diff --git a/extract_f0_print.py b/extract_f0_print.py
index 7685ba5..5a88613 100644
--- a/extract_f0_print.py
+++ b/extract_f0_print.py
@@ -154,7 +154,7 @@ if __name__ == "__main__":
                 f0method,
             ),
         )
-        p.start()
         ps.append(p)
-    for p in ps:
-        p.join()
+        p.start()
+    for i in range(n_p):
+        ps[i].join()
diff --git a/infer-web.py b/infer-web.py
index 0a28d85..01554b6 100644
--- a/infer-web.py
+++ b/infer-web.py
@@ -1,7 +1,6 @@
 import torch, os, traceback, sys, warnings, shutil, numpy as np
 
 os.environ["no_proxy"] = "localhost, 127.0.0.1, ::1"
-from multiprocessing import cpu_count
 import threading
 from time import sleep
 from subprocess import Popen
@@ -25,7 +24,6 @@ import ffmpeg
 
 i18n = I18nAuto()
 # 判断是否有能用来训练和加速推理的N卡
-ncpu = cpu_count()
 ngpu = torch.cuda.device_count()
 gpu_infos = []
 mem = []
@@ -1436,10 +1434,10 @@ with gr.Blocks() as app:
                 )
                 np7 = gr.Slider(
                     minimum=0,
-                    maximum=ncpu,
+                    maximum=config.n_cpu,
                     step=1,
                     label=i18n("提取音高和处理数据使用的CPU进程数"),
-                    value=ncpu,
+                    value=config.n_cpu,
                     interactive=True,
                 )
             with gr.Group():  # 暂时单人的, 后面支持最多4人的#数据处理
diff --git a/train_nsf_sim_cache_sid_load_pretrain.py b/train_nsf_sim_cache_sid_load_pretrain.py
index 910f576..6244831 100644
--- a/train_nsf_sim_cache_sid_load_pretrain.py
+++ b/train_nsf_sim_cache_sid_load_pretrain.py
@@ -66,18 +66,22 @@ class EpochRecorder:
 
 
 def main():
-    # n_gpus = torch.cuda.device_count()
+    n_gpus = torch.cuda.device_count()
     os.environ["MASTER_ADDR"] = "localhost"
     os.environ["MASTER_PORT"] = "51545"
 
-    mp.spawn(
-        run,
-        nprocs=n_gpus,
-        args=(
+    children = []
+    for i in range(n_gpus):
+        subproc = mp.Process(target=run, args=(
+            i,
             n_gpus,
             hps,
-        ),
-    )
+        ))
+        children.append(subproc)
+        subproc.start()
+
+    for i in range(n_gpus):
+        children[i].join()
 
 
 def run(rank, n_gpus, hps):
diff --git a/trainset_preprocess_pipeline_print.py b/trainset_preprocess_pipeline_print.py
index 5c87061..fe1643d 100644
--- a/trainset_preprocess_pipeline_print.py
+++ b/trainset_preprocess_pipeline_print.py
@@ -115,10 +115,10 @@ class PreProcess:
                     p = multiprocessing.Process(
                         target=self.pipeline_mp, args=(infos[i::n_p],)
                     )
-                    p.start()
                     ps.append(p)
-                for p in ps:
-                    p.join()
+                    p.start()
+                for i in range(n_p):
+                    ps[i].join()
         except:
             println("Fail. %s" % traceback.format_exc())