envs.py 92.6 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

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

if TYPE_CHECKING:
    VLLM_HOST_IP: str = ""
16
    VLLM_PORT: int | None = None
17
    VLLM_RPC_BASE_PATH: str = tempfile.gettempdir()
18
    VLLM_USE_MODELSCOPE: bool = False
19
    VLLM_RINGBUFFER_WARNING_INTERVAL: int = 60
20
21
    VLLM_NCCL_SO_PATH: str | None = None
    LD_LIBRARY_PATH: str | None = None
22
    VLLM_ROCM_SLEEP_MEM_CHUNK_SIZE: int = 256
23
    LOCAL_RANK: int = 0
24
    CUDA_VISIBLE_DEVICES: str | None = None
25
    VLLM_ENGINE_ITERATION_TIMEOUT_S: int = 60
26
    VLLM_ENGINE_READY_TIMEOUT_S: int = 600
27
    VLLM_API_KEY: str | None = None
28
    VLLM_DEBUG_LOG_API_SERVER_RESPONSE: bool = False
29
30
31
32
    S3_ACCESS_KEY_ID: str | None = None
    S3_SECRET_ACCESS_KEY: str | None = None
    S3_ENDPOINT_URL: str | None = None
    VLLM_MODEL_REDIRECT_PATH: str | None = None
33
34
    VLLM_CACHE_ROOT: str = os.path.expanduser("~/.cache/vllm")
    VLLM_CONFIG_ROOT: str = os.path.expanduser("~/.config/vllm")
35
36
37
38
    VLLM_USAGE_STATS_SERVER: str = "https://stats.vllm.ai"
    VLLM_NO_USAGE_STATS: bool = False
    VLLM_DO_NOT_TRACK: bool = False
    VLLM_USAGE_SOURCE: str = ""
39
    VLLM_CONFIGURE_LOGGING: bool = True
40
    VLLM_LOGGING_LEVEL: str = "INFO"
41
    VLLM_LOGGING_PREFIX: str = ""
42
    VLLM_LOGGING_STREAM: str = "ext://sys.stdout"
43
    VLLM_LOGGING_CONFIG_PATH: str | None = None
Nick Hill's avatar
Nick Hill committed
44
45
    VLLM_LOGGING_COLOR: str = "auto"
    NO_COLOR: bool = False
46
    VLLM_LOG_STATS_INTERVAL: float = 10.0
47
    VLLM_TRACE_FUNCTION: int = 0
48
49
50
    VLLM_USE_FLASHINFER_SAMPLER: bool | None = None
    VLLM_PP_LAYER_PARTITION: str | None = None
    VLLM_CPU_KVCACHE_SPACE: int | None = 0
51
    VLLM_CPU_OMP_THREADS_BIND: str = ""
52
    VLLM_CPU_NUM_OF_RESERVED_CPU: int | None = None
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 = 16 * 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"] = "spawn"
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_HASHER_ALGORITHM: str = "blake3"
74
    VLLM_TARGET_DEVICE: str = "cuda"
75
    VLLM_MAIN_CUDA_VERSION: str = "12.9"
76
    VLLM_FLOAT32_MATMUL_PRECISION: Literal["highest", "high", "medium"] = "highest"
77
78
    MAX_JOBS: str | None = None
    NVCC_THREADS: str | None = None
79
    VLLM_USE_PRECOMPILED: bool = False
80
    VLLM_SKIP_PRECOMPILED_VERSION_SUFFIX: bool = False
81
    VLLM_DOCKER_BUILD_CONTEXT: bool = False
82
    VLLM_KEEP_ALIVE_ON_ENGINE_DEATH: bool = False
83
    CMAKE_BUILD_TYPE: Literal["Debug", "Release", "RelWithDebInfo"] | None = None
84
    VERBOSE: bool = False
85
    VLLM_ALLOW_LONG_MAX_MODEL_LEN: bool = False
86
    VLLM_RPC_TIMEOUT: int = 10000  # ms
87
    VLLM_HTTP_TIMEOUT_KEEP_ALIVE: int = 5  # seconds
88
89
    VLLM_PLUGINS: list[str] | None = None
    VLLM_LORA_RESOLVER_CACHE_DIR: str | None = None
90
91
92
    # 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
93
    VLLM_TORCH_PROFILER_DIR: str | None = None
94
95
96
97
98
99
100
101
102
103
    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
104
    VLLM_USE_AOT_COMPILE: bool = False
105
    VLLM_USE_BYTECODE_HOOK: bool = False
106
    VLLM_FORCE_AOT_LOAD: bool = False
107
    VLLM_USE_MEGA_AOT_ARTIFACT: bool = False
108
    VLLM_USE_TRITON_AWQ: bool = False
109
    VLLM_ALLOW_RUNTIME_LORA_UPDATING: bool = False
110
    VLLM_SKIP_P2P_CHECK: bool = False
111
    VLLM_DISABLED_KERNELS: list[str] = []
112
    VLLM_DISABLE_PYNCCL: bool = False
113
    VLLM_ROCM_USE_AITER: bool = False
114
    VLLM_ROCM_USE_AITER_PAGED_ATTN: bool = False
115
    VLLM_ROCM_USE_AITER_LINEAR: bool = True
116
    VLLM_ROCM_USE_AITER_MOE: bool = True
117
    VLLM_ROCM_USE_AITER_RMSNORM: bool = True
118
    VLLM_ROCM_USE_AITER_MLA: bool = True
119
    VLLM_ROCM_USE_AITER_MHA: bool = True
120
    VLLM_ROCM_USE_AITER_FP4_ASM_GEMM: bool = False
121
    VLLM_ROCM_USE_AITER_TRITON_ROPE: bool = False
122
    VLLM_ROCM_USE_AITER_FP8BMM: bool = True
123
    VLLM_ROCM_USE_AITER_FP4BMM: 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_ROCM_SHUFFLE_KV_CACHE_LAYOUT: bool = False
132
    VLLM_ENABLE_V1_MULTIPROCESSING: bool = True
133
    VLLM_LOG_BATCHSIZE_INTERVAL: float = -1
134
    VLLM_DISABLE_COMPILE_CACHE: bool = False
zhangshao's avatar
zhangshao committed
135
136
137
    Q_SCALE_CONSTANT: int = 10
    K_SCALE_CONSTANT: int = 10
    V_SCALE_CONSTANT: int = 10
138
    VLLM_SERVER_DEV_MODE: bool = False
139
    VLLM_V1_OUTPUT_PROC_CHUNK_SIZE: int = 128
140
    VLLM_MLA_DISABLE: bool = False
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
guanyu1's avatar
guanyu1 committed
159
    VLLM_1D_MROPE: bool = False
160
    VLLM_TPU_BUCKET_PADDING_GAP: int = 0
161
    VLLM_TPU_MOST_MODEL_LEN: int | None = None
162
    VLLM_TPU_USING_PATHWAYS: bool = False
163
    VLLM_USE_DEEP_GEMM: bool = True
164
    VLLM_MOE_USE_DEEP_GEMM: bool = True
165
    VLLM_USE_DEEP_GEMM_E8M0: bool = True
166
    VLLM_USE_DEEP_GEMM_TMA_ALIGNED_SCALES: bool = True
167
168
169
170
171
    VLLM_DEEP_GEMM_WARMUP: Literal[
        "skip",
        "full",
        "relax",
    ] = "relax"
172
    VLLM_USE_FUSED_MOE_GROUPED_TOPK: bool = True
173
    VLLM_BLOCKSCALE_FP8_GEMM_FLASHINFER: bool = False
174
    VLLM_USE_FLASHINFER_MOE_FP16: bool = False
175
176
    VLLM_USE_FLASHINFER_MOE_FP8: bool = False
    VLLM_USE_FLASHINFER_MOE_FP4: bool = False
177
178
179
    VLLM_FLASHINFER_MOE_BACKEND: Literal["throughput", "latency", "masked_gemm"] = (
        "latency"
    )
180
    VLLM_FLASHINFER_WORKSPACE_BUFFER_SIZE: int = 394 * 1024 * 1024
181
    VLLM_XGRAMMAR_CACHE_MB: int = 0
182
    VLLM_MSGPACK_ZERO_COPY_THRESHOLD: int = 256
183
    VLLM_ALLOW_INSECURE_SERIALIZATION: bool = False
Robert Shaw's avatar
Robert Shaw committed
184
    VLLM_NIXL_SIDE_CHANNEL_HOST: str = "localhost"
185
    VLLM_NIXL_SIDE_CHANNEL_PORT: int = 5600
186
    VLLM_MOONCAKE_BOOTSTRAP_PORT: int = 8998
187
188
189
190
191
    VLLM_ALL2ALL_BACKEND: Literal[
        "naive",
        "pplx",
        "deepep_high_throughput",
        "deepep_low_latency",
192
        "mori",
193
194
195
        "allgather_reducescatter",
        "flashinfer_all2allv",
    ] = "allgather_reducescatter"
196
    VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE: int = 163840
197
    VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS: int = 1
198
    VLLM_SLEEP_WHEN_IDLE: bool = False
199
    VLLM_MQ_MAX_CHUNK_BYTES_MB: int = 16
200
    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: int = 300
201
    VLLM_KV_CACHE_LAYOUT: Literal["NHD", "HND"] | None = None
202
    VLLM_COMPUTE_NANS_IN_LOGITS: bool = False
203
    VLLM_USE_NVFP4_CT_EMULATIONS: bool = False
204
205
206
    VLLM_ROCM_QUICK_REDUCE_QUANTIZATION: Literal[
        "FP", "INT8", "INT6", "INT4", "NONE"
    ] = "NONE"
207
    VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16: bool = True
208
    VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB: int | None = None
209
    VLLM_NIXL_ABORT_REQUEST_TIMEOUT: int = 480
210
211
212
213
    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
214
    VLLM_MOONCAKE_ABORT_REQUEST_TIMEOUT: int = 480
215
    VLLM_ENABLE_CUDAGRAPH_GC: bool = False
216
    VLLM_LOOPBACK_IP: str = ""
217
    VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE: bool = True
218
    VLLM_ENABLE_RESPONSES_API_STORE: bool = False
219
    VLLM_NVFP4_GEMM_BACKEND: str | None = None
220
    VLLM_HAS_FLASHINFER_CUBIN: bool = False
221
222
    VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8: bool = False
    VLLM_USE_FLASHINFER_MOE_MXFP4_BF16: bool = False
xiao-llm's avatar
xiao-llm committed
223
    VLLM_ROCM_FP8_MFMA_PAGE_ATTN: bool = False
224
    VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS: bool = False
225
    VLLM_ALLREDUCE_USE_SYMM_MEM: bool = True
226
    VLLM_TUNED_CONFIG_FOLDER: str | None = None
227
    VLLM_GPT_OSS_SYSTEM_TOOL_MCP_LABELS: set[str] = set()
228
    VLLM_USE_EXPERIMENTAL_PARSER_CONTEXT: bool = False
229
    VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS: bool = False
230
    VLLM_TOOL_JSON_ERROR_AUTOMATIC_RETRY: bool = False
231
    VLLM_CUSTOM_SCOPES_FOR_PROFILING: bool = False
232
    VLLM_NVTX_SCOPES_FOR_PROFILING: bool = False
233
    VLLM_KV_EVENTS_USE_INT_BLOCK_HASHES: bool = True
234
    VLLM_OBJECT_STORAGE_SHM_BUFFER_NAME: str = "VLLM_OBJECT_STORAGE_SHM_BUFFER"
235
    VLLM_DEEPEP_BUFFER_SIZE_MB: int = 1024
236
237
    VLLM_DEEPEP_HIGH_THROUGHPUT_FORCE_INTRA_NODE: bool = False
    VLLM_DEEPEP_LOW_LATENCY_USE_MNNVL: bool = False
238
    VLLM_DBO_COMM_SMS: int = 20
239
240
    VLLM_PATTERN_MATCH_DEBUG: str | None = None
    VLLM_DEBUG_DUMP_PATH: str | None = None
241
242
    VLLM_ENABLE_INDUCTOR_MAX_AUTOTUNE: bool = True
    VLLM_ENABLE_INDUCTOR_COORDINATE_DESCENT_TUNING: bool = True
243
    VLLM_USE_NCCL_SYMM_MEM: bool = False
244
    VLLM_NCCL_INCLUDE_PATH: str | None = None
245
    VLLM_USE_FBGEMM: bool = False
246
    VLLM_GC_DEBUG: str = ""
247
    VLLM_DEBUG_WORKSPACE: bool = False
248
    VLLM_DISABLE_SHARED_EXPERTS_STREAM: bool = False
249
    VLLM_SHARED_EXPERTS_STREAM_TOKEN_THRESHOLD: int = 256
250
    VLLM_COMPILE_CACHE_SAVE_FORMAT: Literal["binary", "unpacked"] = "binary"
Woosuk Kwon's avatar
Woosuk Kwon committed
251
    VLLM_USE_V2_MODEL_RUNNER: bool = False
252
    VLLM_LOG_MODEL_INSPECTION: bool = False
253
    VLLM_DEBUG_MFU_METRICS: bool = False
254
    VLLM_DISABLE_LOG_LOGO: bool = False
255
    VLLM_LORA_DISABLE_PDL: bool = False
256
    
257
    # add envs
zhuwenwen's avatar
zhuwenwen committed
258
    VLLM_OPTEST_URLS_PORT: int | None = None
259
260
    VLLM_OPTEST_MODELS_PATH: str = ""
    VLLM_USE_TRITON_PREFIX_FLASH_ATTN: bool = False
261
    VLLM_USE_FLASH_ATTN_FP8: bool = False
262
    VLLM_USE_QUERY_QUANT: bool = False
263
264
265
266
267
268
    VLLM_USE_FLASH_MLA: bool = False
    VLLM_USE_OPT_OP: bool = False
    VLLM_USE_TC_PAGED_ATTN: bool = False
    VLLM_USE_PA_PRINT_PARAM: bool = False 
    VLLM_SPEC_DECODE_EAGER: bool = False
    VLLM_PCIE_USE_CUSTOM_ALLREDUCE: bool = False
269
    VLLM_CUSTOM_CACHE: bool = False
zhuwenwen's avatar
zhuwenwen committed
270
    VLLM_CUSTOM_ALLREDUCE_SUPPORTED_WORLDSIZE_MAX: int = 16
zhuwenwen's avatar
zhuwenwen committed
271
    VLLM_ENFORCE_EAGER_BS_THRESHOLD: int | None  = None
272
    VLLM_HAS_CONTEXT_DEFAULT: bool = False
273
    VLLM_USE_NN: bool = False
274
    VLLM_ENABLE_TBO: bool = False
275
    VLLM_ENABLE_MOE_FUSED_GATE: bool = False
276
    VLLM_USE_FLASH_ATTN_PA: bool = False
zhuwenwen's avatar
zhuwenwen committed
277
    VLLM_USE_APEX_RN: bool = False
278
    VLLM_USE_GLOBAL_CACHE13: bool = False
279
280
    VLLM_USE_LIGHTOP: bool = False
    VLLM_USE_OPT_CAT: bool = False
zhuwenwen's avatar
zhuwenwen committed
281
282
    VLLM_USE_LIGHTOP_MOE_SUM: bool = False
    VLLM_USE_LIGHTOP_MOE_ALIGN: bool = False
283
    VLLM_USE_MERGE_ATTN_STATES_OPT: bool = False
王敏's avatar
王敏 committed
284
    USE_FUSED_RMS_QUANT: bool = False
285
    USE_FUSED_SILU_MUL_QUANT: bool = False
zhuwenwen's avatar
zhuwenwen committed
286
    VLLM_USE_PD_SPLIT: bool = False
zhuwenwen's avatar
zhuwenwen committed
287
    VLLM_USE_PP_SYNC: bool = False
288
    VLLM_USE_PIECEWISE: bool = False
289
    VLLM_USE_V32_ENCODE: bool = False
290
291
292
    VLLM_USE_FUSE_SILU_AND_MUL: bool = False
    VLLM_USE_OPT_RESHAPE_AND_CACHE: bool = False
    VLLM_USE_TOPK_RENORM: bool = False
293
    VLLM_USE_FUSED_RMS_ROPE: bool = False
294
    VLLM_USE_FUSED_FILL_RMS_CAT: bool = False
295
    VLLM_W8A8_BACKEND: int = 3
jujl1's avatar
jujl1 committed
296
    VLLM_USE_PP_BALANCE = True
297
298
299
300
301
302
    VLLM_MOE_ROUTER_CAPTURE: bool = False
    VLLM_MOE_ROUTER_CAPTURE_DIR: str = "/tmp"
    VLLM_MOE_ROUTER_CAPTURE_RANK: int = -1
    VLLM_MOE_ROUTER_CAPTURE_MAX_LAYERS: int = 0
    VLLM_MOE_ROUTER_CAPTURE_NUM_TOKENS_GT: int = -1
    VLLM_MOE_ROUTER_CAPTURE_NUM_TOKENS_LT: int = -1
王敏's avatar
王敏 committed
303
    VLLM_REJECT_SAMPLE_OPT: bool = False
304
    VLLM_USE_MOE_W16A16_TRITON: bool = False
305
    VLLM_V1_FAST_TOKEN_ID_COPY: bool = False
306
    VLLM_V1_USE_REDUCED_TOPK_TOPP_SAMPLER: bool = False
307
    VLLM_V1_USE_FUSED_QKV_SPLIT_RMS_ROPE_KVSTORE: bool = False
308
    VLLM_USE_FUSED_DTBMM: bool = False # DOUBLE TRANS BMM FP8
309
    VLLM_USE_LIGHTOP_FILL_MOE_ALIGN: bool = False
310
    VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT: bool = False
wujl5's avatar
wujl5 committed
311
    VLLM_USE_CUDA_GRAPH_SIZES: bool = False
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327


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"),
    )


328
def maybe_convert_int(value: str | None) -> int | None:
329
330
331
332
333
    if value is None:
        return None
    return int(value)


334
def maybe_convert_bool(value: str | None) -> bool | None:
335
336
337
338
339
    if value is None:
        return None
    return bool(int(value))


340
341
342
343
def disable_compile_cache() -> bool:
    return bool(int(os.getenv("VLLM_DISABLE_COMPILE_CACHE", "0")))


344
def use_aot_compile() -> bool:
345
346
347
    from vllm.model_executor.layers.batch_invariant import (
        vllm_is_batch_invariant,
    )
zhuwenwen's avatar
zhuwenwen committed
348
    from vllm.platforms import current_platform
349
    from vllm.utils.torch_utils import is_torch_equal_or_newer
350

351
352
    default_value = (
        "1"
zhuwenwen's avatar
zhuwenwen committed
353
354
355
356
357
        if is_torch_equal_or_newer("2.10.0.dev")
        and not disable_compile_cache()
        # Disabling AOT_COMPILE for CPU
        # See: https://github.com/vllm-project/vllm/issues/32033
        and not current_platform.is_cpu()
358
359
360
        else "0"
    )

361
362
363
364
    return (
        not vllm_is_batch_invariant()
        and os.environ.get("VLLM_USE_AOT_COMPILE", default_value) == "1"
    )
365
366


367
def env_with_choices(
368
    env_name: str,
369
370
    default: str | None,
    choices: list[str] | Callable[[], list[str]],
371
    case_sensitive: bool = True,
372
) -> Callable[[], str | None]:
373
374
    """
    Create a lambda that validates environment variable against allowed choices
375

376
377
378
379
380
    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
381

382
383
384
385
    Returns:
        Lambda function for environment_variables dict
    """

386
    def _get_validated_env() -> str | None:
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
        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:
402
403
404
405
            raise ValueError(
                f"Invalid value '{value}' for {env_name}. "
                f"Valid options: {actual_choices}."
            )
406
407
408
409
410
411

        return value

    return _get_validated_env


412
def env_list_with_choices(
413
414
    env_name: str,
    default: list[str],
415
    choices: list[str] | Callable[[], list[str]],
416
417
    case_sensitive: bool = True,
) -> Callable[[], list[str]]:
418
    """
419
    Create a lambda that validates environment variable
420
    containing comma-separated values against allowed choices
421

422
423
424
425
426
    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
427

428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
    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:
457
458
459
460
                raise ValueError(
                    f"Invalid value '{val}' in {env_name}. "
                    f"Valid options: {actual_choices}."
                )
461
462
463
464
465
466

        return values

    return _get_validated_env_list


467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
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


485
def get_vllm_port() -> int | None:
486
    """Get the port from VLLM_PORT environment variable.
487

488
489
    Returns:
        The port number as an integer if VLLM_PORT is set, None otherwise.
490

491
492
493
    Raises:
        ValueError: If VLLM_PORT is a URI, suggest k8s service discovery issue.
    """
494
    if "VLLM_PORT" not in os.environ:
495
496
        return None

497
    port = os.getenv("VLLM_PORT", "0")
498
499
500
501

    try:
        return int(port)
    except ValueError as err:
502
        from urllib3.util import parse_url
503

504
        parsed = parse_url(port)
505
506
507
508
509
510
        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
511
        raise ValueError(f"VLLM_PORT '{port}' must be a valid integer") from err
512
513


514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
def get_env_or_set_default(
    env_name: str,
    default_factory: Callable[[], str],
) -> Callable[[], str]:
    """
    Create a lambda that returns an environment variable value if set,
    or generates and sets a default value using the provided factory function.
    """

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

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

    return _get_or_set_default


Ning Xie's avatar
Ning Xie committed
535
# The start-* and end* here are used by the documentation generator
536
537
# to extract the used env vars.

