[EP] Enable DBO with NIXL EP (#45275)

Signed-off-by: Itay Alroy <ialroy@nvidia.com>

Itay Alroy committed Jun 16, 2026 at 02:37 UTC 3afe659b6bb90b961bf09984166393824a893af9

2 files changed +8 -6

vllm/config/compilation.py

+1 -1

@@ -1197,7 +1197,7 @@ class CompilationConfig:

1197 "are optimized for prefill and are incompatible with CUDA Graphs. "

1198 "In order to use CUDA Graphs for decode-optimized workloads, "

1199 "use --all2all-backend with another option, such as "

1200 - "deepep_low_latency or allgather_reducescatter."

1200 + "deepep_low_latency, nixl_ep, or allgather_reducescatter."

1201 )

1202 self.cudagraph_mode = CUDAGraphMode.NONE

1203

vllm/config/vllm.py

+7 -5

@@ -1457,12 +1457,14 @@ class VllmConfig:

1457 assert a2a_backend in [

1458 "deepep_low_latency",

1459 "deepep_high_throughput",

1460 + "nixl_ep",

1461 ], (

1461 - "Microbatching currently only supports the deepep_low_latency and "

1462 - f"deepep_high_throughput all2all backend. {a2a_backend} is not "

1463 - "supported. To fix use --all2all-backend=deepep_low_latency or "

1464 - "--all2all-backend=deepep_high_throughput and install the DeepEP"

1465 - " kernels."

1462 + "Microbatching currently only supports the deepep_low_latency, "

1463 + "deepep_high_throughput, and nixl_ep all2all backends. "

1464 + f"{a2a_backend} is not supported. To fix use "

1465 + "--all2all-backend=deepep_low_latency, "

1466 + "--all2all-backend=deepep_high_throughput, or "

1467 + "--all2all-backend=nixl_ep and install the matching kernels."

1468 )

1469

1470 if not self.model_config.disable_cascade_attn: