envs.py 81.6 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import functools
5
import json
6
import logging
7
import os
8
import sys
9
import tempfile
10
import uuid
11
12
from collections.abc import Callable
from typing import TYPE_CHECKING, Any, Literal
13
14
15

if TYPE_CHECKING:
    VLLM_HOST_IP: str = ""
16
    VLLM_PORT: int | None = None
17
    VLLM_RPC_BASE_PATH: str = tempfile.gettempdir()
18
    VLLM_USE_MODELSCOPE: bool = False
19
    VLLM_RINGBUFFER_WARNING_INTERVAL: int = 60
20
21
    VLLM_NCCL_SO_PATH: str | None = None
    LD_LIBRARY_PATH: str | None = None
22
    VLLM_ROCM_SLEEP_MEM_CHUNK_SIZE: int = 256
23
    LOCAL_RANK: int = 0
24
    CUDA_VISIBLE_DEVICES: str | None = None
25
    VLLM_ENGINE_ITERATION_TIMEOUT_S: int = 60
26
    VLLM_ENGINE_READY_TIMEOUT_S: int = 600
27
    VLLM_API_KEY: str | None = None
28
    VLLM_DEBUG_LOG_API_SERVER_RESPONSE: bool = False
29
30
31
32
    S3_ACCESS_KEY_ID: str | None = None
    S3_SECRET_ACCESS_KEY: str | None = None
    S3_ENDPOINT_URL: str | None = None
    VLLM_MODEL_REDIRECT_PATH: str | None = None
33
34
    VLLM_CACHE_ROOT: str = os.path.expanduser("~/.cache/vllm")
    VLLM_CONFIG_ROOT: str = os.path.expanduser("~/.config/vllm")
35
36
37
    VLLM_USAGE_STATS_SERVER: str = "https://stats.vllm.ai"
    VLLM_NO_USAGE_STATS: bool = False
    VLLM_DO_NOT_TRACK: bool = False
38
    VLLM_USAGE_SOURCE: str = "production"
39
    VLLM_CONFIGURE_LOGGING: bool = True
40
    VLLM_LOGGING_LEVEL: str = "INFO"
41
    VLLM_LOGGING_PREFIX: str = ""
42
    VLLM_LOGGING_STREAM: str = "ext://sys.stdout"
43
    VLLM_LOGGING_CONFIG_PATH: str | None = None
Nick Hill's avatar
Nick Hill committed
44
45
    VLLM_LOGGING_COLOR: str = "auto"
    NO_COLOR: bool = False
46
    VLLM_LOG_STATS_INTERVAL: float = 10.0
47
    VLLM_TRACE_FUNCTION: int = 0
48
49
50
    VLLM_USE_FLASHINFER_SAMPLER: bool | None = None
    VLLM_PP_LAYER_PARTITION: str | None = None
    VLLM_CPU_KVCACHE_SPACE: int | None = 0
51
    VLLM_CPU_OMP_THREADS_BIND: str = "auto"
52
    VLLM_CPU_NUM_OF_RESERVED_CPU: int | None = None
53
    VLLM_CPU_SGL_KERNEL: bool = False
54
    VLLM_ZENTORCH_WEIGHT_PREPACK: bool = True
55
    VLLM_XLA_CACHE_PATH: str = os.path.join(VLLM_CACHE_ROOT, "xla_cache")
56
    VLLM_XLA_CHECK_RECOMPILATION: bool = False
57
    VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE: Literal["auto", "nccl", "shm"] = "auto"
58
    VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM: bool = False
59
    VLLM_USE_RAY_WRAPPED_PP_COMM: bool = True
60
    VLLM_XLA_USE_SPMD: bool = False
61
    VLLM_WORKER_MULTIPROC_METHOD: Literal["fork", "spawn"] = "fork"
62
    VLLM_ASSETS_CACHE: str = os.path.join(VLLM_CACHE_ROOT, "assets")
63
    VLLM_ASSETS_CACHE_MODEL_CLEAN: bool = False
64
    VLLM_IMAGE_FETCH_TIMEOUT: int = 5
65
    VLLM_VIDEO_FETCH_TIMEOUT: int = 30
66
    VLLM_AUDIO_FETCH_TIMEOUT: int = 10
67
    VLLM_MEDIA_FETCH_MAX_RETRIES: int = 3
68
    VLLM_MEDIA_URL_ALLOW_REDIRECTS: bool = True
69
    VLLM_MEDIA_LOADING_THREAD_COUNT: int = 8
70
    VLLM_MAX_AUDIO_CLIP_FILESIZE_MB: int = 25
71
    VLLM_VIDEO_LOADER_BACKEND: str = "opencv"
72
    VLLM_MEDIA_CONNECTOR: str = "http"
73
    VLLM_MM_HASHER_ALGORITHM: str = "blake3"
74
    VLLM_TARGET_DEVICE: str = "cuda"
75
    VLLM_MAIN_CUDA_VERSION: str = "12.9"
76
    VLLM_FLOAT32_MATMUL_PRECISION: Literal["highest", "high", "medium"] = "highest"
77
    VLLM_BATCH_INVARIANT: bool = False
78
79
    MAX_JOBS: str | None = None
    NVCC_THREADS: str | None = None
80
    VLLM_USE_PRECOMPILED: bool = False
81
    VLLM_SKIP_PRECOMPILED_VERSION_SUFFIX: bool = False
82
    VLLM_DOCKER_BUILD_CONTEXT: bool = False
83
    VLLM_KEEP_ALIVE_ON_ENGINE_DEATH: bool = False
84
    CMAKE_BUILD_TYPE: Literal["Debug", "Release", "RelWithDebInfo"] | None = None
85
    VERBOSE: bool = False
86
    VLLM_ALLOW_LONG_MAX_MODEL_LEN: bool = False
87
    VLLM_RPC_TIMEOUT: int = 10000  # ms
88
    VLLM_HTTP_TIMEOUT_KEEP_ALIVE: int = 5  # seconds
89
90
    VLLM_PLUGINS: list[str] | None = None
    VLLM_LORA_RESOLVER_CACHE_DIR: str | None = None
91
    VLLM_LORA_RESOLVER_HF_REPO_LIST: str | None = None
92
    VLLM_USE_AOT_COMPILE: bool = False
93
    VLLM_USE_BYTECODE_HOOK: bool = True
94
    VLLM_FORCE_AOT_LOAD: bool = False
95
    VLLM_USE_MEGA_AOT_ARTIFACT: bool = False
96
    VLLM_USE_TRITON_AWQ: bool = False
97
    VLLM_ALLOW_RUNTIME_LORA_UPDATING: bool = False
98
    VLLM_SKIP_P2P_CHECK: bool = False
99
    VLLM_DISABLED_KERNELS: list[str] = []
100
    VLLM_ENABLE_FLA_PACKED_RECURRENT_DECODE: bool = True
101
    VLLM_DISABLE_PYNCCL: bool = False
102
    VLLM_USE_OINK_OPS: bool = False
103
    VLLM_ROCM_USE_AITER: bool = False
104
    VLLM_ROCM_USE_AITER_PAGED_ATTN: bool = False
105
    VLLM_ROCM_USE_AITER_LINEAR: bool = True
106
    VLLM_ROCM_USE_AITER_MOE: bool = True
107
    VLLM_ROCM_USE_AITER_RMSNORM: bool = True
108
    VLLM_ROCM_USE_AITER_MLA: bool = True
109
    VLLM_ROCM_USE_AITER_MHA: bool = True
110
    VLLM_ROCM_USE_AITER_FP4_ASM_GEMM: bool = False
111
    VLLM_ROCM_USE_AITER_TRITON_ROPE: bool = False
112
    VLLM_ROCM_USE_AITER_FP8BMM: bool = True
113
    VLLM_ROCM_USE_AITER_FP4BMM: bool = True
114
    VLLM_ROCM_USE_AITER_UNIFIED_ATTENTION: bool = False
115
    VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS: bool = False
116
    VLLM_ROCM_USE_AITER_TRITON_GEMM: bool = True
117
    VLLM_ROCM_USE_SKINNY_GEMM: bool = True
118
    VLLM_ROCM_FP8_PADDING: bool = True
119
    VLLM_ROCM_MOE_PADDING: bool = True
120
    VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT: bool = False
121
    VLLM_ENABLE_V1_MULTIPROCESSING: bool = True
122
    VLLM_LOG_BATCHSIZE_INTERVAL: float = -1
123
    VLLM_DISABLE_COMPILE_CACHE: bool = False
124
    Q_SCALE_CONSTANT: int = 200
125
126
    K_SCALE_CONSTANT: int = 200
    V_SCALE_CONSTANT: int = 100
127
    VLLM_SERVER_DEV_MODE: bool = False
128
    VLLM_V1_OUTPUT_PROC_CHUNK_SIZE: int = 128
129
    VLLM_MLA_DISABLE: bool = False
130
131
    VLLM_RAY_PER_WORKER_GPUS: float = 1.0
    VLLM_RAY_BUNDLE_INDICES: str = ""
132
    VLLM_CUDART_SO_PATH: str | None = None
133
    VLLM_DP_RANK: int = 0
134
    VLLM_DP_RANK_LOCAL: int = -1
135
    VLLM_DP_SIZE: int = 1
136
    VLLM_USE_STANDALONE_COMPILE: bool = True
137
    VLLM_ENABLE_PREGRAD_PASSES: bool = False
138
139
    VLLM_DP_MASTER_IP: str = ""
    VLLM_DP_MASTER_PORT: int = 0
140
    VLLM_MOE_DP_CHUNK_SIZE: int = 256
141
    VLLM_ENABLE_MOE_DP_CHUNK: bool = True
142
    VLLM_RANDOMIZE_DP_DUMMY_INPUTS: bool = False
143
    VLLM_RAY_DP_PACK_STRATEGY: Literal["strict", "fill", "span"] = "strict"
144
145
    VLLM_RAY_EXTRA_ENV_VAR_PREFIXES_TO_COPY: str = ""
    VLLM_RAY_EXTRA_ENV_VARS_TO_COPY: str = ""
146
    VLLM_MARLIN_USE_ATOMIC_ADD: bool = False
147
    VLLM_MARLIN_INPUT_DTYPE: Literal["int8", "fp8"] | None = None
148
    VLLM_MXFP4_USE_MARLIN: bool | None = None
149
    VLLM_DEEPEPLL_NVFP4_DISPATCH: bool = False
150
    VLLM_V1_USE_OUTLINES_CACHE: bool = False
151
    VLLM_TPU_BUCKET_PADDING_GAP: int = 0
152
    VLLM_TPU_MOST_MODEL_LEN: int | None = None
153
    VLLM_TPU_USING_PATHWAYS: bool = False
154
    VLLM_USE_DEEP_GEMM: bool = True
155
    VLLM_MOE_USE_DEEP_GEMM: bool = True
156
    VLLM_USE_DEEP_GEMM_E8M0: bool = True
157
    VLLM_USE_DEEP_GEMM_TMA_ALIGNED_SCALES: bool = True
158
159
160
161
162
    VLLM_DEEP_GEMM_WARMUP: Literal[
        "skip",
        "full",
        "relax",
    ] = "relax"
163
    VLLM_USE_FUSED_MOE_GROUPED_TOPK: bool = True
164
    VLLM_BLOCKSCALE_FP8_GEMM_FLASHINFER: bool = True
165
    VLLM_USE_FLASHINFER_MOE_FP16: bool = False
166
167
    VLLM_USE_FLASHINFER_MOE_FP8: bool = False
    VLLM_USE_FLASHINFER_MOE_FP4: bool = False
168
    VLLM_USE_FLASHINFER_MOE_INT4: bool = False
169
170
171
    VLLM_FLASHINFER_MOE_BACKEND: Literal["throughput", "latency", "masked_gemm"] = (
        "latency"
    )
172
    VLLM_FLASHINFER_ALLREDUCE_BACKEND: Literal["auto", "trtllm", "mnnvl"] = "trtllm"
173
    VLLM_FLASHINFER_WORKSPACE_BUFFER_SIZE: int = 394 * 1024 * 1024
174
    VLLM_XGRAMMAR_CACHE_MB: int = 0
175
    VLLM_MSGPACK_ZERO_COPY_THRESHOLD: int = 256
176
    VLLM_ALLOW_INSECURE_SERIALIZATION: bool = False
177
    VLLM_DISABLE_REQUEST_ID_RANDOMIZATION: bool = False
Robert Shaw's avatar
Robert Shaw committed
178
    VLLM_NIXL_SIDE_CHANNEL_HOST: str = "localhost"
179
    VLLM_NIXL_SIDE_CHANNEL_PORT: int = 5600
180
    VLLM_MOONCAKE_BOOTSTRAP_PORT: int = 8998
181
    VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE: int = 163840
182
    VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS: int = 1
183
    VLLM_MQ_MAX_CHUNK_BYTES_MB: int = 16
184
    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: int = 300
185
    VLLM_KV_CACHE_LAYOUT: Literal["NHD", "HND"] | None = None
186
    VLLM_COMPUTE_NANS_IN_LOGITS: bool = False
187
    VLLM_USE_NVFP4_CT_EMULATIONS: bool = False
188
189
190
    VLLM_ROCM_QUICK_REDUCE_QUANTIZATION: Literal[
        "FP", "INT8", "INT6", "INT4", "NONE"
    ] = "NONE"
191
    VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16: bool = True
192
    VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB: int | None = None
193
    VLLM_NIXL_ABORT_REQUEST_TIMEOUT: int = 480
194
195
196
197
    VLLM_MORIIO_CONNECTOR_READ_MODE: bool = False
    VLLM_MORIIO_QP_PER_TRANSFER: int = 1
    VLLM_MORIIO_POST_BATCH_SIZE: int = -1
    VLLM_MORIIO_NUM_WORKERS: int = 1
198
    VLLM_MOONCAKE_ABORT_REQUEST_TIMEOUT: int = 480
199
    VLLM_ENABLE_CUDAGRAPH_GC: bool = False
200
    VLLM_LOOPBACK_IP: str = ""
201
    VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE: bool = True
202
    VLLM_ENABLE_RESPONSES_API_STORE: bool = False
203
    VLLM_NVFP4_GEMM_BACKEND: str | None = None
204
    VLLM_HAS_FLASHINFER_CUBIN: bool = False
205
206
    VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8: bool = False
    VLLM_USE_FLASHINFER_MOE_MXFP4_BF16: bool = False
xiao-llm's avatar
xiao-llm committed
207
    VLLM_ROCM_FP8_MFMA_PAGE_ATTN: bool = False
208
    VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS: bool = False
209
    VLLM_ALLREDUCE_USE_SYMM_MEM: bool = True
210
    VLLM_ALLREDUCE_USE_FLASHINFER: bool = False
211
    VLLM_TUNED_CONFIG_FOLDER: str | None = None
212
    VLLM_GPT_OSS_SYSTEM_TOOL_MCP_LABELS: set[str] = set()
213
    VLLM_USE_EXPERIMENTAL_PARSER_CONTEXT: bool = False
214
    VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS: bool = False
215
    VLLM_SYSTEM_START_DATE: str | None = None
216
    VLLM_TOOL_JSON_ERROR_AUTOMATIC_RETRY: bool = False
217
    VLLM_CUSTOM_SCOPES_FOR_PROFILING: bool = False
218
    VLLM_NVTX_SCOPES_FOR_PROFILING: bool = False
219
    VLLM_KV_EVENTS_USE_INT_BLOCK_HASHES: bool = True
220
    VLLM_OBJECT_STORAGE_SHM_BUFFER_NAME: str = "VLLM_OBJECT_STORAGE_SHM_BUFFER"
221
    VLLM_DEEPEP_BUFFER_SIZE_MB: int = 1024
222
223
    VLLM_DEEPEP_HIGH_THROUGHPUT_FORCE_INTRA_NODE: bool = False
    VLLM_DEEPEP_LOW_LATENCY_USE_MNNVL: bool = False
224
    VLLM_DBO_COMM_SMS: int = 20
225
226
    VLLM_PATTERN_MATCH_DEBUG: str | None = None
    VLLM_DEBUG_DUMP_PATH: str | None = None
227
228
    VLLM_ENABLE_INDUCTOR_MAX_AUTOTUNE: bool = True
    VLLM_ENABLE_INDUCTOR_COORDINATE_DESCENT_TUNING: bool = True
229
    VLLM_USE_NCCL_SYMM_MEM: bool = False
230
    VLLM_NCCL_INCLUDE_PATH: str | None = None
231
    VLLM_USE_FBGEMM: bool = False
232
    VLLM_GC_DEBUG: str = ""
233
    VLLM_DEBUG_WORKSPACE: bool = False
234
    VLLM_DISABLE_SHARED_EXPERTS_STREAM: bool = False
235
    VLLM_SHARED_EXPERTS_STREAM_TOKEN_THRESHOLD: int = 256
236
    VLLM_COMPILE_CACHE_SAVE_FORMAT: Literal["binary", "unpacked"] = "binary"
Woosuk Kwon's avatar
Woosuk Kwon committed
237
    VLLM_USE_V2_MODEL_RUNNER: bool = False
238
    VLLM_LOG_MODEL_INSPECTION: bool = False
239
    VLLM_DEBUG_MFU_METRICS: bool = False
240
241
    VLLM_WEIGHT_OFFLOADING_DISABLE_PIN_MEMORY: bool = False
    VLLM_WEIGHT_OFFLOADING_DISABLE_UVA: bool = False
242
    VLLM_DISABLE_LOG_LOGO: bool = False
243
    VLLM_LORA_DISABLE_PDL: bool = False
244
245
    VLLM_ENABLE_CUDA_COMPATIBILITY: bool = False
    VLLM_CUDA_COMPATIBILITY_PATH: str | None = None
246
247
    VLLM_ELASTIC_EP_SCALE_UP_LAUNCH: bool = False
    VLLM_ELASTIC_EP_DRAIN_REQUESTS: bool = False
248
    VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS: bool = False
249
    VLLM_NIXL_EP_MAX_NUM_RANKS: int = 32
250
    VLLM_XPU_ENABLE_XPU_GRAPH: bool = False
251