538
# --8<-- [start:env-vars-definition]
539

540
logger = logging.getLogger(__name__)
541

542
environment_variables: dict[str, Callable[[], Any]] = {
543
    # ================== Installation Time Env Vars ==================
544
    # Target device of vLLM, supporting [cuda (by default),
545
    # rocm, cpu]
546
    "VLLM_TARGET_DEVICE": lambda: os.getenv("VLLM_TARGET_DEVICE", "cuda").lower(),
547
    # Main CUDA version of vLLM. This follows PyTorch but can be overridden.
548
    "VLLM_MAIN_CUDA_VERSION": lambda: os.getenv("VLLM_MAIN_CUDA_VERSION", "").lower()
549
    or "12.9",
550
    # Controls PyTorch float32 matmul precision mode within vLLM workers.
551
    # Valid options mirror torch.set_float32_matmul_precision
552
553
    "VLLM_FLOAT32_MATMUL_PRECISION": env_with_choices(
        "VLLM_FLOAT32_MATMUL_PRECISION",
554
555
        "highest",
        ["highest", "high", "medium"],
556
557
        case_sensitive=False,
    ),
558
559
    # Maximum number of compilation jobs to run in parallel.
    # By default this is the number of CPUs
560
    "MAX_JOBS": lambda: os.getenv("MAX_JOBS", None),
561
562
563
    # Number of threads to use for nvcc
    # By default this is 1.
    # If set, `MAX_JOBS` will be reduced to avoid oversubscribing the CPU.
564
    "NVCC_THREADS": lambda: os.getenv("NVCC_THREADS", None),
565
    # If set, vllm will use precompiled binaries (*.so)
566
567
568
569
570
    "VLLM_USE_PRECOMPILED": lambda: os.environ.get("VLLM_USE_PRECOMPILED", "")
    .strip()
    .lower()
    in ("1", "true")
    or bool(os.environ.get("VLLM_PRECOMPILED_WHEEL_LOCATION")),
571
572
573
574
    # 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"))
    ),
575
576
    # Used to mark that setup.py is running in a Docker build context,
    # in order to force the use of precompiled binaries.
577
578
579
580
    "VLLM_DOCKER_BUILD_CONTEXT": lambda: os.environ.get("VLLM_DOCKER_BUILD_CONTEXT", "")
    .strip()
    .lower()
    in ("1", "true"),
581
582
583
    # CMake build type
    # If not set, defaults to "Debug" or "RelWithDebInfo"
    # Available options: "Debug", "Release", "RelWithDebInfo"
584
585
586
    "CMAKE_BUILD_TYPE": env_with_choices(
        "CMAKE_BUILD_TYPE", None, ["Debug", "Release", "RelWithDebInfo"]
    ),
587
    # If set, vllm will print verbose logs during installation
588
    "VERBOSE": lambda: bool(int(os.getenv("VERBOSE", "0"))),
589
    # Root directory for vLLM configuration files
590
    # Defaults to `~/.config/vllm` unless `XDG_CONFIG_HOME` is set
591
592
593
    # 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**.
594
    "VLLM_CONFIG_ROOT": lambda: os.path.expanduser(
595
596
597
        os.getenv(
            "VLLM_CONFIG_ROOT",
            os.path.join(get_default_config_root(), "vllm"),
598
599
        )
    ),
600
    # ================== Runtime Env Vars ==================
601
    # Root directory for vLLM cache files
602
    # Defaults to `~/.cache/vllm` unless `XDG_CACHE_HOME` is set
603
    "VLLM_CACHE_ROOT": lambda: os.path.expanduser(
604
605
606
        os.getenv(
            "VLLM_CACHE_ROOT",
            os.path.join(get_default_cache_root(), "vllm"),
607
608
        )
    ),
609
610
611
612
    # 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.
613
    "VLLM_HOST_IP": lambda: os.getenv("VLLM_HOST_IP", ""),
614
    # used in distributed environment to manually set the communication port
615
616
617
    # 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.
618
    "VLLM_PORT": get_vllm_port,
619
620
    # path used for ipc when the frontend api server is running in
    # multi-processing mode to communicate with the backend engine process.
621
622
623
    "VLLM_RPC_BASE_PATH": lambda: os.getenv(
        "VLLM_RPC_BASE_PATH", tempfile.gettempdir()
    ),
624
625
    # If true, will load models from ModelScope instead of Hugging Face Hub.
    # note that the value is true or false, not numbers
626
627
628
629
    "VLLM_USE_MODELSCOPE": lambda: os.environ.get(
        "VLLM_USE_MODELSCOPE", "False"
    ).lower()
    == "true",
630
    # Interval in seconds to log a warning message when the ring buffer is full
631
632
633
    "VLLM_RINGBUFFER_WARNING_INTERVAL": lambda: int(
        os.environ.get("VLLM_RINGBUFFER_WARNING_INTERVAL", "60")
    ),
634
635
    # path to cudatoolkit home directory, under which should be bin, include,
    # and lib directories.
636
    "CUDA_HOME": lambda: os.environ.get("CUDA_HOME", None),
637
638
    # 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
639
    "VLLM_NCCL_SO_PATH": lambda: os.environ.get("VLLM_NCCL_SO_PATH", None),
640
641
    # 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`
642
    "LD_LIBRARY_PATH": lambda: os.environ.get("LD_LIBRARY_PATH", None),
643
644
    # flag to control if vllm should use triton flash attention
    "VLLM_USE_TRITON_FLASH_ATTN":
645
    lambda: (os.environ.get("VLLM_USE_TRITON_FLASH_ATTN", "False").lower() in
646
             ("true", "1")),
647
648
649
650
    # 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")
    ),
651
    # Feature flag to enable/disable Inductor standalone compile.
652
653
    # In torch <= 2.7 we ignore this flag; in torch >= 2.9 this is
    # enabled by default.
654
    "VLLM_USE_STANDALONE_COMPILE": lambda: os.environ.get(
655
        "VLLM_USE_STANDALONE_COMPILE", "1"
656
657
    )
    == "1",
658
659
    # Debug pattern matching inside custom passes.
    # Should be set to the fx.Node name (e.g. 'getitem_34' or 'scaled_mm_3').
660
661
662
    "VLLM_PATTERN_MATCH_DEBUG": lambda: os.environ.get(
        "VLLM_PATTERN_MATCH_DEBUG", None
    ),
663
664
    # Dump fx graphs to the given directory.
    # It will override CompilationConfig.debug_dump_path if set.
665
    "VLLM_DEBUG_DUMP_PATH": lambda: os.environ.get("VLLM_DEBUG_DUMP_PATH", None),
666
667
668
669
    # 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,
670
671
672
    # Feature flag to enable/disable bytecode in
    # TorchCompileWithNoGuardsWrapper.
    "VLLM_USE_BYTECODE_HOOK": lambda: bool(
673
        int(os.environ.get("VLLM_USE_BYTECODE_HOOK", "0"))
674
    ),
675
676
677
678
    # 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",
679
680
681
682
683
684
685
    # Enable loading compiled models directly from cached standalone compile artifacts
    # without re-splitting graph modules. This reduces overhead during model
    # loading by using reconstruct_serializable_fn_from_mega_artifact.
    "VLLM_USE_MEGA_AOT_ARTIFACT": lambda: os.environ.get(
        "VLLM_USE_MEGA_AOT_ARTIFACT", "0"
    )
    == "1",
686
687
    # local rank of the process in the distributed setting, used to determine
    # the GPU device id
688
    "LOCAL_RANK": lambda: int(os.environ.get("LOCAL_RANK", "0")),
689
    # used to control the visible devices in the distributed setting
690
    "CUDA_VISIBLE_DEVICES": lambda: os.environ.get("CUDA_VISIBLE_DEVICES", None),
691
    # timeout for each iteration in the engine
692
    "VLLM_ENGINE_ITERATION_TIMEOUT_S": lambda: int(
693
        os.environ.get("VLLM_ENGINE_ITERATION_TIMEOUT_S", "120")
694
    ),
695
696
697
698
699
    # 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")
    ),
700
    # API key for vLLM API server
701
    "VLLM_API_KEY": lambda: os.environ.get("VLLM_API_KEY", None),
702
    # Whether to log responses from API Server for debugging
703
704
705
706
    "VLLM_DEBUG_LOG_API_SERVER_RESPONSE": lambda: os.environ.get(
        "VLLM_DEBUG_LOG_API_SERVER_RESPONSE", "False"
    ).lower()
    == "true",
707
    # S3 access information, used for tensorizer to load model from S3
708
709
710
    "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),
711
    # Usage stats collection
712
713
714
715
716
717
718
719
720
721
722
    "VLLM_USAGE_STATS_SERVER": lambda: os.environ.get(
        "VLLM_USAGE_STATS_SERVER", "https://stats.vllm.ai"
    ),
    "VLLM_NO_USAGE_STATS": lambda: os.environ.get("VLLM_NO_USAGE_STATS", "0") == "1",
    "VLLM_DO_NOT_TRACK": lambda: (
        os.environ.get("VLLM_DO_NOT_TRACK", None)
        or os.environ.get("DO_NOT_TRACK", None)
        or "0"
    )
    == "1",
    "VLLM_USAGE_SOURCE": lambda: os.environ.get("VLLM_USAGE_SOURCE", "production"),
723
724
725
726
    # 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
727
728
729
    "VLLM_CONFIGURE_LOGGING": lambda: bool(
        int(os.getenv("VLLM_CONFIGURE_LOGGING", "1"))
    ),
730
    "VLLM_LOGGING_CONFIG_PATH": lambda: os.getenv("VLLM_LOGGING_CONFIG_PATH"),
731
    # this is used for configuring the default logging level
732
    "VLLM_LOGGING_LEVEL": lambda: os.getenv("VLLM_LOGGING_LEVEL", "INFO").upper(),
733
    # this is used for configuring the default logging stream
734
    "VLLM_LOGGING_STREAM": lambda: os.getenv("VLLM_LOGGING_STREAM", "ext://sys.stdout"),
735
    # if set, VLLM_LOGGING_PREFIX will be prepended to all log messages
736
    "VLLM_LOGGING_PREFIX": lambda: os.getenv("VLLM_LOGGING_PREFIX", ""),
Nick Hill's avatar
Nick Hill committed
737
738
739
740
741
    # 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",
742
743
    # If set, vllm will log stats at this interval in seconds
    # If not set, vllm will log stats every 10 seconds.
744
745
746
    "VLLM_LOG_STATS_INTERVAL": lambda: val
    if (val := float(os.getenv("VLLM_LOG_STATS_INTERVAL", "10."))) > 0.0
    else 10.0,
747
748
749
    # Trace function calls
    # If set to 1, vllm will trace function calls
    # Useful for debugging
750
    "VLLM_TRACE_FUNCTION": lambda: int(os.getenv("VLLM_TRACE_FUNCTION", "0")),
751
    # If set, vllm will use flashinfer sampler
752
753
754
755
756
    "VLLM_USE_FLASHINFER_SAMPLER": lambda: bool(
        int(os.environ["VLLM_USE_FLASHINFER_SAMPLER"])
    )
    if "VLLM_USE_FLASHINFER_SAMPLER" in os.environ
    else None,
757
    # Pipeline stage partition strategy
758
    "VLLM_PP_LAYER_PARTITION": lambda: os.getenv("VLLM_PP_LAYER_PARTITION", None),
759
    # (CPU backend only) CPU key-value cache space.
760
    # default is None and will be set as 4 GB
761
762
763
    "VLLM_CPU_KVCACHE_SPACE": lambda: int(os.getenv("VLLM_CPU_KVCACHE_SPACE", "0"))
    if "VLLM_CPU_KVCACHE_SPACE" in os.environ
    else None,
764
765
    # (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 '|'.
766
    "VLLM_CPU_OMP_THREADS_BIND": lambda: os.getenv("VLLM_CPU_OMP_THREADS_BIND", "auto"),
767
768
    # (CPU backend only) CPU cores not used by OMP threads .
    # Those CPU cores will not be used by OMP threads of a rank.
769
770
771
772
773
    "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,
774
    # (CPU backend only) whether to use SGL kernels, optimized for small batch.
775
    "VLLM_CPU_SGL_KERNEL": lambda: bool(int(os.getenv("VLLM_CPU_SGL_KERNEL", "0"))),
776
777
778
779
780
781
782
    # 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
783
784
785
    "VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE": env_with_choices(
        "VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE", "auto", ["auto", "nccl", "shm"]
    ),
786
    # If the env var is set, it enables GPU communication overlap
787
    # (experimental feature) in Ray's Compiled Graph.
788
789
790
    "VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM": lambda: bool(
        int(os.getenv("VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM", "0"))
    ),
791
792
793
    # 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.
794
795
796
    "VLLM_USE_RAY_WRAPPED_PP_COMM": lambda: bool(
        int(os.getenv("VLLM_USE_RAY_WRAPPED_PP_COMM", "1"))
    ),
797
798
    # Use dedicated multiprocess context for workers.
    # Both spawn and fork work
799
    "VLLM_WORKER_MULTIPROC_METHOD": env_with_choices(
800
        "VLLM_WORKER_MULTIPROC_METHOD", "spawn", ["spawn", "fork"]
801
    ),
802
    # Path to the cache for storing downloaded assets
803
    "VLLM_ASSETS_CACHE": lambda: os.path.expanduser(
804
805
806
        os.getenv(
            "VLLM_ASSETS_CACHE",
            os.path.join(get_default_cache_root(), "vllm", "assets"),
807
808
        )
    ),
809
810
    # If the env var is set, we will clean model file in
    # this path $VLLM_ASSETS_CACHE/model_streamer/$model_name
811
812
813
    "VLLM_ASSETS_CACHE_MODEL_CLEAN": lambda: bool(
        int(os.getenv("VLLM_ASSETS_CACHE_MODEL_CLEAN", "0"))
    ),
814
815
    # Timeout for fetching images when serving multimodal models
    # Default is 5 seconds
816
    "VLLM_IMAGE_FETCH_TIMEOUT": lambda: int(os.getenv("VLLM_IMAGE_FETCH_TIMEOUT", "5")),
817
    # Timeout for fetching videos when serving multimodal models
818
    # Default is 30 seconds
819
820
821
    "VLLM_VIDEO_FETCH_TIMEOUT": lambda: int(
        os.getenv("VLLM_VIDEO_FETCH_TIMEOUT", "30")
    ),
822
    # Timeout for fetching audio when serving multimodal models
823
    # Default is 10 seconds
824
825
826
    "VLLM_AUDIO_FETCH_TIMEOUT": lambda: int(
        os.getenv("VLLM_AUDIO_FETCH_TIMEOUT", "10")
    ),
827
828
    # Whether to allow HTTP redirects when fetching from media URLs.
    # Default to True
829
830
831
    "VLLM_MEDIA_URL_ALLOW_REDIRECTS": lambda: bool(
        int(os.getenv("VLLM_MEDIA_URL_ALLOW_REDIRECTS", "1"))
    ),
832
833
834
    # Max number of workers for the thread pool handling
    # media bytes loading. Set to 1 to disable parallel processing.
    # Default is 8
835
836
837
    "VLLM_MEDIA_LOADING_THREAD_COUNT": lambda: int(
        os.getenv("VLLM_MEDIA_LOADING_THREAD_COUNT", "8")
    ),
838
839
840
    # 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
841
842
843
    "VLLM_MAX_AUDIO_CLIP_FILESIZE_MB": lambda: int(
        os.getenv("VLLM_MAX_AUDIO_CLIP_FILESIZE_MB", "25")
    ),
844
845
    # Backend for Video IO
    # - "opencv": Default backend that uses OpenCV stream buffered backend.
Roger Wang's avatar
Roger Wang committed
846
    # - "identity": Returns raw video bytes for model processor to handle.
847
848
849
850
851
    #
    # 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.
852
853
854
    "VLLM_VIDEO_LOADER_BACKEND": lambda: os.getenv(
        "VLLM_VIDEO_LOADER_BACKEND", "opencv"
    ),
855
856
857
858
859
860
861
862
    # 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"),
863
864
865
866
867
868
869
870
871
872
873
    # Hash algorithm for multimodal content hashing.
    # - "blake3": Default, fast cryptographic hash (not FIPS 140-3 compliant)
    # - "sha256": FIPS 140-3 compliant, widely supported
    # - "sha512": FIPS 140-3 compliant, faster on 64-bit systems
    # Use sha256 or sha512 for FIPS compliance in government/enterprise deployments
    "VLLM_MM_HASHER_ALGORITHM": env_with_choices(
        "VLLM_MM_HASHER_ALGORITHM",
        "blake3",
        ["blake3", "sha256", "sha512"],
        case_sensitive=False,
    ),
874
875
    # Path to the XLA persistent cache directory.
    # Only used for XLA devices such as TPUs.
876
    "VLLM_XLA_CACHE_PATH": lambda: os.path.expanduser(
877
        os.getenv(
878
            "VLLM_XLA_CACHE_PATH",
879
            os.path.join(get_default_cache_root(), "vllm", "xla_cache"),
880
881
        )
    ),
882
    # If set, assert on XLA recompilation after each execution step.
883
884
885
    "VLLM_XLA_CHECK_RECOMPILATION": lambda: bool(
        int(os.getenv("VLLM_XLA_CHECK_RECOMPILATION", "0"))
    ),
886
    # Enable SPMD mode for TPU backend.
887
888
    "VLLM_XLA_USE_SPMD": lambda: bool(int(os.getenv("VLLM_XLA_USE_SPMD", "0"))),
    "VLLM_FUSED_MOE_CHUNK_SIZE": lambda: int(
889
        os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", str(16 * 1024))
890
    ),
891
892
893
    # 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.
894
895
896
    "VLLM_ENABLE_FUSED_MOE_ACTIVATION_CHUNKING": lambda: bool(
        int(os.getenv("VLLM_ENABLE_FUSED_MOE_ACTIVATION_CHUNKING", "1"))
    ),
897
898
    # If set, the OpenAI API server will stay alive even after the underlying
    # AsyncLLMEngine errors and stops serving requests
899
    "VLLM_KEEP_ALIVE_ON_ENGINE_DEATH": lambda: bool(
900
        int(os.getenv("VLLM_KEEP_ALIVE_ON_ENGINE_DEATH", "0"))
901
    ),
902
903
904
905
    # 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.
906
907
908
909
    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": lambda: (
        os.environ.get("VLLM_ALLOW_LONG_MAX_MODEL_LEN", "0").strip().lower()
        in ("1", "true")
    ),
910
911
    # If set, forces FP8 Marlin to be used for FP8 quantization regardless
    # of the hardware support for FP8 compute.
912
913
914
915
916
917
918
    "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"
    ),
919
920
    # Time in ms for the zmq client to wait for a response from the backend
    # server for simple data operations
921
    "VLLM_RPC_TIMEOUT": lambda: int(os.getenv("VLLM_RPC_TIMEOUT", "10000")),
922
    # Timeout in seconds for keeping HTTP connections alive in API server
923
924
925
    "VLLM_HTTP_TIMEOUT_KEEP_ALIVE": lambda: int(
        os.environ.get("VLLM_HTTP_TIMEOUT_KEEP_ALIVE", "5")
    ),
926
927
928
    # 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
929
930
931
    "VLLM_PLUGINS": lambda: None
    if "VLLM_PLUGINS" not in os.environ
    else os.environ["VLLM_PLUGINS"].split(","),
932
933
934
    # a local directory to look in for unrecognized LoRA adapters.
    # only works if plugins are enabled and
    # VLLM_ALLOW_RUNTIME_LORA_UPDATING is enabled.
935
936
937
    "VLLM_LORA_RESOLVER_CACHE_DIR": lambda: os.getenv(
        "VLLM_LORA_RESOLVER_CACHE_DIR", None
    ),
938
939
940
    # Enables torch CUDA profiling if set to 1.
    # Deprecated, see profiler_config.
    "VLLM_TORCH_CUDA_PROFILE": lambda: os.getenv("VLLM_TORCH_CUDA_PROFILE"),
941
    # Enables torch profiler if set.
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
    # 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")
968
969
970
    ),
    # Delay number of iterations before starting profiling when using
    # the torch/torch CUDA profiler. If set to 0, will start profiling immediately.
971
972
    # Deprecated, see profiler_config.
    "VLLM_PROFILER_DELAY_ITERS": lambda: (os.getenv("VLLM_PROFILER_DELAY_ITERS")),
973
974
    # Maximum number of iterations to profile when using the torch/torch CUDA profiler.
    # If set to 0, will not limit the number of iterations.
975
    "VLLM_PROFILER_MAX_ITERS": lambda: os.getenv("VLLM_PROFILER_MAX_ITERS"),
976
    # Control whether torch profiler gzip-compresses profiling files.
977
978
    # Deprecated, see profiler_config.
    "VLLM_TORCH_PROFILER_USE_GZIP": lambda: os.getenv("VLLM_TORCH_PROFILER_USE_GZIP"),
979
    # Control whether torch profiler dumps the self_cuda_time_total table.
980
981
982
983
    # 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")
984
    ),
985
    # If set, vLLM will use Triton implementations of AWQ.
986
    "VLLM_USE_TRITON_AWQ": lambda: bool(int(os.getenv("VLLM_USE_TRITON_AWQ", "0"))),
987
    # If set, allow loading or unloading lora adapters in runtime,
988
989
990
991
    "VLLM_ALLOW_RUNTIME_LORA_UPDATING": lambda: (
        os.environ.get("VLLM_ALLOW_RUNTIME_LORA_UPDATING", "0").strip().lower()
        in ("1", "true")
    ),
992
993
994
995
996
997
    # 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
998
    "VLLM_SKIP_P2P_CHECK": lambda: os.getenv("VLLM_SKIP_P2P_CHECK", "1") == "1",
999
1000
1001
1002
    # List of quantization kernels that should be disabled, used for testing
    # and performance comparisons. Currently only affects MPLinearKernel
    # selection
    # (kernels: MacheteLinearKernel, MarlinLinearKernel, ExllamaLinearKernel)
1003
1004
1005
    "VLLM_DISABLED_KERNELS": lambda: []
    if "VLLM_DISABLED_KERNELS" not in os.environ
    else os.environ["VLLM_DISABLED_KERNELS"].split(","),
1006
    # Disable pynccl (using torch.distributed instead)
1007
1008
1009
    "VLLM_DISABLE_PYNCCL": lambda: (
        os.getenv("VLLM_DISABLE_PYNCCL", "False").lower() in ("true", "1")
    ),
1010
1011
    # Disable aiter ops unless specifically enabled.
    # Acts as a parent switch to enable the rest of the other operations.
1012
1013
1014
    "VLLM_ROCM_USE_AITER": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER", "False").lower() in ("true", "1")
    ),
1015
1016
    # Whether to use aiter paged attention.
    # By default is disabled.
1017
1018
1019
    "VLLM_ROCM_USE_AITER_PAGED_ATTN": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_PAGED_ATTN", "False").lower() in ("true", "1")
    ),
1020
1021
1022
    # use aiter linear op if aiter ops are enabled
    # The following list of related ops
    # - scaled_mm (per-tensor / rowwise)
1023
    "VLLM_ROCM_USE_AITER_LINEAR": lambda: (
1024
        os.getenv("VLLM_ROCM_USE_AITER_LINEAR", "False").lower() in ("true", "1")
1025
    ),
1026
1027
    # Whether to use aiter moe ops.
    # By default is enabled.
1028
    "VLLM_ROCM_USE_AITER_MOE": lambda: (
1029
        os.getenv("VLLM_ROCM_USE_AITER_MOE", "False").lower() in ("true", "1")
1030
    ),
1031
    # use aiter rms norm op if aiter ops are enabled.
1032
    "VLLM_ROCM_USE_AITER_RMSNORM": lambda: (
1033
        os.getenv("VLLM_ROCM_USE_AITER_RMSNORM", "False").lower() in ("true", "1")
1034
    ),
1035
1036
    # Whether to use aiter mla ops.
    # By default is enabled.
1037
    "VLLM_ROCM_USE_AITER_MLA": lambda: (
1038
        os.getenv("VLLM_ROCM_USE_AITER_MLA", "False").lower() in ("true", "1")
1039
    ),
1040
1041
    # Whether to use aiter mha ops.
    # By default is enabled.
1042
    "VLLM_ROCM_USE_AITER_MHA": lambda: (
1043
        os.getenv("VLLM_ROCM_USE_AITER_MHA", "False").lower() in ("true", "1")
1044
    ),
1045
1046
    # Whether to use aiter fp4 gemm asm.
    # By default is disabled.
1047
1048
1049
    "VLLM_ROCM_USE_AITER_FP4_ASM_GEMM": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_FP4_ASM_GEMM", "False").lower() in ("true", "1")
    ),
1050
1051
    # Whether to use aiter rope.
    # By default is disabled.
1052
1053
    "VLLM_ROCM_USE_AITER_TRITON_ROPE": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_TRITON_ROPE", "False").lower() in ("true", "1")
1054
    ),
1055
1056
    # Whether to use aiter triton fp8 bmm kernel
    # By default is enabled.
1057
    "VLLM_ROCM_USE_AITER_FP8BMM": lambda: (
1058
        os.getenv("VLLM_ROCM_USE_AITER_FP8BMM", "False").lower() in ("true", "1")
1059
    ),
1060
1061
1062
    # Whether to use aiter triton fp4 bmm kernel
    # By default is enabled.
    "VLLM_ROCM_USE_AITER_FP4BMM": lambda: (
1063
        os.getenv("VLLM_ROCM_USE_AITER_FP4BMM", "False").lower() in ("true", "1")
1064
    ),
1065
1066
1067
1068
1069
    # 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")
    ),
1070
    # Whether to use aiter fusion shared experts ops.
1071
    # By default is disabled.
1072
    "VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS": lambda: (
1073
        os.getenv("VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS", "False").lower()
1074
1075
        in ("true", "1")
    ),
1076
1077
1078
    # Whether to use aiter triton kernels for gemm ops.
    # By default is enabled.
    "VLLM_ROCM_USE_AITER_TRITON_GEMM": lambda: (
1079
        os.getenv("VLLM_ROCM_USE_AITER_TRITON_GEMM", "False").lower() in ("true", "1")
1080
    ),
1081
    # use rocm skinny gemms
1082
1083
1084
    "VLLM_ROCM_USE_SKINNY_GEMM": lambda: (
        os.getenv("VLLM_ROCM_USE_SKINNY_GEMM", "True").lower() in ("true", "1")
    ),
1085
    # Pad the fp8 weights to 256 bytes for ROCm
1086
    "VLLM_ROCM_FP8_PADDING": lambda: bool(int(os.getenv("VLLM_ROCM_FP8_PADDING", "1"))),
1087
