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 14, 00:30 UTC
[Refactor] Remove `VLLM_DEEPEP_LOW_LATENCY_ALLOW_NVLINK` (#27750)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Oct 30, 2025 at 15:32 UTC
a8141fa649d1296488cc5de2b479fed460bb34f4
2 files changed
+1
-8
vllm/distributed/device_communicators/all2all.py
+1
-1
@@ -363,7 +363,7 @@ class DeepEPLLAll2AllManager(DeepEPAll2AllManagerBase):
363
num_rdma_bytes=num_rdma_bytes,
364
low_latency_mode=True,
365
num_qps_per_rank=num_qps_per_rank,
366
- allow_nvlink_for_low_latency_mode=envs.VLLM_DEEPEP_LOW_LATENCY_ALLOW_NVLINK,
366
+ allow_nvlink_for_low_latency_mode=True,
367
allow_mnnvl=envs.VLLM_DEEPEP_LOW_LATENCY_USE_MNNVL,
368
)
369
vllm/envs.py
-7
@@ -207,7 +207,6 @@ if TYPE_CHECKING:
207
VLLM_OBJECT_STORAGE_SHM_BUFFER_NAME: str = "VLLM_OBJECT_STORAGE_SHM_BUFFER"
208
VLLM_DEEPEP_BUFFER_SIZE_MB: int = 1024
209
VLLM_DEEPEP_HIGH_THROUGHPUT_FORCE_INTRA_NODE: bool = False
210
- VLLM_DEEPEP_LOW_LATENCY_ALLOW_NVLINK: bool = True
210
VLLM_DEEPEP_LOW_LATENCY_USE_MNNVL: bool = False
211
VLLM_DBO_COMM_SMS: int = 20
212
VLLM_PATTERN_MATCH_DEBUG: str | None = None
@@ -1400,11 +1399,6 @@ environment_variables: dict[str, Callable[[], Any]] = {
1399
"VLLM_DEEPEP_HIGH_THROUGHPUT_FORCE_INTRA_NODE": lambda: bool(
1400
int(os.getenv("VLLM_DEEPEP_HIGH_THROUGHPUT_FORCE_INTRA_NODE", "0"))
1401
),
1403
- # Allow DeepEP to use nvlink for internode_ll kernel, turn this on for
1404
- # better latency on GB200 like system
1405
- "VLLM_DEEPEP_LOW_LATENCY_ALLOW_NVLINK": lambda: bool(
1406
- int(os.getenv("VLLM_DEEPEP_LOW_LATENCY_ALLOW_NVLINK", "1"))
1407
- ),
1402
# Allow DeepEP to use MNNVL (multi-node nvlink) for internode_ll kernel,
1403
# turn this for better latency on GB200 like system
1404
"VLLM_DEEPEP_LOW_LATENCY_USE_MNNVL": lambda: bool(
@@ -1566,7 +1560,6 @@ def compute_hash() -> str:
1560
"VLLM_NVFP4_GEMM_BACKEND",
1561
"VLLM_USE_FBGEMM",
1562
"VLLM_DEEPEP_HIGH_THROUGHPUT_FORCE_INTRA_NODE",
1569
- "VLLM_DEEPEP_LOW_LATENCY_ALLOW_NVLINK",
1563
"VLLM_DEEPEP_LOW_LATENCY_USE_MNNVL",
1564
]
1565
for key in environment_variables_to_hash: