envs.py 96.4 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
import torch
12
13
from collections.abc import Callable
from typing import TYPE_CHECKING, Any, Literal
14
15
16

if TYPE_CHECKING:
    VLLM_HOST_IP: str = ""
17
    VLLM_PORT: int | None = None
18
    VLLM_RPC_BASE_PATH: str = tempfile.gettempdir()
19
    VLLM_USE_MODELSCOPE: bool = False
20
    VLLM_RINGBUFFER_WARNING_INTERVAL: int = 60
21
22
    VLLM_NCCL_SO_PATH: str | None = None
    LD_LIBRARY_PATH: str | None = None
23
    VLLM_ROCM_SLEEP_MEM_CHUNK_SIZE: int = 256
24
    LOCAL_RANK: int = 0
25
    CUDA_VISIBLE_DEVICES: str | None = None
26
    VLLM_ENGINE_ITERATION_TIMEOUT_S: int = 60
27
    VLLM_ENGINE_READY_TIMEOUT_S: int = 600
28
    VLLM_API_KEY: str | None = None
29
    VLLM_DEBUG_LOG_API_SERVER_RESPONSE: bool = False
30
31
32
33
    S3_ACCESS_KEY_ID: str | None = None
    S3_SECRET_ACCESS_KEY: str | None = None
    S3_ENDPOINT_URL: str | None = None
    VLLM_MODEL_REDIRECT_PATH: str | None = None
34
35
    VLLM_CACHE_ROOT: str = os.path.expanduser("~/.cache/vllm")
    VLLM_CONFIG_ROOT: str = os.path.expanduser("~/.config/vllm")
36
37
38
39
    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 = ""
40
    VLLM_CONFIGURE_LOGGING: bool = True
41
    VLLM_LOGGING_LEVEL: str = "INFO"
42
    VLLM_LOGGING_PREFIX: str = ""
43
    VLLM_LOGGING_STREAM: str = "ext://sys.stdout"
44
    VLLM_LOGGING_CONFIG_PATH: str | None = None
Nick Hill's avatar
Nick Hill committed
45
46
    VLLM_LOGGING_COLOR: str = "auto"
    NO_COLOR: bool = False
47
    VLLM_LOG_STATS_INTERVAL: float = 10.0
48
    VLLM_TRACE_FUNCTION: int = 0
49
50
    VLLM_USE_FLASHINFER_SAMPLER: bool | None = None
    VLLM_PP_LAYER_PARTITION: str | None = None
51
    VLLM_PP_LAYER_PARTITION_D: Optional[str] = None
52
    VLLM_CPU_KVCACHE_SPACE: int | None = 0
53
    VLLM_CPU_OMP_THREADS_BIND: str = ""
54
    VLLM_CPU_NUM_OF_RESERVED_CPU: int | None = None
55
    VLLM_CPU_SGL_KERNEL: bool = False
56
    VLLM_XLA_CACHE_PATH: str = os.path.join(VLLM_CACHE_ROOT, "xla_cache")
57
    VLLM_XLA_CHECK_RECOMPILATION: bool = False
58
    VLLM_FUSED_MOE_CHUNK_SIZE: int = 16 * 1024
59
    VLLM_ENABLE_FUSED_MOE_ACTIVATION_CHUNKING: bool = True
60
    VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE: Literal["auto", "nccl", "shm"] = "auto"
61
    VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM: bool = False
62
    VLLM_USE_RAY_WRAPPED_PP_COMM: bool = True
63
    VLLM_XLA_USE_SPMD: bool = False
64
    VLLM_WORKER_MULTIPROC_METHOD: Literal["fork", "spawn"] = "spawn"
65
    VLLM_ASSETS_CACHE: str = os.path.join(VLLM_CACHE_ROOT, "assets")
66
    VLLM_ASSETS_CACHE_MODEL_CLEAN: bool = False
67
    VLLM_IMAGE_FETCH_TIMEOUT: int = 5
68
    VLLM_VIDEO_FETCH_TIMEOUT: int = 30
69
    VLLM_AUDIO_FETCH_TIMEOUT: int = 10
70
    VLLM_MEDIA_URL_ALLOW_REDIRECTS: bool = True
71
    VLLM_MEDIA_LOADING_THREAD_COUNT: int = 8
72
    VLLM_MAX_AUDIO_CLIP_FILESIZE_MB: int = 25
73
    VLLM_VIDEO_LOADER_BACKEND: str = "opencv"
74
    VLLM_MEDIA_CONNECTOR: str = "http"
75
    VLLM_MM_HASHER_ALGORITHM: str = "blake3"
76
    VLLM_TARGET_DEVICE: str = "cuda"
77
    VLLM_MAIN_CUDA_VERSION: str = "12.9"
78
    VLLM_FLOAT32_MATMUL_PRECISION: Literal["highest", "high", "medium"] = "highest"
79
80
    MAX_JOBS: str | None = None
    NVCC_THREADS: str | None = None
81
    VLLM_USE_PRECOMPILED: bool = False
82
    VLLM_SKIP_PRECOMPILED_VERSION_SUFFIX: bool = False
83
    VLLM_DOCKER_BUILD_CONTEXT: bool = False
84
    VLLM_KEEP_ALIVE_ON_ENGINE_DEATH: bool = False
85
    CMAKE_BUILD_TYPE: Literal["Debug", "Release", "RelWithDebInfo"] | None = None
86
    VERBOSE: bool = False
87
    VLLM_ALLOW_LONG_MAX_MODEL_LEN: bool = False
88
    VLLM_RPC_TIMEOUT: int = 10000  # ms
89
    VLLM_HTTP_TIMEOUT_KEEP_ALIVE: int = 5  # seconds
90
91
    VLLM_PLUGINS: list[str] | None = None
    VLLM_LORA_RESOLVER_CACHE_DIR: str | None = None
92
93
94
    # Deprecated env variables for profiling, kept for backward compatibility
    # See also vllm/config/profiler.py and `--profiler-config` argument
    VLLM_TORCH_CUDA_PROFILE: str | None = None
95
    VLLM_TORCH_PROFILER_DIR: str | None = None
96
97
98
99
100
101
102
103
104
105
    VLLM_TORCH_PROFILER_RECORD_SHAPES: str | None = None
    VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY: str | None = None
    VLLM_TORCH_PROFILER_DISABLE_ASYNC_LLM: str | None = None
    VLLM_TORCH_PROFILER_WITH_STACK: str | None = None
    VLLM_TORCH_PROFILER_WITH_FLOPS: str | None = None
    VLLM_TORCH_PROFILER_USE_GZIP: str | None = None
    VLLM_TORCH_PROFILER_DUMP_CUDA_TIME_TOTAL: str | None = None
    VLLM_PROFILER_DELAY_ITERS: str | None = None
    VLLM_PROFILER_MAX_ITERS: str | None = None
    # End of deprecated env variables for profiling
106
    VLLM_USE_AOT_COMPILE: bool = False
107
    VLLM_USE_BYTECODE_HOOK: bool = False
108
    VLLM_FORCE_AOT_LOAD: bool = False
109
    VLLM_USE_MEGA_AOT_ARTIFACT: bool = False
110
    VLLM_USE_TRITON_AWQ: bool = False
111
    VLLM_ALLOW_RUNTIME_LORA_UPDATING: bool = False
112
    VLLM_SKIP_P2P_CHECK: bool = False
113
    VLLM_DISABLED_KERNELS: list[str] = []
114
    VLLM_DISABLE_PYNCCL: bool = False
115
    VLLM_ROCM_USE_AITER: bool = False
116
    VLLM_ROCM_USE_AITER_PAGED_ATTN: bool = False
117
    VLLM_ROCM_USE_AITER_LINEAR: bool = True
118
    VLLM_ROCM_USE_AITER_MOE: bool = True
119
    VLLM_ROCM_USE_AITER_RMSNORM: bool = True
120
    VLLM_ROCM_USE_AITER_MLA: bool = True
121
    VLLM_ROCM_USE_AITER_MHA: bool = True
122
    VLLM_ROCM_USE_AITER_FP4_ASM_GEMM: bool = False
123
    VLLM_ROCM_USE_AITER_TRITON_ROPE: bool = False
124
    VLLM_ROCM_USE_AITER_FP8BMM: bool = True
125
    VLLM_ROCM_USE_AITER_FP4BMM: bool = True
126
    VLLM_ROCM_USE_AITER_UNIFIED_ATTENTION: bool = False
127
    VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS: bool = False
128
    VLLM_ROCM_USE_AITER_TRITON_GEMM: bool = True
129
    VLLM_ROCM_USE_SKINNY_GEMM: bool = True
130
    VLLM_ROCM_FP8_PADDING: bool = True
131
    VLLM_ROCM_MOE_PADDING: bool = True
132
    VLLM_ROCM_CUSTOM_PAGED_ATTN: bool = True
133
    VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT: bool = False
134
    VLLM_ENABLE_V1_MULTIPROCESSING: bool = True
135
    VLLM_LOG_BATCHSIZE_INTERVAL: float = -1
136
    VLLM_DISABLE_COMPILE_CACHE: bool = False
zhangshao's avatar
zhangshao committed
137
138
139
    Q_SCALE_CONSTANT: int = 10
    K_SCALE_CONSTANT: int = 10
    V_SCALE_CONSTANT: int = 10
140
    VLLM_SERVER_DEV_MODE: bool = False
141
    VLLM_V1_OUTPUT_PROC_CHUNK_SIZE: int = 128
142
    VLLM_MLA_DISABLE: bool = False
143
144
    VLLM_RAY_PER_WORKER_GPUS: float = 1.0
    VLLM_RAY_BUNDLE_INDICES: str = ""
145
    VLLM_CUDART_SO_PATH: str | None = None
146
    VLLM_DP_RANK: int = 0
147
    VLLM_DP_RANK_LOCAL: int = -1
148
    VLLM_DP_SIZE: int = 1
149
    VLLM_USE_STANDALONE_COMPILE: bool = True
150
151
    VLLM_DP_MASTER_IP: str = ""
    VLLM_DP_MASTER_PORT: int = 0
152
    VLLM_MOE_DP_CHUNK_SIZE: int = 256
153
    VLLM_ENABLE_MOE_DP_CHUNK: bool = True
154
    VLLM_RANDOMIZE_DP_DUMMY_INPUTS: bool = False
155
    VLLM_RAY_DP_PACK_STRATEGY: Literal["strict", "fill", "span"] = "strict"
156
    VLLM_MARLIN_USE_ATOMIC_ADD: bool = False
157
    VLLM_MARLIN_INPUT_DTYPE: Literal["int8", "fp8"] | None = None
158
    VLLM_MXFP4_USE_MARLIN: bool | None = None
159
    VLLM_DEEPEPLL_NVFP4_DISPATCH: bool = False
160
    VLLM_V1_USE_OUTLINES_CACHE: bool = False
guanyu1's avatar
guanyu1 committed
161
    VLLM_1D_MROPE: bool = False
162
    VLLM_ENCODER_CACHE_SIZE: int | None = None
163
    VLLM_TPU_BUCKET_PADDING_GAP: int = 0
164
    VLLM_TPU_MOST_MODEL_LEN: int | None = None
165
    VLLM_TPU_USING_PATHWAYS: bool = False
166
    VLLM_USE_DEEP_GEMM: bool = True
167
    VLLM_MOE_USE_DEEP_GEMM: bool = True
168
    VLLM_USE_DEEP_GEMM_E8M0: bool = True
169
    VLLM_USE_DEEP_GEMM_TMA_ALIGNED_SCALES: bool = True
170
171
172
173
174
    VLLM_DEEP_GEMM_WARMUP: Literal[
        "skip",
        "full",
        "relax",
    ] = "relax"
175
    VLLM_USE_FUSED_MOE_GROUPED_TOPK: bool = True
176
    VLLM_BLOCKSCALE_FP8_GEMM_FLASHINFER: bool = False
177
    VLLM_USE_FLASHINFER_MOE_FP16: bool = False
178
179
    VLLM_USE_FLASHINFER_MOE_FP8: bool = False
    VLLM_USE_FLASHINFER_MOE_FP4: bool = False
180
181
182
    VLLM_FLASHINFER_MOE_BACKEND: Literal["throughput", "latency", "masked_gemm"] = (
        "latency"
    )
183
    VLLM_FLASHINFER_WORKSPACE_BUFFER_SIZE: int = 394 * 1024 * 1024
184
    VLLM_XGRAMMAR_CACHE_MB: int = 0
185
    VLLM_MSGPACK_ZERO_COPY_THRESHOLD: int = 256
186
    VLLM_ALLOW_INSECURE_SERIALIZATION: bool = False
187
    VLLM_DISABLE_REQUEST_ID_RANDOMIZATION: bool = False
Robert Shaw's avatar
Robert Shaw committed
188
    VLLM_NIXL_SIDE_CHANNEL_HOST: str = "localhost"
189
    VLLM_NIXL_SIDE_CHANNEL_PORT: int = 5600
190
    VLLM_MOONCAKE_BOOTSTRAP_PORT: int = 8998
191
192
193
194
195
    VLLM_ALL2ALL_BACKEND: Literal[
        "naive",
        "pplx",
        "deepep_high_throughput",
        "deepep_low_latency",
196
        "mori",
197
198
199
        "allgather_reducescatter",
        "flashinfer_all2allv",
    ] = "allgather_reducescatter"
200
    VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE: int = 163840
201
    VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS: int = 1
202
    VLLM_SLEEP_WHEN_IDLE: bool = False
203
    VLLM_MQ_MAX_CHUNK_BYTES_MB: int = 16
204
    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: int = 300
205
    VLLM_KV_CACHE_LAYOUT: Literal["NHD", "HND"] | None = None
206
    VLLM_COMPUTE_NANS_IN_LOGITS: bool = False
207
    VLLM_USE_NVFP4_CT_EMULATIONS: bool = False
208
209
210
    VLLM_ROCM_QUICK_REDUCE_QUANTIZATION: Literal[
        "FP", "INT8", "INT6", "INT4", "NONE"
    ] = "NONE"
211
    VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16: bool = True
212
    VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB: int | None = None
213
    VLLM_NIXL_ABORT_REQUEST_TIMEOUT: int = 480
214
215
216
217
    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
218
    VLLM_MOONCAKE_ABORT_REQUEST_TIMEOUT: int = 480
219
    VLLM_ENABLE_CUDAGRAPH_GC: bool = False
220
    VLLM_LOOPBACK_IP: str = ""
221
    VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE: bool = True
222
    VLLM_ENABLE_RESPONSES_API_STORE: bool = False
223
    VLLM_NVFP4_GEMM_BACKEND: str | None = None
224
    VLLM_HAS_FLASHINFER_CUBIN: bool = False
225
226
    VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8: bool = False
    VLLM_USE_FLASHINFER_MOE_MXFP4_BF16: bool = False
xiao-llm's avatar
xiao-llm committed
227
    VLLM_ROCM_FP8_MFMA_PAGE_ATTN: bool = False
228
    VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS: bool = False
229
    VLLM_ALLREDUCE_USE_SYMM_MEM: bool = True
230
    VLLM_TUNED_CONFIG_FOLDER: str | None = None
231
    VLLM_GPT_OSS_SYSTEM_TOOL_MCP_LABELS: set[str] = set()
232
    VLLM_USE_EXPERIMENTAL_PARSER_CONTEXT: bool = False
233
    VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS: bool = False
234
    VLLM_TOOL_JSON_ERROR_AUTOMATIC_RETRY: bool = False
235
    VLLM_CUSTOM_SCOPES_FOR_PROFILING: bool = False
236
    VLLM_NVTX_SCOPES_FOR_PROFILING: bool = False
237
    VLLM_KV_EVENTS_USE_INT_BLOCK_HASHES: bool = True
238
    VLLM_OBJECT_STORAGE_SHM_BUFFER_NAME: str = "VLLM_OBJECT_STORAGE_SHM_BUFFER"
239
    VLLM_DEEPEP_BUFFER_SIZE_MB: int = 1024
240
241
    VLLM_DEEPEP_HIGH_THROUGHPUT_FORCE_INTRA_NODE: bool = False
    VLLM_DEEPEP_LOW_LATENCY_USE_MNNVL: bool = False
242
    VLLM_DBO_COMM_SMS: int = 20
243
244
    VLLM_PATTERN_MATCH_DEBUG: str | None = None
    VLLM_DEBUG_DUMP_PATH: str | None = None
245
246
    VLLM_ENABLE_INDUCTOR_MAX_AUTOTUNE: bool = True
    VLLM_ENABLE_INDUCTOR_COORDINATE_DESCENT_TUNING: bool = True
247
    VLLM_USE_NCCL_SYMM_MEM: bool = False
248
    VLLM_NCCL_INCLUDE_PATH: str | None = None
249
    VLLM_USE_FBGEMM: bool = False
250
    VLLM_GC_DEBUG: str = ""
251
    VLLM_DEBUG_WORKSPACE: bool = False
252
    VLLM_DISABLE_SHARED_EXPERTS_STREAM: bool = False
253
    VLLM_SHARED_EXPERTS_STREAM_TOKEN_THRESHOLD: int = 256
254
    VLLM_COMPILE_CACHE_SAVE_FORMAT: Literal["binary", "unpacked"] = "binary"
Woosuk Kwon's avatar
Woosuk Kwon committed
255
    VLLM_USE_V2_MODEL_RUNNER: bool = False
256
    VLLM_LOG_MODEL_INSPECTION: bool = False
257
    VLLM_DEBUG_MFU_METRICS: bool = False
258
    VLLM_DISABLE_LOG_LOGO: bool = False
259
    VLLM_LORA_DISABLE_PDL: bool = False
260
    
261
    # add envs
zhuwenwen's avatar
zhuwenwen committed
262
    VLLM_OPTEST_URLS_PORT: int | None = None
263
264
    VLLM_OPTEST_MODELS_PATH: str = ""
    VLLM_USE_TRITON_PREFIX_FLASH_ATTN: bool = False
265
    VLLM_USE_FLASH_ATTN_FP8: bool = False
266
    VLLM_USE_QUERY_QUANT: bool = False
267
268
269
270
271
272
    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
273
    VLLM_CUSTOM_CACHE: bool = False
zhuwenwen's avatar
zhuwenwen committed
274
    VLLM_CUSTOM_ALLREDUCE_SUPPORTED_WORLDSIZE_MAX: int = 16
zhuwenwen's avatar
zhuwenwen committed
275
    VLLM_ENFORCE_EAGER_BS_THRESHOLD: int | None  = None
276
    VLLM_HAS_CONTEXT_DEFAULT: bool = False
277
    VLLM_USE_NN: bool = False
278
    VLLM_ENABLE_TBO: bool = False
279
    VLLM_ENABLE_MOE_FUSED_GATE: bool = False
280
    VLLM_USE_FLASH_ATTN_PA: bool = False
zhuwenwen's avatar
zhuwenwen committed
281
    VLLM_USE_APEX_RN: bool = False
282
    VLLM_USE_GLOBAL_CACHE13: bool = False
283
284
    VLLM_USE_LIGHTOP: bool = False
    VLLM_USE_OPT_CAT: bool = False
zhuwenwen's avatar
zhuwenwen committed
285
286
    VLLM_USE_LIGHTOP_MOE_SUM: bool = False
    VLLM_USE_LIGHTOP_MOE_ALIGN: bool = False
287
    VLLM_USE_MERGE_ATTN_STATES_OPT: bool = False
王敏's avatar
王敏 committed
288
    USE_FUSED_RMS_QUANT: bool = False
xuxz's avatar
xuxz committed
289
290
    VLLM_P2P_ASYNC: bool = False
    VLLM_P2P_BUF_TOKENS: int = 30000
291
    USE_FUSED_SILU_MUL_QUANT: bool = False
zhuwenwen's avatar
zhuwenwen committed
292
    VLLM_USE_PD_SPLIT: bool = False
zhuwenwen's avatar
zhuwenwen committed
293
    VLLM_USE_PP_SYNC: bool = False
294
    VLLM_USE_PIECEWISE: bool = False
295
    VLLM_USE_V32_ENCODE: bool = False
296
297
298
    VLLM_USE_FUSE_SILU_AND_MUL: bool = False
    VLLM_USE_OPT_RESHAPE_AND_CACHE: bool = False
    VLLM_USE_TOPK_RENORM: bool = False
299
    VLLM_USE_FUSED_RMS_ROPE: bool = False
300
    VLLM_USE_FUSED_FILL_RMS_CAT: bool = False
301
    VLLM_USE_CAT_MLA: bool = False
302
    FP8_USE_MIXED_BATCH: bool = False
303
    VLLM_W8A8_BACKEND: int = 3
jujl1's avatar
jujl1 committed
304
    VLLM_USE_PP_BALANCE = True
305
306
307
308
309
310
    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
311
    VLLM_REJECT_SAMPLE_OPT: bool = False
312
    VLLM_USE_MOE_W16A16_TRITON: bool = False
313
    VLLM_V1_FAST_TOKEN_ID_COPY: bool = False
314
    VLLM_V1_USE_REDUCED_TOPK_TOPP_SAMPLER: bool = False
315
    VLLM_V1_USE_FUSED_QKV_SPLIT_RMS_ROPE_KVSTORE: bool = False
316
    VLLM_USE_FUSED_DTBMM: bool = False # DOUBLE TRANS BMM FP8
317
    VLLM_USE_LIGHTOP_FILL_MOE_ALIGN: bool = False
318
    VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT: bool = False
wujl5's avatar
wujl5 committed
319
    VLLM_USE_CUDA_GRAPH_SIZES: bool = False
320
    VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD: bool = False
321
    VLLM_USE_LIGHTOP_FUSED_TOPP_TOPK: bool = False
322
    VLLM_ENABLE_RAY_ASYNC_SCHEDULING: bool = False
wanghl6's avatar
wanghl6 committed
323
324
325
    USE_LIGHTOP_PER_TOKEN_GROUP_QUANT_FP8: bool = False
    USE_LIGHTOP_TOPK: bool = False
    USE_LIGHTOP_CONVERT_REQ_INDEX_TO_GLOBAL_INDEX: bool = False
326
    VLLM_DISABLE_DSA: bool = False
王敏's avatar
王敏 committed
327
328
    VLLM_MLA_CP: bool = False
    VLLM_MLA_CPLB: bool = False
329
330
331
332
333
334
335
336
337
338
339
340
341
342
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"),
    )


