envs.py 78.6 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
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_ROCM_SLEEP_MEM_CHUNK_SIZE: int = 256
22
    VLLM_V1_USE_PREFILL_DECODE_ATTENTION: bool = False
23
    VLLM_FLASH_ATTN_VERSION: int | None = None
24
    LOCAL_RANK: int = 0
25
    CUDA_VISIBLE_DEVICES: str | None = None
26
    VLLM_ENGINE_ITERATION_TIMEOUT_S: int = 60
27
    VLLM_ENGINE_READY_TIMEOUT_S: int = 600
28
    VLLM_API_KEY: str | None = None
29
    VLLM_DEBUG_LOG_API_SERVER_RESPONSE: bool = False
30
31
32
33
    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
34
35
    VLLM_CACHE_ROOT: str = os.path.expanduser("~/.cache/vllm")
    VLLM_CONFIG_ROOT: str = os.path.expanduser("~/.config/vllm")
36
37
    VLLM_USAGE_STATS_SERVER: str = "https://stats.vllm.ai"
    VLLM_NO_USAGE_STATS: bool = False
38
    VLLM_DISABLE_FLASHINFER_PREFILL: bool = False
39
40
    VLLM_DO_NOT_TRACK: bool = False
    VLLM_USAGE_SOURCE: str = ""
41
    VLLM_CONFIGURE_LOGGING: bool = True
42
    VLLM_LOGGING_LEVEL: str = "INFO"
43
    VLLM_LOGGING_PREFIX: str = ""
44
    VLLM_LOGGING_STREAM: str = "ext://sys.stdout"
45
    VLLM_LOGGING_CONFIG_PATH: str | None = None
Nick Hill's avatar
Nick Hill committed
46
47
    VLLM_LOGGING_COLOR: str = "auto"
    NO_COLOR: bool = False
48
    VLLM_LOG_STATS_INTERVAL: float = 10.0
49
    VLLM_TRACE_FUNCTION: int = 0
50
51
52
53
    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
54
    VLLM_CPU_OMP_THREADS_BIND: str = ""
55
    VLLM_CPU_NUM_OF_RESERVED_CPU: int | None = None
56
    VLLM_CPU_SGL_KERNEL: bool = False
57
    VLLM_XLA_CACHE_PATH: str = os.path.join(VLLM_CACHE_ROOT, "xla_cache")
58
    VLLM_XLA_CHECK_RECOMPILATION: bool = False
59
    VLLM_FUSED_MOE_CHUNK_SIZE: int = 16 * 1024
60
    VLLM_ENABLE_FUSED_MOE_ACTIVATION_CHUNKING: bool = True
61
    VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE: Literal["auto", "nccl", "shm"] = "auto"
62
    VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM: bool = False
63
    VLLM_USE_RAY_WRAPPED_PP_COMM: bool = True
64
    VLLM_XLA_USE_SPMD: bool = False
65
    VLLM_WORKER_MULTIPROC_METHOD: Literal["fork", "spawn"] = "fork"
66
    VLLM_ASSETS_CACHE: str = os.path.join(VLLM_CACHE_ROOT, "assets")
67
    VLLM_ASSETS_CACHE_MODEL_CLEAN: bool = False
68
    VLLM_IMAGE_FETCH_TIMEOUT: int = 5
69
    VLLM_VIDEO_FETCH_TIMEOUT: int = 30
70
    VLLM_AUDIO_FETCH_TIMEOUT: int = 10
71
    VLLM_MEDIA_URL_ALLOW_REDIRECTS: bool = True
72
    VLLM_MEDIA_LOADING_THREAD_COUNT: int = 8
73
    VLLM_MAX_AUDIO_CLIP_FILESIZE_MB: int = 25
74
    VLLM_VIDEO_LOADER_BACKEND: str = "opencv"
75
    VLLM_MEDIA_CONNECTOR: str = "http"
76
    VLLM_TARGET_DEVICE: str = "cuda"
77
    VLLM_MAIN_CUDA_VERSION: str = "12.9"
78
    VLLM_FLOAT32_MATMUL_PRECISION: Literal["highest", "high", "medium"] = "highest"
79
80
    MAX_JOBS: str | None = None
    NVCC_THREADS: str | None = None
81
    VLLM_USE_PRECOMPILED: bool = False
82
    VLLM_SKIP_PRECOMPILED_VERSION_SUFFIX: bool = False
83
    VLLM_DOCKER_BUILD_CONTEXT: bool = False
84
    VLLM_KEEP_ALIVE_ON_ENGINE_DEATH: bool = False
85
    CMAKE_BUILD_TYPE: Literal["Debug", "Release", "RelWithDebInfo"] | None = None
86
    VERBOSE: bool = False
87
    VLLM_ALLOW_LONG_MAX_MODEL_LEN: bool = False
88
    VLLM_RPC_TIMEOUT: int = 10000  # ms
89
    VLLM_HTTP_TIMEOUT_KEEP_ALIVE: int = 5  # seconds
90
91
    VLLM_PLUGINS: list[str] | None = None
    VLLM_LORA_RESOLVER_CACHE_DIR: str | None = None
92
93
94
    # Deprecated env variables for profiling, kept for backward compatibility
    # See also vllm/config/profiler.py and `--profiler-config` argument
    VLLM_TORCH_CUDA_PROFILE: str | None = None
95
    VLLM_TORCH_PROFILER_DIR: str | None = None
96
97
98
99
100
101
102
103
104
105
    VLLM_TORCH_PROFILER_RECORD_SHAPES: str | None = None
    VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY: str | None = None
    VLLM_TORCH_PROFILER_DISABLE_ASYNC_LLM: str | None = None
    VLLM_TORCH_PROFILER_WITH_STACK: str | None = None
    VLLM_TORCH_PROFILER_WITH_FLOPS: str | None = None
    VLLM_TORCH_PROFILER_USE_GZIP: str | None = None
    VLLM_TORCH_PROFILER_DUMP_CUDA_TIME_TOTAL: str | None = None
    VLLM_PROFILER_DELAY_ITERS: str | None = None
    VLLM_PROFILER_MAX_ITERS: str | None = None
    # End of deprecated env variables for profiling
106
    VLLM_USE_AOT_COMPILE: bool = False
107
    VLLM_USE_BYTECODE_HOOK: bool = False
108
    VLLM_FORCE_AOT_LOAD: bool = False
109
    VLLM_USE_TRITON_AWQ: bool = False
110
    VLLM_ALLOW_RUNTIME_LORA_UPDATING: bool = False
111
    VLLM_SKIP_P2P_CHECK: bool = False
112
    VLLM_DISABLED_KERNELS: list[str] = []
113
    VLLM_DISABLE_PYNCCL: bool = False
114
    VLLM_ROCM_USE_AITER: bool = False
115
    VLLM_ROCM_USE_AITER_PAGED_ATTN: bool = False
116
    VLLM_ROCM_USE_AITER_LINEAR: bool = True
117
    VLLM_ROCM_USE_AITER_MOE: bool = True
118
    VLLM_ROCM_USE_AITER_RMSNORM: bool = True
119
    VLLM_ROCM_USE_AITER_MLA: bool = True
120
    VLLM_ROCM_USE_AITER_MHA: bool = True
121
    VLLM_ROCM_USE_AITER_FP4_ASM_GEMM: bool = False
122
    VLLM_ROCM_USE_AITER_TRITON_ROPE: bool = False
123
    VLLM_ROCM_USE_AITER_FP8BMM: bool = True
124
    VLLM_ROCM_USE_AITER_UNIFIED_ATTENTION: bool = False
125
    VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS: bool = False
126
    VLLM_ROCM_USE_AITER_TRITON_GEMM: bool = True
127
    VLLM_ROCM_USE_SKINNY_GEMM: bool = True
128
    VLLM_ROCM_FP8_PADDING: bool = True
129
    VLLM_ROCM_MOE_PADDING: bool = True
130
    VLLM_ROCM_CUSTOM_PAGED_ATTN: bool = True
131
    VLLM_ENABLE_V1_MULTIPROCESSING: bool = True
132
    VLLM_LOG_BATCHSIZE_INTERVAL: float = -1
133
    VLLM_DISABLE_COMPILE_CACHE: bool = False
134
    Q_SCALE_CONSTANT: int = 200
135
136
    K_SCALE_CONSTANT: int = 200
    V_SCALE_CONSTANT: int = 100
137
    VLLM_SERVER_DEV_MODE: bool = False
138
    VLLM_V1_OUTPUT_PROC_CHUNK_SIZE: int = 128
139
    VLLM_MLA_DISABLE: bool = False
140
    VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH: int = 32
141
142
    VLLM_RAY_PER_WORKER_GPUS: float = 1.0
    VLLM_RAY_BUNDLE_INDICES: str = ""
143
    VLLM_CUDART_SO_PATH: str | None = None
144
    VLLM_DP_RANK: int = 0
145
    VLLM_DP_RANK_LOCAL: int = -1
146
    VLLM_DP_SIZE: int = 1
147
    VLLM_USE_STANDALONE_COMPILE: bool = True
148
149
    VLLM_DP_MASTER_IP: str = ""
    VLLM_DP_MASTER_PORT: int = 0
150
    VLLM_MOE_DP_CHUNK_SIZE: int = 256
151
    VLLM_ENABLE_MOE_DP_CHUNK: bool = True
152
    VLLM_RANDOMIZE_DP_DUMMY_INPUTS: bool = False
153
    VLLM_RAY_DP_PACK_STRATEGY: Literal["strict", "fill", "span"] = "strict"
154
    VLLM_MARLIN_USE_ATOMIC_ADD: bool = False
155
    VLLM_MARLIN_INPUT_DTYPE: Literal["int8", "fp8"] | None = None
156
    VLLM_MXFP4_USE_MARLIN: bool | None = None
157
    VLLM_DEEPEPLL_NVFP4_DISPATCH: bool = False
158
    VLLM_V1_USE_OUTLINES_CACHE: bool = False
159
    VLLM_TPU_BUCKET_PADDING_GAP: int = 0
160
    VLLM_TPU_MOST_MODEL_LEN: int | None = None
161
    VLLM_TPU_USING_PATHWAYS: bool = False
162
    VLLM_USE_DEEP_GEMM: bool = True
163
    VLLM_MOE_USE_DEEP_GEMM: bool = True
164
    VLLM_USE_DEEP_GEMM_E8M0: bool = True
165
166
167
168
169
    VLLM_DEEP_GEMM_WARMUP: Literal[
        "skip",
        "full",
        "relax",
    ] = "relax"
170
    VLLM_USE_FUSED_MOE_GROUPED_TOPK: bool = True
171
    VLLM_BLOCKSCALE_FP8_GEMM_FLASHINFER: bool = False
172
    VLLM_USE_FLASHINFER_MOE_FP16: bool = False
173
174
    VLLM_USE_FLASHINFER_MOE_FP8: bool = False
    VLLM_USE_FLASHINFER_MOE_FP4: bool = False
175
176
177
    VLLM_FLASHINFER_MOE_BACKEND: Literal["throughput", "latency", "masked_gemm"] = (
        "latency"
    )
178
    VLLM_FLASHINFER_WORKSPACE_BUFFER_SIZE: int = 394 * 1024 * 1024
179
    VLLM_XGRAMMAR_CACHE_MB: int = 0
180
    VLLM_MSGPACK_ZERO_COPY_THRESHOLD: int = 256
181
    VLLM_ALLOW_INSECURE_SERIALIZATION: bool = False
Robert Shaw's avatar
Robert Shaw committed
182
    VLLM_NIXL_SIDE_CHANNEL_HOST: str = "localhost"
183
    VLLM_NIXL_SIDE_CHANNEL_PORT: int = 5600
184
    VLLM_MOONCAKE_BOOTSTRAP_PORT: int = 8998
185
186
187
188
189
190
191
192
    VLLM_ALL2ALL_BACKEND: Literal[
        "naive",
        "pplx",
        "deepep_high_throughput",
        "deepep_low_latency",
        "allgather_reducescatter",
        "flashinfer_all2allv",
    ] = "allgather_reducescatter"
193
    VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE: int = 163840
194
    VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS: int = 1
195
    VLLM_SLEEP_WHEN_IDLE: bool = False
196
    VLLM_MQ_MAX_CHUNK_BYTES_MB: int = 16
197
    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: int = 300
198
    VLLM_KV_CACHE_LAYOUT: Literal["NHD", "HND"] | None = None
199
    VLLM_COMPUTE_NANS_IN_LOGITS: bool = False
200
    VLLM_USE_NVFP4_CT_EMULATIONS: bool = False
201
202
203
    VLLM_ROCM_QUICK_REDUCE_QUANTIZATION: Literal[
        "FP", "INT8", "INT6", "INT4", "NONE"
    ] = "NONE"
204
    VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16: bool = True
205
    VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB: int | None = None
206
    VLLM_NIXL_ABORT_REQUEST_TIMEOUT: int = 480
207
208
209
210
    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
211
    VLLM_MOONCAKE_ABORT_REQUEST_TIMEOUT: int = 480
212
    VLLM_USE_CUDNN_PREFILL: bool = False
213
    VLLM_USE_TRTLLM_RAGGED_DEEPSEEK_PREFILL: bool = False
214
    VLLM_ENABLE_CUDAGRAPH_GC: bool = False
215
    VLLM_LOOPBACK_IP: str = ""
216
    VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE: bool = True
217
    VLLM_ENABLE_RESPONSES_API_STORE: bool = False
218
    VLLM_USE_TRTLLM_ATTENTION: str | None = None
219
    VLLM_NVFP4_GEMM_BACKEND: str | None = None
220
    VLLM_FLASHINFER_DISABLE_Q_QUANTIZATION: bool = False
221
    VLLM_HAS_FLASHINFER_CUBIN: bool = False
222
223
    VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8: bool = False
    VLLM_USE_FLASHINFER_MOE_MXFP4_BF16: bool = False
xiao-llm's avatar
xiao-llm committed
224
    VLLM_ROCM_FP8_MFMA_PAGE_ATTN: bool = False
225
    VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS: bool = False
226
    VLLM_ALLREDUCE_USE_SYMM_MEM: bool = True
227
    VLLM_TUNED_CONFIG_FOLDER: str | None = None
228
    VLLM_GPT_OSS_SYSTEM_TOOL_MCP_LABELS: set[str] = set()
229
    VLLM_USE_EXPERIMENTAL_PARSER_CONTEXT: bool = False
230
    VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS: bool = False
231
    VLLM_TOOL_JSON_ERROR_AUTOMATIC_RETRY: bool = False
232
    VLLM_CUSTOM_SCOPES_FOR_PROFILING: bool = False
233
    VLLM_NVTX_SCOPES_FOR_PROFILING: bool = False
234
    VLLM_KV_EVENTS_USE_INT_BLOCK_HASHES: bool = True
235
    VLLM_OBJECT_STORAGE_SHM_BUFFER_NAME: str = "VLLM_OBJECT_STORAGE_SHM_BUFFER"
236
    VLLM_DEEPEP_BUFFER_SIZE_MB: int = 1024
237
238
    VLLM_DEEPEP_HIGH_THROUGHPUT_FORCE_INTRA_NODE: bool = False
    VLLM_DEEPEP_LOW_LATENCY_USE_MNNVL: bool = False
239
    VLLM_DBO_COMM_SMS: int = 20
240
241
    VLLM_PATTERN_MATCH_DEBUG: str | None = None
    VLLM_DEBUG_DUMP_PATH: str | None = None
242
243
    VLLM_ENABLE_INDUCTOR_MAX_AUTOTUNE: bool = True
    VLLM_ENABLE_INDUCTOR_COORDINATE_DESCENT_TUNING: bool = True
244
    VLLM_USE_NCCL_SYMM_MEM: bool = False
245
    VLLM_NCCL_INCLUDE_PATH: str | None = None
246
    VLLM_USE_FBGEMM: bool = False
247
    VLLM_GC_DEBUG: str = ""
248
    VLLM_DEBUG_WORKSPACE: bool = False
249
    VLLM_DISABLE_SHARED_EXPERTS_STREAM: bool = False
250
    VLLM_SHARED_EXPERTS_STREAM_TOKEN_THRESHOLD: int = 256
251
    VLLM_COMPILE_CACHE_SAVE_FORMAT: Literal["binary", "unpacked"] = "binary"
Woosuk Kwon's avatar
Woosuk Kwon committed
252
    VLLM_USE_V2_MODEL_RUNNER: bool = False
253
    VLLM_DEBUG_MFU_METRICS: bool = False
254

255
256
257
258
259
260
261
262
263
264
265
266
267
268
269

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"),
    )


270
def maybe_convert_int(value: str | None) -> int | None:
271
272
273
274
275
    if value is None:
        return None
    return int(value)


276
def maybe_convert_bool(value: str | None) -> bool | None:
277
278
279
280
281
    if value is None:
        return None
    return bool(int(value))


282
283
284
285
def disable_compile_cache() -> bool:
    return bool(int(os.getenv("VLLM_DISABLE_COMPILE_CACHE", "0")))


286
def use_aot_compile() -> bool:
287
288
289
    from vllm.model_executor.layers.batch_invariant import (
        vllm_is_batch_invariant,
    )
290
    from vllm.utils.torch_utils import is_torch_equal_or_newer
291

292
293
294
295
296
297
    default_value = (
        "1"
        if is_torch_equal_or_newer("2.10.0.dev") and not disable_compile_cache()
        else "0"
    )

298
299
300
301
    return (
        not vllm_is_batch_invariant()
        and os.environ.get("VLLM_USE_AOT_COMPILE", default_value) == "1"
    )
302
303


304
def env_with_choices(
305
    env_name: str,
306
307
    default: str | None,
    choices: list[str] | Callable[[], list[str]],
308
    case_sensitive: bool = True,
309
) -> Callable[[], str | None]:
310
311
    """
    Create a lambda that validates environment variable against allowed choices
312

313
314
315
316
317
    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
318

319
320
321
322
    Returns:
        Lambda function for environment_variables dict
    """

323
    def _get_validated_env() -> str | None:
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
        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:
339
340
341
342
            raise ValueError(
                f"Invalid value '{value}' for {env_name}. "
                f"Valid options: {actual_choices}."
            )
343
344
345
346
347
348

        return value

    return _get_validated_env


349
def env_list_with_choices(
350
351
    env_name: str,
    default: list[str],
352
    choices: list[str] | Callable[[], list[str]],
353
354
    case_sensitive: bool = True,
) -> Callable[[], list[str]]:
355
    """
356
    Create a lambda that validates environment variable
357
    containing comma-separated values against allowed choices
358

359
360
361
362
363
    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
364

365
366
367
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
    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:
394
395
396
397
                raise ValueError(
                    f"Invalid value '{val}' in {env_name}. "
                    f"Valid options: {actual_choices}."
                )
398
399
400
401
402
403

        return values

    return _get_validated_env_list


404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
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


422
def get_vllm_port() -> int | None:
423
    """Get the port from VLLM_PORT environment variable.
424

425
426
    Returns:
        The port number as an integer if VLLM_PORT is set, None otherwise.
427

428
429
430
    Raises:
        ValueError: If VLLM_PORT is a URI, suggest k8s service discovery issue.
    """
431
    if "VLLM_PORT" not in os.environ:
432
433
        return None

434
    port = os.getenv("VLLM_PORT", "0")
435
436
437
438
439

    try:
        return int(port)
    except ValueError as err:
        from urllib.parse import urlparse
440

441
442
443
444
445
446
447
        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
448
        raise ValueError(f"VLLM_PORT '{port}' must be a valid integer") from err
449
450


Ning Xie's avatar
Ning Xie committed
451
# The start-* and end* here are used by the documentation generator
452
453
# to extract the used env vars.

454
# --8<-- [start:env-vars-definition]
455

456
457
logger = logging.getLogger(__name__)

458
environment_variables: dict[str, Callable[[], Any]] = {
459
    # ================== Installation Time Env Vars ==================
460
    # Target device of vLLM, supporting [cuda (by default),
461
    # rocm, cpu]
462
    "VLLM_TARGET_DEVICE": lambda: os.getenv("VLLM_TARGET_DEVICE", "cuda").lower(),
463
    # Main CUDA version of vLLM. This follows PyTorch but can be overridden.
464
    "VLLM_MAIN_CUDA_VERSION": lambda: os.getenv("VLLM_MAIN_CUDA_VERSION", "").lower()
465
    or "12.9",
466
    # Controls PyTorch float32 matmul precision mode within vLLM workers.
467
    # Valid options mirror torch.set_float32_matmul_precision
468
469
    "VLLM_FLOAT32_MATMUL_PRECISION": env_with_choices(
        "VLLM_FLOAT32_MATMUL_PRECISION",
470
471
        "highest",
        ["highest", "high", "medium"],
472
473
        case_sensitive=False,
    ),
474
475
    # Maximum number of compilation jobs to run in parallel.
    # By default this is the number of CPUs
476
    "MAX_JOBS": lambda: os.getenv("MAX_JOBS", None),
477
478
479
    # Number of threads to use for nvcc
    # By default this is 1.
    # If set, `MAX_JOBS` will be reduced to avoid oversubscribing the CPU.
480
    "NVCC_THREADS": lambda: os.getenv("NVCC_THREADS", None),
481
    # If set, vllm will use precompiled binaries (*.so)
482
483
484
485
486
    "VLLM_USE_PRECOMPILED": lambda: os.environ.get("VLLM_USE_PRECOMPILED", "")
    .strip()
    .lower()
    in ("1", "true")
    or bool(os.environ.get("VLLM_PRECOMPILED_WHEEL_LOCATION")),
487
488
489
490
    # 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"))
    ),
491
492
    # Used to mark that setup.py is running in a Docker build context,
    # in order to force the use of precompiled binaries.
493
494
495
496
    "VLLM_DOCKER_BUILD_CONTEXT": lambda: os.environ.get("VLLM_DOCKER_BUILD_CONTEXT", "")
    .strip()
    .lower()
    in ("1", "true"),
497
498
499
    # CMake build type
    # If not set, defaults to "Debug" or "RelWithDebInfo"
    # Available options: "Debug", "Release", "RelWithDebInfo"
500
501
502
    "CMAKE_BUILD_TYPE": env_with_choices(
        "CMAKE_BUILD_TYPE", None, ["Debug", "Release", "RelWithDebInfo"]
    ),
503
    # If set, vllm will print verbose logs during installation
504
    "VERBOSE": lambda: bool(int(os.getenv("VERBOSE", "0"))),
505
    # Root directory for vLLM configuration files
506
    # Defaults to `~/.config/vllm` unless `XDG_CONFIG_HOME` is set
507
508
509
    # 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**.
510
    "VLLM_CONFIG_ROOT": lambda: os.path.expanduser(
511
512
513
        os.getenv(
            "VLLM_CONFIG_ROOT",
            os.path.join(get_default_config_root(), "vllm"),
514
515
        )
    ),
516
    # ================== Runtime Env Vars ==================
517
    # Root directory for vLLM cache files
518
    # Defaults to `~/.cache/vllm` unless `XDG_CACHE_HOME` is set
519
    "VLLM_CACHE_ROOT": lambda: os.path.expanduser(
520
521
522
        os.getenv(
            "VLLM_CACHE_ROOT",
            os.path.join(get_default_cache_root(), "vllm"),
523
524
        )
    ),
525
526
527
528
    # 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.
529
    "VLLM_HOST_IP": lambda: os.getenv("VLLM_HOST_IP", ""),
530
    # used in distributed environment to manually set the communication port
531
532
533
    # 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.
534
    "VLLM_PORT": get_vllm_port,
535
536
    # path used for ipc when the frontend api server is running in
    # multi-processing mode to communicate with the backend engine process.
537
538
539
    "VLLM_RPC_BASE_PATH": lambda: os.getenv(
        "VLLM_RPC_BASE_PATH", tempfile.gettempdir()
    ),
540
541
    # If true, will load models from ModelScope instead of Hugging Face Hub.
    # note that the value is true or false, not numbers
542
543
544
545
    "VLLM_USE_MODELSCOPE": lambda: os.environ.get(
        "VLLM_USE_MODELSCOPE", "False"
    ).lower()
    == "true",
546
    # Interval in seconds to log a warning message when the ring buffer is full
547
548
549
    "VLLM_RINGBUFFER_WARNING_INTERVAL": lambda: int(
        os.environ.get("VLLM_RINGBUFFER_WARNING_INTERVAL", "60")
    ),
550
551
    # path to cudatoolkit home directory, under which should be bin, include,
    # and lib directories.
552
    "CUDA_HOME": lambda: os.environ.get("CUDA_HOME", None),
553
554
    # 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
555
    "VLLM_NCCL_SO_PATH": lambda: os.environ.get("VLLM_NCCL_SO_PATH", None),
556
557
    # 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`
558
    "LD_LIBRARY_PATH": lambda: os.environ.get("LD_LIBRARY_PATH", None),
559
560
561
562
    # 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")
    ),
563
564
    # Use separate prefill and decode kernels for V1 attention instead of
    # the unified triton kernel.
565
566
567
568
    "VLLM_V1_USE_PREFILL_DECODE_ATTENTION": lambda: (
        os.getenv("VLLM_V1_USE_PREFILL_DECODE_ATTENTION", "False").lower()
        in ("true", "1")
    ),
569
570
    # Force vllm to use a specific flash-attention version (2 or 3), only valid
    # when using the flash-attention backend.
571
572
573
    "VLLM_FLASH_ATTN_VERSION": lambda: maybe_convert_int(
        os.environ.get("VLLM_FLASH_ATTN_VERSION", None)
    ),
574
    # Feature flag to enable/disable Inductor standalone compile.
575
576
    # In torch <= 2.7 we ignore this flag; in torch >= 2.9 this is
    # enabled by default.
577
    "VLLM_USE_STANDALONE_COMPILE": lambda: os.environ.get(
578
        "VLLM_USE_STANDALONE_COMPILE", "1"
579
580
    )
    == "1",
581
582
    # Debug pattern matching inside custom passes.
    # Should be set to the fx.Node name (e.g. 'getitem_34' or 'scaled_mm_3').
583
584
585
    "VLLM_PATTERN_MATCH_DEBUG": lambda: os.environ.get(
        "VLLM_PATTERN_MATCH_DEBUG", None
    ),
586
587
    # Dump fx graphs to the given directory.
    # It will override CompilationConfig.debug_dump_path if set.
588
    "VLLM_DEBUG_DUMP_PATH": lambda: os.environ.get("VLLM_DEBUG_DUMP_PATH", None),
589
590
591
592
    # 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,
593
594
595
596
597
    # Feature flag to enable/disable bytecode in
    # TorchCompileWithNoGuardsWrapper.
    "VLLM_USE_BYTECODE_HOOK": lambda: bool(
        int(os.environ.get("VLLM_USE_BYTECODE_HOOK", "1"))
    ),
598
599
600
601
    # 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",
602
603
    # local rank of the process in the distributed setting, used to determine
    # the GPU device id
604
    "LOCAL_RANK": lambda: int(os.environ.get("LOCAL_RANK", "0")),
605
    # used to control the visible devices in the distributed setting
606
    "CUDA_VISIBLE_DEVICES": lambda: os.environ.get("CUDA_VISIBLE_DEVICES", None),
607
    # timeout for each iteration in the engine
608
609
610
    "VLLM_ENGINE_ITERATION_TIMEOUT_S": lambda: int(
        os.environ.get("VLLM_ENGINE_ITERATION_TIMEOUT_S", "60")
    ),
611
612
613
614
615
    # 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")
    ),
616
    # API key for vLLM API server
617
    "VLLM_API_KEY": lambda: os.environ.get("VLLM_API_KEY", None),
618
    # Whether to log responses from API Server for debugging
619
620
621
622
    "VLLM_DEBUG_LOG_API_SERVER_RESPONSE": lambda: os.environ.get(
        "VLLM_DEBUG_LOG_API_SERVER_RESPONSE", "False"
    ).lower()
    == "true",
623
    # S3 access information, used for tensorizer to load model from S3
624
625
626
    "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),
627
    # Usage stats collection
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
    "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"),
643
644
645
646
    # 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
647
648
649
    "VLLM_CONFIGURE_LOGGING": lambda: bool(
        int(os.getenv("VLLM_CONFIGURE_LOGGING", "1"))
    ),
650
    "VLLM_LOGGING_CONFIG_PATH": lambda: os.getenv("VLLM_LOGGING_CONFIG_PATH"),
651
    # this is used for configuring the default logging level
652
    "VLLM_LOGGING_LEVEL": lambda: os.getenv("VLLM_LOGGING_LEVEL", "INFO").upper(),
653
    # this is used for configuring the default logging stream
654
    "VLLM_LOGGING_STREAM": lambda: os.getenv("VLLM_LOGGING_STREAM", "ext://sys.stdout"),
655
    # if set, VLLM_LOGGING_PREFIX will be prepended to all log messages
656
    "VLLM_LOGGING_PREFIX": lambda: os.getenv("VLLM_LOGGING_PREFIX", ""),
Nick Hill's avatar
Nick Hill committed
657
658
659
660
661
    # 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",
662
663
    # If set, vllm will log stats at this interval in seconds
    # If not set, vllm will log stats every 10 seconds.
664
665
666
    "VLLM_LOG_STATS_INTERVAL": lambda: val
    if (val := float(os.getenv("VLLM_LOG_STATS_INTERVAL", "10."))) > 0.0
    else 10.0,
667
668
669
    # Trace function calls
    # If set to 1, vllm will trace function calls
    # Useful for debugging
670
    "VLLM_TRACE_FUNCTION": lambda: int(os.getenv("VLLM_TRACE_FUNCTION", "0")),
671
    # Backend for attention computation
672
    # Example options:
673
674
    # - "TORCH_SDPA": use torch.nn.MultiheadAttention
    # - "FLASH_ATTN": use FlashAttention
675
    # - "FLASHINFER": use flashinfer
676
    # - "FLASHMLA": use FlashMLA
677
    # - "FLASH_ATTN_MLA": use FlashAttention for MLA
678
679
    # - "FLASHINFER_MLA": use FlashInfer for MLA
    # - "CUTLASS_MLA": use CUTLASS for MLA
680
    # All possible options loaded dynamically from AttentionBackendEnum
681
682
683
684
685
    "VLLM_ATTENTION_BACKEND": env_with_choices(
        "VLLM_ATTENTION_BACKEND",
        None,
        lambda: list(
            __import__(
686
687
                "vllm.attention.backends.registry", fromlist=["AttentionBackendEnum"]
            ).AttentionBackendEnum.__members__.keys()
688
689
        ),
    ),
690
    # If set, vllm will use flashinfer sampler
691
692
693
694
695
    "VLLM_USE_FLASHINFER_SAMPLER": lambda: bool(
        int(os.environ["VLLM_USE_FLASHINFER_SAMPLER"])
    )
    if "VLLM_USE_FLASHINFER_SAMPLER" in os.environ
    else None,
696
    # Pipeline stage partition strategy
697
    "VLLM_PP_LAYER_PARTITION": lambda: os.getenv("VLLM_PP_LAYER_PARTITION", None),
698
    # (CPU backend only) CPU key-value cache space.
699
    # default is None and will be set as 4 GB
700
701
702
    "VLLM_CPU_KVCACHE_SPACE": lambda: int(os.getenv("VLLM_CPU_KVCACHE_SPACE", "0"))
    if "VLLM_CPU_KVCACHE_SPACE" in os.environ
    else None,
703
704
    # (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 '|'.
705
    "VLLM_CPU_OMP_THREADS_BIND": lambda: os.getenv("VLLM_CPU_OMP_THREADS_BIND", "auto"),
706
707
    # (CPU backend only) CPU cores not used by OMP threads .
    # Those CPU cores will not be used by OMP threads of a rank.
708
709
710
711
712
    "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,
713
    # (CPU backend only) whether to use SGL kernels, optimized for small batch.
714
    "VLLM_CPU_SGL_KERNEL": lambda: bool(int(os.getenv("VLLM_CPU_SGL_KERNEL", "0"))),
715
716
717
718
719
720
721
    # 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
722
723
724
    "VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE": env_with_choices(
        "VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE", "auto", ["auto", "nccl", "shm"]
    ),
725
    # If the env var is set, it enables GPU communication overlap
726
    # (experimental feature) in Ray's Compiled Graph.
727
728
729
    "VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM": lambda: bool(
        int(os.getenv("VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM", "0"))
    ),
730
731
732
    # 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.
733
734
735
    "VLLM_USE_RAY_WRAPPED_PP_COMM": lambda: bool(
        int(os.getenv("VLLM_USE_RAY_WRAPPED_PP_COMM", "1"))
    ),
736
737
    # Use dedicated multiprocess context for workers.
    # Both spawn and fork work
738
739
740
    "VLLM_WORKER_MULTIPROC_METHOD": env_with_choices(
        "VLLM_WORKER_MULTIPROC_METHOD", "fork", ["spawn", "fork"]
    ),
741
    # Path to the cache for storing downloaded assets
742
    "VLLM_ASSETS_CACHE": lambda: os.path.expanduser(
743
744
745
        os.getenv(
            "VLLM_ASSETS_CACHE",
            os.path.join(get_default_cache_root(), "vllm", "assets"),
746
747
        )
    ),
748
749
    # If the env var is set, we will clean model file in
    # this path $VLLM_ASSETS_CACHE/model_streamer/$model_name
750
751
752
    "VLLM_ASSETS_CACHE_MODEL_CLEAN": lambda: bool(
        int(os.getenv("VLLM_ASSETS_CACHE_MODEL_CLEAN", "0"))
    ),
753
754
    # Timeout for fetching images when serving multimodal models
    # Default is 5 seconds
755
    "VLLM_IMAGE_FETCH_TIMEOUT": lambda: int(os.getenv("VLLM_IMAGE_FETCH_TIMEOUT", "5")),
756
    # Timeout for fetching videos when serving multimodal models
757
    # Default is 30 seconds
758
759
760
    "VLLM_VIDEO_FETCH_TIMEOUT": lambda: int(
        os.getenv("VLLM_VIDEO_FETCH_TIMEOUT", "30")
    ),
761
    # Timeout for fetching audio when serving multimodal models
762
    # Default is 10 seconds
763
764
765
    "VLLM_AUDIO_FETCH_TIMEOUT": lambda: int(
        os.getenv("VLLM_AUDIO_FETCH_TIMEOUT", "10")
    ),
766
767
    # Whether to allow HTTP redirects when fetching from media URLs.
    # Default to True
768
769
770
    "VLLM_MEDIA_URL_ALLOW_REDIRECTS": lambda: bool(
        int(os.getenv("VLLM_MEDIA_URL_ALLOW_REDIRECTS", "1"))
    ),
771
772
773
    # Max number of workers for the thread pool handling
    # media bytes loading. Set to 1 to disable parallel processing.
    # Default is 8
774
775
776
    "VLLM_MEDIA_LOADING_THREAD_COUNT": lambda: int(
        os.getenv("VLLM_MEDIA_LOADING_THREAD_COUNT", "8")
    ),
777
778
779
    # 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
780
781
782
    "VLLM_MAX_AUDIO_CLIP_FILESIZE_MB": lambda: int(
        os.getenv("VLLM_MAX_AUDIO_CLIP_FILESIZE_MB", "25")
    ),
783
784
785
786
787
788
789
    # 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.
790
791
792
    "VLLM_VIDEO_LOADER_BACKEND": lambda: os.getenv(
        "VLLM_VIDEO_LOADER_BACKEND", "opencv"
    ),
793
794
795
796
797
798
799
800
    # 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"),
801
802
    # Path to the XLA persistent cache directory.
    # Only used for XLA devices such as TPUs.
803
    "VLLM_XLA_CACHE_PATH": lambda: os.path.expanduser(
804
        os.getenv(
805
            "VLLM_XLA_CACHE_PATH",
806
            os.path.join(get_default_cache_root(), "vllm", "xla_cache"),
807
808
        )
    ),
809
    # If set, assert on XLA recompilation after each execution step.
810
811
812
    "VLLM_XLA_CHECK_RECOMPILATION": lambda: bool(
        int(os.getenv("VLLM_XLA_CHECK_RECOMPILATION", "0"))
    ),
813
    # Enable SPMD mode for TPU backend.
814
815
    "VLLM_XLA_USE_SPMD": lambda: bool(int(os.getenv("VLLM_XLA_USE_SPMD", "0"))),
    "VLLM_FUSED_MOE_CHUNK_SIZE": lambda: int(
816
        os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", str(16 * 1024))
817
    ),
818
819
820
    # 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.
821
822
823
    "VLLM_ENABLE_FUSED_MOE_ACTIVATION_CHUNKING": lambda: bool(
        int(os.getenv("VLLM_ENABLE_FUSED_MOE_ACTIVATION_CHUNKING", "1"))
    ),
824
825
    # If set, the OpenAI API server will stay alive even after the underlying
    # AsyncLLMEngine errors and stops serving requests
826
    "VLLM_KEEP_ALIVE_ON_ENGINE_DEATH": lambda: bool(
827
        int(os.getenv("VLLM_KEEP_ALIVE_ON_ENGINE_DEATH", "0"))
828
    ),
829
830
831
832
    # 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.
833
834
835
836
    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": lambda: (
        os.environ.get("VLLM_ALLOW_LONG_MAX_MODEL_LEN", "0").strip().lower()
        in ("1", "true")
    ),
837
838
    # If set, forces FP8 Marlin to be used for FP8 quantization regardless
    # of the hardware support for FP8 compute.
839
840
841
842
843
844
845
    "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"
    ),
846
847
    # Time in ms for the zmq client to wait for a response from the backend
    # server for simple data operations
848
    "VLLM_RPC_TIMEOUT": lambda: int(os.getenv("VLLM_RPC_TIMEOUT", "10000")),
849
    # Timeout in seconds for keeping HTTP connections alive in API server
850
851
852
    "VLLM_HTTP_TIMEOUT_KEEP_ALIVE": lambda: int(
        os.environ.get("VLLM_HTTP_TIMEOUT_KEEP_ALIVE", "5")
    ),
853
854
855
    # 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
856
857
858
    "VLLM_PLUGINS": lambda: None
    if "VLLM_PLUGINS" not in os.environ
    else os.environ["VLLM_PLUGINS"].split(","),
859
860
861
    # a local directory to look in for unrecognized LoRA adapters.
    # only works if plugins are enabled and
    # VLLM_ALLOW_RUNTIME_LORA_UPDATING is enabled.
862
863
864
    "VLLM_LORA_RESOLVER_CACHE_DIR": lambda: os.getenv(
        "VLLM_LORA_RESOLVER_CACHE_DIR", None
    ),
865
866
867
    # Enables torch CUDA profiling if set to 1.
    # Deprecated, see profiler_config.
    "VLLM_TORCH_CUDA_PROFILE": lambda: os.getenv("VLLM_TORCH_CUDA_PROFILE"),
868
    # Enables torch profiler if set.
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
    # Deprecated, see profiler_config.
    "VLLM_TORCH_PROFILER_DIR": lambda: os.getenv("VLLM_TORCH_PROFILER_DIR"),
    # Enable torch profiler to record shapes if set to 1.
    # Deprecated, see profiler_config.
    "VLLM_TORCH_PROFILER_RECORD_SHAPES": lambda: (
        os.getenv("VLLM_TORCH_PROFILER_RECORD_SHAPES")
    ),
    # Enable torch profiler to profile memory if set to 1.
    # Deprecated, see profiler_config.
    "VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY": lambda: (
        os.getenv("VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY")
    ),
    # Enable torch profiler to profile stack if set to 1.
    # Deprecated, see profiler_config.
    "VLLM_TORCH_PROFILER_WITH_STACK": lambda: (
        os.getenv("VLLM_TORCH_PROFILER_WITH_STACK")
    ),
    # Enable torch profiler to profile flops if set to 1.
    # Deprecated, see profiler_config.
    "VLLM_TORCH_PROFILER_WITH_FLOPS": lambda: (
        os.getenv("VLLM_TORCH_PROFILER_WITH_FLOPS")
    ),
    # Disable torch profiling of the AsyncLLMEngine process if set to 1.
    # Deprecated, see profiler_config.
    "VLLM_TORCH_PROFILER_DISABLE_ASYNC_LLM": lambda: (
        os.getenv("VLLM_TORCH_PROFILER_DISABLE_ASYNC_LLM")
895
896
897
    ),
    # Delay number of iterations before starting profiling when using
    # the torch/torch CUDA profiler. If set to 0, will start profiling immediately.
898
899
    # Deprecated, see profiler_config.
    "VLLM_PROFILER_DELAY_ITERS": lambda: (os.getenv("VLLM_PROFILER_DELAY_ITERS")),
900
901
    # Maximum number of iterations to profile when using the torch/torch CUDA profiler.
    # If set to 0, will not limit the number of iterations.
902
    "VLLM_PROFILER_MAX_ITERS": lambda: os.getenv("VLLM_PROFILER_MAX_ITERS"),
903
    # Control whether torch profiler gzip-compresses profiling files.
904
905
    # Deprecated, see profiler_config.
    "VLLM_TORCH_PROFILER_USE_GZIP": lambda: os.getenv("VLLM_TORCH_PROFILER_USE_GZIP"),
906
    # Control whether torch profiler dumps the self_cuda_time_total table.
907
908
909
910
    # Set to 0 to disable dumping the table.
    # Deprecated, see profiler_config.
    "VLLM_TORCH_PROFILER_DUMP_CUDA_TIME_TOTAL": lambda: (
        os.getenv("VLLM_TORCH_PROFILER_DUMP_CUDA_TIME_TOTAL")
911
    ),
912
    # If set, vLLM will use Triton implementations of AWQ.
913
    "VLLM_USE_TRITON_AWQ": lambda: bool(int(os.getenv("VLLM_USE_TRITON_AWQ", "0"))),
914
    # If set, allow loading or unloading lora adapters in runtime,
915
916
917
918
    "VLLM_ALLOW_RUNTIME_LORA_UPDATING": lambda: (
        os.environ.get("VLLM_ALLOW_RUNTIME_LORA_UPDATING", "0").strip().lower()
        in ("1", "true")
    ),
919
920
921
922
923
924
    # 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
925
    "VLLM_SKIP_P2P_CHECK": lambda: os.getenv("VLLM_SKIP_P2P_CHECK", "1") == "1",
926
927
928
929
    # List of quantization kernels that should be disabled, used for testing
    # and performance comparisons. Currently only affects MPLinearKernel
    # selection
    # (kernels: MacheteLinearKernel, MarlinLinearKernel, ExllamaLinearKernel)
930
931
932
    "VLLM_DISABLED_KERNELS": lambda: []
    if "VLLM_DISABLED_KERNELS" not in os.environ
    else os.environ["VLLM_DISABLED_KERNELS"].split(","),
933
    # Disable pynccl (using torch.distributed instead)
934
935
936
    "VLLM_DISABLE_PYNCCL": lambda: (
        os.getenv("VLLM_DISABLE_PYNCCL", "False").lower() in ("true", "1")
    ),
937
938
    # Disable aiter ops unless specifically enabled.
    # Acts as a parent switch to enable the rest of the other operations.
939
940
941
    "VLLM_ROCM_USE_AITER": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER", "False").lower() in ("true", "1")
    ),
942
943
    # Whether to use aiter paged attention.
    # By default is disabled.
944
945
946
    "VLLM_ROCM_USE_AITER_PAGED_ATTN": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_PAGED_ATTN", "False").lower() in ("true", "1")
    ),
947
948
949
    # use aiter linear op if aiter ops are enabled
    # The following list of related ops
    # - scaled_mm (per-tensor / rowwise)
950
951
952
    "VLLM_ROCM_USE_AITER_LINEAR": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_LINEAR", "True").lower() in ("true", "1")
    ),
953
954
    # Whether to use aiter moe ops.
    # By default is enabled.
955
956
957
    "VLLM_ROCM_USE_AITER_MOE": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_MOE", "True").lower() in ("true", "1")
    ),
958
    # use aiter rms norm op if aiter ops are enabled.
959
960
961
    "VLLM_ROCM_USE_AITER_RMSNORM": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_RMSNORM", "True").lower() in ("true", "1")
    ),
962
963
    # Whether to use aiter mla ops.
    # By default is enabled.
964
965
966
    "VLLM_ROCM_USE_AITER_MLA": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_MLA", "True").lower() in ("true", "1")
    ),
967
968
    # Whether to use aiter mha ops.
    # By default is enabled.
969
970
971
    "VLLM_ROCM_USE_AITER_MHA": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_MHA", "True").lower() in ("true", "1")
    ),
972
973
    # Whether to use aiter fp4 gemm asm.
    # By default is disabled.
974
975
976
    "VLLM_ROCM_USE_AITER_FP4_ASM_GEMM": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_FP4_ASM_GEMM", "False").lower() in ("true", "1")
    ),
977
978
    # Whether to use aiter rope.
    # By default is disabled.
979
980
    "VLLM_ROCM_USE_AITER_TRITON_ROPE": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_TRITON_ROPE", "False").lower() in ("true", "1")
981
    ),
982
983
    # Whether to use aiter triton fp8 bmm kernel
    # By default is enabled.
984
985
986
    "VLLM_ROCM_USE_AITER_FP8BMM": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_FP8BMM", "True").lower() in ("true", "1")
    ),
987
988
989
990
991
    # 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")
    ),
992
    # Whether to use aiter fusion shared experts ops.
993
    # By default is disabled.
994
    "VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS": lambda: (
995
        os.getenv("VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS", "False").lower()
996
997
        in ("true", "1")
    ),
998
999
1000
1001
1002
    # 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")
    ),
1003
    # use rocm skinny gemms
1004
1005
1006
    "VLLM_ROCM_USE_SKINNY_GEMM": lambda: (
        os.getenv("VLLM_ROCM_USE_SKINNY_GEMM", "True").lower() in ("true", "1")
    ),
1007
    # Pad the fp8 weights to 256 bytes for ROCm
1008
    "VLLM_ROCM_FP8_PADDING": lambda: bool(int(os.getenv("VLLM_ROCM_FP8_PADDING", "1"))),
1009
    # Pad the weights for the moe kernel
1010
    "VLLM_ROCM_MOE_PADDING": lambda: bool(int(os.getenv("VLLM_ROCM_MOE_PADDING", "1"))),
1011
    # custom paged attention kernel for MI3* cards
1012
1013
1014
    "VLLM_ROCM_CUSTOM_PAGED_ATTN": lambda: (
        os.getenv("VLLM_ROCM_CUSTOM_PAGED_ATTN", "True").lower() in ("true", "1")
    ),
1015
1016
1017
    # Custom quick allreduce kernel for MI3* cards
    # Choice of quantization level: FP, INT8, INT6, INT4 or NONE
    # Recommended for large models to get allreduce
1018
1019
1020
1021
1022
    "VLLM_ROCM_QUICK_REDUCE_QUANTIZATION": env_with_choices(
        "VLLM_ROCM_QUICK_REDUCE_QUANTIZATION",
        "NONE",
        ["FP", "INT8", "INT6", "INT4", "NONE"],
    ),
1023
1024
1025
1026
    # 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
1027
1028
1029
1030
    "VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16": lambda: (
        os.getenv("VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16", "True").lower()
        in ("true", "1")
    ),
1031
1032
1033
1034
1035
1036
    # 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.
1037
1038
1039
    "VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB": lambda: maybe_convert_int(
        os.environ.get("VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB", None)
    ),
1040
    # Divisor for dynamic query scale factor calculation for FP8 KV Cache
1041
    "Q_SCALE_CONSTANT": lambda: int(os.getenv("Q_SCALE_CONSTANT", "200")),
1042
    # Divisor for dynamic key scale factor calculation for FP8 KV Cache
1043
    "K_SCALE_CONSTANT": lambda: int(os.getenv("K_SCALE_CONSTANT", "200")),
1044
    # Divisor for dynamic value scale factor calculation for FP8 KV Cache
1045
    "V_SCALE_CONSTANT": lambda: int(os.getenv("V_SCALE_CONSTANT", "100")),
1046
    # If set, enable multiprocessing in LLM for the V1 code path.
1047
1048
1049
1050
1051
1052
    "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")
    ),
1053
    "VLLM_DISABLE_COMPILE_CACHE": disable_compile_cache,
1054
1055
1056
    # If set, vllm will run in development mode, which will enable
    # some additional endpoints for developing and debugging,
    # e.g. `/reset_prefix_cache`
1057
    "VLLM_SERVER_DEV_MODE": lambda: bool(int(os.getenv("VLLM_SERVER_DEV_MODE", "0"))),
1058
1059
1060
1061
1062
1063
1064
    # 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.
1065
1066
1067
    "VLLM_V1_OUTPUT_PROC_CHUNK_SIZE": lambda: int(
        os.getenv("VLLM_V1_OUTPUT_PROC_CHUNK_SIZE", "128")
    ),
1068
    # If set, vLLM will disable the MLA attention optimizations.
1069
    "VLLM_MLA_DISABLE": lambda: bool(int(os.getenv("VLLM_MLA_DISABLE", "0"))),
1070
1071
    # If set, vLLM will pick up the provided Flash Attention MLA
    # max number splits for cuda graph decode
1072
1073
1074
    "VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH": lambda: int(
        os.getenv("VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH", "32")
    ),
1075
1076
1077
    # 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.
1078
1079
1080
    "VLLM_RAY_PER_WORKER_GPUS": lambda: float(
        os.getenv("VLLM_RAY_PER_WORKER_GPUS", "1.0")
    ),
1081
1082
1083
    # 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"
1084
    "VLLM_RAY_BUNDLE_INDICES": lambda: os.getenv("VLLM_RAY_BUNDLE_INDICES", ""),
1085
1086
    # In some system, find_loaded_library() may not work. So we allow users to
    # specify the path through environment variable VLLM_CUDART_SO_PATH.
1087
    "VLLM_CUDART_SO_PATH": lambda: os.getenv("VLLM_CUDART_SO_PATH", None),
1088
    # Rank of the process in the data parallel setting
1089
    "VLLM_DP_RANK": lambda: int(os.getenv("VLLM_DP_RANK", "0")),
1090
1091
    # Rank of the process in the data parallel setting.
    # Defaults to VLLM_DP_RANK when not set.
1092
1093
1094
    "VLLM_DP_RANK_LOCAL": lambda: int(
