"requirements/tpu.txt" did not exist on "2148441fd371faf3e90748b310fdb4500939e527"
envs.py 58.5 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_PP_LAYER_PARTITION_D: Optional[str] = None
46
    VLLM_CPU_KVCACHE_SPACE: int = 0
47
    VLLM_CPU_OMP_THREADS_BIND: str = ""
48
    VLLM_CPU_NUM_OF_RESERVED_CPU: int = 0
49
    VLLM_CPU_MOE_PREPACK: bool = True
50
    VLLM_CPU_SGL_KERNEL: bool = False
51
    VLLM_XLA_CACHE_PATH: str = os.path.join(VLLM_CACHE_ROOT, "xla_cache")
52
    VLLM_XLA_CHECK_RECOMPILATION: bool = False
53
    VLLM_FUSED_MOE_CHUNK_SIZE: int = 64 * 1024
54
    VLLM_ENABLE_FUSED_MOE_ACTIVATION_CHUNKING: bool = True
55
    VLLM_USE_RAY_SPMD_WORKER: bool = False
56
    VLLM_USE_RAY_COMPILED_DAG: bool = False
57
    VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE: str = "auto"
58
    VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM: bool = False
59
    VLLM_XLA_USE_SPMD: bool = False
60
    VLLM_WORKER_MULTIPROC_METHOD: str = "spawn"
61
    VLLM_ASSETS_CACHE: str = os.path.join(VLLM_CACHE_ROOT, "assets")
62
    VLLM_IMAGE_FETCH_TIMEOUT: int = 5
63
    VLLM_VIDEO_FETCH_TIMEOUT: int = 30
64
    VLLM_AUDIO_FETCH_TIMEOUT: int = 10
65
    VLLM_VIDEO_LOADER_BACKEND: str = "opencv"
66
    VLLM_MM_INPUT_CACHE_GIB: int = 8
67
68
69
70
    VLLM_TARGET_DEVICE: str = "cuda"
    MAX_JOBS: Optional[str] = None
    NVCC_THREADS: Optional[str] = None
    VLLM_USE_PRECOMPILED: bool = False
71
    VLLM_TEST_USE_PRECOMPILED_NIGHTLY_WHEEL: bool = False
72
    VLLM_NO_DEPRECATION_WARNING: bool = False
73
    VLLM_KEEP_ALIVE_ON_ENGINE_DEATH: bool = False
74
75
    CMAKE_BUILD_TYPE: Optional[str] = None
    VERBOSE: bool = False
76
    VLLM_ALLOW_LONG_MAX_MODEL_LEN: bool = False
77
    VLLM_RPC_TIMEOUT: int = 10000  # ms
78
    VLLM_HTTP_TIMEOUT_KEEP_ALIVE: int = 5  # seconds
79
    VLLM_PLUGINS: Optional[list[str]] = None
80
    VLLM_LORA_RESOLVER_CACHE_DIR: Optional[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_AITER_MHA: bool = True
95
    VLLM_ROCM_USE_SKINNY_GEMM: bool = True
96
    VLLM_ROCM_FP8_PADDING: bool = True
97
    VLLM_ROCM_MOE_PADDING: bool = True
98
    VLLM_ROCM_CUSTOM_PAGED_ATTN: bool = True
99
    VLLM_QUARK_EMU_MEM_OPT: bool = False
100
    VLLM_ENABLE_V1_MULTIPROCESSING: bool = True
101
    VLLM_LOG_BATCHSIZE_INTERVAL: float = -1
102
    VLLM_DISABLE_COMPILE_CACHE: bool = False
103
    Q_SCALE_CONSTANT: int = 200
104
105
    K_SCALE_CONSTANT: int = 200
    V_SCALE_CONSTANT: int = 100
106
    VLLM_SERVER_DEV_MODE: bool = False
107
    VLLM_V1_OUTPUT_PROC_CHUNK_SIZE: int = 128
108
    VLLM_MLA_DISABLE: bool = False
109
110
    VLLM_RAY_PER_WORKER_GPUS: float = 1.0
    VLLM_RAY_BUNDLE_INDICES: str = ""
111
    VLLM_CUDART_SO_PATH: Optional[str] = None
112
    VLLM_USE_HPU_CONTIGUOUS_CACHE_FETCH: bool = True
113
    VLLM_HPU_USE_DELAYED_SAMPLING: bool = False
114
    VLLM_DP_RANK: int = 0
115
    VLLM_DP_RANK_LOCAL: int = -1
116
117
118
    VLLM_DP_SIZE: int = 1
    VLLM_DP_MASTER_IP: str = ""
    VLLM_DP_MASTER_PORT: int = 0
119
    VLLM_MOE_DP_CHUNK_SIZE: int = 256
120
    VLLM_RANDOMIZE_DP_DUMMY_INPUTS: bool = False
121
    VLLM_MARLIN_USE_ATOMIC_ADD: bool = False
122
    VLLM_V0_USE_OUTLINES_CACHE: bool = False
123
    VLLM_TPU_BUCKET_PADDING_GAP: int = 0
124
    VLLM_TPU_MOST_MODEL_LEN: Optional[int] = None
125
    VLLM_USE_DEEP_GEMM: bool = False
126
    VLLM_XGRAMMAR_CACHE_MB: int = 0
127
    VLLM_MSGPACK_ZERO_COPY_THRESHOLD: int = 256
128
    VLLM_ALLOW_INSECURE_SERIALIZATION: bool = False
Robert Shaw's avatar
Robert Shaw committed
129
130
    VLLM_NIXL_SIDE_CHANNEL_HOST: str = "localhost"
    VLLM_NIXL_SIDE_CHANNEL_PORT: int = 5557
131
    VLLM_ALL2ALL_BACKEND: str = "naive"
yangql's avatar
yangql committed
132
    VLLM_MOE_HT_THRESHOLD: int = 128
133
    VLLM_ALLOW_MNNVL: bool = False
134
    VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE: int = 163840
135
    VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS: int = 1
136
    VLLM_SLEEP_WHEN_IDLE: bool = False
137
    VLLM_MQ_MAX_CHUNK_BYTES_MB: int = 16
138
    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: int = 300
139
    VLLM_KV_CACHE_LAYOUT: Optional[str] = None
140
    VLLM_COMPUTE_NANS_IN_LOGITS: bool = False
141
    VLLM_USE_NVFP4_CT_EMULATIONS: bool = False
142
143
144
    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
145
    
146
147
148
149
150
    # 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
151
    VLLM_USE_FLASH_ATTN_FP8: bool = False
152
    VLLM_USE_FLASH_MLA: bool = False
153
    VLLM_USE_FLASH_MLA_FP8: bool = False
154
155
156
157
158
    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
159
    VLLM_CUSTOM_CACHE: bool = False
zhuwenwen's avatar
zhuwenwen committed
160
    VLLM_CUSTOM_ALLREDUCE_SUPPORTED_WORLDSIZE_MAX: int = 16
161
162
    VLLM_ENFORCE_EAGER_BS_THRESHOLD: Optional[int] = None
    VLLM_HAS_CONTEXT_DEFAULT: bool = False
163
    VLLM_USE_NN: bool = False
164
    VLLM_ENABLE_TBO: bool = False
165
166
    VLLM_TBO_REQ_DELAY_MS: int = 0
    VLLM_TBO_DECODE_BS: int = 0
lizhigong's avatar
lizhigong committed
167
    VLLM_TBO_MIN_TOKENS: int = 200
168
    VLLM_ZERO_OVERHEAD: bool = False
169
    VLLM_ENABLE_MOE_FUSED_GATE: bool = False
170
    VLLM_USE_FLASH_ATTN_PA: bool = False
zhuwenwen's avatar
zhuwenwen committed
171
    VLLM_USE_APEX_RN: bool = False
172
    VLLM_USE_GLOBAL_CACHE13: bool = False
173
    VLLM_USE_LIGHTOP: bool = False
174
    VLLM_USE_OPT_ZEROS: bool = False
175
    VLLM_USE_OPT_CAT: bool = False
176
    VLLM_USE_OPT_MOE_SUM: bool = False
177
    VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD: bool = False
178
179
    VLLM_USE_LIGHTOP_MOE_SUM: bool = False
    VLLM_USE_LIGHTOP_MOE_ALIGN: bool = False
180
    VLLM_USE_MERGE_ATTN_STATES_OPT: bool = False
181
182
    USE_FUSED_RMS_QUANT: bool = False
    USE_FUSED_SILU_MUL_QUANT: bool = False
183
    VLLM_P2P_ASYNC: bool = False
184
    VLLM_P2P_BUF_TOKENS: int = 30000
185
    VLLM_SCHED_ENABLE_MINIMAL_INJECTION: bool = False
zhuwenwen's avatar
zhuwenwen committed
186
    VLLM_USE_PD_SPLIT: bool = False
187
    VLLM_USE_PP_SYNC: bool = False
188
    VLLM_USE_LIGHTOP_FILL_MOE_ALIGN: bool = False
189
    USE_FUSED_CUSTOM_ALL_REDUCE_RMS_QUANT: bool = False
190
    VLLM_USE_PP_BALANCE: bool = False
191
    VLLM_USE_ZERO_MTP: bool = False
192
    VLLM_USE_CUDA_GRAPH_SIZES: bool = False
193
    VLLM_USE_CAT_MLA: bool = False
王敏's avatar
王敏 committed
194
    VLLM_REJECT_SAMPLE_OPT: bool = False
195
    VLLM_USE_FUSE_SILU_AND_MUL: bool = False
196
    VLLM_USE_OPT_RESHAPE_AND_CACHE: bool = False
197
    VLLM_USE_TOPK_RENORM: bool = False
198
    VLLM_PP_DEBUG: bool = False
199
    VLLM_USE_V32_ENCODE: bool = False
200
    VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT: bool = False
laibao's avatar
laibao committed
201
202
    VLLM_USE_FUSED_RMS_ROPE: bool = False
    VLLM_USE_MARLIN_W16A16_MOE:bool = False
203
    VLLM_V1_USE_REDUCED_TOPK_TOPP_SAMPLER: bool = False
204
    VLLM_USE_FUSED_FILL_RMS_CAT:bool = False
王敏's avatar
王敏 committed
205
    VLLM_ENABLE_DEEPEP_HT_DEEPGEMM: bool = True
206
    VLLM_ZERO_OVERHEAD_ENHANCE: bool = False
207
    VLLM_USE_FUSED_QA_KVA_GEMM: bool = False
208
    VLLM_V1_FAST_TOKEN_ID_COPY: bool = False
209
    VLLM_DISABLE_SHARED_EXPERTS_STREAM:bool = True
210
    VLLM_ENABLE_SHARED_EXPERTS_FUSION: bool = False
wujl5's avatar
wujl5 committed
211
212
    VLLM_USE_MOE_W16A16_TRITON: bool = False
    VLLM_USE_FUSED_DTBMM: bool = False
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227

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"),
    )


228
229
230
231
232
233
def maybe_convert_int(value: Optional[str]) -> Optional[int]:
    if value is None:
        return None
    return 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, neuron, cpu]
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
    "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":
290
291
    lambda: bool(os.environ.get("VLLM_USE_PRECOMPILED")) or bool(
        os.environ.get("VLLM_PRECOMPILED_WHEEL_LOCATION")),
292