252
253
254
255
256
257
258
259
260
261
262
263
264
265
266

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"),
    )


267
def maybe_convert_int(value: str | None) -> int | None:
268
269
270
271
272
    if value is None:
        return None
    return int(value)


273
def maybe_convert_bool(value: str | None) -> bool | None:
274
275
276
277
278
    if value is None:
        return None
    return bool(int(value))


279
280
281
282
def disable_compile_cache() -> bool:
    return bool(int(os.getenv("VLLM_DISABLE_COMPILE_CACHE", "0")))


283
def use_aot_compile() -> bool:
284
    from vllm.utils.torch_utils import is_torch_equal_or_newer
285

286
287
    default_value = (
        "1"
288
        if is_torch_equal_or_newer("2.10.0") and not disable_compile_cache()
289
290
291
        else "0"
    )

292
    return (
293
        not bool(int(os.getenv("VLLM_BATCH_INVARIANT", "0")))
294
295
        and os.environ.get("VLLM_USE_AOT_COMPILE", default_value) == "1"
    )
296
297


298
299
300
301
302
303
304
305
306
307
def use_mega_aot_artifact():
    from vllm.utils.torch_utils import is_torch_equal_or_newer

    default_value = (
        "1" if is_torch_equal_or_newer("2.12.0.dev") and use_aot_compile() else "0"
    )

    return os.environ.get("VLLM_USE_MEGA_AOT_ARTIFACT", default_value) == "1"


308
def env_with_choices(
309
    env_name: str,
310
311
    default: str | None,
    choices: list[str] | Callable[[], list[str]],
312
    case_sensitive: bool = True,
313
) -> Callable[[], str | None]:
314
315
    """
    Create a lambda that validates environment variable against allowed choices
316

317
318
319
320
321
    Args:
        env_name: Name of the environment variable
        default: Default value if not set (can be None)
        choices: List of valid string options or callable that returns list
        case_sensitive: Whether validation should be case sensitive
322

323
324
325
326
    Returns:
        Lambda function for environment_variables dict
    """

327
    def _get_validated_env() -> str | None:
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
        value = os.getenv(env_name)
        if value is None:
            return default

        # Resolve choices if it's a callable (for lazy loading)
        actual_choices = choices() if callable(choices) else choices

        if not case_sensitive:
            check_value = value.lower()
            check_choices = [choice.lower() for choice in actual_choices]
        else:
            check_value = value
            check_choices = actual_choices

        if check_value not in check_choices:
343
344
345
346
            raise ValueError(
                f"Invalid value '{value}' for {env_name}. "
                f"Valid options: {actual_choices}."
            )
347
348
349
350
351
352

        return value

    return _get_validated_env


353
def env_list_with_choices(
354
355
    env_name: str,
    default: list[str],
356
    choices: list[str] | Callable[[], list[str]],
357
358
    case_sensitive: bool = True,
) -> Callable[[], list[str]]:
359
    """
360
    Create a lambda that validates environment variable
361
    containing comma-separated values against allowed choices
362

363
364
365
366
367
    Args:
        env_name: Name of the environment variable
        default: Default list of values if not set
        choices: List of valid string options or callable that returns list
        case_sensitive: Whether validation should be case sensitive
368

369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
    Returns:
        Lambda function for environment_variables
        dict that returns list of strings
    """

    def _get_validated_env_list() -> list[str]:
        value = os.getenv(env_name)
        if value is None:
            return default

        # Split comma-separated values and strip whitespace
        values = [v.strip() for v in value.split(",") if v.strip()]

        if not values:
            return default

        # Resolve choices if it's a callable (for lazy loading)
        actual_choices = choices() if callable(choices) else choices

        # Validate each value
        for val in values:
            if not case_sensitive:
                check_value = val.lower()
                check_choices = [choice.lower() for choice in actual_choices]
            else:
                check_value = val
                check_choices = actual_choices

            if check_value not in check_choices:
398
399
400
401
                raise ValueError(
                    f"Invalid value '{val}' in {env_name}. "
                    f"Valid options: {actual_choices}."
                )
402
403
404
405
406
407

        return values

    return _get_validated_env_list


408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
def env_set_with_choices(
    env_name: str,
    default: list[str],
    choices: list[str] | Callable[[], list[str]],
    case_sensitive: bool = True,
) -> Callable[[], set[str]]:
    """
    Creates a lambda which that validates environment variable
    containing comma-separated values against allowed choices which
    returns choices as a set.
    """

    def _get_validated_env_set() -> set[str]:
        return set(env_list_with_choices(env_name, default, choices, case_sensitive)())

    return _get_validated_env_set


426
def get_vllm_port() -> int | None:
427
    """Get the port from VLLM_PORT environment variable.
428

429
430
    Returns:
        The port number as an integer if VLLM_PORT is set, None otherwise.
431

432
433
434
    Raises:
        ValueError: If VLLM_PORT is a URI, suggest k8s service discovery issue.
    """
435
    if "VLLM_PORT" not in os.environ:
436
437
        return None

438
    port = os.getenv("VLLM_PORT", "0")
439
440
441
442

    try:
        return int(port)
    except ValueError as err:
443
        from urllib3.util import parse_url
444

445
        parsed = parse_url(port)
446
447
448
449
450
451
        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
452
        raise ValueError(f"VLLM_PORT '{port}' must be a valid integer") from err
453
454


455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
def get_env_or_set_default(
    env_name: str,
    default_factory: Callable[[], str],
) -> Callable[[], str]:
    """
    Create a lambda that returns an environment variable value if set,
    or generates and sets a default value using the provided factory function.
    """

    def _get_or_set_default() -> str:
        value = os.getenv(env_name)
        if value is not None:
            return value

        default_value = default_factory()
        os.environ[env_name] = default_value
        return default_value

    return _get_or_set_default


Ning Xie's avatar
Ning Xie committed
476
# The start-* and end* here are used by the documentation generator
477
478
# to extract the used env vars.

479
# --8<-- [start:env-vars-definition]
480

481
482
logger = logging.getLogger(__name__)

483
environment_variables: dict[str, Callable[[], Any]] = {
484
    # ================== Installation Time Env Vars ==================
485
    # Target device of vLLM, supporting [cuda (by default),
486
    # rocm, cpu]
487
    "VLLM_TARGET_DEVICE": lambda: os.getenv("VLLM_TARGET_DEVICE", "cuda").lower(),
488
    # Main CUDA version of vLLM. This follows PyTorch but can be overridden.
489
    "VLLM_MAIN_CUDA_VERSION": lambda: os.getenv("VLLM_MAIN_CUDA_VERSION", "").lower()
490
    or "12.9",
491
    # Controls PyTorch float32 matmul precision mode within vLLM workers.
492
    # Valid options mirror torch.set_float32_matmul_precision
493
494
    "VLLM_FLOAT32_MATMUL_PRECISION": env_with_choices(
        "VLLM_FLOAT32_MATMUL_PRECISION",
495
496
        "highest",
        ["highest", "high", "medium"],
497
498
        case_sensitive=False,
    ),
499
500
501
    # Enable batch-invariant mode: deterministic results regardless of
    # batch composition. Requires NVIDIA GPU with compute capability >= 9.0.
    "VLLM_BATCH_INVARIANT": lambda: bool(int(os.getenv("VLLM_BATCH_INVARIANT", "0"))),
502
503
    # Maximum number of compilation jobs to run in parallel.
    # By default this is the number of CPUs
504
    "MAX_JOBS": lambda: os.getenv("MAX_JOBS", None),
505
506
507
    # Number of threads to use for nvcc
    # By default this is 1.
    # If set, `MAX_JOBS` will be reduced to avoid oversubscribing the CPU.
508
    "NVCC_THREADS": lambda: os.getenv("NVCC_THREADS", None),
509
    # If set, vllm will use precompiled binaries (*.so)
510
511
512
513
514
    "VLLM_USE_PRECOMPILED": lambda: os.environ.get("VLLM_USE_PRECOMPILED", "")
    .strip()
    .lower()
    in ("1", "true")
    or bool(os.environ.get("VLLM_PRECOMPILED_WHEEL_LOCATION")),
515
516
517
518
    # If set, skip adding +precompiled suffix to version string
    "VLLM_SKIP_PRECOMPILED_VERSION_SUFFIX": lambda: bool(
        int(os.environ.get("VLLM_SKIP_PRECOMPILED_VERSION_SUFFIX", "0"))
    ),
519
520
    # Used to mark that setup.py is running in a Docker build context,
    # in order to force the use of precompiled binaries.
521
522
523
524
    "VLLM_DOCKER_BUILD_CONTEXT": lambda: os.environ.get("VLLM_DOCKER_BUILD_CONTEXT", "")
    .strip()
    .lower()
    in ("1", "true"),
525
526
527
    # CMake build type
    # If not set, defaults to "Debug" or "RelWithDebInfo"
    # Available options: "Debug", "Release", "RelWithDebInfo"
528
529
530
    "CMAKE_BUILD_TYPE": env_with_choices(
        "CMAKE_BUILD_TYPE", None, ["Debug", "Release", "RelWithDebInfo"]
    ),
531
    # If set, vllm will print verbose logs during installation
532
    "VERBOSE": lambda: bool(int(os.getenv("VERBOSE", "0"))),
533
    # Root directory for vLLM configuration files
534
    # Defaults to `~/.config/vllm` unless `XDG_CONFIG_HOME` is set
535
536
537
    # 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**.
538
    "VLLM_CONFIG_ROOT": lambda: os.path.expanduser(
539
540
541
        os.getenv(
            "VLLM_CONFIG_ROOT",
            os.path.join(get_default_config_root(), "vllm"),
542
543
        )
    ),
544
    # ================== Runtime Env Vars ==================
545
    # Root directory for vLLM cache files
546
    # Defaults to `~/.cache/vllm` unless `XDG_CACHE_HOME` is set
547
    "VLLM_CACHE_ROOT": lambda: os.path.expanduser(
548
549
550
        os.getenv(
            "VLLM_CACHE_ROOT",
            os.path.join(get_default_cache_root(), "vllm"),
551
552
        )
    ),
553
554
555
556
    # 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.
557
    "VLLM_HOST_IP": lambda: os.getenv("VLLM_HOST_IP", ""),
558
    # used in distributed environment to manually set the communication port
559
560
561
    # 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.
562
    "VLLM_PORT": get_vllm_port,
563
564
    # path used for ipc when the frontend api server is running in
    # multi-processing mode to communicate with the backend engine process.
565
566
567
    "VLLM_RPC_BASE_PATH": lambda: os.getenv(
        "VLLM_RPC_BASE_PATH", tempfile.gettempdir()
    ),
568
569
    # If true, will load models from ModelScope instead of Hugging Face Hub.
    # note that the value is true or false, not numbers
570
571
572
573
    "VLLM_USE_MODELSCOPE": lambda: os.environ.get(
        "VLLM_USE_MODELSCOPE", "False"
    ).lower()
    == "true",
574
    # Interval in seconds to log a warning message when the ring buffer is full
575
576
577
    "VLLM_RINGBUFFER_WARNING_INTERVAL": lambda: int(
        os.environ.get("VLLM_RINGBUFFER_WARNING_INTERVAL", "60")
    ),
578
579
    # path to cudatoolkit home directory, under which should be bin, include,
    # and lib directories.
580
    "CUDA_HOME": lambda: os.environ.get("CUDA_HOME", None),
581
582
    # 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
583
    "VLLM_NCCL_SO_PATH": lambda: os.environ.get("VLLM_NCCL_SO_PATH", None),
584
585
    # 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`
586
    "LD_LIBRARY_PATH": lambda: os.environ.get("LD_LIBRARY_PATH", None),
587
588
589
590
    # flag to control the chunk size (in MB) for sleeping memory allocations under ROCm
    "VLLM_ROCM_SLEEP_MEM_CHUNK_SIZE": lambda: int(
        os.environ.get("VLLM_ROCM_SLEEP_MEM_CHUNK_SIZE", "256")
    ),
591
    # Feature flag to enable/disable Inductor standalone compile.
592
593
    # In torch <= 2.7 we ignore this flag; in torch >= 2.9 this is
    # enabled by default.
594
    "VLLM_USE_STANDALONE_COMPILE": lambda: os.environ.get(
595
        "VLLM_USE_STANDALONE_COMPILE", "1"
596
597
    )
    == "1",
598
599
600
601
602
603
604
605
606
    # Inductor's pre-grad passes don't do anything for vLLM.
    # The pre-grad passes get run even on cache-hit and negatively impact
    # vllm cold compile times by O(1s)
    # Can remove this after the following issue gets fixed
    # https://github.com/pytorch/pytorch/issues/174502
    "VLLM_ENABLE_PREGRAD_PASSES": lambda: os.environ.get(
        "VLLM_ENABLE_PREGRAD_PASSES", "0"
    )
    == "1",
607
608
    # Debug pattern matching inside custom passes.
    # Should be set to the fx.Node name (e.g. 'getitem_34' or 'scaled_mm_3').
609
610
611
    "VLLM_PATTERN_MATCH_DEBUG": lambda: os.environ.get(
        "VLLM_PATTERN_MATCH_DEBUG", None
    ),
612
613
    # Dump fx graphs to the given directory.
    # It will override CompilationConfig.debug_dump_path if set.
614
    "VLLM_DEBUG_DUMP_PATH": lambda: os.environ.get("VLLM_DEBUG_DUMP_PATH", None),
615
616
617
618
    # Feature flag to enable/disable AOT compilation. This will ensure
    # compilation is done in warmup phase and the compilation will be
    # reused in subsequent calls.
    "VLLM_USE_AOT_COMPILE": use_aot_compile,
619
620
621
622
623
    # Feature flag to enable/disable bytecode in
    # TorchCompileWithNoGuardsWrapper.
    "VLLM_USE_BYTECODE_HOOK": lambda: bool(
        int(os.environ.get("VLLM_USE_BYTECODE_HOOK", "1"))
    ),
624
625
626
627
    # Force vllm to always load AOT compiled models from disk. Failure
    # to load will result in a hard error when this is enabled.
    # Will be ignored when VLLM_USE_AOT_COMPILE is disabled.
    "VLLM_FORCE_AOT_LOAD": lambda: os.environ.get("VLLM_FORCE_AOT_LOAD", "0") == "1",
628
629
630
    # Enable loading compiled models directly from cached standalone compile artifacts
    # without re-splitting graph modules. This reduces overhead during model
    # loading by using reconstruct_serializable_fn_from_mega_artifact.
631
    "VLLM_USE_MEGA_AOT_ARTIFACT": use_mega_aot_artifact,
632
633
    # local rank of the process in the distributed setting, used to determine
    # the GPU device id
634
    "LOCAL_RANK": lambda: int(os.environ.get("LOCAL_RANK", "0")),
635
    # used to control the visible devices in the distributed setting
636
    "CUDA_VISIBLE_DEVICES": lambda: os.environ.get("CUDA_VISIBLE_DEVICES", None),
637
    # timeout for each iteration in the engine
638
639
640
    "VLLM_ENGINE_ITERATION_TIMEOUT_S": lambda: int(
        os.environ.get("VLLM_ENGINE_ITERATION_TIMEOUT_S", "60")
    ),
641
642
643
644
645
    # Timeout in seconds for waiting for engine cores to become ready
    # during startup. Default is 600 seconds (10 minutes).
    "VLLM_ENGINE_READY_TIMEOUT_S": lambda: int(
        os.environ.get("VLLM_ENGINE_READY_TIMEOUT_S", "600")
    ),
646
    # API key for vLLM API server
647
    "VLLM_API_KEY": lambda: os.environ.get("VLLM_API_KEY", None),
648
    # Whether to log responses from API Server for debugging
649
650
651
652
    "VLLM_DEBUG_LOG_API_SERVER_RESPONSE": lambda: os.environ.get(
        "VLLM_DEBUG_LOG_API_SERVER_RESPONSE", "False"
    ).lower()
    == "true",
653
    # S3 access information, used for tensorizer to load model from S3
654
655
656
    "S3_ACCESS_KEY_ID": lambda: os.environ.get("S3_ACCESS_KEY_ID", None),
    "S3_SECRET_ACCESS_KEY": lambda: os.environ.get("S3_SECRET_ACCESS_KEY", None),
    "S3_ENDPOINT_URL": lambda: os.environ.get("S3_ENDPOINT_URL", None),
657
    # Usage stats collection
658
659
660
661
662
663
664
665
666
667
668
    "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"),
669
670
671
672
    # 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
673
674
675
    "VLLM_CONFIGURE_LOGGING": lambda: bool(
        int(os.getenv("VLLM_CONFIGURE_LOGGING", "1"))
    ),
676
    "VLLM_LOGGING_CONFIG_PATH": lambda: os.getenv("VLLM_LOGGING_CONFIG_PATH"),
677
    # this is used for configuring the default logging level
678
    "VLLM_LOGGING_LEVEL": lambda: os.getenv("VLLM_LOGGING_LEVEL", "INFO").upper(),
679
    # this is used for configuring the default logging stream
680
    "VLLM_LOGGING_STREAM": lambda: os.getenv("VLLM_LOGGING_STREAM", "ext://sys.stdout"),
681
    # if set, VLLM_LOGGING_PREFIX will be prepended to all log messages
682
    "VLLM_LOGGING_PREFIX": lambda: os.getenv("VLLM_LOGGING_PREFIX", ""),
Nick Hill's avatar
Nick Hill committed
683
684
685
686
687
    # Controls colored logging output. Options: "auto" (default, colors when terminal),
    # "1" (always use colors), "0" (never use colors)
    "VLLM_LOGGING_COLOR": lambda: os.getenv("VLLM_LOGGING_COLOR", "auto"),
    # Standard unix flag for disabling ANSI color codes
    "NO_COLOR": lambda: os.getenv("NO_COLOR", "0") != "0",
688
689
    # If set, vllm will log stats at this interval in seconds
    # If not set, vllm will log stats every 10 seconds.
690
691
692
    "VLLM_LOG_STATS_INTERVAL": lambda: val
    if (val := float(os.getenv("VLLM_LOG_STATS_INTERVAL", "10."))) > 0.0
    else 10.0,
693
694
695
    # Trace function calls
    # If set to 1, vllm will trace function calls
    # Useful for debugging
696
    "VLLM_TRACE_FUNCTION": lambda: int(os.getenv("VLLM_TRACE_FUNCTION", "0")),
697
    # If set, vllm will use flashinfer sampler
698
699
700
701
702
    "VLLM_USE_FLASHINFER_SAMPLER": lambda: bool(
        int(os.environ["VLLM_USE_FLASHINFER_SAMPLER"])
    )
    if "VLLM_USE_FLASHINFER_SAMPLER" in os.environ
    else None,
703
    # Pipeline stage partition strategy
704
    "VLLM_PP_LAYER_PARTITION": lambda: os.getenv("VLLM_PP_LAYER_PARTITION", None),
705
    # (CPU backend only) CPU key-value cache space.
706
    # default is None and will be set as 4 GB
707
708
709
    "VLLM_CPU_KVCACHE_SPACE": lambda: int(os.getenv("VLLM_CPU_KVCACHE_SPACE", "0"))
    if "VLLM_CPU_KVCACHE_SPACE" in os.environ
    else None,
710
711
    # (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 '|'.
712
    "VLLM_CPU_OMP_THREADS_BIND": lambda: os.getenv("VLLM_CPU_OMP_THREADS_BIND", "auto"),
713
714
    # (CPU backend only) CPU cores not used by OMP threads .
    # Those CPU cores will not be used by OMP threads of a rank.
715
716
717
718
719
    "VLLM_CPU_NUM_OF_RESERVED_CPU": lambda: int(
        os.getenv("VLLM_CPU_NUM_OF_RESERVED_CPU", "0")
    )
    if "VLLM_CPU_NUM_OF_RESERVED_CPU" in os.environ
    else None,
720
