"tests/kernels/test_rotary_embedding.py" did not exist on "1c046447a6d1ac3c99b9f453796f0d355d673deb"
envs.py 75.2 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import hashlib
5
import json
6
import os
7
import sys
8
import tempfile
9
import torch
10
from typing import TYPE_CHECKING, Any, Callable, Literal, Optional, Union
11
12
13

if TYPE_CHECKING:
    VLLM_HOST_IP: str = ""
14
    VLLM_PORT: Optional[int] = None
15
    VLLM_RPC_BASE_PATH: str = tempfile.gettempdir()
16
    VLLM_USE_MODELSCOPE: bool = False
17
    VLLM_RINGBUFFER_WARNING_INTERVAL: int = 60
18
19
    VLLM_NCCL_SO_PATH: Optional[str] = None
    LD_LIBRARY_PATH: Optional[str] = None
20
    VLLM_USE_TRITON_FLASH_ATTN: bool = True
21
    VLLM_V1_USE_PREFILL_DECODE_ATTENTION: bool = False
22
    VLLM_USE_AITER_UNIFIED_ATTENTION: bool = False
23
    VLLM_FLASH_ATTN_VERSION: Optional[int] = None
24
25
26
27
28
29
30
    LOCAL_RANK: int = 0
    CUDA_VISIBLE_DEVICES: Optional[str] = None
    VLLM_ENGINE_ITERATION_TIMEOUT_S: int = 60
    VLLM_API_KEY: Optional[str] = None
    S3_ACCESS_KEY_ID: Optional[str] = None
    S3_SECRET_ACCESS_KEY: Optional[str] = None
    S3_ENDPOINT_URL: Optional[str] = None
31
    VLLM_MODEL_REDIRECT_PATH: Optional[str] = None
32
33
    VLLM_CACHE_ROOT: str = os.path.expanduser("~/.cache/vllm")
    VLLM_CONFIG_ROOT: str = os.path.expanduser("~/.config/vllm")
34
35
    VLLM_USAGE_STATS_SERVER: str = "https://stats.vllm.ai"
    VLLM_NO_USAGE_STATS: bool = False
36
    VLLM_DISABLE_FLASHINFER_PREFILL: bool = False
37
38
39
    VLLM_DO_NOT_TRACK: bool = False
    VLLM_USAGE_SOURCE: str = ""
    VLLM_CONFIGURE_LOGGING: int = 1
40
    VLLM_LOGGING_LEVEL: str = "INFO"
41
    VLLM_LOGGING_PREFIX: str = ""
42
    VLLM_LOGGING_STREAM: str = "ext://sys.stdout"
43
    VLLM_LOGGING_CONFIG_PATH: Optional[str] = None
44
    VLLM_LOGITS_PROCESSOR_THREADS: Optional[int] = None
45
    VLLM_LOG_STATS_INTERVAL: float = 10.
46
47
    VLLM_TRACE_FUNCTION: int = 0
    VLLM_ATTENTION_BACKEND: Optional[str] = None
48
    VLLM_USE_FLASHINFER_SAMPLER: Optional[bool] = None
49
    VLLM_PP_LAYER_PARTITION: Optional[str] = None
50
    VLLM_CPU_KVCACHE_SPACE: Optional[int] = 0
51
    VLLM_CPU_OMP_THREADS_BIND: str = ""
52
    VLLM_CPU_NUM_OF_RESERVED_CPU: Optional[int] = None
53
    VLLM_CPU_MOE_PREPACK: bool = True
54
    VLLM_CPU_SGL_KERNEL: bool = False
55
    VLLM_XLA_CACHE_PATH: str = os.path.join(VLLM_CACHE_ROOT, "xla_cache")
56
    VLLM_XLA_CHECK_RECOMPILATION: bool = False
57
    VLLM_FUSED_MOE_CHUNK_SIZE: int = 64 * 1024
58
    VLLM_ENABLE_FUSED_MOE_ACTIVATION_CHUNKING: bool = True
59
    VLLM_USE_RAY_SPMD_WORKER: bool = False
60
    VLLM_USE_RAY_COMPILED_DAG: bool = False
61
62
    VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE: Literal["auto", "nccl",
                                                    "shm"] = "auto"
63
    VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM: bool = False
64
    VLLM_USE_RAY_WRAPPED_PP_COMM: bool = True
65
    VLLM_XLA_USE_SPMD: bool = False
66
    VLLM_WORKER_MULTIPROC_METHOD: Literal["fork", "spawn"] = "spawn"
67
    VLLM_ASSETS_CACHE: str = os.path.join(VLLM_CACHE_ROOT, "assets")
68
    VLLM_ASSETS_CACHE_MODEL_CLEAN: bool = False
69
    VLLM_IMAGE_FETCH_TIMEOUT: int = 5
70
    VLLM_VIDEO_FETCH_TIMEOUT: int = 30
71
    VLLM_AUDIO_FETCH_TIMEOUT: int = 10
72
    VLLM_MEDIA_URL_ALLOW_REDIRECTS: bool = True
73
    VLLM_MEDIA_LOADING_THREAD_COUNT: int = 8
74
    VLLM_MAX_AUDIO_CLIP_FILESIZE_MB: int = 25
75
    VLLM_VIDEO_LOADER_BACKEND: str = "opencv"
76
    VLLM_MM_INPUT_CACHE_GIB: int = 4
77
    VLLM_TARGET_DEVICE: str = "cuda"
78
    VLLM_MAIN_CUDA_VERSION: str = "12.8"
79
80
81
    MAX_JOBS: Optional[str] = None
    NVCC_THREADS: Optional[str] = None
    VLLM_USE_PRECOMPILED: bool = False
82
    VLLM_DOCKER_BUILD_CONTEXT: bool = False
83
    VLLM_TEST_USE_PRECOMPILED_NIGHTLY_WHEEL: bool = False
84
    VLLM_KEEP_ALIVE_ON_ENGINE_DEATH: bool = False
85
86
    CMAKE_BUILD_TYPE: Optional[Literal["Debug", "Release",
                                       "RelWithDebInfo"]] = None
87
    VERBOSE: bool = False
88
    VLLM_ALLOW_LONG_MAX_MODEL_LEN: bool = False
89
    VLLM_RPC_TIMEOUT: int = 10000  # ms
90
    VLLM_HTTP_TIMEOUT_KEEP_ALIVE: int = 5  # seconds
91
    VLLM_PLUGINS: Optional[list[str]] = None
92
    VLLM_LORA_RESOLVER_CACHE_DIR: Optional[str] = None
93
    VLLM_TORCH_PROFILER_DIR: Optional[str] = None
94
95
96
97
    VLLM_TORCH_PROFILER_RECORD_SHAPES: bool = False
    VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY: bool = False
    VLLM_TORCH_PROFILER_WITH_STACK: bool = True
    VLLM_TORCH_PROFILER_WITH_FLOPS: bool = False
98
    VLLM_USE_TRITON_AWQ: bool = False
99
    VLLM_ALLOW_RUNTIME_LORA_UPDATING: bool = False
100
    VLLM_SKIP_P2P_CHECK: bool = False
101
    VLLM_DISABLED_KERNELS: list[str] = []
102
    VLLM_DISABLE_NCCL_FOR_DP_SYNCHRONIZATION: bool = False
103
    VLLM_USE_V1: bool = True
104
    VLLM_ROCM_USE_AITER: bool = False
105
    VLLM_ROCM_USE_AITER_PAGED_ATTN: bool = False
106
    VLLM_ROCM_USE_AITER_LINEAR: bool = True
107
    VLLM_ROCM_USE_AITER_MOE: bool = True
108
    VLLM_ROCM_USE_AITER_RMSNORM: bool = True
109
    VLLM_ROCM_USE_AITER_MLA: bool = True
110
    VLLM_ROCM_USE_AITER_MHA: bool = True
111
112
    VLLM_ROCM_USE_AITER_FP4_ASM_GEMM: bool = False
    VLLM_ROCM_USE_TRITON_ROPE: bool = False
113
    VLLM_ROCM_USE_AITER_FP8BMM: bool = True
114
    VLLM_ROCM_USE_SKINNY_GEMM: bool = True
115
    VLLM_ROCM_FP8_PADDING: bool = True
116
    VLLM_ROCM_MOE_PADDING: bool = True
117
    VLLM_ROCM_CUSTOM_PAGED_ATTN: bool = True
118
    VLLM_ENABLE_V1_MULTIPROCESSING: bool = True
119
    VLLM_LOG_BATCHSIZE_INTERVAL: float = -1
120
    VLLM_DISABLE_COMPILE_CACHE: bool = False
zhuwenwen's avatar
zhuwenwen committed
121
122
123
    Q_SCALE_CONSTANT: int = 10
    K_SCALE_CONSTANT: int = 10
    V_SCALE_CONSTANT: int = 10
124
    VLLM_SERVER_DEV_MODE: bool = False
125
    VLLM_V1_OUTPUT_PROC_CHUNK_SIZE: int = 128
126
    VLLM_MLA_DISABLE: bool = False
127
    VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH: int = 32
128
129
    VLLM_RAY_PER_WORKER_GPUS: float = 1.0
    VLLM_RAY_BUNDLE_INDICES: str = ""
130
    VLLM_CUDART_SO_PATH: Optional[str] = None
131
    VLLM_DP_RANK: int = 0
132
    VLLM_DP_RANK_LOCAL: int = -1
133
    VLLM_DP_SIZE: int = 1
134
    VLLM_USE_STANDALONE_COMPILE: bool = False
135
136
    VLLM_DP_MASTER_IP: str = ""
    VLLM_DP_MASTER_PORT: int = 0
137
    VLLM_MOE_DP_CHUNK_SIZE: int = 256
138
    VLLM_RANDOMIZE_DP_DUMMY_INPUTS: bool = False
139
    VLLM_MARLIN_USE_ATOMIC_ADD: bool = False
140
    VLLM_MXFP4_USE_MARLIN: Optional[bool] = None
141
    VLLM_V0_USE_OUTLINES_CACHE: bool = False
142
    VLLM_V1_USE_OUTLINES_CACHE: bool = False
143
    VLLM_TPU_BUCKET_PADDING_GAP: int = 0
144
    VLLM_TPU_MOST_MODEL_LEN: Optional[int] = None
145
    VLLM_TPU_USING_PATHWAYS: bool = False
146
    VLLM_USE_DEEP_GEMM: bool = True
147
    VLLM_USE_DEEP_GEMM_E8M0: bool = True
148
    VLLM_USE_DEEP_GEMM_E8M0_HOPPER: bool = False
149
    VLLM_SKIP_DEEP_GEMM_WARMUP: bool = False
150
    VLLM_USE_FUSED_MOE_GROUPED_TOPK: bool = True
151
    VLLM_USE_FLASHINFER_MOE_FP16: bool = False
152
153
    VLLM_USE_FLASHINFER_MOE_FP8: bool = False
    VLLM_USE_FLASHINFER_MOE_FP4: bool = False
154
155
    VLLM_FLASHINFER_MOE_BACKEND: Literal["throughput",
                                         "latency"] = "throughput"
156
    VLLM_XGRAMMAR_CACHE_MB: int = 0
157
    VLLM_MSGPACK_ZERO_COPY_THRESHOLD: int = 256
158
    VLLM_ALLOW_INSECURE_SERIALIZATION: bool = False
Robert Shaw's avatar
Robert Shaw committed
159
160
    VLLM_NIXL_SIDE_CHANNEL_HOST: str = "localhost"
    VLLM_NIXL_SIDE_CHANNEL_PORT: int = 5557
161
162
163
    VLLM_ALL2ALL_BACKEND: Literal["naive", "pplx",
                                  "deepep_high_throughput",
                                  "deepep_low_latency",
164
165
                                  "allgather_reducescatter",
                                  "flashinfer_all2allv"] = \
166
                                  "allgather_reducescatter"
167
    VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE: int = 163840
168
    VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS: int = 1
169
    VLLM_SLEEP_WHEN_IDLE: bool = False
170
    VLLM_MQ_MAX_CHUNK_BYTES_MB: int = 16
171
    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: int = 300
172
    VLLM_KV_CACHE_LAYOUT: Optional[Literal["NHD", "HND"]] = None
173
    VLLM_COMPUTE_NANS_IN_LOGITS: bool = False
174
    VLLM_USE_NVFP4_CT_EMULATIONS: bool = False
175
176
    VLLM_ROCM_QUICK_REDUCE_QUANTIZATION: Literal["FP", "INT8", "INT6", "INT4",
                                                 "NONE"] = "NONE"
177
178
    VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16: bool = True
    VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB: Optional[int] = None
179
    VLLM_NIXL_ABORT_REQUEST_TIMEOUT: int = 120
180
    VLLM_USE_CUDNN_PREFILL: bool = False
181
    VLLM_ENABLE_CUDAGRAPH_GC: bool = False
182
    VLLM_LOOPBACK_IP: str = ""
183
    VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE: bool = False
184
    VLLM_ENABLE_RESPONSES_API_STORE: bool = False
185
    VLLM_USE_TRTLLM_ATTENTION: Optional[str] = None
186
    VLLM_FLASHINFER_DISABLE_Q_QUANTIZATION: bool = False
187
    VLLM_HAS_FLASHINFER_CUBIN: bool = False
188
189
    VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8: bool = False
    VLLM_USE_FLASHINFER_MOE_MXFP4_BF16: bool = False
xiao-llm's avatar
xiao-llm committed
190
    VLLM_ROCM_FP8_MFMA_PAGE_ATTN: bool = False
191
    VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS: bool = False
192
    VLLM_ALLREDUCE_USE_SYMM_MEM: bool = True
193
    VLLM_TUNED_CONFIG_FOLDER: Optional[str] = None
194
    VLLM_DISABLE_PAD_FOR_CUDAGRAPH: bool = False
195
    VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS: bool = False
196
    VLLM_CUSTOM_SCOPES_FOR_PROFILING: bool = False
197
    VLLM_NVTX_SCOPES_FOR_PROFILING: bool = False
198
    VLLM_KV_EVENTS_USE_INT_BLOCK_HASHES: bool = True
199
    VLLM_OBJECT_STORAGE_SHM_BUFFER_NAME: str = "VLLM_OBJECT_STORAGE_SHM_BUFFER"
200
201
    VLLM_DEEPEP_BUFFER_SIZE_MB: int = 1024
    VLLM_DBO_COMM_SMS: int = 20
202
    GPT_OSS_SYSTEM_TOOL_MCP_LABELS: list[str] = []
203
    VLLM_PATTERN_MATCH_DEBUG: Optional[str] = None
204
205
    VLLM_ENABLE_INDUCTOR_MAX_AUTOTUNE: bool = True
    VLLM_ENABLE_INDUCTOR_COORDINATE_DESCENT_TUNING: bool = True
206
207
    VLLM_USE_NCCL_SYMM_MEM: bool = False
    VLLM_NCCL_INCLUDE_PATH: Optional[str] = None
208
    VLLM_USE_FBGEMM: bool = False
zhuwenwen's avatar
zhuwenwen committed
209
    
210
211
212
213
    # add envs
    VLLM_OPTEST_URLS_PORT: Optional[int] = None
    VLLM_OPTEST_MODELS_PATH: str = ""
    VLLM_USE_TRITON_PREFIX_FLASH_ATTN: bool = False
zhuwenwen's avatar
zhuwenwen committed
214
    VLLM_USE_QUERY_QUANT: bool = False
215
216
217
218
219
    VLLM_USE_FLASH_MLA: bool = False
    VLLM_USE_OPT_OP: bool = False
    VLLM_USE_PA_PRINT_PARAM: bool = False 
    VLLM_SPEC_DECODE_EAGER: bool = False
    VLLM_PCIE_USE_CUSTOM_ALLREDUCE: bool = False
220
    VLLM_CUSTOM_CACHE: bool = False
zhuwenwen's avatar
zhuwenwen committed
221
    VLLM_CUSTOM_ALLREDUCE_SUPPORTED_WORLDSIZE_MAX: int = 16
222
223
    VLLM_ENFORCE_EAGER_BS_THRESHOLD: Optional[int] = None
    VLLM_HAS_CONTEXT_DEFAULT: bool = False
224
    VLLM_USE_NN: bool = False
225
    VLLM_ENABLE_TBO: bool = False
226
    VLLM_ENABLE_MOE_FUSED_GATE: bool = False
227
    VLLM_USE_FLASH_ATTN_PA: bool = False
zhuwenwen's avatar
zhuwenwen committed
228
    VLLM_USE_APEX_RN: bool = False
229
    VLLM_USE_GLOBAL_CACHE13: bool = False
230
    VLLM_USE_LIGHTOP: bool = False
231
    VLLM_USE_OPT_ZEROS: bool = False
232
    VLLM_USE_OPT_CAT: bool = False
zhuwenwen's avatar
zhuwenwen committed
233
    VLLM_USE_OPT_MOE_SUM: bool = False
234
    VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD: bool = False
zhuwenwen's avatar
zhuwenwen committed
235
236
    VLLM_USE_LIGHTOP_MOE_SUM: bool = False
    VLLM_USE_LIGHTOP_MOE_ALIGN: bool = False
237
    VLLM_USE_MERGE_ATTN_STATES_OPT: bool = False
zhuwenwen's avatar
zhuwenwen committed
238
239
    USE_FUSED_RMS_QUANT: bool = False
    USE_FUSED_SILU_MUL_QUANT: bool = False
zhuwenwen's avatar
zhuwenwen committed
240
    VLLM_USE_PD_SPLIT: bool = False
zhuwenwen's avatar
zhuwenwen committed
241
    VLLM_USE_PP_SYNC: bool = False
242
    VLLM_USE_PIECEWISE: bool = False
243
    VLLM_USE_V32_ENCODE: bool = False
244
245
246
    VLLM_USE_FUSE_SILU_AND_MUL: bool = False
    VLLM_USE_OPT_RESHAPE_AND_CACHE: bool = False
    VLLM_USE_TOPK_RENORM: bool = False
247
    VLLM_USE_FUSED_RMS_ROPE: bool = False
248
    VLLM_V1_FAST_TOKEN_ID_COPY: bool = False
249
    VLLM_V1_USE_REDUCED_TOPK_TOPP_SAMPLER: bool = False
250
    VLLM_W8A8_BACKEND: int = 3
251
    VLLM_USE_MOE_W16A16_TRITON: bool = False
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266

def get_default_cache_root():
    return os.getenv(
        "XDG_CACHE_HOME",
        os.path.join(os.path.expanduser("~"), ".cache"),
    )


def get_default_config_root():
    return os.getenv(
        "XDG_CONFIG_HOME",
        os.path.join(os.path.expanduser("~"), ".config"),
    )


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


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


279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
def env_with_choices(
        env_name: str,
        default: Optional[str],
        choices: Union[list[str], Callable[[], list[str]]],
        case_sensitive: bool = True) -> Callable[[], Optional[str]]:
    """
    Create a lambda that validates environment variable against allowed choices
    
    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
        
    Returns:
        Lambda function for environment_variables dict
    """

    def _get_validated_env() -> Optional[str]:
        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:
            raise ValueError(f"Invalid value '{value}' for {env_name}. "
                             f"Valid options: {actual_choices}.")

        return value

    return _get_validated_env


321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
def env_list_with_choices(
        env_name: str,
        default: list[str],
        choices: Union[list[str], Callable[[], list[str]]],
        case_sensitive: bool = True) -> Callable[[], list[str]]:
    """
    Create a lambda that validates environment variable 
    containing comma-separated values against allowed choices
    
    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
        
    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:
                raise ValueError(f"Invalid value '{val}' in {env_name}. "
                                 f"Valid options: {actual_choices}.")

        return values

    return _get_validated_env_list


373
374
def get_vllm_port() -> Optional[int]:
    """Get the port from VLLM_PORT environment variable.
375

376
377
    Returns:
        The port number as an integer if VLLM_PORT is set, None otherwise.
378

379
380
381
382
383
384
385
386
387
388
389
390
    Raises:
        ValueError: If VLLM_PORT is a URI, suggest k8s service discovery issue.
    """
    if 'VLLM_PORT' not in os.environ:
        return None

    port = os.getenv('VLLM_PORT', '0')

    try:
        return int(port)
    except ValueError as err:
        from urllib.parse import urlparse
391
392
393
394
395
396
397
        parsed = urlparse(port)
        if parsed.scheme:
            raise ValueError(
                f"VLLM_PORT '{port}' appears to be a URI. "
                "This may be caused by a Kubernetes service discovery issue,"
                "check the warning in: https://docs.vllm.ai/en/stable/serving/env_vars.html"
            ) from None
398
399
400
401
        raise ValueError(
            f"VLLM_PORT '{port}' must be a valid integer") from err


402
403
404
# The begin-* and end* here are used by the documentation generator
# to extract the used env vars.

405
# --8<-- [start:env-vars-definition]
406

407
environment_variables: dict[str, Callable[[], Any]] = {
408
409
410

    # ================== Installation Time Env Vars ==================

411
    # Target device of vLLM, supporting [cuda (by default),
412
    # rocm, cpu]
413
    "VLLM_TARGET_DEVICE":
414
    lambda: os.getenv("VLLM_TARGET_DEVICE", "cuda").lower(),
415

416
417
418
419
420
    # Main CUDA version of vLLM, supporting [12.6, 12.8, 12.9],
    # 12.8 is the default. This follows PyTorch but can be overridden.
    "VLLM_MAIN_CUDA_VERSION":
    lambda: os.getenv("VLLM_MAIN_CUDA_VERSION", "").lower() or "12.8",

421
422
423
424
425
426
427
428
429
430
431
432
433
    # Maximum number of compilation jobs to run in parallel.
    # By default this is the number of CPUs
    "MAX_JOBS":
    lambda: os.getenv("MAX_JOBS", None),

    # Number of threads to use for nvcc
    # By default this is 1.
    # If set, `MAX_JOBS` will be reduced to avoid oversubscribing the CPU.
    "NVCC_THREADS":
    lambda: os.getenv("NVCC_THREADS", None),

    # If set, vllm will use precompiled binaries (*.so)
    "VLLM_USE_PRECOMPILED":
434
435
436
437
438
439
440
441
    lambda: os.environ.get("VLLM_USE_PRECOMPILED", "").strip().lower() in
    ("1", "true") or bool(os.environ.get("VLLM_PRECOMPILED_WHEEL_LOCATION")),