293
294
295
296
297
298
    # 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"))
                 ),

299
300
301
302
303
304
305
306
307
308
    # 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'))),

309
    # Root directory for vLLM configuration files
310
    # Defaults to `~/.config/vllm` unless `XDG_CONFIG_HOME` is set
311
312
313
314
    # 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":
315
316
317
318
319
    lambda: os.path.expanduser(
        os.getenv(
            "VLLM_CONFIG_ROOT",
            os.path.join(get_default_config_root(), "vllm"),
        )),
320
321
322

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

323
    # Root directory for vLLM cache files
324
325
326
327
328
329
330
331
    # 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"),
        )),

332
333
334
335
    # 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.
336
    'VLLM_HOST_IP':
337
    lambda: os.getenv('VLLM_HOST_IP', ""),
338

339
    # used in distributed environment to manually set the communication port
340
341
342
    # 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.
343
    'VLLM_PORT':
344
    get_vllm_port,
345

346
347
348
349
    # 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()),
350

351
352
353
354
355
    # 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",

356
357
358
359
    # 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")),

360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
    # 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":
377
    lambda: (os.environ.get("VLLM_USE_TRITON_FLASH_ATTN", "False").lower() in
378
379
             ("true", "1")),

380
381
382
383
384
385
386
    # 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")),

387
388
389
390
391
    # 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)),

392
393
394
395
396
    # Internal flag to enable Dynamo fullgraph capture
    "VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE":
    lambda: bool(
        os.environ.get("VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE", "1") != "0"),

397
398
399
400
401
    # 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",
402

403
404
405
406
407
408
409
410
411
412
413
    # 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":
414
    lambda: int(os.environ.get("VLLM_ENGINE_ITERATION_TIMEOUT_S", "120")),
415

416
    # API key for vLLM API server
417
418
419
    "VLLM_API_KEY":
    lambda: os.environ.get("VLLM_API_KEY", None),

420
421
    # Whether to log responses from API Server for debugging
    "VLLM_DEBUG_LOG_API_SERVER_RESPONSE":
422
423
    lambda: os.environ.get("VLLM_DEBUG_LOG_API_SERVER_RESPONSE", "False"
                           ).lower() == "true",
424

425
426
    # S3 access information, used for tensorizer to load model from S3
    "S3_ACCESS_KEY_ID":
427
    lambda: os.environ.get("S3_ACCESS_KEY_ID", None),
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
    "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"),

453
454
    # this is used for configuring the default logging level
    "VLLM_LOGGING_LEVEL":
455
    lambda: os.getenv("VLLM_LOGGING_LEVEL", "INFO").upper(),
456

457
458
459
460
    # if set, VLLM_LOGGING_PREFIX will be prepended to all log messages
    "VLLM_LOGGING_PREFIX":
    lambda: os.getenv("VLLM_LOGGING_PREFIX", ""),

461
462
463
464
465
466
467
468
    # 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,

469
470
471
472
473
474
475
476
477
478
479
480
    # 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
481
    # - "FLASHINFER": use flashinfer
482
    # - "FLASHMLA": use FlashMLA
483
484
485
    "VLLM_ATTENTION_BACKEND":
    lambda: os.getenv("VLLM_ATTENTION_BACKEND", None),

486
487
    # If set, vllm will use flashinfer sampler
    "VLLM_USE_FLASHINFER_SAMPLER":
488
489
    lambda: bool(int(os.environ["VLLM_USE_FLASHINFER_SAMPLER"]))
    if "VLLM_USE_FLASHINFER_SAMPLER" in os.environ else None,
490

491
492
493
494
495
    # 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"))),

496
497
498
499
    # Pipeline stage partition strategy
    "VLLM_PP_LAYER_PARTITION":
    lambda: os.getenv("VLLM_PP_LAYER_PARTITION", None),

500
501
502
503
    # Pipeline stage partition strategy
    "VLLM_PP_LAYER_PARTITION_D":
    lambda: os.getenv("VLLM_PP_LAYER_PARTITION_D", None),

504
    # (CPU backend only) CPU key-value cache space.
505
    # default is 4 GiB
506
507
508
    "VLLM_CPU_KVCACHE_SPACE":
    lambda: int(os.getenv("VLLM_CPU_KVCACHE_SPACE", "0")),

509
510
511
    # (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":
512
513
514
515
516
517
    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")),
518

519
520
521
522
523
    # (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"))),
524

525
526
527
528
    # (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"))),

529
530
531
532
533
    # 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":
534
    lambda: bool(int(os.getenv("VLLM_USE_RAY_SPMD_WORKER", "0"))),
535

536
537
538
    # If the env var is set, it uses the Ray's Compiled Graph
    # (previously known as ADAG) API which optimizes the
    # control plane overhead.
539
    # Run vLLM with VLLM_USE_RAY_COMPILED_DAG=1 to enable it.
540
541
    # Note that this variable is set to 1 in V1 by default
    # when ray distributed executor is used.
542
    "VLLM_USE_RAY_COMPILED_DAG":
543
544
    lambda: bool(int(os.getenv("VLLM_USE_RAY_COMPILED_DAG", "0"))),

545
546
547
548
549
550
551
552
553
554
    # 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"),
555

556
    # If the env var is set, it enables GPU communication overlap
557
    # (experimental feature) in Ray's Compiled Graph. This flag is ignored if
558
559
    # VLLM_USE_RAY_COMPILED_DAG is not set.
    "VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM":
560
    lambda: bool(int(os.getenv("VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM", "0"))
561
562
                 ),

563
564
565
    # Use dedicated multiprocess context for workers.
    # Both spawn and fork work
    "VLLM_WORKER_MULTIPROC_METHOD":
566
    lambda: os.getenv("VLLM_WORKER_MULTIPROC_METHOD", "spawn"),
567

568
569
570
571
572
573
574
575
    # 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"),
        )),

576
577
578
579
    # 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")),
580

581
    # Timeout for fetching videos when serving multimodal models
582
    # Default is 30 seconds
583
    "VLLM_VIDEO_FETCH_TIMEOUT":
584
    lambda: int(os.getenv("VLLM_VIDEO_FETCH_TIMEOUT", "30")),
585

586
    # Timeout for fetching audio when serving multimodal models
587
    # Default is 10 seconds
588
    "VLLM_AUDIO_FETCH_TIMEOUT":
589
    lambda: int(os.getenv("VLLM_AUDIO_FETCH_TIMEOUT", "10")),
590

591
592
593
594
595
596
597
598
599
600
    # 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"),

601
    # Cache size (in GiB) for multimodal input cache
602
    # Default is 4 GiB
603
    "VLLM_MM_INPUT_CACHE_GIB":
604
    lambda: int(os.getenv("VLLM_MM_INPUT_CACHE_GIB", "4")),
605

606
607
608
    # Path to the XLA persistent cache directory.
    # Only used for XLA devices such as TPUs.
    "VLLM_XLA_CACHE_PATH":
609
610
    lambda: os.path.expanduser(
        os.getenv(
611
            "VLLM_XLA_CACHE_PATH",
612
613
            os.path.join(get_default_cache_root(), "vllm", "xla_cache"),
        )),
614
615
616
617

    # If set, assert on XLA recompilation after each execution step.
    "VLLM_XLA_CHECK_RECOMPILATION":
    lambda: bool(int(os.getenv("VLLM_XLA_CHECK_RECOMPILATION", "0"))),
618
619
620
621

    # Enable SPMD mode for TPU backend.
    "VLLM_XLA_USE_SPMD":
    lambda: bool(int(os.getenv("VLLM_XLA_USE_SPMD", "0"))),
622
    "VLLM_FUSED_MOE_CHUNK_SIZE":
623
    lambda: int(os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", "32768")),
624
625
626
627
628
629
    # 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"))),
630
631
632
633

    # If set, vllm will skip the deprecation warnings.
    "VLLM_NO_DEPRECATION_WARNING":
    lambda: bool(int(os.getenv("VLLM_NO_DEPRECATION_WARNING", "0"))),
634

635
636
637
638
639
    # 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)),

640
641
642
643
644
645
646
647
    # 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")),
648
649
650
651
652
653
654

    # 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")),
655
656
    "VLLM_TEST_FORCE_LOAD_FORMAT":
    lambda: os.getenv("VLLM_TEST_FORCE_LOAD_FORMAT", "dummy"),
657

658
659
    # Time in ms for the zmq client to wait for a response from the backend
    # server for simple data operations
660
661
    "VLLM_RPC_TIMEOUT":
    lambda: int(os.getenv("VLLM_RPC_TIMEOUT", "10000")),
662

663
664
665
666
    # 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")),

667
668
669
670
671
672
    # 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(","),
673

674
675
676
677
678
679
    # 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),

680
681
682
683
684
    # 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", "."))),
685
686
687
688

    # If set, vLLM will use Triton implementations of AWQ.
    "VLLM_USE_TRITON_AWQ":
    lambda: bool(int(os.getenv("VLLM_USE_TRITON_AWQ", "0"))),
689
690
691
692
693
694

    # 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")),
695
696
697
698
699

    # 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
700
     ("1", "true")),
701
702
703
704
705
706
    # 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",
707

708
709
710
711
712
713
714
    # 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(","),
715
716
717

    # If set, use the V1 code path.
    "VLLM_USE_V1":
zhuwenwen's avatar
zhuwenwen committed
718
    lambda: bool(int(os.getenv("VLLM_USE_V1", "1"))),
719

720
721
722
723
724
    # 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")),
725

726
727
728
729
730
731
    # 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")),

732
733
734
735
736
737
738
    # 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")),

739
740
741
742
743
744
    # 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")),

745
746
747
748
749
    # 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")),

750
751
752
753
754
    # 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")),
755
756
757
758
759
760
761

    # 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")),

762
763
764
765
766
    # use rocm skinny gemms
    "VLLM_ROCM_USE_SKINNY_GEMM":
    lambda: (os.getenv("VLLM_ROCM_USE_SKINNY_GEMM", "True").lower() in
             ("true", "1")),

767
768
769
    # Pad the fp8 weights to 256 bytes for ROCm
    "VLLM_ROCM_FP8_PADDING":
    lambda: bool(int(os.getenv("VLLM_ROCM_FP8_PADDING", "1"))),
770

771
772
    # Pad the weights for the moe kernel
    "VLLM_ROCM_MOE_PADDING":
zhuwenwen's avatar
zhuwenwen committed
773
    lambda: bool(int(os.getenv("VLLM_ROCM_MOE_PADDING", "0"))),
774

775
776
777
778
779
    # 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")),

780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
    # 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)),

805
806
807
808
809
810
811
812
    # 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"))),

813
814
815
    # Divisor for dynamic query scale factor calculation for FP8 KV Cache
    "Q_SCALE_CONSTANT":
    lambda: int(os.getenv("Q_SCALE_CONSTANT", "200")),
816
817
818
819
820
821
    # 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")),
822