    # (CPU backend only) whether to use SGL kernels, optimized for small batch.
721
    "VLLM_CPU_SGL_KERNEL": lambda: bool(int(os.getenv("VLLM_CPU_SGL_KERNEL", "0"))),
722
723
724
725
726
    # (Zen CPU backend) eagerly prepack weights into ZenDNN blocked layout
    # at model load time. Eliminates per-inference layout conversion overhead.
    "VLLM_ZENTORCH_WEIGHT_PREPACK": lambda: bool(
        int(os.getenv("VLLM_ZENTORCH_WEIGHT_PREPACK", "1"))
    ),
727
728
729
730
731
732
733
    # 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
734
735
736
    "VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE": env_with_choices(
        "VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE", "auto", ["auto", "nccl", "shm"]
    ),
737
    # If the env var is set, it enables GPU communication overlap
738
    # (experimental feature) in Ray's Compiled Graph.
739
740
741
    "VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM": lambda: bool(
        int(os.getenv("VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM", "0"))
    ),
742
743
744
    # If the env var is set, it uses a Ray Communicator wrapping
    # vLLM's pipeline parallelism communicator to interact with Ray's
    # Compiled Graph. Otherwise, it uses Ray's NCCL communicator.
745
746
747
    "VLLM_USE_RAY_WRAPPED_PP_COMM": lambda: bool(
        int(os.getenv("VLLM_USE_RAY_WRAPPED_PP_COMM", "1"))
    ),
748
749
    # Use dedicated multiprocess context for workers.
    # Both spawn and fork work
750
751
752
    "VLLM_WORKER_MULTIPROC_METHOD": env_with_choices(
        "VLLM_WORKER_MULTIPROC_METHOD", "fork", ["spawn", "fork"]
    ),
753
    # Path to the cache for storing downloaded assets
754
    "VLLM_ASSETS_CACHE": lambda: os.path.expanduser(
755
756
757
        os.getenv(
            "VLLM_ASSETS_CACHE",
            os.path.join(get_default_cache_root(), "vllm", "assets"),
758
759
        )
    ),
760
761
    # If the env var is set, we will clean model file in
    # this path $VLLM_ASSETS_CACHE/model_streamer/$model_name
762
763
764
    "VLLM_ASSETS_CACHE_MODEL_CLEAN": lambda: bool(
        int(os.getenv("VLLM_ASSETS_CACHE_MODEL_CLEAN", "0"))
    ),
765
766
    # Timeout for fetching images when serving multimodal models
    # Default is 5 seconds
767
    "VLLM_IMAGE_FETCH_TIMEOUT": lambda: int(os.getenv("VLLM_IMAGE_FETCH_TIMEOUT", "5")),
768
    # Timeout for fetching videos when serving multimodal models
769
    # Default is 30 seconds
770
771
772
    "VLLM_VIDEO_FETCH_TIMEOUT": lambda: int(
        os.getenv("VLLM_VIDEO_FETCH_TIMEOUT", "30")
    ),
773
    # Timeout for fetching audio when serving multimodal models
774
    # Default is 10 seconds
775
776
777
    "VLLM_AUDIO_FETCH_TIMEOUT": lambda: int(
        os.getenv("VLLM_AUDIO_FETCH_TIMEOUT", "10")
    ),
778
779
780
781
782
    # Maximum number of retries for fetching media (images, audio, video)
    # from URLs. Each retry quadruples the timeout. Default is 3.
    "VLLM_MEDIA_FETCH_MAX_RETRIES": lambda: int(
        os.getenv("VLLM_MEDIA_FETCH_MAX_RETRIES", "3")
    ),
783
784
    # Whether to allow HTTP redirects when fetching from media URLs.
    # Default to True
785
786
787
    "VLLM_MEDIA_URL_ALLOW_REDIRECTS": lambda: bool(
        int(os.getenv("VLLM_MEDIA_URL_ALLOW_REDIRECTS", "1"))
    ),
788
789
790
    # Max number of workers for the thread pool handling
    # media bytes loading. Set to 1 to disable parallel processing.
    # Default is 8
791
792
793
    "VLLM_MEDIA_LOADING_THREAD_COUNT": lambda: int(
        os.getenv("VLLM_MEDIA_LOADING_THREAD_COUNT", "8")
    ),
794
795
796
    # Maximum filesize in MB for a single audio file when processing
    # speech-to-text requests. Files larger than this will be rejected.
    # Default is 25 MB
797
798
799
    "VLLM_MAX_AUDIO_CLIP_FILESIZE_MB": lambda: int(
        os.getenv("VLLM_MAX_AUDIO_CLIP_FILESIZE_MB", "25")
    ),
800
801
    # Backend for Video IO
    # - "opencv": Default backend that uses OpenCV stream buffered backend.
Roger Wang's avatar
Roger Wang committed
802
    # - "identity": Returns raw video bytes for model processor to handle.
803
804
805
806
807
    #
    # 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.
808
809
810
    "VLLM_VIDEO_LOADER_BACKEND": lambda: os.getenv(
        "VLLM_VIDEO_LOADER_BACKEND", "opencv"
    ),
811
812
813
814
815
816
817
818
    # Media connector implementation.
    # - "http": Default connector that supports fetching media via HTTP.
    #
    # Custom implementations can be registered
    # via `@MEDIA_CONNECTOR_REGISTRY.register("my_custom_media_connector")` and
    # imported at runtime.
    # If a non-existing backend is used, an AssertionError will be thrown.
    "VLLM_MEDIA_CONNECTOR": lambda: os.getenv("VLLM_MEDIA_CONNECTOR", "http"),
819
820
821
822
823
824
825
826
827
828
829
    # Hash algorithm for multimodal content hashing.
    # - "blake3": Default, fast cryptographic hash (not FIPS 140-3 compliant)
    # - "sha256": FIPS 140-3 compliant, widely supported
    # - "sha512": FIPS 140-3 compliant, faster on 64-bit systems
    # Use sha256 or sha512 for FIPS compliance in government/enterprise deployments
    "VLLM_MM_HASHER_ALGORITHM": env_with_choices(
        "VLLM_MM_HASHER_ALGORITHM",
        "blake3",
        ["blake3", "sha256", "sha512"],
        case_sensitive=False,
    ),
830
831
    # Path to the XLA persistent cache directory.
    # Only used for XLA devices such as TPUs.
832
    "VLLM_XLA_CACHE_PATH": lambda: os.path.expanduser(
833
        os.getenv(
834
            "VLLM_XLA_CACHE_PATH",
835
            os.path.join(get_default_cache_root(), "vllm", "xla_cache"),
836
837
        )
    ),
838
    # If set, assert on XLA recompilation after each execution step.
839
840
841
    "VLLM_XLA_CHECK_RECOMPILATION": lambda: bool(
        int(os.getenv("VLLM_XLA_CHECK_RECOMPILATION", "0"))
    ),
842
    # Enable SPMD mode for TPU backend.
843
    "VLLM_XLA_USE_SPMD": lambda: bool(int(os.getenv("VLLM_XLA_USE_SPMD", "0"))),
844
845
    # If set, the OpenAI API server will stay alive even after the underlying
    # AsyncLLMEngine errors and stops serving requests
846
    "VLLM_KEEP_ALIVE_ON_ENGINE_DEATH": lambda: bool(
847
        int(os.getenv("VLLM_KEEP_ALIVE_ON_ENGINE_DEATH", "0"))
848
    ),
849
850
851
852
    # 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.
853
854
855
856
    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": lambda: (
        os.environ.get("VLLM_ALLOW_LONG_MAX_MODEL_LEN", "0").strip().lower()
        in ("1", "true")
    ),
857
858
    # If set, forces FP8 Marlin to be used for FP8 quantization regardless
    # of the hardware support for FP8 compute.
859
860
861
862
863
864
865
    "VLLM_TEST_FORCE_FP8_MARLIN": lambda: (
        os.environ.get("VLLM_TEST_FORCE_FP8_MARLIN", "0").strip().lower()
        in ("1", "true")
    ),
    "VLLM_TEST_FORCE_LOAD_FORMAT": lambda: os.getenv(
        "VLLM_TEST_FORCE_LOAD_FORMAT", "dummy"
    ),
866
867
    # Time in ms for the zmq client to wait for a response from the backend
    # server for simple data operations
868
    "VLLM_RPC_TIMEOUT": lambda: int(os.getenv("VLLM_RPC_TIMEOUT", "10000")),
869
    # Timeout in seconds for keeping HTTP connections alive in API server
870
871
872
    "VLLM_HTTP_TIMEOUT_KEEP_ALIVE": lambda: int(
        os.environ.get("VLLM_HTTP_TIMEOUT_KEEP_ALIVE", "5")
    ),
873
874
875
    # 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
876
877
878
    "VLLM_PLUGINS": lambda: None
    if "VLLM_PLUGINS" not in os.environ
    else os.environ["VLLM_PLUGINS"].split(","),
879
880
881
    # a local directory to look in for unrecognized LoRA adapters.
    # only works if plugins are enabled and
    # VLLM_ALLOW_RUNTIME_LORA_UPDATING is enabled.
882
883
884
    "VLLM_LORA_RESOLVER_CACHE_DIR": lambda: os.getenv(
        "VLLM_LORA_RESOLVER_CACHE_DIR", None
    ),
885
886
887
888
889
890
891
    # A remote HF repo(s) containing one or more LoRA adapters, which
    # may be downloaded and leveraged as needed. Only works if plugins
    # are enabled and VLLM_ALLOW_RUNTIME_LORA_UPDATING is enabled.
    # Values should be comma separated.
    "VLLM_LORA_RESOLVER_HF_REPO_LIST": lambda: os.getenv(
        "VLLM_LORA_RESOLVER_HF_REPO_LIST", None
    ),
892
    # If set, vLLM will use Triton implementations of AWQ.
893
    "VLLM_USE_TRITON_AWQ": lambda: bool(int(os.getenv("VLLM_USE_TRITON_AWQ", "0"))),
894
    # If set, allow loading or unloading lora adapters in runtime,
895
896
897
898
    "VLLM_ALLOW_RUNTIME_LORA_UPDATING": lambda: (
        os.environ.get("VLLM_ALLOW_RUNTIME_LORA_UPDATING", "0").strip().lower()
        in ("1", "true")
    ),
899
900
901
902
903
904
    # We assume drivers can report p2p status correctly.
    # If the program hangs when using custom allreduce,
    # potantially caused by a bug in the driver (535 series),
    # if might be helpful to set VLLM_SKIP_P2P_CHECK=0
    # so that vLLM can verify if p2p is actually working.
    # See https://github.com/vllm-project/vllm/blob/a9b15c606fea67a072416ea0ea115261a2756058/vllm/distributed/device_communicators/custom_all_reduce_utils.py#L101-L108 for details. # noqa
905
    "VLLM_SKIP_P2P_CHECK": lambda: os.getenv("VLLM_SKIP_P2P_CHECK", "1") == "1",
906
907
908
909
    # List of quantization kernels that should be disabled, used for testing
    # and performance comparisons. Currently only affects MPLinearKernel
    # selection
    # (kernels: MacheteLinearKernel, MarlinLinearKernel, ExllamaLinearKernel)
910
911
912
    "VLLM_DISABLED_KERNELS": lambda: []
    if "VLLM_DISABLED_KERNELS" not in os.environ
    else os.environ["VLLM_DISABLED_KERNELS"].split(","),
913
914
915
    "VLLM_ENABLE_FLA_PACKED_RECURRENT_DECODE": lambda: bool(
        int(os.getenv("VLLM_ENABLE_FLA_PACKED_RECURRENT_DECODE", "1"))
    ),
916
    # Disable pynccl (using torch.distributed instead)
917
918
919
    "VLLM_DISABLE_PYNCCL": lambda: (
        os.getenv("VLLM_DISABLE_PYNCCL", "False").lower() in ("true", "1")
    ),
920
921
922
923
924
    # Optional: enable external Oink custom ops (e.g., Blackwell RMSNorm).
    # Disabled by default.
    "VLLM_USE_OINK_OPS": lambda: (
        os.getenv("VLLM_USE_OINK_OPS", "False").lower() in ("true", "1")
    ),
925
926
    # Disable aiter ops unless specifically enabled.
    # Acts as a parent switch to enable the rest of the other operations.
927
928
929
    "VLLM_ROCM_USE_AITER": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER", "False").lower() in ("true", "1")
    ),
930
931
    # Whether to use aiter paged attention.
    # By default is disabled.
932
933
934
    "VLLM_ROCM_USE_AITER_PAGED_ATTN": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_PAGED_ATTN", "False").lower() in ("true", "1")
    ),
935
936
937
    # use aiter linear op if aiter ops are enabled
    # The following list of related ops
    # - scaled_mm (per-tensor / rowwise)
938
939
940
    "VLLM_ROCM_USE_AITER_LINEAR": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_LINEAR", "True").lower() in ("true", "1")
    ),
941
942
    # Whether to use aiter moe ops.
    # By default is enabled.
943
944
945
    "VLLM_ROCM_USE_AITER_MOE": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_MOE", "True").lower() in ("true", "1")
    ),
946
    # use aiter rms norm op if aiter ops are enabled.
947
948
949
    "VLLM_ROCM_USE_AITER_RMSNORM": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_RMSNORM", "True").lower() in ("true", "1")
    ),
950
951
    # Whether to use aiter mla ops.
    # By default is enabled.
952
953
954
    "VLLM_ROCM_USE_AITER_MLA": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_MLA", "True").lower() in ("true", "1")
    ),
955
956
    # Whether to use aiter mha ops.
    # By default is enabled.
957
958
959
    "VLLM_ROCM_USE_AITER_MHA": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_MHA", "True").lower() in ("true", "1")
    ),
960
961
    # Whether to use aiter fp4 gemm asm.
    # By default is disabled.
962
963
964