343
def maybe_convert_int(value: str | None) -> int | None:
344
345
346
347
348
    if value is None:
        return None
    return int(value)


349
def maybe_convert_bool(value: str | None) -> bool | None:
350
351
352
353
354
    if value is None:
        return None
    return bool(int(value))


355
356
357
358
def disable_compile_cache() -> bool:
    return bool(int(os.getenv("VLLM_DISABLE_COMPILE_CACHE", "0")))


359
def use_aot_compile() -> bool:
360
361
362
    from vllm.model_executor.layers.batch_invariant import (
        vllm_is_batch_invariant,
    )
zhuwenwen's avatar
zhuwenwen committed
363
    from vllm.platforms import current_platform
364
    from vllm.utils.torch_utils import is_torch_equal_or_newer
365

366
367
    default_value = (
        "1"
zhuwenwen's avatar
zhuwenwen committed
368
369
370
371
372
        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()
373
374
375
        else "0"
    )

376
377
378
379
    return (
        not vllm_is_batch_invariant()
        and os.environ.get("VLLM_USE_AOT_COMPILE", default_value) == "1"
    )
380
381


382
def env_with_choices(
383
    env_name: str,
384
385
    default: str | None,
    choices: list[str] | Callable[[], list[str]],
386
    case_sensitive: bool = True,
387
) -> Callable[[], str | None]:
388
389
    """
    Create a lambda that validates environment variable against allowed choices
390

391
392
393
394
395
    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
396

397
398
399
400
    Returns:
        Lambda function for environment_variables dict
    """

401
    def _get_validated_env() -> str | None:
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
        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:
417
418
419
420
            raise ValueError(
                f"Invalid value '{value}' for {env_name}. "
                f"Valid options: {actual_choices}."
            )
421
422
423
424
425
426

        return value

    return _get_validated_env


427
def env_list_with_choices(
428
429
    env_name: str,
    default: list[str],
430
    choices: list[str] | Callable[[], list[str]],
431
432
    case_sensitive: bool = True,
) -> Callable[[], list[str]]:
433
    """
434
    Create a lambda that validates environment variable
435
    containing comma-separated values against allowed choices
436

437
438
439
440
441
    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
442

443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
    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:
472
473
474
475
                raise ValueError(
                    f"Invalid value '{val}' in {env_name}. "
                    f"Valid options: {actual_choices}."
                )
476
477
478
479
480
481

        return values

    return _get_validated_env_list


482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
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


500
def get_vllm_port() -> int | None:
501
    """Get the port from VLLM_PORT environment variable.
502

503
504
    Returns:
        The port number as an integer if VLLM_PORT is set, None otherwise.
505

506
507
508
    Raises:
        ValueError: If VLLM_PORT is a URI, suggest k8s service discovery issue.
    """
509
    if "VLLM_PORT" not in os.environ:
510
511
        return None

512
    port = os.getenv("VLLM_PORT", "0")
513
514
515
516

    try:
        return int(port)
    except ValueError as err:
517
        from urllib3.util import parse_url
518

519
        parsed = parse_url(port)
520
521
522
523
524
525
        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
526
        raise ValueError(f"VLLM_PORT '{port}' must be a valid integer") from err
527
528


529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
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
550
# The start-* and end* here are used by the documentation generator
551
552
# to extract the used env vars.

553
# --8<-- [start:env-vars-definition]
554

555
logger = logging.getLogger(__name__)
556

557
environment_variables: dict[str, Callable[[], Any]] = {
558
    # ================== Installation Time Env Vars ==================
559
    # Target device of vLLM, supporting [cuda (by default),
560
    # rocm, cpu]
561
    "VLLM_TARGET_DEVICE": lambda: os.getenv("VLLM_TARGET_DEVICE", "cuda").lower(),
562
    # Main CUDA version of vLLM. This follows PyTorch but can be overridden.
563
    "VLLM_MAIN_CUDA_VERSION": lambda: os.getenv("VLLM_MAIN_CUDA_VERSION", "").lower()
564
    or "12.9",
565
    # Controls PyTorch float32 matmul precision mode within vLLM workers.
566
    # Valid options mirror torch.set_float32_matmul_precision
567
568
    "VLLM_FLOAT32_MATMUL_PRECISION": env_with_choices(
        "VLLM_FLOAT32_MATMUL_PRECISION",
569
570
        "highest",
        ["highest", "high", "medium"],
571
572
        case_sensitive=False,
    ),
573
574
    # Maximum number of compilation jobs to run in parallel.
    # By default this is the number of CPUs
575
    "MAX_JOBS": lambda: os.getenv("MAX_JOBS", None),
576
577
578
    # Number of threads to use for nvcc
    # By default this is 1.
    # If set, `MAX_JOBS` will be reduced to avoid oversubscribing the CPU.
579
    "NVCC_THREADS": lambda: os.getenv("NVCC_THREADS", None),
580
    # If set, vllm will use precompiled binaries (*.so)
581
582
583
584
585
    "VLLM_USE_PRECOMPILED": lambda: os.environ.get("VLLM_USE_PRECOMPILED", "")
    .strip()
    .lower()
    in ("1", "true")
    or bool(os.environ.get("VLLM_PRECOMPILED_WHEEL_LOCATION")),
586
587
588
589
    # 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"))
    ),
590
591
    # Used to mark that setup.py is running in a Docker build context,
    # in order to force the use of precompiled binaries.
592
593
594
595
    "VLLM_DOCKER_BUILD_CONTEXT": lambda: os.environ.get("VLLM_DOCKER_BUILD_CONTEXT", "")
    .strip()
    .lower()
    in ("1", "true"),
596
597
598
    # CMake build type
    # If not set, defaults to "Debug" or "RelWithDebInfo"
    # Available options: "Debug", "Release", "RelWithDebInfo"
599
600
601
    "CMAKE_BUILD_TYPE": env_with_choices(
        "CMAKE_BUILD_TYPE", None, ["Debug", "Release", "RelWithDebInfo"]
    ),
602
    # If set, vllm will print verbose logs during installation
603
    "VERBOSE": lambda: bool(int(os.getenv("VERBOSE", "0"))),
604
    # Root directory for vLLM configuration files
605
    # Defaults to `~/.config/vllm` unless `XDG_CONFIG_HOME` is set
606
607
608
    # 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**.
609
    "VLLM_CONFIG_ROOT": lambda: os.path.expanduser(
610
611
612
        os.getenv(
            "VLLM_CONFIG_ROOT",
            os.path.join(get_default_config_root(), "vllm"),
613
614
        )
    ),
615
    # ================== Runtime Env Vars ==================
616
    # Root directory for vLLM cache files
617
    # Defaults to `~/.cache/vllm` unless `XDG_CACHE_HOME` is set
618
    "VLLM_CACHE_ROOT": lambda: os.path.expanduser(
619
620
621
        os.getenv(
            "VLLM_CACHE_ROOT",
            os.path.join(get_default_cache_root(), "vllm"),
622
623
        )
    ),
624
625
626
627
    # 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.
628
    "VLLM_HOST_IP": lambda: os.getenv("VLLM_HOST_IP", ""),
629
    # used in distributed environment to manually set the communication port
630
631
632
    # 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.
633
    "VLLM_PORT": get_vllm_port,
634
635
    # path used for ipc when the frontend api server is running in
    # multi-processing mode to communicate with the backend engine process.
636
637
638
    "VLLM_RPC_BASE_PATH": lambda: os.getenv(
        "VLLM_RPC_BASE_PATH", tempfile.gettempdir()
    ),
639
640
    # If true, will load models from ModelScope instead of Hugging Face Hub.
    # note that the value is true or false, not numbers
641
642
643
644
    "VLLM_USE_MODELSCOPE": lambda: os.environ.get(
        "VLLM_USE_MODELSCOPE", "False"
    ).lower()
    == "true",
645
    # Interval in seconds to log a warning message when the ring buffer is full
646
647
648
    "VLLM_RINGBUFFER_WARNING_INTERVAL": lambda: int(
        os.environ.get("VLLM_RINGBUFFER_WARNING_INTERVAL", "60")
    ),
649
650
    # path to cudatoolkit home directory, under which should be bin, include,
    # and lib directories.
651
    "CUDA_HOME": lambda: os.environ.get("CUDA_HOME", None),
652
653
    # 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
654
    "VLLM_NCCL_SO_PATH": lambda: os.environ.get("VLLM_NCCL_SO_PATH", None),
655
656
    # 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`
657
    "LD_LIBRARY_PATH": lambda: os.environ.get("LD_LIBRARY_PATH", None),
658
659
    # flag to control if vllm should use triton flash attention
    "VLLM_USE_TRITON_FLASH_ATTN":
660
    lambda: (os.environ.get("VLLM_USE_TRITON_FLASH_ATTN", "False").lower() in
661
             ("true", "1")),
662
663
664
665
    # 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")
    ),
666
    # Feature flag to enable/disable Inductor standalone compile.
667
668
    # In torch <= 2.7 we ignore this flag; in torch >= 2.9 this is
    # enabled by default.
669
    "VLLM_USE_STANDALONE_COMPILE": lambda: os.environ.get(
670
        "VLLM_USE_STANDALONE_COMPILE", "1"
671
672
    )
    == "1",
673
674
    # Debug pattern matching inside custom passes.
    # Should be set to the fx.Node name (e.g. 'getitem_34' or 'scaled_mm_3').
675
676
677
    "VLLM_PATTERN_MATCH_DEBUG": lambda: os.environ.get(
        "VLLM_PATTERN_MATCH_DEBUG", None
    ),
678
679
    # Dump fx graphs to the given directory.
    # It will override CompilationConfig.debug_dump_path if set.
680
    "VLLM_DEBUG_DUMP_PATH": lambda: os.environ.get("VLLM_DEBUG_DUMP_PATH", None),
681
682
683
684
    # 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,
685
686
687
    # Feature flag to enable/disable bytecode in
    # TorchCompileWithNoGuardsWrapper.
    "VLLM_USE_BYTECODE_HOOK": lambda: bool(
688
        int(os.environ.get("VLLM_USE_BYTECODE_HOOK", "0"))
689
    ),
690
691
692
693
    # 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",
694
695
696
697
698
699
700
    # 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",
701
702
    # local rank of the process in the distributed setting, used to determine
    # the GPU device id
703
    "LOCAL_RANK": lambda: int(os.environ.get("LOCAL_RANK", "0")),
704
    # used to control the visible devices in the distributed setting
705
    "CUDA_VISIBLE_DEVICES": lambda: os.environ.get("CUDA_VISIBLE_DEVICES", None),
706
    # timeout for each iteration in the engine
707
    "VLLM_ENGINE_ITERATION_TIMEOUT_S": lambda: int(
708
        os.environ.get("VLLM_ENGINE_ITERATION_TIMEOUT_S", "120")
709
    ),
710
711
712
713
714
    # 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")
    ),
715
    # API key for vLLM API server
716
    "VLLM_API_KEY": lambda: os.environ.get("VLLM_API_KEY", None),
717
    # Whether to log responses from API Server for debugging
718
719
720
721
    "VLLM_DEBUG_LOG_API_SERVER_RESPONSE": lambda: os.environ.get(
        "VLLM_DEBUG_LOG_API_SERVER_RESPONSE", "False"
    ).lower()
    == "true",
722
    # S3 access information, used for tensorizer to load model from S3
723
724
725
    "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),
726
    # Usage stats collection
727
728
729
730
731
732
733
734
735
736
737
    "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"),
738
739
740
741
    # 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
742
743
744
    "VLLM_CONFIGURE_LOGGING": lambda: bool(
        int(os.getenv("VLLM_CONFIGURE_LOGGING", "1"))
    ),
745
    "VLLM_LOGGING_CONFIG_PATH": lambda: os.getenv("VLLM_LOGGING_CONFIG_PATH"),
746
    # this is used for configuring the default logging level
747
    "VLLM_LOGGING_LEVEL": lambda: os.getenv("VLLM_LOGGING_LEVEL", "INFO").upper(),
748
    # this is used for configuring the default logging stream
749
    "VLLM_LOGGING_STREAM": lambda: os.getenv("VLLM_LOGGING_STREAM", "ext://sys.stdout"),
750
    # if set, VLLM_LOGGING_PREFIX will be prepended to all log messages
751
    "VLLM_LOGGING_PREFIX": lambda: os.getenv("VLLM_LOGGING_PREFIX", ""),
Nick Hill's avatar
Nick Hill committed
752
753
754
755
756
    # 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",
757
758
    # If set, vllm will log stats at this interval in seconds
    # If not set, vllm will log stats every 10 seconds.
759
760
761
    "VLLM_LOG_STATS_INTERVAL": lambda: val
    if (val := float(os.getenv("VLLM_LOG_STATS_INTERVAL", "10."))) > 0.0
    else 10.0,
762
763
764
    # Trace function calls
    # If set to 1, vllm will trace function calls
    # Useful for debugging
765
    "VLLM_TRACE_FUNCTION": lambda: int(os.getenv("VLLM_TRACE_FUNCTION", "0")),
766
    # If set, vllm will use flashinfer sampler
767
768
769
770
771
    "VLLM_USE_FLASHINFER_SAMPLER": lambda: bool(
        int(os.environ["VLLM_USE_FLASHINFER_SAMPLER"])
    )
    if "VLLM_USE_FLASHINFER_SAMPLER" in os.environ
    else None,
772
    # Pipeline stage partition strategy
773
    "VLLM_PP_LAYER_PARTITION": lambda: os.getenv("VLLM_PP_LAYER_PARTITION", None),
774
775
776
777
778
    