    # Pad the weights for the moe kernel
1088
    "VLLM_ROCM_MOE_PADDING": lambda: bool(int(os.getenv("VLLM_ROCM_MOE_PADDING", "0"))),
1089
    # custom paged attention kernel for MI3* cards
1090
1091
1092
    "VLLM_ROCM_CUSTOM_PAGED_ATTN": lambda: (
        os.getenv("VLLM_ROCM_CUSTOM_PAGED_ATTN", "True").lower() in ("true", "1")
    ),
1093
1094
1095
1096
    # Whether to use the shuffled kv cache layout
    "VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT": lambda: (
        os.getenv("VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT", "False").lower() in ("true", "1")
    ),
1097
1098
1099
    # Custom quick allreduce kernel for MI3* cards
    # Choice of quantization level: FP, INT8, INT6, INT4 or NONE
    # Recommended for large models to get allreduce
1100
1101
1102
1103
1104
    "VLLM_ROCM_QUICK_REDUCE_QUANTIZATION": env_with_choices(
        "VLLM_ROCM_QUICK_REDUCE_QUANTIZATION",
        "NONE",
        ["FP", "INT8", "INT6", "INT4", "NONE"],
    ),
1105
1106
1107
1108
    # 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
1109
1110
1111
1112
    "VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16": lambda: (
        os.getenv("VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16", "True").lower()
        in ("true", "1")
    ),
1113
1114
1115
1116
1117
1118
    # 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.
1119
1120
1121
    "VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB": lambda: maybe_convert_int(
        os.environ.get("VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB", None)
    ),
1122
    # Divisor for dynamic query scale factor calculation for FP8 KV Cache
zhangshao's avatar
zhangshao committed
1123
    "Q_SCALE_CONSTANT": lambda: int(os.getenv("Q_SCALE_CONSTANT", "10")),
1124
    # Divisor for dynamic key scale factor calculation for FP8 KV Cache
zhangshao's avatar
zhangshao committed
1125
    "K_SCALE_CONSTANT": lambda: int(os.getenv("K_SCALE_CONSTANT", "10")),
1126
    # Divisor for dynamic value scale factor calculation for FP8 KV Cache
zhangshao's avatar
zhangshao committed
1127
    "V_SCALE_CONSTANT": lambda: int(os.getenv("V_SCALE_CONSTANT", "10")),
1128
    # If set, enable multiprocessing in LLM for the V1 code path.
1129
1130
1131
1132
1133
1134
    "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")
    ),
1135
    "VLLM_DISABLE_COMPILE_CACHE": disable_compile_cache,
1136
1137
1138
    # If set, vllm will run in development mode, which will enable
    # some additional endpoints for developing and debugging,
    # e.g. `/reset_prefix_cache`
1139
    "VLLM_SERVER_DEV_MODE": lambda: bool(int(os.getenv("VLLM_SERVER_DEV_MODE", "0"))),
1140
1141
1142
1143
1144
1145
1146
    # 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.
1147
1148
1149
    "VLLM_V1_OUTPUT_PROC_CHUNK_SIZE": lambda: int(
        os.getenv("VLLM_V1_OUTPUT_PROC_CHUNK_SIZE", "128")
    ),
1150
    # If set, vLLM will disable the MLA attention optimizations.
1151
    "VLLM_MLA_DISABLE": lambda: bool(int(os.getenv("VLLM_MLA_DISABLE", "0"))),
1152
    # If set, vLLM will pick up the provided Flash Attention MLA
1153
1154
1155
    # 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.
1156
1157
1158
    "VLLM_RAY_PER_WORKER_GPUS": lambda: float(
        os.getenv("VLLM_RAY_PER_WORKER_GPUS", "1.0")
    ),
1159
1160
1161
    # 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"
1162
    "VLLM_RAY_BUNDLE_INDICES": lambda: os.getenv("VLLM_RAY_BUNDLE_INDICES", ""),
1163
1164
    # In some system, find_loaded_library() may not work. So we allow users to
    # specify the path through environment variable VLLM_CUDART_SO_PATH.
1165
    "VLLM_CUDART_SO_PATH": lambda: os.getenv("VLLM_CUDART_SO_PATH", None),
1166
    # Rank of the process in the data parallel setting
1167
    "VLLM_DP_RANK": lambda: int(os.getenv("VLLM_DP_RANK", "0")),
1168
1169
    # Rank of the process in the data parallel setting.
    # Defaults to VLLM_DP_RANK when not set.
1170
1171
1172
    "VLLM_DP_RANK_LOCAL": lambda: int(
        os.getenv("VLLM_DP_RANK_LOCAL", sys.modules[__name__].VLLM_DP_RANK)
    ),
1173
    # World size of the data parallel setting
1174
    "VLLM_DP_SIZE": lambda: int(os.getenv("VLLM_DP_SIZE", "1")),
1175
    # IP address of the master node in the data parallel setting
1176
    "VLLM_DP_MASTER_IP": lambda: os.getenv("VLLM_DP_MASTER_IP", "127.0.0.1"),
1177
    # Port of the master node in the data parallel setting
1178
    "VLLM_DP_MASTER_PORT": lambda: int(os.getenv("VLLM_DP_MASTER_PORT", "0")),
1179
1180
1181
1182
1183
    # 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.
1184
    "VLLM_MOE_DP_CHUNK_SIZE": lambda: int(os.getenv("VLLM_MOE_DP_CHUNK_SIZE", "256")),
1185
1186
1187
    "VLLM_ENABLE_MOE_DP_CHUNK": lambda: bool(
        int(os.getenv("VLLM_ENABLE_MOE_DP_CHUNK", "1"))
    ),
1188
    # Randomize inputs during dummy runs when using Data Parallel
1189
1190
1191
1192
    "VLLM_RANDOMIZE_DP_DUMMY_INPUTS": lambda: os.environ.get(
        "VLLM_RANDOMIZE_DP_DUMMY_INPUTS", "0"
    )
    == "1",
1193
1194
1195
1196
1197
1198
1199
    # 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;
1200
1201
1202
    # - "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;
1203
1204
1205
1206
    # 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"
    ),
1207
    # Whether to use S3 path for model loading in CI via RunAI Streamer
1208
    "VLLM_CI_USE_S3": lambda: os.environ.get("VLLM_CI_USE_S3", "0") == "1",
1209
    # Use model_redirect to redirect the model name to a local folder.
1210
1211
1212
1213
1214
    # `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
1215
1216
1217
    "VLLM_MODEL_REDIRECT_PATH": lambda: os.environ.get(
        "VLLM_MODEL_REDIRECT_PATH", None
    ),
1218
    # Whether to use atomicAdd reduce in gptq/awq marlin kernel.
1219
1220
1221
1222
    "VLLM_MARLIN_USE_ATOMIC_ADD": lambda: os.environ.get(
        "VLLM_MARLIN_USE_ATOMIC_ADD", "0"
    )
    == "1",
1223
    # Whether to use marlin kernel in mxfp4 quantization method
1224
1225
1226
    "VLLM_MXFP4_USE_MARLIN": lambda: maybe_convert_bool(
        os.environ.get("VLLM_MXFP4_USE_MARLIN", None)
    ),
1227
1228
1229
1230
    # The activation dtype for marlin kernel
    "VLLM_MARLIN_INPUT_DTYPE": env_with_choices(
        "VLLM_MARLIN_INPUT_DTYPE", None, ["int8", "fp8"]
    ),
1231
1232
1233
1234
1235
1236
    # 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"))
    ),
1237
1238
1239
    # 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.
1240
1241
1242
1243
    "VLLM_V1_USE_OUTLINES_CACHE": lambda: os.environ.get(
        "VLLM_V1_USE_OUTLINES_CACHE", "0"
    )
    == "1",
1244
1245
    # Gap between padding buckets for the forward pass. So we have
    # 8, we will run forward pass with [16, 24, 32, ...].
1246
1247
1248
1249
1250
1251
1252
1253
    "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)
    ),
1254
    # Whether using Pathways
1255
1256
1257
    "VLLM_TPU_USING_PATHWAYS": lambda: bool(
        "proxy" in os.getenv("JAX_PLATFORMS", "").lower()
    ),
1258
    # Allow use of DeepGemm kernels for fused moe ops.
1259
    "VLLM_USE_DEEP_GEMM": lambda: bool(int(os.getenv("VLLM_USE_DEEP_GEMM", "1"))),
1260
1261
1262
1263
    # 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"))
    ),
1264
    # Whether to use E8M0 scaling when DeepGEMM is used on Blackwell GPUs.
1265
1266
1267
    "VLLM_USE_DEEP_GEMM_E8M0": lambda: bool(
        int(os.getenv("VLLM_USE_DEEP_GEMM_E8M0", "1"))
    ),
1268
1269
1270
1271
    # Whether to create TMA-aligned scale tensor when DeepGEMM is used.
    "VLLM_USE_DEEP_GEMM_TMA_ALIGNED_SCALES": lambda: bool(
        int(os.getenv("VLLM_USE_DEEP_GEMM_TMA_ALIGNED_SCALES", "1"))
    ),
1272
1273
1274
1275
    # 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.
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
    # 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",
        ],
1291
    ),
1292
    # Whether to use fused grouped_topk used for MoE expert selection.
1293
1294
1295
    "VLLM_USE_FUSED_MOE_GROUPED_TOPK": lambda: bool(
        int(os.getenv("VLLM_USE_FUSED_MOE_GROUPED_TOPK", "1"))
    ),
1296
1297
1298
1299
1300
    # 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"))
    ),
1301
    # Allow use of FlashInfer MoE kernels for fused moe ops.
1302
1303
1304
    "VLLM_USE_FLASHINFER_MOE_FP16": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_FP16", "0"))
    ),
1305
    # Allow use of FlashInfer MoE kernels for fused moe ops.
1306
1307
1308
    "VLLM_USE_FLASHINFER_MOE_FP8": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_FP8", "0"))
    ),
1309
    # Allow use of FlashInfer CUTLASS kernels for fused moe ops.
1310
1311
1312
    "VLLM_USE_FLASHINFER_MOE_FP4": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_FP4", "0"))
    ),
1313
1314
    # If set to 1, use the FlashInfer
    # MXFP8 (activation) x MXFP4 (weight) MoE backend.
1315
1316
1317
    "VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8", "0"))
    ),
1318
1319
1320
1321
    # 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.
1322
1323
1324
    "VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS", "0"))
    ),
1325
1326
    # If set to 1, use the FlashInfer
    # BF16 (activation) x MXFP4 (weight) MoE backend.
1327
1328
1329
    "VLLM_USE_FLASHINFER_MOE_MXFP4_BF16": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_MXFP4_BF16", "0"))
    ),
1330
1331
1332
    # 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.
1333
    "VLLM_XGRAMMAR_CACHE_MB": lambda: int(os.getenv("VLLM_XGRAMMAR_CACHE_MB", "512")),
1334
1335
1336
1337
1338
1339
1340
    # 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.
1341
1342
1343
    "VLLM_MSGPACK_ZERO_COPY_THRESHOLD": lambda: int(
        os.getenv("VLLM_MSGPACK_ZERO_COPY_THRESHOLD", "256")
    ),
1344
1345
1346
    # 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.
1347
1348
1349
    "VLLM_ALLOW_INSECURE_SERIALIZATION": lambda: bool(
        int(os.getenv("VLLM_ALLOW_INSECURE_SERIALIZATION", "0"))
    ),
Robert Shaw's avatar
Robert Shaw committed
1350
    # IP address used for NIXL handshake between remote agents.
1351
1352
1353
    "VLLM_NIXL_SIDE_CHANNEL_HOST": lambda: os.getenv(
        "VLLM_NIXL_SIDE_CHANNEL_HOST", "localhost"
    ),
Robert Shaw's avatar
Robert Shaw committed
1354
    # Port used for NIXL handshake between remote agents.
1355
1356
1357
    "VLLM_NIXL_SIDE_CHANNEL_PORT": lambda: int(
        os.getenv("VLLM_NIXL_SIDE_CHANNEL_PORT", "5600")
    ),
1358
1359
1360
1361
    # Port used for Mooncake handshake between remote agents.
    "VLLM_MOONCAKE_BOOTSTRAP_PORT": lambda: int(
        os.getenv("VLLM_MOONCAKE_BOOTSTRAP_PORT", "8998")
    ),
1362
1363
    # [DEPRECATED - will be removed in v0.15.0] all2all backend for vllm's
    # expert parallel communication. Use --all2all-backend CLI argument instead.
1364
    # Available options:
1365
1366
1367
    # - "naive": naive all2all implementation using broadcasts
    # - "allgather_reducescatter": all2all implementation based on allgather and
    #  reducescatter
1368
    # - "pplx": use pplx kernels
1369
1370
    # - "deepep_high_throughput", use deepep high-throughput kernels
    # - "deepep_low_latency", use deepep low-latency kernels
1371
    # - "mori", use MoRI kernels
1372
    # - "flashinfer_all2allv", use flashinfer alltoallv kernels for mnnvl
1373
1374
    "VLLM_ALL2ALL_BACKEND": env_with_choices(
        "VLLM_ALL2ALL_BACKEND",
1375
        None,
1376
1377
1378
1379
1380
        [
            "naive",
            "pplx",
            "deepep_high_throughput",
            "deepep_low_latency",
1381
            "mori",
1382
1383
1384
1385
            "allgather_reducescatter",
            "flashinfer_all2allv",
        ],
    ),
1386
1387
    # Flashinfer MoE backend for vLLM's fused Mixture-of-Experts support.
    # Both require compute capability 10.0 or above.
1388
1389
1390
1391
1392
    # Available options:
    # - "throughput":  [default]
    #     Uses CUTLASS kernels optimized for high-throughput batch inference.
    # - "latency":
    #     Uses TensorRT-LLM kernels optimized for low-latency inference.
1393
    "VLLM_FLASHINFER_MOE_BACKEND": env_with_choices(
1394
1395
1396
        "VLLM_FLASHINFER_MOE_BACKEND",
        "latency",
        ["throughput", "latency", "masked_gemm"],
1397
    ),
1398
1399
1400
1401
    # 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))
    ),
1402
1403
1404
1405
    # 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.
1406
1407
1408
    "VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE": lambda: int(
        os.getenv("VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE", "163840")
    ),
1409
1410
1411
1412
    # 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> }
1413
    # Unspecified world sizes will fall back to
1414
    #     { 2: 64, 4: 1, <everything else>: 0.5 }
1415
1416
1417
    "VLLM_FLASHINFER_ALLREDUCE_FUSION_THRESHOLDS_MB": lambda: json.loads(
        os.getenv("VLLM_FLASHINFER_ALLREDUCE_FUSION_THRESHOLDS_MB", "{}")
    ),
1418
1419
1420
    # MoE routing strategy selector.
    # See `RoutingSimulator.get_available_strategies()` # for available
    # strategies.
1421
    # Custom routing strategies can be registered by
1422
1423
    # RoutingSimulator.register_strategy()
    # Note: custom strategies may not produce correct model outputs
1424
1425
1426
    "VLLM_MOE_ROUTING_SIMULATION_STRATEGY": lambda: os.environ.get(
        "VLLM_MOE_ROUTING_SIMULATION_STRATEGY", ""
    ).lower(),
1427
    # Regex timeout for use by the vLLM tool parsing plugins.
1428
1429
1430
    "VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS": lambda: int(
        os.getenv("VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS", "1")
    ),
1431
1432
    # Reduce CPU usage when vLLM is idle. Enabling this will incur small
    # latency penalty when a request eventually comes.
1433
    "VLLM_SLEEP_WHEN_IDLE": lambda: bool(int(os.getenv("VLLM_SLEEP_WHEN_IDLE", "0"))),
1434
1435
1436
    # 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.
1437
1438
1439
    "VLLM_MQ_MAX_CHUNK_BYTES_MB": lambda: int(
        os.getenv("VLLM_MQ_MAX_CHUNK_BYTES_MB", "16")
    ),
1440
1441
    # Timeout in seconds for execute_model RPC calls in multiprocessing
    # executor (only applies when TP > 1).
1442
1443
1444
    "VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS": lambda: int(
        os.getenv("VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS", "300")
    ),
1445
1446
1447
1448
1449
1450
1451
    # 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.
1452
1453
1454
    "VLLM_KV_CACHE_LAYOUT": env_with_choices(
        "VLLM_KV_CACHE_LAYOUT", None, ["NHD", "HND"]
    ),
1455
1456
1457
    # 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.
1458
1459
1460
    "VLLM_COMPUTE_NANS_IN_LOGITS": lambda: bool(
        int(os.getenv("VLLM_COMPUTE_NANS_IN_LOGITS", "0"))
    ),
1461
1462
1463
    # Controls whether or not emulations are used for NVFP4
    # generations on machines < 100 for compressed-tensors
    # models
1464
1465
1466
    "VLLM_USE_NVFP4_CT_EMULATIONS": lambda: bool(
        int(os.getenv("VLLM_USE_NVFP4_CT_EMULATIONS", "0"))
    ),
1467
1468
1469
1470
    # 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.
1471
1472
1473
    "VLLM_NIXL_ABORT_REQUEST_TIMEOUT": lambda: int(
        os.getenv("VLLM_NIXL_ABORT_REQUEST_TIMEOUT", "480")
    ),
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
    # 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")),
1488
1489
1490
1491
    # Timeout (in seconds) for MooncakeConnector in PD disaggregated setup.
    "VLLM_MOONCAKE_ABORT_REQUEST_TIMEOUT": lambda: int(
        os.getenv("VLLM_MOONCAKE_ABORT_REQUEST_TIMEOUT", "480")
    ),
1492
1493
    # If set, it means we pre-downloaded cubin files and flashinfer will
    # read the cubin files directly.
1494
1495
1496
    "VLLM_HAS_FLASHINFER_CUBIN": lambda: bool(
        int(os.getenv("VLLM_HAS_FLASHINFER_CUBIN", "0"))
    ),
1497
1498
1499
1500
    # Supported options:
    # - "flashinfer-cudnn": use flashinfer cudnn GEMM backend
    # - "flashinfer-trtllm": use flashinfer trtllm GEMM backend
    # - "flashinfer-cutlass": use flashinfer cutlass GEMM backend
1501
    # - "marlin": use marlin GEMM backend (for GPUs without native FP4 support)
1502
1503
1504
1505
    # - <none>: automatically pick an available backend
    "VLLM_NVFP4_GEMM_BACKEND": env_with_choices(
        "VLLM_NVFP4_GEMM_BACKEND",
        None,
1506
1507
1508
1509
1510
        [
            "flashinfer-cudnn",
            "flashinfer-trtllm",
            "flashinfer-cutlass",
            "cutlass",
1511
            "marlin",
1512
        ],
1513
    ),
1514
1515
1516
    # 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.
1517
1518
1519
    "VLLM_ENABLE_CUDAGRAPH_GC": lambda: bool(
        int(os.getenv("VLLM_ENABLE_CUDAGRAPH_GC", "0"))
    ),
1520
    # Used to force set up loopback IP
1521
    "VLLM_LOOPBACK_IP": lambda: os.getenv("VLLM_LOOPBACK_IP", ""),
1522
1523
1524
    # Used to set the process name prefix for vLLM processes.
    # This is useful for debugging and monitoring purposes.
    # The default value is "VLLM".
1525
    "VLLM_PROCESS_NAME_PREFIX": lambda: os.getenv("VLLM_PROCESS_NAME_PREFIX", "VLLM"),
1526
1527
1528
1529
1530
1531
1532
    # 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.
1533
    "VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE": lambda: bool(
1534
        int(os.getenv("VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE", "1"))
1535
    ),
1536
1537
    # 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
1538
1539
    # messages for those requests in memory. By default, this is disabled (0),
    # and the "store" option is ignored.
1540
1541
1542
1543
1544
    # 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.
1545
1546
1547
    "VLLM_ENABLE_RESPONSES_API_STORE": lambda: bool(
        int(os.getenv("VLLM_ENABLE_RESPONSES_API_STORE", "0"))
    ),
xiao-llm's avatar
xiao-llm committed
1548
    # If set, use the fp8 mfma in rocm paged attention.
1549
1550
1551
    "VLLM_ROCM_FP8_MFMA_PAGE_ATTN": lambda: bool(
        int(os.getenv("VLLM_ROCM_FP8_MFMA_PAGE_ATTN", "0"))
    ),
1552
    # Whether to use pytorch symmetric memory for allreduce
1553
    "VLLM_ALLREDUCE_USE_SYMM_MEM": lambda: bool(
1554
        int(os.getenv("VLLM_ALLREDUCE_USE_SYMM_MEM", "1"))
1555
    ),
1556
1557
1558
1559
    # 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"))
    ),
1560
    # Allows vllm to find tuned config under customized folder
1561
    "VLLM_TUNED_CONFIG_FOLDER": lambda: os.getenv("VLLM_TUNED_CONFIG_FOLDER", None),
1562
1563
1564
1565
1566
1567
1568
1569
1570
    # 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"],
    ),
1571
    # Allows harmony instructions to be injected on system messages
1572
1573
1574
    "VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS": lambda: bool(
        int(os.getenv("VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS", "0"))
    ),
1575
1576
1577
1578
1579
1580
    # 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"))
    ),
1581
    # Add optional custom scopes for profiling, disable to avoid overheads
1582
1583
1584
    "VLLM_CUSTOM_SCOPES_FOR_PROFILING": lambda: bool(
        int(os.getenv("VLLM_CUSTOM_SCOPES_FOR_PROFILING", "0"))
    ),
1585
    # Add optional nvtx scopes for profiling, disable to avoid overheads
1586
1587
1588
    "VLLM_NVTX_SCOPES_FOR_PROFILING": lambda: bool(
        int(os.getenv("VLLM_NVTX_SCOPES_FOR_PROFILING", "0"))
    ),
1589
1590
    # Represent block hashes in KV cache events as 64-bit integers instead of
    # raw bytes. Defaults to True for backward compatibility.
1591
1592
1593
    "VLLM_KV_EVENTS_USE_INT_BLOCK_HASHES": lambda: bool(
        int(os.getenv("VLLM_KV_EVENTS_USE_INT_BLOCK_HASHES", "1"))
    ),
1594
1595
    # Name of the shared memory buffer used for object storage.
    # Only effective when mm_config.mm_processor_cache_type == "shm".
1596
1597
1598
1599
1600
    # Automatically generates a unique UUID-based name per process tree
    # if not explicitly set.
    "VLLM_OBJECT_STORAGE_SHM_BUFFER_NAME": get_env_or_set_default(
        "VLLM_OBJECT_STORAGE_SHM_BUFFER_NAME",
        lambda: f"VLLM_OBJECT_STORAGE_SHM_BUFFER_{uuid.uuid4().hex}",
1601
    ),
1602
    # The size in MB of the buffers (NVL and RDMA) used by DeepEP
1603
1604
1605
    "VLLM_DEEPEP_BUFFER_SIZE_MB": lambda: int(
        os.getenv("VLLM_DEEPEP_BUFFER_SIZE_MB", "1024")
    ),
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
    # 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"))
    ),
1617
1618
    # 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
1619
    "VLLM_DBO_COMM_SMS": lambda: int(os.getenv("VLLM_DBO_COMM_SMS", "20")),
1620
1621
1622
    # 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)
1623
1624
1625
    "VLLM_ENABLE_INDUCTOR_MAX_AUTOTUNE": lambda: bool(
        int(os.getenv("VLLM_ENABLE_INDUCTOR_MAX_AUTOTUNE", "1"))
    ),
1626
1627
    # If set to 1, enable coordinate_descent_tuning;
    # By default, this is enabled (1)
1628
1629
1630
    "VLLM_ENABLE_INDUCTOR_COORDINATE_DESCENT_TUNING": lambda: bool(
        int(os.getenv("VLLM_ENABLE_INDUCTOR_COORDINATE_DESCENT_TUNING", "1"))
    ),
1631
    # Flag to enable NCCL symmetric memory allocation and registration
1632
1633
1634
    "VLLM_USE_NCCL_SYMM_MEM": lambda: bool(
        int(os.getenv("VLLM_USE_NCCL_SYMM_MEM", "0"))
    ),
1635
    # NCCL header path
1636
    "VLLM_NCCL_INCLUDE_PATH": lambda: os.environ.get("VLLM_NCCL_INCLUDE_PATH", None),
1637
1638
    # Flag to enable FBGemm kernels on model execution
    "VLLM_USE_FBGEMM": lambda: bool(int(os.getenv("VLLM_USE_FBGEMM", "0"))),
1639
1640
1641
1642
1643
1644
    # 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", ""),
1645
1646
1647
    # Debug workspace allocations.
    # logging of workspace resize operations.
    "VLLM_DEBUG_WORKSPACE": lambda: bool(int(os.getenv("VLLM_DEBUG_WORKSPACE", "0"))),
1648
    # Disables parallel execution of shared_experts via separate cuda stream
1649
1650
    "VLLM_DISABLE_SHARED_EXPERTS_STREAM": lambda: bool(
        int(os.getenv("VLLM_DISABLE_SHARED_EXPERTS_STREAM", "0"))
1651
    ),
1652
1653
1654
1655
1656
1657
1658
    # 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))
    ),
1659
1660
1661
1662
1663
1664
1665
1666
1667
    # 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
1668
1669
1670
1671
    # Flag to enable v2 model runner.
    "VLLM_USE_V2_MODEL_RUNNER": lambda: bool(
        int(os.getenv("VLLM_USE_V2_MODEL_RUNNER", "0"))
    ),
1672
1673
1674
1675
1676
1677
    # Log model inspection after loading.
    # If enabled, logs a transformers-style hierarchical view of the model
    # with quantization methods and attention backends.
    "VLLM_LOG_MODEL_INSPECTION": lambda: bool(
        int(os.getenv("VLLM_LOG_MODEL_INSPECTION", "0"))
    ),
1678
1679
1680
1681
    # Debug logging for --enable-mfu-metrics
    "VLLM_DEBUG_MFU_METRICS": lambda: bool(
        int(os.getenv("VLLM_DEBUG_MFU_METRICS", "0"))
    ),
1682
1683
    # Disable logging of vLLM logo at server startup time.
    "VLLM_DISABLE_LOG_LOGO": lambda: bool(int(os.getenv("VLLM_DISABLE_LOG_LOGO", "0"))),
1684
1685
1686
    # Disable PDL for LoRA, as enabling PDL with LoRA on SM100 causes
    # Triton compilation to fail.
    "VLLM_LORA_DISABLE_PDL": lambda: bool(int(os.getenv("VLLM_LORA_DISABLE_PDL", "0"))),
1687
    
1688
1689
1690
    # add envs
    
    # used in optest environment to manually set the https port
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
    'VLLM_OPTEST_URLS_PORT':
    lambda: int(os.getenv('VLLM_OPTEST_URLS_PORT', '8000'))
    if 'VLLM_OPTEST_URLS_PORT' in os.environ else None,
    
    # Path to the optest models.
    # If set, will load models from local path instead of Hugging Face Hub.
    'VLLM_OPTEST_MODELS_PATH':
    lambda: os.getenv('VLLM_OPTEST_MODELS_PATH', "") or os.getenv("OPTEST_MODELS_PATH", ""),
    
    # flag to control if vllm should use triton prefix flash attention
    "VLLM_USE_TRITON_PREFIX_FLASH_ATTN":
    lambda: (os.environ.get("VLLM_USE_TRITON_PREFIX_FLASH_ATTN", "False").lower() in
             ("true", "1")),
    
1705
1706
1707
1708
    # If set, vLLM will use FLASH ATTN fp8 attention optimizations.
    "VLLM_USE_FLASH_ATTN_FP8":
    lambda: bool(int(os.getenv("VLLM_USE_FLASH_ATTN_FP8", "0"))),
    
1709
1710
1711
1712
1713
    # flag to control if vllm should use q quant
    "VLLM_USE_QUERY_QUANT":
    lambda: (os.environ.get("VLLM_USE_QUERY_QUANT", "False").lower() in
             ("true", "1")),
    
zhuwenwen's avatar
zhuwenwen committed
1714
1715
1716
1717
    # If set, vLLM will use FLASH MLA attention optimizations.
    "VLLM_USE_FLASH_MLA":
    lambda: bool(int(os.getenv("VLLM_USE_FLASH_MLA", "1"))),
    
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
    # flag to control vllm to use optimized kernels
    "VLLM_USE_OPT_OP":
    lambda: (os.environ.get("VLLM_USE_OPT_OP", "True").lower() in
             ("true", "1")),
    
    # flag to control vllm to use optimized tc paged attn kernels
    "VLLM_USE_TC_PAGED_ATTN":
    lambda: (os.environ.get("VLLM_USE_TC_PAGED_ATTN", "True").lower() in
             ("true", "1")),
    
    # flag to control if vllm print pa parameters
    "VLLM_USE_PA_PRINT_PARAM":
    lambda: (os.environ.get("VLLM_USE_PA_PRINT_PARAM", "False").lower() in
             ("true", "1")),
    
    # If set, vLLM will disable the draft model in cudagraph mode.
    "VLLM_SPEC_DECODE_EAGER":
    lambda: bool(int(os.getenv("VLLM_SPEC_DECODE_EAGER", "0"))),
    
    # flag to control vllm to use optimized kernels
    "VLLM_PCIE_USE_CUSTOM_ALLREDUCE":
1739
    lambda: bool(int(os.environ.get("VLLM_PCIE_USE_CUSTOM_ALLREDUCE", "1"))),
1740
    
1741
1742
    # flag to control vllm to use optimized kernels
    "VLLM_CUSTOM_CACHE":
1743
    lambda: bool(int(os.environ.get("VLLM_CUSTOM_CACHE", "1"))),
1744
    
zhuwenwen's avatar
zhuwenwen committed
1745
1746
1747
1748
    # flag to control vllm to use optimized kernels
    "VLLM_CUSTOM_ALLREDUCE_SUPPORTED_WORLDSIZE_MAX":
    lambda: int(os.getenv("VLLM_CUSTOM_ALLREDUCE_SUPPORTED_WORLDSIZE_MAX", "16")),
    
1749
1750
1751
    # If set, vLLM will disable the draft model in cudagraph mode.
    "VLLM_ENFORCE_EAGER_BS_THRESHOLD":
    lambda: int(os.environ.get("VLLM_ENFORCE_EAGER_BS_THRESHOLD", "-1")),
1752

1753
1754
1755
1756
1757
1758
    # 'has_comtext' is a variable in common.py, which is calculated
    # by metadata by default. However, it may introduce synchronization 
    # and affect performance, so it is directly assigned as False. 
    # If there are any problems during use, use environment variables 
    # to restore the default usage.
    "VLLM_HAS_CONTEXT_DEFAULT":
zhuwenwen's avatar
zhuwenwen committed
1759
    lambda: bool(int(os.getenv("VLLM_HAS_CONTEXT_DEFAULT", "1"))),
1760
1761
1762
    
    # If set, vLLM will transpose weight to use nn layout
    "VLLM_USE_NN":
zhuwenwen's avatar
zhuwenwen committed
1763
    lambda: (os.environ.get("VLLM_USE_NN", "True").lower() in
1764
             ("true", "1")),
1765

1766
1767
1768
    # Enable two batch overlap.
    "VLLM_ENABLE_TBO":
    lambda: bool(int(os.getenv("VLLM_ENABLE_TBO", "0"))),
1769
1770
1771

    # If set, vLLM will enable the moe_fused_gate kernel.
    "VLLM_ENABLE_MOE_FUSED_GATE":
zhuwenwen's avatar
zhuwenwen committed
1772
    lambda: bool(int(os.getenv("VLLM_ENABLE_MOE_FUSED_GATE", "1"))),
zhuwenwen's avatar
zhuwenwen committed
1773
    
1774
1775
    # vLLM will use FlashAttention Backend for page attention computation on rocm
    "VLLM_USE_FLASH_ATTN_PA":
zhuwenwen's avatar
zhuwenwen committed
1776
    lambda: (os.environ.get("VLLM_USE_FLASH_ATTN_PA", "True").lower() in
zhuwenwen's avatar
zhuwenwen committed
1777
             ("true", "1")),
zhuwenwen's avatar
zhuwenwen committed
1778
1779
1780
1781
1782
    
    # vLLM will use apex for rmsnorm
    "VLLM_USE_APEX_RN":
    lambda: (os.environ.get("VLLM_USE_APEX_RN", "False").lower() in
             ("true", "1")),
1783
1784
1785
    
    # vLLM will use global cache for moe
    "VLLM_USE_GLOBAL_CACHE13":
1786
        lambda: (os.environ.get("VLLM_USE_GLOBAL_CACHE13", "False").lower() in
1787
                 ("true", "1")),
1788
        
1789
1790
1791
    # vLLM will use lightop for deepseek-v3
    "VLLM_USE_LIGHTOP":
        lambda: (os.environ.get("VLLM_USE_LIGHTOP", "False").lower() in
1792
                 ("true", "1")),
1793
        
