trunk/1444a820223be3c5b27914ba733ba76d2e7961ef: [TokenSwitch] Hold `topk_idx` in `torch.int64` (#195985)
- PyTorch: 911 events in the last 90 days
- PyTorch: 900th Release in the last 90 days
- Previous: earlier the same day · trunk/274ab543b884df9f95c147f27133797a452de833
What happened
because NCCL_EP_ALGO_HIGH_THROUGHPUT is hardcoded at https://github.com/pytorch/pytorch/blob/5405e96f56b2caf08548e7c614d0e9c576bcc8dc/torch/csrc/distributed/c10d/symm_mem/nccl_ep.cu#L134 and the algorithm seems to expect topk_idx to be int64 as per https://github.com/NVIDIA/nccl-extensions/blob/850d97a66ab6882ed0fca3856ce620f0d0ed7f93/nccl_ep/nccl_ep.cc#L4339 . Pull Request resolved: #195985 Approved by: https://git…
Summary assembled by rule from the sources below