
692
    "VLLM_RPC_TIMEOUT":
    lambda: int(os.getenv("VLLM_RPC_TIMEOUT", "10000")),
693

694
695
696
697
    # 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")),

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

705
706
707
708
709
710
    # 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),

711
712
713
714
    # Enables torch profiler if set.
    # Both AsyncLLM's CPU traces as well as workers'
    # traces (CPU & GPU) will be saved under this directory.
    # Note that it must be an absolute path.
715
716
    "VLLM_TORCH_PROFILER_DIR":
    lambda: (None if os.getenv("VLLM_TORCH_PROFILER_DIR", None) is None else os
717
718
             .path.abspath(os.path.expanduser(os.getenv(
        "VLLM_TORCH_PROFILER_DIR", ".")))),
719

720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
    # 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"),

745
746
747
    # If set, vLLM will use Triton implementations of AWQ.
    "VLLM_USE_TRITON_AWQ":
    lambda: bool(int(os.getenv("VLLM_USE_TRITON_AWQ", "0"))),
748
749
750
751
752
753

    # 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")),
754

755
756
757
758
759
760
    # 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
761
    "VLLM_SKIP_P2P_CHECK":
762
    lambda: os.getenv("VLLM_SKIP_P2P_CHECK", "1") == "1",
763

764
765
766
767
768
769
770
    # 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(","),
771
772
773

    # If set, use the V1 code path.
    "VLLM_USE_V1":
zhuwenwen's avatar
zhuwenwen committed
774
    lambda: bool(int(os.getenv("VLLM_USE_V1", "1"))),
775

776
777
778
779
780
    # 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")),
781

782
783
784
785
786
787
    # 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")),

788
789
790
791
792
793
794
    # 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")),

795
796
797
798
799
800
    # 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")),

801
802
803
804
805
    # 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")),

806
807
808
809
810
    # 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")),
811
812
813
814
815
816
817

    # 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")),

818
819
820
821
822
823
    # Whether to use aiter triton fp8 bmm kernel
    # By default is enabled.
    "VLLM_ROCM_USE_AITER_FP8BMM":
    lambda: (os.getenv("VLLM_ROCM_USE_AITER_FP8BMM", "True").lower() in
             ("true", "1")),

824
825
826
827
828
    # use rocm skinny gemms
    "VLLM_ROCM_USE_SKINNY_GEMM":
    lambda: (os.getenv("VLLM_ROCM_USE_SKINNY_GEMM", "True").lower() in
             ("true", "1")),

829
830
831
    # Pad the fp8 weights to 256 bytes for ROCm
    "VLLM_ROCM_FP8_PADDING":
    lambda: bool(int(os.getenv("VLLM_ROCM_FP8_PADDING", "1"))),
832

833
834
    # Pad the weights for the moe kernel
    "VLLM_ROCM_MOE_PADDING":
zhuwenwen's avatar
zhuwenwen committed
835
    lambda: bool(int(os.getenv("VLLM_ROCM_MOE_PADDING", "0"))),
836

837
838
839
840
841
    # 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")),

842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
    # 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)),

867
868
869
    # Divisor for dynamic query scale factor calculation for FP8 KV Cache
    "Q_SCALE_CONSTANT":
    lambda: int(os.getenv("Q_SCALE_CONSTANT", "200")),
870
871
872
873
874
875
    # 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")),
876

877
878
    # If set, enable multiprocessing in LLM for the V1 code path.
    "VLLM_ENABLE_V1_MULTIPROCESSING":
879
    lambda: bool(int(os.getenv("VLLM_ENABLE_V1_MULTIPROCESSING", "1"))),
880
881
    "VLLM_LOG_BATCHSIZE_INTERVAL":
    lambda: float(os.getenv("VLLM_LOG_BATCHSIZE_INTERVAL", "-1")),
882
883
    "VLLM_DISABLE_COMPILE_CACHE":
    lambda: bool(int(os.getenv("VLLM_DISABLE_COMPILE_CACHE", "0"))),
884
885
886
887
888
889

    # 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"))),
890
891
892
893
894
895
896
897
898
899

    # 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")),
900
901
902
903
904

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

905
906
907
908
909
910
911
912
913
914
915
916
    # 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", ""),

917
918
919
920
    # 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),
921

922
923
924
925
    # Rank of the process in the data parallel setting
    "VLLM_DP_RANK":
    lambda: int(os.getenv("VLLM_DP_RANK", "0")),

926
927
928
929
930
931
    # 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)),

932
933
934
935
936
937
938
939
940
941
942
    # 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")),
943

944
945
946
947
948
949
950
951
    # 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")),

952
953
954
955
    # 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",

956
957
958
    # 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",
959

960
    # Use model_redirect to redirect the model name to a local folder.
961
962
963
964
965
    # `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
966
967
968
    "VLLM_MODEL_REDIRECT_PATH":
    lambda: os.environ.get("VLLM_MODEL_REDIRECT_PATH", None),

969
970
971
    # 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",
972

973
974
975
976
    # Whether to use marlin kernel in mxfp4 quantization method
    "VLLM_MXFP4_USE_MARLIN":
    lambda: maybe_convert_bool(os.environ.get("VLLM_MXFP4_USE_MARLIN", None)),

977
978
979
980
981
    # 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",
982

983
984
985
986
987
988
    # 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",

989
990
991
992
    # 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"])
993
    if "VLLM_TPU_BUCKET_PADDING_GAP" in os.environ else 0,
994
995
    "VLLM_TPU_MOST_MODEL_LEN":
    lambda: maybe_convert_int(os.environ.get("VLLM_TPU_MOST_MODEL_LEN", None)),
996

997
998
999
1000
    # Whether using Pathways
    "VLLM_TPU_USING_PATHWAYS":
    lambda: bool("proxy" in os.getenv("JAX_PLATFORMS", "").lower()),

1001
1002
1003
    # Allow use of DeepGemm kernels for fused moe ops.
    "VLLM_USE_DEEP_GEMM":
    lambda: bool(int(os.getenv("VLLM_USE_DEEP_GEMM", "0"))),
1004

1005
1006
1007
    # Whether to use E8M0 scaling when DeepGEMM is used on Blackwell GPUs.
    "VLLM_USE_DEEP_GEMM_E8M0":
    lambda: bool(int(os.getenv("VLLM_USE_DEEP_GEMM_E8M0", "1"))),
1008
1009
1010
1011
    # TODO(wentao): unify the two E8M0 flags after verifying the correctness.
    # Whether to use E8M0 scaling when DeepGEMM is used on Hopper GPUs.
    "VLLM_USE_DEEP_GEMM_E8M0_HOPPER":
    lambda: bool(int(os.getenv("VLLM_USE_DEEP_GEMM_E8M0_HOPPER", "0"))),
1012
1013
1014
1015
1016
1017
1018
1019
    # 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"))),

1020
1021
1022
1023
    # Whether to use fused grouped_topk used for MoE expert selection.
    "VLLM_USE_FUSED_MOE_GROUPED_TOPK":
    lambda: bool(int(os.getenv("VLLM_USE_FUSED_MOE_GROUPED_TOPK", "1"))),

1024
1025
1026
1027
    # 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"))),

1028
    # Allow use of FlashInfer CUTLASS kernels for fused moe ops.
1029
1030
    "VLLM_USE_FLASHINFER_MOE_FP4":
    lambda: bool(int(os.getenv("VLLM_USE_FLASHINFER_MOE_FP4", "0"))),
1031

1032
1033
1034
1035
1036
    # 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"))),

1037
1038
1039
1040
1041
1042
1043
1044
1045
    # If set to 1, use the FlashInfer CUTLASS backend for
    # MXFP8 (activation) x MXFP4 (weight) MoE.
    # This is separate from the TRTLLMGEN path controlled by
    # VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8.
    "VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS":
    lambda: bool(int(
        os.getenv("VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS", "0")
        )),

1046
1047
1048
1049
1050
    # 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"))),

1051
1052
1053
1054
1055
    # 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")),
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065

    # 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")),
1066
1067
1068
1069
1070
1071

    # 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
1072
1073
1074
1075
1076
1077
1078
1079

    # 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")),
1080
1081

    # all2all backend for vllm's expert parallel communication
1082
1083
1084
    # Available options:
    # - "naive": naive all2all implementation using all-reduce
    # - "pplx": use pplx kernels
1085
1086
    # - "deepep_high_throughput", use deepep high-throughput kernels
    # - "deepep_low_latency", use deepep low-latency kernels
1087
1088
    "VLLM_ALL2ALL_BACKEND":
    lambda: os.getenv("VLLM_ALL2ALL_BACKEND", "naive"),
1089

1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
    # 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"
    ),

1104
1105
1106
1107
1108
1109
    # 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")),
1110

1111
1112
1113
1114
    # Specifies the thresholds of the communicated tensor sizes under which
    # vllm should use flashinfer fused allreduce. The variable should be a
    # JSON with the following format:
    #     { <world size>: <max size in mb> }
1115
    # Unspecified world sizes will fall back to
1116
1117
1118
1119
1120
    #     { 2: 64, 4: 1, <everything else>: 0.5 }
    "VLLM_FLASHINFER_ALLREDUCE_FUSION_THRESHOLDS_MB":
    lambda: json.loads(os.getenv(
        "VLLM_FLASHINFER_ALLREDUCE_FUSION_THRESHOLDS_MB", "{}")),

1121
1122
1123
1124
1125
1126
1127
1128
1129
    # 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(),

1130
1131
1132
    # 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")),
1133
1134
1135
1136
1137

    # 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"))),
1138
1139
1140
1141
1142
1143

    # 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")),
1144
    
1145
1146
1147
1148
1149
    # 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")),

1150
1151
1152
1153
1154
1155
1156
1157
    # 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":
1158
1159
1160
1161
1162
1163
1164
    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"))),
1165
1166
1167
1168
1169

    # Controls whether or not emulations are used for NVFP4
    # generations on machines < 100 for compressed-tensors
    # models
    "VLLM_USE_NVFP4_CT_EMULATIONS":
zhuwenwen's avatar
zhuwenwen committed
1170
1171
    lambda: bool(int(os.getenv("VLLM_USE_NVFP4_CT_EMULATIONS", "0"))),
    
1172
1173
1174
1175
1176
    # 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":
1177
1178
    lambda: int(os.getenv("VLLM_NIXL_ABORT_REQUEST_TIMEOUT", "120")),

1179
1180
1181
1182
    # Controls whether or not to use cudnn prefill
    "VLLM_USE_CUDNN_PREFILL":
    lambda: bool(int(os.getenv("VLLM_USE_CUDNN_PREFILL", "0"))),

1183
1184
1185
    # If set to 1, use the TRTLLM attention backend in flashinfer.
    "VLLM_USE_TRTLLM_ATTENTION":
    lambda: os.getenv("VLLM_USE_TRTLLM_ATTENTION", None),
1186

1187
1188
1189
1190
    # If set to 1, when we use fp8 kv, we do not quantize Q to fp8
    "VLLM_FLASHINFER_DISABLE_Q_QUANTIZATION":
    lambda: bool(int(os.getenv("VLLM_FLASHINFER_DISABLE_Q_QUANTIZATION", "0"))),

1191
1192
1193
1194
1195
    # If set, it means we pre-downloaded cubin files and flashinfer will
    # read the cubin files directly.
    "VLLM_HAS_FLASHINFER_CUBIN":
    lambda: os.getenv("VLLM_HAS_FLASHINFER_CUBIN", False),

1196
1197
1198
1199
1200
    # If set to 1, force the use of TRTLLM FP4 GEMM backend in flashinfer.
    # Otherwise, uses the first available of: flashinfer cutlass GEMM,
    # vllm cutlass GEMM, marlin GEMM.
    "VLLM_USE_TRTLLM_FP4_GEMM":
    lambda: bool(int(os.getenv("VLLM_USE_TRTLLM_FP4_GEMM", "0"))),
1201

1202
1203
1204
1205
1206
1207
    # 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"))),

1208
1209
1210
1211
1212
1213
    # Disable padding to CUDA graph capture batch sizes.
    # TODO(wentao): https://github.com/vllm-project/vllm/issues/23378
    # After the issue is fixed, we can remove this flag.
    "VLLM_DISABLE_PAD_FOR_CUDAGRAPH":
    lambda: bool(int(os.getenv("VLLM_DISABLE_PAD_FOR_CUDAGRAPH", "0"))),

1214
1215
1216
    # Used to force set up loopback IP
    "VLLM_LOOPBACK_IP":
    lambda: os.getenv("VLLM_LOOPBACK_IP", ""),
1217
    
1218
1219
1220
1221
1222
    # 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"),
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233

    # 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"))),
1234
1235
1236

    # 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
1237
1238
    # messages for those requests in memory. By default, this is disabled (0),
    # and the "store" option is ignored.
1239
1240
1241
1242
1243
1244
1245
    # 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"))),
1246

1247
1248
1249
1250
    # Whether to use pytorch symmetric memory for allreduce
    "VLLM_ALLREDUCE_USE_SYMM_MEM":
    lambda: bool(int(os.getenv("VLLM_ALLREDUCE_USE_SYMM_MEM", "0"))),

1251
1252
1253
    # Allows vllm to find tuned config under customized folder
    "VLLM_TUNED_CONFIG_FOLDER":
    lambda: os.getenv("VLLM_TUNED_CONFIG_FOLDER", None),
1254
    
1255
1256
1257
1258
1259
1260
1261
1262
1263
    # Allows vllm use container tool
    "VLLM_GPT_OSS_USE_CONTAINER_TOOL":
    lambda: bool(int(os.getenv("VLLM_GPT_OSS_USE_CONTAINER_TOOL", "0"))),

    # Allows harmony instructions to be injected on system messages
    "VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS":
    lambda: bool(
        int(os.getenv("VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS", "0"))),

1264
1265
1266
    # Add optional custom scopes for profiling, disable to avoid overheads
    "VLLM_CUSTOM_SCOPES_FOR_PROFILING":
    lambda: bool(int(os.getenv("VLLM_CUSTOM_SCOPES_FOR_PROFILING", "0"))),
1267
1268
1269
1270
1271

    # Represent block hashes in KV cache events as 64-bit integers instead of
    # raw bytes. Defaults to True for backward compatibility.
    "VLLM_KV_EVENTS_USE_INT_BLOCK_HASHES":
    lambda: bool(int(os.getenv("VLLM_KV_EVENTS_USE_INT_BLOCK_HASHES", "1"))),
1272
    
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
     # used in optest environment to manually set the https port
    'VLLM_OPTEST_URLS_PORT':
    lambda: int(os.getenv('VLLM_OPTEST_URLS_PORT', '8000'))
    if 'VLLM_OPTEST_URLS_PORT' in os.environ else None,
    
    # Path to the optest models.
    # If set, will load models from local path instead of Hugging Face Hub.
    'VLLM_OPTEST_MODELS_PATH':
    lambda: os.getenv('VLLM_OPTEST_MODELS_PATH', "") or os.getenv("OPTEST_MODELS_PATH", ""),
    
    # flag to control if vllm should use triton prefix flash attention
    "VLLM_USE_TRITON_PREFIX_FLASH_ATTN":
    lambda: (os.environ.get("VLLM_USE_TRITON_PREFIX_FLASH_ATTN", "False").lower() in
             ("true", "1")),
    
zhuwenwen's avatar
zhuwenwen committed
1288
1289
1290
1291
1292
1293
1294
1295
    # If set, vLLM will use optimized MLA attention optimizations.
    "VLLM_USE_TRITON_OPT_MLA":
    lambda: bool(int(os.getenv("VLLM_USE_TRITON_OPT_MLA", "0"))),
    
    # If set, vLLM will use FLASH MLA attention optimizations.
    "VLLM_USE_FLASH_MLA":
    lambda: bool(int(os.getenv("VLLM_USE_FLASH_MLA", "1"))),
    
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
    # flag to control vllm to use optimized kernels
    "VLLM_USE_OPT_OP":
    lambda: (os.environ.get("VLLM_USE_OPT_OP", "True").lower() in
             ("true", "1")),
    
    # flag to control vllm to use optimized tc paged attn kernels
    "VLLM_USE_TC_PAGED_ATTN":
    lambda: (os.environ.get("VLLM_USE_TC_PAGED_ATTN", "True").lower() in
             ("true", "1")),
    
    # flag to control if vllm print pa parameters
    "VLLM_USE_PA_PRINT_PARAM":
    lambda: (os.environ.get("VLLM_USE_PA_PRINT_PARAM", "False").lower() in
             ("true", "1")),
    
1311
1312
1313
1314
1315
1316
    # If set, vLLM will use tree-style speculative decoding.
    "VLLM_TREE_DECODING":
    lambda: 
    (os.environ.get("VLLM_TREE_DECODING", "0").strip().lower() in
     ("1", "true")),
    
1317
1318
1319
1320
1321
1322
1323
1324
    # If set, vLLM will disable the draft model in cudagraph mode.
    "VLLM_SPEC_DECODE_EAGER":
    lambda: bool(int(os.getenv("VLLM_SPEC_DECODE_EAGER", "0"))),
    
    # flag to control vllm to use optimized kernels
    "VLLM_PCIE_USE_CUSTOM_ALLREDUCE":
    lambda: bool(int(os.environ.get("VLLM_PCIE_USE_CUSTOM_ALLREDUCE", "0"))),
    
zhuwenwen's avatar
zhuwenwen committed
1325
1326
1327
1328
    # flag to control vllm to use optimized kernels
    "VLLM_CUSTOM_ALLREDUCE_SUPPORTED_WORLDSIZE_MAX":
    lambda: int(os.getenv("VLLM_CUSTOM_ALLREDUCE_SUPPORTED_WORLDSIZE_MAX", "16")),
    
1329
1330
1331
    # If set, vLLM will disable the draft model in cudagraph mode.
    "VLLM_ENFORCE_EAGER_BS_THRESHOLD":
    lambda: int(os.environ.get("VLLM_ENFORCE_EAGER_BS_THRESHOLD", "-1")),
1332

1333
1334
1335
1336
1337
1338
1339
    # 'has_comtext' is a variable in common.py, which is calculated
    # by metadata by default. However, it may introduce synchronization 
    # and affect performance, so it is directly assigned as False. 
    # If there are any problems during use, use environment variables 
    # to restore the default usage.
    "VLLM_HAS_CONTEXT_DEFAULT":
    lambda: bool(int(os.getenv("VLLM_HAS_CONTEXT_DEFAULT", "0"))),
1340
1341
1342
    
    # If set, vLLM will transpose weight to use nn layout
    "VLLM_USE_NN":
zhuwenwen's avatar
zhuwenwen committed
1343
    lambda: (os.environ.get("VLLM_USE_NN", "True").lower() in
1344
             ("true", "1")),
1345

1346
1347
1348
    # Enable two batch overlap.
    "VLLM_ENABLE_TBO":
    lambda: bool(int(os.getenv("VLLM_ENABLE_TBO", "0"))),
1349
    
1350
1351
1352
    # set delay on server when only one requet, the purpose is to merge a larger batch.
    "VLLM_TBO_REQ_DELAY_MS":
    lambda: int(os.getenv("VLLM_TBO_REQ_DELAY_MS", "0")),
1353

1354
1355
1356
    # set the minimum batch size to enable TBO in decode, if < 2 , disable TBO in decode.
    "VLLM_TBO_DECODE_BS":
    lambda: int(os.getenv("VLLM_TBO_DECODE_BS", "0")),
zhuwenwen's avatar
zhuwenwen committed
1357
1358
1359
1360
    
    # set the minimum tokens size for each mini-batch to enable TBO on v1, default is 200.
    "VLLM_TBO_MIN_TOKENS":
    lambda: int(os.getenv("VLLM_TBO_MIN_TOKENS", "200")),
1361

1362
1363
1364
    # Enable zero overhead scheduler.
    "VLLM_ZERO_OVERHEAD":
    lambda: bool(int(os.getenv("VLLM_ZERO_OVERHEAD", "0"))),
1365
1366
1367

    # If set, vLLM will enable the moe_fused_gate kernel.
    "VLLM_ENABLE_MOE_FUSED_GATE":
zhuwenwen's avatar
zhuwenwen committed
1368
    lambda: bool(int(os.getenv("VLLM_ENABLE_MOE_FUSED_GATE", "1"))),
zhuwenwen's avatar
zhuwenwen committed
1369
    
1370
1371
    # vLLM will use FlashAttention Backend for page attention computation on rocm
    "VLLM_USE_FLASH_ATTN_PA":
zhuwenwen's avatar
zhuwenwen committed
1372
    lambda: (os.environ.get("VLLM_USE_FLASH_ATTN_PA", "True").lower() in
zhuwenwen's avatar
zhuwenwen committed
1373
             ("true", "1")),
zhuwenwen's avatar
zhuwenwen committed
1374
1375
1376
1377
1378
    
    # vLLM will use apex for rmsnorm
    "VLLM_USE_APEX_RN":
    lambda: (os.environ.get("VLLM_USE_APEX_RN", "False").lower() in
             ("true", "1")),
1379
1380
1381
    
    # vLLM will use global cache for moe
    "VLLM_USE_GLOBAL_CACHE13":
1382
        lambda: (os.environ.get("VLLM_USE_GLOBAL_CACHE13", "False").lower() in
1383
                 ("true", "1")),
1384
1385
1386
1387
1388
        
