envs.py 33.8 KB
Newer Older
1
2
# SPDX-License-Identifier: Apache-2.0

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

if TYPE_CHECKING:
    VLLM_HOST_IP: str = ""
11
    VLLM_PORT: Optional[int] = None
12
13
    VLLM_OPTEST_URLS_PORT: Optional[int] = None
    VLLM_OPTEST_MODELS_PATH: str = ""
14
    VLLM_RPC_BASE_PATH: str = tempfile.gettempdir()
15
    VLLM_USE_MODELSCOPE: bool = False
16
    VLLM_RINGBUFFER_WARNING_INTERVAL: int = 60
17
18
19
    VLLM_NCCL_SO_PATH: Optional[str] = None
    LD_LIBRARY_PATH: Optional[str] = None
    VLLM_USE_TRITON_FLASH_ATTN: bool = False
20
    VLLM_USE_TRITON_OPT_MLA: bool = False
21
    VLLM_USE_FLASH_MLA: bool = False
zhuwenwen's avatar
zhuwenwen committed
22
    VLLM_USE_OPT_OP: bool = False
23
    VLLM_USE_TC_PAGED_ATTN: bool = False
24
    VLLM_USE_PA_PRINT_PARAM: bool = False 
25
    VLLM_SPEC_DECODE_EAGER: bool = False
zhuwenwen's avatar
zhuwenwen committed
26
    VLLM_PCIE_USE_CUSTOM_ALLREDUCE: bool = False
zhuwenwen's avatar
zhuwenwen committed
27
    VLLM_ENFORCE_EAGER_BS_THRESHOLD: Optional[int] = None
28
    VLLM_FLASH_ATTN_VERSION: Optional[int] = None
29
30
31
32
33
34
35
    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
36
    VLLM_MODEL_REDIRECT_PATH: Optional[str] = None
37
38
    VLLM_CACHE_ROOT: str = os.path.expanduser("~/.cache/vllm")
    VLLM_CONFIG_ROOT: str = os.path.expanduser("~/.config/vllm")
39
40
41
42
43
    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
44
    VLLM_LOGGING_LEVEL: str = "INFO"
45
    VLLM_LOGGING_PREFIX: str = ""
46
    VLLM_LOGGING_CONFIG_PATH: Optional[str] = None
47
    VLLM_LOGITS_PROCESSOR_THREADS: Optional[int] = None
48
49
    VLLM_TRACE_FUNCTION: int = 0
    VLLM_ATTENTION_BACKEND: Optional[str] = None
50
    VLLM_USE_FLASHINFER_SAMPLER: Optional[bool] = None
51
    VLLM_FLASHINFER_FORCE_TENSOR_CORES: bool = False
52
    VLLM_PP_LAYER_PARTITION: Optional[str] = None
53
    VLLM_CPU_KVCACHE_SPACE: int = 0
54
    VLLM_CPU_OMP_THREADS_BIND: str = ""
55
    VLLM_CPU_MOE_PREPACK: bool = True
56
    VLLM_XLA_CACHE_PATH: str = os.path.join(VLLM_CACHE_ROOT, "xla_cache")
57
    VLLM_XLA_CHECK_RECOMPILATION: bool = False
58
    VLLM_FUSED_MOE_CHUNK_SIZE: int = 64 * 1024
59
    VLLM_USE_RAY_SPMD_WORKER: bool = False
60
    VLLM_USE_RAY_COMPILED_DAG: bool = False
61
    VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE: str = "auto"
62
    VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM: 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_MM_INPUT_CACHE_GIB: int = 8
69
70
71
72
    VLLM_TARGET_DEVICE: str = "cuda"
    MAX_JOBS: Optional[str] = None
    NVCC_THREADS: Optional[str] = None
    VLLM_USE_PRECOMPILED: bool = False
73
    VLLM_TEST_USE_PRECOMPILED_NIGHTLY_WHEEL: bool = False
74
    VLLM_NO_DEPRECATION_WARNING: bool = False
75
    VLLM_KEEP_ALIVE_ON_ENGINE_DEATH: bool = False
76
77
    CMAKE_BUILD_TYPE: Optional[str] = None
    VERBOSE: bool = False
78
    VLLM_ALLOW_LONG_MAX_MODEL_LEN: bool = False
79
    VLLM_RPC_TIMEOUT: int = 10000  # ms
80
    VLLM_PLUGINS: Optional[list[str]] = None
81
    VLLM_TORCH_PROFILER_DIR: Optional[str] = None
82
    VLLM_USE_TRITON_AWQ: bool = False
83
    VLLM_ALLOW_RUNTIME_LORA_UPDATING: bool = False
84
    VLLM_TREE_DECODING: bool = False
85
    VLLM_SKIP_P2P_CHECK: bool = False
86
    VLLM_DISABLED_KERNELS: list[str] = []
87
    VLLM_USE_V1: bool = True
88
    VLLM_ROCM_USE_AITER: bool = False
89
    VLLM_ROCM_USE_AITER_PAGED_ATTN: bool = False
90
    VLLM_ROCM_USE_AITER_LINEAR: bool = True
91
    VLLM_ROCM_USE_AITER_MOE: bool = True
92
    VLLM_ROCM_USE_AITER_RMSNORM: bool = True
93
    VLLM_ROCM_USE_AITER_MLA: 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_ENABLE_V1_MULTIPROCESSING: bool = True
99
    VLLM_LOG_BATCHSIZE_INTERVAL: float = -1
100
    VLLM_DISABLE_COMPILE_CACHE: bool = False
101
    Q_SCALE_CONSTANT: int = 200
102
103
    K_SCALE_CONSTANT: int = 200
    V_SCALE_CONSTANT: int = 100
104
    VLLM_SERVER_DEV_MODE: bool = False
105
    VLLM_V1_OUTPUT_PROC_CHUNK_SIZE: int = 128
106
    VLLM_MLA_DISABLE: bool = False
107
    VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON: 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_MARLIN_USE_ATOMIC_ADD: bool = False
119
    VLLM_V0_USE_OUTLINES_CACHE: bool = False
120
    VLLM_TPU_BUCKET_PADDING_GAP: int = 0
121
    VLLM_USE_DEEP_GEMM: bool = False
122
    VLLM_XGRAMMAR_CACHE_MB: int = 0
123
    VLLM_MSGPACK_ZERO_COPY_THRESHOLD: int = 256
124

125
126
127
128
129
130
131
132
133
134
135
136
137
138
139

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"),
    )


140
141
142
143
144
145
def maybe_convert_int(value: Optional[str]) -> Optional[int]:
    if value is None:
        return None
    return int(value)


146
147
148
149
150
# The begin-* and end* here are used by the documentation generator
# to extract the used env vars.

# begin-env-vars-definition

151
environment_variables: dict[str, Callable[[], Any]] = {
152
153
154

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

155
    # Target device of vLLM, supporting [cuda (by default),
156
    # rocm, neuron, cpu]
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
    "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":
173
174
    lambda: bool(os.environ.get("VLLM_USE_PRECOMPILED")) or bool(
        os.environ.get("VLLM_PRECOMPILED_WHEEL_LOCATION")),
175

176
177
178
179
180
181
    # 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"))
                 ),

182
183
184
185
186
187
188
189
190
191
    # 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'))),

192
    # Root directory for vLLM configuration files
193
    # Defaults to `~/.config/vllm` unless `XDG_CONFIG_HOME` is set
194
195
196
197
    # 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":
198
199
200
201
202
    lambda: os.path.expanduser(
        os.getenv(
            "VLLM_CONFIG_ROOT",
            os.path.join(get_default_config_root(), "vllm"),
        )),
203
204
205

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

206
    # Root directory for vLLM cache files
207
208
209
210
211
212
213
214
    # 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"),
        )),

215
216
217
218
    # 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.
219
    'VLLM_HOST_IP':
220
    lambda: os.getenv('VLLM_HOST_IP', ""),
221

222
    # used in distributed environment to manually set the communication port
223
224
225
    # 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.
226
227
228
229
    # '0' is used to make mypy happy
    'VLLM_PORT':
    lambda: int(os.getenv('VLLM_PORT', '0'))
    if 'VLLM_PORT' in os.environ else None,
230
231
232
233
234
235
236
237
238
239
    
    # 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", ""),
240

241
242
243
244
    # 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()),
245

246
247
248
249
250
    # 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",

251
252
253
254
    # 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")),

255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
    # 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":
272
    lambda: (os.environ.get("VLLM_USE_TRITON_FLASH_ATTN", "False").lower() in
273
             ("true", "1")),
274
    
zhuwenwen's avatar
zhuwenwen committed
275
276
277
278
279
    # 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")),
    
280
281
    # flag to control vllm to use optimized tc paged attn kernels
    "VLLM_USE_TC_PAGED_ATTN":
zhuwenwen's avatar
zhuwenwen committed
282
    lambda: (os.environ.get("VLLM_USE_TC_PAGED_ATTN", "True").lower() in
283
284
             ("true", "1")),
    
285
286
287
288
    # 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")),
289
290
291
292
    
    # 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"))),
zhuwenwen's avatar
zhuwenwen committed
293
294
295
296
    
    # 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"))),
297

298
299
300
301
    # Force vllm to use a specific flash-attention version (2 or 3), only valid
    # when using the flash-attention backend.
    "VLLM_FLASH_ATTN_VERSION":
    lambda: maybe_convert_int(os.environ.get("VLLM_FLASH_ATTN_VERSION", None)),
zhuwenwen's avatar
zhuwenwen committed
302
303
304
305
    
    # 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")),
306

307
308
309
310
311
    # Internal flag to enable Dynamo fullgraph capture
    "VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE":
    lambda: bool(
        os.environ.get("VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE", "1") != "0"),

312
313
314
315
316
317
318
319
320
321
322
    # 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":
323
    lambda: int(os.environ.get("VLLM_ENGINE_ITERATION_TIMEOUT_S", "120")),
324

325
    # API key for vLLM API server
326
327
328
    "VLLM_API_KEY":
    lambda: os.environ.get("VLLM_API_KEY", None),

329
330
331
332
333
    # Whether to log responses from API Server for debugging
    "VLLM_DEBUG_LOG_API_SERVER_RESPONSE":
    lambda: os.environ.get("VLLM_DEBUG_LOG_API_SERVER_RESPONSE", "False").
    lower() == "true",

334
335
    # S3 access information, used for tensorizer to load model from S3
    "S3_ACCESS_KEY_ID":
336
    lambda: os.environ.get("S3_ACCESS_KEY_ID", None),
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
    "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"),

362
363
    # this is used for configuring the default logging level
    "VLLM_LOGGING_LEVEL":
364
    lambda: os.getenv("VLLM_LOGGING_LEVEL", "INFO").upper(),
365

366
367
368
369
    # if set, VLLM_LOGGING_PREFIX will be prepended to all log messages
    "VLLM_LOGGING_PREFIX":
    lambda: os.getenv("VLLM_LOGGING_PREFIX", ""),

370
371
372
373
374
375
376
377
    # 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,

378
379
380
381
382
383
384
385
386
387
388
389
    # 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
390
    # - "FLASHINFER": use flashinfer
391
    # - "FLASHMLA": use FlashMLA
392
393
394
    "VLLM_ATTENTION_BACKEND":
    lambda: os.getenv("VLLM_ATTENTION_BACKEND", None),

395
396
    # If set, vllm will use flashinfer sampler
    "VLLM_USE_FLASHINFER_SAMPLER":
397
398
    lambda: bool(int(os.environ["VLLM_USE_FLASHINFER_SAMPLER"]))
    if "VLLM_USE_FLASHINFER_SAMPLER" in os.environ else None,
399

400
401
402
403
404
    # 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"))),

405
406
407
408
    # Pipeline stage partition strategy
    "VLLM_PP_LAYER_PARTITION":
    lambda: os.getenv("VLLM_PP_LAYER_PARTITION", None),

409
    # (CPU backend only) CPU key-value cache space.
410
    # default is 4 GiB
411
412
413
    "VLLM_CPU_KVCACHE_SPACE":
    lambda: int(os.getenv("VLLM_CPU_KVCACHE_SPACE", "0")),

414
415
416
417
418
    # (CPU backend only) CPU core ids bound by OpenMP threads, e.g., "0-31",
    # "0,1,2", "0-31,33". CPU cores of different ranks are separated by '|'.
    "VLLM_CPU_OMP_THREADS_BIND":
    lambda: os.getenv("VLLM_CPU_OMP_THREADS_BIND", "all"),

419
420
421
422
423
    # (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"))),
424

425
426
427
428
429
    # 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":
430
    lambda: bool(int(os.getenv("VLLM_USE_RAY_SPMD_WORKER", "0"))),
431

432
433
434
    # If the env var is set, it uses the Ray's Compiled Graph
    # (previously known as ADAG) API which optimizes the
    # control plane overhead.
435
    # Run vLLM with VLLM_USE_RAY_COMPILED_DAG=1 to enable it.
436
437
    # Note that this variable is set to 1 in V1 by default
    # when ray distributed executor is used.
438
    "VLLM_USE_RAY_COMPILED_DAG":
439
440
    lambda: bool(int(os.getenv("VLLM_USE_RAY_COMPILED_DAG", "0"))),