        os.getenv("VLLM_DP_RANK_LOCAL", sys.modules[__name__].VLLM_DP_RANK)
    ),
1095
    # World size of the data parallel setting
1096
    "VLLM_DP_SIZE": lambda: int(os.getenv("VLLM_DP_SIZE", "1")),
1097
    # IP address of the master node in the data parallel setting
1098
    "VLLM_DP_MASTER_IP": lambda: os.getenv("VLLM_DP_MASTER_IP", "127.0.0.1"),
1099
    # Port of the master node in the data parallel setting
1100
    "VLLM_DP_MASTER_PORT": lambda: int(os.getenv("VLLM_DP_MASTER_PORT", "0")),
1101
1102
1103
1104
1105
    # 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.
1106
    "VLLM_MOE_DP_CHUNK_SIZE": lambda: int(os.getenv("VLLM_MOE_DP_CHUNK_SIZE", "256")),
1107
1108
1109
    "VLLM_ENABLE_MOE_DP_CHUNK": lambda: bool(
        int(os.getenv("VLLM_ENABLE_MOE_DP_CHUNK", "1"))
    ),
1110
    # Randomize inputs during dummy runs when using Data Parallel
1111
1112
1113
1114
    "VLLM_RANDOMIZE_DP_DUMMY_INPUTS": lambda: os.environ.get(
        "VLLM_RANDOMIZE_DP_DUMMY_INPUTS", "0"
    )
    == "1",
1115
1116
1117
1118
1119
1120
1121
    # 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;
1122
1123
1124
    # - "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;
1125
1126
1127
1128
    # 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"
    ),
1129
    # Whether to use S3 path for model loading in CI via RunAI Streamer
1130
    "VLLM_CI_USE_S3": lambda: os.environ.get("VLLM_CI_USE_S3", "0") == "1",
1131
    # Use model_redirect to redirect the model name to a local folder.
1132
1133
1134
1135
1136
    # `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
1137
1138
1139
    "VLLM_MODEL_REDIRECT_PATH": lambda: os.environ.get(
        "VLLM_MODEL_REDIRECT_PATH", None
    ),
1140
    # Whether to use atomicAdd reduce in gptq/awq marlin kernel.
1141
1142
1143
1144
    "VLLM_MARLIN_USE_ATOMIC_ADD": lambda: os.environ.get(
        "VLLM_MARLIN_USE_ATOMIC_ADD", "0"
    )
    == "1",
1145
    # Whether to use marlin kernel in mxfp4 quantization method
1146
1147
1148
    "VLLM_MXFP4_USE_MARLIN": lambda: maybe_convert_bool(
        os.environ.get("VLLM_MXFP4_USE_MARLIN", None)
    ),
1149
1150
1151
1152
    # The activation dtype for marlin kernel
    "VLLM_MARLIN_INPUT_DTYPE": env_with_choices(
        "VLLM_MARLIN_INPUT_DTYPE", None, ["int8", "fp8"]
    ),
1153
1154
1155
1156
1157
1158
    # 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"))
    ),
1159
1160
1161
    # 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.
1162
1163
1164
1165
    "VLLM_V1_USE_OUTLINES_CACHE": lambda: os.environ.get(
        "VLLM_V1_USE_OUTLINES_CACHE", "0"
    )
    == "1",
1166
1167
    # Gap between padding buckets for the forward pass. So we have
    # 8, we will run forward pass with [16, 24, 32, ...].
1168
1169
1170
1171
1172
1173
1174
1175
    "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)
    ),
1176
    # Whether using Pathways
1177
1178
1179
    "VLLM_TPU_USING_PATHWAYS": lambda: bool(
        "proxy" in os.getenv("JAX_PLATFORMS", "").lower()
    ),
1180
    # Allow use of DeepGemm kernels for fused moe ops.
1181
    "VLLM_USE_DEEP_GEMM": lambda: bool(int(os.getenv("VLLM_USE_DEEP_GEMM", "1"))),
1182
1183
1184
1185
    # 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"))
    ),
1186
    # Whether to use E8M0 scaling when DeepGEMM is used on Blackwell GPUs.
1187
1188
1189
    "VLLM_USE_DEEP_GEMM_E8M0": lambda: bool(
        int(os.getenv("VLLM_USE_DEEP_GEMM_E8M0", "1"))
    ),
1190
1191
1192
1193
    # 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.
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
    # 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",
        ],
1209
    ),
1210
    # Whether to use fused grouped_topk used for MoE expert selection.
1211
1212
1213
    "VLLM_USE_FUSED_MOE_GROUPED_TOPK": lambda: bool(
        int(os.getenv("VLLM_USE_FUSED_MOE_GROUPED_TOPK", "1"))
    ),
1214
1215
1216
1217
1218
    # 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(
        int(os.getenv("VLLM_BLOCKSCALE_FP8_GEMM_FLASHINFER", "0"))
    ),
1219
    # Allow use of FlashInfer MoE kernels for fused moe ops.
1220
1221
1222
    "VLLM_USE_FLASHINFER_MOE_FP16": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_FP16", "0"))
    ),
1223
    # Allow use of FlashInfer MoE kernels for fused moe ops.
1224
1225
1226
    "VLLM_USE_FLASHINFER_MOE_FP8": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_FP8", "0"))
    ),
1227
    # Allow use of FlashInfer CUTLASS kernels for fused moe ops.
1228
1229
1230
    "VLLM_USE_FLASHINFER_MOE_FP4": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_FP4", "0"))
    ),
1231
1232
    # If set to 1, use the FlashInfer
    # MXFP8 (activation) x MXFP4 (weight) MoE backend.
1233
1234
1235
    "VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8", "0"))
    ),
1236
1237
1238
1239
    # 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.
1240
1241
1242
    "VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS", "0"))
    ),
1243
1244
    # If set to 1, use the FlashInfer
    # BF16 (activation) x MXFP4 (weight) MoE backend.
1245
1246
1247
    "VLLM_USE_FLASHINFER_MOE_MXFP4_BF16": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_MXFP4_BF16", "0"))
    ),
1248
1249
1250
    # 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.
1251
    "VLLM_XGRAMMAR_CACHE_MB": lambda: int(os.getenv("VLLM_XGRAMMAR_CACHE_MB", "512")),
1252
1253
1254
1255
1256
1257
1258
    # 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.
1259
1260
1261
    "VLLM_MSGPACK_ZERO_COPY_THRESHOLD": lambda: int(
        os.getenv("VLLM_MSGPACK_ZERO_COPY_THRESHOLD", "256")
    ),
1262
1263
1264
    # 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.
1265
1266
1267
    "VLLM_ALLOW_INSECURE_SERIALIZATION": lambda: bool(
        int(os.getenv("VLLM_ALLOW_INSECURE_SERIALIZATION", "0"))
    ),
Robert Shaw's avatar
Robert Shaw committed
1268
    # IP address used for NIXL handshake between remote agents.
1269
1270
1271
    "VLLM_NIXL_SIDE_CHANNEL_HOST": lambda: os.getenv(
        "VLLM_NIXL_SIDE_CHANNEL_HOST", "localhost"
    ),
Robert Shaw's avatar
Robert Shaw committed
1272
    # Port used for NIXL handshake between remote agents.
1273
1274
1275
    "VLLM_NIXL_SIDE_CHANNEL_PORT": lambda: int(
        os.getenv("VLLM_NIXL_SIDE_CHANNEL_PORT", "5600")
    ),
1276
1277
1278
1279
    # Port used for Mooncake handshake between remote agents.
    "VLLM_MOONCAKE_BOOTSTRAP_PORT": lambda: int(
        os.getenv("VLLM_MOONCAKE_BOOTSTRAP_PORT", "8998")
    ),
1280
1281
    # [DEPRECATED - will be removed in v0.15.0] all2all backend for vllm's
    # expert parallel communication. Use --all2all-backend CLI argument instead.
1282
    # Available options:
1283
1284
1285
    # - "naive": naive all2all implementation using broadcasts
    # - "allgather_reducescatter": all2all implementation based on allgather and
    #  reducescatter
1286
    # - "pplx": use pplx kernels
1287
1288
    # - "deepep_high_throughput", use deepep high-throughput kernels
    # - "deepep_low_latency", use deepep low-latency kernels
1289
    # - "flashinfer_all2allv", use flashinfer alltoallv kernels for mnnvl
1290
1291
    "VLLM_ALL2ALL_BACKEND": env_with_choices(
        "VLLM_ALL2ALL_BACKEND",
1292
        None,
1293
1294
1295
1296
1297
1298
1299
1300
1301
        [
            "naive",
            "pplx",
            "deepep_high_throughput",
            "deepep_low_latency",
            "allgather_reducescatter",
            "flashinfer_all2allv",
        ],
    ),
1302
1303
    # Flashinfer MoE backend for vLLM's fused Mixture-of-Experts support.
    # Both require compute capability 10.0 or above.
1304
1305
1306
1307
1308
    # Available options:
    # - "throughput":  [default]
    #     Uses CUTLASS kernels optimized for high-throughput batch inference.
    # - "latency":
    #     Uses TensorRT-LLM kernels optimized for low-latency inference.
1309
    "VLLM_FLASHINFER_MOE_BACKEND": env_with_choices(
1310
1311
1312
        "VLLM_FLASHINFER_MOE_BACKEND",
        "latency",
        ["throughput", "latency", "masked_gemm"],
1313
    ),
1314
1315
1316
1317
    # 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))
    ),
1318
1319
1320
1321
    # 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.
1322
1323
1324
    "VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE": lambda: int(
        os.getenv("VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE", "163840")
    ),
1325
1326
1327
1328
    # 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> }
1329
    # Unspecified world sizes will fall back to
1330
    #     { 2: 64, 4: 1, <everything else>: 0.5 }
1331
1332
1333
    "VLLM_FLASHINFER_ALLREDUCE_FUSION_THRESHOLDS_MB": lambda: json.loads(
        os.getenv("VLLM_FLASHINFER_ALLREDUCE_FUSION_THRESHOLDS_MB", "{}")
    ),
1334
1335
1336
    # MoE routing strategy selector.
    # See `RoutingSimulator.get_available_strategies()` # for available
    # strategies.
1337
    # Custom routing strategies can be registered by
1338
1339
    # RoutingSimulator.register_strategy()
    # Note: custom strategies may not produce correct model outputs
1340
1341
1342
    "VLLM_MOE_ROUTING_SIMULATION_STRATEGY": lambda: os.environ.get(
        "VLLM_MOE_ROUTING_SIMULATION_STRATEGY", ""
    ).lower(),
1343
    # Regex timeout for use by the vLLM tool parsing plugins.
1344
1345
1346
    "VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS": lambda: int(
        os.getenv("VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS", "1")
    ),
1347
1348
    # Reduce CPU usage when vLLM is idle. Enabling this will incur small
    # latency penalty when a request eventually comes.
1349
    "VLLM_SLEEP_WHEN_IDLE": lambda: bool(int(os.getenv("VLLM_SLEEP_WHEN_IDLE", "0"))),
1350
1351
1352
    # 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.
1353
1354
1355
    "VLLM_MQ_MAX_CHUNK_BYTES_MB": lambda: int(
        os.getenv("VLLM_MQ_MAX_CHUNK_BYTES_MB", "16")
    ),
1356
1357
    # Timeout in seconds for execute_model RPC calls in multiprocessing
    # executor (only applies when TP > 1).
1358
1359
1360
    "VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS": lambda: int(
        os.getenv("VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS", "300")
    ),
1361
1362
1363
1364
1365
1366
1367
    # 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.
1368
1369
1370
    "VLLM_KV_CACHE_LAYOUT": env_with_choices(
        "VLLM_KV_CACHE_LAYOUT", None, ["NHD", "HND"]
    ),
1371
1372
1373
    # 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.
1374
1375
1376
    "VLLM_COMPUTE_NANS_IN_LOGITS": lambda: bool(
        int(os.getenv("VLLM_COMPUTE_NANS_IN_LOGITS", "0"))
    ),
1377
1378
1379
    # Controls whether or not emulations are used for NVFP4
    # generations on machines < 100 for compressed-tensors
    # models
1380
1381
1382
    "VLLM_USE_NVFP4_CT_EMULATIONS": lambda: bool(
        int(os.getenv("VLLM_USE_NVFP4_CT_EMULATIONS", "0"))
    ),
1383
1384
1385
1386
    # 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.
1387
1388
1389
    "VLLM_NIXL_ABORT_REQUEST_TIMEOUT": lambda: int(
        os.getenv("VLLM_NIXL_ABORT_REQUEST_TIMEOUT", "480")
    ),
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
    # 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")),
1404
1405
1406
1407
    # Timeout (in seconds) for MooncakeConnector in PD disaggregated setup.
    "VLLM_MOONCAKE_ABORT_REQUEST_TIMEOUT": lambda: int(
        os.getenv("VLLM_MOONCAKE_ABORT_REQUEST_TIMEOUT", "480")
    ),
1408
    # Controls whether or not to use cudnn prefill
1409
1410
1411
    "VLLM_USE_CUDNN_PREFILL": lambda: bool(
        int(os.getenv("VLLM_USE_CUDNN_PREFILL", "0"))
    ),
1412
1413
1414
1415
    # 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"))
    ),
1416
1417
1418
    # 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.
1419
1420
1421
1422
1423
    "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")
    ),
1424
    # If set to 1, when we use fp8 kv, we do not quantize Q to fp8
1425
1426
1427
    "VLLM_FLASHINFER_DISABLE_Q_QUANTIZATION": lambda: bool(
        int(os.getenv("VLLM_FLASHINFER_DISABLE_Q_QUANTIZATION", "0"))
    ),
1428
1429
    # If set, it means we pre-downloaded cubin files and flashinfer will
    # read the cubin files directly.
1430
1431
1432
    "VLLM_HAS_FLASHINFER_CUBIN": lambda: bool(
        int(os.getenv("VLLM_HAS_FLASHINFER_CUBIN", "0"))
    ),
1433
1434
1435
1436
1437
1438
1439
1440
    # 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,
1441
        ["flashinfer-cudnn", "flashinfer-trtllm", "flashinfer-cutlass", "cutlass"],
1442
    ),
1443
1444
1445
    # 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.
1446
1447
1448
    "VLLM_ENABLE_CUDAGRAPH_GC": lambda: bool(
        int(os.getenv("VLLM_ENABLE_CUDAGRAPH_GC", "0"))
    ),
1449
    # Used to force set up loopback IP
1450
    "VLLM_LOOPBACK_IP": lambda: os.getenv("VLLM_LOOPBACK_IP", ""),
1451
1452
1453
    # Used to set the process name prefix for vLLM processes.
    # This is useful for debugging and monitoring purposes.
    # The default value is "VLLM".
1454
    "VLLM_PROCESS_NAME_PREFIX": lambda: os.getenv("VLLM_PROCESS_NAME_PREFIX", "VLLM"),
1455
1456
1457
1458
1459
1460
1461
    # 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.
1462
    "VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE": lambda: bool(
1463
        int(os.getenv("VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE", "1"))
1464
    ),
1465
1466
    # 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
1467
1468
    # messages for those requests in memory. By default, this is disabled (0),
    # and the "store" option is ignored.
1469
1470
1471
1472
1473
    # 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.
1474
1475
1476
    "VLLM_ENABLE_RESPONSES_API_STORE": lambda: bool(
        int(os.getenv("VLLM_ENABLE_RESPONSES_API_STORE", "0"))
    ),
xiao-llm's avatar
xiao-llm committed
1477
    # If set, use the fp8 mfma in rocm paged attention.
1478
1479
1480
    "VLLM_ROCM_FP8_MFMA_PAGE_ATTN": lambda: bool(
        int(os.getenv("VLLM_ROCM_FP8_MFMA_PAGE_ATTN", "0"))
    ),
1481
    # Whether to use pytorch symmetric memory for allreduce
1482
    "VLLM_ALLREDUCE_USE_SYMM_MEM": lambda: bool(
1483
        int(os.getenv("VLLM_ALLREDUCE_USE_SYMM_MEM", "1"))
1484
    ),
1485
1486
1487
1488
    # 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"))
    ),
1489
    # Allows vllm to find tuned config under customized folder
1490
    "VLLM_TUNED_CONFIG_FOLDER": lambda: os.getenv("VLLM_TUNED_CONFIG_FOLDER", None),
1491
1492
1493
1494
1495
1496
1497
1498
1499
    # 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"],
    ),
1500
    # Allows harmony instructions to be injected on system messages
1501
1502
1503
    "VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS": lambda: bool(
        int(os.getenv("VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS", "0"))
    ),
1504
1505
1506
1507
1508
1509
    # 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"))
    ),
1510
    # Add optional custom scopes for profiling, disable to avoid overheads
1511
1512
1513
    "VLLM_CUSTOM_SCOPES_FOR_PROFILING": lambda: bool(
        int(os.getenv("VLLM_CUSTOM_SCOPES_FOR_PROFILING", "0"))
    ),
1514
    # Add optional nvtx scopes for profiling, disable to avoid overheads
1515
1516
1517
    "VLLM_NVTX_SCOPES_FOR_PROFILING": lambda: bool(
        int(os.getenv("VLLM_NVTX_SCOPES_FOR_PROFILING", "0"))
    ),
1518
1519
    # Represent block hashes in KV cache events as 64-bit integers instead of
    # raw bytes. Defaults to True for backward compatibility.
1520
1521
1522
    "VLLM_KV_EVENTS_USE_INT_BLOCK_HASHES": lambda: bool(
        int(os.getenv("VLLM_KV_EVENTS_USE_INT_BLOCK_HASHES", "1"))
    ),
1523
1524
    # Name of the shared memory buffer used for object storage.
    # Only effective when mm_config.mm_processor_cache_type == "shm".
1525
1526
1527
    "VLLM_OBJECT_STORAGE_SHM_BUFFER_NAME": lambda: os.getenv(
        "VLLM_OBJECT_STORAGE_SHM_BUFFER_NAME", "VLLM_OBJECT_STORAGE_SHM_BUFFER"
    ),
1528
    # The size in MB of the buffers (NVL and RDMA) used by DeepEP
1529
1530
1531
    "VLLM_DEEPEP_BUFFER_SIZE_MB": lambda: int(
        os.getenv("VLLM_DEEPEP_BUFFER_SIZE_MB", "1024")
    ),
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
    # 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"))
    ),
1543
1544
    # 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
1545
    "VLLM_DBO_COMM_SMS": lambda: int(os.getenv("VLLM_DBO_COMM_SMS", "20")),
1546
1547
1548
    # 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)
1549
1550
1551
    "VLLM_ENABLE_INDUCTOR_MAX_AUTOTUNE": lambda: bool(
        int(os.getenv("VLLM_ENABLE_INDUCTOR_MAX_AUTOTUNE", "1"))
    ),
1552
1553
    # If set to 1, enable coordinate_descent_tuning;
    # By default, this is enabled (1)
1554
1555
1556
    "VLLM_ENABLE_INDUCTOR_COORDINATE_DESCENT_TUNING": lambda: bool(
        int(os.getenv("VLLM_ENABLE_INDUCTOR_COORDINATE_DESCENT_TUNING", "1"))
    ),
1557
    # Flag to enable NCCL symmetric memory allocation and registration
1558
1559
1560
    "VLLM_USE_NCCL_SYMM_MEM": lambda: bool(
        int(os.getenv("VLLM_USE_NCCL_SYMM_MEM", "0"))
    ),
1561
    # NCCL header path
1562
    "VLLM_NCCL_INCLUDE_PATH": lambda: os.environ.get("VLLM_NCCL_INCLUDE_PATH", None),
1563
1564
    # Flag to enable FBGemm kernels on model execution
    "VLLM_USE_FBGEMM": lambda: bool(int(os.getenv("VLLM_USE_FBGEMM", "0"))),
1565
1566
1567
1568
1569
1570
    # 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", ""),
1571
1572
1573
    # Debug workspace allocations.
    # logging of workspace resize operations.
    "VLLM_DEBUG_WORKSPACE": lambda: bool(int(os.getenv("VLLM_DEBUG_WORKSPACE", "0"))),
1574
    # Disables parallel execution of shared_experts via separate cuda stream
1575
1576
    "VLLM_DISABLE_SHARED_EXPERTS_STREAM": lambda: bool(
        int(os.getenv("VLLM_DISABLE_SHARED_EXPERTS_STREAM", "0"))
1577
    ),
1578
1579
1580
1581
1582
1583
1584
    # 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))
    ),
1585
1586
1587
1588
1589
1590
1591
1592
1593
    # 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
1594
1595
1596
1597
    # Flag to enable v2 model runner.
    "VLLM_USE_V2_MODEL_RUNNER": lambda: bool(
        int(os.getenv("VLLM_USE_V2_MODEL_RUNNER", "0"))
    ),
1598
1599
1600
1601
    # Debug logging for --enable-mfu-metrics
    "VLLM_DEBUG_MFU_METRICS": lambda: bool(
        int(os.getenv("VLLM_DEBUG_MFU_METRICS", "0"))
    ),
1602
1603
}