    # Used to mark that setup.py is running in a Docker build context,
    # in order to force the use of precompiled binaries.
    "VLLM_DOCKER_BUILD_CONTEXT":
    lambda: os.environ.get("VLLM_DOCKER_BUILD_CONTEXT", "").strip().lower() in
    ("1", "true"),
442

443
444
445
446
447
448
    # Whether to force using nightly wheel in python build.
    # This is used for testing the nightly wheel in python build.
    "VLLM_TEST_USE_PRECOMPILED_NIGHTLY_WHEEL":
    lambda: bool(int(os.getenv("VLLM_TEST_USE_PRECOMPILED_NIGHTLY_WHEEL", "0"))
                 ),

449
450
451
452
    # CMake build type
    # If not set, defaults to "Debug" or "RelWithDebInfo"
    # Available options: "Debug", "Release", "RelWithDebInfo"
    "CMAKE_BUILD_TYPE":
453
454
    env_with_choices("CMAKE_BUILD_TYPE", None,
        ["Debug", "Release", "RelWithDebInfo"]),
455
456
457
458
459

    # If set, vllm will print verbose logs during installation
    "VERBOSE":
    lambda: bool(int(os.getenv('VERBOSE', '0'))),

460
    # Root directory for vLLM configuration files
461
    # Defaults to `~/.config/vllm` unless `XDG_CONFIG_HOME` is set
462
463
464
465
    # 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**.
    "VLLM_CONFIG_ROOT":
466
467
468
469
470
    lambda: os.path.expanduser(
        os.getenv(
            "VLLM_CONFIG_ROOT",
            os.path.join(get_default_config_root(), "vllm"),
        )),
471
472
473

    # ================== Runtime Env Vars ==================

474
    # Root directory for vLLM cache files
475
476
477
478
479
480
481
482
    # Defaults to `~/.cache/vllm` unless `XDG_CACHE_HOME` is set
    "VLLM_CACHE_ROOT":
    lambda: os.path.expanduser(
        os.getenv(
            "VLLM_CACHE_ROOT",
            os.path.join(get_default_cache_root(), "vllm"),
        )),

483
484
485
486
    # 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.
487
    'VLLM_HOST_IP':
488
    lambda: os.getenv('VLLM_HOST_IP', ""),
489

490
    # used in distributed environment to manually set the communication port
491
492
493
    # 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.
494
    'VLLM_PORT':
495
    get_vllm_port,
496

497
498
499
500
    # path used for ipc when the frontend api server is running in
    # multi-processing mode to communicate with the backend engine process.
    'VLLM_RPC_BASE_PATH':
    lambda: os.getenv('VLLM_RPC_BASE_PATH', tempfile.gettempdir()),
501

502
503
504
505
506
    # If true, will load models from ModelScope instead of Hugging Face Hub.
    # note that the value is true or false, not numbers
    "VLLM_USE_MODELSCOPE":
    lambda: os.environ.get("VLLM_USE_MODELSCOPE", "False").lower() == "true",

507
508
509
510
    # Interval in seconds to log a warning message when the ring buffer is full
    "VLLM_RINGBUFFER_WARNING_INTERVAL":
    lambda: int(os.environ.get("VLLM_RINGBUFFER_WARNING_INTERVAL", "60")),

511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
    # path to cudatoolkit home directory, under which should be bin, include,
    # and lib directories.
    "CUDA_HOME":
    lambda: os.environ.get("CUDA_HOME", None),

    # 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
    "VLLM_NCCL_SO_PATH":
    lambda: os.environ.get("VLLM_NCCL_SO_PATH", None),

    # 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`
    "LD_LIBRARY_PATH":
    lambda: os.environ.get("LD_LIBRARY_PATH", None),

    # flag to control if vllm should use triton flash attention
    "VLLM_USE_TRITON_FLASH_ATTN":
528
    lambda: (os.environ.get("VLLM_USE_TRITON_FLASH_ATTN", "False").lower() in
529
530
             ("true", "1")),

531
532
533
534
535
536
537
    # Use separate prefill and decode kernels for V1 attention instead of
    # the unified triton kernel.
    "VLLM_V1_USE_PREFILL_DECODE_ATTENTION":
    lambda:
    (os.getenv("VLLM_V1_USE_PREFILL_DECODE_ATTENTION", "False").lower() in
     ("true", "1")),

538
539
540
541
542
543
    # Use AITER triton unified attention for V1 attention
    "VLLM_USE_AITER_UNIFIED_ATTENTION":
    lambda:
    (os.getenv("VLLM_USE_AITER_UNIFIED_ATTENTION", "False").lower() in
     ("true", "1")),

544
545
546
547
548
    # Force vllm to use a specific flash-attention version (2 or 3), only valid
    # when using the flash-attention backend.
    "VLLM_FLASH_ATTN_VERSION":
    lambda: maybe_convert_int(os.environ.get("VLLM_FLASH_ATTN_VERSION", None)),

549
550
    # Feature flag to enable/disable Inductor standalone compile.
    # In torch <= 2.7 we ignore this flag; in torch >= 2.8 this is
551
    # disabled by default.
552
    "VLLM_USE_STANDALONE_COMPILE":
553
    lambda: os.environ.get("VLLM_USE_STANDALONE_COMPILE", "0") == "1",
554

555
556
557
558
    # Debug pattern matching inside custom passes.
    # Should be set to the fx.Node name (e.g. 'getitem_34' or 'scaled_mm_3').
    "VLLM_PATTERN_MATCH_DEBUG":
    lambda: os.environ.get("VLLM_PATTERN_MATCH_DEBUG", None),
559

560
561
562
563
564
565
566
567
568
569
570
    # local rank of the process in the distributed setting, used to determine
    # the GPU device id
    "LOCAL_RANK":
    lambda: int(os.environ.get("LOCAL_RANK", "0")),

    # used to control the visible devices in the distributed setting
    "CUDA_VISIBLE_DEVICES":
    lambda: os.environ.get("CUDA_VISIBLE_DEVICES", None),

    # timeout for each iteration in the engine
    "VLLM_ENGINE_ITERATION_TIMEOUT_S":
571
    lambda: int(os.environ.get("VLLM_ENGINE_ITERATION_TIMEOUT_S", "120")),
572

573
    # API key for vLLM API server
574
575
576
    "VLLM_API_KEY":
    lambda: os.environ.get("VLLM_API_KEY", None),

577
578
    # Whether to log responses from API Server for debugging
    "VLLM_DEBUG_LOG_API_SERVER_RESPONSE":
579
580
    lambda: os.environ.get("VLLM_DEBUG_LOG_API_SERVER_RESPONSE", "False"
                           ).lower() == "true",
581

582
583
    # S3 access information, used for tensorizer to load model from S3
    "S3_ACCESS_KEY_ID":
584
    lambda: os.environ.get("S3_ACCESS_KEY_ID", None),
585
586
587
588
589
590
591
592
593
594
    "S3_SECRET_ACCESS_KEY":
    lambda: os.environ.get("S3_SECRET_ACCESS_KEY", None),
    "S3_ENDPOINT_URL":
    lambda: os.environ.get("S3_ENDPOINT_URL", None),

    # Usage stats collection
    "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",
595
596
    "VLLM_DISABLE_FLASHINFER_PREFILL":
    lambda: os.environ.get("VLLM_DISABLE_FLASHINFER_PREFILL", "0") == "1",
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
    "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"),

    # 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
    "VLLM_CONFIGURE_LOGGING":
    lambda: int(os.getenv("VLLM_CONFIGURE_LOGGING", "1")),
    "VLLM_LOGGING_CONFIG_PATH":
    lambda: os.getenv("VLLM_LOGGING_CONFIG_PATH"),

612
613
    # this is used for configuring the default logging level
    "VLLM_LOGGING_LEVEL":
614
    lambda: os.getenv("VLLM_LOGGING_LEVEL", "INFO").upper(),
615

616
617
618
619
    # this is used for configuring the default logging stream
    "VLLM_LOGGING_STREAM":
    lambda: os.getenv("VLLM_LOGGING_STREAM", "ext://sys.stdout"),

620
621
622
623
    # if set, VLLM_LOGGING_PREFIX will be prepended to all log messages
    "VLLM_LOGGING_PREFIX":
    lambda: os.getenv("VLLM_LOGGING_PREFIX", ""),

624
625
626
627
628
629
630
631
    # if set, vllm will call logits processors in a thread pool with this many
    # threads. This is useful when using custom logits processors that either
    # (a) launch additional CUDA kernels or (b) do significant CPU-bound work
    # while not holding the python GIL, or both.
    "VLLM_LOGITS_PROCESSOR_THREADS":
    lambda: int(os.getenv("VLLM_LOGITS_PROCESSOR_THREADS", "0"))
    if "VLLM_LOGITS_PROCESSOR_THREADS" in os.environ else None,

632
633
634
635
636
637
    # If set, vllm will log stats at this interval in seconds
    # If not set, vllm will log stats every 10 seconds.
    "VLLM_LOG_STATS_INTERVAL":
    lambda: val if (val := float(os.getenv("VLLM_LOG_STATS_INTERVAL", "10.")))
        > 0. else 10.,

638
639
640
641
642
643
644
    # Trace function calls
    # If set to 1, vllm will trace function calls
    # Useful for debugging
    "VLLM_TRACE_FUNCTION":
    lambda: int(os.getenv("VLLM_TRACE_FUNCTION", "0")),

    # Backend for attention computation
645
    # Example options:
646
647
648
    # - "TORCH_SDPA": use torch.nn.MultiheadAttention
    # - "FLASH_ATTN": use FlashAttention
    # - "XFORMERS": use XFormers
649
    # - "FLASHINFER": use flashinfer
650
    # - "FLASHMLA": use FlashMLA
651
    # - "FLASH_ATTN_MLA": use FlashAttention for MLA
652
653
    # - "FLASHINFER_MLA": use FlashInfer for MLA
    # - "CUTLASS_MLA": use CUTLASS for MLA
654
    # All possible options loaded dynamically from _Backend enum
655
    "VLLM_ATTENTION_BACKEND":
656
657
658
    env_with_choices("VLLM_ATTENTION_BACKEND", None,
                     lambda: list(__import__('vllm.platforms.interface', \
                        fromlist=['_Backend'])._Backend.__members__.keys())),
659

660
661
    # If set, vllm will use flashinfer sampler
    "VLLM_USE_FLASHINFER_SAMPLER":
662
663
    lambda: bool(int(os.environ["VLLM_USE_FLASHINFER_SAMPLER"]))
    if "VLLM_USE_FLASHINFER_SAMPLER" in os.environ else None,
664

665
666
667
668
    # Pipeline stage partition strategy
    "VLLM_PP_LAYER_PARTITION":
    lambda: os.getenv("VLLM_PP_LAYER_PARTITION", None),

669
    # (CPU backend only) CPU key-value cache space.
670
    # default is None and will be set as 4 GB
671
    "VLLM_CPU_KVCACHE_SPACE":
672
673
    lambda: int(os.getenv("VLLM_CPU_KVCACHE_SPACE", "0"))
    if "VLLM_CPU_KVCACHE_SPACE" in os.environ else None,
674

675
676
677
    # (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 '|'.
    "VLLM_CPU_OMP_THREADS_BIND":
678
679
680
681
682
    lambda: os.getenv("VLLM_CPU_OMP_THREADS_BIND", "auto"),

