envs.py 46.7 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

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

if TYPE_CHECKING:
    VLLM_HOST_IP: str = ""
12
    VLLM_PORT: Optional[int] = None
13
    VLLM_RPC_BASE_PATH: str = tempfile.gettempdir()
14
    VLLM_USE_MODELSCOPE: bool = False
15
    VLLM_RINGBUFFER_WARNING_INTERVAL: int = 60
16
17
    VLLM_NCCL_SO_PATH: Optional[str] = None
    LD_LIBRARY_PATH: Optional[str] = None
18
    VLLM_USE_TRITON_FLASH_ATTN: bool = True
19
    VLLM_V1_USE_PREFILL_DECODE_ATTENTION: bool = False
20
    VLLM_FLASH_ATTN_VERSION: Optional[int] = None
21
22
23
24
25
26
27
    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
28
    VLLM_MODEL_REDIRECT_PATH: Optional[str] = None
29
30
    VLLM_CACHE_ROOT: str = os.path.expanduser("~/.cache/vllm")
    VLLM_CONFIG_ROOT: str = os.path.expanduser("~/.config/vllm")
31
32
33
34
35
    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
36
    VLLM_LOGGING_LEVEL: str = "INFO"
37
    VLLM_LOGGING_PREFIX: str = ""
38
    VLLM_LOGGING_CONFIG_PATH: Optional[str] = None
39
    VLLM_LOGITS_PROCESSOR_THREADS: Optional[int] = None
40
41
    VLLM_TRACE_FUNCTION: int = 0
    VLLM_ATTENTION_BACKEND: Optional[str] = None
42
    VLLM_USE_FLASHINFER_SAMPLER: Optional[bool] = None
43
    VLLM_FLASHINFER_FORCE_TENSOR_CORES: bool = False
44
    VLLM_PP_LAYER_PARTITION: Optional[str] = None
45
    VLLM_CPU_KVCACHE_SPACE: int = 0
46
    VLLM_CPU_OMP_THREADS_BIND: str = ""
47
    VLLM_CPU_NUM_OF_RESERVED_CPU: int = 0
48
    VLLM_CPU_MOE_PREPACK: bool = True
49
    VLLM_CPU_SGL_KERNEL: bool = False
50
    VLLM_XLA_CACHE_PATH: str = os.path.join(VLLM_CACHE_ROOT, "xla_cache")
51
    VLLM_XLA_CHECK_RECOMPILATION: bool = False
52
    VLLM_FUSED_MOE_CHUNK_SIZE: int = 64 * 1024
53
    VLLM_ENABLE_FUSED_MOE_ACTIVATION_CHUNKING: bool = True
54
    VLLM_USE_RAY_SPMD_WORKER: bool = False
55
    VLLM_USE_RAY_COMPILED_DAG: bool = False
56
    VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE: str = "auto"
57
    VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM: bool = False
58
    VLLM_XLA_USE_SPMD: bool = False
59
    VLLM_WORKER_MULTIPROC_METHOD: str = "spawn"
60
    VLLM_ASSETS_CACHE: str = os.path.join(VLLM_CACHE_ROOT, "assets")
61
    VLLM_IMAGE_FETCH_TIMEOUT: int = 5
62
    VLLM_VIDEO_FETCH_TIMEOUT: int = 30
63
    VLLM_AUDIO_FETCH_TIMEOUT: int = 10
64
    VLLM_VIDEO_LOADER_BACKEND: str = "opencv"
65
    VLLM_MM_INPUT_CACHE_GIB: int = 8
66
67
68
69
    VLLM_TARGET_DEVICE: str = "cuda"
    MAX_JOBS: Optional[str] = None
    NVCC_THREADS: Optional[str] = None
    VLLM_USE_PRECOMPILED: bool = False
70
    VLLM_TEST_USE_PRECOMPILED_NIGHTLY_WHEEL: bool = False
71
    VLLM_NO_DEPRECATION_WARNING: bool = False
72
    VLLM_KEEP_ALIVE_ON_ENGINE_DEATH: bool = False
73
74
    CMAKE_BUILD_TYPE: Optional[str] = None
    VERBOSE: bool = False
75
    VLLM_ALLOW_LONG_MAX_MODEL_LEN: bool = False
76
    VLLM_RPC_TIMEOUT: int = 10000  # ms
77
    VLLM_HTTP_TIMEOUT_KEEP_ALIVE: int = 5  # seconds
