A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 9, 01:15 UTC
[Misc] Remove deprecated VLLM_ALL2ALL_BACKEND environment variable (#33535)
Signed-off-by: carlory <baofa.fan@daocloud.io> Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
杨朱 · Kiki committed
Feb 3, 2026 at 15:01 UTC
b95cc5014dc7b260e5c70ae33d1b30c54d11306d
4 files changed
+2
-47
.buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_ep_eplb.sh
+1
-1
@@ -43,7 +43,6 @@ trap cleanup EXIT
43
44
for BACK in "${BACKENDS[@]}"; do
45
VLLM_DEEP_GEMM_WARMUP=skip \
46
- VLLM_ALL2ALL_BACKEND=$BACK \
46
vllm serve "$MODEL" \
47
--enforce-eager \
48
--tensor-parallel-size 2 \
@@ -52,6 +51,7 @@ for BACK in "${BACKENDS[@]}"; do
51
--enable-eplb \
52
--trust-remote-code \
53
--max-model-len 2048 \
54
+ --all2all-backend $BACK \
55
--port $PORT &
56
SERVER_PID=$!
57
wait_for_server $PORT
tests/kernels/moe/modular_kernel_tools/common.py
+1
-4
@@ -150,10 +150,7 @@ class Config:
150
"VLLM_USE_DEEP_GEMM": str(int(self.needs_deep_gemm())),
151
}
152
153
- backend = self.all2all_backend()
154
- vllm_config.parallel_config.all2all_backend = backend
155
- if backend is not None:
156
- env_dict.update({"VLLM_ALL2ALL_BACKEND": backend})
153
+ vllm_config.parallel_config.all2all_backend = self.all2all_backend()
154
155
if self.fused_moe_chunk_size is not None:
156
env_dict.update(
vllm/config/parallel.py
-9
@@ -549,15 +549,6 @@ class ParallelConfig:
549
return hash_factors(factors)
550
551
def __post_init__(self) -> None:
552
- # Set all2all_backend from env var if not specified, with deprecation warning
553
- if envs.is_set("VLLM_ALL2ALL_BACKEND"):
554
- logger.warning_once(
555
- "VLLM_ALL2ALL_BACKEND environment variable is deprecated and "
556
- "will be removed in v0.15.0. Please use the "
557
- "--all2all-backend command-line argument instead."
558
- )
559
- self.all2all_backend = envs.VLLM_ALL2ALL_BACKEND
560
-
552
# Continue with the rest of the initialization
553
self.world_size = (
554
self.pipeline_parallel_size
vllm/envs.py
-33
@@ -171,15 +171,6 @@ if TYPE_CHECKING:
171
VLLM_NIXL_SIDE_CHANNEL_HOST: str = "localhost"
172
VLLM_NIXL_SIDE_CHANNEL_PORT: int = 5600
173
VLLM_MOONCAKE_BOOTSTRAP_PORT: int = 8998
174
- VLLM_ALL2ALL_BACKEND: Literal[
175
- "naive",
176
- "pplx",
177
- "deepep_high_throughput",
178
- "deepep_low_latency",
179
- "mori",
180
- "allgather_reducescatter",
181
- "flashinfer_all2allv",
182
- ] = "allgather_reducescatter"
174
VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE: int = 163840
175
VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS: int = 1
176
VLLM_SLEEP_WHEN_IDLE: bool = False
@@ -1292,30 +1283,6 @@ environment_variables: dict[str, Callable[[], Any]] = {
1283
"VLLM_MOONCAKE_BOOTSTRAP_PORT": lambda: int(
1284
os.getenv("VLLM_MOONCAKE_BOOTSTRAP_PORT", "8998")
1285
),
1295
- # [DEPRECATED - will be removed in v0.15.0] all2all backend for vllm's
1296
- # expert parallel communication. Use --all2all-backend CLI argument instead.
1297
- # Available options:
1298
- # - "naive": naive all2all implementation using broadcasts
1299
- # - "allgather_reducescatter": all2all implementation based on allgather and
1300
- # reducescatter
1301
- # - "pplx": use pplx kernels
1302
- # - "deepep_high_throughput", use deepep high-throughput kernels
1303
- # - "deepep_low_latency", use deepep low-latency kernels
1304
- # - "mori", use MoRI kernels
1305
- # - "flashinfer_all2allv", use flashinfer alltoallv kernels for mnnvl
1306
- "VLLM_ALL2ALL_BACKEND": env_with_choices(
1307
- "VLLM_ALL2ALL_BACKEND",
1308
- None,
1309
- [
1310
- "naive",
1311
- "pplx",
1312
- "deepep_high_throughput",
1313
- "deepep_low_latency",
1314
- "mori",
1315
- "allgather_reducescatter",
1316
- "flashinfer_all2allv",
1317
- ],
1318
- ),
1286
# Flashinfer MoE backend for vLLM's fused Mixture-of-Experts support.
1287
# Both require compute capability 10.0 or above.
1288
# Available options: