NCCL is not supported with MIG, so a MIG slice cannot participate in a tensor-parallel group - MIG and TP are mutually exclusive as shipped

Planning error rather than a runtime crash: an operator who partitions an H100 into MIG instances and then tries to run TP=2 inside them gets NCCL unsupported-with-MIG rather than a working deployment. The NVIDIA MIG User Guide lists it flatly under Application Considerations: 'NCCL is currently not supported with MIG.' The request to lift this has been open on NCCL since at least 2025 (NCCL #1716), whose stated motivation is exactly this use case: 'It'd much cheaper to rent 1x A100 or H100 split it 4 way and do a 4-way gpu testing on CI, than paying for 4 gpus.'

MIG instances are exposed as separate CUDA devices with their own GPU and Compute Instance scoping, and the vendor documents a matching set of restrictions that make collective communication across them non-functional: with driver R570 only P2P between MIG instances on the same GPU is supported - 'P2P between MIG instances on different GPUs, or between MIG instances to non-MIG mode GPU devices are not supported'; 'CUDA IPC across GPU instances is not supported. CUDA IPC across Compute instances is supported.' A collective library needs P2P or IPC across the group, so NCCL support was never enabled on MIG.

Do not combine MIG with TP or EP. Either partition the GPU with MIG and run one engine per slice at TP=1 (the supported multi-tenant pattern), or leave MIG off and shard with TP across physical GPUs. If the need is many-GPU CI, MIG alone is the answer and the TP axis has to be dropped. Note CUDA IPC works across Compute instances within one GPU Instance, so MPS-style co-location of several ranks inside one GI is the closest supported approximation.

The same section of the guide blocks CUDA IPC across GPU Instances but permits it across Compute Instances - that asymmetry is the mechanism behind mgpu-mps-max-clients-scale-down-on-mig. Do not read 'MIG supports up to seven instances' as 'MIG gives you a 7-GPU TP group'.