    # Pipeline stage partition strategy
    "VLLM_PP_LAYER_PARTITION_D":
    lambda: os.getenv("VLLM_PP_LAYER_PARTITION_D", None),

779
    # (CPU backend only) CPU key-value cache space.
780
    # default is None and will be set as 4 GB
781
782
783
    "VLLM_CPU_KVCACHE_SPACE": lambda: int(os.getenv("VLLM_CPU_KVCACHE_SPACE", "0"))
    if "VLLM_CPU_KVCACHE_SPACE" in os.environ
    else None,
784
785
    # (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 '|'.
786
    "VLLM_CPU_OMP_THREADS_BIND": lambda: os.getenv("VLLM_CPU_OMP_THREADS_BIND", "auto"),
787
788
    # (CPU backend only) CPU cores not used by OMP threads .
    # Those CPU cores will not be used by OMP threads of a rank.
789
790
791
792
793
    "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,
794
    # (CPU backend only) whether to use SGL kernels, optimized for small batch.
795
    "VLLM_CPU_SGL_KERNEL": lambda: bool(int(os.getenv("VLLM_CPU_SGL_KERNEL", "0"))),
796
797
798
799
800
801
802
    # 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
803
804
805
    "VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE": env_with_choices(
        "VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE", "auto", ["auto", "nccl", "shm"]
    ),
806
    # If the env var is set, it enables GPU communication overlap
807
    # (experimental feature) in Ray's Compiled Graph.
808
809
810
    "VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM": lambda: bool(
        int(os.getenv("VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM", "0"))
    ),
811
812
813
    # 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.
814
815
816
    "VLLM_USE_RAY_WRAPPED_PP_COMM": lambda: bool(
        int(os.getenv("VLLM_USE_RAY_WRAPPED_PP_COMM", "1"))
    ),
817
818
    # Use dedicated multiprocess context for workers.
    # Both spawn and fork work
819
    "VLLM_WORKER_MULTIPROC_METHOD": env_with_choices(
820
        "VLLM_WORKER_MULTIPROC_METHOD", "spawn", ["spawn", "fork"]
821
    ),
822
    # Path to the cache for storing downloaded assets
823
    "VLLM_ASSETS_CACHE": lambda: os.path.expanduser(
824
825
826
        os.getenv(
            "VLLM_ASSETS_CACHE",
            os.path.join(get_default_cache_root(), "vllm", "assets"),
827
828
        )
    ),
829
830
    # If the env var is set, we will clean model file in
    # this path $VLLM_ASSETS_CACHE/model_streamer/$model_name
831
832
833
    "VLLM_ASSETS_CACHE_MODEL_CLEAN": lambda: bool(
        int(os.getenv("VLLM_ASSETS_CACHE_MODEL_CLEAN", "0"))
    ),
834
835
    # Timeout for fetching images when serving multimodal models
    # Default is 5 seconds
836
    "VLLM_IMAGE_FETCH_TIMEOUT": lambda: int(os.getenv("VLLM_IMAGE_FETCH_TIMEOUT", "5")),
837
    # Timeout for fetching videos when serving multimodal models
838
    # Default is 30 seconds
839
840
841
    "VLLM_VIDEO_FETCH_TIMEOUT": lambda: int(
        os.getenv("VLLM_VIDEO_FETCH_TIMEOUT", "30")
    ),
842
    # Timeout for fetching audio when serving multimodal models
843
    # Default is 10 seconds
844
845
846
    "VLLM_AUDIO_FETCH_TIMEOUT": lambda: int(
        os.getenv("VLLM_AUDIO_FETCH_TIMEOUT", "10")
    ),
847
848
    # Whether to allow HTTP redirects when fetching from media URLs.
    # Default to True
849
850
851
    "VLLM_MEDIA_URL_ALLOW_REDIRECTS": lambda: bool(
        int(os.getenv("VLLM_MEDIA_URL_ALLOW_REDIRECTS", "1"))
    ),
852
853
854
    # Max number of workers for the thread pool handling
    # media bytes loading. Set to 1 to disable parallel processing.
    # Default is 8
855
856
857
    "VLLM_MEDIA_LOADING_THREAD_COUNT": lambda: int(
        os.getenv("VLLM_MEDIA_LOADING_THREAD_COUNT", "8")
    ),
858
859
860
    # 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
861
862
863
    "VLLM_MAX_AUDIO_CLIP_FILESIZE_MB": lambda: int(
        os.getenv("VLLM_MAX_AUDIO_CLIP_FILESIZE_MB", "25")
    ),
864
865
    # Backend for Video IO
    # - "opencv": Default backend that uses OpenCV stream buffered backend.
Roger Wang's avatar
Roger Wang committed
866
    # - "identity": Returns raw video bytes for model processor to handle.
867
868
869
870
871
    #
    # 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.
872
873
874
    "VLLM_VIDEO_LOADER_BACKEND": lambda: os.getenv(
        "VLLM_VIDEO_LOADER_BACKEND", "opencv"
    ),
875
876
877
878
879
880
881
882
    # 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"),
883
884
885
886
887
888
889
890
891
892
893
    # 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,
    ),
894
895
    # Path to the XLA persistent cache directory.
    # Only used for XLA devices such as TPUs.
896
    "VLLM_XLA_CACHE_PATH": lambda: os.path.expanduser(
897
        os.getenv(
898
            "VLLM_XLA_CACHE_PATH",
899
            os.path.join(get_default_cache_root(), "vllm", "xla_cache"),
900
901
        )
    ),
902
    # If set, assert on XLA recompilation after each execution step.
903
904
905
    "VLLM_XLA_CHECK_RECOMPILATION": lambda: bool(
        int(os.getenv("VLLM_XLA_CHECK_RECOMPILATION", "0"))
    ),
906
    # Enable SPMD mode for TPU backend.
907
908
    "VLLM_XLA_USE_SPMD": lambda: bool(int(os.getenv("VLLM_XLA_USE_SPMD", "0"))),
    "VLLM_FUSED_MOE_CHUNK_SIZE": lambda: int(
909
        os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", str(16 * 1024))
910
    ),
911
912
913
    # 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.
914
915
916
    "VLLM_ENABLE_FUSED_MOE_ACTIVATION_CHUNKING": lambda: bool(
        int(os.getenv("VLLM_ENABLE_FUSED_MOE_ACTIVATION_CHUNKING", "1"))
    ),
917
918
    # If set, the OpenAI API server will stay alive even after the underlying
    # AsyncLLMEngine errors and stops serving requests
919
    "VLLM_KEEP_ALIVE_ON_ENGINE_DEATH": lambda: bool(
920
        int(os.getenv("VLLM_KEEP_ALIVE_ON_ENGINE_DEATH", "0"))
921
    ),
922
923
924
925
    # 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.
926
927
928
929
    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": lambda: (
        os.environ.get("VLLM_ALLOW_LONG_MAX_MODEL_LEN", "0").strip().lower()
        in ("1", "true")
    ),
930
931
    # If set, forces FP8 Marlin to be used for FP8 quantization regardless
    # of the hardware support for FP8 compute.
932
933
934
935
936
937
938
    "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"
    ),
939
940
    # Time in ms for the zmq client to wait for a response from the backend
    # server for simple data operations
941
    "VLLM_RPC_TIMEOUT": lambda: int(os.getenv("VLLM_RPC_TIMEOUT", "10000")),
942
    # Timeout in seconds for keeping HTTP connections alive in API server
943
944
945
    "VLLM_HTTP_TIMEOUT_KEEP_ALIVE": lambda: int(
        os.environ.get("VLLM_HTTP_TIMEOUT_KEEP_ALIVE", "5")
    ),
946
947
948
    # 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
949
950
951
    "VLLM_PLUGINS": lambda: None
    if "VLLM_PLUGINS" not in os.environ
    else os.environ["VLLM_PLUGINS"].split(","),
952
953
954
    # a local directory to look in for unrecognized LoRA adapters.
    # only works if plugins are enabled and
    # VLLM_ALLOW_RUNTIME_LORA_UPDATING is enabled.
955
956
957
    "VLLM_LORA_RESOLVER_CACHE_DIR": lambda: os.getenv(
        "VLLM_LORA_RESOLVER_CACHE_DIR", None
    ),
958
959
960
    # Enables torch CUDA profiling if set to 1.
    # Deprecated, see profiler_config.
    "VLLM_TORCH_CUDA_PROFILE": lambda: os.getenv("VLLM_TORCH_CUDA_PROFILE"),
961
    # Enables torch profiler if set.
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
    # 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")
988
989
990
    ),
    # Delay number of iterations before starting profiling when using
    # the torch/torch CUDA profiler. If set to 0, will start profiling immediately.
991
992
    # Deprecated, see profiler_config.
    "VLLM_PROFILER_DELAY_ITERS": lambda: (os.getenv("VLLM_PROFILER_DELAY_ITERS")),
993
994
    # Maximum number of iterations to profile when using the torch/torch CUDA profiler.
    # If set to 0, will not limit the number of iterations.
995
    "VLLM_PROFILER_MAX_ITERS": lambda: os.getenv("VLLM_PROFILER_MAX_ITERS"),
996
    # Control whether torch profiler gzip-compresses profiling files.
997
998
    # Deprecated, see profiler_config.
    "VLLM_TORCH_PROFILER_USE_GZIP": lambda: os.getenv("VLLM_TORCH_PROFILER_USE_GZIP"),
999
    # Control whether torch profiler dumps the self_cuda_time_total table.
1000
1001
1002
1003
    # 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")
1004
    ),
1005
    # If set, vLLM will use Triton implementations of AWQ.
1006
    "VLLM_USE_TRITON_AWQ": lambda: bool(int(os.getenv("VLLM_USE_TRITON_AWQ", "0"))),
1007
    # If set, allow loading or unloading lora adapters in runtime,
1008
1009
1010
1011
    "VLLM_ALLOW_RUNTIME_LORA_UPDATING": lambda: (
        os.environ.get("VLLM_ALLOW_RUNTIME_LORA_UPDATING", "0").strip().lower()
        in ("1", "true")
    ),
1012
1013
1014
1015
1016
1017
    # 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
1018
    "VLLM_SKIP_P2P_CHECK": lambda: os.getenv("VLLM_SKIP_P2P_CHECK", "1") == "1",
1019
1020
1021
1022
    # List of quantization kernels that should be disabled, used for testing
    # and performance comparisons. Currently only affects MPLinearKernel
    # selection
    # (kernels: MacheteLinearKernel, MarlinLinearKernel, ExllamaLinearKernel)
1023
1024
1025
    "VLLM_DISABLED_KERNELS": lambda: []
    if "VLLM_DISABLED_KERNELS" not in os.environ
    else os.environ["VLLM_DISABLED_KERNELS"].split(","),
1026
    # Disable pynccl (using torch.distributed instead)
1027
1028
1029
    "VLLM_DISABLE_PYNCCL": lambda: (
        os.getenv("VLLM_DISABLE_PYNCCL", "False").lower() in ("true", "1")
    ),
1030
1031
    # Disable aiter ops unless specifically enabled.
    # Acts as a parent switch to enable the rest of the other operations.
1032
1033
1034
    "VLLM_ROCM_USE_AITER": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER", "False").lower() in ("true", "1")
    ),
1035
1036
    # Whether to use aiter paged attention.
    # By default is disabled.
1037
1038
1039
    "VLLM_ROCM_USE_AITER_PAGED_ATTN": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_PAGED_ATTN", "False").lower() in ("true", "1")
    ),
1040
1041
1042
    # use aiter linear op if aiter ops are enabled
    # The following list of related ops
    # - scaled_mm (per-tensor / rowwise)
1043
    "VLLM_ROCM_USE_AITER_LINEAR": lambda: (
1044
        os.getenv("VLLM_ROCM_USE_AITER_LINEAR", "False").lower() in ("true", "1")
1045
    ),
1046
1047
    # Whether to use aiter moe ops.
    # By default is enabled.
1048
    "VLLM_ROCM_USE_AITER_MOE": lambda: (
1049
        os.getenv("VLLM_ROCM_USE_AITER_MOE", "False").lower() in ("true", "1")
1050
    ),
1051
    # use aiter rms norm op if aiter ops are enabled.
1052
    "VLLM_ROCM_USE_AITER_RMSNORM": lambda: (
1053
        os.getenv("VLLM_ROCM_USE_AITER_RMSNORM", "False").lower() in ("true", "1")
1054
    ),
1055
1056
    # Whether to use aiter mla ops.
    # By default is enabled.
1057
    "VLLM_ROCM_USE_AITER_MLA": lambda: (
1058
        os.getenv("VLLM_ROCM_USE_AITER_MLA", "False").lower() in ("true", "1")
1059
    ),
1060
1061
    # Whether to use aiter mha ops.
    # By default is enabled.
1062
    "VLLM_ROCM_USE_AITER_MHA": lambda: (
1063
        os.getenv("VLLM_ROCM_USE_AITER_MHA", "False").lower() in ("true", "1")
1064
    ),
1065
1066
    # Whether to use aiter fp4 gemm asm.
    # By default is disabled.
1067
1068
1069
    "VLLM_ROCM_USE_AITER_FP4_ASM_GEMM": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_FP4_ASM_GEMM", "False").lower() in ("true", "1")
    ),
1070
1071
    # Whether to use aiter rope.
    # By default is disabled.
1072
1073
    "VLLM_ROCM_USE_AITER_TRITON_ROPE": lambda: (
        os.getenv("VLLM_ROCM_USE_AITER_TRITON_ROPE", "False").lower() in ("true", "1")
1074
    ),
1075
1076
    # Whether to use aiter triton fp8 bmm kernel
    # By default is enabled.
1077
    "VLLM_ROCM_USE_AITER_FP8BMM": lambda: (
1078
        os.getenv("VLLM_ROCM_USE_AITER_FP8BMM", "False").lower() in ("true", "1")
1079
    ),
1080
1081
1082
    # Whether to use aiter triton fp4 bmm kernel
    # By default is enabled.
    "VLLM_ROCM_USE_AITER_FP4BMM": lambda: (
1083
        os.getenv("VLLM_ROCM_USE_AITER_FP4BMM", "False").lower() in ("true", "1")
1084
    ),
1085
1086
1087
1088
1089
    # 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")
    ),
1090
    # Whether to use aiter fusion shared experts ops.
1091
    # By default is disabled.
1092
    "VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS": lambda: (
1093
        os.getenv("VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS", "False").lower()
1094
1095
        in ("true", "1")
    ),
1096
1097
1098
    # Whether to use aiter triton kernels for gemm ops.
    # By default is enabled.
    "VLLM_ROCM_USE_AITER_TRITON_GEMM": lambda: (
1099
        os.getenv("VLLM_ROCM_USE_AITER_TRITON_GEMM", "False").lower() in ("true", "1")
1100
    ),
1101
    # use rocm skinny gemms
1102
1103
1104
    "VLLM_ROCM_USE_SKINNY_GEMM": lambda: (
        os.getenv("VLLM_ROCM_USE_SKINNY_GEMM", "True").lower() in ("true", "1")
    ),
1105
    # Pad the fp8 weights to 256 bytes for ROCm
1106
    "VLLM_ROCM_FP8_PADDING": lambda: bool(int(os.getenv("VLLM_ROCM_FP8_PADDING", "1"))),
1107
    # Pad the weights for the moe kernel
1108
    "VLLM_ROCM_MOE_PADDING": lambda: bool(int(os.getenv("VLLM_ROCM_MOE_PADDING", "0"))),
1109
    # custom paged attention kernel for MI3* cards
1110
1111
1112
    "VLLM_ROCM_CUSTOM_PAGED_ATTN": lambda: (
        os.getenv("VLLM_ROCM_CUSTOM_PAGED_ATTN", "True").lower() in ("true", "1")
    ),
1113
1114
1115
1116
    # 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")
    ),
1117
1118
1119
    # Custom quick allreduce kernel for MI3* cards
    # Choice of quantization level: FP, INT8, INT6, INT4 or NONE
    # Recommended for large models to get allreduce
1120
1121
1122
1123
1124
    "VLLM_ROCM_QUICK_REDUCE_QUANTIZATION": env_with_choices(
        "VLLM_ROCM_QUICK_REDUCE_QUANTIZATION",
        "NONE",
        ["FP", "INT8", "INT6", "INT4", "NONE"],
    ),
1125
1126
1127
1128
    # 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
1129
1130
1131
1132
    "VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16": lambda: (
        os.getenv("VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16", "True").lower()
        in ("true", "1")
    ),
1133
1134
1135
1136
1137
1138
    # 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.
1139
1140
1141
    "VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB": lambda: maybe_convert_int(
        os.environ.get("VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB", None)
    ),
1142
    # Divisor for dynamic query scale factor calculation for FP8 KV Cache
zhangshao's avatar
zhangshao committed
1143
    "Q_SCALE_CONSTANT": lambda: int(os.getenv("Q_SCALE_CONSTANT", "10")),
1144
    # Divisor for dynamic key scale factor calculation for FP8 KV Cache
zhangshao's avatar
zhangshao committed
1145
    "K_SCALE_CONSTANT": lambda: int(os.getenv("K_SCALE_CONSTANT", "10")),
1146
    # Divisor for dynamic value scale factor calculation for FP8 KV Cache
zhangshao's avatar
zhangshao committed
1147
    "V_SCALE_CONSTANT": lambda: int(os.getenv("V_SCALE_CONSTANT", "10")),
1148
    # If set, enable multiprocessing in LLM for the V1 code path.
1149
1150
1151
1152
1153
1154
    "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")
    ),
1155
    "VLLM_DISABLE_COMPILE_CACHE": disable_compile_cache,
1156
1157
1158
    # If set, vllm will run in development mode, which will enable
    # some additional endpoints for developing and debugging,
    # e.g. `/reset_prefix_cache`
1159
    "VLLM_SERVER_DEV_MODE": lambda: bool(int(os.getenv("VLLM_SERVER_DEV_MODE", "0"))),
1160
1161
1162
1163
1164
1165
1166
    # 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.
1167
1168
1169
    "VLLM_V1_OUTPUT_PROC_CHUNK_SIZE": lambda: int(
        os.getenv("VLLM_V1_OUTPUT_PROC_CHUNK_SIZE", "128")
    ),
1170
    # If set, vLLM will disable the MLA attention optimizations.
1171
    "VLLM_MLA_DISABLE": lambda: bool(int(os.getenv("VLLM_MLA_DISABLE", "0"))),
1172
    # If set, vLLM will pick up the provided Flash Attention MLA
1173
1174
1175
    # 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.
1176
1177
1178
    "VLLM_RAY_PER_WORKER_GPUS": lambda: float(
        os.getenv("VLLM_RAY_PER_WORKER_GPUS", "1.0")
    ),
1179
1180
1181
    # 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"
1182
    "VLLM_RAY_BUNDLE_INDICES": lambda: os.getenv("VLLM_RAY_BUNDLE_INDICES", ""),
1183
1184
    # In some system, find_loaded_library() may not work. So we allow users to
    # specify the path through environment variable VLLM_CUDART_SO_PATH.
1185
    "VLLM_CUDART_SO_PATH": lambda: os.getenv("VLLM_CUDART_SO_PATH", None),
1186
    # Rank of the process in the data parallel setting
1187
    "VLLM_DP_RANK": lambda: int(os.getenv("VLLM_DP_RANK", "0")),
1188
1189
    # Rank of the process in the data parallel setting.
    # Defaults to VLLM_DP_RANK when not set.
1190
1191
1192
    "VLLM_DP_RANK_LOCAL": lambda: int(
        os.getenv("VLLM_DP_RANK_LOCAL", sys.modules[__name__].VLLM_DP_RANK)
    ),
1193
    # World size of the data parallel setting
1194
    "VLLM_DP_SIZE": lambda: int(os.getenv("VLLM_DP_SIZE", "1")),
1195
    # IP address of the master node in the data parallel setting
1196
    "VLLM_DP_MASTER_IP": lambda: os.getenv("VLLM_DP_MASTER_IP", "127.0.0.1"),
1197
    # Port of the master node in the data parallel setting
1198
    "VLLM_DP_MASTER_PORT": lambda: int(os.getenv("VLLM_DP_MASTER_PORT", "0")),
1199
1200
1201
1202
1203
    # 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.
1204
    "VLLM_MOE_DP_CHUNK_SIZE": lambda: int(os.getenv("VLLM_MOE_DP_CHUNK_SIZE", "256")),
1205
1206
1207
    "VLLM_ENABLE_MOE_DP_CHUNK": lambda: bool(
        int(os.getenv("VLLM_ENABLE_MOE_DP_CHUNK", "1"))
    ),
1208
    # Randomize inputs during dummy runs when using Data Parallel
1209
1210
1211
1212
    "VLLM_RANDOMIZE_DP_DUMMY_INPUTS": lambda: os.environ.get(
        "VLLM_RANDOMIZE_DP_DUMMY_INPUTS", "0"
    )
    == "1",
1213
1214
1215
1216
1217
1218
1219
    # 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;
1220
1221
1222
    # - "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;
1223
1224
1225
1226
    # 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"
    ),
1227
    # Whether to use S3 path for model loading in CI via RunAI Streamer
1228
    "VLLM_CI_USE_S3": lambda: os.environ.get("VLLM_CI_USE_S3", "0") == "1",
1229
    # Use model_redirect to redirect the model name to a local folder.
1230
1231
1232
1233
1234
    # `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
1235
1236
1237
    "VLLM_MODEL_REDIRECT_PATH": lambda: os.environ.get(
        "VLLM_MODEL_REDIRECT_PATH", None
    ),
1238
    # Whether to use atomicAdd reduce in gptq/awq marlin kernel.
1239
1240
1241
1242
    "VLLM_MARLIN_USE_ATOMIC_ADD": lambda: os.environ.get(
        "VLLM_MARLIN_USE_ATOMIC_ADD", "0"
    )
    == "1",
1243
    # Whether to use marlin kernel in mxfp4 quantization method
1244
1245
1246
    "VLLM_MXFP4_USE_MARLIN": lambda: maybe_convert_bool(
        os.environ.get("VLLM_MXFP4_USE_MARLIN", None)
    ),
1247
1248
1249
1250
    # The activation dtype for marlin kernel
    "VLLM_MARLIN_INPUT_DTYPE": env_with_choices(
        "VLLM_MARLIN_INPUT_DTYPE", None, ["int8", "fp8"]
    ),
1251
1252
1253
1254
1255
1256
    # 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"))
    ),
1257
1258
1259
    # 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.
1260
1261
1262
1263
    "VLLM_V1_USE_OUTLINES_CACHE": lambda: os.environ.get(
        "VLLM_V1_USE_OUTLINES_CACHE", "0"
    )
    == "1",
1264
1265
    # Gap between padding buckets for the forward pass. So we have
    # 8, we will run forward pass with [16, 24, 32, ...].
1266
1267
1268
1269
1270
1271
1272
1273
    "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)
    ),
1274
    # Whether using Pathways
1275
1276
1277
    "VLLM_TPU_USING_PATHWAYS": lambda: bool(
        "proxy" in os.getenv("JAX_PLATFORMS", "").lower()
    ),
1278
    # Allow use of DeepGemm kernels for fused moe ops.
1279
    "VLLM_USE_DEEP_GEMM": lambda: bool(int(os.getenv("VLLM_USE_DEEP_GEMM", "1"))),
1280
1281
1282
1283
    # 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"))
    ),
1284
    # Whether to use E8M0 scaling when DeepGEMM is used on Blackwell GPUs.
1285
1286
1287
    "VLLM_USE_DEEP_GEMM_E8M0": lambda: bool(
        int(os.getenv("VLLM_USE_DEEP_GEMM_E8M0", "1"))
    ),
1288
1289
1290
1291
    # 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"))
    ),
1292
1293
1294
1295
    # 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.
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
    # 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",
        ],
1311
    ),
1312
    # Whether to use fused grouped_topk used for MoE expert selection.
1313
1314
1315
    "VLLM_USE_FUSED_MOE_GROUPED_TOPK": lambda: bool(
        int(os.getenv("VLLM_USE_FUSED_MOE_GROUPED_TOPK", "1"))
    ),
1316
1317
1318
1319
1320
    # 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"))
    ),
1321
    # Allow use of FlashInfer MoE kernels for fused moe ops.
1322
1323
1324
    "VLLM_USE_FLASHINFER_MOE_FP16": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_FP16", "0"))
    ),
1325
    # Allow use of FlashInfer MoE kernels for fused moe ops.
1326
1327
1328
    "VLLM_USE_FLASHINFER_MOE_FP8": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_FP8", "0"))
    ),
1329
    # Allow use of FlashInfer CUTLASS kernels for fused moe ops.
1330
1331
1332
    "VLLM_USE_FLASHINFER_MOE_FP4": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_FP4", "0"))
    ),
1333
1334
    # If set to 1, use the FlashInfer
    # MXFP8 (activation) x MXFP4 (weight) MoE backend.
1335
1336
1337
    "VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8", "0"))
    ),
1338
1339
1340
1341
    # 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.
1342
1343
1344
    "VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS", "0"))
    ),
1345
1346
    # If set to 1, use the FlashInfer
    # BF16 (activation) x MXFP4 (weight) MoE backend.
1347
1348
1349
    "VLLM_USE_FLASHINFER_MOE_MXFP4_BF16": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_MXFP4_BF16", "0"))
    ),
1350
1351
1352
    # 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.
1353
    "VLLM_XGRAMMAR_CACHE_MB": lambda: int(os.getenv("VLLM_XGRAMMAR_CACHE_MB", "512")),
1354
1355
1356
1357
1358
1359
1360
    # 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.
1361
1362
1363
    "VLLM_MSGPACK_ZERO_COPY_THRESHOLD": lambda: int(
        os.getenv("VLLM_MSGPACK_ZERO_COPY_THRESHOLD", "256")
    ),
1364
1365
1366
    # 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.
1367
1368
1369
    "VLLM_ALLOW_INSECURE_SERIALIZATION": lambda: bool(
        int(os.getenv("VLLM_ALLOW_INSECURE_SERIALIZATION", "0"))
    ),
1370
1371
1372
1373
1374
    # Temporary: skip adding random suffix to internal request IDs. May be
    # needed for KV connectors that match request IDs across instances.
    "VLLM_DISABLE_REQUEST_ID_RANDOMIZATION": lambda: bool(
        int(os.getenv("VLLM_DISABLE_REQUEST_ID_RANDOMIZATION", "1"))
    ),
Robert Shaw's avatar
Robert Shaw committed
1375
    # IP address used for NIXL handshake between remote agents.
1376
1377
1378
    "VLLM_NIXL_SIDE_CHANNEL_HOST": lambda: os.getenv(
        "VLLM_NIXL_SIDE_CHANNEL_HOST", "localhost"
    ),
Robert Shaw's avatar
Robert Shaw committed
1379
    # Port used for NIXL handshake between remote agents.
1380
1381
1382
    "VLLM_NIXL_SIDE_CHANNEL_PORT": lambda: int(
        os.getenv("VLLM_NIXL_SIDE_CHANNEL_PORT", "5600")
    ),
1383
1384
1385
1386
    # Port used for Mooncake handshake between remote agents.
    "VLLM_MOONCAKE_BOOTSTRAP_PORT": lambda: int(
        os.getenv("VLLM_MOONCAKE_BOOTSTRAP_PORT", "8998")
    ),
1387
1388
    # [DEPRECATED - will be removed in v0.15.0] all2all backend for vllm's
    # expert parallel communication. Use --all2all-backend CLI argument instead.
1389
    # Available options:
1390
1391
1392
    # - "naive": naive all2all implementation using broadcasts
    # - "allgather_reducescatter": all2all implementation based on allgather and
    #  reducescatter
1393
    # - "pplx": use pplx kernels
1394
1395
    # - "deepep_high_throughput", use deepep high-throughput kernels
    # - "deepep_low_latency", use deepep low-latency kernels
1396
    # - "mori", use MoRI kernels
1397
    # - "flashinfer_all2allv", use flashinfer alltoallv kernels for mnnvl
1398
1399
    "VLLM_ALL2ALL_BACKEND": env_with_choices(
        "VLLM_ALL2ALL_BACKEND",
1400
        None,
1401
1402
1403
1404
1405
        [
            "naive",
            "pplx",
            "deepep_high_throughput",
            "deepep_low_latency",
1406
            "mori",
1407
1408
1409
1410
            "allgather_reducescatter",
            "flashinfer_all2allv",
        ],
    ),
1411
1412
    # Flashinfer MoE backend for vLLM's fused Mixture-of-Experts support.
    # Both require compute capability 10.0 or above.
1413
1414
1415
1416
1417
    # Available options:
    # - "throughput":  [default]
    #     Uses CUTLASS kernels optimized for high-throughput batch inference.
    # - "latency":
    #     Uses TensorRT-LLM kernels optimized for low-latency inference.
1418
    "VLLM_FLASHINFER_MOE_BACKEND": env_with_choices(
1419
1420
1421
        "VLLM_FLASHINFER_MOE_BACKEND",
        "latency",
        ["throughput", "latency", "masked_gemm"],
1422
    ),
1423
1424
1425
1426
    # 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))
    ),
1427
1428
1429
1430
    # 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.
1431
1432
1433
    "VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE": lambda: int(
        os.getenv("VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE", "163840")
    ),
1434
1435
1436
1437
    # 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> }
1438
    # Unspecified world sizes will fall back to
1439
    #     { 2: 64, 4: 1, <everything else>: 0.5 }
1440
1441
1442
    "VLLM_FLASHINFER_ALLREDUCE_FUSION_THRESHOLDS_MB": lambda: json.loads(
        os.getenv("VLLM_FLASHINFER_ALLREDUCE_FUSION_THRESHOLDS_MB", "{}")
    ),
1443
1444
1445
    # MoE routing strategy selector.
    # See `RoutingSimulator.get_available_strategies()` # for available
    # strategies.
1446
    # Custom routing strategies can be registered by
1447
1448
    # RoutingSimulator.register_strategy()
    # Note: custom strategies may not produce correct model outputs
1449
1450
1451
    "VLLM_MOE_ROUTING_SIMULATION_STRATEGY": lambda: os.environ.get(
        "VLLM_MOE_ROUTING_SIMULATION_STRATEGY", ""
    ).lower(),
1452
    # Regex timeout for use by the vLLM tool parsing plugins.
1453
1454
1455
    "VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS": lambda: int(
        os.getenv("VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS", "1")
    ),
1456
1457
    # Reduce CPU usage when vLLM is idle. Enabling this will incur small
    # latency penalty when a request eventually comes.
1458
    "VLLM_SLEEP_WHEN_IDLE": lambda: bool(int(os.getenv("VLLM_SLEEP_WHEN_IDLE", "0"))),
1459
1460
1461
    # 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.
1462
1463
1464
    "VLLM_MQ_MAX_CHUNK_BYTES_MB": lambda: int(
        os.getenv("VLLM_MQ_MAX_CHUNK_BYTES_MB", "16")
    ),
1465
1466
    # Timeout in seconds for execute_model RPC calls in multiprocessing
    # executor (only applies when TP > 1).
1467
1468
1469
    "VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS": lambda: int(
        os.getenv("VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS", "300")
    ),
1470
1471
1472
1473
1474
1475
1476
    # 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.
1477
1478
1479
    "VLLM_KV_CACHE_LAYOUT": env_with_choices(
        "VLLM_KV_CACHE_LAYOUT", None, ["NHD", "HND"]
    ),
1480
1481
1482
    # 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.
1483
1484
1485
    "VLLM_COMPUTE_NANS_IN_LOGITS": lambda: bool(
        int(os.getenv("VLLM_COMPUTE_NANS_IN_LOGITS", "0"))
    ),
1486
1487
1488
    # Controls whether or not emulations are used for NVFP4
    # generations on machines < 100 for compressed-tensors
    # models
1489
1490
1491
    "VLLM_USE_NVFP4_CT_EMULATIONS": lambda: bool(
        int(os.getenv("VLLM_USE_NVFP4_CT_EMULATIONS", "0"))
    ),
1492
1493
1494
1495
    # 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.
1496
1497
1498
    "VLLM_NIXL_ABORT_REQUEST_TIMEOUT": lambda: int(
        os.getenv("VLLM_NIXL_ABORT_REQUEST_TIMEOUT", "480")
    ),
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
    # 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")),
1513
1514
1515
1516
    # Timeout (in seconds) for MooncakeConnector in PD disaggregated setup.
    "VLLM_MOONCAKE_ABORT_REQUEST_TIMEOUT": lambda: int(
        os.getenv("VLLM_MOONCAKE_ABORT_REQUEST_TIMEOUT", "480")
    ),
1517
1518
    # If set, it means we pre-downloaded cubin files and flashinfer will
    # read the cubin files directly.
1519
1520
1521
    "VLLM_HAS_FLASHINFER_CUBIN": lambda: bool(
        int(os.getenv("VLLM_HAS_FLASHINFER_CUBIN", "0"))
    ),
1522
1523
1524
1525
    # Supported options:
    # - "flashinfer-cudnn": use flashinfer cudnn GEMM backend
    # - "flashinfer-trtllm": use flashinfer trtllm GEMM backend
    # - "flashinfer-cutlass": use flashinfer cutlass GEMM backend
1526
    # - "marlin": use marlin GEMM backend (for GPUs without native FP4 support)
1527
1528
1529
1530
    # - <none>: automatically pick an available backend
    "VLLM_NVFP4_GEMM_BACKEND": env_with_choices(
        "VLLM_NVFP4_GEMM_BACKEND",
        None,
1531
1532
1533
1534
1535
        [
            "flashinfer-cudnn",
            "flashinfer-trtllm",
            "flashinfer-cutlass",
            "cutlass",
1536
            "marlin",
1537
        ],
1538
    ),
1539
1540
1541
    # 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.
1542
1543
1544
    "VLLM_ENABLE_CUDAGRAPH_GC": lambda: bool(
        int(os.getenv("VLLM_ENABLE_CUDAGRAPH_GC", "0"))
    ),
1545
    # Used to force set up loopback IP
1546
    "VLLM_LOOPBACK_IP": lambda: os.getenv("VLLM_LOOPBACK_IP", ""),
1547
1548
1549
    # Used to set the process name prefix for vLLM processes.
    # This is useful for debugging and monitoring purposes.
    # The default value is "VLLM".
1550
    "VLLM_PROCESS_NAME_PREFIX": lambda: os.getenv("VLLM_PROCESS_NAME_PREFIX", "VLLM"),
1551
1552
1553
1554
1555
1556
1557
    # 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.
1558
    "VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE": lambda: bool(
1559
        int(os.getenv("VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE", "1"))
1560
    ),
1561
1562
    # 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
1563
1564
    # messages for those requests in memory. By default, this is disabled (0),
    # and the "store" option is ignored.
1565
1566
1567
1568
1569
    # 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.
1570
1571
1572
    "VLLM_ENABLE_RESPONSES_API_STORE": lambda: bool(
        int(os.getenv("VLLM_ENABLE_RESPONSES_API_STORE", "0"))
    ),
xiao-llm's avatar
xiao-llm committed
1573
    # If set, use the fp8 mfma in rocm paged attention.
1574
1575
1576
    "VLLM_ROCM_FP8_MFMA_PAGE_ATTN": lambda: bool(
        int(os.getenv("VLLM_ROCM_FP8_MFMA_PAGE_ATTN", "0"))
    ),
1577
    # Whether to use pytorch symmetric memory for allreduce
1578
    "VLLM_ALLREDUCE_USE_SYMM_MEM": lambda: bool(
1579
        int(os.getenv("VLLM_ALLREDUCE_USE_SYMM_MEM", "1"))
1580
    ),
1581
1582
1583
1584
    # 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"))
    ),
1585
    # Allows vllm to find tuned config under customized folder
1586
    "VLLM_TUNED_CONFIG_FOLDER": lambda: os.getenv("VLLM_TUNED_CONFIG_FOLDER", None),
1587
1588
1589
1590
1591
1592
1593
1594
1595
    # 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"],
    ),
1596
    # Allows harmony instructions to be injected on system messages
1597
1598
1599
    "VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS": lambda: bool(
        int(os.getenv("VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS", "0"))
    ),
1600
1601
1602
1603
1604
1605
    # 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"))
    ),
1606
    # Add optional custom scopes for profiling, disable to avoid overheads
1607
1608
1609
    "VLLM_CUSTOM_SCOPES_FOR_PROFILING": lambda: bool(
        int(os.getenv("VLLM_CUSTOM_SCOPES_FOR_PROFILING", "0"))
    ),
1610
    # Add optional nvtx scopes for profiling, disable to avoid overheads
1611
1612
1613
    "VLLM_NVTX_SCOPES_FOR_PROFILING": lambda: bool(
        int(os.getenv("VLLM_NVTX_SCOPES_FOR_PROFILING", "0"))
    ),
1614
1615
    # Represent block hashes in KV cache events as 64-bit integers instead of
    # raw bytes. Defaults to True for backward compatibility.
1616
1617
1618
    "VLLM_KV_EVENTS_USE_INT_BLOCK_HASHES": lambda: bool(
        int(os.getenv("VLLM_KV_EVENTS_USE_INT_BLOCK_HASHES", "1"))
    ),
1619
1620
    # Name of the shared memory buffer used for object storage.
    # Only effective when mm_config.mm_processor_cache_type == "shm".
1621
1622
1623
1624
1625
    # 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}",
1626
    ),
1627
    # The size in MB of the buffers (NVL and RDMA) used by DeepEP
1628
1629
1630
    "VLLM_DEEPEP_BUFFER_SIZE_MB": lambda: int(
        os.getenv("VLLM_DEEPEP_BUFFER_SIZE_MB", "1024")
    ),
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
    # 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"))
    ),
1642
1643
    # 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
1644
    "VLLM_DBO_COMM_SMS": lambda: int(os.getenv("VLLM_DBO_COMM_SMS", "20")),
1645
1646
1647
    # 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)
1648
1649
1650
    "VLLM_ENABLE_INDUCTOR_MAX_AUTOTUNE": lambda: bool(
        int(os.getenv("VLLM_ENABLE_INDUCTOR_MAX_AUTOTUNE", "1"))
    ),
1651
1652
    # If set to 1, enable coordinate_descent_tuning;
    # By default, this is enabled (1)
1653
1654
1655
    "VLLM_ENABLE_INDUCTOR_COORDINATE_DESCENT_TUNING": lambda: bool(
        int(os.getenv("VLLM_ENABLE_INDUCTOR_COORDINATE_DESCENT_TUNING", "1"))
    ),
1656
    # Flag to enable NCCL symmetric memory allocation and registration
1657
1658
1659
    "VLLM_USE_NCCL_SYMM_MEM": lambda: bool(
        int(os.getenv("VLLM_USE_NCCL_SYMM_MEM", "0"))
    ),
1660
    # NCCL header path
1661
    "VLLM_NCCL_INCLUDE_PATH": lambda: os.environ.get("VLLM_NCCL_INCLUDE_PATH", None),
1662
1663
    # Flag to enable FBGemm kernels on model execution
    "VLLM_USE_FBGEMM": lambda: bool(int(os.getenv("VLLM_USE_FBGEMM", "0"))),
1664
1665
1666
1667
1668
1669
    # 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", ""),
1670
1671
1672
    # Debug workspace allocations.
    # logging of workspace resize operations.
    "VLLM_DEBUG_WORKSPACE": lambda: bool(int(os.getenv("VLLM_DEBUG_WORKSPACE", "0"))),
1673
    # Disables parallel execution of shared_experts via separate cuda stream
1674
1675
    "VLLM_DISABLE_SHARED_EXPERTS_STREAM": lambda: bool(
        int(os.getenv("VLLM_DISABLE_SHARED_EXPERTS_STREAM", "0"))
1676
    ),
1677
1678
1679
1680
1681
1682
1683
    # 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))
    ),
1684
1685
1686
1687
1688
1689
1690
1691
1692
    # 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
1693
1694
1695
1696
    # Flag to enable v2 model runner.
    "VLLM_USE_V2_MODEL_RUNNER": lambda: bool(
        int(os.getenv("VLLM_USE_V2_MODEL_RUNNER", "0"))
    ),
1697
1698
1699
1700
1701
1702
    # 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"))
    ),
1703
1704
1705
1706
    # Debug logging for --enable-mfu-metrics
    "VLLM_DEBUG_MFU_METRICS": lambda: bool(
        int(os.getenv("VLLM_DEBUG_MFU_METRICS", "0"))
    ),
1707
1708
    # Disable logging of vLLM logo at server startup time.
    "VLLM_DISABLE_LOG_LOGO": lambda: bool(int(os.getenv("VLLM_DISABLE_LOG_LOGO", "0"))),
1709
1710
1711
    # 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"))),
1712
    