441
442
443
444
445
446
447
448
449
450
    # 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"),
451

452
    # If the env var is set, it enables GPU communication overlap
453
    # (experimental feature) in Ray's Compiled Graph. This flag is ignored if
454
455
    # VLLM_USE_RAY_COMPILED_DAG is not set.
    "VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM":
456
    lambda: bool(int(os.getenv("VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM", "0"))
457
458
                 ),

459
460
461
    # Use dedicated multiprocess context for workers.
    # Both spawn and fork work
    "VLLM_WORKER_MULTIPROC_METHOD":
462
    lambda: os.getenv("VLLM_WORKER_MULTIPROC_METHOD", "spawn"),
463

464
465
466
467
468
469
470
471
    # 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"),
        )),

472
473
474
475
    # 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")),
476

477
    # Timeout for fetching videos when serving multimodal models
478
    # Default is 30 seconds
479
    "VLLM_VIDEO_FETCH_TIMEOUT":
480
    lambda: int(os.getenv("VLLM_VIDEO_FETCH_TIMEOUT", "30")),
481

482
    # Timeout for fetching audio when serving multimodal models
483
    # Default is 10 seconds
484
    "VLLM_AUDIO_FETCH_TIMEOUT":
485
    lambda: int(os.getenv("VLLM_AUDIO_FETCH_TIMEOUT", "10")),
486

487
    # Cache size (in GiB) for multimodal input cache
488
    # Default is 4 GiB
489
    "VLLM_MM_INPUT_CACHE_GIB":
490
    lambda: int(os.getenv("VLLM_MM_INPUT_CACHE_GIB", "4")),
491

492
493
494
    # Path to the XLA persistent cache directory.
    # Only used for XLA devices such as TPUs.
    "VLLM_XLA_CACHE_PATH":
495
496
    lambda: os.path.expanduser(
        os.getenv(
497
            "VLLM_XLA_CACHE_PATH",
498
499
            os.path.join(get_default_cache_root(), "vllm", "xla_cache"),
        )),
500
501
502
503

    # If set, assert on XLA recompilation after each execution step.
    "VLLM_XLA_CHECK_RECOMPILATION":
    lambda: bool(int(os.getenv("VLLM_XLA_CHECK_RECOMPILATION", "0"))),
504
    "VLLM_FUSED_MOE_CHUNK_SIZE":
505
    lambda: int(os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", "32768")),
506
507
508
509

    # If set, vllm will skip the deprecation warnings.
    "VLLM_NO_DEPRECATION_WARNING":
    lambda: bool(int(os.getenv("VLLM_NO_DEPRECATION_WARNING", "0"))),
510

511
512
513
514
515
    # 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)),

516
517
518
519
520
521
522
523
    # 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")),
524
525
526
527
528
529
530

    # 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")),
531
532
    "VLLM_TEST_FORCE_LOAD_FORMAT":
    lambda: os.getenv("VLLM_TEST_FORCE_LOAD_FORMAT", "dummy"),
533

534
535
    # Time in ms for the zmq client to wait for a response from the backend
    # server for simple data operations
536
537
    "VLLM_RPC_TIMEOUT":
    lambda: int(os.getenv("VLLM_RPC_TIMEOUT", "10000")),
538

539
540
541
542
543
544
    # 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(","),
545
546
547
548
549
550

    # 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", "."))),
551
552
553
554

    # If set, vLLM will use Triton implementations of AWQ.
    "VLLM_USE_TRITON_AWQ":
    lambda: bool(int(os.getenv("VLLM_USE_TRITON_AWQ", "0"))),
555
556
557
558
559
560

    # 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")),
561
562
563
564
565

    # 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
566
     ("1", "true")),
567
568
569
570
571
572
    # 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",
573

574
575
576
577
578
579
580
    # 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(","),
581
582
583

    # If set, use the V1 code path.
    "VLLM_USE_V1":
584
    lambda: bool(int(os.getenv("VLLM_USE_V1", "1"))),
585

586
587
588
589
590
    # 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")),
591

592
593
594
595
596
597
    # 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")),

598
599
600
601
602
603
604
    # 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")),

605
606
607
608
609
610
    # 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")),

611
612
613
614
615
    # 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")),

