[NPU] test TransformerTTS with NPU

Author: windstamp

examples/ljspeech/tts1/local/train.sh

@@ -8,5 +8,5 @@ python3 ${BIN_DIR}/train.py \
     --dev-metadata=dump/dev/norm/metadata.jsonl \
     --config=${config_path} \
     --output-dir=${train_output_path} \
-    --ngpu=2 \
+    --ngpu=1 \
     --phones-dict=dump/phone_id_map.txt

examples/ljspeech/tts1/run.sh

@@ -3,7 +3,7 @@
 set -e
 source path.sh
 
-gpus=0,1
+gpus=0
 stage=0
 stop_stage=100
 

paddlespeech/t2s/exps/transformer_tts/train.py

@@ -42,10 +42,12 @@
 def train_sp(args, config):
     # decides device type and whether to run in parallel
     # setup running environment correctly
-    if (not paddle.is_compiled_with_cuda()) or args.ngpu == 0:
-        paddle.set_device("cpu")
-    else:
+    if paddle.is_compiled_with_cuda() and args.ngpu > 0:
         paddle.set_device("gpu")
+    elif paddle.is_compiled_with_npu() and args.ngpu > 0:
+        paddle.set_device("npu")
+    else:
+        paddle.set_device("cpu")
     world_size = paddle.distributed.get_world_size()
     if world_size > 1:
         paddle.distributed.init_parallel_env()