推理batch优化

GPT_SoVITS/AR/models/t2s_model.py

@@ -321,6 +321,11 @@ class Text2SemanticDecoder(nn.Module):
         # 错位
         return targets[:, :-1], targets[:, 1:]
 
+    def check_end(self, logits, samples):
+        if torch.all(torch.argmax(logits, dim=-1) == self.EOS) or torch.all(samples[:, 0] == self.EOS):
+
+            return True
+
     def infer_panel(
         self,
         x,  #####全部文本token
@@ -338,8 +343,8 @@ class Text2SemanticDecoder(nn.Module):
 
         # AR Decoder
         y = prompts
-        
-        x_len = x.shape[1]
+        ref_y_len = y.shape[1] if y is not None else 0
+        x_len = x_lens.max()
         x_attn_mask = torch.zeros((x_len, x_len), dtype=torch.bool)
         stop = False
         # print(1111111,self.num_layers)
@@ -385,7 +390,7 @@ class Text2SemanticDecoder(nn.Module):
         xy_attn_mask = torch.concat([x_attn_mask_pad, y_attn_mask], dim=0).to(
             x.device
         )
-        
+        batch_end_index = [0 for _ in range(x.shape[0])]
 
         for idx in tqdm(range(1500)):
 
@@ -397,17 +402,18 @@ class Text2SemanticDecoder(nn.Module):
             if(idx==0):###第一次跑不能EOS否则没有了
                 logits = logits[:, :-1]  ###刨除1024终止符号的概率
             samples = sample(
-                logits[0], y, top_k=top_k, top_p=top_p, repetition_penalty=1.35, temperature=temperature
-            )[0].unsqueeze(0)
+                logits, y, top_k=top_k, top_p=top_p, repetition_penalty=1.35, temperature=temperature
+            )[0]
             # 本次生成的 semantic_ids 和之前的 y 构成新的 y
             # print(samples.shape)#[1,1]#第一个1是bs
-            y = torch.concat([y, samples], dim=1) 
 
+            y = torch.concat([y, samples], dim=1) 
+            # print(y)
             if early_stop_num != -1 and (y.shape[1] - prefix_len) > early_stop_num:
                 print("use early stop num:", early_stop_num)
                 stop = True
 
-            if torch.argmax(logits, dim=-1)[0] == self.EOS or samples[0, 0] == self.EOS:
+            if self.check_end(logits, samples):
                 # print(torch.argmax(logits, dim=-1)[0] == self.EOS, samples[0, 0] == self.EOS)
                 stop = True
             if stop:
@@ -443,6 +449,7 @@ class Text2SemanticDecoder(nn.Module):
             xy_attn_mask = torch.zeros(
                 (1, x_len + y_len), dtype=torch.bool, device=xy_pos.device
             )
+        batch_end_index = [min(torch.where(by == self.EOS)[0]) if self.EOS in by else idx for by in y ] 
         if ref_free:
-            return y[:, :-1], 0
-        return y[:, :-1], idx-1
+            return y[:, :-1], [0 for _ in range(y.shape[0])], batch_end_index
+        return y[:, :-1], [ref_y_len for _ in range(y.shape[0])], batch_end_index

GPT_SoVITS/AR/models/utils.py

@@ -115,17 +115,17 @@ def logits_to_probs(
     top_p: Optional[int] = None,
     repetition_penalty: float = 1.0,
 ):
-    if previous_tokens is not None:
-        previous_tokens = previous_tokens.squeeze()
+    # if previous_tokens is not None:
+    #     previous_tokens = previous_tokens.squeeze()
     # print(logits.shape,previous_tokens.shape)
     # pdb.set_trace()
     if previous_tokens is not None and repetition_penalty != 1.0:
         previous_tokens = previous_tokens.long()
-        score = torch.gather(logits, dim=0, index=previous_tokens)
+        score = torch.gather(logits, dim=1, index=previous_tokens)
         score = torch.where(
             score < 0, score * repetition_penalty, score / repetition_penalty
         )
-        logits.scatter_(dim=0, index=previous_tokens, src=score)
+        logits.scatter_(dim=1, index=previous_tokens, src=score)
 
     if top_p is not None and top_p < 1.0:
         sorted_logits, sorted_indices = torch.sort(logits, descending=True)
@@ -133,9 +133,9 @@ def logits_to_probs(
             torch.nn.functional.softmax(sorted_logits, dim=-1), dim=-1
         )
         sorted_indices_to_remove = cum_probs > top_p
-        sorted_indices_to_remove[0] = False  # keep at least one option
+        sorted_indices_to_remove[:, 0] = False  # keep at least one option
         indices_to_remove = sorted_indices_to_remove.scatter(
-            dim=0, index=sorted_indices, src=sorted_indices_to_remove
+            dim=1, index=sorted_indices, src=sorted_indices_to_remove
         )
         logits = logits.masked_fill(indices_to_remove, -float("Inf"))
 
@@ -143,7 +143,7 @@ def logits_to_probs(
 
     if top_k is not None:
         v, _ = torch.topk(logits, min(top_k, logits.size(-1)))
-        pivot = v.select(-1, -1).unsqueeze(-1)
+        pivot = v[: , -1].unsqueeze(-1)
         logits = torch.where(logits < pivot, -float("Inf"), logits)
 
     probs = torch.nn.functional.softmax(logits, dim=-1)

GPT_SoVITS/AR/modules/patched_mha_with_cache.py

@@ -448,7 +448,6 @@ def multi_head_attention_forward_patched(
         k = k.view(bsz, num_heads, src_len, head_dim)
         v = v.view(bsz, num_heads, src_len, head_dim)
 
-        # with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=True, enable_mem_efficient=True):
         attn_output = scaled_dot_product_attention(
             q, k, v, attn_mask, dropout_p, is_causal
             )

GPT_SoVITS/inference_webui.py

@@ -66,6 +66,7 @@ from time import time as ttime
 from module.mel_processing import spectrogram_torch
 from my_utils import load_audio
 from tools.i18n.i18n import I18nAuto
+from utils import tensor_padding
 
 i18n = I18nAuto()
 
@@ -311,7 +312,7 @@ def merge_short_text_in_array(texts, threshold):
             result[len(result) - 1] += text
     return result
 
-def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language, how_to_cut=i18n("不切"), top_k=20, top_p=0.6, temperature=0.6, ref_free = False):
+def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language, how_to_cut=i18n("不切"), top_k=20, top_p=0.6, temperature=0.6, ref_free = False, infer_batch = 1):
     if prompt_text is None or len(prompt_text) == 0:
         ref_free = True
     t0 = ttime()
@@ -371,13 +372,28 @@ def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language,
     if not ref_free:
         phones1,bert1,norm_text1=get_phones_and_bert(prompt_text, prompt_language)
 