616
617
618
619
620
    # 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")),
621
622
623
624
625
    # use rocm skinny gemms
    "VLLM_ROCM_USE_SKINNY_GEMM":
    lambda: (os.getenv("VLLM_ROCM_USE_SKINNY_GEMM", "True").lower() in
             ("true", "1")),

626
627
628
    # Pad the fp8 weights to 256 bytes for ROCm
    "VLLM_ROCM_FP8_PADDING":
    lambda: bool(int(os.getenv("VLLM_ROCM_FP8_PADDING", "1"))),
629

630
631
    # Pad the weights for the moe kernel
    "VLLM_ROCM_MOE_PADDING":
zhuwenwen's avatar
zhuwenwen committed
632
    lambda: bool(int(os.getenv("VLLM_ROCM_MOE_PADDING", "0"))),
633

634
635
636
637
638
    # 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")),

639
640
641
    # Divisor for dynamic query scale factor calculation for FP8 KV Cache
    "Q_SCALE_CONSTANT":
    lambda: int(os.getenv("Q_SCALE_CONSTANT", "200")),
642
643
644
645
646
647
    # 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")),
648

649
650
    # If set, enable multiprocessing in LLM for the V1 code path.
    "VLLM_ENABLE_V1_MULTIPROCESSING":
651
    lambda: bool(int(os.getenv("VLLM_ENABLE_V1_MULTIPROCESSING", "1"))),
652
653
    "VLLM_LOG_BATCHSIZE_INTERVAL":
    lambda: float(os.getenv("VLLM_LOG_BATCHSIZE_INTERVAL", "-1")),
654
655
    "VLLM_DISABLE_COMPILE_CACHE":
    lambda: bool(int(os.getenv("VLLM_DISABLE_COMPILE_CACHE", "0"))),
656
657
658
659
660
661

    # 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"))),
662
663
664
665
666
667
668
669
670
671

    # 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")),
672
673
674
675

    # If set, vLLM will disable the MLA attention optimizations.
    "VLLM_MLA_DISABLE":
    lambda: bool(int(os.getenv("VLLM_MLA_DISABLE", "0"))),
676
677
678
679
    
    # 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"))),
680
681
682
683
    
    # If set, vLLM will use FLASH MLA attention optimizations.
    "VLLM_USE_FLASH_MLA":
    lambda: bool(int(os.getenv("VLLM_USE_FLASH_MLA", "1"))),
684

685
686
687
688
689
    # If set, vLLM will use the Triton implementation of moe_align_block_size,
    # i.e. moe_align_block_size_triton in fused_moe.py.
    "VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON":
    lambda: bool(int(os.getenv("VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON", "0"))
                 ),
690

691
692
693
694
695
696
697
698
699
700
701
702
    # 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", ""),

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

    # 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"),
714

715
716
717
718
719
720
    # 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"),

721
722
723
724
    # Rank of the process in the data parallel setting
    "VLLM_DP_RANK":
    lambda: int(os.getenv("VLLM_DP_RANK", "0")),

725
726
727
728
729
730
    # 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)),

731
732
733
734
735
736
737
738
739
740
741
    # 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")),
742
743
744
745

    # 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",
746

747
    # Use model_redirect to redirect the model name to a local folder.
748
749
750
751
752
    # `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
753
754
755
    "VLLM_MODEL_REDIRECT_PATH":
    lambda: os.environ.get("VLLM_MODEL_REDIRECT_PATH", None),

756
757
758
    # 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",
759
760
761
762
763
764

    # 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",
765

766
767
768
769
    # 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"])
770
    if "VLLM_TPU_BUCKET_PADDING_GAP" in os.environ else 0,
771
772
773
774

    # Allow use of DeepGemm kernels for fused moe ops.
    "VLLM_USE_DEEP_GEMM":
    lambda: bool(int(os.getenv("VLLM_USE_DEEP_GEMM", "0"))),
775
776
777
778
779
780

    # 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")),
781
782
783
784
785
786
787
788
789
790

    # 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")),
791
792
}

793
794
# end-env-vars-definition

795

796
def __getattr__(name: str):
797
798
799
800
801
802
803
804
    # 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())
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820


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"
821
822
823
824
825
826
827
828


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
829
    graphs, so it is included in the factors list. The env vars that
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
    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",
    ]
    for key in environment_variables_to_hash:
        if key in environment_variables:
            factorize(key)

858
859
    hash_str = hashlib.md5(str(factors).encode(),
                           usedforsecurity=False).hexdigest()
860

861
    return hash_str