1713
1714
1715
    # add envs
    
    # used in optest environment to manually set the https port
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
    '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")),
    
1730
1731
1732
1733
    # 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"))),
    
1734
1735
1736
1737
1738
    # 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
1739
1740
1741
1742
    # If set, vLLM will use FLASH MLA attention optimizations.
    "VLLM_USE_FLASH_MLA":
    lambda: bool(int(os.getenv("VLLM_USE_FLASH_MLA", "1"))),
    
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
    # 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":
1764
    lambda: bool(int(os.environ.get("VLLM_PCIE_USE_CUSTOM_ALLREDUCE", "1"))),
1765
    
1766
1767
    # flag to control vllm to use optimized kernels
    "VLLM_CUSTOM_CACHE":
1768
    lambda: bool(int(os.environ.get("VLLM_CUSTOM_CACHE", "1"))),
1769
    
zhuwenwen's avatar
zhuwenwen committed
1770
1771
1772
1773
    # 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")),
    
1774
1775
1776
    # 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")),
1777

1778
1779
1780
1781
1782
1783
    # '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
1784
    lambda: bool(int(os.getenv("VLLM_HAS_CONTEXT_DEFAULT", "1"))),
1785
1786
1787
    
    # If set, vLLM will transpose weight to use nn layout
    "VLLM_USE_NN":
zhuwenwen's avatar
zhuwenwen committed
1788
    lambda: (os.environ.get("VLLM_USE_NN", "True").lower() in
1789
             ("true", "1")),
1790

1791
1792
1793
    # Enable two batch overlap.
    "VLLM_ENABLE_TBO":
    lambda: bool(int(os.getenv("VLLM_ENABLE_TBO", "0"))),
1794
1795
1796