    # (CPU backend only) CPU cores not used by OMP threads .
    # Those CPU cores will not be used by OMP threads of a rank.
    "VLLM_CPU_NUM_OF_RESERVED_CPU":
683
684
    lambda: int(os.getenv("VLLM_CPU_NUM_OF_RESERVED_CPU", "0"))
    if "VLLM_CPU_NUM_OF_RESERVED_CPU" in os.environ else None,
685

686
687
688
689
690
    # (CPU backend only) whether to use prepack for MoE layer. This will be
    # passed to ipex.llm.modules.GatedMLPMOE. On unsupported CPUs, you might
    # need to set this to "0" (False).
    "VLLM_CPU_MOE_PREPACK":
    lambda: bool(int(os.getenv("VLLM_CPU_MOE_PREPACK", "1"))),
691

692
693
694
695
    # (CPU backend only) whether to use SGL kernels, optimized for small batch.
    "VLLM_CPU_SGL_KERNEL":
    lambda: bool(int(os.getenv("VLLM_CPU_SGL_KERNEL", "0"))),

696
697
698
699
700
    # If the env var is set, then all workers will execute as separate
    # processes from the engine, and we use the same mechanism to trigger
    # execution on all workers.
    # Run vLLM with VLLM_USE_RAY_SPMD_WORKER=1 to enable it.
    "VLLM_USE_RAY_SPMD_WORKER":
701
    lambda: bool(int(os.getenv("VLLM_USE_RAY_SPMD_WORKER", "0"))),
702

703
704
705
    # If the env var is set, it uses the Ray's Compiled Graph
    # (previously known as ADAG) API which optimizes the
    # control plane overhead.
706
    # Run vLLM with VLLM_USE_RAY_COMPILED_DAG=1 to enable it.
707
708
    # Note that this variable is set to 1 in V1 by default
    # when ray distributed executor is used.
709
    "VLLM_USE_RAY_COMPILED_DAG":
710
711
    lambda: bool(int(os.getenv("VLLM_USE_RAY_COMPILED_DAG", "0"))),

712
713
714
715
716
717
718
719
720
    # 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
    # This flag is ignored if VLLM_USE_RAY_COMPILED_DAG is not set.
    "VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE":
721
722
    env_with_choices("VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE", "auto",
        ["auto", "nccl", "shm"]),
723

724
    # If the env var is set, it enables GPU communication overlap
725
    # (experimental feature) in Ray's Compiled Graph. This flag is ignored if
726
727
    # VLLM_USE_RAY_COMPILED_DAG is not set.
    "VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM":
728
    lambda: bool(int(os.getenv("VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM", "0"))
729
730
                 ),

731
732
733
734
735
736
737
    # 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.
    # This flag is ignored if VLLM_USE_RAY_COMPILED_DAG is not set.
    "VLLM_USE_RAY_WRAPPED_PP_COMM":
    lambda: bool(int(os.getenv("VLLM_USE_RAY_WRAPPED_PP_COMM", "1"))),

738
739
740
    # Use dedicated multiprocess context for workers.
    # Both spawn and fork work
    "VLLM_WORKER_MULTIPROC_METHOD":
741
    env_with_choices("VLLM_WORKER_MULTIPROC_METHOD", "spawn",
742
       ["spawn", "fork"]),
743

744
745
746
747
748
749
750
751
    # Path to the cache for storing downloaded assets
    "VLLM_ASSETS_CACHE":
    lambda: os.path.expanduser(
        os.getenv(
            "VLLM_ASSETS_CACHE",
            os.path.join(get_default_cache_root(), "vllm", "assets"),
        )),

752
753
754
755
756
    # If the env var is set, we will clean model file in
    # this path $VLLM_ASSETS_CACHE/model_streamer/$model_name
    "VLLM_ASSETS_CACHE_MODEL_CLEAN":
    lambda: bool(int(os.getenv("VLLM_ASSETS_CACHE_MODEL_CLEAN", "0"))),

757
758
759
760
    # Timeout for fetching images when serving multimodal models
    # Default is 5 seconds
    "VLLM_IMAGE_FETCH_TIMEOUT":
    lambda: int(os.getenv("VLLM_IMAGE_FETCH_TIMEOUT", "5")),
761

762
    # Timeout for fetching videos when serving multimodal models
763
    # Default is 30 seconds
764
    "VLLM_VIDEO_FETCH_TIMEOUT":
765
    lambda: int(os.getenv("VLLM_VIDEO_FETCH_TIMEOUT", "30")),
766

767
    # Timeout for fetching audio when serving multimodal models
768
    # Default is 10 seconds
769
    "VLLM_AUDIO_FETCH_TIMEOUT":
770
    lambda: int(os.getenv("VLLM_AUDIO_FETCH_TIMEOUT", "10")),
771

772
773
774
775
776
    # Whether to allow HTTP redirects when fetching from media URLs.
    # Default to True
    "VLLM_MEDIA_URL_ALLOW_REDIRECTS":
    lambda: bool(int(os.getenv("VLLM_MEDIA_URL_ALLOW_REDIRECTS", "1"))),

777
778
779
780
781
782
    # Max number of workers for the thread pool handling
    # media bytes loading. Set to 1 to disable parallel processing.
    # Default is 8
    "VLLM_MEDIA_LOADING_THREAD_COUNT":
    lambda: int(os.getenv("VLLM_MEDIA_LOADING_THREAD_COUNT", "8")),

783
784
785
786
787
788
    # 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
    "VLLM_MAX_AUDIO_CLIP_FILESIZE_MB":
    lambda: int(os.getenv("VLLM_MAX_AUDIO_CLIP_FILESIZE_MB", "25")),

789
790
791
792
793
794
795
796
797
798
    # Backend for Video IO
    # - "opencv": Default backend that uses OpenCV stream buffered backend.
    #
    # Custom backend implementations can be registered
    # via `@VIDEO_LOADER_REGISTRY.register("my_custom_video_loader")` and
    # imported at runtime.
    # If a non-existing backend is used, an AssertionError will be thrown.
    "VLLM_VIDEO_LOADER_BACKEND":
    lambda: os.getenv("VLLM_VIDEO_LOADER_BACKEND", "opencv"),

799
    # [DEPRECATED] Cache size (in GiB per process) for multimodal input cache
800
    # Default is 4 GiB per API process + 4 GiB per engine core process
801
    "VLLM_MM_INPUT_CACHE_GIB":
802
    lambda: int(os.getenv("VLLM_MM_INPUT_CACHE_GIB", "4")),
803

804
805
806
    # Path to the XLA persistent cache directory.
    # Only used for XLA devices such as TPUs.
    "VLLM_XLA_CACHE_PATH":
807
808
    lambda: os.path.expanduser(
        os.getenv(
809
            "VLLM_XLA_CACHE_PATH",
810
811
            os.path.join(get_default_cache_root(), "vllm", "xla_cache"),
        )),
812
813
814
815

    # If set, assert on XLA recompilation after each execution step.
    "VLLM_XLA_CHECK_RECOMPILATION":
    lambda: bool(int(os.getenv("VLLM_XLA_CHECK_RECOMPILATION", "0"))),
816
817
818
819

    # Enable SPMD mode for TPU backend.
    "VLLM_XLA_USE_SPMD":
    lambda: bool(int(os.getenv("VLLM_XLA_USE_SPMD", "0"))),
820
    "VLLM_FUSED_MOE_CHUNK_SIZE":
821
    lambda: int(os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", "32768")),
822
823
824
825
826
827
    # 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.
    "VLLM_ENABLE_FUSED_MOE_ACTIVATION_CHUNKING":
    lambda: bool(
        int(os.getenv("VLLM_ENABLE_FUSED_MOE_ACTIVATION_CHUNKING", "1"))),
828

829
830
831
832
833
    # If set, the OpenAI API server will stay alive even after the underlying
    # AsyncLLMEngine errors and stops serving requests
    "VLLM_KEEP_ALIVE_ON_ENGINE_DEATH":
    lambda: bool(os.getenv("VLLM_KEEP_ALIVE_ON_ENGINE_DEATH", 0)),

834
835
836
837
838
839
840
841
    # 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.
    "VLLM_ALLOW_LONG_MAX_MODEL_LEN":
    lambda:
    (os.environ.get("VLLM_ALLOW_LONG_MAX_MODEL_LEN", "0").strip().lower() in
     ("1", "true")),
842
843
844
845
846
847
848

    # If set, forces FP8 Marlin to be used for FP8 quantization regardless
    # of the hardware support for FP8 compute.
    "VLLM_TEST_FORCE_FP8_MARLIN":
    lambda:
    (os.environ.get("VLLM_TEST_FORCE_FP8_MARLIN", "0").strip().lower() in
     ("1", "true")),
849
850
    "VLLM_TEST_FORCE_LOAD_FORMAT":
    lambda: os.getenv("VLLM_TEST_FORCE_LOAD_FORMAT", "dummy"),
851

852
853
    # Time in ms for the zmq client to wait for a response from the backend
    # server for simple data operations
854
855
    "VLLM_RPC_TIMEOUT":
    lambda: int(os.getenv("VLLM_RPC_TIMEOUT", "10000")),
856

857
858
859
860
    # Timeout in seconds for keeping HTTP connections alive in API server
    "VLLM_HTTP_TIMEOUT_KEEP_ALIVE":
    lambda: int(os.environ.get("VLLM_HTTP_TIMEOUT_KEEP_ALIVE", "5")),

861
862
863
864
865
866
    # 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
    "VLLM_PLUGINS":
    lambda: None if "VLLM_PLUGINS" not in os.environ else os.environ[
        "VLLM_PLUGINS"].split(","),
867

868
869
870
871
872
873
    # a local directory to look in for unrecognized LoRA adapters.
    # only works if plugins are enabled and
    # VLLM_ALLOW_RUNTIME_LORA_UPDATING is enabled.
    "VLLM_LORA_RESOLVER_CACHE_DIR":
    lambda: os.getenv("VLLM_LORA_RESOLVER_CACHE_DIR", None),

874
875
876
877
    # Enables torch profiler if set.
    # Both AsyncLLM's CPU traces as well as workers'
    # traces (CPU & GPU) will be saved under this directory.
    # Note that it must be an absolute path.
878
879
    "VLLM_TORCH_PROFILER_DIR":
    lambda: (None if os.getenv("VLLM_TORCH_PROFILER_DIR", None) is None else os
880
881
             .path.abspath(os.path.expanduser(os.getenv(
        "VLLM_TORCH_PROFILER_DIR", ".")))),
882

883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
    # Enable torch profiler to record shapes if set
    # VLLM_TORCH_PROFILER_RECORD_SHAPES=1. If not set, torch profiler will
    # not record shapes.
    "VLLM_TORCH_PROFILER_RECORD_SHAPES":
    lambda: bool(os.getenv("VLLM_TORCH_PROFILER_RECORD_SHAPES", "0") != "0"),

    # Enable torch profiler to profile memory if set
    # VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY=1. If not set, torch profiler
    # will not profile memory.
    "VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY":
    lambda: bool(
        os.getenv("VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY", "0") != "0"),

    # Enable torch profiler to profile stack if set
    # VLLM_TORCH_PROFILER_WITH_STACK=1. If not set, torch profiler WILL
    # profile stack by default.
    "VLLM_TORCH_PROFILER_WITH_STACK":
    lambda: bool(os.getenv("VLLM_TORCH_PROFILER_WITH_STACK", "1") != "0"),

