blackteay/GPT-SoVITS
1
0
Fork 0
mirror of https://github.com/RVC-Boss/GPT-SoVITS.git synced 2025-10-08 07:49:59 +08:00
Code Issues Packages Projects Releases Wiki Activity

同步了0306更新

Browse Source
This commit is contained in:
XTer 2024-03-07 14:31:33 +08:00
parent a2fd134288
commit 74f81f3009
3 changed files with 156 additions and 78 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

153
GPT_SoVITS/AR/models/t2s_model.py
View File

@ -22,6 +22,11 @@ from torch import nn
from torch.nn import functional as F
from torchmetrics.classification import MulticlassAccuracy
try:
from flash_attn import flash_attn_with_kvcache
except ImportError:
flash_attn_with_kvcache = None
default_config = {
"embedding_dim": 512,
"hidden_dim": 512,
@ -321,6 +326,90 @@ class Text2SemanticDecoder(nn.Module):
# 错位
return targets[:, :-1], targets[:, 1:]
def infer_one_step(self, x, xy_attn_mask, k_cache, v_cache, cache_seqlens):
hidden_dim = x.shape[-1]
for layer_id in range(self.num_layers):
layer = self.h.layers[layer_id]
q, k, v = F.linear(
x,
layer.self_attn.in_proj_weight,
layer.self_attn.in_proj_bias
).chunk(3, dim=-1)
batch_size = q.shape[0]
q_len = q.shape[1]
if flash_attn_with_kvcache is None:
past_k = k_cache[layer_id]
past_v = v_cache[layer_id]
if past_k is not None:
k = torch.cat([past_k, k], 1)
v = torch.cat([past_v, v], 1)
k_cache[layer_id] = k
v_cache[layer_id] = v
kv_len = k.shape[1]
q = q.view(batch_size, q_len, layer.self_attn.num_heads, -1).transpose(1, 2)
k = k.view(batch_size, kv_len, layer.self_attn.num_heads, -1).transpose(1, 2)
v = v.view(batch_size, kv_len, layer.self_attn.num_heads, -1).transpose(1, 2)
if xy_attn_mask is None:
attn = F.scaled_dot_product_attention(q, k, v)
else:
attn = F.scaled_dot_product_attention(q, k, v, ~xy_attn_mask)
attn = attn.permute(2, 0, 1, 3).reshape(-1, hidden_dim)
else:
q = q.view(batch_size, q_len, layer.self_attn.num_heads, -1)
k = k.view(batch_size, q_len, layer.self_attn.num_heads, -1)
v = v.view(batch_size, q_len, layer.self_attn.num_heads, -1)
if xy_attn_mask is None:
attn = flash_attn_with_kvcache(q, k_cache[layer_id], v_cache[layer_id], k, v, cache_seqlens=cache_seqlens, causal=True)
else:
# NOTE: there's a slight difference with the result produced by SDPA.
x_len = (~xy_attn_mask).sum(1)[0].item()
attn_x = flash_attn_with_kvcache(
q[:, :x_len],
k_cache[layer_id],
v_cache[layer_id],
k[:, :x_len],
v[:, :x_len],
cache_seqlens=cache_seqlens,
causal=False
)
attn_y = flash_attn_with_kvcache(
q[:, x_len:],
k_cache[layer_id],
v_cache[layer_id],
k[:, x_len:],
v[:, x_len:],
cache_seqlens=cache_seqlens + x_len,
causal=True
)
attn = torch.cat([attn_x, attn_y], dim=1)
attn = attn.view(-1, hidden_dim)
attn_out = F.linear(attn, layer.self_attn.out_proj.weight, layer.self_attn.out_proj.bias)
x = layer.norm1(x + attn_out, None)
x = layer.norm2(x + layer.linear2(F.relu(layer.linear1(x))), None)
xy_dec = x
logits = self.ar_predict_layer(
xy_dec[:, -1]
)
return logits
def infer_panel(
self,
x, #####全部文本token
@ -343,17 +432,13 @@ class Text2SemanticDecoder(nn.Module):
x_attn_mask = torch.zeros((x_len, x_len), dtype=torch.bool)
stop = False
# print(1111111,self.num_layers)
cache = {
"all_stage": self.num_layers,
"k": [None] * self.num_layers, ###根据配置自己手写
"v": [None] * self.num_layers,
# "xy_pos":None,##y_pos位置编码每次都不一样的没法缓存，每次都要重新拼xy_pos.主要还是写法原因，其实是可以历史统一一样的，但也没啥计算量就不管了
"y_emb": None, ##只需要对最新的samples求emb，再拼历史的就行
# "logits":None,###原版就已经只对结尾求再拼接了，不用管
# "xy_dec":None,###不需要，本来只需要最后一个做logits
"first_infer": 1,
"stage": 0,
}
if flash_attn_with_kvcache is not None:
k_cache = [torch.empty(x.shape[0], 2048, 16, 32, dtype=x.dtype, device=x.device) for _ in range(self.num_layers)]
v_cache = [torch.empty(x.shape[0], 2048, 16, 32, dtype=x.dtype, device=x.device) for _ in range(self.num_layers)]
else:
k_cache = [None] * self.num_layers
v_cache = [None] * self.num_layers
################### first step ##########################
if y is not None:
y_emb = self.ar_audio_embedding(y)
@ -361,7 +446,6 @@ class Text2SemanticDecoder(nn.Module):
prefix_len = y.shape[1]
y_pos = self.ar_audio_position(y_emb)
xy_pos = torch.concat([x, y_pos], dim=1)
cache["y_emb"] = y_emb
ref_free = False
else:
y_emb = None
@ -386,21 +470,23 @@ class Text2SemanticDecoder(nn.Module):
x.device
)
cache_seqlens = torch.zeros(x.shape[0], dtype=torch.int32, device=x.device)
for idx in tqdm(range(1500)):
logits = self.infer_one_step(xy_pos, xy_attn_mask, k_cache, v_cache, cache_seqlens)
xy_dec, _ = self.h((xy_pos, None), mask=xy_attn_mask, cache=cache)
logits = self.ar_predict_layer(
xy_dec[:, -1]
) ##不用改，如果用了cache的默认就是只有一帧，取最后一帧一样的
# samples = topk_sampling(logits, top_k=top_k, top_p=1.0, temperature=temperature)
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
if idx == 0:
cache_seqlens += xy_pos.shape[1]
else:
cache_seqlens += 1
xy_attn_mask = None
if idx == 0:
logits = logits[:, :-1]
samples = sample(
logits[0], y, top_k=top_k, top_p=top_p, repetition_penalty=1.35, temperature=temperature
)[0].unsqueeze(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)
if early_stop_num != -1 and (y.shape[1] - prefix_len) > early_stop_num:
@ -408,12 +494,8 @@ class Text2SemanticDecoder(nn.Module):
stop = True
if torch.argmax(logits, dim=-1)[0] == self.EOS or samples[0, 0] == self.EOS:
# print(torch.argmax(logits, dim=-1)[0] == self.EOS, samples[0, 0] == self.EOS)
stop = True
if stop:
# if prompts.shape[1] == y.shape[1]:
# y = torch.concat([y, torch.zeros_like(samples)], dim=1)
# print("bad zero prediction")
if y.shape[1] == 0:
y = torch.concat([y, torch.zeros_like(samples)], dim=1)
print("bad zero prediction")
@ -421,28 +503,9 @@ class Text2SemanticDecoder(nn.Module):
break
####################### update next step ###################################
cache["first_infer"] = 0
if cache["y_emb"] is not None:
y_emb = torch.cat(
[cache["y_emb"], self.ar_audio_embedding(y[:, -1:])], dim = 1
)
cache["y_emb"] = y_emb
y_pos = self.ar_audio_position(y_emb)
xy_pos = y_pos[:, -1:]
else:
y_emb = self.ar_audio_embedding(y[:, -1:])
cache["y_emb"] = y_emb
y_pos = self.ar_audio_position(y_emb)
xy_pos = y_pos
y_len = y_pos.shape[1]
xy_pos = y_emb * self.ar_audio_position.x_scale + self.ar_audio_position.alpha * self.ar_audio_position.pe[:, prompts.shape[1] + idx]
###最右边一列（是错的）
# xy_attn_mask=torch.ones((1, x_len+y_len), dtype=torch.bool,device=xy_pos.device)
# xy_attn_mask[:,-1]=False
###最下面一行（是对的）
xy_attn_mask = torch.zeros(
(1, x_len + y_len), dtype=torch.bool, device=xy_pos.device
)
if ref_free:
return y[:, :-1], 0
return y[:, :-1], idx - 1

13
docs/cn/Changelog_CN.md
View File

@ -135,6 +135,19 @@
4-colab修复不开启公网url
### 20240306更新
1-推理加速50%（RTX3090+pytorch2.2.1+cu11.8+win10+py39 tested）https://github.com/RVC-Boss/GPT-SoVITS/pull/672
2-如果用faster whisper非中文ASR不再需要先下中文funasr模型
3-修复uvr5去混响模型 是否混响 反的 https://github.com/RVC-Boss/GPT-SoVITS/pull/610
4-faster whisper如果无cuda可用自动cpu推理 https://github.com/RVC-Boss/GPT-SoVITS/pull/675
5-修改is_half的判断使在Mac上能正常CPU推理 https://github.com/RVC-Boss/GPT-SoVITS/pull/573
todolist：
1-中文多音字推理优化

2
tools/uvr5/uvr5_weights/.gitignore vendored Normal file
View File

@ -0,0 +1,2 @@
*
!.gitignore