    "VLLM_ROCM_USE_AITER_FP4_ASM_GEMM": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_FP4_ASM_GEMM", "False").lower() in ("true", "1")
    ),
965
    # Whether to use aiter rope.
966
    # By default is disabled.
967
    "VLLM_ROCM_USE_AITER_TRITON_ROPE": lambda: (
968
        os.getenv("VLLM_ROCM_USE_AITER_TRITON_ROPE", "False").lower() in ("true", "1")
969
    ),
970
971
    # Whether to use aiter triton fp8 bmm kernel
    # By default is enabled.
972
973
974
    "VLLM_ROCM_USE_AITER_FP8BMM": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_FP8BMM", "True").lower() in ("true", "1")
    ),
975
976
977
978
979
    # Whether to use aiter triton fp4 bmm kernel
    # By default is enabled.
    "VLLM_ROCM_USE_AITER_FP4BMM": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_FP4BMM", "True").lower() in ("true", "1")
    ),
980
981
982
983
984
    # Use AITER triton unified attention for V1 attention
    "VLLM_ROCM_USE_AITER_UNIFIED_ATTENTION": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_UNIFIED_ATTENTION", "False").lower()
        in ("true", "1")
    ),
985
    # Whether to use aiter fusion shared experts ops.
986
    # By default is disabled.
987
    "VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS": lambda: (
988
        os.getenv("VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS", "False").lower()
989
990
        in ("true", "1")
    ),
991
992
993
994
995
    # Whether to use aiter triton kernels for gemm ops.
    # By default is enabled.
    "VLLM_ROCM_USE_AITER_TRITON_GEMM": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_TRITON_GEMM", "True").lower() in ("true", "1")
    ),
996
    # use rocm skinny gemms
997
998
999
    "VLLM_ROCM_USE_SKINNY_GEMM": lambda: (
        os.getenv("VLLM_ROCM_USE_SKINNY_GEMM", "True").lower() in ("true", "1")
    ),
1000
    # Pad the fp8 weights to 256 bytes for ROCm
1001
    "VLLM_ROCM_FP8_PADDING": lambda: bool(int(os.getenv("VLLM_ROCM_FP8_PADDING", "1"))),
1002
    # Pad the weights for the moe kernel
1003
    "VLLM_ROCM_MOE_PADDING": lambda: bool(int(os.getenv("VLLM_ROCM_MOE_PADDING", "1"))),
1004
1005
1006
1007
    # Whether to use the shuffled kv cache layout
    "VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT": lambda: (
        os.getenv("VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT", "False").lower() in ("true", "1")
    ),
1008
1009
1010
    # Custom quick allreduce kernel for MI3* cards
    # Choice of quantization level: FP, INT8, INT6, INT4 or NONE
    # Recommended for large models to get allreduce
1011
1012
1013
1014
1015
    "VLLM_ROCM_QUICK_REDUCE_QUANTIZATION": env_with_choices(
        "VLLM_ROCM_QUICK_REDUCE_QUANTIZATION",
        "NONE",
        ["FP", "INT8", "INT6", "INT4", "NONE"],
    ),
1016
1017
1018
1019
    # 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
1020
1021
1022
1023
    "VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16": lambda: (
        os.getenv("VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16", "True").lower()
        in ("true", "1")
    ),
1024
1025
1026
1027
1028
1029
    # 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.
1030
1031
1032
    "VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB": lambda: maybe_convert_int(
        os.environ.get("VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB", None)
    ),
1033
    # Divisor for dynamic query scale factor calculation for FP8 KV Cache
1034
    "Q_SCALE_CONSTANT": lambda: int(os.getenv("Q_SCALE_CONSTANT", "200")),
1035
    # Divisor for dynamic key scale factor calculation for FP8 KV Cache
1036
    "K_SCALE_CONSTANT": lambda: int(os.getenv("K_SCALE_CONSTANT", "200")),
1037
    # Divisor for dynamic value scale factor calculation for FP8 KV Cache
1038
    "V_SCALE_CONSTANT": lambda: int(os.getenv("V_SCALE_CONSTANT", "100")),
1039
    # If set, enable multiprocessing in LLM for the V1 code path.
1040
1041
1042
1043
1044
1045
    "VLLM_ENABLE_V1_MULTIPROCESSING": lambda: bool(
        int(os.getenv("VLLM_ENABLE_V1_MULTIPROCESSING", "1"))
    ),
    "VLLM_LOG_BATCHSIZE_INTERVAL": lambda: float(
        os.getenv("VLLM_LOG_BATCHSIZE_INTERVAL", "-1")
    ),
1046
    "VLLM_DISABLE_COMPILE_CACHE": disable_compile_cache,
1047
1048
1049
    # If set, vllm will run in development mode, which will enable
    # some additional endpoints for developing and debugging,
    # e.g. `/reset_prefix_cache`
1050
    "VLLM_SERVER_DEV_MODE": lambda: bool(int(os.getenv("VLLM_SERVER_DEV_MODE", "0"))),
1051
1052
1053
1054
1055
1056
1057
    # 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.
1058
1059
1060
    "VLLM_V1_OUTPUT_PROC_CHUNK_SIZE": lambda: int(
        os.getenv("VLLM_V1_OUTPUT_PROC_CHUNK_SIZE", "128")
    ),
1061
    # If set, vLLM will disable the MLA attention optimizations.
1062
    "VLLM_MLA_DISABLE": lambda: bool(int(os.getenv("VLLM_MLA_DISABLE", "0"))),
1063
    # If set, vLLM will pick up the provided Flash Attention MLA
1064
1065
1066
    # 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.
1067
1068
1069
    "VLLM_RAY_PER_WORKER_GPUS": lambda: float(
        os.getenv("VLLM_RAY_PER_WORKER_GPUS", "1.0")
    ),
1070
1071
1072
    # 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"
1073
    "VLLM_RAY_BUNDLE_INDICES": lambda: os.getenv("VLLM_RAY_BUNDLE_INDICES", ""),
1074
1075
    # In some system, find_loaded_library() may not work. So we allow users to
    # specify the path through environment variable VLLM_CUDART_SO_PATH.
1076
    "VLLM_CUDART_SO_PATH": lambda: os.getenv("VLLM_CUDART_SO_PATH", None),
1077
    # Rank of the process in the data parallel setting
1078
    "VLLM_DP_RANK": lambda: int(os.getenv("VLLM_DP_RANK", "0")),
1079
1080
    # Rank of the process in the data parallel setting.
    # Defaults to VLLM_DP_RANK when not set.
1081
1082
1083
    "VLLM_DP_RANK_LOCAL": lambda: int(
        os.getenv("VLLM_DP_RANK_LOCAL", sys.modules[__name__].VLLM_DP_RANK)
    ),
1084
    # World size of the data parallel setting
1085
    "VLLM_DP_SIZE": lambda: int(os.getenv("VLLM_DP_SIZE", "1")),
1086
    # IP address of the master node in the data parallel setting
1087
    "VLLM_DP_MASTER_IP": lambda: os.getenv("VLLM_DP_MASTER_IP", "127.0.0.1"),
1088
    # Port of the master node in the data parallel setting
1089
    "VLLM_DP_MASTER_PORT": lambda: int(os.getenv("VLLM_DP_MASTER_PORT", "0")),
1090
1091
1092
1093
1094
    # 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.
1095
    "VLLM_MOE_DP_CHUNK_SIZE": lambda: int(os.getenv("VLLM_MOE_DP_CHUNK_SIZE", "256")),
1096
1097
1098
    "VLLM_ENABLE_MOE_DP_CHUNK": lambda: bool(
        int(os.getenv("VLLM_ENABLE_MOE_DP_CHUNK", "1"))
    ),
1099
    # Randomize inputs during dummy runs when using Data Parallel
1100
1101
1102
1103
    "VLLM_RANDOMIZE_DP_DUMMY_INPUTS": lambda: os.environ.get(
        "VLLM_RANDOMIZE_DP_DUMMY_INPUTS", "0"
    )
    == "1",
1104
1105
1106
1107
1108
1109
1110
    # Strategy to pack the data parallel ranks for Ray.
    # Available options:
    # - "fill":
    #   for DP master node, allocate exactly data-parallel-size-local DP ranks,
    #   for non-master nodes, allocate as many DP ranks as can fit;
    # - "strict":
    #   allocate exactly data-parallel-size-local DP ranks to each picked node;
1111
1112
1113
    # - "span":
    #   Should be used only when a single DP rank requires multiple nodes.
    #   allocate one DP rank over as many nodes as required for set world_size;
1114
1115
1116
1117
    # This environment variable is ignored if data-parallel-backend is not Ray.
    "VLLM_RAY_DP_PACK_STRATEGY": lambda: os.getenv(
        "VLLM_RAY_DP_PACK_STRATEGY", "strict"
    ),
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
    # Comma-separated *additional* prefixes of env vars to copy from the
    # driver to Ray workers.  These are merged with the built-in defaults
    # defined in ``vllm.ray.ray_env`` (VLLM_, etc.).  Example: "MYLIB_,OTHER_"
    "VLLM_RAY_EXTRA_ENV_VAR_PREFIXES_TO_COPY": lambda: os.getenv(
        "VLLM_RAY_EXTRA_ENV_VAR_PREFIXES_TO_COPY", ""
    ),
    # Comma-separated *additional* individual env var names to copy from
    # the driver to Ray workers.  Merged with the built-in defaults
    # defined in ``vllm.ray.ray_env`` (PYTHONHASHSEED).
    # Example: "MY_SECRET,MY_FLAG"
    "VLLM_RAY_EXTRA_ENV_VARS_TO_COPY": lambda: os.getenv(
        "VLLM_RAY_EXTRA_ENV_VARS_TO_COPY", ""
    ),
1131
    # Whether to use S3 path for model loading in CI via RunAI Streamer
1132
    "VLLM_CI_USE_S3": lambda: os.environ.get("VLLM_CI_USE_S3", "0") == "1",
1133
    # Use model_redirect to redirect the model name to a local folder.
1134
1135
1136
1137
1138
    # `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
1139
1140
1141
    "VLLM_MODEL_REDIRECT_PATH": lambda: os.environ.get(
        "VLLM_MODEL_REDIRECT_PATH", None
    ),
1142
    # Whether to use atomicAdd reduce in gptq/awq marlin kernel.
1143
1144
1145
1146
    "VLLM_MARLIN_USE_ATOMIC_ADD": lambda: os.environ.get(
        "VLLM_MARLIN_USE_ATOMIC_ADD", "0"
    )
    == "1",
1147
    # Whether to use marlin kernel in mxfp4 quantization method
1148
1149
1150
    "VLLM_MXFP4_USE_MARLIN": lambda: maybe_convert_bool(
        os.environ.get("VLLM_MXFP4_USE_MARLIN", None)
    ),
1151
1152
1153
1154
    # The activation dtype for marlin kernel
    "VLLM_MARLIN_INPUT_DTYPE": env_with_choices(
        "VLLM_MARLIN_INPUT_DTYPE", None, ["int8", "fp8"]
    ),
1155
1156
1157
1158
1159
1160
    # Whether to use DeepEPLL kernels for NVFP4 quantization and dispatch method
    # only supported on Blackwell GPUs and with
    # https://github.com/deepseek-ai/DeepEP/pull/341
    "VLLM_DEEPEPLL_NVFP4_DISPATCH": lambda: bool(
        int(os.getenv("VLLM_DEEPEPLL_NVFP4_DISPATCH", "0"))
    ),
1161
1162
1163
    # Whether to turn on the outlines cache for V1
    # This cache is unbounded and on disk, so it's not safe to use in
    # an environment with potentially malicious users.
1164
1165
1166
1167
    "VLLM_V1_USE_OUTLINES_CACHE": lambda: os.environ.get(
        "VLLM_V1_USE_OUTLINES_CACHE", "0"
    )
    == "1",
1168
1169
    # Gap between padding buckets for the forward pass. So we have
    # 8, we will run forward pass with [16, 24, 32, ...].
1170
1171
1172
1173
1174
1175
1176
1177
    "VLLM_TPU_BUCKET_PADDING_GAP": lambda: int(
        os.environ["VLLM_TPU_BUCKET_PADDING_GAP"]
    )
    if "VLLM_TPU_BUCKET_PADDING_GAP" in os.environ
    else 0,
    "VLLM_TPU_MOST_MODEL_LEN": lambda: maybe_convert_int(
        os.environ.get("VLLM_TPU_MOST_MODEL_LEN", None)
    ),
1178
    # Whether using Pathways
1179
1180
1181
    "VLLM_TPU_USING_PATHWAYS": lambda: bool(
        "proxy" in os.getenv("JAX_PLATFORMS", "").lower()
    ),
1182
    # Allow use of DeepGemm kernels for fused moe ops.
1183
    "VLLM_USE_DEEP_GEMM": lambda: bool(int(os.getenv("VLLM_USE_DEEP_GEMM", "1"))),
1184
1185
1186
1187
    # Allow use of DeepGemm specifically for MoE fused ops (overrides only MoE).
    "VLLM_MOE_USE_DEEP_GEMM": lambda: bool(
        int(os.getenv("VLLM_MOE_USE_DEEP_GEMM", "1"))
    ),
1188
    # Whether to use E8M0 scaling when DeepGEMM is used on Blackwell GPUs.
1189
1190
1191
    "VLLM_USE_DEEP_GEMM_E8M0": lambda: bool(
        int(os.getenv("VLLM_USE_DEEP_GEMM_E8M0", "1"))
    ),
1192
1193
1194
1195
    # Whether to create TMA-aligned scale tensor when DeepGEMM is used.
    "VLLM_USE_DEEP_GEMM_TMA_ALIGNED_SCALES": lambda: bool(
        int(os.getenv("VLLM_USE_DEEP_GEMM_TMA_ALIGNED_SCALES", "1"))
    ),
1196
1197
1198
1199
    # DeepGemm JITs the kernels on-demand. The warmup attempts to make DeepGemm
    # JIT all the required kernels before model execution so there is no
    # JIT'ing in the hot-path. However, this warmup increases the engine
    # startup time by a couple of minutes.
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
    # Available options:
    #  - "skip"  : Skip warmup.
    #  - "full"  : Warmup deepgemm by running all possible gemm shapes the
    #   engine could encounter.
    #  - "relax" : Select gemm shapes to run based on some heuristics. The
    #   heuristic aims to have the same effect as running all possible gemm
    #   shapes, but provides no guarantees.
    "VLLM_DEEP_GEMM_WARMUP": env_with_choices(
        "VLLM_DEEP_GEMM_WARMUP",
        "relax",
        [
            "skip",
            "full",
            "relax",
        ],
1215
    ),
1216
    # Whether to use fused grouped_topk used for MoE expert selection.
1217
1218
1219
    "VLLM_USE_FUSED_MOE_GROUPED_TOPK": lambda: bool(
        int(os.getenv("VLLM_USE_FUSED_MOE_GROUPED_TOPK", "1"))
    ),
1220
1221
1222
    # Allow use of FlashInfer FP8 block-scale GEMM for linear layers.
    # This uses TensorRT-LLM kernels and requires SM90+ (Hopper).
    "VLLM_BLOCKSCALE_FP8_GEMM_FLASHINFER": lambda: bool(
1223
        int(os.getenv("VLLM_BLOCKSCALE_FP8_GEMM_FLASHINFER", "1"))
1224
    ),
1225
    # Allow use of FlashInfer BF16 MoE kernels for fused moe ops.
1226
1227
1228
    "VLLM_USE_FLASHINFER_MOE_FP16": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_FP16", "0"))
    ),
1229
    # Allow use of FlashInfer FP8 MoE kernels for fused moe ops.
1230
1231
1232
    "VLLM_USE_FLASHINFER_MOE_FP8": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_FP8", "0"))
    ),
1233
    # Allow use of FlashInfer NVFP4 MoE kernels for fused moe ops.
1234
1235
1236
    "VLLM_USE_FLASHINFER_MOE_FP4": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_FP4", "0"))
    ),
1237
1238
1239
1240
    # Allow use of FlashInfer MxInt4 MoE kernels for fused moe ops.
    "VLLM_USE_FLASHINFER_MOE_INT4": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_INT4", "0"))
    ),
1241
1242
    # If set to 1, use the FlashInfer
    # MXFP8 (activation) x MXFP4 (weight) MoE backend.
1243
1244
1245
    "VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8", "0"))
    ),
1246
1247
1248
1249
    # If set to 1, use the FlashInfer CUTLASS backend for
    # MXFP8 (activation) x MXFP4 (weight) MoE.
    # This is separate from the TRTLLMGEN path controlled by
    # VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8.
1250
1251
1252
    "VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS", "0"))
    ),
1253
1254
    # If set to 1, use the FlashInfer
    # BF16 (activation) x MXFP4 (weight) MoE backend.
1255
1256
1257
    "VLLM_USE_FLASHINFER_MOE_MXFP4_BF16": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_MXFP4_BF16", "0"))
    ),
1258
1259
1260
    # 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.
1261
    "VLLM_XGRAMMAR_CACHE_MB": lambda: int(os.getenv("VLLM_XGRAMMAR_CACHE_MB", "512")),
1262
1263
1264
1265
1266
1267
1268
    # 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.
1269
1270
1271
    "VLLM_MSGPACK_ZERO_COPY_THRESHOLD": lambda: int(
        os.getenv("VLLM_MSGPACK_ZERO_COPY_THRESHOLD", "256")
    ),
1272
1273
1274
    # 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.
1275
1276
1277
    "VLLM_ALLOW_INSECURE_SERIALIZATION": lambda: bool(
        int(os.getenv("VLLM_ALLOW_INSECURE_SERIALIZATION", "0"))
    ),
1278
1279
1280
1281
1282
    # Temporary: skip adding random suffix to internal request IDs. May be
    # needed for KV connectors that match request IDs across instances.
    "VLLM_DISABLE_REQUEST_ID_RANDOMIZATION": lambda: bool(
        int(os.getenv("VLLM_DISABLE_REQUEST_ID_RANDOMIZATION", "0"))
    ),
Robert Shaw's avatar
Robert Shaw committed
1283
    # IP address used for NIXL handshake between remote agents.
1284
1285
1286
    "VLLM_NIXL_SIDE_CHANNEL_HOST": lambda: os.getenv(
        "VLLM_NIXL_SIDE_CHANNEL_HOST", "localhost"
    ),
Robert Shaw's avatar
Robert Shaw committed
1287
    # Port used for NIXL handshake between remote agents.
1288
1289
1290
    "VLLM_NIXL_SIDE_CHANNEL_PORT": lambda: int(
        os.getenv("VLLM_NIXL_SIDE_CHANNEL_PORT", "5600")
    ),
1291
1292
1293
1294
    # Port used for Mooncake handshake between remote agents.
    "VLLM_MOONCAKE_BOOTSTRAP_PORT": lambda: int(
        os.getenv("VLLM_MOONCAKE_BOOTSTRAP_PORT", "8998")
    ),
1295
1296
    # Flashinfer MoE backend for vLLM's fused Mixture-of-Experts support.
    # Both require compute capability 10.0 or above.
1297
1298
1299
1300
1301
    # Available options:
    # - "throughput":  [default]
    #     Uses CUTLASS kernels optimized for high-throughput batch inference.
    # - "latency":
    #     Uses TensorRT-LLM kernels optimized for low-latency inference.
1302
    "VLLM_FLASHINFER_MOE_BACKEND": env_with_choices(
1303
1304
1305
        "VLLM_FLASHINFER_MOE_BACKEND",
        "latency",
        ["throughput", "latency", "masked_gemm"],
1306
    ),
1307
1308
1309
    # Flashinfer fused allreduce backend.
    # "auto" will default to "mnnvl", which performs mostly same/better than "trtllm".
    # But "mnnvl" backend does not support fuse with quantization.
1310
1311
1312
    # TODO: Default is "trtllm" right now because "mnnvl" has issues with cudagraph:
    # https://github.com/vllm-project/vllm/issues/35772
    # Should switch back to "auto" if the issue is resolved.
1313
1314
    "VLLM_FLASHINFER_ALLREDUCE_BACKEND": env_with_choices(
        "VLLM_FLASHINFER_ALLREDUCE_BACKEND",
1315
        "trtllm",
1316
1317
        ["auto", "trtllm", "mnnvl"],
    ),
1318
1319
1320
1321
    # Control the workspace buffer size for the FlashInfer backend.
    "VLLM_FLASHINFER_WORKSPACE_BUFFER_SIZE": lambda: int(
        os.getenv("VLLM_FLASHINFER_WORKSPACE_BUFFER_SIZE", str(394 * 1024 * 1024))
    ),
1322
1323
1324
1325
    # 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.
1326
1327
1328
    "VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE": lambda: int(
        os.getenv("VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE", "163840")
    ),
1329
1330
1331
1332
    # Specifies the thresholds of the communicated tensor sizes under which
    # vllm should use flashinfer fused allreduce. The variable should be a
    # JSON with the following format:
    #     { <world size>: <max size in mb> }
1333
    # Unspecified world sizes will fall back to
1334
    #     { 2: 64, 4: 1, <everything else>: 0.5 }
1335
1336
1337
    "VLLM_FLASHINFER_ALLREDUCE_FUSION_THRESHOLDS_MB": lambda: json.loads(
        os.getenv("VLLM_FLASHINFER_ALLREDUCE_FUSION_THRESHOLDS_MB", "{}")
    ),
1338
1339
1340
    # MoE routing strategy selector.
    # See `RoutingSimulator.get_available_strategies()` # for available
    # strategies.
1341
    # Custom routing strategies can be registered by
1342
1343
    # RoutingSimulator.register_strategy()
    # Note: custom strategies may not produce correct model outputs
1344
1345
1346
    "VLLM_MOE_ROUTING_SIMULATION_STRATEGY": lambda: os.environ.get(
        "VLLM_MOE_ROUTING_SIMULATION_STRATEGY", ""
    ).lower(),
1347
    # Regex timeout for use by the vLLM tool parsing plugins.
1348
1349
1350
    "VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS": lambda: int(
        os.getenv("VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS", "1")
    ),
1351
1352
1353
    # 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.
1354
1355
1356
    "VLLM_MQ_MAX_CHUNK_BYTES_MB": lambda: int(
        os.getenv("VLLM_MQ_MAX_CHUNK_BYTES_MB", "16")
    ),
1357
1358
    # Timeout in seconds for execute_model RPC calls in multiprocessing
    # executor (only applies when TP > 1).
1359
1360
1361
    "VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS": lambda: int(
        os.getenv("VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS", "300")
    ),
1362
1363
1364
1365
1366
1367
1368
    # 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.
1369
1370
1371
    "VLLM_KV_CACHE_LAYOUT": env_with_choices(
        "VLLM_KV_CACHE_LAYOUT", None, ["NHD", "HND"]
    ),
1372
1373
1374
    # 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.
1375
1376
1377
    "VLLM_COMPUTE_NANS_IN_LOGITS": lambda: bool(
        int(os.getenv("VLLM_COMPUTE_NANS_IN_LOGITS", "0"))
    ),
1378
1379
1380
    # Controls whether or not emulations are used for NVFP4
    # generations on machines < 100 for compressed-tensors
    # models
1381
1382
1383
    "VLLM_USE_NVFP4_CT_EMULATIONS": lambda: bool(
        int(os.getenv("VLLM_USE_NVFP4_CT_EMULATIONS", "0"))
    ),
1384
1385
1386
1387
    # Time (in seconds) after which the KV cache on the producer side is
    # automatically cleared if no READ notification is received from the
    # consumer. This is only applicable when using NixlConnector in a
    # disaggregated decode-prefill setup.
1388
1389
1390
    "VLLM_NIXL_ABORT_REQUEST_TIMEOUT": lambda: int(
        os.getenv("VLLM_NIXL_ABORT_REQUEST_TIMEOUT", "480")
    ),
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
    # Controls the read mode for the Mori-IO connector
    "VLLM_MORIIO_CONNECTOR_READ_MODE": lambda: (
        os.getenv("VLLM_MORIIO_CONNECTOR_READ_MODE", "False").lower() in ("true", "1")
    ),
    # Controls the QP (Queue Pair) per transfer configuration for the Mori-IO connector
    "VLLM_MORIIO_QP_PER_TRANSFER": lambda: int(
        os.getenv("VLLM_MORIIO_QP_PER_TRANSFER", "1")
    ),
    # Controls the post-processing batch size for the Mori-IO connector
    "VLLM_MORIIO_POST_BATCH_SIZE": lambda: int(
        os.getenv("VLLM_MORIIO_POST_BATCH_SIZE", "-1")
    ),
    # Controls the number of workers for Mori operations for the Mori-IO connector
    "VLLM_MORIIO_NUM_WORKERS": lambda: int(os.getenv("VLLM_MORIIO_NUM_WORKERS", "1")),
1405
1406
1407
1408
    # Timeout (in seconds) for MooncakeConnector in PD disaggregated setup.
    "VLLM_MOONCAKE_ABORT_REQUEST_TIMEOUT": lambda: int(
        os.getenv("VLLM_MOONCAKE_ABORT_REQUEST_TIMEOUT", "480")
    ),
1409
1410
    # If set, it means we pre-downloaded cubin files and flashinfer will
    # read the cubin files directly.
1411
1412
1413
    "VLLM_HAS_FLASHINFER_CUBIN": lambda: bool(
        int(os.getenv("VLLM_HAS_FLASHINFER_CUBIN", "0"))
    ),
1414
1415
1416
1417
    # Supported options:
    # - "flashinfer-cudnn": use flashinfer cudnn GEMM backend
    # - "flashinfer-trtllm": use flashinfer trtllm GEMM backend
    # - "flashinfer-cutlass": use flashinfer cutlass GEMM backend
1418
    # - "marlin": use marlin GEMM backend (for GPUs without native FP4 support)
1419
1420
1421
1422
    # - <none>: automatically pick an available backend
    "VLLM_NVFP4_GEMM_BACKEND": env_with_choices(
        "VLLM_NVFP4_GEMM_BACKEND",
        None,
1423
1424
1425
1426
1427
        [
            "flashinfer-cudnn",
            "flashinfer-trtllm",
            "flashinfer-cutlass",
            "cutlass",
1428
            "marlin",
1429
        ],
1430
    ),
1431
1432
1433
    # Controls garbage collection during CUDA graph capture.
    # If set to 0 (default), enables GC freezing to speed up capture time.
    # If set to 1, allows GC to run during capture.
1434
1435
1436
    "VLLM_ENABLE_CUDAGRAPH_GC": lambda: bool(
        int(os.getenv("VLLM_ENABLE_CUDAGRAPH_GC", "0"))
    ),
1437
    # Used to force set up loopback IP
1438
    "VLLM_LOOPBACK_IP": lambda: os.getenv("VLLM_LOOPBACK_IP", ""),
1439
1440
1441
    # Used to set the process name prefix for vLLM processes.
    # This is useful for debugging and monitoring purposes.
    # The default value is "VLLM".
1442
    "VLLM_PROCESS_NAME_PREFIX": lambda: os.getenv("VLLM_PROCESS_NAME_PREFIX", "VLLM"),
1443
1444
1445
1446
1447
1448
1449
    # Allow chunked local attention with hybrid kv cache manager.
    # Currently using the Hybrid KV cache manager with chunked local attention
    # in the Llama4 models (the only models currently using chunked local attn)
    # causes a latency regression. For this reason, we disable it by default.
    # This flag is used to allow users to enable it if they want to (to save on
    # kv-cache memory usage and enable longer contexts)
    # TODO(lucas): Remove this flag once latency regression is resolved.
1450
    "VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE": lambda: bool(
1451
        int(os.getenv("VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE", "1"))