    # Enable torch profiler to profile flops if set
    # VLLM_TORCH_PROFILER_WITH_FLOPS=1. If not set, torch profiler will
    # not profile flops.
    "VLLM_TORCH_PROFILER_WITH_FLOPS":
    lambda: bool(os.getenv("VLLM_TORCH_PROFILER_WITH_FLOPS", "0") != "0"),

908
909
910
    # If set, vLLM will use Triton implementations of AWQ.
    "VLLM_USE_TRITON_AWQ":
    lambda: bool(int(os.getenv("VLLM_USE_TRITON_AWQ", "0"))),
911
912
913
914
915
916

    # If set, allow loading or unloading lora adapters in runtime,
    "VLLM_ALLOW_RUNTIME_LORA_UPDATING":
    lambda:
    (os.environ.get("VLLM_ALLOW_RUNTIME_LORA_UPDATING", "0").strip().lower() in
     ("1", "true")),
917

918
919
920
921
922
923
    # 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
924
    "VLLM_SKIP_P2P_CHECK":
925
    lambda: os.getenv("VLLM_SKIP_P2P_CHECK", "1") == "1",
926

927
928
929
930
931
932
933
    # List of quantization kernels that should be disabled, used for testing
    # and performance comparisons. Currently only affects MPLinearKernel
    # selection
    # (kernels: MacheteLinearKernel, MarlinLinearKernel, ExllamaLinearKernel)
    "VLLM_DISABLED_KERNELS":
    lambda: [] if "VLLM_DISABLED_KERNELS" not in os.environ else os.environ[
        "VLLM_DISABLED_KERNELS"].split(","),
934

935
936
937
938
939
940
941
    # Swaps the all reduce backend that we use to coordinate the DP padding
    # information from NCCL to gloo.
    "VLLM_DISABLE_NCCL_FOR_DP_SYNCHRONIZATION":
    lambda:
    (os.getenv("VLLM_DISABLE_NCCL_FOR_DP_SYNCHRONIZATION", "False").lower() in
             ("true", "1")),

942
943
    # If set, use the V1 code path.
    "VLLM_USE_V1":
zhuwenwen's avatar
zhuwenwen committed
944
    lambda: bool(int(os.getenv("VLLM_USE_V1", "1"))),
945

946
947
948
949
950
    # Disable aiter ops unless specifically enabled.
    # Acts as a parent switch to enable the rest of the other operations.
    "VLLM_ROCM_USE_AITER":
    lambda: (os.getenv("VLLM_ROCM_USE_AITER", "False").lower() in
             ("true", "1")),
951

952
953
954
955
956
957
    # Whether to use aiter paged attention.
    # By default is disabled.
    "VLLM_ROCM_USE_AITER_PAGED_ATTN":
    lambda: (os.getenv("VLLM_ROCM_USE_AITER_PAGED_ATTN", "False").lower() in
             ("true", "1")),

958
959
960
961
962
963
964
    # use aiter linear op if aiter ops are enabled
    # The following list of related ops
    # - scaled_mm (per-tensor / rowwise)
    "VLLM_ROCM_USE_AITER_LINEAR":
    lambda: (os.getenv("VLLM_ROCM_USE_AITER_LINEAR", "True").lower() in
             ("true", "1")),

965
966
967
968
969
970
    # Whether to use aiter moe ops.
    # By default is enabled.
    "VLLM_ROCM_USE_AITER_MOE":
    lambda: (os.getenv("VLLM_ROCM_USE_AITER_MOE", "True").lower() in
             ("true", "1")),

971
972
973
974
975
    # use aiter rms norm op if aiter ops are enabled.
    "VLLM_ROCM_USE_AITER_RMSNORM":
    lambda: (os.getenv("VLLM_ROCM_USE_AITER_RMSNORM", "True").lower() in
             ("true", "1")),

976
977
978
979
980
    # Whether to use aiter mla ops.
    # By default is enabled.
    "VLLM_ROCM_USE_AITER_MLA":
    lambda: (os.getenv("VLLM_ROCM_USE_AITER_MLA", "True").lower() in
             ("true", "1")),
981
982
983
984
985
986
987

    # Whether to use aiter mha ops.
    # By default is enabled.
    "VLLM_ROCM_USE_AITER_MHA":
    lambda: (os.getenv("VLLM_ROCM_USE_AITER_MHA", "True").lower() in
             ("true", "1")),

988
989
990
991
992
993
994
995
996
997
998
999
    # Whether to use aiter fp4 gemm asm.
    # By default is disabled.
    "VLLM_ROCM_USE_AITER_FP4_ASM_GEMM":
    lambda: (os.getenv("VLLM_ROCM_USE_AITER_FP4_ASM_GEMM", "False").lower() in
             ("true", "1")),

    # Whether to use aiter rope.
    # By default is disabled.
    "VLLM_ROCM_USE_TRITON_ROPE":
    lambda: (os.getenv("VLLM_ROCM_USE_TRITON_ROPE", "False").lower() in
             ("true", "1")),

1000
1001
1002
1003
1004
1005
    # Whether to use aiter triton fp8 bmm kernel
    # By default is enabled.
    "VLLM_ROCM_USE_AITER_FP8BMM":
    lambda: (os.getenv("VLLM_ROCM_USE_AITER_FP8BMM", "True").lower() in
             ("true", "1")),

1006
1007
1008
1009
1010
    # use rocm skinny gemms
    "VLLM_ROCM_USE_SKINNY_GEMM":
    lambda: (os.getenv("VLLM_ROCM_USE_SKINNY_GEMM", "True").lower() in
             ("true", "1")),

1011
1012
1013
    # Pad the fp8 weights to 256 bytes for ROCm
    "VLLM_ROCM_FP8_PADDING":
    lambda: bool(int(os.getenv("VLLM_ROCM_FP8_PADDING", "1"))),
1014

1015
1016
    # Pad the weights for the moe kernel
    "VLLM_ROCM_MOE_PADDING":
zhuwenwen's avatar
zhuwenwen committed
1017
    lambda: bool(int(os.getenv("VLLM_ROCM_MOE_PADDING", "0"))),
1018

1019
1020
1021
1022
1023
    # custom paged attention kernel for MI3* cards
    "VLLM_ROCM_CUSTOM_PAGED_ATTN":
    lambda: (os.getenv("VLLM_ROCM_CUSTOM_PAGED_ATTN", "True").lower() in
             ("true", "1")),

1024
1025
1026
1027
    # Custom quick allreduce kernel for MI3* cards
    # Choice of quantization level: FP, INT8, INT6, INT4 or NONE
    # Recommended for large models to get allreduce
    "VLLM_ROCM_QUICK_REDUCE_QUANTIZATION":
1028
1029
    env_with_choices("VLLM_ROCM_QUICK_REDUCE_QUANTIZATION", "NONE",
                            ["FP", "INT8", "INT6", "INT4", "NONE"]),
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049

    # 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
    "VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16":
    lambda:
    (os.getenv("VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16", "True").lower() in
     ("true", "1")),

    # 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.
    "VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB":
    lambda: maybe_convert_int(
        os.environ.get("VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB", None)),

1050
1051
    # Divisor for dynamic query scale factor calculation for FP8 KV Cache
    "Q_SCALE_CONSTANT":
zhuwenwen's avatar
zhuwenwen committed
1052
    lambda: int(os.getenv("Q_SCALE_CONSTANT", "10")),
1053
1054
    # Divisor for dynamic key scale factor calculation for FP8 KV Cache
    "K_SCALE_CONSTANT":
zhuwenwen's avatar
zhuwenwen committed
1055
    lambda: int(os.getenv("K_SCALE_CONSTANT", "10")),
1056
1057
    # Divisor for dynamic value scale factor calculation for FP8 KV Cache
    "V_SCALE_CONSTANT":
zhuwenwen's avatar
zhuwenwen committed
1058
    lambda: int(os.getenv("V_SCALE_CONSTANT", "10")),
1059

1060
1061
    # If set, enable multiprocessing in LLM for the V1 code path.
    "VLLM_ENABLE_V1_MULTIPROCESSING":
1062
    lambda: bool(int(os.getenv("VLLM_ENABLE_V1_MULTIPROCESSING", "1"))),
1063
1064
    "VLLM_LOG_BATCHSIZE_INTERVAL":
    lambda: float(os.getenv("VLLM_LOG_BATCHSIZE_INTERVAL", "-1")),
1065
1066
    "VLLM_DISABLE_COMPILE_CACHE":
    lambda: bool(int(os.getenv("VLLM_DISABLE_COMPILE_CACHE", "0"))),
1067
1068
1069
1070
1071
1072

    # If set, vllm will run in development mode, which will enable
    # some additional endpoints for developing and debugging,
    # e.g. `/reset_prefix_cache`
    "VLLM_SERVER_DEV_MODE":
    lambda: bool(int(os.getenv("VLLM_SERVER_DEV_MODE", "0"))),
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082

    # 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.
    "VLLM_V1_OUTPUT_PROC_CHUNK_SIZE":
    lambda: int(os.getenv("VLLM_V1_OUTPUT_PROC_CHUNK_SIZE", "128")),
1083
1084
1085
1086
1087

    # If set, vLLM will disable the MLA attention optimizations.
    "VLLM_MLA_DISABLE":
    lambda: bool(int(os.getenv("VLLM_MLA_DISABLE", "0"))),

1088
1089
1090
1091
    # If set, vLLM will pick up the provided Flash Attention MLA
    # max number splits for cuda graph decode
    "VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH":
    lambda: int(os.getenv("VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH",
1092
                          "32")),
1093

1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
    # 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.
    "VLLM_RAY_PER_WORKER_GPUS":
    lambda: float(os.getenv("VLLM_RAY_PER_WORKER_GPUS", "1.0")),

    # 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"
    "VLLM_RAY_BUNDLE_INDICES":
    lambda: os.getenv("VLLM_RAY_BUNDLE_INDICES", ""),

1106
1107
1108
1109
    # In some system, find_loaded_library() may not work. So we allow users to
    # specify the path through environment variable VLLM_CUDART_SO_PATH.
    "VLLM_CUDART_SO_PATH":
    lambda: os.getenv("VLLM_CUDART_SO_PATH", None),
1110

1111
1112
1113
1114
    # Rank of the process in the data parallel setting
    "VLLM_DP_RANK":
    lambda: int(os.getenv("VLLM_DP_RANK", "0")),

1115
1116
1117
1118
1119
1120
    # Rank of the process in the data parallel setting.
    # Defaults to VLLM_DP_RANK when not set.
    "VLLM_DP_RANK_LOCAL":
    lambda: int(
        os.getenv("VLLM_DP_RANK_LOCAL", sys.modules[__name__].VLLM_DP_RANK)),

1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
    # World size of the data parallel setting
    "VLLM_DP_SIZE":
    lambda: int(os.getenv("VLLM_DP_SIZE", "1")),

    # IP address of the master node in the data parallel setting
    "VLLM_DP_MASTER_IP":
    lambda: os.getenv("VLLM_DP_MASTER_IP", "127.0.0.1"),

    # Port of the master node in the data parallel setting
    "VLLM_DP_MASTER_PORT":
    lambda: int(os.getenv("VLLM_DP_MASTER_PORT", "0")),
1132

1133
1134
1135
1136
1137
1138
1139
1140
    # 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.
    "VLLM_MOE_DP_CHUNK_SIZE":
    lambda: int(os.getenv("VLLM_MOE_DP_CHUNK_SIZE", "256")),

