solution history
Proposed fix: [Accelerate / torch.distributed] Training hangs then times out (NCCL watchdog) on gather/reduce: shape mismatch across ranks - "Cannot apply desired operation due to shape mismatches"
Revision reasons are not recorded in this slice.