support cpu device for pruned-array benchmark (#1874)

Summary:
Pull Request resolved: #1874

as titled, we need to pass whether cpu is used to make pruned-array cpu benchmark working

Differential Revision: D47344641

fbshipit-source-id: 5405b671948ca6e531284f0d4bc1c266aa13ab8b

fbgemm_gpu/bench/split_table_batched_embeddings_benchmark.py

@@ -2305,6 +2305,7 @@ def pruned_array(  # noqa C901
         E,
         requests_data_file=requests_data_file,
         tables=tables,
+        use_cpu=True if device == "cpu" else False,
     )
     requests = [(a.int().to(device), b.int().to(device), c) for (a, b, c) in requests]
 

fbgemm_gpu/fbgemm_gpu/split_embedding_utils.py

@@ -111,6 +111,7 @@ def generate_requests(  # noqa C901
     # and mu_L
     sigma_L: Optional[int] = None,
     emulate_pruning: bool = False,
+    use_cpu: bool = False,
 ) -> List[Tuple[torch.IntTensor, torch.IntTensor, Optional[torch.Tensor]]]:
     if requests_data_file is not None:
         indices_tensor, offsets_tensor, lengths_tensor = torch.load(requests_data_file)
@@ -309,7 +310,9 @@ def generate_requests(  # noqa C901
                 )  # per sample weights will always be FP32
             )
             rs.append(
-                get_table_batched_offsets_from_dense(all_indices[it].view(T, B, L))
+                get_table_batched_offsets_from_dense(
+                    all_indices[it].view(T, B, L), use_cpu=use_cpu
+                )
                 + (weights_tensor,)
             )
     return rs