78
    VLLM_PLUGINS: Optional[list[str]] = None
79
    VLLM_LORA_RESOLVER_CACHE_DIR: Optional[str] = None
80
    VLLM_TORCH_PROFILER_DIR: Optional[str] = None
81
    VLLM_USE_TRITON_AWQ: bool = False
82
    VLLM_ALLOW_RUNTIME_LORA_UPDATING: bool = False
83
    VLLM_TREE_DECODING: bool = False
84
    VLLM_SKIP_P2P_CHECK: bool = False
85
    VLLM_DISABLED_KERNELS: list[str] = []
86
    VLLM_USE_V1: bool = True
87
    VLLM_ROCM_USE_AITER: bool = False
88
    VLLM_ROCM_USE_AITER_PAGED_ATTN: bool = False
89
    VLLM_ROCM_USE_AITER_LINEAR: bool = True
90
    VLLM_ROCM_USE_AITER_MOE: bool = True
91
    VLLM_ROCM_USE_AITER_RMSNORM: bool = True
92
    VLLM_ROCM_USE_AITER_MLA: bool = True
93
    VLLM_ROCM_USE_AITER_MHA: bool = True
94
    VLLM_ROCM_USE_SKINNY_GEMM: bool = True
95
    VLLM_ROCM_FP8_PADDING: bool = True
96
    VLLM_ROCM_MOE_PADDING: bool = True
97
    VLLM_ROCM_CUSTOM_PAGED_ATTN: bool = True
98
    VLLM_QUARK_EMU_MEM_OPT: bool = False
99
    VLLM_ENABLE_V1_MULTIPROCESSING: bool = True
100
    VLLM_LOG_BATCHSIZE_INTERVAL: float = -1
101
    VLLM_DISABLE_COMPILE_CACHE: bool = False
102
    Q_SCALE_CONSTANT: int = 200
103
104
    K_SCALE_CONSTANT: int = 200
    V_SCALE_CONSTANT: int = 100
105
    VLLM_SERVER_DEV_MODE: bool = False
106
    VLLM_V1_OUTPUT_PROC_CHUNK_SIZE: int = 128
107
    VLLM_MLA_DISABLE: bool = False
108
109
    VLLM_RAY_PER_WORKER_GPUS: float = 1.0
    VLLM_RAY_BUNDLE_INDICES: str = ""
110
    VLLM_CUDART_SO_PATH: Optional[str] = None
111
    VLLM_USE_HPU_CONTIGUOUS_CACHE_FETCH: bool = True
112
    VLLM_HPU_USE_DELAYED_SAMPLING: bool = False
113
    VLLM_DP_RANK: int = 0
114
    VLLM_DP_RANK_LOCAL: int = -1
115
116
117
    VLLM_DP_SIZE: int = 1
    VLLM_DP_MASTER_IP: str = ""
    VLLM_DP_MASTER_PORT: int = 0
118
    VLLM_MOE_DP_CHUNK_SIZE: int = 256
119
    VLLM_RANDOMIZE_DP_DUMMY_INPUTS: bool = False
120
    VLLM_MARLIN_USE_ATOMIC_ADD: bool = False
121
    VLLM_V0_USE_OUTLINES_CACHE: bool = False
122
    VLLM_TPU_BUCKET_PADDING_GAP: int = 0
123
    VLLM_TPU_MOST_MODEL_LEN: Optional[int] = None
124
    VLLM_USE_DEEP_GEMM: bool = False
125
    VLLM_XGRAMMAR_CACHE_MB: int = 0
126
    VLLM_MSGPACK_ZERO_COPY_THRESHOLD: int = 256
127
    VLLM_ALLOW_INSECURE_SERIALIZATION: bool = False
Robert Shaw's avatar
Robert Shaw committed
128
129
    VLLM_NIXL_SIDE_CHANNEL_HOST: str = "localhost"
    VLLM_NIXL_SIDE_CHANNEL_PORT: int = 5557
130
    VLLM_ALL2ALL_BACKEND: str = "naive"
131
    VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE: int = 163840
132
    VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS: int = 1
133
    VLLM_SLEEP_WHEN_IDLE: bool = False
134
    VLLM_MQ_MAX_CHUNK_BYTES_MB: int = 16
135
    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: int = 300
136
    VLLM_KV_CACHE_LAYOUT: Optional[str] = None
137
    VLLM_COMPUTE_NANS_IN_LOGITS: bool = False
138
    VLLM_USE_NVFP4_CT_EMULATIONS: bool = False
139
140
141
    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
zhuwenwen's avatar
zhuwenwen committed
142
    
143
144
145
146
147
148
149
150
151
152
153
    # 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 
    VLLM_SPEC_DECODE_EAGER: bool = False
    VLLM_PCIE_USE_CUSTOM_ALLREDUCE: bool = False
zhuwenwen's avatar
zhuwenwen committed
154
    VLLM_CUSTOM_ALLREDUCE_SUPPORTED_WORLDSIZE_MAX: int = 16
155
156
    VLLM_ENFORCE_EAGER_BS_THRESHOLD: Optional[int] = None
    VLLM_HAS_CONTEXT_DEFAULT: bool = False
157
    VLLM_USE_NN: bool = False
158
    VLLM_ENABLE_TBO: bool = False
159
160
    VLLM_TBO_REQ_DELAY_MS: int = 0
    VLLM_TBO_DECODE_BS: int = 0
lizhigong's avatar
lizhigong committed
161
    VLLM_TBO_MIN_TOKENS: int = 200
162
    VLLM_ZERO_OVERHEAD: bool = False
163
    VLLM_ENABLE_MOE_FUSED_GATE: bool = False
164
    VLLM_USE_FLASH_ATTN_PA: bool = False
zhuwenwen's avatar
zhuwenwen committed
165
    VLLM_USE_APEX_RN: bool = False
166
    VLLM_USE_GLOBAL_CACHE13: bool = False
王敏's avatar
王敏 committed
167
    VLLM_USE_ALLTOALL_EP: bool = False
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182

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"),
    )


183
184
185
186
187
188
def maybe_convert_int(value: Optional[str]) -> Optional[int]:
    if value is None:
        return None
    return int(value)


189
190
def get_vllm_port() -> Optional[int]:
    """Get the port from VLLM_PORT environment variable.
191

192
193
    Returns:
        The port number as an integer if VLLM_PORT is set, None otherwise.
194

195
196
197
198
199
200
201
202
203
204
205
206
    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
207
208
209
210
211
212
213
        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
214
215
216
217
        raise ValueError(
            f"VLLM_PORT '{port}' must be a valid integer") from err


218
219
220
# The begin-* and end* here are used by the documentation generator
# to extract the used env vars.

221
# --8<-- [start:env-vars-definition]
222

223
environment_variables: dict[str, Callable[[], Any]] = {
224
225
226

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

227
    # Target device of vLLM, supporting [cuda (by default),
228
    # rocm, neuron, cpu]
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
    "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":
245
246
    lambda: bool(os.environ.get("VLLM_USE_PRECOMPILED")) or bool(
        os.environ.get("VLLM_PRECOMPILED_WHEEL_LOCATION")),
247

248
249
250
251
252
253
    # 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"))
                 ),

254
255
256
257
258
259
260
261
262
263
    # 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'))),

264
    # Root directory for vLLM configuration files
265
    # Defaults to `~/.config/vllm` unless `XDG_CONFIG_HOME` is set
266
267
268
269
    # 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":
270
271
272
273
274
    lambda: os.path.expanduser(
        os.getenv(
            "VLLM_CONFIG_ROOT",
            os.path.join(get_default_config_root(), "vllm"),
        )),
275
276
277

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

278
    # Root directory for vLLM cache files
279
280
281
282
283
284
285
286
    # 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"),
        )),

287
288
289
290
    # 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.
291
    'VLLM_HOST_IP':
292
    lambda: os.getenv('VLLM_HOST_IP', ""),
293

294
    # used in distributed environment to manually set the communication port
295
296
297
    # 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.
298
    'VLLM_PORT':
299
    get_vllm_port,
300

301
302
303
304
    # 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()),
305

306
307
308
309
310
    # 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",

311
312
313
314
    # 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")),

315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
    # 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":
332
    lambda: (os.environ.get("VLLM_USE_TRITON_FLASH_ATTN", "False").lower() in
333
334
             ("true", "1")),

335
336
337
338
339
340
341
    # 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")),

342
343
344
345
346
    # 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)),