    # If set, vLLM will enable the moe_fused_gate kernel.
    "VLLM_ENABLE_MOE_FUSED_GATE":
zhuwenwen's avatar
zhuwenwen committed
1797
    lambda: bool(int(os.getenv("VLLM_ENABLE_MOE_FUSED_GATE", "1"))),
zhuwenwen's avatar
zhuwenwen committed
1798
    
1799
1800
    # vLLM will use FlashAttention Backend for page attention computation on rocm
    "VLLM_USE_FLASH_ATTN_PA":
zhuwenwen's avatar
zhuwenwen committed
1801
    lambda: (os.environ.get("VLLM_USE_FLASH_ATTN_PA", "True").lower() in
zhuwenwen's avatar
zhuwenwen committed
1802
             ("true", "1")),
zhuwenwen's avatar
zhuwenwen committed
1803
1804
1805
1806
1807
    
    # vLLM will use apex for rmsnorm
    "VLLM_USE_APEX_RN":
    lambda: (os.environ.get("VLLM_USE_APEX_RN", "False").lower() in
             ("true", "1")),
1808
1809
1810
    
    # vLLM will use global cache for moe
    "VLLM_USE_GLOBAL_CACHE13":
1811
        lambda: (os.environ.get("VLLM_USE_GLOBAL_CACHE13", "False").lower() in
1812
                 ("true", "1")),
1813
        
1814
1815
1816
    # vLLM will use lightop for deepseek-v3
    "VLLM_USE_LIGHTOP":
        lambda: (os.environ.get("VLLM_USE_LIGHTOP", "False").lower() in
1817
                 ("true", "1")),
1818
        
1819
1820
1821
    # 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
1822
                 ("true", "1")), 
zhuwenwen's avatar
zhuwenwen committed
1823
1824
1825
1826
1827
1828
1829
    # 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
1830
1831
1832
1833
                 ("true", "1")),    
    # vllm will use fused cat and mla
    "VLLM_USE_CAT_MLA":
        lambda: (os.getenv('VLLM_USE_CAT_MLA', 'False').lower() in
1834
1835
1836
1837
1838
                 ("true", "1")),
    # vllm will use fused cat and mla
    "FP8_USE_MIXED_BATCH":
        lambda: (os.getenv('FP8_USE_MIXED_BATCH', 'False').lower() in
                 ("true", "1")),                                     
1839
1840
1841
1842
    # 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
1843
1844
    # vllm will use rmsquant fused op
    "USE_FUSED_RMS_QUANT":
王敏's avatar
王敏 committed
1845
        lambda: bool(int(os.getenv("USE_FUSED_RMS_QUANT", "0"))),
xuxz's avatar
xuxz committed
1846
1847
1848
1849
1850
1851
    # vllm pd separation will be used async
    "VLLM_P2P_ASYNC":
    lambda: bool(int(os.getenv("VLLM_P2P_ASYNC", "0"))),
    # pd separation p2p async buf tokens
    "VLLM_P2P_BUF_TOKENS":
    lambda: int(os.getenv("VLLM_P2P_BUF_TOKENS", "30000")),
1852
1853
    # vllm will use silu_mul_quant fused op
    "USE_FUSED_SILU_MUL_QUANT":
1854
1855
        lambda: (os.getenv("USE_FUSED_SILU_MUL_QUANT", "False").lower() in
                ("true", "1")),
1856

zhuwenwen's avatar
zhuwenwen committed
1857
1858
    # vLLM will split prefill and decode, not mix up
    "VLLM_USE_PD_SPLIT":
1859
        lambda: (os.environ.get("VLLM_USE_PD_SPLIT", "True").lower() in
zhuwenwen's avatar
zhuwenwen committed
1860
                 ("true", "1")), 
zhuwenwen's avatar
zhuwenwen committed
1861
1862
1863
1864
    # vLLM will sync to avoid pp vmfault
    "VLLM_USE_PP_SYNC":
        lambda: (os.environ.get("VLLM_USE_PP_SYNC", "False").lower() in
                 ("true", "1")), 
1865
1866
    # vLLM will use piecewise
    "VLLM_USE_PIECEWISE":
王敏's avatar
王敏 committed
1867
        lambda: (os.environ.get("VLLM_USE_PIECEWISE", "False").lower() in
1868
                 ("true", "1")), 
1869
1870
    # vllm will use encoding_dsv32.py for dpsk-v32
    "VLLM_USE_V32_ENCODE":
1871
        lambda: (os.environ.get('VLLM_USE_V32_ENCODE', 'False').lower() in
1872
                 ("true", "1")),  
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
    # 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
1885
        (os.environ.get("VLLM_USE_TOPK_RENORM", "False").lower() in
1886
                ("true", "1")),
1887
1888
    # vLLM will use fused RMS + RoPE kernel
    "VLLM_USE_FUSED_RMS_ROPE":
1889
        lambda: (os.environ.get("VLLM_USE_FUSED_RMS_ROPE", "True").lower() in
1890
                 ("true", "1")),
1891
1892
1893
1894
    # 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
1895
1896
1897
    "VLLM_USE_PP_BALANCE":
        lambda: (os.environ.get("VLLM_USE_PP_BALANCE", "True").lower() in
                 ("true", "1")),
1898
1899
1900
1901
1902
    # W8A8 GEMM backend selection for vLLM quantized models.
    # lightop/triton: 1
    # cutlass: 2 (will remove in the future)
    # blaslt: 3 (default)
    # rocblas: others
1903
1904
1905
    "VLLM_W8A8_BACKEND": lambda: int(
            1 if "gfx928" in torch.cuda.get_device_properties("cuda").gcnArchName.split(':')[0] else os.getenv("VLLM_W8A8_BACKEND", "3")
    ),
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
    # 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
1927
1928
1929
1930
1931

    # vllm will use optimized reject sample
    "VLLM_REJECT_SAMPLE_OPT":
        lambda: (os.getenv('VLLM_REJECT_SAMPLE_OPT', 'True').lower() in
                 ("true", "1")),
1932
1933
1934
1935
    # 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
1936
1937
    "VLLM_1D_MROPE":
        lambda: (os.environ.get("VLLM_1D_MROPE", "0").lower() in ("true", "1")),
1938
1939
    "VLLM_ENCODER_CACHE_SIZE":
        lambda: maybe_convert_int(os.environ.get("VLLM_ENCODER_CACHE_SIZE", None)),
1940
1941
1942
1943
    #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")),
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
    # 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")
        ),
1957
1958
1959
1960
    # 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")),
1961
1962
1963
1964
1965

    #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")),
1966
1967
1968
1969
1970
    # 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")),

1971
1972
1973
1974
    # 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
1975
1976
1977
1978
    # 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")),
1979

1980
1981
1982
1983
1984
    # vLLM will use lightop fused moe_sum + mul + add (bias + factor)
    "VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD":
        lambda: (os.environ.get("VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD",
                                "False").lower() in ("true", "1")),

1985
1986
1987
1988
    #If set to 1/True, enable fused topk topk kernel in lightop
    "VLLM_USE_LIGHTOP_FUSED_TOPP_TOPK":
        lambda: (os.environ.get("VLLM_USE_LIGHTOP_FUSED_TOPP_TOPK", "False").lower() in
                    ("true", "1")),
1989
1990
1991
1992
1993

    #If set to 1/True, enable async scheduling in ray distribute mode
    "VLLM_ENABLE_RAY_ASYNC_SCHEDULING":
        lambda: (os.environ.get("VLLM_ENABLE_RAY_ASYNC_SCHEDULING", "False").lower() in
                    ("true", "1")),
1994

fanwl's avatar
fanwl committed
1995
1996
1997
1998
    #If set to 1/True, enable the flash attention unified path.
    "VLLM_V1_USE_FA_UNIFIED_ATTN_2D":
        lambda: (os.environ.get("VLLM_V1_USE_FA_UNIFIED_ATTN_2D", "False").lower() in
                    ("true", "1")),
wanghl6's avatar
wanghl6 committed
1999
2000
2001
2002
2003
2004
2005
2006
2007
    "USE_LIGHTOP_PER_TOKEN_GROUP_QUANT_FP8":
        lambda: (os.environ.get("USE_LIGHTOP_PER_TOKEN_GROUP_QUANT_FP8", "False").lower() in
                    ("true", "1")),   
    "USE_LIGHTOP_TOPK":
        lambda: (os.environ.get("USE_LIGHTOP_TOPK", "False").lower() in
                    ("true", "1")), 
    "USE_LIGHTOP_CONVERT_REQ_INDEX_TO_GLOBAL_INDEX":
        lambda: (os.environ.get("USE_LIGHTOP_CONVERT_REQ_INDEX_TO_GLOBAL_INDEX", "False").lower() in
                    ("true", "1")),               
2008
2009
2010
    #If set to 1/True, disenable the DSA.
    "VLLM_DISABLE_DSA":
        lambda: (os.environ.get("VLLM_DISABLE_DSA", "False").lower() in
王敏's avatar
王敏 committed
2011
2012
                    ("true", "1")),
# If set to 1/True, enable mla context parallel
王敏's avatar
王敏 committed
2013
2014
2015
2016
2017
    "VLLM_MLA_CP":
        lambda: (os.environ.get("VLLM_MLA_CP", "False").lower() in
                    ("true", "1")),
    "VLLM_MLA_CPLB":
        lambda: (os.environ.get("VLLM_MLA_CPLB", "False").lower() in
王敏's avatar
王敏 committed
2018
                    ("true", "1")),
2019
2020
}

2021
# --8<-- [end:env-vars-definition]
2022

2023

2024
def __getattr__(name: str):
2025
2026
2027
2028
2029
2030
    """
    Gets environment variables lazily.

    NOTE: After enable_envs_cache() invocation (which triggered after service
    initialization), all environment variables will be cached.
    """
2031
2032
2033
2034
2035
    if name in environment_variables:
        return environment_variables[name]()
    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")


2036
2037
2038
2039
2040
2041
def _is_envs_cache_enabled() -> bool:
    """Checked if __getattr__ is wrapped with functools.cache"""
    global __getattr__
    return hasattr(__getattr__, "cache_clear")


2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
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.
    """
2052
2053
2054
    if _is_envs_cache_enabled():
        # Avoid wrapping functools.cache multiple times
        return
2055
2056
2057
2058
2059
2060
2061
2062
2063
    # Tag __getattr__ with functools.cache
    global __getattr__
    __getattr__ = functools.cache(__getattr__)

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


2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
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__


2075
2076
def __dir__():
    return list(environment_variables.keys())
2077
2078
2079
2080
2081
2082
2083
2084
2085


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}")


2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
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",
2108
        "VLLM_FORCE_AOT_LOAD",
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
        "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
2119
        "VLLM_LOGGING_COLOR",
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
        "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",
2136
        "VLLM_OBJECT_STORAGE_SHM_BUFFER_NAME",
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
        "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
2150
        "NO_COLOR",
2151
        "VLLM_W8A8_BACKEND",
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
    }

    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
2170

2171
        factors[factor] = normalize_value(raw)
2172

2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
    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",
2193
    ]
2194

2195
2196
    for var in ray_noset_env_vars:
        factors[var] = normalize_value(os.getenv(var))
2197

2198
    return factors