Spaces:
Running
on
Zero
Running
on
Zero
XXXXRT666
commited on
Commit
·
4ae2215
1
Parent(s):
301f27c
- AR/models/embedding.py +0 -45
- AR/models/structs.py +10 -2
- AR/models/t2s_model_abc.py +54 -2
- AR/models/t2s_model_flash_attn.py +59 -37
- inference_webui.py +59 -92
AR/models/embedding.py
CHANGED
@@ -33,51 +33,6 @@ class TokenEmbedding(nn.Module):
|
|
33 |
return x
|
34 |
|
35 |
|
36 |
-
class SinePositionalEmbedding(nn.Module):
|
37 |
-
def __init__(
|
38 |
-
self,
|
39 |
-
embedding_dim: int,
|
40 |
-
dropout: float = 0.0,
|
41 |
-
scale: bool = False,
|
42 |
-
alpha: bool = False,
|
43 |
-
):
|
44 |
-
super().__init__()
|
45 |
-
self.embedding_dim = embedding_dim
|
46 |
-
self.x_scale = math.sqrt(embedding_dim) if scale else 1.0
|
47 |
-
self.alpha = nn.Parameter(torch.ones(1), requires_grad=alpha)
|
48 |
-
self.dropout = torch.nn.Dropout(p=dropout)
|
49 |
-
|
50 |
-
self.reverse = False
|
51 |
-
self.pe = None
|
52 |
-
self.extend_pe(torch.tensor(0.0).expand(1, 4000))
|
53 |
-
|
54 |
-
def extend_pe(self, x):
|
55 |
-
"""Reset the positional encodings."""
|
56 |
-
if self.pe is not None:
|
57 |
-
if self.pe.size(1) >= x.size(1):
|
58 |
-
if self.pe.dtype != x.dtype or self.pe.device != x.device:
|
59 |
-
self.pe = self.pe.to(dtype=x.dtype, device=x.device)
|
60 |
-
return
|
61 |
-
pe = torch.zeros(x.size(1), self.embedding_dim)
|
62 |
-
if self.reverse:
|
63 |
-
position = torch.arange(x.size(1) - 1, -1, -1.0, dtype=torch.float32).unsqueeze(1)
|
64 |
-
else:
|
65 |
-
position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
|
66 |
-
div_term = torch.exp(
|
67 |
-
torch.arange(0, self.embedding_dim, 2, dtype=torch.float32) * -(math.log(10000.0) / self.embedding_dim)
|
68 |
-
)
|
69 |
-
pe[:, 0::2] = torch.sin(position * div_term)
|
70 |
-
pe[:, 1::2] = torch.cos(position * div_term)
|
71 |
-
pe = pe.unsqueeze(0)
|
72 |
-
self.pe = pe.to(device=x.device, dtype=x.dtype).detach()
|
73 |
-
|
74 |
-
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
75 |
-
self.extend_pe(x)
|
76 |
-
output = x.unsqueeze(-1) if x.ndim == 2 else x
|
77 |
-
output = output * self.x_scale + self.alpha * self.pe[:, : x.size(1)]
|
78 |
-
return self.dropout(output)
|
79 |
-
|
80 |
-
|
81 |
class SinePositionalEmbeddingNested(nn.Module):
|
82 |
def __init__(
|
83 |
self,
|
|
|
33 |
return x
|
34 |
|
35 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
36 |
class SinePositionalEmbeddingNested(nn.Module):
|
37 |
def __init__(
|
38 |
self,
|
AR/models/structs.py
CHANGED
@@ -5,11 +5,11 @@ Modified From https://github.com/XXXXRT666/GPT-SoVITS
|
|
5 |
from __future__ import annotations
|
6 |
|
7 |
from dataclasses import dataclass
|
8 |
-
from typing import List, Literal, Optional
|
9 |
|
10 |
import torch
|
11 |
|
12 |
-
from AR.models.t2s_model_abc import Sampler, T2SDecoderABC
|
13 |
|
14 |
Tensor = torch.Tensor
|
15 |
|
@@ -53,6 +53,7 @@ class T2SSession:
|
|
53 |
self.y_len = y_len
|
54 |
|
55 |
# Cache
|
|
|
56 |
self.sampler = Sampler(bsz, decoder.vocab_size)
|
57 |
|
58 |
# Forward args
|
@@ -66,6 +67,11 @@ class T2SSession:
|
|
66 |
self.input_pos = torch.zeros_like(self.prefill_len)
|
67 |
self.input_pos.add_(self.prefill_len)
|
68 |
|
|
|
|
|
|
|
|
|
|
|
69 |
# EOS
|
70 |
self.completed = torch.Tensor([False] * len(self.x)).bool().to(device)
|
71 |
self.y_results: List[Tensor] = [None] * len(self.x) # type: ignore
|
@@ -81,3 +87,5 @@ class T2SSession:
|
|
81 |
mask[-y_len:, -y_len:] = ~torch.triu(torch.ones(y_len, y_len, dtype=torch.bool), diagonal=1)
|
82 |
attn_mask.append(mask)
|
83 |
self.attn_mask_nested = torch.nested.nested_tensor(attn_mask)
|
|
|
|
|
|
5 |
from __future__ import annotations
|
6 |
|
7 |
from dataclasses import dataclass
|
8 |
+
from typing import List, Literal, MutableSequence, Optional
|
9 |
|
10 |
import torch
|
11 |
|
12 |
+
from AR.models.t2s_model_abc import KVCacheABC, Sampler, T2SDecoderABC
|
13 |
|
14 |
Tensor = torch.Tensor
|
15 |
|
|
|
53 |
self.y_len = y_len
|
54 |
|
55 |
# Cache
|
56 |
+
self.kv_cache: MutableSequence[KVCacheABC]
|
57 |
self.sampler = Sampler(bsz, decoder.vocab_size)
|
58 |
|
59 |
# Forward args
|
|
|
67 |
self.input_pos = torch.zeros_like(self.prefill_len)
|
68 |
self.input_pos.add_(self.prefill_len)
|
69 |
|
70 |
+
# CUDA Graph
|
71 |
+
self.graph: Optional[torch.cuda.CUDAGraph] = None
|
72 |
+
self.xy_pos_: Tensor
|
73 |
+
self.xy_dec_: Tensor
|
74 |
+
|
75 |
# EOS
|
76 |
self.completed = torch.Tensor([False] * len(self.x)).bool().to(device)
|
77 |
self.y_results: List[Tensor] = [None] * len(self.x) # type: ignore
|
|
|
87 |
mask[-y_len:, -y_len:] = ~torch.triu(torch.ones(y_len, y_len, dtype=torch.bool), diagonal=1)
|
88 |
attn_mask.append(mask)
|
89 |
self.attn_mask_nested = torch.nested.nested_tensor(attn_mask)
|
90 |
+
|
91 |
+
self.id: int = -1
|
AR/models/t2s_model_abc.py
CHANGED
@@ -5,10 +5,10 @@ Modified From https://github.com/XXXXRT666/GPT-SoVITS
|
|
5 |
from __future__ import annotations
|
6 |
|
7 |
import os
|
8 |
-
import
|
9 |
from abc import ABC, abstractmethod
|
10 |
from contextlib import nullcontext
|
11 |
-
from typing import Any, Dict, List, MutableSequence,
|
12 |
|
13 |
import torch
|
14 |
import torch._inductor.config
|
@@ -85,6 +85,10 @@ class KVCacheABC(ABC, nn.Module):
|
|
85 |
@abstractmethod
|
86 |
def prefill_kv(self, k_val: Tensor, v_val: Tensor, bs: int) -> None: ...
|
87 |
|
|
|
|
|
|
|
|
|
88 |
def forward(self):
|
89 |
raise NotImplementedError()
|
90 |
|
@@ -363,6 +367,8 @@ class T2SDecoderABC(ABC, nn.Module):
|
|
363 |
|
364 |
self.kv_class: Type[KVCacheNHD] | Type[KVCacheHND]
|
365 |
|
|
|
|
|
366 |
self._register_load_state_dict_pre_hook(self.load_hook)
|
367 |
|
368 |
def load_hook(self, state_dict, prefix, *args):
|
@@ -396,6 +402,7 @@ class T2SDecoderABC(ABC, nn.Module):
|
|
396 |
self.h.compile(fullgraph=True, mode="reduce-overhead")
|
397 |
|
398 |
def capture(self, input_pos: Tensor, x: Tensor, x_dec: Tensor, *args, **kwds) -> CUDAGraph:
|
|
|
399 |
s = torch.cuda.Stream()
|
400 |
s.wait_stream(torch.cuda.current_stream())
|
401 |
|
@@ -419,6 +426,51 @@ class T2SDecoderABC(ABC, nn.Module):
|
|
419 |
def post_forward(self, idx: int, session: Any) -> None: ...
|
420 |
|
421 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
422 |
class TorchProfiler:
|
423 |
def __init__(self, debug: bool, log_dir: str = "./profiler") -> None:
|
424 |
self.debug = debug
|
|
|
5 |
from __future__ import annotations
|
6 |
|
7 |
import os
|
8 |
+
import random
|
9 |
from abc import ABC, abstractmethod
|
10 |
from contextlib import nullcontext
|
11 |
+
from typing import Any, Dict, List, MutableSequence, Tuple, Type
|
12 |
|
13 |
import torch
|
14 |
import torch._inductor.config
|
|
|
85 |
@abstractmethod
|
86 |
def prefill_kv(self, k_val: Tensor, v_val: Tensor, bs: int) -> None: ...
|
87 |
|
88 |
+
def sync_cache(self, kv_cache: KVCacheABC):
|
89 |
+
self.k_cache.copy_(kv_cache.k_cache)
|
90 |
+
self.v_cache.copy_(kv_cache.v_cache)
|
91 |
+
|
92 |
def forward(self):
|
93 |
raise NotImplementedError()
|
94 |
|
|
|
367 |
|
368 |
self.kv_class: Type[KVCacheNHD] | Type[KVCacheHND]
|
369 |
|
370 |
+
self.GraphCache: CUDAGraphCacheABC | None
|
371 |
+
|
372 |
self._register_load_state_dict_pre_hook(self.load_hook)
|
373 |
|
374 |
def load_hook(self, state_dict, prefix, *args):
|
|
|
402 |
self.h.compile(fullgraph=True, mode="reduce-overhead")
|
403 |
|
404 |
def capture(self, input_pos: Tensor, x: Tensor, x_dec: Tensor, *args, **kwds) -> CUDAGraph:
|
405 |
+
assert torch.cuda.is_available()
|
406 |
s = torch.cuda.Stream()
|
407 |
s.wait_stream(torch.cuda.current_stream())
|
408 |
|
|
|
426 |
def post_forward(self, idx: int, session: Any) -> None: ...
|
427 |
|
428 |
|
429 |
+
class CUDAGraphCacheABC(ABC):
|
430 |
+
def __init__(
|
431 |
+
self,
|
432 |
+
decoder: T2SDecoderABC,
|
433 |
+
device: torch.device = torch.device("cpu"),
|
434 |
+
dtype: torch.dtype = torch.float32,
|
435 |
+
) -> None:
|
436 |
+
assert torch.cuda.is_available()
|
437 |
+
|
438 |
+
self.assigned: bool = False
|
439 |
+
|
440 |
+
self.decoder: T2SDecoderABC = decoder
|
441 |
+
self.kv_cache: MutableSequence[KVCacheABC] = decoder.init_cache(1)
|
442 |
+
self.xy_pos = torch.rand((1, 1, decoder.embedding_dim), device=device).to(dtype)
|
443 |
+
self.xy_dec = torch.rand((1, 1, decoder.embedding_dim), device=device).to(dtype)
|
444 |
+
self.input_pos = torch.tensor([10]).int().cuda()
|
445 |
+
self.graph: torch.cuda.CUDAGraph | None = None
|
446 |
+
|
447 |
+
self.id: int = random.randint(1, 2**32 - 1)
|
448 |
+
|
449 |
+
def assign_graph(self, session: Any):
|
450 |
+
if self.graph is None:
|
451 |
+
args, kwds = self.decoder.pre_forward(session)
|
452 |
+
graph = self.decoder.capture(self.input_pos, self.xy_pos, self.xy_dec, *args, **kwds)
|
453 |
+
self.graph = graph
|
454 |
+
|
455 |
+
if self.assigned is False:
|
456 |
+
self.get_cache_graph(session)
|
457 |
+
session.id = self.id
|
458 |
+
self.assigned = True
|
459 |
+
else:
|
460 |
+
self.capture_new_graph(session)
|
461 |
+
|
462 |
+
@abstractmethod
|
463 |
+
def release_graph(self, session: Any): ...
|
464 |
+
|
465 |
+
@abstractmethod
|
466 |
+
def get_cache_graph(self, session: Any):
|
467 |
+
pass
|
468 |
+
|
469 |
+
@abstractmethod
|
470 |
+
def capture_new_graph(self, session: Any):
|
471 |
+
pass
|
472 |
+
|
473 |
+
|
474 |
class TorchProfiler:
|
475 |
def __init__(self, debug: bool, log_dir: str = "./profiler") -> None:
|
476 |
self.debug = debug
|
AR/models/t2s_model_flash_attn.py
CHANGED
@@ -2,13 +2,13 @@
|
|
2 |
Modified From https://github.com/XXXXRT666/GPT-SoVITS
|
3 |
"""
|
4 |
|
|
|
5 |
import os
|
6 |
import time
|
7 |
import traceback
|
8 |
-
from typing import Dict, List,
|
9 |
|
10 |
import flash_attn # type: ignore
|
11 |
-
import gradio as gr
|
12 |
import torch
|
13 |
import torch.nn as nn
|
14 |
from tqdm import tqdm
|
@@ -20,6 +20,7 @@ from AR.models.embedding import TokenEmbedding
|
|
20 |
from AR.models.structs import T2SRequest, T2SResult, T2SSession
|
21 |
from AR.models.t2s_model_abc import (
|
22 |
AttentionABC,
|
|
|
23 |
FeedForward,
|
24 |
KVCacheABC,
|
25 |
KVCacheNHD,
|
@@ -121,6 +122,7 @@ class T2SDecoder(T2SDecoderABC):
|
|
121 |
max_batch_size=10,
|
122 |
**kwds,
|
123 |
) -> None:
|
|
|
124 |
super().__init__()
|
125 |
|
126 |
hidden_dim = config["model"]["hidden_dim"]
|
@@ -205,6 +207,42 @@ class T2SDecoder(T2SDecoderABC):
|
|
205 |
return list(), dict()
|
206 |
|
207 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
208 |
class CUDAGraphRunner:
|
209 |
def __init__(
|
210 |
self,
|
@@ -212,70 +250,51 @@ class CUDAGraphRunner:
|
|
212 |
device: torch.device = torch.device("cpu"),
|
213 |
dtype: torch.dtype = torch.float32,
|
214 |
) -> None:
|
215 |
-
assert device.type
|
216 |
-
assert dtype in {torch.float16, torch.bfloat16, torch.float32}
|
217 |
self.device = device
|
218 |
self.dtype = dtype
|
219 |
|
220 |
-
self.decoder_path: os.PathLike
|
221 |
self.decoder_model: T2SDecoderABC = decoder_model.to(self.device, self.dtype)
|
222 |
|
223 |
-
self.
|
224 |
-
self.xy_pos_ = torch.rand((1, 1, decoder_model.embedding_dim), device=device).to(dtype)
|
225 |
-
self.xy_dec_ = torch.rand((1, 1, decoder_model.embedding_dim), device=device).to(dtype)
|
226 |
-
self.kv_cache = decoder_model.init_cache(1)
|
227 |
-
self.input_pos = torch.tensor([10]).int().cuda()
|
228 |
|
229 |
def _handle_request(self, request: T2SRequest):
|
230 |
with self.device:
|
231 |
-
for i in self.kv_cache:
|
232 |
-
i.empty()
|
233 |
-
|
234 |
decoder = self.decoder_model
|
235 |
session = T2SSession(decoder, request, device=self.device, dtype=self.dtype)
|
236 |
-
self.input_pos.copy_(session.input_pos)
|
237 |
|
238 |
t1 = 0.0
|
239 |
infer_speed = 0.0
|
240 |
-
|
241 |
-
bsz = y.size(0)
|
242 |
torch_profiler = TorchProfiler(request.debug)
|
243 |
with torch_profiler.profiler():
|
244 |
for idx in tqdm(range(1500)):
|
245 |
if idx == 0:
|
246 |
-
|
|
|
247 |
xy_dec = torch.stack([t[[-1]] for t in xy_dec.unbind()])
|
248 |
else:
|
249 |
-
if request.use_cuda_graph and
|
250 |
-
self.
|
251 |
-
args, kwds = decoder.pre_forward(session)
|
252 |
-
self.graph = decoder.capture(
|
253 |
-
self.input_pos,
|
254 |
-
self.xy_pos_,
|
255 |
-
self.xy_dec_,
|
256 |
-
kv_caches=self.kv_cache,
|
257 |
-
*args,
|
258 |
-
**kwds,
|
259 |
-
)
|
260 |
|
261 |
with torch_profiler.record("AR"):
|
262 |
-
if
|
263 |
-
|
264 |
-
|
265 |
-
xy_dec =
|
266 |
else:
|
267 |
args, kwds = decoder.pre_forward(session)
|
268 |
xy_dec = decoder.h.forward(
|
269 |
-
|
270 |
session.xy_pos,
|
271 |
-
|
272 |
*args,
|
273 |
**kwds,
|
274 |
)
|
275 |
|
276 |
decoder.post_forward(idx, session)
|
277 |
logits = decoder.ar_predict_layer(xy_dec[:, -1])
|
278 |
-
|
279 |
|
280 |
if idx == 0:
|
281 |
logits[:, -1] = float("-inf")
|
@@ -322,7 +341,7 @@ class CUDAGraphRunner:
|
|
322 |
request.early_stop_num != -1
|
323 |
and (session.y.size(1) - session.y_len) > request.early_stop_num
|
324 |
) or idx == 1499:
|
325 |
-
for i in range(bsz):
|
326 |
if not session.completed[i].item():
|
327 |
session.y_results[i] = session.y[i, session.y_len :]
|
328 |
session.completed[i] = True
|
@@ -330,7 +349,7 @@ class CUDAGraphRunner:
|
|
330 |
|
331 |
with torch_profiler.record("NextPos"):
|
332 |
y_emb = decoder.ar_audio_embedding(session.y[:, -1:])
|
333 |
-
session.xy_pos = decoder.ar_audio_position.forward(
|
334 |
|
335 |
if idx == 2:
|
336 |
torch_profiler.start()
|
@@ -359,8 +378,11 @@ class CUDAGraphRunner:
|
|
359 |
torch.xpu.empty_cache()
|
360 |
case "mtia":
|
361 |
torch.mtia.empty_cache()
|
|
|
|
|
362 |
|
363 |
torch_profiler.end()
|
|
|
364 |
return session.y_results[: request.valid_length], infer_speed
|
365 |
|
366 |
def generate(self, request: T2SRequest):
|
|
|
2 |
Modified From https://github.com/XXXXRT666/GPT-SoVITS
|
3 |
"""
|
4 |
|
5 |
+
import gc
|
6 |
import os
|
7 |
import time
|
8 |
import traceback
|
9 |
+
from typing import Dict, List, Tuple
|
10 |
|
11 |
import flash_attn # type: ignore
|
|
|
12 |
import torch
|
13 |
import torch.nn as nn
|
14 |
from tqdm import tqdm
|
|
|
20 |
from AR.models.structs import T2SRequest, T2SResult, T2SSession
|
21 |
from AR.models.t2s_model_abc import (
|
22 |
AttentionABC,
|
23 |
+
CUDAGraphCacheABC,
|
24 |
FeedForward,
|
25 |
KVCacheABC,
|
26 |
KVCacheNHD,
|
|
|
122 |
max_batch_size=10,
|
123 |
**kwds,
|
124 |
) -> None:
|
125 |
+
assert torch.cuda.is_available()
|
126 |
super().__init__()
|
127 |
|
128 |
hidden_dim = config["model"]["hidden_dim"]
|
|
|
207 |
return list(), dict()
|
208 |
|
209 |
|
210 |
+
class CUDAGraphCache(CUDAGraphCacheABC):
|
211 |
+
def __init__(
|
212 |
+
self,
|
213 |
+
decoder: T2SDecoderABC,
|
214 |
+
device: torch.device = torch.device("cpu"),
|
215 |
+
dtype: torch.dtype = torch.float32,
|
216 |
+
) -> None:
|
217 |
+
super().__init__(decoder, device, dtype)
|
218 |
+
|
219 |
+
def release_graph(self, session: T2SSession):
|
220 |
+
if session.id != self.id:
|
221 |
+
self.assigned = False
|
222 |
+
else:
|
223 |
+
del session.graph, session.xy_pos_, session.xy_dec_, session.input_pos, session.kv_cache
|
224 |
+
|
225 |
+
def get_cache_graph(self, session: T2SSession):
|
226 |
+
assert self.graph
|
227 |
+
session.graph = self.graph
|
228 |
+
|
229 |
+
session.xy_pos_ = self.xy_pos
|
230 |
+
session.xy_dec_ = self.xy_dec
|
231 |
+
session.input_pos = self.input_pos.copy_(session.input_pos)
|
232 |
+
|
233 |
+
for cache, cache_ in zip(self.kv_cache, session.kv_cache):
|
234 |
+
cache.sync_cache(cache_)
|
235 |
+
|
236 |
+
def capture_new_graph(self, session: T2SSession):
|
237 |
+
session.xy_pos_ = self.xy_pos.clone()
|
238 |
+
session.xy_dec_ = self.xy_dec.clone()
|
239 |
+
session.input_pos = self.input_pos.clone().copy_(session.input_pos)
|
240 |
+
|
241 |
+
args, kwds = self.decoder.pre_forward(session)
|
242 |
+
graph = self.decoder.capture(self.input_pos, self.xy_pos, self.xy_dec, *args, **kwds)
|
243 |
+
session.graph = graph
|
244 |
+
|
245 |
+
|
246 |
class CUDAGraphRunner:
|
247 |
def __init__(
|
248 |
self,
|
|
|
250 |
device: torch.device = torch.device("cpu"),
|
251 |
dtype: torch.dtype = torch.float32,
|
252 |
) -> None:
|
253 |
+
assert device.type == "cuda"
|
|
|
254 |
self.device = device
|
255 |
self.dtype = dtype
|
256 |
|
|
|
257 |
self.decoder_model: T2SDecoderABC = decoder_model.to(self.device, self.dtype)
|
258 |
|
259 |
+
self.graphcache = CUDAGraphCache(decoder_model, device, dtype)
|
|
|
|
|
|
|
|
|
260 |
|
261 |
def _handle_request(self, request: T2SRequest):
|
262 |
with self.device:
|
|
|
|
|
|
|
263 |
decoder = self.decoder_model
|
264 |
session = T2SSession(decoder, request, device=self.device, dtype=self.dtype)
|
|
|
265 |
|
266 |
t1 = 0.0
|
267 |
infer_speed = 0.0
|
268 |
+
|
|
|
269 |
torch_profiler = TorchProfiler(request.debug)
|
270 |
with torch_profiler.profiler():
|
271 |
for idx in tqdm(range(1500)):
|
272 |
if idx == 0:
|
273 |
+
session.kv_cache = decoder.init_cache(session.bsz)
|
274 |
+
xy_dec = decoder.h.prefill(session.xy_pos, session.attn_mask_nested, session.kv_cache)
|
275 |
xy_dec = torch.stack([t[[-1]] for t in xy_dec.unbind()])
|
276 |
else:
|
277 |
+
if request.use_cuda_graph and session.graph is None and torch.cuda.is_available():
|
278 |
+
self.graphcache.assign_graph(session)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
279 |
|
280 |
with torch_profiler.record("AR"):
|
281 |
+
if session.graph:
|
282 |
+
session.xy_pos_.copy_(session.xy_pos)
|
283 |
+
session.graph.replay()
|
284 |
+
xy_dec = session.xy_dec_.clone()
|
285 |
else:
|
286 |
args, kwds = decoder.pre_forward(session)
|
287 |
xy_dec = decoder.h.forward(
|
288 |
+
session.input_pos,
|
289 |
session.xy_pos,
|
290 |
+
session.kv_cache,
|
291 |
*args,
|
292 |
**kwds,
|
293 |
)
|
294 |
|
295 |
decoder.post_forward(idx, session)
|
296 |
logits = decoder.ar_predict_layer(xy_dec[:, -1])
|
297 |
+
session.input_pos.add_(1)
|
298 |
|
299 |
if idx == 0:
|
300 |
logits[:, -1] = float("-inf")
|
|
|
341 |
request.early_stop_num != -1
|
342 |
and (session.y.size(1) - session.y_len) > request.early_stop_num
|
343 |
) or idx == 1499:
|
344 |
+
for i in range(session.bsz):
|
345 |
if not session.completed[i].item():
|
346 |
session.y_results[i] = session.y[i, session.y_len :]
|
347 |
session.completed[i] = True
|
|
|
349 |
|
350 |
with torch_profiler.record("NextPos"):
|
351 |
y_emb = decoder.ar_audio_embedding(session.y[:, -1:])
|
352 |
+
session.xy_pos = decoder.ar_audio_position.forward(session.input_pos - session.x_lens, y_emb)
|
353 |
|
354 |
if idx == 2:
|
355 |
torch_profiler.start()
|
|
|
378 |
torch.xpu.empty_cache()
|
379 |
case "mtia":
|
380 |
torch.mtia.empty_cache()
|
381 |
+
case "cpu":
|
382 |
+
gc.collect()
|
383 |
|
384 |
torch_profiler.end()
|
385 |
+
self.graphcache.release_graph(session)
|
386 |
return session.y_results[: request.valid_length], infer_speed
|
387 |
|
388 |
def generate(self, request: T2SRequest):
|
inference_webui.py
CHANGED
@@ -1,7 +1,47 @@
|
|
|
|
1 |
import os
|
|
|
|
|
|
|
2 |
|
3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
4 |
from huggingface_hub import snapshot_download
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
5 |
|
6 |
snapshot_download(
|
7 |
repo_id="lj1995/GPT-SoVITS",
|
@@ -27,75 +67,20 @@ snapshot_download(
|
|
27 |
allow_patterns="v2Pro/s2Gv2ProPlus.pth",
|
28 |
local_dir="pretrained_models",
|
29 |
)
|
30 |
-
import logging
|
31 |
-
import traceback
|
32 |
-
|
33 |
-
logging.getLogger("markdown_it").setLevel(logging.ERROR)
|
34 |
-
logging.getLogger("urllib3").setLevel(logging.ERROR)
|
35 |
-
logging.getLogger("httpcore").setLevel(logging.ERROR)
|
36 |
-
logging.getLogger("httpx").setLevel(logging.ERROR)
|
37 |
-
logging.getLogger("asyncio").setLevel(logging.ERROR)
|
38 |
-
logging.getLogger("charset_normalizer").setLevel(logging.ERROR)
|
39 |
-
logging.getLogger("torchaudio._extension").setLevel(logging.ERROR)
|
40 |
-
logging.getLogger("multipart.multipart").setLevel(logging.ERROR)
|
41 |
-
logging.getLogger("python_multipart.multipart").setLevel(logging.ERROR)
|
42 |
-
logging.getLogger("split_lang.split.splitter").setLevel(logging.ERROR)
|
43 |
-
|
44 |
-
import nltk
|
45 |
-
import torchaudio
|
46 |
-
|
47 |
-
from text.LangSegmenter import LangSegmenter
|
48 |
-
|
49 |
-
nltk.download("averaged_perceptron_tagger_eng")
|
50 |
-
import json
|
51 |
-
import os
|
52 |
-
import pdb
|
53 |
-
import re
|
54 |
-
import sys
|
55 |
-
import threading
|
56 |
-
|
57 |
-
import LangSegment
|
58 |
-
import spaces
|
59 |
-
import torch
|
60 |
-
|
61 |
-
lock = threading.Lock()
|
62 |
|
63 |
version = "v2" # os.environ.get("version","v2")
|
64 |
cnhubert_base_path = os.environ.get("cnhubert_base_path", "pretrained_models/chinese-hubert-base")
|
65 |
bert_path = os.environ.get("bert_path", "pretrained_models/chinese-roberta-wwm-ext-large")
|
66 |
-
|
67 |
-
punctuation = set(["!", "?", "…", ",", ".", "-", " "])
|
68 |
-
import gradio as gr
|
69 |
-
import gradio.themes as themes
|
70 |
-
import librosa
|
71 |
-
import numpy as np
|
72 |
-
from gradio.themes.utils import fonts
|
73 |
-
from transformers import AutoModelForMaskedLM, AutoTokenizer
|
74 |
-
|
75 |
-
from feature_extractor import cnhubert
|
76 |
-
|
77 |
cnhubert.cnhubert_base_path = cnhubert_base_path
|
78 |
|
79 |
-
|
80 |
|
81 |
-
from AR.models.structs import T2SRequest
|
82 |
-
from AR.models.t2s_model_flash_attn import CUDAGraphRunner
|
83 |
-
from module.mel_processing import spectrogram_torch
|
84 |
-
from module.models import SynthesizerTrn
|
85 |
-
from text import cleaned_text_to_sequence
|
86 |
-
from text.cleaner import clean_text
|
87 |
-
from tools.i18n.i18n import I18nAuto, scan_language_list
|
88 |
-
from tools.my_utils import load_audio
|
89 |
|
90 |
-
# language=os.environ.get("language","Auto")
|
91 |
-
# language=sys.argv[-1] if sys.argv[-1] in scan_language_list() else language
|
92 |
i18n = I18nAuto(language="Auto")
|
93 |
|
94 |
-
# os.environ['PYTORCH_ENABLE_MPS_FALLBACK'] = '1' # 确保直接启动推理UI时也能够设置。
|
95 |
-
|
96 |
if torch.cuda.is_available():
|
97 |
device = "cuda"
|
98 |
-
is_half = True
|
99 |
else:
|
100 |
device = "cpu"
|
101 |
is_half = False
|
@@ -125,7 +110,7 @@ dict_language = dict_language_v1 if version == "v1" else dict_language_v2
|
|
125 |
|
126 |
tokenizer = AutoTokenizer.from_pretrained(bert_path)
|
127 |
bert_model = AutoModelForMaskedLM.from_pretrained(bert_path)
|
128 |
-
if is_half
|
129 |
bert_model = bert_model.half().to(device)
|
130 |
else:
|
131 |
bert_model = bert_model.to(device)
|
@@ -176,7 +161,7 @@ class DictToAttrRecursive(dict):
|
|
176 |
|
177 |
|
178 |
ssl_model = cnhubert.get_model()
|
179 |
-
if is_half
|
180 |
ssl_model = ssl_model.half().to(device)
|
181 |
else:
|
182 |
ssl_model = ssl_model.to(device)
|
@@ -248,7 +233,7 @@ def change_gpt_weights(gpt_path):
|
|
248 |
|
249 |
|
250 |
change_gpt_weights("pretrained_models/s1v3.ckpt")
|
251 |
-
|
252 |
|
253 |
sv_cn_model = SV(device, is_half)
|
254 |
|
@@ -288,7 +273,7 @@ def get_spepc(hps, filename, dtype, device, is_v2pro=False):
|
|
288 |
center=False,
|
289 |
)
|
290 |
spec = spec.to(dtype)
|
291 |
-
if is_v2pro
|
292 |
audio = resample(audio, sr1, 16000, device).to(dtype)
|
293 |
return spec, audio
|
294 |
|
@@ -300,7 +285,7 @@ def clean_text_inf(text, language, version):
|
|
300 |
return phones, word2ph, norm_text
|
301 |
|
302 |
|
303 |
-
dtype = torch.float16 if is_half
|
304 |
|
305 |
|
306 |
def get_bert_inf(phones, word2ph, norm_text, language):
|
@@ -310,27 +295,13 @@ def get_bert_inf(phones, word2ph, norm_text, language):
|
|
310 |
else:
|
311 |
bert = torch.zeros(
|
312 |
(1024, len(phones)),
|
313 |
-
dtype=torch.float16 if is_half
|
314 |
).to(device)
|
315 |
|
316 |
return bert
|
317 |
|
318 |
|
319 |
-
splits = {
|
320 |
-
",",
|
321 |
-
"。",
|
322 |
-
"?",
|
323 |
-
"!",
|
324 |
-
",",
|
325 |
-
".",
|
326 |
-
"?",
|
327 |
-
"!",
|
328 |
-
"~",
|
329 |
-
":",
|
330 |
-
":",
|
331 |
-
"—",
|
332 |
-
"…",
|
333 |
-
}
|
334 |
|
335 |
|
336 |
def get_first(text):
|
@@ -339,9 +310,6 @@ def get_first(text):
|
|
339 |
return text
|
340 |
|
341 |
|
342 |
-
from text import chinese
|
343 |
-
|
344 |
-
|
345 |
def get_phones_and_bert(text, language, version, final=False):
|
346 |
if language in {"en", "all_zh", "all_ja", "all_ko", "all_yue"}:
|
347 |
formattext = text
|
@@ -363,7 +331,7 @@ def get_phones_and_bert(text, language, version, final=False):
|
|
363 |
phones, word2ph, norm_text = clean_text_inf(formattext, language, version)
|
364 |
bert = torch.zeros(
|
365 |
(1024, len(phones)),
|
366 |
-
dtype=torch.float16 if is_half
|
367 |
).to(device)
|
368 |
elif language in {"zh", "ja", "ko", "yue", "auto", "auto_yue"}:
|
369 |
textlist = []
|
@@ -475,7 +443,7 @@ def get_tts_wav(
|
|
475 |
print(i18n("实际输入的目标文本:"), text)
|
476 |
zero_wav = np.zeros(
|
477 |
int(hps.data.sampling_rate * 0.3),
|
478 |
-
dtype=np.float16 if is_half
|
479 |
)
|
480 |
if not ref_free:
|
481 |
with torch.no_grad():
|
@@ -485,7 +453,7 @@ def get_tts_wav(
|
|
485 |
raise OSError(i18n("参考音频在3~10秒范围外,请更换!"))
|
486 |
wav16k = torch.from_numpy(wav16k)
|
487 |
zero_wav_torch = torch.from_numpy(zero_wav)
|
488 |
-
if is_half
|
489 |
wav16k = wav16k.half().to(device)
|
490 |
zero_wav_torch = zero_wav_torch.half().to(device)
|
491 |
else:
|
@@ -544,10 +512,10 @@ def get_tts_wav(
|
|
544 |
t2 = ttime()
|
545 |
# cache_key="%s-%s-%s-%s-%s-%s-%s-%s"%(ref_wav_path,prompt_text,prompt_language,text,text_language,top_k,top_p,temperature)
|
546 |
# print(cache.keys(),if_freeze)
|
547 |
-
if i_text in cache and if_freeze
|
548 |
pred_semantic = cache[i_text]
|
549 |
else:
|
550 |
-
with torch.no_grad()
|
551 |
t2s_request = T2SRequest(
|
552 |
[all_phoneme_ids.squeeze(0)],
|
553 |
all_phoneme_len,
|
@@ -564,9 +532,8 @@ def get_tts_wav(
|
|
564 |
t2s_result = t2s_model.generate(t2s_request)
|
565 |
|
566 |
if t2s_result.exception is not None:
|
567 |
-
print(t2s_result.exception)
|
568 |
print(t2s_result.traceback)
|
569 |
-
raise
|
570 |
|
571 |
infer_speed.append(t2s_result.infer_speed)
|
572 |
pred_semantic = t2s_result.result
|
@@ -608,8 +575,8 @@ def get_tts_wav(
|
|
608 |
t.extend([t2 - t1, t3 - t2, t4 - t3])
|
609 |
t1 = ttime()
|
610 |
print("%.3f\t%.3f\t%.3f\t%.3f" % (t[0], sum(t[1::3]), sum(t[2::3]), sum(t[3::3])))
|
611 |
-
gr.Info(f"
|
612 |
-
gr.Info("%.3f\t%.3f\t%.3f\t%.3f" % (t[0], sum(t[1::3]), sum(t[2::3]), sum(t[3::3])),
|
613 |
yield hps.data.sampling_rate, (np.concatenate(audio_opt, 0) * 32768).astype(np.int16)
|
614 |
|
615 |
|
@@ -713,7 +680,7 @@ def cut5(inp):
|
|
713 |
|
714 |
def custom_sort_key(s):
|
715 |
# 使用正则表达式提取字符串中的数字部分和非数字部分
|
716 |
-
parts = re.split("(\d+)", s)
|
717 |
# 将数字部分转换为整数,非数字部分保持不变
|
718 |
parts = [int(part) if part.isdigit() else part for part in parts]
|
719 |
return parts
|
|
|
1 |
+
import logging
|
2 |
import os
|
3 |
+
import re
|
4 |
+
import traceback
|
5 |
+
from time import time as ttime
|
6 |
|
7 |
+
import gradio as gr
|
8 |
+
import gradio.themes as themes
|
9 |
+
import librosa
|
10 |
+
import nltk
|
11 |
+
import numpy as np
|
12 |
+
import spaces
|
13 |
+
import torch
|
14 |
+
import torchaudio
|
15 |
+
from gradio.themes.utils import fonts
|
16 |
from huggingface_hub import snapshot_download
|
17 |
+
from transformers.models.auto.modeling_auto import AutoModelForMaskedLM
|
18 |
+
from transformers.models.auto.tokenization_auto import AutoTokenizer
|
19 |
+
|
20 |
+
from AR.models.structs import T2SRequest
|
21 |
+
from AR.models.t2s_model_flash_attn import CUDAGraphRunner
|
22 |
+
from feature_extractor import cnhubert
|
23 |
+
from module.mel_processing import spectrogram_torch
|
24 |
+
from module.models import SynthesizerTrn
|
25 |
+
from sv import SV
|
26 |
+
from text import chinese, cleaned_text_to_sequence
|
27 |
+
from text.cleaner import clean_text
|
28 |
+
from text.LangSegmenter import LangSegmenter
|
29 |
+
from tools.i18n.i18n import I18nAuto
|
30 |
+
|
31 |
+
logging.getLogger("markdown_it").setLevel(logging.ERROR)
|
32 |
+
logging.getLogger("urllib3").setLevel(logging.ERROR)
|
33 |
+
logging.getLogger("httpcore").setLevel(logging.ERROR)
|
34 |
+
logging.getLogger("httpx").setLevel(logging.ERROR)
|
35 |
+
logging.getLogger("asyncio").setLevel(logging.ERROR)
|
36 |
+
logging.getLogger("charset_normalizer").setLevel(logging.ERROR)
|
37 |
+
logging.getLogger("torchaudio._extension").setLevel(logging.ERROR)
|
38 |
+
logging.getLogger("multipart.multipart").setLevel(logging.ERROR)
|
39 |
+
logging.getLogger("python_multipart.multipart").setLevel(logging.ERROR)
|
40 |
+
logging.getLogger("split_lang.split.splitter").setLevel(logging.ERROR)
|
41 |
+
|
42 |
+
os.makedirs("pretrained_models", exist_ok=True)
|
43 |
+
|
44 |
+
nltk.download("averaged_perceptron_tagger_eng")
|
45 |
|
46 |
snapshot_download(
|
47 |
repo_id="lj1995/GPT-SoVITS",
|
|
|
67 |
allow_patterns="v2Pro/s2Gv2ProPlus.pth",
|
68 |
local_dir="pretrained_models",
|
69 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
70 |
|
71 |
version = "v2" # os.environ.get("version","v2")
|
72 |
cnhubert_base_path = os.environ.get("cnhubert_base_path", "pretrained_models/chinese-hubert-base")
|
73 |
bert_path = os.environ.get("bert_path", "pretrained_models/chinese-roberta-wwm-ext-large")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
74 |
cnhubert.cnhubert_base_path = cnhubert_base_path
|
75 |
|
76 |
+
punctuation = set(["!", "?", "…", ",", ".", "-", " "])
|
77 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
78 |
|
|
|
|
|
79 |
i18n = I18nAuto(language="Auto")
|
80 |
|
|
|
|
|
81 |
if torch.cuda.is_available():
|
82 |
device = "cuda"
|
83 |
+
is_half = True
|
84 |
else:
|
85 |
device = "cpu"
|
86 |
is_half = False
|
|
|
110 |
|
111 |
tokenizer = AutoTokenizer.from_pretrained(bert_path)
|
112 |
bert_model = AutoModelForMaskedLM.from_pretrained(bert_path)
|
113 |
+
if is_half is True:
|
114 |
bert_model = bert_model.half().to(device)
|
115 |
else:
|
116 |
bert_model = bert_model.to(device)
|
|
|
161 |
|
162 |
|
163 |
ssl_model = cnhubert.get_model()
|
164 |
+
if is_half is True:
|
165 |
ssl_model = ssl_model.half().to(device)
|
166 |
else:
|
167 |
ssl_model = ssl_model.to(device)
|
|
|
233 |
|
234 |
|
235 |
change_gpt_weights("pretrained_models/s1v3.ckpt")
|
236 |
+
|
237 |
|
238 |
sv_cn_model = SV(device, is_half)
|
239 |
|
|
|
273 |
center=False,
|
274 |
)
|
275 |
spec = spec.to(dtype)
|
276 |
+
if is_v2pro is True:
|
277 |
audio = resample(audio, sr1, 16000, device).to(dtype)
|
278 |
return spec, audio
|
279 |
|
|
|
285 |
return phones, word2ph, norm_text
|
286 |
|
287 |
|
288 |
+
dtype = torch.float16 if is_half is True else torch.float32
|
289 |
|
290 |
|
291 |
def get_bert_inf(phones, word2ph, norm_text, language):
|
|
|
295 |
else:
|
296 |
bert = torch.zeros(
|
297 |
(1024, len(phones)),
|
298 |
+
dtype=torch.float16 if is_half is True else torch.float32,
|
299 |
).to(device)
|
300 |
|
301 |
return bert
|
302 |
|
303 |
|
304 |
+
splits = {",", "。", "?", "!", ",", ".", "?", "!", "~", ":", ":", "—", "…"}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
305 |
|
306 |
|
307 |
def get_first(text):
|
|
|
310 |
return text
|
311 |
|
312 |
|
|
|
|
|
|
|
313 |
def get_phones_and_bert(text, language, version, final=False):
|
314 |
if language in {"en", "all_zh", "all_ja", "all_ko", "all_yue"}:
|
315 |
formattext = text
|
|
|
331 |
phones, word2ph, norm_text = clean_text_inf(formattext, language, version)
|
332 |
bert = torch.zeros(
|
333 |
(1024, len(phones)),
|
334 |
+
dtype=torch.float16 if is_half is True else torch.float32,
|
335 |
).to(device)
|
336 |
elif language in {"zh", "ja", "ko", "yue", "auto", "auto_yue"}:
|
337 |
textlist = []
|
|
|
443 |
print(i18n("实际输入的目标文本:"), text)
|
444 |
zero_wav = np.zeros(
|
445 |
int(hps.data.sampling_rate * 0.3),
|
446 |
+
dtype=np.float16 if is_half is True else np.float32,
|
447 |
)
|
448 |
if not ref_free:
|
449 |
with torch.no_grad():
|
|
|
453 |
raise OSError(i18n("参考音频在3~10秒范围外,请更换!"))
|
454 |
wav16k = torch.from_numpy(wav16k)
|
455 |
zero_wav_torch = torch.from_numpy(zero_wav)
|
456 |
+
if is_half is True:
|
457 |
wav16k = wav16k.half().to(device)
|
458 |
zero_wav_torch = zero_wav_torch.half().to(device)
|
459 |
else:
|
|
|
512 |
t2 = ttime()
|
513 |
# cache_key="%s-%s-%s-%s-%s-%s-%s-%s"%(ref_wav_path,prompt_text,prompt_language,text,text_language,top_k,top_p,temperature)
|
514 |
# print(cache.keys(),if_freeze)
|
515 |
+
if i_text in cache and if_freeze is True:
|
516 |
pred_semantic = cache[i_text]
|
517 |
else:
|
518 |
+
with torch.no_grad():
|
519 |
t2s_request = T2SRequest(
|
520 |
[all_phoneme_ids.squeeze(0)],
|
521 |
all_phoneme_len,
|
|
|
532 |
t2s_result = t2s_model.generate(t2s_request)
|
533 |
|
534 |
if t2s_result.exception is not None:
|
|
|
535 |
print(t2s_result.traceback)
|
536 |
+
raise t2s_result.exception
|
537 |
|
538 |
infer_speed.append(t2s_result.infer_speed)
|
539 |
pred_semantic = t2s_result.result
|
|
|
575 |
t.extend([t2 - t1, t3 - t2, t4 - t3])
|
576 |
t1 = ttime()
|
577 |
print("%.3f\t%.3f\t%.3f\t%.3f" % (t[0], sum(t[1::3]), sum(t[2::3]), sum(t[3::3])))
|
578 |
+
gr.Info(f"{sum(infer_speed) / len(infer_speed):.2f} Token/s", title="Infer Speed")
|
579 |
+
gr.Info("%.3f\t%.3f\t%.3f\t%.3f" % (t[0], sum(t[1::3]), sum(t[2::3]), sum(t[3::3])), title="Time Stamps")
|
580 |
yield hps.data.sampling_rate, (np.concatenate(audio_opt, 0) * 32768).astype(np.int16)
|
581 |
|
582 |
|
|
|
680 |
|
681 |
def custom_sort_key(s):
|
682 |
# 使用正则表达式提取字符串中的数字部分和非数字部分
|
683 |
+
parts = re.split(r"(\d+)", s)
|
684 |
# 将数字部分转换为整数,非数字部分保持不变
|
685 |
parts = [int(part) if part.isdigit() else part for part in parts]
|
686 |
return parts
|