1141
1142
1143
1144
    # Randomize inputs during dummy runs when using Data Parallel
    "VLLM_RANDOMIZE_DP_DUMMY_INPUTS":
    lambda: os.environ.get("VLLM_RANDOMIZE_DP_DUMMY_INPUTS", "0") == "1",

1145
1146
1147
    # Whether to use S3 path for model loading in CI via RunAI Streamer
    "VLLM_CI_USE_S3":
    lambda: os.environ.get("VLLM_CI_USE_S3", "0") == "1",
1148

1149
    # Use model_redirect to redirect the model name to a local folder.
1150
1151
1152
1153
1154
    # `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
1155
1156
1157
    "VLLM_MODEL_REDIRECT_PATH":
    lambda: os.environ.get("VLLM_MODEL_REDIRECT_PATH", None),

1158
1159
1160
    # Whether to use atomicAdd reduce in gptq/awq marlin kernel.
    "VLLM_MARLIN_USE_ATOMIC_ADD":
    lambda: os.environ.get("VLLM_MARLIN_USE_ATOMIC_ADD", "0") == "1",
1161

1162
1163
1164
1165
    # Whether to use marlin kernel in mxfp4 quantization method
    "VLLM_MXFP4_USE_MARLIN":
    lambda: maybe_convert_bool(os.environ.get("VLLM_MXFP4_USE_MARLIN", None)),

1166
1167
1168
1169
1170
    # Whether to turn on the outlines cache for V0
    # This cache is unbounded and on disk, so it's not safe to use in
    # an environment with potentially malicious users.
    "VLLM_V0_USE_OUTLINES_CACHE":
    lambda: os.environ.get("VLLM_V0_USE_OUTLINES_CACHE", "0") == "1",
1171

1172
1173
1174
1175
1176
1177
    # 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.
    "VLLM_V1_USE_OUTLINES_CACHE":
    lambda: os.environ.get("VLLM_V1_USE_OUTLINES_CACHE", "0") == "1",

1178
1179
1180
1181
    # Gap between padding buckets for the forward pass. So we have
    # 8, we will run forward pass with [16, 24, 32, ...].
    "VLLM_TPU_BUCKET_PADDING_GAP":
    lambda: int(os.environ["VLLM_TPU_BUCKET_PADDING_GAP"])
1182
    if "VLLM_TPU_BUCKET_PADDING_GAP" in os.environ else 0,
1183
1184
    "VLLM_TPU_MOST_MODEL_LEN":
    lambda: maybe_convert_int(os.environ.get("VLLM_TPU_MOST_MODEL_LEN", None)),
1185

1186
1187
1188
1189
    # Whether using Pathways
    "VLLM_TPU_USING_PATHWAYS":
    lambda: bool("proxy" in os.getenv("JAX_PLATFORMS", "").lower()),

1190
1191
    # Allow use of DeepGemm kernels for fused moe ops.
    "VLLM_USE_DEEP_GEMM":
1192
    lambda: bool(int(os.getenv("VLLM_USE_DEEP_GEMM", "1"))),
1193

1194
1195
1196
    # Whether to use E8M0 scaling when DeepGEMM is used on Blackwell GPUs.
    "VLLM_USE_DEEP_GEMM_E8M0":
    lambda: bool(int(os.getenv("VLLM_USE_DEEP_GEMM_E8M0", "1"))),
1197
1198
1199
1200
    # TODO(wentao): unify the two E8M0 flags after verifying the correctness.
    # Whether to use E8M0 scaling when DeepGEMM is used on Hopper GPUs.
    "VLLM_USE_DEEP_GEMM_E8M0_HOPPER":
    lambda: bool(int(os.getenv("VLLM_USE_DEEP_GEMM_E8M0_HOPPER", "0"))),
1201
1202
1203
1204
1205
1206
1207
1208
    # 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.
    # Set `VLLM_SKIP_DEEP_GEMM_WARMUP` to disable the warmup.
    "VLLM_SKIP_DEEP_GEMM_WARMUP":
    lambda: bool(int(os.getenv("VLLM_SKIP_DEEP_GEMM_WARMUP", "0"))),

1209
1210
1211
1212
    # Whether to use fused grouped_topk used for MoE expert selection.
    "VLLM_USE_FUSED_MOE_GROUPED_TOPK":
    lambda: bool(int(os.getenv("VLLM_USE_FUSED_MOE_GROUPED_TOPK", "1"))),

1213
1214
1215
1216
    # Allow use of FlashInfer MoE kernels for fused moe ops.
    "VLLM_USE_FLASHINFER_MOE_FP16":
    lambda: bool(int(os.getenv("VLLM_USE_FLASHINFER_MOE_FP16", "0"))),

1217
1218
1219
1220
    # Allow use of FlashInfer MoE kernels for fused moe ops.
    "VLLM_USE_FLASHINFER_MOE_FP8":
    lambda: bool(int(os.getenv("VLLM_USE_FLASHINFER_MOE_FP8", "0"))),

1221
    # Allow use of FlashInfer CUTLASS kernels for fused moe ops.
1222
1223
    "VLLM_USE_FLASHINFER_MOE_FP4":
    lambda: bool(int(os.getenv("VLLM_USE_FLASHINFER_MOE_FP4", "0"))),
1224

1225
1226
1227
1228
1229
    # If set to 1, use the FlashInfer
    # MXFP8 (activation) x MXFP4 (weight) MoE backend.
    "VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8":
    lambda: bool(int(os.getenv("VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8", "0"))),

1230
1231
1232
1233
1234
1235
1236
1237
1238
    # 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.
    "VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS":
    lambda: bool(int(
        os.getenv("VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS", "0")
        )),

1239
1240
1241
1242
1243
    # If set to 1, use the FlashInfer
    # BF16 (activation) x MXFP4 (weight) MoE backend.
    "VLLM_USE_FLASHINFER_MOE_MXFP4_BF16":
    lambda: bool(int(os.getenv("VLLM_USE_FLASHINFER_MOE_MXFP4_BF16", "0"))),

1244
1245
1246
1247
1248
    # 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.
    "VLLM_XGRAMMAR_CACHE_MB":
    lambda: int(os.getenv("VLLM_XGRAMMAR_CACHE_MB", "512")),
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258

    # Control the threshold for msgspec to use 'zero copy' for
    # serialization/deserialization of tensors. Tensors below
    # this limit will be encoded into the msgpack buffer, and
    # tensors above will instead be sent via a separate message.
    # While the sending side still actually copies the tensor
    # in all cases, on the receiving side, tensors above this
    # limit will actually be zero-copy decoded.
    "VLLM_MSGPACK_ZERO_COPY_THRESHOLD":
    lambda: int(os.getenv("VLLM_MSGPACK_ZERO_COPY_THRESHOLD", "256")),
1259
1260
1261
1262
1263
1264

    # If set, allow insecure serialization using pickle.
    # This is useful for environments where it is deemed safe to use the
    # insecure method and it is needed for some reason.
    "VLLM_ALLOW_INSECURE_SERIALIZATION":
    lambda: bool(int(os.getenv("VLLM_ALLOW_INSECURE_SERIALIZATION", "0"))),
Robert Shaw's avatar
Robert Shaw committed
1265
1266
1267
1268
1269
1270
1271
1272

    # IP address used for NIXL handshake between remote agents.
    "VLLM_NIXL_SIDE_CHANNEL_HOST":
    lambda: os.getenv("VLLM_NIXL_SIDE_CHANNEL_HOST", "localhost"),

    # Port used for NIXL handshake between remote agents.
    "VLLM_NIXL_SIDE_CHANNEL_PORT":
    lambda: int(os.getenv("VLLM_NIXL_SIDE_CHANNEL_PORT", "5557")),
1273
1274

    # all2all backend for vllm's expert parallel communication
1275
    # Available options:
1276
1277
1278
    # - "naive": naive all2all implementation using broadcasts
    # - "allgather_reducescatter": all2all implementation based on allgather and
    #  reducescatter
1279
    # - "pplx": use pplx kernels
1280
1281
    # - "deepep_high_throughput", use deepep high-throughput kernels
    # - "deepep_low_latency", use deepep low-latency kernels
1282
    # - "flashinfer_all2allv", use flashinfer alltoallv kernels for mnnvl
1283
    "VLLM_ALL2ALL_BACKEND":
1284
    env_with_choices("VLLM_ALL2ALL_BACKEND", "allgather_reducescatter",
1285
                     ["naive", "pplx",
1286
1287
                     "deepep_high_throughput",
                     "deepep_low_latency",
1288
1289
                     "allgather_reducescatter",
                     "flashinfer_all2allv"]),
1290

1291
1292
    # Flashinfer MoE backend for vLLM's fused Mixture-of-Experts support.
    # Both require compute capability 10.0 or above.
1293
1294
1295
1296
1297
    # Available options:
    # - "throughput":  [default]
    #     Uses CUTLASS kernels optimized for high-throughput batch inference.
    # - "latency":
    #     Uses TensorRT-LLM kernels optimized for low-latency inference.
1298
1299
1300
    "VLLM_FLASHINFER_MOE_BACKEND":
    env_with_choices("VLLM_FLASHINFER_MOE_BACKEND", "throughput",
    ["throughput", "latency"]),
1301

1302
1303
1304
1305
1306
1307
    # 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.
    "VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE":
    lambda: int(os.getenv("VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE", "163840")),
1308

1309
1310
1311
1312
    # 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> }
1313
    # Unspecified world sizes will fall back to
1314
1315
1316
1317
1318
    #     { 2: 64, 4: 1, <everything else>: 0.5 }
    "VLLM_FLASHINFER_ALLREDUCE_FUSION_THRESHOLDS_MB":
    lambda: json.loads(os.getenv(
        "VLLM_FLASHINFER_ALLREDUCE_FUSION_THRESHOLDS_MB", "{}")),

1319
1320
1321
1322
1323
1324
1325
1326
1327
    # MoE routing strategy selector.
    # See `RoutingSimulator.get_available_strategies()` # for available
    # strategies.
    # Cutstom routing strategies can be registered by
    # RoutingSimulator.register_strategy()
    # Note: custom strategies may not produce correct model outputs
    "VLLM_MOE_ROUTING_SIMULATION_STRATEGY":
    lambda: os.environ.get("VLLM_MOE_ROUTING_SIMULATION_STRATEGY", "").lower(),

1328
1329
1330
    # Regex timeout for use by the vLLM tool parsing plugins.
    "VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS":
    lambda: int(os.getenv("VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS", "1")),
1331
1332
1333
1334
1335

    # Reduce CPU usage when vLLM is idle. Enabling this will incur small
    # latency penalty when a request eventually comes.
    "VLLM_SLEEP_WHEN_IDLE":
    lambda: bool(int(os.getenv("VLLM_SLEEP_WHEN_IDLE", "0"))),
1336
1337
1338
1339
1340
1341

    # 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.
    "VLLM_MQ_MAX_CHUNK_BYTES_MB":
    lambda: int(os.getenv("VLLM_MQ_MAX_CHUNK_BYTES_MB", "16")),
1342
    
1343
1344
1345
1346
1347
    # Timeout in seconds for execute_model RPC calls in multiprocessing
    # executor (only applies when TP > 1).
    "VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS":
    lambda: int(os.getenv("VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS", "300")),