347
348
349
350
351
    # Internal flag to enable Dynamo fullgraph capture
    "VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE":
    lambda: bool(
        os.environ.get("VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE", "1") != "0"),

352
353
354
355
356
    # 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",
357

358
359
360
361
362
363
364
365
366
367
368
    # 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":
369
    lambda: int(os.environ.get("VLLM_ENGINE_ITERATION_TIMEOUT_S", "120")),
370

371
    # API key for vLLM API server
372
373
374
    "VLLM_API_KEY":
    lambda: os.environ.get("VLLM_API_KEY", None),

375
376
    # Whether to log responses from API Server for debugging
    "VLLM_DEBUG_LOG_API_SERVER_RESPONSE":
377
378
    lambda: os.environ.get("VLLM_DEBUG_LOG_API_SERVER_RESPONSE", "False"
                           ).lower() == "true",
379

380
381
    # S3 access information, used for tensorizer to load model from S3
    "S3_ACCESS_KEY_ID":
382
    lambda: os.environ.get("S3_ACCESS_KEY_ID", None),
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
    "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"),

408
409
    # this is used for configuring the default logging level
    "VLLM_LOGGING_LEVEL":
410
    lambda: os.getenv("VLLM_LOGGING_LEVEL", "INFO").upper(),
411

412
413
414
415
    # if set, VLLM_LOGGING_PREFIX will be prepended to all log messages
    "VLLM_LOGGING_PREFIX":
    lambda: os.getenv("VLLM_LOGGING_PREFIX", ""),

416
417
418
419
420
421
422
423
    # 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,

424
425
426
427
428
429
430
431
432
433
434
435
    # 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
436
    # - "FLASHINFER": use flashinfer
437
    # - "FLASHMLA": use FlashMLA
438
439
440
    "VLLM_ATTENTION_BACKEND":
    lambda: os.getenv("VLLM_ATTENTION_BACKEND", None),

441
442
    # If set, vllm will use flashinfer sampler
    "VLLM_USE_FLASHINFER_SAMPLER":
443
444
    lambda: bool(int(os.environ["VLLM_USE_FLASHINFER_SAMPLER"]))
    if "VLLM_USE_FLASHINFER_SAMPLER" in os.environ else None,
445

446
447
448
449
450
    # 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"))),

451
452
453
454
    # Pipeline stage partition strategy
    "VLLM_PP_LAYER_PARTITION":
    lambda: os.getenv("VLLM_PP_LAYER_PARTITION", None),

455
    # (CPU backend only) CPU key-value cache space.
456
    # default is 4 GiB
457
458
459
    "VLLM_CPU_KVCACHE_SPACE":
    lambda: int(os.getenv("VLLM_CPU_KVCACHE_SPACE", "0")),

460
461
462
    # (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":
463
464
465
466
467
468
    lambda: os.getenv("VLLM_CPU_OMP_THREADS_BIND", "auto"),

    # (CPU backend only) CPU cores not used by OMP threads .
    # Those CPU cores will not be used by OMP threads of a rank.
    "VLLM_CPU_NUM_OF_RESERVED_CPU":
    lambda: int(os.getenv("VLLM_CPU_NUM_OF_RESERVED_CPU", "0")),
469

470
471
472
473
474
    # (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"))),
475

476
477
478
479
    # (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"))),

480
481
482
483
484
    # 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":
485
    lambda: bool(int(os.getenv("VLLM_USE_RAY_SPMD_WORKER", "0"))),
486

487
488
489
    # If the env var is set, it uses the Ray's Compiled Graph
    # (previously known as ADAG) API which optimizes the
    # control plane overhead.
490
    # Run vLLM with VLLM_USE_RAY_COMPILED_DAG=1 to enable it.
491
492
    # Note that this variable is set to 1 in V1 by default
    # when ray distributed executor is used.
493
    "VLLM_USE_RAY_COMPILED_DAG":
494
495
    lambda: bool(int(os.getenv("VLLM_USE_RAY_COMPILED_DAG", "0"))),

496
497
498
499
500
501
502
503
504
505
    # 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"),
506

507
    # If the env var is set, it enables GPU communication overlap
508
    # (experimental feature) in Ray's Compiled Graph. This flag is ignored if
509
510
    # VLLM_USE_RAY_COMPILED_DAG is not set.
    "VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM":
511
    lambda: bool(int(os.getenv("VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM", "0"))
512
513
                 ),

514
515
516
    # Use dedicated multiprocess context for workers.
    # Both spawn and fork work
    "VLLM_WORKER_MULTIPROC_METHOD":
517
    lambda: os.getenv("VLLM_WORKER_MULTIPROC_METHOD", "spawn"),
518

519
520
521
522
523
524
525
526
    # 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"),
        )),

527
528
529
530
    # 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")),
531

532
    # Timeout for fetching videos when serving multimodal models
533
    # Default is 30 seconds
534
    "VLLM_VIDEO_FETCH_TIMEOUT":
535
    lambda: int(os.getenv("VLLM_VIDEO_FETCH_TIMEOUT", "30")),
536

537
    # Timeout for fetching audio when serving multimodal models
538
    # Default is 10 seconds
539
    "VLLM_AUDIO_FETCH_TIMEOUT":
540
    lambda: int(os.getenv("VLLM_AUDIO_FETCH_TIMEOUT", "10")),
541

542
543
544
545
546
547
548
549
550
551
    # 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"),

552
    # Cache size (in GiB) for multimodal input cache
553
    # Default is 4 GiB
554
    "VLLM_MM_INPUT_CACHE_GIB":
555
    lambda: int(os.getenv("VLLM_MM_INPUT_CACHE_GIB", "4")),
556

557
558
559
    # Path to the XLA persistent cache directory.
    # Only used for XLA devices such as TPUs.
    "VLLM_XLA_CACHE_PATH":
560
561
    lambda: os.path.expanduser(
        os.getenv(
562
            "VLLM_XLA_CACHE_PATH",
563
564
            os.path.join(get_default_cache_root(), "vllm", "xla_cache"),
        )),
565
566
567
568

    # If set, assert on XLA recompilation after each execution step.
    "VLLM_XLA_CHECK_RECOMPILATION":
    lambda: bool(int(os.getenv("VLLM_XLA_CHECK_RECOMPILATION", "0"))),
569
570
571
572

    # Enable SPMD mode for TPU backend.
    "VLLM_XLA_USE_SPMD":
    lambda: bool(int(os.getenv("VLLM_XLA_USE_SPMD", "0"))),
573
    "VLLM_FUSED_MOE_CHUNK_SIZE":
574
    lambda: int(os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", "32768")),
575
576
577
578
579
580
    # 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"))),
581
582
583
584

    # If set, vllm will skip the deprecation warnings.
    "VLLM_NO_DEPRECATION_WARNING":
    lambda: bool(int(os.getenv("VLLM_NO_DEPRECATION_WARNING", "0"))),
585

586
587
588
589
590
    # 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)),

591
592
593
594
595
596
597
598
    # 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")),
599
600
601
602
603
604
605

    # 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")),
606
607
    "VLLM_TEST_FORCE_LOAD_FORMAT":
    lambda: os.getenv("VLLM_TEST_FORCE_LOAD_FORMAT", "dummy"),
608

609
610
    # Time in ms for the zmq client to wait for a response from the backend
    # server for simple data operations
611
612
    "VLLM_RPC_TIMEOUT":
    lambda: int(os.getenv("VLLM_RPC_TIMEOUT", "10000")),
613

614
615
616
617
    # 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")),

618
619
620
621
622
623
    # 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(","),
624

625
626
627
628
629
630
    # 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),

631
632
633
634
635
    # 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", "."))),
636
637
638
639

    # If set, vLLM will use Triton implementations of AWQ.
    "VLLM_USE_TRITON_AWQ":
    lambda: bool(int(os.getenv("VLLM_USE_TRITON_AWQ", "0"))),
640
641
642
643
644
645

    # 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")),
646
647
648
649
650

    # If set, vLLM will use tree-style speculative decoding.
    "VLLM_TREE_DECODING":
    lambda: 
    (os.environ.get("VLLM_TREE_DECODING", "0").strip().lower() in
zhuwenwen's avatar
zhuwenwen committed
651
     ("1", "true")),
652
653
654
655
656
657
    # 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",
658

659
660
661
662
663
664
665
    # 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(","),
666
667
668

    # If set, use the V1 code path.
    "VLLM_USE_V1":
zhuwenwen's avatar
zhuwenwen committed
669
    lambda: bool(int(os.getenv("VLLM_USE_V1", "1"))),
670

671
672
673
674
675
    # 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")),
676

677
678
679
680
681
682
    # 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")),

683
684
685
686
687
688
689
    # 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")),

690
691
692
693
694
695
    # 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")),

696
697
698
699
700
    # 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")),

701
702
703
704
705
    # 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")),
706
707
708
709
710
711
712

    # 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")),

713
714
715
716
717
    # use rocm skinny gemms
    "VLLM_ROCM_USE_SKINNY_GEMM":
    lambda: (os.getenv("VLLM_ROCM_USE_SKINNY_GEMM", "True").lower() in
             ("true", "1")),

718
719
720
    # Pad the fp8 weights to 256 bytes for ROCm
    "VLLM_ROCM_FP8_PADDING":
    lambda: bool(int(os.getenv("VLLM_ROCM_FP8_PADDING", "1"))),
721

722
723
    # Pad the weights for the moe kernel
    "VLLM_ROCM_MOE_PADDING":
zhuwenwen's avatar
zhuwenwen committed
724
    lambda: bool(int(os.getenv("VLLM_ROCM_MOE_PADDING", "0"))),
725

726
727
728
729
730
    # 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")),

731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
    # 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)),

756
757
758
759
760
761
762
763
    # 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"))),

764
765
766
    # Divisor for dynamic query scale factor calculation for FP8 KV Cache
    "Q_SCALE_CONSTANT":
    lambda: int(os.getenv("Q_SCALE_CONSTANT", "200")),
767
768
769
770
771
772
    # 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")),
773

774
775
    # If set, enable multiprocessing in LLM for the V1 code path.
    "VLLM_ENABLE_V1_MULTIPROCESSING":
776
    lambda: bool(int(os.getenv("VLLM_ENABLE_V1_MULTIPROCESSING", "1"))),
777
778
    "VLLM_LOG_BATCHSIZE_INTERVAL":
    lambda: float(os.getenv("VLLM_LOG_BATCHSIZE_INTERVAL", "-1")),
779
780
    "VLLM_DISABLE_COMPILE_CACHE":
    lambda: bool(int(os.getenv("VLLM_DISABLE_COMPILE_CACHE", "0"))),
781
782
783
784
785
786

    # 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"))),
787
788
789
790
791
792
793
794
795
796

    # 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")),
797
798
799
800
801

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

802
803
804
805
806
807
808
809
810
811
812
813
    # 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", ""),

814
815
816
817
    # 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),
818
819
820
821
822
823
824

    # 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"),
825

826
827
828
829
830
831
    # 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"),

832
833
834
835
    # Rank of the process in the data parallel setting
    "VLLM_DP_RANK":
    lambda: int(os.getenv("VLLM_DP_RANK", "0")),

836
837
838
839
840
841
    # 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)),

842
843
844
845
846
847
848
849
850
851
852
    # 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")),
853

854
855
856
857
858
859
860
861
    # 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")),

862
863
864
865
    # 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",

866
867
868
    # 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",
869

870
    # Use model_redirect to redirect the model name to a local folder.
871
872
873
874
875
    # `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
876
877
878
    "VLLM_MODEL_REDIRECT_PATH":
    lambda: os.environ.get("VLLM_MODEL_REDIRECT_PATH", None),

879
880
881
    # 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",
882
883
884
885
886
887

    # 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",
888

889
890
891
892
    # 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"])
893
    if "VLLM_TPU_BUCKET_PADDING_GAP" in os.environ else 0,
894
895
    "VLLM_TPU_MOST_MODEL_LEN":
    lambda: maybe_convert_int(os.environ.get("VLLM_TPU_MOST_MODEL_LEN", None)),
896
897
898
899

    # Allow use of DeepGemm kernels for fused moe ops.
    "VLLM_USE_DEEP_GEMM":
    lambda: bool(int(os.getenv("VLLM_USE_DEEP_GEMM", "0"))),
900
901
902
903
904
905

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

    # 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")),
916
917
918
919
920
921

    # 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
922
923
924
925
926
927
928
929

    # 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")),
930
931

    # all2all backend for vllm's expert parallel communication
932
933
934
    # Available options:
    # - "naive": naive all2all implementation using all-reduce
    # - "pplx": use pplx kernels
935
936
    # - "deepep_high_throughput", use deepep high-throughput kernels
    # - "deepep_low_latency", use deepep low-latency kernels
937
938
    "VLLM_ALL2ALL_BACKEND":
    lambda: os.getenv("VLLM_ALL2ALL_BACKEND", "naive"),
939
940
941
942
943
944
945

