[Misc] Remove orphaned env vars and stale env-var references (#44749)
Signed-off-by: Daoyuan Li <94409450+DaoyuanLi2816@users.noreply.github.com>
Daoyuan Li committed
Jul 16, 2026 at 17:00 UTC
b88abb5036aa6b7191a3e9341ab652e6562a49d2
3 files changed
+1
-33
docs/configuration/env_vars.md
+1
-1
@@ -5,7 +5,7 @@ vLLM uses the following environment variables to configure the system:
5
!!! warning
6
Please note that `VLLM_PORT` and `VLLM_HOST_IP` set the port and ip for vLLM's **internal usage**. It is not the port and ip for the API server. If you use `--host $VLLM_HOST_IP` and `--port $VLLM_PORT` to start the API server, it will not work.
7
8
- All environment variables used by vLLM are prefixed with `VLLM_`. **Special care should be taken for Kubernetes users**: please do not name the service as `vllm`, otherwise environment variables set by Kubernetes might conflict with vLLM's environment variables, because [Kubernetes sets environment variables for each service with the capitalized service name as the prefix](https://kubernetes.io/docs/concepts/services-networking/service/#environment-variables).
8
+ Most vLLM-specific environment variables are prefixed with `VLLM_` (a handful of standard names — for example `CUDA_VISIBLE_DEVICES`, `MAX_JOBS`, `S3_ACCESS_KEY_ID`/`S3_SECRET_ACCESS_KEY`/`S3_ENDPOINT_URL`, `DO_NOT_TRACK`, `NO_COLOR` — are also read directly when set). **Special care should be taken for Kubernetes users**: please do not name the service as `vllm`, otherwise environment variables set by Kubernetes might conflict with vLLM's environment variables, because [Kubernetes sets environment variables for each service with the capitalized service name as the prefix](https://kubernetes.io/docs/concepts/services-networking/service/#environment-variables).
9
10
```python
11
--8<-- "vllm/envs.py:env-vars-definition"
docs/serving/expert_parallel_deployment.md
-2
@@ -220,8 +220,6 @@ For multi-node deployment, add these EPLB flags to each node's command. We recom
220
221
- Use simulator flags `VLLM_MOE_ROUTING_SIMULATION_STRATEGY=uniform_random` and `VLLM_RANDOMIZE_DP_DUMMY_INPUTS=1` so token routing is balanced across EP ranks.
222
223
-- Increasing `VLLM_MOE_DP_CHUNK_SIZE` may increase throughput by increasing the maximum batch size for inter-rank token transfers. This may cause DeepEP to throw `assert self.nvshmem_qp_depth >= (num_max_dispatch_tokens_per_rank + 1) * 2`, which can be fixed by increasing environment variable `NVSHMEM_QP_DEPTH`.
224
-
223
## Disaggregated Serving (Prefill/Decode Split)
224
225
For production deployments requiring strict SLA guarantees for time-to-first-token and inter-token latency, disaggregated serving allows independent scaling of prefill and decode operations.
vllm/envs.py
-30
@@ -127,7 +127,6 @@ if TYPE_CHECKING:
127
VLLM_MXFP8_EMULATION_DEQUANT_AT_LOAD: bool = True
128
VLLM_ROCM_USE_AITER: bool = False
129
VLLM_ROCM_USE_AITER_CUSTOM_AR: bool = True
130
- VLLM_ROCM_USE_AITER_PAGED_ATTN: bool = False
130
VLLM_ROCM_USE_AITER_LINEAR: bool = True
131
VLLM_ROCM_USE_AITER_LINEAR_HIPBMM: bool = False
132
VLLM_ROCM_USE_AITER_MOE: bool = True
@@ -182,8 +181,6 @@ if TYPE_CHECKING:
181
VLLM_HUMMING_MOE_GEMM_TYPE: Literal["indexed", "grouped", "auto"] | None = None
182
VLLM_DEEPEPLL_NVFP4_DISPATCH: bool = False
183
VLLM_V1_USE_OUTLINES_CACHE: bool = False
185
- VLLM_TPU_BUCKET_PADDING_GAP: int = 0
186
- VLLM_TPU_MOST_MODEL_LEN: int | None = None
184
VLLM_TPU_USING_PATHWAYS: bool = False
185
VLLM_USE_DEEP_GEMM: bool = True
186
VLLM_MOE_USE_DEEP_GEMM: bool = True
@@ -1194,11 +1191,6 @@ environment_variables: dict[str, Callable[[], Any]] = {
1191
"VLLM_ROCM_USE_AITER_CUSTOM_AR": lambda: (
1192
os.getenv("VLLM_ROCM_USE_AITER_CUSTOM_AR", "True").lower() in ("true", "1")
1193
),
1197
- # Whether to use aiter paged attention.