1452
    ),
1453
1454
    # Enables support for the "store" option in the OpenAI Responses API.
    # When set to 1, vLLM's OpenAI server will retain the input and output
1455
1456
    # messages for those requests in memory. By default, this is disabled (0),
    # and the "store" option is ignored.
1457
1458
1459
1460
1461
    # NOTE/WARNING:
    # 1. Messages are kept in memory only (not persisted to disk) and will be
    #    lost when the vLLM server shuts down.
    # 2. Enabling this option will cause a memory leak, as stored messages are
    #    never removed from memory until the server terminates.
1462
1463
1464
    "VLLM_ENABLE_RESPONSES_API_STORE": lambda: bool(
        int(os.getenv("VLLM_ENABLE_RESPONSES_API_STORE", "0"))
    ),
xiao-llm's avatar
xiao-llm committed
1465
    # If set, use the fp8 mfma in rocm paged attention.
1466
1467
1468
    "VLLM_ROCM_FP8_MFMA_PAGE_ATTN": lambda: bool(
        int(os.getenv("VLLM_ROCM_FP8_MFMA_PAGE_ATTN", "0"))
    ),
1469
    # Whether to use pytorch symmetric memory for allreduce
1470
    "VLLM_ALLREDUCE_USE_SYMM_MEM": lambda: bool(
1471
        int(os.getenv("VLLM_ALLREDUCE_USE_SYMM_MEM", "1"))
1472
    ),
1473
1474
1475
1476
    # Whether to use FlashInfer allreduce
    "VLLM_ALLREDUCE_USE_FLASHINFER": lambda: bool(
        int(os.getenv("VLLM_ALLREDUCE_USE_FLASHINFER", "0"))
    ),
1477
1478
1479
1480
    # Experimental: use this to enable MCP tool calling for non harmony models
    "VLLM_USE_EXPERIMENTAL_PARSER_CONTEXT": lambda: bool(
        int(os.getenv("VLLM_USE_EXPERIMENTAL_PARSER_CONTEXT", "0"))
    ),
1481
    # Allows vllm to find tuned config under customized folder
1482
    "VLLM_TUNED_CONFIG_FOLDER": lambda: os.getenv("VLLM_TUNED_CONFIG_FOLDER", None),
1483
1484
1485
1486
1487
1488
1489
1490
1491
    # Valid values are container,code_interpreter,web_search_preview
    # ex VLLM_GPT_OSS_SYSTEM_TOOL_MCP_LABELS=container,code_interpreter
    # If the server_label of your mcp tool is not in this list it will
    # be completely ignored.
    "VLLM_GPT_OSS_SYSTEM_TOOL_MCP_LABELS": env_set_with_choices(
        "VLLM_GPT_OSS_SYSTEM_TOOL_MCP_LABELS",
        default=[],
        choices=["container", "code_interpreter", "web_search_preview"],
    ),
1492
    # Allows harmony instructions to be injected on system messages
1493
1494
1495
    "VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS": lambda: bool(
        int(os.getenv("VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS", "0"))
    ),
1496
1497
1498
1499
1500
1501
    # Pin the conversation start date injected into the Harmony system
    # message. When unset the current date is used, which introduces
    # non-determinism (different tokens -> different model behaviour at
    # temperature=0). Set to an ISO date string, e.g. "2023-09-12",
    # for reproducible inference or testing.
    "VLLM_SYSTEM_START_DATE": lambda: os.getenv("VLLM_SYSTEM_START_DATE", None),
1502
1503
1504
1505
1506
1507
    # Enable automatic retry when tool call JSON parsing fails
    # If enabled, returns an error message to the model to retry
    # If disabled (default), raises an exception and fails the request
    "VLLM_TOOL_JSON_ERROR_AUTOMATIC_RETRY": lambda: bool(
        int(os.getenv("VLLM_TOOL_JSON_ERROR_AUTOMATIC_RETRY", "0"))
    ),
1508
    # Add optional custom scopes for profiling, disable to avoid overheads
1509
1510
1511
    "VLLM_CUSTOM_SCOPES_FOR_PROFILING": lambda: bool(
        int(os.getenv("VLLM_CUSTOM_SCOPES_FOR_PROFILING", "0"))
    ),
1512
    # Add optional nvtx scopes for profiling, disable to avoid overheads
1513
1514
1515
    "VLLM_NVTX_SCOPES_FOR_PROFILING": lambda: bool(
        int(os.getenv("VLLM_NVTX_SCOPES_FOR_PROFILING", "0"))
    ),
1516
1517
    # Represent block hashes in KV cache events as 64-bit integers instead of
    # raw bytes. Defaults to True for backward compatibility.
1518
1519
1520
    "VLLM_KV_EVENTS_USE_INT_BLOCK_HASHES": lambda: bool(
        int(os.getenv("VLLM_KV_EVENTS_USE_INT_BLOCK_HASHES", "1"))
    ),
1521
1522
    # Name of the shared memory buffer used for object storage.
    # Only effective when mm_config.mm_processor_cache_type == "shm".
1523
1524
1525
1526
1527
    # Automatically generates a unique UUID-based name per process tree
    # if not explicitly set.
    "VLLM_OBJECT_STORAGE_SHM_BUFFER_NAME": get_env_or_set_default(
        "VLLM_OBJECT_STORAGE_SHM_BUFFER_NAME",
        lambda: f"VLLM_OBJECT_STORAGE_SHM_BUFFER_{uuid.uuid4().hex}",
1528
    ),
1529
    # The size in MB of the buffers (NVL and RDMA) used by DeepEP
1530
1531
1532
    "VLLM_DEEPEP_BUFFER_SIZE_MB": lambda: int(
        os.getenv("VLLM_DEEPEP_BUFFER_SIZE_MB", "1024")
    ),
1533
    # Force DeepEP to use intranode kernel for inter-node communication in
Jiayi Yan's avatar
Jiayi Yan committed
1534
    # high throughput mode. This is useful archive higher prefill throughput
1535
1536
1537
1538
1539
1540
1541
1542
1543
    # on system supports multi-node nvlink (e.g GB200).
    "VLLM_DEEPEP_HIGH_THROUGHPUT_FORCE_INTRA_NODE": lambda: bool(
        int(os.getenv("VLLM_DEEPEP_HIGH_THROUGHPUT_FORCE_INTRA_NODE", "0"))
    ),
    # Allow DeepEP to use MNNVL (multi-node nvlink) for internode_ll kernel,
    # turn this for better latency on GB200 like system
    "VLLM_DEEPEP_LOW_LATENCY_USE_MNNVL": lambda: bool(
        int(os.getenv("VLLM_DEEPEP_LOW_LATENCY_USE_MNNVL", "0"))
    ),
1544
1545
    # The number of SMs to allocate for communication kernels when running DBO
    # the rest of the SMs on the device will be allocated to compute
1546
    "VLLM_DBO_COMM_SMS": lambda: int(os.getenv("VLLM_DBO_COMM_SMS", "20")),
1547
1548
1549
    # Enable max_autotune & coordinate_descent_tuning in inductor_config
    # to compile static shapes passed from compile_sizes in compilation_config
    # If set to 1, enable max_autotune; By default, this is enabled (1)
1550
1551
1552
    "VLLM_ENABLE_INDUCTOR_MAX_AUTOTUNE": lambda: bool(
        int(os.getenv("VLLM_ENABLE_INDUCTOR_MAX_AUTOTUNE", "1"))
    ),
1553
1554
    # If set to 1, enable coordinate_descent_tuning;
    # By default, this is enabled (1)
1555
1556
1557
    "VLLM_ENABLE_INDUCTOR_COORDINATE_DESCENT_TUNING": lambda: bool(
        int(os.getenv("VLLM_ENABLE_INDUCTOR_COORDINATE_DESCENT_TUNING", "1"))
    ),
1558
    # Flag to enable NCCL symmetric memory allocation and registration
1559
1560
1561
    "VLLM_USE_NCCL_SYMM_MEM": lambda: bool(
        int(os.getenv("VLLM_USE_NCCL_SYMM_MEM", "0"))
    ),
1562
    # NCCL header path
1563
    "VLLM_NCCL_INCLUDE_PATH": lambda: os.environ.get("VLLM_NCCL_INCLUDE_PATH", None),
1564
1565
    # Flag to enable FBGemm kernels on model execution
    "VLLM_USE_FBGEMM": lambda: bool(int(os.getenv("VLLM_USE_FBGEMM", "0"))),
1566
1567
1568
1569
1570
1571
    # GC debug config
    # - VLLM_GC_DEBUG=0: disable GC debugger
    # - VLLM_GC_DEBUG=1: enable GC debugger with gc.collect elpased times
    # - VLLM_GC_DEBUG='{"top_objects":5}': enable GC debugger with
    #                                      top 5 collected objects
    "VLLM_GC_DEBUG": lambda: os.getenv("VLLM_GC_DEBUG", ""),
1572
1573
1574
    # Debug workspace allocations.
    # logging of workspace resize operations.
    "VLLM_DEBUG_WORKSPACE": lambda: bool(int(os.getenv("VLLM_DEBUG_WORKSPACE", "0"))),
1575
    # Disables parallel execution of shared_experts via separate cuda stream
1576
1577
    "VLLM_DISABLE_SHARED_EXPERTS_STREAM": lambda: bool(
        int(os.getenv("VLLM_DISABLE_SHARED_EXPERTS_STREAM", "0"))
1578
    ),
1579
1580
1581
1582
1583
1584
1585
    # Limits when we run shared_experts in a separate stream.
    # We found out that for large batch sizes, the separate stream
    # execution is not beneficial (most likely because of the input clone)
    # TODO(alexm-redhat): Tune to be more dynamic based on GPU type
    "VLLM_SHARED_EXPERTS_STREAM_TOKEN_THRESHOLD": lambda: int(
        int(os.getenv("VLLM_SHARED_EXPERTS_STREAM_TOKEN_THRESHOLD", 256))
    ),
1586
1587
1588
1589
1590
1591
1592
1593
1594
    # Format for saving torch.compile cache artifacts
    # - "binary": saves as binary file
    #     Safe for multiple vllm serve processes accessing the same torch compile cache.
    # - "unpacked": saves as directory structure (for inspection/debugging)
    #     NOT multiprocess safe - race conditions may occur with multiple processes.
    #     Allows viewing and setting breakpoints in Inductor's code output files.
    "VLLM_COMPILE_CACHE_SAVE_FORMAT": env_with_choices(
        "VLLM_COMPILE_CACHE_SAVE_FORMAT", "binary", ["binary", "unpacked"]
    ),
Woosuk Kwon's avatar
Woosuk Kwon committed
1595
1596
1597
1598
    # Flag to enable v2 model runner.
    "VLLM_USE_V2_MODEL_RUNNER": lambda: bool(
        int(os.getenv("VLLM_USE_V2_MODEL_RUNNER", "0"))
    ),
1599
1600
1601
1602
1603
1604
    # Log model inspection after loading.
    # If enabled, logs a transformers-style hierarchical view of the model
    # with quantization methods and attention backends.
    "VLLM_LOG_MODEL_INSPECTION": lambda: bool(
        int(os.getenv("VLLM_LOG_MODEL_INSPECTION", "0"))
    ),
1605
1606
1607
1608
    # Debug logging for --enable-mfu-metrics
    "VLLM_DEBUG_MFU_METRICS": lambda: bool(
        int(os.getenv("VLLM_DEBUG_MFU_METRICS", "0"))
    ),
1609
1610
1611
1612
1613
1614
1615
1616
    # Disable using pytorch's pin memory for CPU offloading.
    "VLLM_WEIGHT_OFFLOADING_DISABLE_PIN_MEMORY": lambda: bool(
        int(os.getenv("VLLM_WEIGHT_OFFLOADING_DISABLE_PIN_MEMORY", "0"))
    ),
    # Disable using UVA (Unified Virtual Addressing) for CPU offloading.
    "VLLM_WEIGHT_OFFLOADING_DISABLE_UVA": lambda: bool(
        int(os.getenv("VLLM_WEIGHT_OFFLOADING_DISABLE_UVA", "0"))
    ),
1617
1618
    # Disable logging of vLLM logo at server startup time.
    "VLLM_DISABLE_LOG_LOGO": lambda: bool(int(os.getenv("VLLM_DISABLE_LOG_LOGO", "0"))),
1619
1620
1621
    # Disable PDL for LoRA, as enabling PDL with LoRA on SM100 causes
    # Triton compilation to fail.
    "VLLM_LORA_DISABLE_PDL": lambda: bool(int(os.getenv("VLLM_LORA_DISABLE_PDL", "0"))),
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
    # Enable CUDA compatibility mode for datacenter GPUs with older
    # driver versions than the CUDA toolkit major version of vLLM.
    "VLLM_ENABLE_CUDA_COMPATIBILITY": lambda: (
        os.environ.get("VLLM_ENABLE_CUDA_COMPATIBILITY", "0").strip().lower()
        in ("1", "true")
    ),
    # Path to the CUDA compatibility libraries when CUDA compatibility is enabled.
    "VLLM_CUDA_COMPATIBILITY_PATH": lambda: os.environ.get(
        "VLLM_CUDA_COMPATIBILITY_PATH", None
    ),
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
    # Whether it is a scale up launch engine for elastic EP,
    # Should only be set by EngineCoreClient.
    "VLLM_ELASTIC_EP_SCALE_UP_LAUNCH": lambda: bool(
        int(os.getenv("VLLM_ELASTIC_EP_SCALE_UP_LAUNCH", "0"))
    ),
    # Whether to wait for all requests to drain before sending the
    # scaling command in elastic EP.
    "VLLM_ELASTIC_EP_DRAIN_REQUESTS": lambda: bool(
        int(os.getenv("VLLM_ELASTIC_EP_DRAIN_REQUESTS", "0"))
    ),
1642
1643
1644
1645
1646
1647
    # If set to 1, enable CUDA graph memory estimation during memory profiling.
    # This profiles CUDA graph memory usage to provide more accurate KV cache
    # memory allocation. Disabled by default to preserve existing behavior.
    "VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS": lambda: bool(
        int(os.getenv("VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS", "0"))
    ),
1648
1649
1650
    # NIXL EP environment variables
    "VLLM_NIXL_EP_MAX_NUM_RANKS": lambda: int(
        os.getenv("VLLM_NIXL_EP_MAX_NUM_RANKS", "32")
1651
1652
1653
1654
    ),
    # Whether enable XPU graph on Intel GPU
    "VLLM_XPU_ENABLE_XPU_GRAPH": lambda: bool(
        int(os.getenv("VLLM_XPU_ENABLE_XPU_GRAPH", "0"))
1655
    ),
1656
1657
}