    # 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")),
946
947
948
949

    # 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")),
950
951
952
953
954

    # 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"))),
955
956
957
958
959
960

    # 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")),
961
    
962
963
964
965
966
    # 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")),

967
968
969
970
971
972
973
974
    # 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":
975
976
977
978
979
980
981
    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"))),
982
983
984
985
986

    # 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
987
988
    lambda: bool(int(os.getenv("VLLM_USE_NVFP4_CT_EMULATIONS", "0"))),
    
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
     # 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
1004
1005
1006
1007
1008
1009
1010
1011
    # 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"))),
    
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
    # 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")),
    
    # 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
1035
1036
1037
1038
    # 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")),
    
1039
1040
1041
    # 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")),
1042

1043
1044
1045
1046
1047
1048
1049
    # '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"))),
1050
1051
1052
    
    # If set, vLLM will transpose weight to use nn layout
    "VLLM_USE_NN":
zhuwenwen's avatar
zhuwenwen committed
1053
    lambda: (os.environ.get("VLLM_USE_NN", "True").lower() in
1054
             ("true", "1")),
1055

1056
1057
1058
    # Enable two batch overlap.
    "VLLM_ENABLE_TBO":
    lambda: bool(int(os.getenv("VLLM_ENABLE_TBO", "0"))),
1059
    
1060
1061
1062
    # 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")),
1063

1064
1065
1066
1067
    # 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")),

lizhigong's avatar
lizhigong committed
1068
1069
1070
1071
    # 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")),

1072
1073
1074
    # Enable zero overhead scheduler.
    "VLLM_ZERO_OVERHEAD":
    lambda: bool(int(os.getenv("VLLM_ZERO_OVERHEAD", "0"))),
1075
1076
1077
1078

    # If set, vLLM will enable the moe_fused_gate kernel.
    "VLLM_ENABLE_MOE_FUSED_GATE":
    lambda: bool(int(os.getenv("VLLM_ENABLE_MOE_FUSED_GATE", "1"))),
zhuwenwen's avatar
zhuwenwen committed
1079
    
1080
1081
    # vLLM will use FlashAttention Backend for page attention computation on rocm
    "VLLM_USE_FLASH_ATTN_PA":
zhuwenwen's avatar
zhuwenwen committed
1082
    lambda: (os.environ.get("VLLM_USE_FLASH_ATTN_PA", "True").lower() in
zhuwenwen's avatar
zhuwenwen committed
1083
             ("true", "1")),
zhuwenwen's avatar
zhuwenwen committed
1084
1085
1086
1087
1088
    
    # vLLM will use apex for rmsnorm
    "VLLM_USE_APEX_RN":
    lambda: (os.environ.get("VLLM_USE_APEX_RN", "False").lower() in
             ("true", "1")),
1089
1090
    # vLLM will use global cache for moe
    "VLLM_USE_GLOBAL_CACHE13":
zhuwenwen's avatar
zhuwenwen committed
1091
        lambda: (os.environ.get("VLLM_USE_GLOBAL_CACHE13", "False").lower() in
1092
                 ("true", "1")),
王敏's avatar
王敏 committed
1093
1094
1095
1096
    # vLLM will use all_to_all ep mode
    "VLLM_USE_ALLTOALL_EP":
        lambda: (os.environ.get("VLLM_USE_ALLTOALL_EP", "True").lower() in
                 ("true", "1")),
1097
1098
}

1099
# --8<-- [end:env-vars-definition]
1100

1101

1102
def __getattr__(name: str):
1103
1104
1105
1106
1107
1108
1109
1110
    # 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())
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126


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"
1127
1128
1129
1130
1131
1132
1133
1134


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
1135
    graphs, so it is included in the factors list. The env vars that
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
    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",
1159
        "VLLM_USE_STANDALONE_COMPILE",
1160
        "VLLM_FUSED_MOE_CHUNK_SIZE",
1161
1162
1163
1164
1165
    ]
    for key in environment_variables_to_hash:
        if key in environment_variables:
            factorize(key)

1166
1167
    hash_str = hashlib.md5(str(factors).encode(),
                           usedforsecurity=False).hexdigest()
1168

zhuwenwen's avatar
zhuwenwen committed
1169
    return hash_str