update Gradient Checkpointing to reduce VRAM usage (#2040)

* update Gradient Checkpointing to reduce VRAM usage

* fix inference

GPT_SoVITS/configs/s2.json

@@ -18,7 +18,8 @@
     "warmup_epochs": 0,
     "c_mel": 45,
     "c_kl": 1.0,
-    "text_low_lr_rate": 0.4
+    "text_low_lr_rate": 0.4, 
+    "grad_ckpt": false
   },
   "data": {
     "max_wav_value": 32768.0,

GPT_SoVITS/f5_tts/model/backbones/dit.py

@@ -12,6 +12,7 @@ from __future__ import annotations
 import torch
 from torch import nn
 import torch.nn.functional as F
+from torch.utils.checkpoint import checkpoint
 
 from x_transformers.x_transformers import RotaryEmbedding
 
@@ -121,6 +122,14 @@ class DiT(nn.Module):
         self.norm_out = AdaLayerNormZero_Final(dim)  # final modulation
         self.proj_out = nn.Linear(dim, mel_dim)
 
+    def ckpt_wrapper(self, module):
+        # https://github.com/chuanyangjin/fast-DiT/blob/main/models.py
+        def ckpt_forward(*inputs):
+            outputs = module(*inputs)
+            return outputs
+
+        return ckpt_forward
+
     def forward(#x, prompt_x, x_lens, t, style,cond
         self,#d is channel,n is T
         x0: float["b n d"],  # nosied input audio  # noqa: F722
@@ -129,11 +138,12 @@ class DiT(nn.Module):
         time: float["b"] | float[""],  # time step  # noqa: F821 F722
             dt_base_bootstrap,
         text0,  # : int["b nt"]  # noqa: F722#####condition feature
-
+        use_grad_ckpt,  # bool
         ###no-use
         drop_audio_cond=False,  # cfg for cond audio
         drop_text=False,  # cfg for text
         # mask: bool["b n"] | None = None,  # noqa: F722
+        
     ):
 
         x=x0.transpose(2,1)
@@ -158,7 +168,10 @@ class DiT(nn.Module):
             residual = x
 
         for block in self.transformer_blocks:
-            x = block(x, t, mask=mask, rope=rope)
+            if use_grad_ckpt:
+                x = checkpoint(self.ckpt_wrapper(block), x, t, mask, rope, use_reentrant=False)
+            else:
+                x = block(x, t, mask=mask, rope=rope)
 
         if self.long_skip_connection is not None:
             x = self.long_skip_connection(torch.cat((x, residual), dim=-1))

GPT_SoVITS/module/models.py

@@ -1089,15 +1089,15 @@ class CFM(torch.nn.Module):
             t_tensor = torch.ones(x.shape[0], device=x.device,dtype=mu.dtype) * t
             d_tensor = torch.ones(x.shape[0], device=x.device,dtype=mu.dtype) * d
             # v_pred = model(x, t_tensor, d_tensor, **extra_args)
-            v_pred = self.estimator(x, prompt_x, x_lens, t_tensor,d_tensor, mu,drop_audio_cond=False,drop_text=False).transpose(2, 1)
+            v_pred = self.estimator(x, prompt_x, x_lens, t_tensor,d_tensor, mu, use_grad_ckpt=False,drop_audio_cond=False,drop_text=False).transpose(2, 1)
             if inference_cfg_rate>1e-5:
-                neg = self.estimator(x, prompt_x, x_lens, t_tensor, d_tensor, mu, drop_audio_cond=True, drop_text=True).transpose(2, 1)
+                neg = self.estimator(x, prompt_x, x_lens, t_tensor, d_tensor, mu, use_grad_ckpt=False, drop_audio_cond=True, drop_text=True).transpose(2, 1)
                 v_pred=v_pred+(v_pred-neg)*inference_cfg_rate
             x = x + d * v_pred
             t = t + d
             x[:, :, :prompt_len] = 0
         return x
-    def forward(self, x1, x_lens, prompt_lens, mu):
+    def forward(self, x1, x_lens, prompt_lens, mu, use_grad_ckpt):
         b, _, t = x1.shape
 
         # random timestep
@@ -1117,16 +1117,16 @@ class CFM(torch.nn.Module):
             d_input = d.clone()
             d_input[d_input < 1e-2] = 0
             # with torch.no_grad():
-            v_pred_1 = self.estimator(xt, prompt, x_lens, t, d_input, mu).transpose(2, 1).detach()
+            v_pred_1 = self.estimator(xt, prompt, x_lens, t, d_input, mu, use_grad_ckpt).transpose(2, 1).detach()
             # v_pred_1 = self.diffusion(xt, t, d_input, cond=conditioning).detach()
             x_mid = xt + d[:, None, None] * v_pred_1
             # v_pred_2 = self.diffusion(x_mid, t+d, d_input, cond=conditioning).detach()
-            v_pred_2 = self.estimator(x_mid, prompt, x_lens, t+d, d_input, mu).transpose(2, 1).detach()
+            v_pred_2 = self.estimator(x_mid, prompt, x_lens, t+d, d_input, mu, use_grad_ckpt).transpose(2, 1).detach()
             vt = (v_pred_1 + v_pred_2) / 2
             vt = vt.detach()
             dt = 2*d
 
-        vt_pred = self.estimator(xt, prompt, x_lens, t,dt, mu).transpose(2,1)
+        vt_pred = self.estimator(xt, prompt, x_lens, t,dt, mu, use_grad_ckpt).transpose(2,1)
         loss = 0
 
         # print(45555555,estimator_out.shape,u.shape,x_lens,prompt_lens)#45555555 torch.Size([7, 465, 100]) torch.Size([7, 100, 465]) tensor([461, 461, 451, 451, 442, 442, 442], device='cuda:0') tensor([ 96,  93, 185,  59, 244, 262, 294], device='cuda:0')
@@ -1220,7 +1220,7 @@ class SynthesizerTrnV3(nn.Module):
         self.linear_mel=nn.Conv1d(inter_channels2,100,1,stride=1)
         self.cfm = CFM(100,DiT(**dict(dim=1024, depth=22, heads=16, ff_mult=2, text_dim=inter_channels2, conv_layers=4)),)#text_dim is condition feature dim
 
-    def forward(self, ssl, y, mel,ssl_lengths,y_lengths, text, text_lengths,mel_lengths):#ssl_lengths no need now
+    def forward(self, ssl, y, mel,ssl_lengths,y_lengths, text, text_lengths,mel_lengths, use_grad_ckpt):#ssl_lengths no need now
         with autocast(enabled=False):
             y_mask = torch.unsqueeze(commons.sequence_mask(y_lengths, y.size(2)), 1).to(y.dtype)
             ge = self.ref_enc(y[:,:704] * y_mask, y_mask)
@@ -1245,7 +1245,7 @@ class SynthesizerTrnV3(nn.Module):
         minn=min(mel.shape[-1],fea.shape[-1])
         mel=mel[:,:,:minn]
         fea=fea[:,:,:minn]
-        cfm_loss= self.cfm(mel, mel_lengths, prompt_len, fea)
+        cfm_loss= self.cfm(mel, mel_lengths, prompt_len, fea, use_grad_ckpt)
         return cfm_loss
 
     @torch.no_grad()

GPT_SoVITS/s2_train_v3.py

@@ -304,7 +304,7 @@ def train_and_evaluate(
             text, text_lengths = text.to(device), text_lengths.to(device)
 
         with autocast(enabled=hps.train.fp16_run):
-            cfm_loss = net_g(ssl, spec, mel,ssl_lengths,spec_lengths, text, text_lengths,mel_lengths)
+            cfm_loss = net_g(ssl, spec, mel,ssl_lengths,spec_lengths, text, text_lengths,mel_lengths, use_grad_ckpt=hps.train.grad_ckpt)
             loss_gen_all=cfm_loss
         optim_g.zero_grad()
         scaler.scale(loss_gen_all).backward()

webui.py

@@ -10,6 +10,7 @@ import json,yaml,torch,pdb,re,shutil
 import platform
 import psutil
 import signal
+os.environ['TORCH_DISTRIBUTED_DEBUG'] = 'INFO'
 torch.manual_seed(233333)
 tmp = os.path.join(now_dir, "TEMP")
 os.makedirs(tmp, exist_ok=True)
@@ -327,7 +328,7 @@ def close_denoise():
     return "已终止语音降噪进程", {"__type__":"update","visible":True}, {"__type__":"update","visible":False}
 
 p_train_SoVITS=None
-def open1Ba(batch_size,total_epoch,exp_name,text_low_lr_rate,if_save_latest,if_save_every_weights,save_every_epoch,gpu_numbers1Ba,pretrained_s2G,pretrained_s2D):
+def open1Ba(batch_size,total_epoch,exp_name,text_low_lr_rate,if_save_latest,if_save_every_weights,save_every_epoch,gpu_numbers1Ba,pretrained_s2G,pretrained_s2D,if_grad_ckpt):
     global p_train_SoVITS
     if(p_train_SoVITS==None):
         with open("GPT_SoVITS/configs/s2.json")as f:
@@ -349,6 +350,7 @@ def open1Ba(batch_size,total_epoch,exp_name,text_low_lr_rate,if_save_latest,if_s
         data["train"]["if_save_every_weights"]=if_save_every_weights
         data["train"]["save_every_epoch"]=save_every_epoch
         data["train"]["gpu_numbers"]=gpu_numbers1Ba
+        data["train"]["grad_ckpt"]=if_grad_ckpt
         data["model"]["version"]=version
         data["data"]["exp_dir"]=data["s2_ckpt_dir"]=s2_dir
         data["save_weight_dir"]=SoVITS_weight_root[int(version[-1])-1]
@@ -779,7 +781,7 @@ def switch_version(version_):
     else:
         gr.Warning(i18n(f'未下载{version.upper()}模型'))
     set_default()
-    return  {'__type__':'update', 'value':pretrained_sovits_name[int(version[-1])-1]}, {'__type__':'update', 'value':pretrained_sovits_name[int(version[-1])-1].replace("s2G","s2D")}, {'__type__':'update', 'value':pretrained_gpt_name[int(version[-1])-1]}, {'__type__':'update', 'value':pretrained_gpt_name[int(version[-1])-1]}, {'__type__':'update', 'value':pretrained_sovits_name[int(version[-1])-1]},{'__type__':'update',"value":default_batch_size,"maximum":default_max_batch_size},{'__type__':'update',"value":default_sovits_epoch,"maximum":max_sovits_epoch},{'__type__':'update',"value":default_sovits_save_every_epoch,"maximum":max_sovits_save_every_epoch},{'__type__':'update',"interactive":True if version!="v3"else False}
+    return  {'__type__':'update', 'value':pretrained_sovits_name[int(version[-1])-1]}, {'__type__':'update', 'value':pretrained_sovits_name[int(version[-1])-1].replace("s2G","s2D")}, {'__type__':'update', 'value':pretrained_gpt_name[int(version[-1])-1]}, {'__type__':'update', 'value':pretrained_gpt_name[int(version[-1])-1]}, {'__type__':'update', 'value':pretrained_sovits_name[int(version[-1])-1]},{'__type__':'update',"value":default_batch_size,"maximum":default_max_batch_size},{'__type__':'update',"value":default_sovits_epoch,"maximum":max_sovits_epoch},{'__type__':'update',"value":default_sovits_save_every_epoch,"maximum":max_sovits_save_every_epoch},{'__type__':'update',"interactive":True if version!="v3"else False},{'__type__':'update',"interactive":True if version == "v3" else False}
 
 if os.path.exists('GPT_SoVITS/text/G2PWModel'):...
 else:
@@ -1016,6 +1018,7 @@ with gr.Blocks(title="GPT-SoVITS WebUI") as app:
                         with gr.Column():                   
                             if_save_latest = gr.Checkbox(label=i18n("是否仅保存最新的ckpt文件以节省硬盘空间"), value=True, interactive=True, show_label=True)
                             if_save_every_weights = gr.Checkbox(label=i18n("是否在每次保存时间点将最终小模型保存至weights文件夹"), value=True, interactive=True, show_label=True)
+                            if_grad_ckpt = gr.Checkbox(label="是否开启梯度检查点节省显存占用", value=False, interactive=True if version == "v3" else False, show_label=True) # 只有V3s2可以用
                         with gr.Row():
                             gpu_numbers1Ba = gr.Textbox(label=i18n("GPU卡号以-分割，每个卡号一个进程"), value="%s" % (gpus), interactive=True)
                 with gr.Row():
@@ -1045,7 +1048,7 @@ with gr.Blocks(title="GPT-SoVITS WebUI") as app:
                         button1Bb_close = gr.Button(i18n("终止GPT训练"), variant="primary",visible=False)
                     with gr.Row():
                         info1Bb=gr.Textbox(label=i18n("GPT训练进程输出信息"))
-            button1Ba_open.click(open1Ba, [batch_size,total_epoch,exp_name,text_low_lr_rate,if_save_latest,if_save_every_weights,save_every_epoch,gpu_numbers1Ba,pretrained_s2G,pretrained_s2D], [info1Ba,button1Ba_open,button1Ba_close])
+            button1Ba_open.click(open1Ba, [batch_size,total_epoch,exp_name,text_low_lr_rate,if_save_latest,if_save_every_weights,save_every_epoch,gpu_numbers1Ba,pretrained_s2G,pretrained_s2D,if_grad_ckpt], [info1Ba,button1Ba_open,button1Ba_close])
             button1Ba_close.click(close1Ba, [], [info1Ba,button1Ba_open,button1Ba_close])
             button1Bb_open.click(open1Bb, [batch_size1Bb,total_epoch1Bb,exp_name,if_dpo,if_save_latest1Bb,if_save_every_weights1Bb,save_every_epoch1Bb,gpu_numbers1Bb,pretrained_s1],   [info1Bb,button1Bb_open,button1Bb_close])
             button1Bb_close.click(close1Bb, [], [info1Bb,button1Bb_open,button1Bb_close])
@@ -1069,7 +1072,7 @@ with gr.Blocks(title="GPT-SoVITS WebUI") as app:
                         tts_info = gr.Textbox(label=i18n("TTS推理WebUI进程输出信息"))
                     open_tts.click(change_tts_inference, [bert_pretrained_dir,cnhubert_base_dir,gpu_number_1C,GPT_dropdown,SoVITS_dropdown, batched_infer_enabled], [tts_info,open_tts,close_tts])
                     close_tts.click(change_tts_inference, [bert_pretrained_dir,cnhubert_base_dir,gpu_number_1C,GPT_dropdown,SoVITS_dropdown, batched_infer_enabled], [tts_info,open_tts,close_tts])
-            version_checkbox.change(switch_version,[version_checkbox],[pretrained_s2G,pretrained_s2D,pretrained_s1,GPT_dropdown,SoVITS_dropdown,batch_size,total_epoch,save_every_epoch,text_low_lr_rate])
+            version_checkbox.change(switch_version,[version_checkbox],[pretrained_s2G,pretrained_s2D,pretrained_s1,GPT_dropdown,SoVITS_dropdown,batch_size,total_epoch,save_every_epoch,text_low_lr_rate, if_grad_ckpt])
         with gr.TabItem(i18n("2-GPT-SoVITS-变声")):gr.Markdown(value=i18n("施工中，请静候佳音"))
     app.queue().launch(#concurrency_count=511, max_size=1022
         server_name="0.0.0.0",