Summary
The DeepEP v2 all-to-all setup cannot determine NCCL GIN support, raised at tests/kernels/moe/parallel_utils.py:248 inside torch.multiprocessing.spawn, failing all 20 tests in the B200 FusedMoE kernel cluster. The sibling message 'DeepEPv2 communicator properties query failed; networking capability could not be determined.' from vllm/distributed/device_communicators/all2all.py:1074 is the same probe failing via another call path on both ranks. A capability query returning an unusable answer points at the NCCL that torch ships, but a B200-only fabric or driver configuration property would present identically, so this needs confirming against a same-day main build before filing.
Signature
RuntimeError: Failed to determine NCCL GIN support
Affected job clusters
:nvidia: (B200) FusedMoE Layer Kernels:nvidia: (B200) Distributed
Evidence it is torch-nightly specific
Fails on torch-nightly build #89321 and passes on the same-commit baseline #89273 (commit ceb87de065b9).
Representative jobs
Suggested routing
pytorch/pytorch (agent confidence: classification medium, new-failure medium)
Filed automatically by the vLLM torch-nightly triage workflow. The root cause above was produced by an automated analysis of the Buildkite log tail and has not been human-verified.
Summary
The DeepEP v2 all-to-all setup cannot determine NCCL GIN support, raised at tests/kernels/moe/parallel_utils.py:248 inside torch.multiprocessing.spawn, failing all 20 tests in the B200 FusedMoE kernel cluster. The sibling message 'DeepEPv2 communicator properties query failed; networking capability could not be determined.' from vllm/distributed/device_communicators/all2all.py:1074 is the same probe failing via another call path on both ranks. A capability query returning an unusable answer points at the NCCL that torch ships, but a B200-only fabric or driver configuration property would present identically, so this needs confirming against a same-day main build before filing.
Signature
Affected job clusters
:nvidia: (B200) FusedMoE Layer Kernels:nvidia: (B200) DistributedEvidence it is torch-nightly specific
Fails on torch-nightly build #89321 and passes on the same-commit baseline #89273 (commit
ceb87de065b9).Representative jobs
Suggested routing
pytorch/pytorch (agent confidence: classification medium, new-failure medium)
Filed automatically by the vLLM torch-nightly triage workflow. The root cause above was produced by an automated analysis of the Buildkite log tail and has not been human-verified.