1794
1795
1796
    # vLLM will use opt cat for deepseek-v3
    "VLLM_USE_OPT_CAT":
        lambda: (os.environ.get("VLLM_USE_OPT_CAT", "True").lower() in
zhuwenwen's avatar
zhuwenwen committed
1797
                 ("true", "1")), 
zhuwenwen's avatar
zhuwenwen committed
1798
1799
1800
1801
1802
1803
1804
1805
    # vLLM will use lightop moe_sum 
    "VLLM_USE_LIGHTOP_MOE_SUM":
        lambda: (os.environ.get("VLLM_USE_LIGHTOP_MOE_SUM", "False").lower() in
                 ("true", "1")),  
    # vLLM will use lightop moe_align_block_size 
    "VLLM_USE_LIGHTOP_MOE_ALIGN":
        lambda: (os.environ.get("VLLM_USE_LIGHTOP_MOE_ALIGN", "False").lower() in
                 ("true", "1")),     
1806
1807
1808
1809
    # vLLM will use opt merge_aatn_states,not triton
    "VLLM_USE_MERGE_ATTN_STATES_OPT":
        lambda: (os.environ.get("VLLM_USE_MERGE_ATTN_STATES_OPT", "True").lower() in
                 ("true", "1")),  
jujl1's avatar
jujl1 committed
1810
1811
    # vllm will use rmsquant fused op
    "USE_FUSED_RMS_QUANT":
王敏's avatar
王敏 committed
1812
        lambda: bool(int(os.getenv("USE_FUSED_RMS_QUANT", "0"))),
1813
1814
1815
        
    # vllm will use silu_mul_quant fused op
    "USE_FUSED_SILU_MUL_QUANT":
1816
1817
        lambda: (os.getenv("USE_FUSED_SILU_MUL_QUANT", "False").lower() in
                ("true", "1")),
1818

zhuwenwen's avatar
zhuwenwen committed
1819
1820
    # vLLM will split prefill and decode, not mix up
    "VLLM_USE_PD_SPLIT":
1821
        lambda: (os.environ.get("VLLM_USE_PD_SPLIT", "True").lower() in
zhuwenwen's avatar
zhuwenwen committed
1822
                 ("true", "1")), 
zhuwenwen's avatar
zhuwenwen committed
1823
1824
1825
1826
    # vLLM will sync to avoid pp vmfault
    "VLLM_USE_PP_SYNC":
        lambda: (os.environ.get("VLLM_USE_PP_SYNC", "False").lower() in
                 ("true", "1")), 
1827
1828
1829
1830
    # vLLM will use piecewise
    "VLLM_USE_PIECEWISE":
        lambda: (os.environ.get("VLLM_USE_PIECEWISE", "True").lower() in
                 ("true", "1")), 
1831
1832
    # vllm will use encoding_dsv32.py for dpsk-v32
    "VLLM_USE_V32_ENCODE":
1833
        lambda: (os.environ.get('VLLM_USE_V32_ENCODE', 'False').lower() in
1834
                 ("true", "1")),  
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
    # vLLM will use fused silu+mul kernel (fp16 + qwen3-30b)
    "VLLM_USE_FUSE_SILU_AND_MUL":
        lambda: (os.environ.get("VLLM_USE_FUSE_SILU_AND_MUL", "False").lower() in
                 ("true", "1")),
    # vLLM will use optimized reshape_and_cache kernel when enabled (fp16 + qwen3-30b)
    "VLLM_USE_OPT_RESHAPE_AND_CACHE":
        lambda:
        (os.environ.get("VLLM_USE_OPT_RESHAPE_AND_CACHE", "False").lower() in
                ("true", "1")),
    # vLLM will use optimized topk_softmax + renormalize
    "VLLM_USE_TOPK_RENORM":
        lambda:
zhuwenwen's avatar
zhuwenwen committed
1847
        (os.environ.get("VLLM_USE_TOPK_RENORM", "False").lower() in
1848
                ("true", "1")),
1849
1850
    # vLLM will use fused RMS + RoPE kernel
    "VLLM_USE_FUSED_RMS_ROPE":
1851
        lambda: (os.environ.get("VLLM_USE_FUSED_RMS_ROPE", "True").lower() in
1852
                 ("true", "1")),
1853
1854
1855
1856
    # vLLM will use lightop for dpsk mtp fill + rms*2 + cat
    "VLLM_USE_FUSED_FILL_RMS_CAT":
        lambda: (os.environ.get("VLLM_USE_FUSED_FILL_RMS_CAT", "False").lower() in
                 ("true", "1")),
jujl1's avatar
jujl1 committed
1857
1858
1859
    "VLLM_USE_PP_BALANCE":
        lambda: (os.environ.get("VLLM_USE_PP_BALANCE", "True").lower() in
                 ("true", "1")),
1860
1861
1862
1863
1864
1865
    # W8A8 GEMM backend selection for vLLM quantized models.
    # lightop/triton: 1
    # cutlass: 2 (will remove in the future)
    # blaslt: 3 (default)
    # rocblas: others
    "VLLM_W8A8_BACKEND": lambda: int(os.getenv("VLLM_W8A8_BACKEND", "3")),
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
    # Capture MoE router logits for debugging/analysis.
    "VLLM_MOE_ROUTER_CAPTURE":
    lambda: (os.getenv("VLLM_MOE_ROUTER_CAPTURE", "0").lower() in ("true", "1")),
    # Output directory for MoE router capture dumps.
    "VLLM_MOE_ROUTER_CAPTURE_DIR":
    lambda: os.environ.get(
        "VLLM_MOE_ROUTER_CAPTURE_DIR",
        "/tmp",
    ),
    # Capture only the specified rank; set to -1 to capture all ranks.
    "VLLM_MOE_ROUTER_CAPTURE_RANK":
    lambda: int(os.environ.get("VLLM_MOE_ROUTER_CAPTURE_RANK", "-1")),
    # Max number of MoE layers to record per process (0 = unlimited).
    "VLLM_MOE_ROUTER_CAPTURE_MAX_LAYERS":
    lambda: int(os.environ.get("VLLM_MOE_ROUTER_CAPTURE_MAX_LAYERS", "0")),
    # Only capture when num_tokens > N (negative disables).
    "VLLM_MOE_ROUTER_CAPTURE_NUM_TOKENS_GT":
    lambda: int(os.environ.get("VLLM_MOE_ROUTER_CAPTURE_NUM_TOKENS_GT", "-1")),
    # Only capture when num_tokens < N (0 disables).
    "VLLM_MOE_ROUTER_CAPTURE_NUM_TOKENS_LT":
    lambda: int(os.environ.get("VLLM_MOE_ROUTER_CAPTURE_NUM_TOKENS_LT", "-1")),
王敏's avatar
王敏 committed
1887
1888
1889
1890
1891

    # vllm will use optimized reject sample
    "VLLM_REJECT_SAMPLE_OPT":
        lambda: (os.getenv('VLLM_REJECT_SAMPLE_OPT', 'True').lower() in
                 ("true", "1")),
1892
1893
1894
1895
    # Force using Triton MoE path (disable Marlin W16A16 MoE).
    "VLLM_USE_MOE_W16A16_TRITON":
        lambda: (os.environ.get("VLLM_USE_MOE_W16A16_TRITON", "0").lower() in
                 ("true", "1")),
guanyu1's avatar
guanyu1 committed
1896
1897
    "VLLM_1D_MROPE":
        lambda: (os.environ.get("VLLM_1D_MROPE", "0").lower() in ("true", "1")),
1898
1899
1900
1901
    #If set to 1/True, enable the V1 fast token-id copy path in InputBatch.
    "VLLM_V1_FAST_TOKEN_ID_COPY":
        lambda: (os.environ.get("VLLM_V1_FAST_TOKEN_ID_COPY", "False").lower() in
                    ("true", "1")),
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
    # If set to 1/True, enable reduced top-k/top-p sampling fast path in the
    # V1 PyTorch-native sampler path.
    #
    # Recommended when both top_k is enabled and top_p < 1.0 (nucleus
    # sampling). Not recommended for top-k only (top_p == 1.0) due to
    # potential behavior differences when the k-th logit is tied.
    "VLLM_V1_USE_REDUCED_TOPK_TOPP_SAMPLER":
        lambda: (
            os.environ.get(
                "VLLM_V1_USE_REDUCED_TOPK_TOPP_SAMPLER", "False"
            ).lower()
            in ("true", "1")
        ),
1915
1916
1917
1918
    # vLLM will use lightop fill + moe_align_block_size
    "VLLM_USE_LIGHTOP_FILL_MOE_ALIGN":
        lambda: (os.environ.get("VLLM_USE_LIGHTOP_FILL_MOE_ALIGN", "False").lower() in
                 ("true", "1")),
1919
1920
1921
1922
1923

    #If set to 1/True, enable fuse split qkv+rmsnorm+rope+kv update just like glm4.7 moe attention.
    "VLLM_V1_USE_FUSED_QKV_SPLIT_RMS_ROPE_KVSTORE":
        lambda: (os.environ.get("VLLM_V1_USE_FUSED_QKV_SPLIT_RMS_ROPE_KVSTORE", "False").lower() in
                    ("true", "1")),
1924
1925
1926
1927
1928
    # DeepSeek MLA: fused rmsnorm + contiguous + rope + concat_and_cache_mla
    "VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT":
        lambda: (os.environ.get("VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT",
                                "False").lower() in ("true", "1")),

1929
1930
1931
1932
    # DOUBLE TRANSPOSE BMM FP8 format use in NMZ DeepSeek models
    "VLLM_USE_FUSED_DTBMM":
        lambda: (os.environ.get("VLLM_USE_FUSED_DTBMM", "False").lower() in
                ("true", "1")),
wujl5's avatar
wujl5 committed
1933
1934
1935
1936
    # vllm will use 1-24,32,40,48... (not only 1 2 4 8 16)
    "VLLM_USE_CUDA_GRAPH_SIZES":
        lambda: (os.getenv("VLLM_USE_CUDA_GRAPH_SIZES", "False").lower() in
                ("true", "1")),
1937
1938
}

1939
# --8<-- [end:env-vars-definition]
1940

1941

1942
def __getattr__(name: str):
1943
1944
1945
1946
1947
1948
    """
    Gets environment variables lazily.

    NOTE: After enable_envs_cache() invocation (which triggered after service
    initialization), all environment variables will be cached.
    """
1949
1950
1951
1952
1953
    if name in environment_variables:
        return environment_variables[name]()
    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")


1954
1955
1956
1957
1958
1959
def _is_envs_cache_enabled() -> bool:
    """Checked if __getattr__ is wrapped with functools.cache"""
    global __getattr__
    return hasattr(__getattr__, "cache_clear")


1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
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.
    """
1970
1971
1972
    if _is_envs_cache_enabled():
        # Avoid wrapping functools.cache multiple times
        return
1973
1974
1975
1976
1977
1978
1979
1980
1981
    # Tag __getattr__ with functools.cache
    global __getattr__
    __getattr__ = functools.cache(__getattr__)

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


1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
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__


1993
1994
def __dir__():
    return list(environment_variables.keys())
1995
1996
1997
1998
1999
2000
2001
2002
2003


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}")


2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
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",
2026
        "VLLM_FORCE_AOT_LOAD",
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
        "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
2037
        "VLLM_LOGGING_COLOR",
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
        "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",
2054
        "VLLM_OBJECT_STORAGE_SHM_BUFFER_NAME",
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
        "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
2068
        "NO_COLOR",
2069
        "VLLM_W8A8_BACKEND",
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
    }

    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
2088

2089
        factors[factor] = normalize_value(raw)
2090

2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
    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",
2111
    ]
2112

2113
2114
    for var in ray_noset_env_vars:
        factors[var] = normalize_value(os.getenv(var))
2115

2116
    return factors