update

egs/aishell/paraformer/run.sh

@@ -180,7 +180,6 @@ if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
                 --input_size $feats_dim \
                 --ngpu $gpu_num \
                 --num_worker_count $count \
-                --multiprocessing_distributed true \
                 --dist_init_method $init_method \
                 --dist_world_size $world_size \
                 --dist_rank $rank \

funasr/bin/train.py

@@ -77,6 +77,12 @@ def get_parser():
         help="Whether to use the find_unused_parameters in "
              "torch.nn.parallel.DistributedDataParallel ",
     )
+    parser.add_argument(
+        "--gpu_id",
+        type=int,
+        default=0,
+        help="local gpu id.",
+    )
 
     # cudnn related
     parser.add_argument(
@@ -399,6 +405,7 @@ if __name__ == '__main__':
     torch.backends.cudnn.deterministic = args.cudnn_deterministic
 
     # ddp init
+    os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpu_id)
     args.distributed = args.dist_world_size > 1
     distributed_option = build_distributed(args)
 

funasr/datasets/small_datasets/sequence_iter_factory.py

@@ -62,7 +62,7 @@ class SequenceIterFactory(AbsIterFactory):
         # sampler
         dataset_conf = args.dataset_conf
         batch_sampler = LengthBatchSampler(
-            batch_bins=dataset_conf["batch_size"],
+            batch_bins=dataset_conf["batch_size"] * args.ngpu,
             shape_files=shape_files,
             sort_in_batch=dataset_conf["sort_in_batch"] if hasattr(dataset_conf, "sort_in_batch") else "descending",
             sort_batch=dataset_conf["sort_batch"] if hasattr(dataset_conf, "sort_batch") else "ascending",