envs.py 73.1 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 hashlib
6
import json
7
import os
8
import sys
9
import tempfile
10
11
from collections.abc import Callable
from typing import TYPE_CHECKING, Any, Literal
12
13
14

if TYPE_CHECKING:
    VLLM_HOST_IP: str = ""
15
    VLLM_PORT: int | None = None
16
    VLLM_RPC_BASE_PATH: str = tempfile.gettempdir()
17
    VLLM_USE_MODELSCOPE: bool = False
18
    VLLM_RINGBUFFER_WARNING_INTERVAL: int = 60
19
20
    VLLM_NCCL_SO_PATH: str | None = None
    LD_LIBRARY_PATH: str | None = None
21
    VLLM_V1_USE_PREFILL_DECODE_ATTENTION: bool = False
22
    VLLM_FLASH_ATTN_VERSION: int | None = None
23
    LOCAL_RANK: int = 0
24
    CUDA_VISIBLE_DEVICES: str | None = None
25
    VLLM_ENGINE_ITERATION_TIMEOUT_S: int = 60
26
    VLLM_API_KEY: str | None = None
27
    VLLM_DEBUG_LOG_API_SERVER_RESPONSE: bool = False
28
29
30
31
    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
32
33
    VLLM_CACHE_ROOT: str = os.path.expanduser("~/.cache/vllm")
    VLLM_CONFIG_ROOT: str = os.path.expanduser("~/.config/vllm")
34
35
    VLLM_USAGE_STATS_SERVER: str = "https://stats.vllm.ai"
    VLLM_NO_USAGE_STATS: bool = False
36
    VLLM_DISABLE_FLASHINFER_PREFILL: bool = False
37
38
39
    VLLM_DO_NOT_TRACK: bool = False
    VLLM_USAGE_SOURCE: str = ""
    VLLM_CONFIGURE_LOGGING: int = 1
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
44
    VLLM_LOG_STATS_INTERVAL: float = 10.0
45
    VLLM_TRACE_FUNCTION: int = 0
46
47
48
49
    VLLM_ATTENTION_BACKEND: str | None = None
    VLLM_USE_FLASHINFER_SAMPLER: bool | None = None
    VLLM_PP_LAYER_PARTITION: str | None = None
    VLLM_CPU_KVCACHE_SPACE: int | None = 0
50
    VLLM_CPU_OMP_THREADS_BIND: str = ""
51
    VLLM_CPU_NUM_OF_RESERVED_CPU: int | None = None
52
    VLLM_CPU_MOE_PREPACK: bool = True
53
    VLLM_CPU_SGL_KERNEL: bool = False
54
    VLLM_XLA_CACHE_PATH: str = os.path.join(VLLM_CACHE_ROOT, "xla_cache")
55
    VLLM_XLA_CHECK_RECOMPILATION: bool = False
56
    VLLM_FUSED_MOE_CHUNK_SIZE: int = 64 * 1024
57
    VLLM_ENABLE_FUSED_MOE_ACTIVATION_CHUNKING: bool = True
58
    VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE: Literal["auto", "nccl", "shm"] = "auto"
59
    VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM: bool = False
60
    VLLM_USE_RAY_WRAPPED_PP_COMM: bool = True
61
    VLLM_XLA_USE_SPMD: bool = False
62
    VLLM_WORKER_MULTIPROC_METHOD: Literal["fork", "spawn"] = "fork"
63
    VLLM_ASSETS_CACHE: str = os.path.join(VLLM_CACHE_ROOT, "assets")
64
    VLLM_ASSETS_CACHE_MODEL_CLEAN: bool = False
65
    VLLM_IMAGE_FETCH_TIMEOUT: int = 5
66
    VLLM_VIDEO_FETCH_TIMEOUT: int = 30
67
    VLLM_AUDIO_FETCH_TIMEOUT: int = 10
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_INPUT_CACHE_GIB: int = 4
74
    VLLM_TARGET_DEVICE: str = "cuda"
75
    VLLM_MAIN_CUDA_VERSION: str = "12.8"
76
77
    MAX_JOBS: str | None = None
    NVCC_THREADS: str | None = None
78
    VLLM_USE_PRECOMPILED: bool = False
79
    VLLM_DOCKER_BUILD_CONTEXT: bool = False
80
    VLLM_TEST_USE_PRECOMPILED_NIGHTLY_WHEEL: bool = False
81
    VLLM_KEEP_ALIVE_ON_ENGINE_DEATH: bool = False
82
    CMAKE_BUILD_TYPE: Literal["Debug", "Release", "RelWithDebInfo"] | None = None
83
    VERBOSE: bool = False
84
    VLLM_ALLOW_LONG_MAX_MODEL_LEN: bool = False
85
    VLLM_RPC_TIMEOUT: int = 10000  # ms
86
    VLLM_HTTP_TIMEOUT_KEEP_ALIVE: int = 5  # seconds
87
88
    VLLM_PLUGINS: list[str] | None = None
    VLLM_LORA_RESOLVER_CACHE_DIR: str | None = None
89
    VLLM_TORCH_CUDA_PROFILE: bool = False
90
    VLLM_TORCH_PROFILER_DIR: str | None = None
91
92
    VLLM_TORCH_PROFILER_RECORD_SHAPES: bool = False
    VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY: bool = False
93
94
    VLLM_USE_AOT_COMPILE: bool = False
    VLLM_FORCE_AOT_LOAD: bool = False
95
96
    VLLM_TORCH_PROFILER_WITH_STACK: bool = True
    VLLM_TORCH_PROFILER_WITH_FLOPS: bool = False
97
    VLLM_USE_TRITON_AWQ: bool = False
98
    VLLM_ALLOW_RUNTIME_LORA_UPDATING: bool = False
99
    VLLM_SKIP_P2P_CHECK: bool = False
100
    VLLM_DISABLED_KERNELS: list[str] = []
101
    VLLM_DISABLE_PYNCCL: bool = False
102
    VLLM_ROCM_USE_AITER: bool = False
103
    VLLM_ROCM_USE_AITER_PAGED_ATTN: bool = False
104
    VLLM_ROCM_USE_AITER_LINEAR: bool = True
105
    VLLM_ROCM_USE_AITER_MOE: bool = True
106
    VLLM_ROCM_USE_AITER_RMSNORM: bool = True
107
    VLLM_ROCM_USE_AITER_MLA: bool = True
108
    VLLM_ROCM_USE_AITER_MHA: bool = True
109
    VLLM_ROCM_USE_AITER_FP4_ASM_GEMM: bool = False
110
    VLLM_ROCM_USE_AITER_TRITON_ROPE: bool = False
111
    VLLM_ROCM_USE_AITER_FP8BMM: bool = True
112
    VLLM_ROCM_USE_AITER_UNIFIED_ATTENTION: bool = False
113
    VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS: bool = True
114
    VLLM_ROCM_USE_AITER_TRITON_GEMM: bool = True
115
    VLLM_ROCM_USE_SKINNY_GEMM: bool = True
116
    VLLM_ROCM_FP8_PADDING: bool = True
117
    VLLM_ROCM_MOE_PADDING: bool = True
118
    VLLM_ROCM_CUSTOM_PAGED_ATTN: bool = True
119
    VLLM_ENABLE_V1_MULTIPROCESSING: bool = True
120
    VLLM_LOG_BATCHSIZE_INTERVAL: float = -1
121
    VLLM_DISABLE_COMPILE_CACHE: bool = False
122
    Q_SCALE_CONSTANT: int = 200
123
124
    K_SCALE_CONSTANT: int = 200
    V_SCALE_CONSTANT: int = 100
125
    VLLM_SERVER_DEV_MODE: bool = False
126
    VLLM_V1_OUTPUT_PROC_CHUNK_SIZE: int = 128
127
    VLLM_MLA_DISABLE: bool = False
128
    VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH: int = 32
129
130
    VLLM_RAY_PER_WORKER_GPUS: float = 1.0
    VLLM_RAY_BUNDLE_INDICES: str = ""
131
    VLLM_CUDART_SO_PATH: str | None = None
132
    VLLM_DP_RANK: int = 0
133
    VLLM_DP_RANK_LOCAL: int = -1
134
    VLLM_DP_SIZE: int = 1
135
    VLLM_USE_STANDALONE_COMPILE: bool = True
136
137
    VLLM_DP_MASTER_IP: str = ""
    VLLM_DP_MASTER_PORT: int = 0
138
    VLLM_MOE_DP_CHUNK_SIZE: int = 256
139
    VLLM_RANDOMIZE_DP_DUMMY_INPUTS: bool = False
140
    VLLM_RAY_DP_PACK_STRATEGY: Literal["strict", "fill", "span"] = "strict"
141
    VLLM_MARLIN_USE_ATOMIC_ADD: bool = False
142
    VLLM_MXFP4_USE_MARLIN: bool | None = None
143
    VLLM_V1_USE_OUTLINES_CACHE: bool = False
144
    VLLM_TPU_BUCKET_PADDING_GAP: int = 0
145
    VLLM_TPU_MOST_MODEL_LEN: int | None = None
146
    VLLM_TPU_USING_PATHWAYS: bool = False
147
    VLLM_USE_DEEP_GEMM: bool = True
148
    VLLM_MOE_USE_DEEP_GEMM: bool = True
149
    VLLM_USE_DEEP_GEMM_E8M0: bool = True
150
151
152
153
154
    VLLM_DEEP_GEMM_WARMUP: Literal[
        "skip",
        "full",
        "relax",
    ] = "relax"
155
    VLLM_USE_FUSED_MOE_GROUPED_TOPK: bool = True
156
    VLLM_USE_FLASHINFER_MOE_FP16: bool = False
157
158
    VLLM_USE_FLASHINFER_MOE_FP8: bool = False
    VLLM_USE_FLASHINFER_MOE_FP4: bool = False
159
    VLLM_FLASHINFER_MOE_BACKEND: Literal["throughput", "latency"] = "latency"
160
    VLLM_FLASHINFER_WORKSPACE_BUFFER_SIZE: int = 394 * 1024 * 1024
161
    VLLM_XGRAMMAR_CACHE_MB: int = 0
162
    VLLM_MSGPACK_ZERO_COPY_THRESHOLD: int = 256
163
    VLLM_ALLOW_INSECURE_SERIALIZATION: bool = False
Robert Shaw's avatar
Robert Shaw committed
164
    VLLM_NIXL_SIDE_CHANNEL_HOST: str = "localhost"
165
    VLLM_NIXL_SIDE_CHANNEL_PORT: int = 5600
166
167
168
169
170
171
172
173
    VLLM_ALL2ALL_BACKEND: Literal[
        "naive",
        "pplx",
        "deepep_high_throughput",
        "deepep_low_latency",
        "allgather_reducescatter",
        "flashinfer_all2allv",
    ] = "allgather_reducescatter"
174
    VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE: int = 163840
175
    VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS: int = 1
176
    VLLM_SLEEP_WHEN_IDLE: bool = False
177
    VLLM_MQ_MAX_CHUNK_BYTES_MB: int = 16
178
    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: int = 300
179
    VLLM_KV_CACHE_LAYOUT: Literal["NHD", "HND"] | None = None
180
    VLLM_COMPUTE_NANS_IN_LOGITS: bool = False
181
    VLLM_USE_NVFP4_CT_EMULATIONS: bool = False
182
183
184
    VLLM_ROCM_QUICK_REDUCE_QUANTIZATION: Literal[
        "FP", "INT8", "INT6", "INT4", "NONE"
    ] = "NONE"
185
    VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16: bool = True
186
    VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB: int | None = None
187
    VLLM_NIXL_ABORT_REQUEST_TIMEOUT: int = 480
188
    VLLM_USE_CUDNN_PREFILL: bool = False
189
    VLLM_USE_TRTLLM_RAGGED_DEEPSEEK_PREFILL: bool = False
190
    VLLM_ENABLE_CUDAGRAPH_GC: bool = False
191
    VLLM_LOOPBACK_IP: str = ""
192
    VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE: bool = False
193
    VLLM_ENABLE_RESPONSES_API_STORE: bool = False
194
    VLLM_USE_TRTLLM_ATTENTION: str | None = None
195
    VLLM_NVFP4_GEMM_BACKEND: str | None = None
196
    VLLM_FLASHINFER_DISABLE_Q_QUANTIZATION: bool = False
197
    VLLM_HAS_FLASHINFER_CUBIN: bool = False
198
199
    VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8: bool = False
    VLLM_USE_FLASHINFER_MOE_MXFP4_BF16: bool = False
xiao-llm's avatar
xiao-llm committed
200
    VLLM_ROCM_FP8_MFMA_PAGE_ATTN: bool = False
201
    VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS: bool = False
202
    VLLM_ALLREDUCE_USE_SYMM_MEM: bool = True
203
    VLLM_TUNED_CONFIG_FOLDER: str | None = None
204
    VLLM_GPT_OSS_SYSTEM_TOOL_MCP_LABELS: set[str] = set()
205
    VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS: bool = False
206
    VLLM_TOOL_JSON_ERROR_AUTOMATIC_RETRY: bool = False
207
    VLLM_CUSTOM_SCOPES_FOR_PROFILING: bool = False
208
    VLLM_NVTX_SCOPES_FOR_PROFILING: bool = False
209
    VLLM_KV_EVENTS_USE_INT_BLOCK_HASHES: bool = True
210
    VLLM_OBJECT_STORAGE_SHM_BUFFER_NAME: str = "VLLM_OBJECT_STORAGE_SHM_BUFFER"
211
    VLLM_DEEPEP_BUFFER_SIZE_MB: int = 1024
212
213
    VLLM_DEEPEP_HIGH_THROUGHPUT_FORCE_INTRA_NODE: bool = False
    VLLM_DEEPEP_LOW_LATENCY_USE_MNNVL: bool = False
214
    VLLM_DBO_COMM_SMS: int = 20
215
216
    VLLM_PATTERN_MATCH_DEBUG: str | None = None
    VLLM_DEBUG_DUMP_PATH: str | None = None
217
218
    VLLM_ENABLE_INDUCTOR_MAX_AUTOTUNE: bool = True
    VLLM_ENABLE_INDUCTOR_COORDINATE_DESCENT_TUNING: bool = True
219
    VLLM_USE_NCCL_SYMM_MEM: bool = False
220
    VLLM_NCCL_INCLUDE_PATH: str | None = None
221
    VLLM_USE_FBGEMM: bool = False
222
    VLLM_GC_DEBUG: str = ""
223
    VLLM_DISABLE_SHARED_EXPERTS_STREAM: bool = False
224
    VLLM_COMPILE_CACHE_SAVE_FORMAT: Literal["binary", "unpacked"] = "binary"
225
    VLLM_FLAT_LOGPROBS: bool = False
226

227
228
229
230
231
232
233
234
235
236
237
238
239
240
241

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"),
    )


242
def maybe_convert_int(value: str | None) -> int | None:
243
244
245
246
247
    if value is None:
        return None
    return int(value)


248
def maybe_convert_bool(value: str | None) -> bool | None:
249
250
251
252
253
    if value is None:
        return None
    return bool(int(value))


254
255
256
257
def disable_compile_cache() -> bool:
    return bool(int(os.getenv("VLLM_DISABLE_COMPILE_CACHE", "0")))


258
def use_aot_compile() -> bool:
259
260
261
    from vllm.model_executor.layers.batch_invariant import (
        vllm_is_batch_invariant,
    )
262
    from vllm.utils.torch_utils import is_torch_equal_or_newer
263

264
265
266
267
268
269
    default_value = (
        "1"
        if is_torch_equal_or_newer("2.10.0.dev") and not disable_compile_cache()
        else "0"
    )

270
271
272
273
    return (
        not vllm_is_batch_invariant()
        and os.environ.get("VLLM_USE_AOT_COMPILE", default_value) == "1"
    )
274
275


276
def env_with_choices(
277
    env_name: str,
278
279
    default: str | None,
    choices: list[str] | Callable[[], list[str]],
280
    case_sensitive: bool = True,
281
) -> Callable[[], str | None]:
282
283
    """
    Create a lambda that validates environment variable against allowed choices
284

285
286
287
288
289
    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
290

291
292
293
294
    Returns:
        Lambda function for environment_variables dict
    """

295
    def _get_validated_env() -> str | None:
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
        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:
311
312
313
314
            raise ValueError(
                f"Invalid value '{value}' for {env_name}. "
                f"Valid options: {actual_choices}."
            )
315
316
317
318
319
320

        return value

    return _get_validated_env


321
def env_list_with_choices(
322
323
    env_name: str,
    default: list[str],
324
    choices: list[str] | Callable[[], list[str]],
325
326
    case_sensitive: bool = True,
) -> Callable[[], list[str]]:
327
    """
328
    Create a lambda that validates environment variable
329
    containing comma-separated values against allowed choices
330

331
332
333
334
335
    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
336

337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
    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:
366
367
368
369
                raise ValueError(
                    f"Invalid value '{val}' in {env_name}. "
                    f"Valid options: {actual_choices}."
                )
370
371
372
373
374
375

        return values

    return _get_validated_env_list


376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
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


394
def get_vllm_port() -> int | None:
395
    """Get the port from VLLM_PORT environment variable.
396

397
398
    Returns:
        The port number as an integer if VLLM_PORT is set, None otherwise.
399

400
401
402
    Raises:
        ValueError: If VLLM_PORT is a URI, suggest k8s service discovery issue.
    """
403
    if "VLLM_PORT" not in os.environ:
404
405
        return None

406
    port = os.getenv("VLLM_PORT", "0")
407
408
409
410
411

    try:
        return int(port)
    except ValueError as err:
        from urllib.parse import urlparse
412

413
414
415
416
417
418
419
        parsed = urlparse(port)
        if parsed.scheme:
            raise ValueError(
                f"VLLM_PORT '{port}' appears to be a URI. "
                "This may be caused by a Kubernetes service discovery issue,"
                "check the warning in: https://docs.vllm.ai/en/stable/serving/env_vars.html"
            ) from None
420
        raise ValueError(f"VLLM_PORT '{port}' must be a valid integer") from err
421
422


423
424
425
# The begin-* and end* here are used by the documentation generator
# to extract the used env vars.

426
# --8<-- [start:env-vars-definition]
427

428
environment_variables: dict[str, Callable[[], Any]] = {
429
    # ================== Installation Time Env Vars ==================
430
    # Target device of vLLM, supporting [cuda (by default),
431
    # rocm, cpu]
432
    "VLLM_TARGET_DEVICE": lambda: os.getenv("VLLM_TARGET_DEVICE", "cuda").lower(),
433
434
    # Main CUDA version of vLLM, supporting [12.6, 12.8, 12.9],
    # 12.8 is the default. This follows PyTorch but can be overridden.
435
436
    "VLLM_MAIN_CUDA_VERSION": lambda: os.getenv("VLLM_MAIN_CUDA_VERSION", "").lower()
    or "12.8",
437
438
    # Maximum number of compilation jobs to run in parallel.
    # By default this is the number of CPUs
439
    "MAX_JOBS": lambda: os.getenv("MAX_JOBS", None),
440
441
442
    # Number of threads to use for nvcc
    # By default this is 1.
    # If set, `MAX_JOBS` will be reduced to avoid oversubscribing the CPU.
443
    "NVCC_THREADS": lambda: os.getenv("NVCC_THREADS", None),
444
    # If set, vllm will use precompiled binaries (*.so)
445
446
447
448
449
    "VLLM_USE_PRECOMPILED": lambda: os.environ.get("VLLM_USE_PRECOMPILED", "")
    .strip()
    .lower()
    in ("1", "true")
    or bool(os.environ.get("VLLM_PRECOMPILED_WHEEL_LOCATION")),
450
451
    # Used to mark that setup.py is running in a Docker build context,
    # in order to force the use of precompiled binaries.
452
453
454
455
    "VLLM_DOCKER_BUILD_CONTEXT": lambda: os.environ.get("VLLM_DOCKER_BUILD_CONTEXT", "")
    .strip()
    .lower()
    in ("1", "true"),
456
457
    # Whether to force using nightly wheel in python build.
    # This is used for testing the nightly wheel in python build.
458
459
460
    "VLLM_TEST_USE_PRECOMPILED_NIGHTLY_WHEEL": lambda: bool(
        int(os.getenv("VLLM_TEST_USE_PRECOMPILED_NIGHTLY_WHEEL", "0"))
    ),
461
462
463
    # CMake build type
    # If not set, defaults to "Debug" or "RelWithDebInfo"
    # Available options: "Debug", "Release", "RelWithDebInfo"
464
465
466
    "CMAKE_BUILD_TYPE": env_with_choices(
        "CMAKE_BUILD_TYPE", None, ["Debug", "Release", "RelWithDebInfo"]
    ),
467
    # If set, vllm will print verbose logs during installation
468
    "VERBOSE": lambda: bool(int(os.getenv("VERBOSE", "0"))),
469
    # Root directory for vLLM configuration files
470
    # Defaults to `~/.config/vllm` unless `XDG_CONFIG_HOME` is set
471
472
473
    # 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**.
474
    "VLLM_CONFIG_ROOT": lambda: os.path.expanduser(
475
476
477
        os.getenv(
            "VLLM_CONFIG_ROOT",
            os.path.join(get_default_config_root(), "vllm"),
478
479
        )
    ),
480
    # ================== Runtime Env Vars ==================
481
    # Root directory for vLLM cache files
482
    # Defaults to `~/.cache/vllm` unless `XDG_CACHE_HOME` is set
483
    "VLLM_CACHE_ROOT": lambda: os.path.expanduser(
484
485
486
        os.getenv(
            "VLLM_CACHE_ROOT",
            os.path.join(get_default_cache_root(), "vllm"),
487
488
        )
    ),
489
490
491
492
    # 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.
493
    "VLLM_HOST_IP": lambda: os.getenv("VLLM_HOST_IP", ""),
494
    # used in distributed environment to manually set the communication port
495
496
497
    # 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.
498
    "VLLM_PORT": get_vllm_port,
499
500
    # path used for ipc when the frontend api server is running in
    # multi-processing mode to communicate with the backend engine process.
501
502
503
    "VLLM_RPC_BASE_PATH": lambda: os.getenv(
        "VLLM_RPC_BASE_PATH", tempfile.gettempdir()
    ),
504
505
    # If true, will load models from ModelScope instead of Hugging Face Hub.
    # note that the value is true or false, not numbers
506
507
508
509
    "VLLM_USE_MODELSCOPE": lambda: os.environ.get(
        "VLLM_USE_MODELSCOPE", "False"
    ).lower()
    == "true",
510
    # Interval in seconds to log a warning message when the ring buffer is full
511
512
513
    "VLLM_RINGBUFFER_WARNING_INTERVAL": lambda: int(
        os.environ.get("VLLM_RINGBUFFER_WARNING_INTERVAL", "60")
    ),
514
515
    # path to cudatoolkit home directory, under which should be bin, include,
    # and lib directories.
516
    "CUDA_HOME": lambda: os.environ.get("CUDA_HOME", None),
517
518
    # 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
519
    "VLLM_NCCL_SO_PATH": lambda: os.environ.get("VLLM_NCCL_SO_PATH", None),
520
521
    # 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`
522
    "LD_LIBRARY_PATH": lambda: os.environ.get("LD_LIBRARY_PATH", None),
523
524
    # Use separate prefill and decode kernels for V1 attention instead of
    # the unified triton kernel.
525
526
527
528
    "VLLM_V1_USE_PREFILL_DECODE_ATTENTION": lambda: (
        os.getenv("VLLM_V1_USE_PREFILL_DECODE_ATTENTION", "False").lower()
        in ("true", "1")
    ),
529
530
    # Force vllm to use a specific flash-attention version (2 or 3), only valid
    # when using the flash-attention backend.
531
532
533
    "VLLM_FLASH_ATTN_VERSION": lambda: maybe_convert_int(
        os.environ.get("VLLM_FLASH_ATTN_VERSION", None)
    ),
534
    # Feature flag to enable/disable Inductor standalone compile.
535
536
    # In torch <= 2.7 we ignore this flag; in torch >= 2.9 this is
    # enabled by default.
537
    "VLLM_USE_STANDALONE_COMPILE": lambda: os.environ.get(
538
        "VLLM_USE_STANDALONE_COMPILE", "1"
539
540
    )
    == "1",
541
542
    # Debug pattern matching inside custom passes.
    # Should be set to the fx.Node name (e.g. 'getitem_34' or 'scaled_mm_3').
543
544
545
    "VLLM_PATTERN_MATCH_DEBUG": lambda: os.environ.get(
        "VLLM_PATTERN_MATCH_DEBUG", None
    ),
546
547
    # Dump fx graphs to the given directory.
    # It will override CompilationConfig.debug_dump_path if set.
548
    "VLLM_DEBUG_DUMP_PATH": lambda: os.environ.get("VLLM_DEBUG_DUMP_PATH", None),
549
550
551
552
553
554
555
556
    # 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,
    # 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",
557
558
    # local rank of the process in the distributed setting, used to determine
    # the GPU device id
559
    "LOCAL_RANK": lambda: int(os.environ.get("LOCAL_RANK", "0")),
560
    # used to control the visible devices in the distributed setting
561
    "CUDA_VISIBLE_DEVICES": lambda: os.environ.get("CUDA_VISIBLE_DEVICES", None),
562
    # timeout for each iteration in the engine
563
564
565
    "VLLM_ENGINE_ITERATION_TIMEOUT_S": lambda: int(
        os.environ.get("VLLM_ENGINE_ITERATION_TIMEOUT_S", "60")
    ),
566
    # API key for vLLM API server
567
    "VLLM_API_KEY": lambda: os.environ.get("VLLM_API_KEY", None),
568
    # Whether to log responses from API Server for debugging
569
570
571
572
    "VLLM_DEBUG_LOG_API_SERVER_RESPONSE": lambda: os.environ.get(
        "VLLM_DEBUG_LOG_API_SERVER_RESPONSE", "False"
    ).lower()
    == "true",
573
    # S3 access information, used for tensorizer to load model from S3
574
575
576
    "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),
577
    # Usage stats collection
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
    "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_DISABLE_FLASHINFER_PREFILL": lambda: os.environ.get(
        "VLLM_DISABLE_FLASHINFER_PREFILL", "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"),
593
594
595
596
    # 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
597
598
    "VLLM_CONFIGURE_LOGGING": lambda: int(os.getenv("VLLM_CONFIGURE_LOGGING", "1")),
    "VLLM_LOGGING_CONFIG_PATH": lambda: os.getenv("VLLM_LOGGING_CONFIG_PATH"),
599
    # this is used for configuring the default logging level
600
    "VLLM_LOGGING_LEVEL": lambda: os.getenv("VLLM_LOGGING_LEVEL", "INFO").upper(),
601
    # this is used for configuring the default logging stream
602
    "VLLM_LOGGING_STREAM": lambda: os.getenv("VLLM_LOGGING_STREAM", "ext://sys.stdout"),
603
    # if set, VLLM_LOGGING_PREFIX will be prepended to all log messages
604
    "VLLM_LOGGING_PREFIX": lambda: os.getenv("VLLM_LOGGING_PREFIX", ""),
605
606
    # If set, vllm will log stats at this interval in seconds
    # If not set, vllm will log stats every 10 seconds.
607
608
609
    "VLLM_LOG_STATS_INTERVAL": lambda: val
    if (val := float(os.getenv("VLLM_LOG_STATS_INTERVAL", "10."))) > 0.0
    else 10.0,
610
611
612
    # Trace function calls
    # If set to 1, vllm will trace function calls
    # Useful for debugging
613
    "VLLM_TRACE_FUNCTION": lambda: int(os.getenv("VLLM_TRACE_FUNCTION", "0")),
614
    # Backend for attention computation
615
    # Example options:
616
617
618
    # - "TORCH_SDPA": use torch.nn.MultiheadAttention
    # - "FLASH_ATTN": use FlashAttention
    # - "XFORMERS": use XFormers
619
    # - "FLASHINFER": use flashinfer
620
    # - "FLASHMLA": use FlashMLA
621
    # - "FLASH_ATTN_MLA": use FlashAttention for MLA
622
623
    # - "FLASHINFER_MLA": use FlashInfer for MLA
    # - "CUTLASS_MLA": use CUTLASS for MLA
624
    # All possible options loaded dynamically from AttentionBackendEnum
625
626
627
628
629
    "VLLM_ATTENTION_BACKEND": env_with_choices(
        "VLLM_ATTENTION_BACKEND",
        None,
        lambda: list(
            __import__(
630
631
                "vllm.attention.backends.registry", fromlist=["AttentionBackendEnum"]
            ).AttentionBackendEnum.__members__.keys()
632
633
        ),
    ),
634
    # If set, vllm will use flashinfer sampler
635
636
637
638
639
    "VLLM_USE_FLASHINFER_SAMPLER": lambda: bool(
        int(os.environ["VLLM_USE_FLASHINFER_SAMPLER"])
    )
    if "VLLM_USE_FLASHINFER_SAMPLER" in os.environ
    else None,
640
    # Pipeline stage partition strategy
641
    "VLLM_PP_LAYER_PARTITION": lambda: os.getenv("VLLM_PP_LAYER_PARTITION", None),
642
    # (CPU backend only) CPU key-value cache space.
643
    # default is None and will be set as 4 GB
644
645
646
    "VLLM_CPU_KVCACHE_SPACE": lambda: int(os.getenv("VLLM_CPU_KVCACHE_SPACE", "0"))
    if "VLLM_CPU_KVCACHE_SPACE" in os.environ
    else None,
647
648
    # (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 '|'.
649
    "VLLM_CPU_OMP_THREADS_BIND": lambda: os.getenv("VLLM_CPU_OMP_THREADS_BIND", "auto"),
650
651
    # (CPU backend only) CPU cores not used by OMP threads .
    # Those CPU cores will not be used by OMP threads of a rank.
652
653
654
655
656
    "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,
657
658
659
    # (CPU backend only) whether to use prepack for MoE layer. This will be
    # passed to ipex.llm.modules.GatedMLPMOE. On unsupported CPUs, you might
    # need to set this to "0" (False).
660
    "VLLM_CPU_MOE_PREPACK": lambda: bool(int(os.getenv("VLLM_CPU_MOE_PREPACK", "1"))),
661
    # (CPU backend only) whether to use SGL kernels, optimized for small batch.
662
    "VLLM_CPU_SGL_KERNEL": lambda: bool(int(os.getenv("VLLM_CPU_SGL_KERNEL", "0"))),
663
664
665
666
667
668
669
    # 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
670
671
672
    "VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE": env_with_choices(
        "VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE", "auto", ["auto", "nccl", "shm"]
    ),
673
    # If the env var is set, it enables GPU communication overlap
674
    # (experimental feature) in Ray's Compiled Graph.
675
676
677
    "VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM": lambda: bool(
        int(os.getenv("VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM", "0"))
    ),
678
679
680
    # 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.
681
682
683
    "VLLM_USE_RAY_WRAPPED_PP_COMM": lambda: bool(
        int(os.getenv("VLLM_USE_RAY_WRAPPED_PP_COMM", "1"))
    ),
684
685
    # Use dedicated multiprocess context for workers.
    # Both spawn and fork work
686
687
688
    "VLLM_WORKER_MULTIPROC_METHOD": env_with_choices(
        "VLLM_WORKER_MULTIPROC_METHOD", "fork", ["spawn", "fork"]
    ),
689
    # Path to the cache for storing downloaded assets
690
    "VLLM_ASSETS_CACHE": lambda: os.path.expanduser(
691
692
693
        os.getenv(
            "VLLM_ASSETS_CACHE",
            os.path.join(get_default_cache_root(), "vllm", "assets"),
694
695
        )
    ),
696
697
    # If the env var is set, we will clean model file in
    # this path $VLLM_ASSETS_CACHE/model_streamer/$model_name
698
699
700
    "VLLM_ASSETS_CACHE_MODEL_CLEAN": lambda: bool(
        int(os.getenv("VLLM_ASSETS_CACHE_MODEL_CLEAN", "0"))
    ),
701
702
    # Timeout for fetching images when serving multimodal models
    # Default is 5 seconds
703
    "VLLM_IMAGE_FETCH_TIMEOUT": lambda: int(os.getenv("VLLM_IMAGE_FETCH_TIMEOUT", "5")),
704
    # Timeout for fetching videos when serving multimodal models
705
    # Default is 30 seconds
706
707
708
    "VLLM_VIDEO_FETCH_TIMEOUT": lambda: int(
        os.getenv("VLLM_VIDEO_FETCH_TIMEOUT", "30")
    ),
709
    # Timeout for fetching audio when serving multimodal models
710
    # Default is 10 seconds
711
712
713
    "VLLM_AUDIO_FETCH_TIMEOUT": lambda: int(
        os.getenv("VLLM_AUDIO_FETCH_TIMEOUT", "10")
    ),
714
715
    # Whether to allow HTTP redirects when fetching from media URLs.
    # Default to True
716
717
718
    "VLLM_MEDIA_URL_ALLOW_REDIRECTS": lambda: bool(
        int(os.getenv("VLLM_MEDIA_URL_ALLOW_REDIRECTS", "1"))
    ),
719
720
721
    # Max number of workers for the thread pool handling
    # media bytes loading. Set to 1 to disable parallel processing.
    # Default is 8
722
723
724
    "VLLM_MEDIA_LOADING_THREAD_COUNT": lambda: int(
        os.getenv("VLLM_MEDIA_LOADING_THREAD_COUNT", "8")
    ),
725
726
727
    # 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
728
729
730
    "VLLM_MAX_AUDIO_CLIP_FILESIZE_MB": lambda: int(
        os.getenv("VLLM_MAX_AUDIO_CLIP_FILESIZE_MB", "25")
    ),
731
732
733
734
735
736
737
    # Backend for Video IO
    # - "opencv": Default backend that uses OpenCV stream buffered backend.
    #
    # Custom backend implementations can be registered
    # via `@VIDEO_LOADER_REGISTRY.register("my_custom_video_loader")` and
    # imported at runtime.
    # If a non-existing backend is used, an AssertionError will be thrown.
738
739
740
    "VLLM_VIDEO_LOADER_BACKEND": lambda: os.getenv(
        "VLLM_VIDEO_LOADER_BACKEND", "opencv"
    ),
741
742
743
744
745
746
747
748
    # 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"),
749
    # [DEPRECATED] Cache size (in GiB per process) for multimodal input cache
750
    # Default is 4 GiB per API process + 4 GiB per engine core process
751
    "VLLM_MM_INPUT_CACHE_GIB": lambda: int(os.getenv("VLLM_MM_INPUT_CACHE_GIB", "4")),
752
753
    # Path to the XLA persistent cache directory.
    # Only used for XLA devices such as TPUs.
754
    "VLLM_XLA_CACHE_PATH": lambda: os.path.expanduser(
755
        os.getenv(
756
            "VLLM_XLA_CACHE_PATH",
757
            os.path.join(get_default_cache_root(), "vllm", "xla_cache"),
758
759
        )
    ),
760
    # If set, assert on XLA recompilation after each execution step.
761
762
763
    "VLLM_XLA_CHECK_RECOMPILATION": lambda: bool(
        int(os.getenv("VLLM_XLA_CHECK_RECOMPILATION", "0"))
    ),
764
    # Enable SPMD mode for TPU backend.
765
766
767
768
    "VLLM_XLA_USE_SPMD": lambda: bool(int(os.getenv("VLLM_XLA_USE_SPMD", "0"))),
    "VLLM_FUSED_MOE_CHUNK_SIZE": lambda: int(
        os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", "32768")
    ),
769
770
771
    # Control whether to use fused MoE activation chunking. Current chunking
    # logic is incompatible with torch.compile and causes IMA. See issue
    # https://github.com/vllm-project/vllm/issues/19631.
772
773
774
    "VLLM_ENABLE_FUSED_MOE_ACTIVATION_CHUNKING": lambda: bool(
        int(os.getenv("VLLM_ENABLE_FUSED_MOE_ACTIVATION_CHUNKING", "1"))
    ),
775
776
    # If set, the OpenAI API server will stay alive even after the underlying
    # AsyncLLMEngine errors and stops serving requests
777
    "VLLM_KEEP_ALIVE_ON_ENGINE_DEATH": lambda: bool(
778
        int(os.getenv("VLLM_KEEP_ALIVE_ON_ENGINE_DEATH", "0"))
779
    ),
780
781
782
783
    # 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.
784
785
786
787
    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": lambda: (
        os.environ.get("VLLM_ALLOW_LONG_MAX_MODEL_LEN", "0").strip().lower()
        in ("1", "true")
    ),
788
789
    # If set, forces FP8 Marlin to be used for FP8 quantization regardless
    # of the hardware support for FP8 compute.
790
791
792
793
794
795
796
    "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"
    ),
797
798
    # Time in ms for the zmq client to wait for a response from the backend
    # server for simple data operations
799
    "VLLM_RPC_TIMEOUT": lambda: int(os.getenv("VLLM_RPC_TIMEOUT", "10000")),
800
    # Timeout in seconds for keeping HTTP connections alive in API server
801
802
803
    "VLLM_HTTP_TIMEOUT_KEEP_ALIVE": lambda: int(
        os.environ.get("VLLM_HTTP_TIMEOUT_KEEP_ALIVE", "5")
    ),
804
805
806
    # 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
807
808
809
    "VLLM_PLUGINS": lambda: None
    if "VLLM_PLUGINS" not in os.environ
    else os.environ["VLLM_PLUGINS"].split(","),
810
811
812
    # a local directory to look in for unrecognized LoRA adapters.
    # only works if plugins are enabled and
    # VLLM_ALLOW_RUNTIME_LORA_UPDATING is enabled.
813
814
815
    "VLLM_LORA_RESOLVER_CACHE_DIR": lambda: os.getenv(
        "VLLM_LORA_RESOLVER_CACHE_DIR", None
    ),
816
817
818
819
820
    # Enables torch CUDA profiling if set.
    # On NVIDIA GPUs, this will start/stop cudaProfilerApi when triggered.
    "VLLM_TORCH_CUDA_PROFILE": lambda: bool(
        os.getenv("VLLM_TORCH_CUDA_PROFILE", "0") != "0"
    ),
821
822
823
824
    # Enables torch profiler if set.
    # Both AsyncLLM's CPU traces as well as workers'
    # traces (CPU & GPU) will be saved under this directory.
    # Note that it must be an absolute path.
825
826
    "VLLM_TORCH_PROFILER_DIR": lambda: (
        None
827
828
829
830
831
        if (val := os.getenv("VLLM_TORCH_PROFILER_DIR")) is None
        else (
            val
            if val.startswith("gs://") and val[5:] and val[5] != "/"
            else os.path.abspath(os.path.expanduser(val))
832
833
        )
    ),
834
835
836
    # Enable torch profiler to record shapes if set
    # VLLM_TORCH_PROFILER_RECORD_SHAPES=1. If not set, torch profiler will
    # not record shapes.
837
838
839
    "VLLM_TORCH_PROFILER_RECORD_SHAPES": lambda: bool(
        os.getenv("VLLM_TORCH_PROFILER_RECORD_SHAPES", "0") != "0"
    ),
840
841
842
    # Enable torch profiler to profile memory if set
    # VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY=1. If not set, torch profiler
    # will not profile memory.
843
844
845
    "VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY": lambda: bool(
        os.getenv("VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY", "0") != "0"
    ),
846
847
848
    # Enable torch profiler to profile stack if set
    # VLLM_TORCH_PROFILER_WITH_STACK=1. If not set, torch profiler WILL
    # profile stack by default.
849
850
851
    "VLLM_TORCH_PROFILER_WITH_STACK": lambda: bool(
        os.getenv("VLLM_TORCH_PROFILER_WITH_STACK", "1") != "0"
    ),
852
853
854
    # Enable torch profiler to profile flops if set
    # VLLM_TORCH_PROFILER_WITH_FLOPS=1. If not set, torch profiler will
    # not profile flops.
855
856
857
    "VLLM_TORCH_PROFILER_WITH_FLOPS": lambda: bool(
        os.getenv("VLLM_TORCH_PROFILER_WITH_FLOPS", "0") != "0"
    ),
858
    # If set, vLLM will use Triton implementations of AWQ.
859
    "VLLM_USE_TRITON_AWQ": lambda: bool(int(os.getenv("VLLM_USE_TRITON_AWQ", "0"))),
860
    # If set, allow loading or unloading lora adapters in runtime,
861
862
863
864
    "VLLM_ALLOW_RUNTIME_LORA_UPDATING": lambda: (
        os.environ.get("VLLM_ALLOW_RUNTIME_LORA_UPDATING", "0").strip().lower()
        in ("1", "true")
    ),
865
866
867
868
869
870
    # 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
871
    "VLLM_SKIP_P2P_CHECK": lambda: os.getenv("VLLM_SKIP_P2P_CHECK", "1") == "1",
872
873
874
875
    # List of quantization kernels that should be disabled, used for testing
    # and performance comparisons. Currently only affects MPLinearKernel
    # selection
    # (kernels: MacheteLinearKernel, MarlinLinearKernel, ExllamaLinearKernel)
876
877
878
    "VLLM_DISABLED_KERNELS": lambda: []
    if "VLLM_DISABLED_KERNELS" not in os.environ
    else os.environ["VLLM_DISABLED_KERNELS"].split(","),
879
    # Disable pynccl (using torch.distributed instead)
880
881
882
    "VLLM_DISABLE_PYNCCL": lambda: (
        os.getenv("VLLM_DISABLE_PYNCCL", "False").lower() in ("true", "1")
    ),
883
884
    # Disable aiter ops unless specifically enabled.
    # Acts as a parent switch to enable the rest of the other operations.
885
886
887
    "VLLM_ROCM_USE_AITER": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER", "False").lower() in ("true", "1")
    ),
888
889
    # Whether to use aiter paged attention.
    # By default is disabled.
890
891
892
    "VLLM_ROCM_USE_AITER_PAGED_ATTN": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_PAGED_ATTN", "False").lower() in ("true", "1")
    ),
893
894
895
    # use aiter linear op if aiter ops are enabled
    # The following list of related ops
    # - scaled_mm (per-tensor / rowwise)
896
897
898
    "VLLM_ROCM_USE_AITER_LINEAR": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_LINEAR", "True").lower() in ("true", "1")
    ),
899
900
    # Whether to use aiter moe ops.
    # By default is enabled.
901
902
903
    "VLLM_ROCM_USE_AITER_MOE": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_MOE", "True").lower() in ("true", "1")
    ),
904
    # use aiter rms norm op if aiter ops are enabled.
905
906
907
    "VLLM_ROCM_USE_AITER_RMSNORM": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_RMSNORM", "True").lower() in ("true", "1")
    ),
908
909
    # Whether to use aiter mla ops.
    # By default is enabled.
910
911
912
    "VLLM_ROCM_USE_AITER_MLA": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_MLA", "True").lower() in ("true", "1")
    ),
913
914
    # Whether to use aiter mha ops.
    # By default is enabled.
915
916
917
    "VLLM_ROCM_USE_AITER_MHA": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_MHA", "True").lower() in ("true", "1")
    ),
918
919
    # Whether to use aiter fp4 gemm asm.
    # By default is disabled.
920
921
922
    "VLLM_ROCM_USE_AITER_FP4_ASM_GEMM": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_FP4_ASM_GEMM", "False").lower() in ("true", "1")
    ),
923
924
    # Whether to use aiter rope.
    # By default is disabled.
925
926
    "VLLM_ROCM_USE_AITER_TRITON_ROPE": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_TRITON_ROPE", "False").lower() in ("true", "1")
927
    ),
928
929
    # Whether to use aiter triton fp8 bmm kernel
    # By default is enabled.
930
931
932
    "VLLM_ROCM_USE_AITER_FP8BMM": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_FP8BMM", "True").lower() in ("true", "1")
    ),
933
934
935
936
937
    # 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")
    ),
938
939
940
941
942
943
    # Whether to use aiter fusion shared experts ops.
    # By default is enabled.
    "VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS", "True").lower()
        in ("true", "1")
    ),
944
945
946
947
948
    # 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")
    ),
949
    # use rocm skinny gemms
950
951
952
    "VLLM_ROCM_USE_SKINNY_GEMM": lambda: (
        os.getenv("VLLM_ROCM_USE_SKINNY_GEMM", "True").lower() in ("true", "1")
    ),
953
    # Pad the fp8 weights to 256 bytes for ROCm
954
    "VLLM_ROCM_FP8_PADDING": lambda: bool(int(os.getenv("VLLM_ROCM_FP8_PADDING", "1"))),
955
    # Pad the weights for the moe kernel
956
    "VLLM_ROCM_MOE_PADDING": lambda: bool(int(os.getenv("VLLM_ROCM_MOE_PADDING", "1"))),
957
    # custom paged attention kernel for MI3* cards
958
959
960
    "VLLM_ROCM_CUSTOM_PAGED_ATTN": lambda: (
        os.getenv("VLLM_ROCM_CUSTOM_PAGED_ATTN", "True").lower() in ("true", "1")
    ),
961
962
963
    # Custom quick allreduce kernel for MI3* cards
    # Choice of quantization level: FP, INT8, INT6, INT4 or NONE
    # Recommended for large models to get allreduce
964
965
966
967
968
    "VLLM_ROCM_QUICK_REDUCE_QUANTIZATION": env_with_choices(
        "VLLM_ROCM_QUICK_REDUCE_QUANTIZATION",
        "NONE",
        ["FP", "INT8", "INT6", "INT4", "NONE"],
    ),
969
970
971
972
    # 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
973
974
975
976
    "VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16": lambda: (
        os.getenv("VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16", "True").lower()
        in ("true", "1")
    ),
977
978
979
980
981
982
    # 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.
983
984
985
    "VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB": lambda: maybe_convert_int(
        os.environ.get("VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB", None)
    ),
986
    # Divisor for dynamic query scale factor calculation for FP8 KV Cache
987
    "Q_SCALE_CONSTANT": lambda: int(os.getenv("Q_SCALE_CONSTANT", "200")),
988
    # Divisor for dynamic key scale factor calculation for FP8 KV Cache
989
    "K_SCALE_CONSTANT": lambda: int(os.getenv("K_SCALE_CONSTANT", "200")),
990
    # Divisor for dynamic value scale factor calculation for FP8 KV Cache
991
    "V_SCALE_CONSTANT": lambda: int(os.getenv("V_SCALE_CONSTANT", "100")),
992
    # If set, enable multiprocessing in LLM for the V1 code path.
993
994
995
996
997
998
    "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")
    ),
999
    "VLLM_DISABLE_COMPILE_CACHE": disable_compile_cache,
1000
1001
1002
    # If set, vllm will run in development mode, which will enable
    # some additional endpoints for developing and debugging,
    # e.g. `/reset_prefix_cache`
1003
    "VLLM_SERVER_DEV_MODE": lambda: bool(int(os.getenv("VLLM_SERVER_DEV_MODE", "0"))),
1004
1005
1006
1007
1008
1009
1010
    # 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.
1011
1012
1013
    "VLLM_V1_OUTPUT_PROC_CHUNK_SIZE": lambda: int(
        os.getenv("VLLM_V1_OUTPUT_PROC_CHUNK_SIZE", "128")
    ),
1014
    # If set, vLLM will disable the MLA attention optimizations.
1015
    "VLLM_MLA_DISABLE": lambda: bool(int(os.getenv("VLLM_MLA_DISABLE", "0"))),
1016
1017
    # If set, vLLM will pick up the provided Flash Attention MLA
    # max number splits for cuda graph decode
1018
1019
1020
    "VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH": lambda: int(
        os.getenv("VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH", "32")
    ),
1021
1022
1023
    # 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.
1024
1025
1026
    "VLLM_RAY_PER_WORKER_GPUS": lambda: float(
        os.getenv("VLLM_RAY_PER_WORKER_GPUS", "1.0")
    ),
1027
1028
1029
    # 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"
1030
    "VLLM_RAY_BUNDLE_INDICES": lambda: os.getenv("VLLM_RAY_BUNDLE_INDICES", ""),
1031
1032
    # In some system, find_loaded_library() may not work. So we allow users to
    # specify the path through environment variable VLLM_CUDART_SO_PATH.
1033
    "VLLM_CUDART_SO_PATH": lambda: os.getenv("VLLM_CUDART_SO_PATH", None),
1034
    # Rank of the process in the data parallel setting
1035
    "VLLM_DP_RANK": lambda: int(os.getenv("VLLM_DP_RANK", "0")),
1036
1037
    # Rank of the process in the data parallel setting.
    # Defaults to VLLM_DP_RANK when not set.
1038
1039
1040
    "VLLM_DP_RANK_LOCAL": lambda: int(
        os.getenv("VLLM_DP_RANK_LOCAL", sys.modules[__name__].VLLM_DP_RANK)
    ),
1041
    # World size of the data parallel setting
1042
    "VLLM_DP_SIZE": lambda: int(os.getenv("VLLM_DP_SIZE", "1")),
1043
    # IP address of the master node in the data parallel setting
1044
    "VLLM_DP_MASTER_IP": lambda: os.getenv("VLLM_DP_MASTER_IP", "127.0.0.1"),
1045
    # Port of the master node in the data parallel setting
1046
    "VLLM_DP_MASTER_PORT": lambda: int(os.getenv("VLLM_DP_MASTER_PORT", "0")),
1047
1048
1049
1050
1051
    # 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.
1052
    "VLLM_MOE_DP_CHUNK_SIZE": lambda: int(os.getenv("VLLM_MOE_DP_CHUNK_SIZE", "256")),
1053
    # Randomize inputs during dummy runs when using Data Parallel
1054
1055
1056
1057
    "VLLM_RANDOMIZE_DP_DUMMY_INPUTS": lambda: os.environ.get(
        "VLLM_RANDOMIZE_DP_DUMMY_INPUTS", "0"
    )
    == "1",
1058
1059
1060
1061
1062
1063
1064
    # 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;
1065
1066
1067
    # - "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;
1068
1069
1070
1071
    # 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"
    ),
1072
    # Whether to use S3 path for model loading in CI via RunAI Streamer
1073
    "VLLM_CI_USE_S3": lambda: os.environ.get("VLLM_CI_USE_S3", "0") == "1",
1074
    # Use model_redirect to redirect the model name to a local folder.
1075
1076
1077
1078
1079
    # `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
1080
1081
1082
    "VLLM_MODEL_REDIRECT_PATH": lambda: os.environ.get(
        "VLLM_MODEL_REDIRECT_PATH", None
    ),
1083
    # Whether to use atomicAdd reduce in gptq/awq marlin kernel.
1084
1085
1086
1087
    "VLLM_MARLIN_USE_ATOMIC_ADD": lambda: os.environ.get(
        "VLLM_MARLIN_USE_ATOMIC_ADD", "0"
    )
    == "1",
1088
    # Whether to use marlin kernel in mxfp4 quantization method
1089
1090
1091
    "VLLM_MXFP4_USE_MARLIN": lambda: maybe_convert_bool(
        os.environ.get("VLLM_MXFP4_USE_MARLIN", None)
    ),
1092
1093
1094
    # 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.
1095
1096
1097
1098
    "VLLM_V1_USE_OUTLINES_CACHE": lambda: os.environ.get(
        "VLLM_V1_USE_OUTLINES_CACHE", "0"
    )
    == "1",
1099
1100
    # Gap between padding buckets for the forward pass. So we have
    # 8, we will run forward pass with [16, 24, 32, ...].
1101
1102
1103
1104
1105
1106
1107
1108
    "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)
    ),
1109
    # Whether using Pathways
1110
1111
1112
    "VLLM_TPU_USING_PATHWAYS": lambda: bool(
        "proxy" in os.getenv("JAX_PLATFORMS", "").lower()
    ),
1113
    # Allow use of DeepGemm kernels for fused moe ops.
1114
    "VLLM_USE_DEEP_GEMM": lambda: bool(int(os.getenv("VLLM_USE_DEEP_GEMM", "1"))),
1115
1116
1117
1118
    # 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"))
    ),
1119
    # Whether to use E8M0 scaling when DeepGEMM is used on Blackwell GPUs.
1120
1121
1122
    "VLLM_USE_DEEP_GEMM_E8M0": lambda: bool(
        int(os.getenv("VLLM_USE_DEEP_GEMM_E8M0", "1"))
    ),
1123
1124
1125
1126
    # 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.
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
    # 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",
        ],
1142
    ),
1143
    # Whether to use fused grouped_topk used for MoE expert selection.
1144
1145
1146
    "VLLM_USE_FUSED_MOE_GROUPED_TOPK": lambda: bool(
        int(os.getenv("VLLM_USE_FUSED_MOE_GROUPED_TOPK", "1"))
    ),
1147
    # Allow use of FlashInfer MoE kernels for fused moe ops.
1148
1149
1150
    "VLLM_USE_FLASHINFER_MOE_FP16": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_FP16", "0"))
    ),
1151
    # Allow use of FlashInfer MoE kernels for fused moe ops.
1152
1153
1154
    "VLLM_USE_FLASHINFER_MOE_FP8": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_FP8", "0"))
    ),
1155
    # Allow use of FlashInfer CUTLASS kernels for fused moe ops.
1156
1157
1158
    "VLLM_USE_FLASHINFER_MOE_FP4": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_FP4", "0"))
    ),
1159
1160
    # If set to 1, use the FlashInfer
    # MXFP8 (activation) x MXFP4 (weight) MoE backend.
1161
1162
1163
    "VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8", "0"))
    ),
1164
1165
1166
1167
    # 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.
1168
1169
1170
    "VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS", "0"))
    ),
1171
1172
    # If set to 1, use the FlashInfer
    # BF16 (activation) x MXFP4 (weight) MoE backend.
1173
1174
1175
    "VLLM_USE_FLASHINFER_MOE_MXFP4_BF16": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_MXFP4_BF16", "0"))
    ),
1176
1177
1178
    # 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.
1179
    "VLLM_XGRAMMAR_CACHE_MB": lambda: int(os.getenv("VLLM_XGRAMMAR_CACHE_MB", "512")),
1180
1181
1182
1183
1184
1185
1186
    # 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.
1187
1188
1189
    "VLLM_MSGPACK_ZERO_COPY_THRESHOLD": lambda: int(
        os.getenv("VLLM_MSGPACK_ZERO_COPY_THRESHOLD", "256")
    ),
1190
1191
1192
    # 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.
1193
1194
1195
    "VLLM_ALLOW_INSECURE_SERIALIZATION": lambda: bool(
        int(os.getenv("VLLM_ALLOW_INSECURE_SERIALIZATION", "0"))
    ),
Robert Shaw's avatar
Robert Shaw committed
1196
    # IP address used for NIXL handshake between remote agents.
1197
1198
1199
    "VLLM_NIXL_SIDE_CHANNEL_HOST": lambda: os.getenv(
        "VLLM_NIXL_SIDE_CHANNEL_HOST", "localhost"
    ),
Robert Shaw's avatar
Robert Shaw committed
1200
    # Port used for NIXL handshake between remote agents.
1201
1202
1203
    "VLLM_NIXL_SIDE_CHANNEL_PORT": lambda: int(
        os.getenv("VLLM_NIXL_SIDE_CHANNEL_PORT", "5600")
    ),
1204
    # all2all backend for vllm's expert parallel communication
1205
    # Available options:
1206
1207
1208
    # - "naive": naive all2all implementation using broadcasts
    # - "allgather_reducescatter": all2all implementation based on allgather and
    #  reducescatter
1209
    # - "pplx": use pplx kernels
1210
1211
    # - "deepep_high_throughput", use deepep high-throughput kernels
    # - "deepep_low_latency", use deepep low-latency kernels
1212
    # - "flashinfer_all2allv", use flashinfer alltoallv kernels for mnnvl
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
    "VLLM_ALL2ALL_BACKEND": env_with_choices(
        "VLLM_ALL2ALL_BACKEND",
        "allgather_reducescatter",
        [
            "naive",
            "pplx",
            "deepep_high_throughput",
            "deepep_low_latency",
            "allgather_reducescatter",
            "flashinfer_all2allv",
        ],
    ),
1225
1226
    # Flashinfer MoE backend for vLLM's fused Mixture-of-Experts support.
    # Both require compute capability 10.0 or above.
1227
1228
1229
1230
1231
    # Available options:
    # - "throughput":  [default]
    #     Uses CUTLASS kernels optimized for high-throughput batch inference.
    # - "latency":
    #     Uses TensorRT-LLM kernels optimized for low-latency inference.
1232
    "VLLM_FLASHINFER_MOE_BACKEND": env_with_choices(
1233
        "VLLM_FLASHINFER_MOE_BACKEND", "latency", ["throughput", "latency"]
1234
    ),
1235
1236
1237
1238
    # 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))
    ),
1239
1240
1241
1242
    # 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.
1243
1244
1245
    "VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE": lambda: int(
        os.getenv("VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE", "163840")
    ),
1246
1247
1248
1249
    # 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> }
1250
    # Unspecified world sizes will fall back to
1251
    #     { 2: 64, 4: 1, <everything else>: 0.5 }
1252
1253
1254
    "VLLM_FLASHINFER_ALLREDUCE_FUSION_THRESHOLDS_MB": lambda: json.loads(
        os.getenv("VLLM_FLASHINFER_ALLREDUCE_FUSION_THRESHOLDS_MB", "{}")
    ),
1255
1256
1257
1258
1259
1260
    # MoE routing strategy selector.
    # See `RoutingSimulator.get_available_strategies()` # for available
    # strategies.
    # Cutstom routing strategies can be registered by
    # RoutingSimulator.register_strategy()
    # Note: custom strategies may not produce correct model outputs
1261
1262
1263
    "VLLM_MOE_ROUTING_SIMULATION_STRATEGY": lambda: os.environ.get(
        "VLLM_MOE_ROUTING_SIMULATION_STRATEGY", ""
    ).lower(),
1264
    # Regex timeout for use by the vLLM tool parsing plugins.
1265
1266
1267
    "VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS": lambda: int(
        os.getenv("VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS", "1")
    ),
1268
1269
    # Reduce CPU usage when vLLM is idle. Enabling this will incur small
    # latency penalty when a request eventually comes.
1270
    "VLLM_SLEEP_WHEN_IDLE": lambda: bool(int(os.getenv("VLLM_SLEEP_WHEN_IDLE", "0"))),
1271
1272
1273
    # 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.
1274
1275
1276
    "VLLM_MQ_MAX_CHUNK_BYTES_MB": lambda: int(
        os.getenv("VLLM_MQ_MAX_CHUNK_BYTES_MB", "16")
    ),
1277
1278
    # Timeout in seconds for execute_model RPC calls in multiprocessing
    # executor (only applies when TP > 1).
1279
1280
1281
    "VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS": lambda: int(
        os.getenv("VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS", "300")
    ),
1282
1283
1284
1285
1286
1287
1288
    # 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.
1289
1290
1291
    "VLLM_KV_CACHE_LAYOUT": env_with_choices(
        "VLLM_KV_CACHE_LAYOUT", None, ["NHD", "HND"]
    ),
1292
1293
1294
    # 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.
1295
1296
1297
    "VLLM_COMPUTE_NANS_IN_LOGITS": lambda: bool(
        int(os.getenv("VLLM_COMPUTE_NANS_IN_LOGITS", "0"))
    ),
1298
1299
1300
    # Controls whether or not emulations are used for NVFP4
    # generations on machines < 100 for compressed-tensors
    # models
1301
1302
1303
    "VLLM_USE_NVFP4_CT_EMULATIONS": lambda: bool(
        int(os.getenv("VLLM_USE_NVFP4_CT_EMULATIONS", "0"))
    ),
1304
1305
1306
1307
    # 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.
1308
1309
1310
    "VLLM_NIXL_ABORT_REQUEST_TIMEOUT": lambda: int(
        os.getenv("VLLM_NIXL_ABORT_REQUEST_TIMEOUT", "480")
    ),
1311
    # Controls whether or not to use cudnn prefill
1312
1313
1314
    "VLLM_USE_CUDNN_PREFILL": lambda: bool(
        int(os.getenv("VLLM_USE_CUDNN_PREFILL", "0"))
    ),
1315
1316
1317
1318
    # Controls whether to use TRT-LLM ragged DeepSeek prefill
    "VLLM_USE_TRTLLM_RAGGED_DEEPSEEK_PREFILL": lambda: bool(
        int(os.getenv("VLLM_USE_TRTLLM_RAGGED_DEEPSEEK_PREFILL", "0"))
    ),
1319
1320
1321
    # If set to 1/True, use the TRTLLM attention backend in flashinfer.
    # If set to 0/False, use the default attention backend in flashinfer.
    # If not set, auto-detect the attention backend in flashinfer.
1322
1323
1324
1325
1326
    "VLLM_USE_TRTLLM_ATTENTION": lambda: (
        None
        if "VLLM_USE_TRTLLM_ATTENTION" not in os.environ
        else os.environ["VLLM_USE_TRTLLM_ATTENTION"].lower() in ("1", "true")
    ),
1327
    # If set to 1, when we use fp8 kv, we do not quantize Q to fp8
1328
1329
1330
    "VLLM_FLASHINFER_DISABLE_Q_QUANTIZATION": lambda: bool(
        int(os.getenv("VLLM_FLASHINFER_DISABLE_Q_QUANTIZATION", "0"))
    ),
1331
1332
    # If set, it means we pre-downloaded cubin files and flashinfer will
    # read the cubin files directly.
1333
1334
1335
    "VLLM_HAS_FLASHINFER_CUBIN": lambda: bool(
        int(os.getenv("VLLM_HAS_FLASHINFER_CUBIN", "0"))
    ),
1336
1337
1338
1339
1340
1341
1342
1343
    # Supported options:
    # - "flashinfer-cudnn": use flashinfer cudnn GEMM backend
    # - "flashinfer-trtllm": use flashinfer trtllm GEMM backend
    # - "flashinfer-cutlass": use flashinfer cutlass GEMM backend
    # - <none>: automatically pick an available backend
    "VLLM_NVFP4_GEMM_BACKEND": env_with_choices(
        "VLLM_NVFP4_GEMM_BACKEND",
        None,
1344
        ["flashinfer-cudnn", "flashinfer-trtllm", "flashinfer-cutlass", "cutlass"],
1345
    ),
1346
1347
1348
    # 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.
1349
1350
1351
    "VLLM_ENABLE_CUDAGRAPH_GC": lambda: bool(
        int(os.getenv("VLLM_ENABLE_CUDAGRAPH_GC", "0"))
    ),
1352
    # Used to force set up loopback IP
1353
    "VLLM_LOOPBACK_IP": lambda: os.getenv("VLLM_LOOPBACK_IP", ""),
1354
1355
1356
    # Used to set the process name prefix for vLLM processes.
    # This is useful for debugging and monitoring purposes.
    # The default value is "VLLM".
1357
    "VLLM_PROCESS_NAME_PREFIX": lambda: os.getenv("VLLM_PROCESS_NAME_PREFIX", "VLLM"),
1358
1359
1360
1361
1362
1363
1364
    # 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.
1365
1366
1367
    "VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE": lambda: bool(
        int(os.getenv("VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE", "0"))
    ),
1368
1369
    # 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
1370
1371
    # messages for those requests in memory. By default, this is disabled (0),
    # and the "store" option is ignored.
1372
1373
1374
1375
1376
    # 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.
1377
1378
1379
    "VLLM_ENABLE_RESPONSES_API_STORE": lambda: bool(
        int(os.getenv("VLLM_ENABLE_RESPONSES_API_STORE", "0"))
    ),
xiao-llm's avatar
xiao-llm committed
1380
    # If set, use the fp8 mfma in rocm paged attention.
1381
1382
1383
    "VLLM_ROCM_FP8_MFMA_PAGE_ATTN": lambda: bool(
        int(os.getenv("VLLM_ROCM_FP8_MFMA_PAGE_ATTN", "0"))
    ),
1384
    # Whether to use pytorch symmetric memory for allreduce
1385
    "VLLM_ALLREDUCE_USE_SYMM_MEM": lambda: bool(
1386
        int(os.getenv("VLLM_ALLREDUCE_USE_SYMM_MEM", "1"))
1387
    ),
1388
    # Allows vllm to find tuned config under customized folder
1389
    "VLLM_TUNED_CONFIG_FOLDER": lambda: os.getenv("VLLM_TUNED_CONFIG_FOLDER", None),
1390
1391
1392
1393
1394
1395
1396
1397
1398
    # 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"],
    ),
1399
    # Allows harmony instructions to be injected on system messages
1400
1401
1402
    "VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS": lambda: bool(
        int(os.getenv("VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS", "0"))
    ),
1403
1404
1405
1406
1407
1408
    # 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"))
    ),
1409
    # Add optional custom scopes for profiling, disable to avoid overheads
1410
1411
1412
    "VLLM_CUSTOM_SCOPES_FOR_PROFILING": lambda: bool(
        int(os.getenv("VLLM_CUSTOM_SCOPES_FOR_PROFILING", "0"))
    ),
1413
    # Add optional nvtx scopes for profiling, disable to avoid overheads
1414
1415
1416
    "VLLM_NVTX_SCOPES_FOR_PROFILING": lambda: bool(
        int(os.getenv("VLLM_NVTX_SCOPES_FOR_PROFILING", "0"))
    ),
1417
1418
    # Represent block hashes in KV cache events as 64-bit integers instead of
    # raw bytes. Defaults to True for backward compatibility.
1419
1420
1421
    "VLLM_KV_EVENTS_USE_INT_BLOCK_HASHES": lambda: bool(
        int(os.getenv("VLLM_KV_EVENTS_USE_INT_BLOCK_HASHES", "1"))
    ),
1422
1423
    # Name of the shared memory buffer used for object storage.
    # Only effective when mm_config.mm_processor_cache_type == "shm".
1424
1425
1426
    "VLLM_OBJECT_STORAGE_SHM_BUFFER_NAME": lambda: os.getenv(
        "VLLM_OBJECT_STORAGE_SHM_BUFFER_NAME", "VLLM_OBJECT_STORAGE_SHM_BUFFER"
    ),
1427
    # The size in MB of the buffers (NVL and RDMA) used by DeepEP
1428
1429
1430
    "VLLM_DEEPEP_BUFFER_SIZE_MB": lambda: int(
        os.getenv("VLLM_DEEPEP_BUFFER_SIZE_MB", "1024")
    ),
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
    # Force DeepEP to use intranode kernel for inter-node communication in
    # high throughput mode. This is useful archive higher prefill throuhgput
    # 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"))
    ),
1442
1443
    # 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
1444
    "VLLM_DBO_COMM_SMS": lambda: int(os.getenv("VLLM_DBO_COMM_SMS", "20")),
1445
1446
1447
    # 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)
1448
1449
1450
    "VLLM_ENABLE_INDUCTOR_MAX_AUTOTUNE": lambda: bool(
        int(os.getenv("VLLM_ENABLE_INDUCTOR_MAX_AUTOTUNE", "1"))
    ),
1451
1452
    # If set to 1, enable coordinate_descent_tuning;
    # By default, this is enabled (1)
1453
1454
1455
    "VLLM_ENABLE_INDUCTOR_COORDINATE_DESCENT_TUNING": lambda: bool(
        int(os.getenv("VLLM_ENABLE_INDUCTOR_COORDINATE_DESCENT_TUNING", "1"))
    ),
1456
    # Flag to enable NCCL symmetric memory allocation and registration
1457
1458
1459
    "VLLM_USE_NCCL_SYMM_MEM": lambda: bool(
        int(os.getenv("VLLM_USE_NCCL_SYMM_MEM", "0"))
    ),
1460
    # NCCL header path
1461
    "VLLM_NCCL_INCLUDE_PATH": lambda: os.environ.get("VLLM_NCCL_INCLUDE_PATH", None),
1462
1463
    # Flag to enable FBGemm kernels on model execution
    "VLLM_USE_FBGEMM": lambda: bool(int(os.getenv("VLLM_USE_FBGEMM", "0"))),
1464
1465
1466
1467
1468
1469
    # 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", ""),
1470
    # Disables parallel execution of shared_experts via separate cuda stream
1471
1472
    "VLLM_DISABLE_SHARED_EXPERTS_STREAM": lambda: bool(
        int(os.getenv("VLLM_DISABLE_SHARED_EXPERTS_STREAM", "0"))
1473
    ),
1474
1475
1476
1477
1478
1479
1480
1481
1482
    # 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"]
    ),
1483
    # Flag to enable FlatLogprobs whose GC overhead is significantly smaller than
1484
1485
    # the original list[dict[int, Logprob]] approach.
    # After enabled, PromptLogprobs and SampleLogprobs would populated as
1486
1487
    # FlatLogprobs.
    "VLLM_FLAT_LOGPROBS": lambda: bool(int(os.getenv("VLLM_FLAT_LOGPROBS", "0"))),
1488
1489
}