1348
1349
1350
1351
1352
1353
1354
1355
    # 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.
    "VLLM_KV_CACHE_LAYOUT":
1356
    env_with_choices("VLLM_KV_CACHE_LAYOUT", None, ["NHD", "HND"]),
1357
1358
1359
1360
1361
1362

    # 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.
    "VLLM_COMPUTE_NANS_IN_LOGITS":
    lambda: bool(int(os.getenv("VLLM_COMPUTE_NANS_IN_LOGITS", "0"))),
1363
1364
1365
1366
1367

    # Controls whether or not emulations are used for NVFP4
    # generations on machines < 100 for compressed-tensors
    # models
    "VLLM_USE_NVFP4_CT_EMULATIONS":
zhuwenwen's avatar
zhuwenwen committed
1368
1369
    lambda: bool(int(os.getenv("VLLM_USE_NVFP4_CT_EMULATIONS", "0"))),
    
1370
1371
1372
1373
1374
    # 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.
    "VLLM_NIXL_ABORT_REQUEST_TIMEOUT":
1375
1376
    lambda: int(os.getenv("VLLM_NIXL_ABORT_REQUEST_TIMEOUT", "120")),

1377
1378
1379
1380
    # Controls whether or not to use cudnn prefill
    "VLLM_USE_CUDNN_PREFILL":
    lambda: bool(int(os.getenv("VLLM_USE_CUDNN_PREFILL", "0"))),

1381
1382
1383
    # If set to 1/True, use the TRTLLM attention backend in flashinfer.
    # If set to 0/False, use the default attention backend in flashinfer.
    # If not set, auto-detect the attention backend in flashinfer.
1384
    "VLLM_USE_TRTLLM_ATTENTION":
1385
1386
    lambda: (None if "VLLM_USE_TRTLLM_ATTENTION" not in os.environ else
             os.environ["VLLM_USE_TRTLLM_ATTENTION"].lower() in ("1", "true")),
1387

1388
1389
1390
1391
    # If set to 1, when we use fp8 kv, we do not quantize Q to fp8
    "VLLM_FLASHINFER_DISABLE_Q_QUANTIZATION":
    lambda: bool(int(os.getenv("VLLM_FLASHINFER_DISABLE_Q_QUANTIZATION", "0"))),

1392
1393
1394
1395
1396
    # If set, it means we pre-downloaded cubin files and flashinfer will
    # read the cubin files directly.
    "VLLM_HAS_FLASHINFER_CUBIN":
    lambda: os.getenv("VLLM_HAS_FLASHINFER_CUBIN", False),

1397
1398
1399
1400
1401
    # If set to 1, force the use of TRTLLM FP4 GEMM backend in flashinfer.
    # Otherwise, uses the first available of: flashinfer cutlass GEMM,
    # vllm cutlass GEMM, marlin GEMM.
    "VLLM_USE_TRTLLM_FP4_GEMM":
    lambda: bool(int(os.getenv("VLLM_USE_TRTLLM_FP4_GEMM", "0"))),
1402

1403
1404
1405
1406
1407
1408
    # 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.
    "VLLM_ENABLE_CUDAGRAPH_GC":
    lambda: bool(int(os.getenv("VLLM_ENABLE_CUDAGRAPH_GC", "0"))),

1409
1410
1411
1412
1413
1414
    # Disable padding to CUDA graph capture batch sizes.
    # TODO(wentao): https://github.com/vllm-project/vllm/issues/23378
    # After the issue is fixed, we can remove this flag.
    "VLLM_DISABLE_PAD_FOR_CUDAGRAPH":
    lambda: bool(int(os.getenv("VLLM_DISABLE_PAD_FOR_CUDAGRAPH", "0"))),

1415
1416
1417
    # Used to force set up loopback IP
    "VLLM_LOOPBACK_IP":
    lambda: os.getenv("VLLM_LOOPBACK_IP", ""),
1418
    
1419
1420
1421
1422
1423
    # Used to set the process name prefix for vLLM processes.
    # This is useful for debugging and monitoring purposes.
    # The default value is "VLLM".
    "VLLM_PROCESS_NAME_PREFIX":
    lambda: os.getenv("VLLM_PROCESS_NAME_PREFIX", "VLLM"),
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434

    # 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.
    "VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE":
    lambda: bool(int(os.getenv(\
            "VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE", "0"))),
1435
1436
1437

    # 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
1438
1439
    # messages for those requests in memory. By default, this is disabled (0),
    # and the "store" option is ignored.
1440
1441
1442
1443
1444
1445
1446
    # 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.
    "VLLM_ENABLE_RESPONSES_API_STORE":
    lambda: bool(int(os.getenv("VLLM_ENABLE_RESPONSES_API_STORE", "0"))),
1447

xiao-llm's avatar
xiao-llm committed
1448
1449
1450
1451
    # If set, use the fp8 mfma in rocm paged attention.
    "VLLM_ROCM_FP8_MFMA_PAGE_ATTN":
    lambda: bool(int(os.getenv("VLLM_ROCM_FP8_MFMA_PAGE_ATTN", "0"))),

1452
1453
    # Whether to use pytorch symmetric memory for allreduce
    "VLLM_ALLREDUCE_USE_SYMM_MEM":
1454
    lambda: bool(int(os.getenv("VLLM_ALLREDUCE_USE_SYMM_MEM", "1"))),
1455

1456
1457
1458
    # Allows vllm to find tuned config under customized folder
    "VLLM_TUNED_CONFIG_FOLDER":
    lambda: os.getenv("VLLM_TUNED_CONFIG_FOLDER", None),
1459
1460
1461
1462
1463
1464

    # Allows harmony instructions to be injected on system messages
    "VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS":
    lambda: bool(
        int(os.getenv("VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS", "0"))),

1465
1466
1467
    # Add optional custom scopes for profiling, disable to avoid overheads
    "VLLM_CUSTOM_SCOPES_FOR_PROFILING":
    lambda: bool(int(os.getenv("VLLM_CUSTOM_SCOPES_FOR_PROFILING", "0"))),
1468

1469
1470
1471
1472
    # Add optional nvtx scopes for profiling, disable to avoid overheads
    "VLLM_NVTX_SCOPES_FOR_PROFILING":
    lambda: bool(int(os.getenv("VLLM_NVTX_SCOPES_FOR_PROFILING", "0"))),

1473
1474
1475
1476
    # Represent block hashes in KV cache events as 64-bit integers instead of
    # raw bytes. Defaults to True for backward compatibility.
    "VLLM_KV_EVENTS_USE_INT_BLOCK_HASHES":
    lambda: bool(int(os.getenv("VLLM_KV_EVENTS_USE_INT_BLOCK_HASHES", "1"))),
1477
1478
1479
1480
1481
1482

    # Name of the shared memory buffer used for object storage.
    # Only effective when mm_config.mm_processor_cache_type == "shm".
    "VLLM_OBJECT_STORAGE_SHM_BUFFER_NAME":
    lambda: os.getenv("VLLM_OBJECT_STORAGE_SHM_BUFFER_NAME",
                      "VLLM_OBJECT_STORAGE_SHM_BUFFER"),
1483

1484
1485
1486
1487
1488
1489
1490
1491
1492
    # The size in MB of the buffers (NVL and RDMA) used by DeepEP
    "VLLM_DEEPEP_BUFFER_SIZE_MB":
    lambda: int(os.getenv("VLLM_DEEPEP_BUFFER_SIZE_MB", "1024")),

    # 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
    "VLLM_DBO_COMM_SMS":
    lambda: int(os.getenv("VLLM_DBO_COMM_SMS", "20")),

1493
1494
1495
1496
1497
1498
1499
    # Valid values are container,code_interpreter,web_search_preview
    # ex GPT_OSS_SYSTEM_TOOL_MCP_LABELS=container,code_interpreter
    "GPT_OSS_SYSTEM_TOOL_MCP_LABELS":
    env_list_with_choices("GPT_OSS_SYSTEM_TOOL_MCP_LABELS", [],
                            ["container",
                            "code_interpreter",
                            "web_search_preview"]),
1500

1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
    # 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)
    "VLLM_ENABLE_INDUCTOR_MAX_AUTOTUNE":
    lambda: bool(int(os.getenv("VLLM_ENABLE_INDUCTOR_MAX_AUTOTUNE", "1"))),
    # If set to 1, enable coordinate_descent_tuning;
    # By default, this is enabled (1)
    "VLLM_ENABLE_INDUCTOR_COORDINATE_DESCENT_TUNING":
    lambda: bool(int(os.getenv("VLLM_ENABLE_INDUCTOR_COORDINATE_DESCENT_TUNING",
        "1"))),

1512
1513
1514
1515
1516
1517
1518
    # Flag to enable NCCL symmetric memory allocation and registration
    "VLLM_USE_NCCL_SYMM_MEM":
    lambda: bool(int(os.getenv("VLLM_USE_NCCL_SYMM_MEM", "0"))),

    # NCCL header path
    "VLLM_NCCL_INCLUDE_PATH":
    lambda: os.environ.get("VLLM_NCCL_INCLUDE_PATH", None),
1519
1520
    # Flag to enable FBGemm kernels on model execution
    "VLLM_USE_FBGEMM": lambda: bool(int(os.getenv("VLLM_USE_FBGEMM", "0"))),
1521
    
1522
1523
1524
    # add envs
    
    # used in optest environment to manually set the https port
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
    '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")),
    
zhuwenwen's avatar
zhuwenwen committed
1539
1540
1541
1542
1543
1544
    
    # 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
1545
1546
1547
1548
    # If set, vLLM will use FLASH MLA attention optimizations.
    "VLLM_USE_FLASH_MLA":
    lambda: bool(int(os.getenv("VLLM_USE_FLASH_MLA", "1"))),
    
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
    # 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 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":
1565
    lambda: bool(int(os.environ.get("VLLM_PCIE_USE_CUSTOM_ALLREDUCE", "1"))),
1566
    
1567
1568
    # flag to control vllm to use optimized kernels
    "VLLM_CUSTOM_CACHE":
1569
    lambda: bool(int(os.environ.get("VLLM_CUSTOM_CACHE", "1"))),
1570
    
zhuwenwen's avatar
zhuwenwen committed
1571
1572
1573
1574
    # 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")),
    
1575
1576
1577
    # 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")),
1578

1579
1580
1581
1582
1583
1584
    # '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
1585
    lambda: bool(int(os.getenv("VLLM_HAS_CONTEXT_DEFAULT", "1"))),
1586
1587
1588
    
    # If set, vLLM will transpose weight to use nn layout
    "VLLM_USE_NN":
zhuwenwen's avatar
zhuwenwen committed
1589
    lambda: (os.environ.get("VLLM_USE_NN", "True").lower() in
1590
             ("true", "1")),
1591

1592
1593
1594
    # Enable two batch overlap.
    "VLLM_ENABLE_TBO":
    lambda: bool(int(os.getenv("VLLM_ENABLE_TBO", "0"))),
1595
1596
1597

    # If set, vLLM will enable the moe_fused_gate kernel.
    "VLLM_ENABLE_MOE_FUSED_GATE":
zhuwenwen's avatar
zhuwenwen committed
1598
    lambda: bool(int(os.getenv("VLLM_ENABLE_MOE_FUSED_GATE", "1"))),
zhuwenwen's avatar
zhuwenwen committed
1599
    
1600
1601
    # vLLM will use FlashAttention Backend for page attention computation on rocm
    "VLLM_USE_FLASH_ATTN_PA":
zhuwenwen's avatar
zhuwenwen committed
1602
    lambda: (os.environ.get("VLLM_USE_FLASH_ATTN_PA", "True").lower() in
zhuwenwen's avatar
zhuwenwen committed
1603
             ("true", "1")),
zhuwenwen's avatar
zhuwenwen committed
1604
1605
1606
1607
1608
    
    # vLLM will use apex for rmsnorm
    "VLLM_USE_APEX_RN":
    lambda: (os.environ.get("VLLM_USE_APEX_RN", "False").lower() in
             ("true", "1")),
1609
1610
1611
    
    # vLLM will use global cache for moe
    "VLLM_USE_GLOBAL_CACHE13":
1612
        lambda: (os.environ.get("VLLM_USE_GLOBAL_CACHE13", "False").lower() in
1613
                 ("true", "1")),
1614
        
1615
1616
1617
    # vLLM will use lightop for deepseek-v3
    "VLLM_USE_LIGHTOP":
        lambda: (os.environ.get("VLLM_USE_LIGHTOP", "False").lower() in
1618
                 ("true", "1")),
1619
        
1620
1621
1622
1623
1624
    # vLLM will use elenmentwise not triton_
    "VLLM_USE_OPT_ZEROS":
        lambda: (os.environ.get("VLLM_USE_OPT_ZEROS", "False").lower() in
                 ("true", "1")),
        
1625
1626
1627
    # vLLM will use opt cat for deepseek-v3
    "VLLM_USE_OPT_CAT":
        lambda: (os.environ.get("VLLM_USE_OPT_CAT", "True").lower() in
1628
                 ("true", "1")),  
zhuwenwen's avatar
zhuwenwen committed
1629
1630
1631
1632
    # vLLM will use triton moe_sum 
    "VLLM_USE_OPT_MOE_SUM":
        lambda: (os.environ.get("VLLM_USE_OPT_MOE_SUM", "False").lower() in
                 ("true", "1")),  
1633
1634
1635
1636
    # vLLM will use lightop moe_sum_mul_add
    "VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD":
        lambda: (os.environ.get("VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD",
                                "False").lower() in ("true", "1")),
zhuwenwen's avatar
zhuwenwen committed
1637
1638
1639
1640
1641
1642
1643
1644
    # vLLM will use lightop moe_sum 
    "VLLM_USE_LIGHTOP_MOE_SUM":
        lambda: (os.environ.get("VLLM_USE_LIGHTOP_MOE_SUM", "False").lower() in
                 ("true", "1")),  
    # vLLM will use lightop moe_align_block_size 
    "VLLM_USE_LIGHTOP_MOE_ALIGN":
        lambda: (os.environ.get("VLLM_USE_LIGHTOP_MOE_ALIGN", "False").lower() in
                 ("true", "1")),     
1645
1646
1647
1648
    # 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")),  
zhuwenwen's avatar
zhuwenwen committed
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
    # vllm will use rmsquant fused op 
    "USE_FUSED_RMS_QUANT": 
    lambda: bool(int(os.getenv("USE_FUSED_RMS_QUANT", "0"))),
    
    # vllm will use silu_mul_quant fused op,
    # This variable has a default value of true, 
    # but it is still controlled by CRQ and RQ.
    "USE_FUSED_SILU_MUL_QUANT":
    lambda: bool(int(os.getenv("USE_FUSED_SILU_MUL_QUANT", "0"))),

zhuwenwen's avatar
zhuwenwen committed
1659
1660
    # vLLM will split prefill and decode, not mix up
    "VLLM_USE_PD_SPLIT":
zhuwenwen's avatar
zhuwenwen committed
1661
        lambda: (os.environ.get("VLLM_USE_PD_SPLIT", "True").lower() in
zhuwenwen's avatar
zhuwenwen committed
1662
                 ("true", "1")), 
zhuwenwen's avatar
zhuwenwen committed
1663
1664
1665
1666
    # vLLM will sync to avoid pp vmfault
    "VLLM_USE_PP_SYNC":
        lambda: (os.environ.get("VLLM_USE_PP_SYNC", "False").lower() in
                 ("true", "1")), 
1667
1668
    # vLLM will use piecewise
    "VLLM_USE_PIECEWISE":
1669
        lambda: (os.environ.get("VLLM_USE_PIECEWISE", "True").lower() in
1670
                 ("true", "1")), 
1671
1672
    # vllm will use encoding_dsv32.py for dpsk-v32
    "VLLM_USE_V32_ENCODE":
1673
        lambda: (os.environ.get('VLLM_USE_V32_ENCODE', 'False').lower() in
1674
                 ("true", "1")),  
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
    # 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
1687
        (os.environ.get("VLLM_USE_TOPK_RENORM", "True").lower() in
1688
                ("true", "1")),
1689
1690
    # vLLM will use fused RMS + RoPE kernel
    "VLLM_USE_FUSED_RMS_ROPE":
1691
        lambda: (os.environ.get("VLLM_USE_FUSED_RMS_ROPE", "False").lower() in
1692
                 ("true", "1")),
1693
1694
1695
1696
    # vLLM will use fast token id copy
    "VLLM_V1_FAST_TOKEN_ID_COPY":
        lambda: (os.environ.get("VLLM_V1_FAST_TOKEN_ID_COPY", "False").lower() in
                 ("true", "1")),
1697
1698
1699
1700
1701
    # If set to 1/True, enable the reduced top-k/top-p sampling path in the
    # V1 PyTorch-native sampler.
    "VLLM_V1_USE_REDUCED_TOPK_TOPP_SAMPLER":
        lambda: (os.getenv("VLLM_V1_USE_REDUCED_TOPK_TOPP_SAMPLER",
                           "0").lower() in ("true", "1")),
1702
1703
1704
1705
1706
1707
                           
    # W8A8 GEMM backend selection for vLLM quantized models.
    # lightop/triton: 1
    # cutlass: 2 (will remove in the future)
    # blaslt: 3 (default)
    # rocblas: others
1708
1709
1710
    "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")),

1711
1712
1713
1714
    # 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")),
1715
1716
}

1717
# --8<-- [end:env-vars-definition]
1718

1719

1720
def __getattr__(name: str):
1721
1722
1723
1724
1725
1726
1727
1728
    # lazy evaluation of environment variables
    if name in environment_variables:
        return environment_variables[name]()
    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")


def __dir__():
    return list(environment_variables.keys())
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744


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}")


def set_vllm_use_v1(use_v1: bool):
    if is_set("VLLM_USE_V1"):
        raise ValueError(
            "Should not call set_vllm_use_v1() if VLLM_USE_V1 is set "
            "explicitly by the user. Please raise this as a Github "
            "Issue and explicitly set VLLM_USE_V1=0 or 1.")
    os.environ["VLLM_USE_V1"] = "1" if use_v1 else "0"
1745
1746
1747
1748
1749
1750
1751
1752


def compute_hash() -> str:
    """
    WARNING: Whenever a new key is added to this environment
    variables, ensure that it is included in the factors list if
    it affects the computation graph. For example, different values
    of VLLM_PP_LAYER_PARTITION will generate different computation
1753
    graphs, so it is included in the factors list. The env vars that
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
    affect the choice of different kernels or attention backends should
    also be included in the factors list.
    """

    # The values of envs may affects the computation graph.
    # TODO(DefTruth): hash all environment variables?
    # for key in environment_variables:
    #     factorize(key)
    environment_variables_to_hash = [
        "VLLM_PP_LAYER_PARTITION",
        "VLLM_MLA_DISABLE",
1765
        "VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH",
1766
1767
1768
1769
        "VLLM_USE_TRITON_FLASH_ATTN",
        "VLLM_USE_TRITON_AWQ",
        "VLLM_DP_RANK",
        "VLLM_DP_SIZE",
1770
        "VLLM_USE_STANDALONE_COMPILE",
1771
        "VLLM_FUSED_MOE_CHUNK_SIZE",
1772
1773
1774
1775
1776
1777
1778
        "VLLM_FLASHINFER_MOE_BACKEND",
        "VLLM_V1_USE_PREFILL_DECODE_ATTENTION",
        "VLLM_USE_AITER_UNIFIED_ATTENTION",
        "VLLM_ATTENTION_BACKEND",
        "VLLM_USE_FLASHINFER_SAMPLER",
        "VLLM_DISABLED_KERNELS",
        "VLLM_USE_DEEP_GEMM",
1779
1780
        "VLLM_USE_DEEP_GEMM_E8M0",
        "VLLM_USE_DEEP_GEMM_E8M0_HOPPER",
1781
        "VLLM_USE_TRTLLM_FP4_GEMM",
1782
        "VLLM_USE_FUSED_MOE_GROUPED_TOPK",
1783
        "VLLM_USE_FLASHINFER_MOE_FP16",
1784
1785
1786
        "VLLM_USE_FLASHINFER_MOE_FP8",
        "VLLM_USE_FLASHINFER_MOE_FP4",
        "VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8",
1787
        "VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS",
1788
1789
1790
        "VLLM_USE_FLASHINFER_MOE_MXFP4_BF16",
        "VLLM_USE_CUDNN_PREFILL",
        "VLLM_USE_TRTLLM_ATTENTION",
1791
        "VLLM_FLASHINFER_DISABLE_Q_QUANTIZATION",
1792
1793
1794
1795
1796
1797
1798
        "VLLM_ROCM_USE_AITER",
        "VLLM_ROCM_USE_AITER_PAGED_ATTN",
        "VLLM_ROCM_USE_AITER_LINEAR",
        "VLLM_ROCM_USE_AITER_MOE",
        "VLLM_ROCM_USE_AITER_RMSNORM",
        "VLLM_ROCM_USE_AITER_MLA",
        "VLLM_ROCM_USE_AITER_MHA",
1799
1800
        "VLLM_ROCM_USE_AITER_FP4_ASM_GEMM",
        "VLLM_ROCM_USE_TRITON_ROPE",
1801
        "VLLM_ROCM_USE_AITER_FP8BMM",
1802
1803
1804
1805
1806
1807
1808
        "VLLM_ROCM_USE_SKINNY_GEMM",
        "VLLM_ROCM_FP8_PADDING",
        "VLLM_ROCM_MOE_PADDING",
        "VLLM_ROCM_CUSTOM_PAGED_ATTN",
        "VLLM_ROCM_QUICK_REDUCE_QUANTIZATION",
        "VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16",
        "VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB",
xiao-llm's avatar
xiao-llm committed
1809
        "VLLM_ROCM_FP8_MFMA_PAGE_ATTN",
1810
1811
        "VLLM_ENABLE_INDUCTOR_MAX_AUTOTUNE",
        "VLLM_ENABLE_INDUCTOR_COORDINATE_DESCENT_TUNING",
1812
        "VLLM_USE_FBGEMM",
1813
        "VLLM_W8A8_BACKEND",
1814
1815
    ]
    for key in environment_variables_to_hash:
1816
1817
1818
1819
1820
1821
1822
1823
        # if this goes out of sync with environment_variables,
        # it's not a user error, it's a bug
        assert key in environment_variables, \
            "Please update environment_variables_to_hash in envs.py"

    factors = [
        environment_variables[key]() for key in environment_variables_to_hash
    ]
1824

1825
1826
    hash_str = hashlib.md5(str(factors).encode(),
                           usedforsecurity=False).hexdigest()
1827

1828
    return hash_str