opensource/ChatTTS
3
0
Fork 0
mirror of https://github.com/2noise/ChatTTS.git synced 2026-08-31 01:12:28 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix(gpt): compile failed

Browse Source
This commit is contained in:
源文雨
2024-07-06 00:02:55 +09:00
parent 37674fbd3a
commit 6e80df2bcc
2 changed files with 21 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ChatTTS/core.py
+9 -2
View File
@@ -259,6 +259,7 @@ class Chat:
device = select_device()
self.logger.info("use device %s", str(device))
self.device = device
self.compile = compile
if vocos_config_path:
vocos = (
@@ -548,7 +549,10 @@ class Chat:
input_ids, attention_mask, text_mask = self._text_to_token(text, gpt.device_gpt)
emb = gpt(input_ids, text_mask)
with torch.inference_mode(not self.compile):
with torch.no_grad():
emb = gpt(input_ids, text_mask)
del text_mask
if params.spk_emb is not None:
@@ -612,7 +616,10 @@ class Chat:
repetition_penalty=params.repetition_penalty,
)
emb = gpt(input_ids, text_mask)
with torch.inference_mode(not self.compile):
with torch.no_grad():
emb = gpt(input_ids, text_mask)
del text_mask
result = next(
ChatTTS/model/gpt.py
+12 -9
View File
@@ -366,6 +366,7 @@ class GPT(nn.Module):
show_tqdm=True,
ensure_non_empty=True,
stream_batch=24,
compile=False,
context=Context(),
):
@@ -435,15 +436,17 @@ class GPT(nn.Module):
model_input.to(self.device_gpt, self.gpt.dtype)
outputs: BaseModelOutputWithPast = self.gpt(
attention_mask=model_input.attention_mask,
position_ids=model_input.position_ids,
past_key_values=model_input.past_key_values,
inputs_embeds=model_input.inputs_embeds,
use_cache=model_input.use_cache,
output_attentions=return_attn,
cache_position=model_input.cache_position,
)
with torch.inference_mode(not compile):
with torch.no_grad():
outputs: BaseModelOutputWithPast = self.gpt(
attention_mask=model_input.attention_mask,
position_ids=model_input.position_ids,
past_key_values=model_input.past_key_values,
inputs_embeds=model_input.inputs_embeds,
use_cache=model_input.use_cache,
output_attentions=return_attn,
cache_position=model_input.cache_position,
)
del_all(model_input)
attentions.append(outputs.attentions)
hidden_states = outputs.last_hidden_state.to(