823
824
    # If set, enable multiprocessing in LLM for the V1 code path.
    "VLLM_ENABLE_V1_MULTIPROCESSING":
825
    lambda: bool(int(os.getenv("VLLM_ENABLE_V1_MULTIPROCESSING", "1"))),
826
827
    "VLLM_LOG_BATCHSIZE_INTERVAL":
    lambda: float(os.getenv("VLLM_LOG_BATCHSIZE_INTERVAL", "-1")),
828
829
    "VLLM_DISABLE_COMPILE_CACHE":
    lambda: bool(int(os.getenv("VLLM_DISABLE_COMPILE_CACHE", "0"))),
830
831
832
833
834
835

    # 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"))),
836
837
838
839
840
841
842
843
844
845

    # 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")),
846
847
848
849
850

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

851
852
853
854
855
856
857
858
859
860
861
862
    # 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", ""),

863
864
865
866
    # 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),
867
868
869
870
871
872
873

    # 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"),
874

875
876
877
878
879
880
    # 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"),

881
882
883
884
    # Rank of the process in the data parallel setting
    "VLLM_DP_RANK":
    lambda: int(os.getenv("VLLM_DP_RANK", "0")),

885
886
887
888
889
890
    # 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)),

891
892
893
894
895
896
897
898
899
900
901
    # 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")),
902

903
904
905
906
907
908
909
910
    # 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")),

911
912
913
914
    # 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",

915
916
917
    # 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",
918

919
    # Use model_redirect to redirect the model name to a local folder.
920
921
922
923
924
    # `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
925
926
927
    "VLLM_MODEL_REDIRECT_PATH":
    lambda: os.environ.get("VLLM_MODEL_REDIRECT_PATH", None),

928
929
930
    # 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",
931
932
933
934
935
936

    # 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",
937

938
939
940
941
    # 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"])
942
    if "VLLM_TPU_BUCKET_PADDING_GAP" in os.environ else 0,
943
944
    "VLLM_TPU_MOST_MODEL_LEN":
    lambda: maybe_convert_int(os.environ.get("VLLM_TPU_MOST_MODEL_LEN", None)),
945
946
947
948

    # Allow use of DeepGemm kernels for fused moe ops.
    "VLLM_USE_DEEP_GEMM":
    lambda: bool(int(os.getenv("VLLM_USE_DEEP_GEMM", "0"))),
949
950
951
952
953
954

    # 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")),
955
956
957
958
959
960
961
962
963
964

    # 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")),
965
966
967
968
969
970

    # 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
971
972
973
974
975
976
977
978

    # 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")),
979
980

    # all2all backend for vllm's expert parallel communication
981
982
983
    # Available options:
    # - "naive": naive all2all implementation using all-reduce
    # - "pplx": use pplx kernels
984
985
    # - "deepep_high_throughput", use deepep high-throughput kernels
    # - "deepep_low_latency", use deepep low-latency kernels
986
987
    "VLLM_ALL2ALL_BACKEND":
    lambda: os.getenv("VLLM_ALL2ALL_BACKEND", "naive"),
988
    
yangql's avatar
yangql committed
989
990
991
    # VLLM_MOE_HT_THRESHOLD
    "VLLM_MOE_HT_THRESHOLD":
    lambda: int(os.getenv("VLLM_MOE_HT_THRESHOLD", "128")),
992
993
994
995
    # use ALLOW_MNNVL
    "VLLM_ALLOW_MNNVL":
    lambda: (os.environ.get("VLLM_ALLOW_MNNVL", "False").lower() in
             ("true", "1")),
996
997
998
999
1000
1001
1002

    # 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")),
1003
1004
1005
1006

    # 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")),
1007
1008
1009
1010
1011

    # 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"))),
1012
1013
1014
1015
1016
1017

    # 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")),
1018
    
1019
1020
1021
1022
1023
    # 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")),

1024
1025
1026
1027
1028
1029
1030
1031
    # 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":
1032
1033
1034
1035
1036
1037
1038
    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"))),
1039
1040
1041
1042
1043

    # 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
1044
1045
    lambda: bool(int(os.getenv("VLLM_USE_NVFP4_CT_EMULATIONS", "0"))),
    
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
     # 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
1061
1062
1063
1064
    # 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"))),
    
1065
1066
1067
1068
    # If set, vLLM will use FLASH ATTN fp8 attention optimizations.
    "VLLM_USE_FLASH_ATTN_FP8":
    lambda: bool(int(os.getenv("VLLM_USE_FLASH_ATTN_FP8", "0"))),
    
zhuwenwen's avatar
zhuwenwen committed
1069
1070
1071
1072
    # If set, vLLM will use FLASH MLA attention optimizations.
    "VLLM_USE_FLASH_MLA":
    lambda: bool(int(os.getenv("VLLM_USE_FLASH_MLA", "1"))),
    
1073
1074
1075
1076
    # If set, vLLM will use FLASH MLA fp8 attention optimizations.
    "VLLM_USE_FLASH_MLA_FP8":
    lambda: bool(int(os.getenv("VLLM_USE_FLASH_MLA_FP8", "0"))),
    
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
    # 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"))),
1099
1100
1101

    # flag to control vllm to use optimized kernels
    "VLLM_CUSTOM_CACHE":
1102
    lambda: bool(int(os.environ.get("VLLM_CUSTOM_CACHE", "0"))),
1103
    
zhuwenwen's avatar
zhuwenwen committed
1104
1105
1106
1107
    # 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")),
    
1108
1109
1110
    # 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")),
1111

1112
1113
1114
1115
1116
1117
    # '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":
zhuwenwen's avatar
zhuwenwen committed
1118
    lambda: bool(int(os.getenv("VLLM_HAS_CONTEXT_DEFAULT", "1"))),
1119
1120
1121
    
    # If set, vLLM will transpose weight to use nn layout
    "VLLM_USE_NN":
zhuwenwen's avatar
zhuwenwen committed
1122
    lambda: (os.environ.get("VLLM_USE_NN", "True").lower() in
1123
             ("true", "1")),
1124

1125
1126
1127
    # Enable two batch overlap.
    "VLLM_ENABLE_TBO":
    lambda: bool(int(os.getenv("VLLM_ENABLE_TBO", "0"))),
1128
    
1129
1130
1131
    # 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")),
1132

1133
1134
1135
1136
    # 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
1137
1138
1139
1140
    # 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")),

1141
1142
1143
    # Enable zero overhead scheduler.
    "VLLM_ZERO_OVERHEAD":
    lambda: bool(int(os.getenv("VLLM_ZERO_OVERHEAD", "0"))),
1144
1145
1146
1147

    # 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
1148
    
1149
1150
    # vLLM will use FlashAttention Backend for page attention computation on rocm
    "VLLM_USE_FLASH_ATTN_PA":
zhuwenwen's avatar
zhuwenwen committed
1151
    lambda: (os.environ.get("VLLM_USE_FLASH_ATTN_PA", "True").lower() in
zhuwenwen's avatar
zhuwenwen committed
1152
             ("true", "1")),
1153
    
zhuwenwen's avatar
zhuwenwen committed
1154
1155
1156
1157
    # vLLM will use apex for rmsnorm
    "VLLM_USE_APEX_RN":
    lambda: (os.environ.get("VLLM_USE_APEX_RN", "False").lower() in
             ("true", "1")),
1158
    
1159
1160
    # vLLM will use global cache for moe
    "VLLM_USE_GLOBAL_CACHE13":
zhuwenwen's avatar
zhuwenwen committed
1161
        lambda: (os.environ.get("VLLM_USE_GLOBAL_CACHE13", "False").lower() in
1162
                 ("true", "1")),
1163
        
zhuwenwen's avatar
zhuwenwen committed
1164
    # vLLM will use lightop for deepseek-v3
1165
1166
    "VLLM_USE_LIGHTOP":
        lambda: (os.environ.get("VLLM_USE_LIGHTOP", "False").lower() in
1167
                 ("true", "1")),
1168
        
1169
1170
    # vLLM will use elenmentwise not triton_
    "VLLM_USE_OPT_ZEROS":
1171
        lambda: (os.environ.get("VLLM_USE_OPT_ZEROS", "True").lower() in
1172
                 ("true", "1")),
1173
        
zhuwenwen's avatar
zhuwenwen committed
1174
    # vLLM will use opt cat for deepseek-v3
1175
    "VLLM_USE_OPT_CAT":
zhuwenwen's avatar
zhuwenwen committed
1176
        lambda: (os.environ.get("VLLM_USE_OPT_CAT", "False").lower() in
1177
                 ("true", "1")),  
1178
        
1179
1180
1181
1182
    # vLLM will use triton moe_sum 
    "VLLM_USE_OPT_MOE_SUM":
        lambda: (os.environ.get("VLLM_USE_OPT_MOE_SUM", "False").lower() in
                 ("true", "1")),  
1183
1184
        
    # vLLM will use lightop moe_sum_mul_add for deepseek-v3
1185
    "VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD":
1186
        lambda: (os.environ.get("VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD", "False").lower() in
1187
                 ("true", "1")),  
1188
1189
        
    # vLLM will use lightop moe_sum (qwen3-30b)
1190
    "VLLM_USE_LIGHTOP_MOE_SUM":
1191
        lambda: (os.environ.get("VLLM_USE_LIGHTOP_MOE_SUM", "False").lower() in
1192
                 ("true", "1")),  
1193
1194
        
    # vLLM will use lightop moe_align_block_size (qwen3-30b)
1195
    "VLLM_USE_LIGHTOP_MOE_ALIGN":
1196
        lambda: (os.environ.get("VLLM_USE_LIGHTOP_MOE_ALIGN", "False").lower() in
1197
                 ("true", "1")),    
1198
        
zhuwenwen's avatar
zhuwenwen committed
1199
    # vLLM will use opt merge_aatn_states, not triton
1200
1201
1202
    "VLLM_USE_MERGE_ATTN_STATES_OPT":
        lambda: (os.environ.get("VLLM_USE_MERGE_ATTN_STATES_OPT", "True").lower() in
                 ("true", "1")),  
1203
        
1204
1205
    # vllm will use rmsquant fused op 
    "USE_FUSED_RMS_QUANT": 
1206
    lambda: bool(int(os.getenv("USE_FUSED_RMS_QUANT", "0"))),
1207
1208
1209
1210
    # vllm will use silu_mul_quant fused op,
    # This variable has a default value of true, 
    # but it is still controlled by CRQ and RQ.
    "USE_FUSED_SILU_MUL_QUANT":
1211
    lambda: bool(int(os.getenv("USE_FUSED_SILU_MUL_QUANT", "0"))),
1212
    
1213
1214
1215
    # vllm pd separation will be used async
    "VLLM_P2P_ASYNC":
    lambda: bool(int(os.getenv("VLLM_P2P_ASYNC", "0"))),
1216

1217
1218
1219
    # pd separation p2p async buf tokens
    "VLLM_P2P_BUF_TOKENS":
    lambda: int(os.getenv("VLLM_P2P_BUF_TOKENS", "30000")),
1220

1221
1222
1223
1224
    # vllm will enable minimal injection for pipeline parallel scheduling
    "VLLM_SCHED_ENABLE_MINIMAL_INJECTION":
        lambda: (os.getenv("VLLM_SCHED_ENABLE_MINIMAL_INJECTION", "0").lower() in
                 ("true", "1")),
1225

1226
1227
    # vLLM will split prefill and decode, not mix up
    "VLLM_USE_PD_SPLIT":
zhuwenwen's avatar
zhuwenwen committed
1228
        lambda: (os.environ.get("VLLM_USE_PD_SPLIT", "False").lower() in
1229
                 ("true", "1")), 
1230