    # vLLM will use lightop for moe_fused_gate and moe_align_block_size
    "VLLM_USE_LIGHT_OP":
        lambda: (os.environ.get("VLLM_USE_LIGHT_OP", "False").lower() in
                 ("true", "1")),
1389
1390
1391
1392
1393
        
    # vLLM will use global cache for moe
    "VLLM_USE_TRITON_CAT":
        lambda: (os.environ.get("VLLM_USE_TRITON_CAT", "False").lower() in
                 ("true", "1")),  
1394
1395
}

1396
# --8<-- [end:env-vars-definition]
1397

1398

1399
def __getattr__(name: str):
1400
1401
1402
1403
1404
1405
1406
1407
    # 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())
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423


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"
1424
1425
1426
1427
1428
1429
1430
1431


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
1432
    graphs, so it is included in the factors list. The env vars that
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
    affect the choice of different kernels or attention backends should
    also be included in the factors list.
    """

    # The values of envs may affects the computation graph.
    # TODO(DefTruth): hash all environment variables?
    # for key in environment_variables:
    #     factorize(key)
    environment_variables_to_hash = [
        "VLLM_PP_LAYER_PARTITION",
        "VLLM_MLA_DISABLE",
        "VLLM_USE_TRITON_FLASH_ATTN",
        "VLLM_USE_TRITON_AWQ",
        "VLLM_DP_RANK",
        "VLLM_DP_SIZE",
1448
        "VLLM_USE_STANDALONE_COMPILE",
1449
        "VLLM_FUSED_MOE_CHUNK_SIZE",
1450
1451
1452
1453
1454
1455
1456
        "VLLM_FLASHINFER_MOE_BACKEND",
        "VLLM_V1_USE_PREFILL_DECODE_ATTENTION",
        "VLLM_USE_AITER_UNIFIED_ATTENTION",
        "VLLM_ATTENTION_BACKEND",
        "VLLM_USE_FLASHINFER_SAMPLER",
        "VLLM_DISABLED_KERNELS",
        "VLLM_USE_DEEP_GEMM",
1457
1458
        "VLLM_USE_DEEP_GEMM_E8M0",
        "VLLM_USE_DEEP_GEMM_E8M0_HOPPER",
1459
        "VLLM_USE_TRTLLM_FP4_GEMM",
1460
        "VLLM_USE_FUSED_MOE_GROUPED_TOPK",
1461
1462
1463
        "VLLM_USE_FLASHINFER_MOE_FP8",
        "VLLM_USE_FLASHINFER_MOE_FP4",
        "VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8",
1464
        "VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS",
1465
1466
1467
        "VLLM_USE_FLASHINFER_MOE_MXFP4_BF16",
        "VLLM_USE_CUDNN_PREFILL",
        "VLLM_USE_TRTLLM_ATTENTION",
1468
        "VLLM_FLASHINFER_DISABLE_Q_QUANTIZATION",
1469
1470
1471
1472
1473
1474
1475
        "VLLM_ROCM_USE_AITER",
        "VLLM_ROCM_USE_AITER_PAGED_ATTN",
        "VLLM_ROCM_USE_AITER_LINEAR",
        "VLLM_ROCM_USE_AITER_MOE",
        "VLLM_ROCM_USE_AITER_RMSNORM",
        "VLLM_ROCM_USE_AITER_MLA",
        "VLLM_ROCM_USE_AITER_MHA",
1476
        "VLLM_ROCM_USE_AITER_FP8BMM",
1477
1478
1479
1480
1481
1482
1483
        "VLLM_ROCM_USE_SKINNY_GEMM",
        "VLLM_ROCM_FP8_PADDING",
        "VLLM_ROCM_MOE_PADDING",
        "VLLM_ROCM_CUSTOM_PAGED_ATTN",
        "VLLM_ROCM_QUICK_REDUCE_QUANTIZATION",
        "VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16",
        "VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB",
1484
1485
    ]
    for key in environment_variables_to_hash:
1486
1487
1488
1489
1490
1491
1492
1493
        # if this goes out of sync with environment_variables,
        # it's not a user error, it's a bug
        assert key in environment_variables, \
            "Please update environment_variables_to_hash in envs.py"

    factors = [
        environment_variables[key]() for key in environment_variables_to_hash
    ]
1494

1495
1496
    hash_str = hashlib.md5(str(factors).encode(),
                           usedforsecurity=False).hexdigest()
1497

1498
    return hash_str