How can I use sharp with Pytorch(Tensor Parallel)

[47oo]

My sharp version is 3.5
When I run NCCL+Sharp with TP(https://pytorch.org/docs/stable/distributed.tensor.parallel.html ),it turns out that with TP the model is divided into two and each requests its own set of resources, and a conflict occurs sharp + nccl.How to use SHARP with NCCL properly with tensor parallelism factor equal 2?