1231
1232
    # vLLM will sync to avoid pp vmfault
    "VLLM_USE_PP_SYNC":
1233
        lambda: (os.environ.get("VLLM_USE_PP_SYNC", "True").lower() in
1234
                 ("true", "1")),
1235

1236
    # vLLM will use lightop to fuse fill and moe align (dpsk-v3 + qwen3-30b)
1237
    "VLLM_USE_LIGHTOP_FILL_MOE_ALIGN":
1238
        lambda: (os.environ.get("VLLM_USE_LIGHTOP_FILL_MOE_ALIGN", "False").lower() in
zhuwenwen's avatar
zhuwenwen committed
1239
                 ("true", "1")), 
1240

1241
1242
1243
1244
    # vllm will use custom-allreduce rmsquant fused op
    "USE_FUSED_CUSTOM_ALL_REDUCE_RMS_QUANT": 
    lambda: (os.getenv('USE_FUSED_CUSTOM_ALL_REDUCE_RMS_QUANT', '0').lower() in
             ("true", "1")),
1245

1246
1247
1248
    "VLLM_USE_PP_BALANCE":
        lambda: (os.getenv('VLLM_USE_PP_BALANCE', '1').lower() in
                 ("true", "1")),
1249
1250

    "VLLM_USE_ZERO_MTP":
1251
        lambda: (os.getenv('VLLM_USE_ZERO_MTP', '1').lower() in
1252
1253
                 ("true", "1")),

1254
    # vllm will use 1-24... (not only 1 2 4 8 16 24)
1255
    "VLLM_USE_CUDA_GRAPH_SIZES":
1256
        lambda: (os.getenv('VLLM_USE_CUDA_GRAPH_SIZES', 'True').lower() in
1257
                 ("true", "1")),
1258
1259
1260
        
    # vllm will use fused cat and mla
    "VLLM_USE_CAT_MLA":
zhuwenwen's avatar
zhuwenwen committed
1261
        lambda: (os.getenv('VLLM_USE_CAT_MLA', 'False').lower() in
王敏's avatar
王敏 committed
1262
1263
1264
1265
                 ("true", "1")),

    # vllm will use fused cat and mla
    "VLLM_REJECT_SAMPLE_OPT":
zhuwenwen's avatar
zhuwenwen committed
1266
        lambda: (os.getenv('VLLM_REJECT_SAMPLE_OPT', 'True').lower() in
王敏's avatar
王敏 committed
1267
                 ("true", "1")),      
zhuwenwen's avatar
zhuwenwen committed
1268

1269
    # vLLM will use fused silu+mul kernel (fp16 + qwen3-30b)
1270
    "VLLM_USE_FUSE_SILU_AND_MUL":
zhuwenwen's avatar
zhuwenwen committed
1271
        lambda: (os.environ.get("VLLM_USE_FUSE_SILU_AND_MUL", "False").lower() in
1272
                 ("true", "1")),
1273
1274
        
     # vLLM will use optimized reshape_and_cache kernel when enabled (fp16 + qwen3-30b)
1275
1276
    "VLLM_USE_OPT_RESHAPE_AND_CACHE":
        lambda:
1277
        (os.environ.get("VLLM_USE_OPT_RESHAPE_AND_CACHE", "False").lower() in
1278
                ("true", "1")),
1279
1280
1281
1282
        
    # vLLM will use optimized topk_softmax + renormalize
    "VLLM_USE_TOPK_RENORM":
        lambda:
zhuwenwen's avatar
zhuwenwen committed
1283
        (os.environ.get("VLLM_USE_TOPK_RENORM", "True").lower() in
1284
                ("true", "1")),
1285
1286
1287
1288
    "VLLM_PP_DEBUG":
        lambda:
        (os.environ.get("VLLM_PP_DEBUG", "False").lower() in
         ("true", "1")),
1289
        
1290
1291
1292
1293
1294
    # vllm will use encoding_dsv32.py for dpsk-v32
    "VLLM_USE_V32_ENCODE":
        lambda: (os.getenv('VLLM_USE_V32_ENCODE', 'False').lower() in
                 ("true", "1")),  
        
1295
1296
1297
1298
    # vllm will use fused rmsnorm + contiguous + rope(for dpsk-v3) + concat_and_cache_mla
    "VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT":
        lambda: (os.getenv('VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT', 'False').lower() in
                 ("true", "1")),  
1299
        
laibao's avatar
laibao committed
1300
1301
1302
1303
1304
1305
1306
1307
    # vLLM will use fused RMS + RoPE kernel
    "VLLM_USE_FUSED_RMS_ROPE":
        lambda: (os.environ.get("VLLM_USE_FUSED_RMS_ROPE", "False").lower() in
                 ("true", "1")),
    # vLLM will use Marlin W16A16 kernel for MoE experts
    "VLLM_USE_MARLIN_W16A16_MOE":
        lambda: (os.environ.get("VLLM_USE_MARLIN_W16A16_MOE", "False").lower() in
                 ("true", "1")),
1308
1309
1310
    # vLLM will use lightop for dpsk mtp fill + rms*2 + cat
    "VLLM_USE_FUSED_FILL_RMS_CAT":
        lambda: (os.environ.get("VLLM_USE_FUSED_FILL_RMS_CAT", "False").lower() in
1311
                 ("true", "1")),
1312
                
1313
1314
1315
1316
1317
    # If set to 1/True, enable the reduced top-k/top-p sampling path in the
    # V1 PyTorch-native sampler.
    "VLLM_V1_USE_REDUCED_TOPK_TOPP_SAMPLER":
        lambda: (os.getenv("VLLM_V1_USE_REDUCED_TOPK_TOPP_SAMPLER",
                           "0").lower() in ("true", "1")),
1318
                 
王敏's avatar
王敏 committed
1319
1320
1321
1322
    # vLLM will use deepgemm kernel for deepep ht mode
    "VLLM_ENABLE_DEEPEP_HT_DEEPGEMM":
        lambda: (os.getenv('VLLM_ENABLE_DEEPEP_HT_DEEPGEMM', '1').lower() in
                 ("true", "1")),
1323
                 
1324
1325
1326
1327
1328
    # Only quantized DeepSeek models supported.
    # Unquantized versions are not supported.
    "VLLM_USE_FUSED_QA_KVA_GEMM":
        lambda: (os.environ.get("VLLM_USE_FUSED_QA_KVA_GEMM", "False").lower() in
                ("true", "1")),
1329
1330
1331
    "VLLM_ZERO_OVERHEAD_ENHANCE":
        lambda: (os.getenv('VLLM_ZERO_OVERHEAD_ENHANCE', '0').lower() in
                 ("true", "1")),
1332
1333
1334
1335
    # vLLM will use fast token id copy
    "VLLM_V1_FAST_TOKEN_ID_COPY":
        lambda: (os.environ.get("VLLM_V1_FAST_TOKEN_ID_COPY", "False").lower() in
                 ("true", "1")),
1336

1337
1338
1339
    # shared experts overlap with routed experts
    # VLLM_DISABLE_SHARED_EXPERTS_STREAM = 1 disable shared experts overlap
    # VLLM_DISABLE_SHARED_EXPERTS_STREAM = 0 enable shared experts overlap
1340
1341
1342
    "VLLM_DISABLE_SHARED_EXPERTS_STREAM": lambda: bool(
        int(os.getenv("VLLM_DISABLE_SHARED_EXPERTS_STREAM", "1"))
    ),
1343
1344
1345
1346
1347
1348
    # shared experts fusion
    # VLLM_ENABLE_SHARED_EXPERTS_FUSION = 1 enable shared experts fusion
    # VLLM_ENABLE_SHARED_EXPERTS_FUSION = 0 disable shared experts fusion
    "VLLM_ENABLE_SHARED_EXPERTS_FUSION": lambda: bool(
        int(os.getenv("VLLM_ENABLE_SHARED_EXPERTS_FUSION", "0"))
    ),
wujl5's avatar
wujl5 committed
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386

    # W8A8 GEMM backend selection for vLLM quantized models.
    # lightop/triton: 1
    # cutlass: 2 (will remove in the future)
    # blaslt: 3 (default)
    # rocblas: others
    "VLLM_W8A8_BACKEND": lambda: int(os.getenv("VLLM_W8A8_BACKEND", "3")),

    # Capture MoE router logits for debugging/analysis.
    "VLLM_MOE_ROUTER_CAPTURE":
    lambda: (os.getenv("VLLM_MOE_ROUTER_CAPTURE", "0").lower() in ("true", "1")),
    # Output directory for MoE router capture dumps.
    "VLLM_MOE_ROUTER_CAPTURE_DIR":
    lambda: os.environ.get(
        "VLLM_MOE_ROUTER_CAPTURE_DIR",
        "/tmp",
    ),
    # Capture only the specified rank; set to -1 to capture all ranks.
    "VLLM_MOE_ROUTER_CAPTURE_RANK":
    lambda: int(os.environ.get("VLLM_MOE_ROUTER_CAPTURE_RANK", "-1")),
    # Max number of MoE layers to record per process (0 = unlimited).
    "VLLM_MOE_ROUTER_CAPTURE_MAX_LAYERS":
    lambda: int(os.environ.get("VLLM_MOE_ROUTER_CAPTURE_MAX_LAYERS", "0")),
    # Only capture when num_tokens > N (negative disables).
    "VLLM_MOE_ROUTER_CAPTURE_NUM_TOKENS_GT":
    lambda: int(os.environ.get("VLLM_MOE_ROUTER_CAPTURE_NUM_TOKENS_GT", "-1")),
    # Only capture when num_tokens < N (0 disables).
    "VLLM_MOE_ROUTER_CAPTURE_NUM_TOKENS_LT":
    lambda: int(os.environ.get("VLLM_MOE_ROUTER_CAPTURE_NUM_TOKENS_LT", "-1")),
    # Force using Triton MoE path (disable Marlin W16A16 MoE).
    "VLLM_USE_MOE_W16A16_TRITON":
        lambda: (os.environ.get("VLLM_USE_MOE_W16A16_TRITON", "0").lower() in
                 ("true", "1")),
    
    # Only quantized DeepSeek models supported.
    "VLLM_USE_FUSED_DTBMM":
        lambda: (os.environ.get("VLLM_USE_FUSED_DTBMM", "False").lower() in
                ("true", "1")),
1387
1388
}

1389
# --8<-- [end:env-vars-definition]
1390

1391

1392
def __getattr__(name: str):
1393
1394
1395
1396
1397
1398
1399
1400
    # 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())
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416


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"
1417
1418
1419
1420
1421
1422
1423
1424


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
1425
    graphs, so it is included in the factors list. The env vars that
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
    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",
1449
        "VLLM_USE_STANDALONE_COMPILE",
1450
        "VLLM_FUSED_MOE_CHUNK_SIZE",
1451
1452
1453
1454
1455
    ]
    for key in environment_variables_to_hash:
        if key in environment_variables:
            factorize(key)

1456
1457
    hash_str = hashlib.md5(str(factors).encode(),
                           usedforsecurity=False).hexdigest()
1458

zhuwenwen's avatar
zhuwenwen committed
1459
    return hash_str