1490
# --8<-- [end:env-vars-definition]
1491

1492

1493
def __getattr__(name: str):
1494
1495
1496
1497
1498
1499
    """
    Gets environment variables lazily.

    NOTE: After enable_envs_cache() invocation (which triggered after service
    initialization), all environment variables will be cached.
    """
1500
1501
1502
1503
1504
    if name in environment_variables:
        return environment_variables[name]()
    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")


1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
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.
    """
    # Tag __getattr__ with functools.cache
    global __getattr__
    __getattr__ = functools.cache(__getattr__)

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


1524
1525
def __dir__():
    return list(environment_variables.keys())
1526
1527
1528
1529
1530
1531
1532
1533
1534


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}")


1535
1536
1537
1538
1539
1540
def compute_hash() -> str:
    """
    WARNING: Whenever a new key is added to this environment
    variables, ensure that it is included in the factors list if
    it affects the computation graph. For example, different values
    of VLLM_PP_LAYER_PARTITION will generate different computation
1541
    graphs, so it is included in the factors list. The env vars that
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
    affect the choice of different kernels or attention backends should
    also be included in the factors list.
    """

    # The values of envs may affects the computation graph.
    # TODO(DefTruth): hash all environment variables?
    # for key in environment_variables:
    #     factorize(key)
    environment_variables_to_hash = [
        "VLLM_PP_LAYER_PARTITION",
        "VLLM_MLA_DISABLE",
1553
        "VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH",
1554
1555
1556
        "VLLM_USE_TRITON_AWQ",
        "VLLM_DP_RANK",
        "VLLM_DP_SIZE",
1557
        "VLLM_USE_STANDALONE_COMPILE",
1558
        "VLLM_FUSED_MOE_CHUNK_SIZE",
1559
1560
1561
1562
1563
1564
        "VLLM_FLASHINFER_MOE_BACKEND",
        "VLLM_V1_USE_PREFILL_DECODE_ATTENTION",
        "VLLM_ATTENTION_BACKEND",
        "VLLM_USE_FLASHINFER_SAMPLER",
        "VLLM_DISABLED_KERNELS",
        "VLLM_USE_DEEP_GEMM",
1565
        "VLLM_MOE_USE_DEEP_GEMM",
1566
        "VLLM_USE_DEEP_GEMM_E8M0",
1567
        "VLLM_USE_FUSED_MOE_GROUPED_TOPK",
1568
        "VLLM_USE_FLASHINFER_MOE_FP16",
1569
1570
1571
        "VLLM_USE_FLASHINFER_MOE_FP8",
        "VLLM_USE_FLASHINFER_MOE_FP4",
        "VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8",
1572
        "VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS",
1573
        "VLLM_USE_FLASHINFER_MOE_MXFP4_BF16",
1574
        "VLLM_FLASHINFER_WORKSPACE_BUFFER_SIZE",
1575
        "VLLM_USE_CUDNN_PREFILL",
1576
        "VLLM_USE_TRTLLM_RAGGED_DEEPSEEK_PREFILL",
1577
        "VLLM_USE_TRTLLM_ATTENTION",
1578
        "VLLM_FLASHINFER_DISABLE_Q_QUANTIZATION",
1579
1580
1581
1582
1583
1584
1585
        "VLLM_ROCM_USE_AITER",
        "VLLM_ROCM_USE_AITER_PAGED_ATTN",
        "VLLM_ROCM_USE_AITER_LINEAR",
        "VLLM_ROCM_USE_AITER_MOE",
        "VLLM_ROCM_USE_AITER_RMSNORM",
        "VLLM_ROCM_USE_AITER_MLA",
        "VLLM_ROCM_USE_AITER_MHA",
1586
        "VLLM_ROCM_USE_AITER_FP4_ASM_GEMM",
1587
        "VLLM_ROCM_USE_AITER_TRITON_ROPE",
1588
        "VLLM_ROCM_USE_AITER_FP8BMM",
1589
        "VLLM_ROCM_USE_AITER_UNIFIED_ATTENTION",
1590
        "VLLM_ROCM_USE_AITER_TRITON_GEMM",
1591
1592
1593
1594
1595
1596
1597
        "VLLM_ROCM_USE_SKINNY_GEMM",
        "VLLM_ROCM_FP8_PADDING",
        "VLLM_ROCM_MOE_PADDING",
        "VLLM_ROCM_CUSTOM_PAGED_ATTN",
        "VLLM_ROCM_QUICK_REDUCE_QUANTIZATION",
        "VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16",
        "VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB",
xiao-llm's avatar
xiao-llm committed
1598
        "VLLM_ROCM_FP8_MFMA_PAGE_ATTN",
1599
1600
        "VLLM_ENABLE_INDUCTOR_MAX_AUTOTUNE",
        "VLLM_ENABLE_INDUCTOR_COORDINATE_DESCENT_TUNING",
1601
        "VLLM_NVFP4_GEMM_BACKEND",
1602
        "VLLM_USE_FBGEMM",
1603
1604
        "VLLM_DEEPEP_HIGH_THROUGHPUT_FORCE_INTRA_NODE",
        "VLLM_DEEPEP_LOW_LATENCY_USE_MNNVL",
1605
1606
    ]
    for key in environment_variables_to_hash:
1607
1608
        # if this goes out of sync with environment_variables,
        # it's not a user error, it's a bug
1609
        assert key in environment_variables, (
1610
            "Please update environment_variables_to_hash in envs.py"
1611
        )
1612

1613
    factors = [environment_variables[key]() for key in environment_variables_to_hash]
1614

1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
    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",
    ]
    factors.extend([os.getenv(var) for var in ray_noset_env_vars])

1638
    hash_str = hashlib.md5(str(factors).encode(), usedforsecurity=False).hexdigest()
1639
1640

    return hash_str