1198
- # By default is disabled.
1199
- "VLLM_ROCM_USE_AITER_PAGED_ATTN": lambda: (
1200
- os.getenv("VLLM_ROCM_USE_AITER_PAGED_ATTN", "False").lower() in ("true", "1")
1201
- ),
1194
# use aiter linear op if aiter ops are enabled
1195
# The following list of related ops
1196
# - scaled_mm (per-tensor / rowwise)
@@ -1432,8 +1424,6 @@ environment_variables: dict[str, Callable[[], Any]] = {
1424
"VLLM_RAY_EXTRA_ENV_VARS_TO_COPY": lambda: os.getenv(
1425
"VLLM_RAY_EXTRA_ENV_VARS_TO_COPY", ""
1426
),
1435
- # Whether to use S3 path for model loading in CI via RunAI Streamer
1436
- "VLLM_CI_USE_S3": lambda: os.environ.get("VLLM_CI_USE_S3", "0") == "1",
1427
# Use model_redirect to redirect the model name to a local folder.
1428
# `model_redirect` can be a json file mapping the model between
1429
# repo_id and local folder:
@@ -1482,16 +1472,6 @@ environment_variables: dict[str, Callable[[], Any]] = {
1472
"VLLM_V1_USE_OUTLINES_CACHE": lambda: (
1473
os.environ.get("VLLM_V1_USE_OUTLINES_CACHE", "0") == "1"
1474
),
1485
- # Gap between padding buckets for the forward pass. So we have
1486
- # 8, we will run forward pass with [16, 24, 32, ...].
1487
- "VLLM_TPU_BUCKET_PADDING_GAP": lambda: (
1488
- int(os.environ["VLLM_TPU_BUCKET_PADDING_GAP"])
1489
- if "VLLM_TPU_BUCKET_PADDING_GAP" in os.environ
1490
- else 0
1491
- ),
1492
- "VLLM_TPU_MOST_MODEL_LEN": lambda: maybe_convert_int(
1493
- os.environ.get("VLLM_TPU_MOST_MODEL_LEN", None)
1494
- ),
1475
# Whether using Pathways
1476
"VLLM_TPU_USING_PATHWAYS": lambda: bool(
1477
"proxy" in os.getenv("JAX_PLATFORMS", "").lower()
@@ -1667,15 +1647,6 @@ environment_variables: dict[str, Callable[[], Any]] = {
1647
"VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE": lambda: int(
1648
os.getenv("VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE", "163840")
1649
),
1670
- # Specifies the thresholds of the communicated tensor sizes under which
1671
- # vllm should use flashinfer fused allreduce. The variable should be a
1672
- # JSON with the following format:
1673
- # { <world size>: <max size in mb> }
1674
- # Unspecified world sizes will fall back to
1675
- # { 2: 64, 4: 1, <everything else>: 0.5 }
1676
- "VLLM_FLASHINFER_ALLREDUCE_FUSION_THRESHOLDS_MB": lambda: json.loads(
1677
- os.getenv("VLLM_FLASHINFER_ALLREDUCE_FUSION_THRESHOLDS_MB", "{}")
1678
- ),
1650
# MoE routing strategy selector.
1651
# See `RoutingSimulator.get_available_strategies()` # for available
1652
# strategies.
@@ -2147,7 +2118,6 @@ def compile_factors() -> dict[str, object]:
2118
"VLLM_DP_MASTER_PORT",
2119
"VLLM_NIXL_SIDE_CHANNEL_HOST",
2120
"VLLM_RANDOMIZE_DP_DUMMY_INPUTS",
2150
- "VLLM_CI_USE_S3",
2121
"VLLM_MODEL_REDIRECT_PATH",
2122
"VLLM_HOST_IP",
2123
"VLLM_FORCE_AOT_LOAD",