fix devices during backprop

hivemind/moe/client/moe.py

@@ -309,16 +309,18 @@ def backward(cls, ctx, *raw_grads: torch.Tensor) -> Tuple[Optional[torch.Tensor]
 
         num_samples, max_experts = dummy_grad_mask.shape
 
-        inputs_per_expert = zip(*(tensor[alive_ii].split(1, dim=0) for tensor in flat_inputs_cpu))
+        alive_ii_cpu = alive_ii.cpu()
+        alive_jj_cpu = alive_jj.cpu()
+        inputs_per_expert = zip(*(tensor[alive_ii_cpu].split(1, dim=0) for tensor in flat_inputs_cpu))
         grad_outputs_per_expert = zip(
-            *(tensor[alive_ii, alive_jj].split(1, dim=0) for tensor in flat_grad_outputs_cpu)
+            *(tensor[alive_ii_cpu, alive_jj_cpu].split(1, dim=0) for tensor in flat_grad_outputs_cpu)
         )
         backward_schema = tuple(nested_flatten((info["forward_schema"], info["outputs_schema"])))
 
         # dispatch tasks to all remote experts, collect responses
         pending_tasks = {}
         for i, j, inputs_ij, grad_outputs_ij in zip(
-            alive_ii.cpu().numpy(), alive_jj.cpu().numpy(), inputs_per_expert, grad_outputs_per_expert
+            alive_ii_cpu.numpy(), alive_jj_cpu.numpy(), inputs_per_expert, grad_outputs_per_expert
         ):
             expert: RemoteExpert = expert_per_sample[i.item()][j.item()]
             stub = get_server_stub(expert.p2p, expert.peer_id)