1604
# --8<-- [end:env-vars-definition]
1605

1606

1607
def __getattr__(name: str):
1608
1609
1610
1611
1612
1613
    """
    Gets environment variables lazily.

    NOTE: After enable_envs_cache() invocation (which triggered after service
    initialization), all environment variables will be cached.
    """
1614
1615
1616
1617
1618
    if name in environment_variables:
        return environment_variables[name]()
    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")


1619
1620
1621
1622
1623
1624
def _is_envs_cache_enabled() -> bool:
    """Checked if __getattr__ is wrapped with functools.cache"""
    global __getattr__
    return hasattr(__getattr__, "cache_clear")


1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
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.
    """
1635
1636
1637
    if _is_envs_cache_enabled():
        # Avoid wrapping functools.cache multiple times
        return
1638
1639
1640
1641
1642
1643
1644
1645
1646
    # Tag __getattr__ with functools.cache
    global __getattr__
    __getattr__ = functools.cache(__getattr__)

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


1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
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():
        __getattr__ = __getattr__.__wrapped__


1658
1659
def __dir__():
    return list(environment_variables.keys())
1660
1661
1662
1663
1664
1665
1666
1667
1668


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}")


1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
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",
1691
        "VLLM_FORCE_AOT_LOAD",
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
        "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
1702
        "VLLM_LOGGING_COLOR",
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
        "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_SLEEP_WHEN_IDLE",
        "VLLM_IMAGE_FETCH_TIMEOUT",
        "VLLM_VIDEO_FETCH_TIMEOUT",
        "VLLM_AUDIO_FETCH_TIMEOUT",
        "VLLM_MEDIA_URL_ALLOW_REDIRECTS",
        "VLLM_MEDIA_LOADING_THREAD_COUNT",
        "VLLM_MAX_AUDIO_CLIP_FILESIZE_MB",
        "VLLM_VIDEO_LOADER_BACKEND",
        "VLLM_MEDIA_CONNECTOR",
        "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_OMP_THREADS_BIND",
        "VLLM_CPU_NUM_OF_RESERVED_CPU",
        "VLLM_CPU_MOE_PREPACK",
        "VLLM_CPU_SGL_KERNEL",
        "VLLM_TEST_FORCE_LOAD_FORMAT",
        "LOCAL_RANK",
        "CUDA_VISIBLE_DEVICES",
Nick Hill's avatar
Nick Hill committed
1732
        "NO_COLOR",
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
    }

    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
1751

1752
        factors[factor] = normalize_value(raw)
1753

1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
    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",
    ]

1776
1777
    for var in ray_noset_env_vars:
        factors[var] = normalize_value(os.getenv(var))
1778

1779
    return factors