1658

1659
# --8<-- [end:env-vars-definition]
1660

1661

1662
def __getattr__(name: str):
1663
1664
1665
1666
1667
1668
    """
    Gets environment variables lazily.

    NOTE: After enable_envs_cache() invocation (which triggered after service
    initialization), all environment variables will be cached.
    """
1669
1670
1671
1672
1673
    if name in environment_variables:
        return environment_variables[name]()
    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")


1674
1675
1676
1677
1678
1679
def _is_envs_cache_enabled() -> bool:
    """Checked if __getattr__ is wrapped with functools.cache"""
    global __getattr__
    return hasattr(__getattr__, "cache_clear")


1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
def enable_envs_cache() -> None:
    """
    Enables caching of environment variables. This is useful for performance
    reasons, as it avoids the need to re-evaluate environment variables on
    every call.

    NOTE: Currently, it's invoked after service initialization to reduce
    runtime overhead. This also means that environment variables should NOT
    be updated after the service is initialized.
    """
1690
1691
1692
    if _is_envs_cache_enabled():
        # Avoid wrapping functools.cache multiple times
        return
1693
1694
1695
1696
1697
1698
1699
1700
1701
    # Tag __getattr__ with functools.cache
    global __getattr__
    __getattr__ = functools.cache(__getattr__)

    # Cache all environment variables
    for key in environment_variables:
        __getattr__(key)


1702
1703
1704
1705
1706
1707
1708
1709
def disable_envs_cache() -> None:
    """
    Resets the environment variables cache. It could be used to isolate environments
    between unit tests.
    """
    global __getattr__
    # If __getattr__ is wrapped by functions.cache, unwrap the caching layer.
    if _is_envs_cache_enabled():
1710
        assert hasattr(__getattr__, "__wrapped__")
1711
1712
1713
        __getattr__ = __getattr__.__wrapped__


1714
1715
def __dir__():
    return list(environment_variables.keys())
1716
1717
1718
1719
1720
1721
1722
1723
1724


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}")


1725
1726
1727
1728
1729
1730
1731
1732
1733
def validate_environ(hard_fail: bool) -> None:
    for env in os.environ:
        if env.startswith("VLLM_") and env not in environment_variables:
            if hard_fail:
                raise ValueError(f"Unknown vLLM environment variable detected: {env}")
            else:
                logger.warning("Unknown vLLM environment variable detected: %s", env)


1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
def compile_factors() -> dict[str, object]:
    """Return env vars used for torch.compile cache keys.

    Start with every known vLLM env var; drop entries in `ignored_factors`;
    hash everything else. This keeps the cache key aligned across workers."""

    ignored_factors: set[str] = {
        "MAX_JOBS",
        "VLLM_RPC_BASE_PATH",
        "VLLM_USE_MODELSCOPE",
        "VLLM_RINGBUFFER_WARNING_INTERVAL",
        "VLLM_DEBUG_DUMP_PATH",
        "VLLM_PORT",
        "VLLM_CACHE_ROOT",
        "LD_LIBRARY_PATH",
        "VLLM_SERVER_DEV_MODE",
        "VLLM_DP_MASTER_IP",
        "VLLM_DP_MASTER_PORT",
        "VLLM_RANDOMIZE_DP_DUMMY_INPUTS",
        "VLLM_CI_USE_S3",
        "VLLM_MODEL_REDIRECT_PATH",
        "VLLM_HOST_IP",
1756
        "VLLM_FORCE_AOT_LOAD",
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
        "S3_ACCESS_KEY_ID",
        "S3_SECRET_ACCESS_KEY",
        "S3_ENDPOINT_URL",
        "VLLM_USAGE_STATS_SERVER",
        "VLLM_NO_USAGE_STATS",
        "VLLM_DO_NOT_TRACK",
        "VLLM_LOGGING_LEVEL",
        "VLLM_LOGGING_PREFIX",
        "VLLM_LOGGING_STREAM",
        "VLLM_LOGGING_CONFIG_PATH",
Nick Hill's avatar
Nick Hill committed
1767
        "VLLM_LOGGING_COLOR",
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
        "VLLM_LOG_STATS_INTERVAL",
        "VLLM_DEBUG_LOG_API_SERVER_RESPONSE",
        "VLLM_TUNED_CONFIG_FOLDER",
        "VLLM_ENGINE_ITERATION_TIMEOUT_S",
        "VLLM_HTTP_TIMEOUT_KEEP_ALIVE",
        "VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS",
        "VLLM_KEEP_ALIVE_ON_ENGINE_DEATH",
        "VLLM_IMAGE_FETCH_TIMEOUT",
        "VLLM_VIDEO_FETCH_TIMEOUT",
        "VLLM_AUDIO_FETCH_TIMEOUT",
1778
        "VLLM_MEDIA_FETCH_MAX_RETRIES",
1779
1780
1781
1782
1783
        "VLLM_MEDIA_URL_ALLOW_REDIRECTS",
        "VLLM_MEDIA_LOADING_THREAD_COUNT",
        "VLLM_MAX_AUDIO_CLIP_FILESIZE_MB",
        "VLLM_VIDEO_LOADER_BACKEND",
        "VLLM_MEDIA_CONNECTOR",
1784
        "VLLM_OBJECT_STORAGE_SHM_BUFFER_NAME",
1785
1786
1787
1788
1789
1790
1791
        "VLLM_ASSETS_CACHE",
        "VLLM_ASSETS_CACHE_MODEL_CLEAN",
        "VLLM_WORKER_MULTIPROC_METHOD",
        "VLLM_ENABLE_V1_MULTIPROCESSING",
        "VLLM_V1_OUTPUT_PROC_CHUNK_SIZE",
        "VLLM_CPU_KVCACHE_SPACE",
        "VLLM_CPU_MOE_PREPACK",
1792
        "VLLM_ZENTORCH_WEIGHT_PREPACK",
1793
        "VLLM_TEST_FORCE_LOAD_FORMAT",
1794
1795
        "VLLM_ENABLE_CUDA_COMPATIBILITY",
        "VLLM_CUDA_COMPATIBILITY_PATH",
1796
1797
        "LOCAL_RANK",
        "CUDA_VISIBLE_DEVICES",
Nick Hill's avatar
Nick Hill committed
1798
        "NO_COLOR",
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
    }

    from vllm.config.utils import normalize_value

    factors: dict[str, object] = {}
    for factor, getter in environment_variables.items():
        if factor in ignored_factors:
            continue

        try:
            raw = getter()
        except Exception as exc:  # pragma: no cover - defensive logging
            logger.warning(
                "Skipping environment variable %s while hashing compile factors: %s",
                factor,
                exc,
            )
            continue
1817

1818
        factors[factor] = normalize_value(raw)
1819

1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
    ray_noset_env_vars = [
        # Refer to
        # https://github.com/ray-project/ray/blob/c584b1ea97b00793d1def71eaf81537d70efba42/python/ray/_private/accelerators/nvidia_gpu.py#L11
        # https://github.com/ray-project/ray/blob/c584b1ea97b00793d1def71eaf81537d70efba42/python/ray/_private/accelerators/amd_gpu.py#L11
        # https://github.com/ray-project/ray/blob/b97d21dab233c2bd8ed7db749a82a1e594222b5c/python/ray/_private/accelerators/amd_gpu.py#L10
        # https://github.com/ray-project/ray/blob/c584b1ea97b00793d1def71eaf81537d70efba42/python/ray/_private/accelerators/npu.py#L12
        # https://github.com/ray-project/ray/blob/c584b1ea97b00793d1def71eaf81537d70efba42/python/ray/_private/accelerators/hpu.py#L12
        # https://github.com/ray-project/ray/blob/c584b1ea97b00793d1def71eaf81537d70efba42/python/ray/_private/accelerators/neuron.py#L14
        # https://github.com/ray-project/ray/blob/c584b1ea97b00793d1def71eaf81537d70efba42/python/ray/_private/accelerators/tpu.py#L38
        # https://github.com/ray-project/ray/blob/c584b1ea97b00793d1def71eaf81537d70efba42/python/ray/_private/accelerators/intel_gpu.py#L10
        # https://github.com/ray-project/ray/blob/c584b1ea97b00793d1def71eaf81537d70efba42/python/ray/_private/accelerators/rbln.py#L10
        "RAY_EXPERIMENTAL_NOSET_CUDA_VISIBLE_DEVICES",
        "RAY_EXPERIMENTAL_NOSET_ROCR_VISIBLE_DEVICES",
        "RAY_EXPERIMENTAL_NOSET_HIP_VISIBLE_DEVICES",
        "RAY_EXPERIMENTAL_NOSET_ASCEND_RT_VISIBLE_DEVICES",
        "RAY_EXPERIMENTAL_NOSET_HABANA_VISIBLE_MODULES",
        "RAY_EXPERIMENTAL_NOSET_NEURON_RT_VISIBLE_CORES",
        "RAY_EXPERIMENTAL_NOSET_TPU_VISIBLE_CHIPS",
        "RAY_EXPERIMENTAL_NOSET_ONEAPI_DEVICE_SELECTOR",
        "RAY_EXPERIMENTAL_NOSET_RBLN_RT_VISIBLE_DEVICES",
    ]

1842
1843
    for var in ray_noset_env_vars:
        factors[var] = normalize_value(os.getenv(var))
1844

1845
    return factors