-    for text in texts:
+    batch_num = len(texts) // infer_batch
+    if len(texts) % infer_batch != 0:
+        batch_num += 1
+    for i in range(batch_num):
+        if i == batch_num - 1:
+            text = texts[i * infer_batch:]
+        else:
+            text = texts[i * infer_batch: (i + 1) * infer_batch]
+    # for text in texts:
         # 解决输入目标文本的空行导致报错的问题
-        if (len(text.strip()) == 0):
+        # 过滤每一个需要合成的文本
+        all_phoneme_ids_batch = []
+        all_bert_batch = []
+        all_phoneme_len_batch = []
+        all_vits_phones2 = []
+        for T in text:
+            if (len(T.strip()) == 0):
                 continue
-        if (text[-1] not in splits): text += "。" if text_language != "en" else "."
-        print(i18n("实际输入的目标文本(每句):"), text)
-        phones2,bert2,norm_text2=get_phones_and_bert(text, text_language)
+            if (T[-1] not in splits): T += "。" if text_language != "en" else "."
+            print(i18n("实际输入的目标文本(每句):"), T)
+            phones2,bert2,norm_text2=get_phones_and_bert(T, text_language)
+            all_vits_phones2+=phones2
             print(i18n("前端处理后的文本(每句):"), norm_text2)
             if not ref_free:
                 bert = torch.cat([bert1, bert2], 1)
@@ -385,14 +401,19 @@ def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language,
             else:
                 bert = bert2
                 all_phoneme_ids = torch.LongTensor(phones2).to(device).unsqueeze(0)
-
             bert = bert.to(device).unsqueeze(0)
-        all_phoneme_len = torch.tensor([all_phoneme_ids.shape[-1]]).to(device)
-        prompt = prompt_semantic.unsqueeze(0).to(device)
+            all_phoneme_ids_batch.append(all_phoneme_ids)
+            all_bert_batch.append(bert) 
+            all_phoneme_len_batch.append(all_phoneme_ids.shape[-1])
+
+        all_phoneme_ids = tensor_padding(all_phoneme_ids_batch)
+        bert = tensor_padding(all_bert_batch)
+        all_phoneme_len = torch.tensor(all_phoneme_len_batch).to(device)
+        prompt = prompt_semantic.unsqueeze(0).to(device).expand(all_phoneme_ids.shape[0], -1, -1)
         t2 = ttime()
         with torch.no_grad():
             # pred_semantic = t2s_model.model.infer(
-            pred_semantic, idx = t2s_model.model.infer_panel(
+            pred_semantic, sidx, eidx = t2s_model.model.infer_panel(
                 all_phoneme_ids,
                 all_phoneme_len,
                 None if ref_free else prompt,
@@ -405,9 +426,7 @@ def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language,
             )
         t3 = ttime()
         # print(pred_semantic.shape,idx)
-        pred_semantic = pred_semantic[:, -idx:].unsqueeze(
-            0
-        )  # .unsqueeze(0)#mq要多unsqueeze一次
+        pred_semantic = torch.cat([pred[si:ei] for pred, si, ei in zip(pred_semantic, sidx, eidx)], dim=0).unsqueeze(0).unsqueeze(0)  # .unsqueeze(0)#mq要多unsqueeze一次
         refer = get_spepc(hps, ref_wav_path)  # .to(device)
         if is_half == True:
             refer = refer.half().to(device)
@@ -416,7 +435,7 @@ def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language,
         # audio = vq_model.decode(pred_semantic, all_phoneme_ids, refer).detach().cpu().numpy()[0, 0]
         audio = (
             vq_model.decode(
-                pred_semantic, torch.LongTensor(phones2).to(device).unsqueeze(0), refer
+                pred_semantic, torch.LongTensor(all_vits_phones2).to(device).unsqueeze(0), refer
             )
                 .detach()
                 .cpu()
@@ -588,12 +607,14 @@ with gr.Blocks(title="GPT-SoVITS WebUI") as app:
                 top_k = gr.Slider(minimum=1,maximum=100,step=1,label=i18n("top_k"),value=5,interactive=True)
                 top_p = gr.Slider(minimum=0,maximum=1,step=0.05,label=i18n("top_p"),value=1,interactive=True)
                 temperature = gr.Slider(minimum=0,maximum=1,step=0.05,label=i18n("temperature"),value=1,interactive=True)
+            with gr.Row():
+                batch_size = gr.Slider(minimum=1,maximum=32,step=1,label=i18n("batch_size"),value=1,interactive=True)
                 inference_button = gr.Button(i18n("合成语音"), variant="primary")
             output = gr.Audio(label=i18n("输出的语音"))
 
         inference_button.click(
             get_tts_wav,
-            [inp_ref, prompt_text, prompt_language, text, text_language, how_to_cut, top_k, top_p, temperature, ref_text_free],
+            [inp_ref, prompt_text, prompt_language, text, text_language, how_to_cut, top_k, top_p, temperature, ref_text_free, batch_size],
             [output],
         )
 

GPT_SoVITS/utils.py

@@ -6,6 +6,7 @@ import logging
 import json
 import subprocess
 import traceback
+from typing import List
 
 import librosa
 import numpy as np
@@ -270,7 +271,6 @@ def clean_checkpoints(path_to_models="logs/44k/", n_ckpts_to_keep=2, sort_by_tim
     del_routine = lambda x: [os.remove(x), del_info(x)]
     rs = [del_routine(fn) for fn in to_del]
 
-
 def get_hparams_from_dir(model_dir):
     config_save_path = os.path.join(model_dir, "config.json")
     with open(config_save_path, "r") as f:
@@ -331,6 +331,12 @@ def get_logger(model_dir, filename="train.log"):
     return logger
 
 
+def tensor_padding(tensor_list: List[torch.Tensor]):
+    max_len = max([t.shape[-1] for t in tensor_list])
+    tensor_list = [torch.cat([t, torch.zeros(*list(t.shape[:-1]), max_len - t.size(-1)).to(t)], dim=-1) for t in tensor_list]
+
+    return torch.cat(tensor_list, dim=0)
+
 class HParams:
     def __init__(self, **kwargs):
         for k, v in kwargs.items():

gweight.txt

@@ -0,0 +1 @@
+GPT_SoVITS/pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt

sweight.txt

@@ -0,0 +1 @@
+GPT_SoVITS/pretrained_models/s2G488k.pth