umbragit/Qwen3-TTS
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

fix: padding value process bug in tokenizer decode

Browse Source
This commit is contained in:
Xiong Wang
2026-02-06 12:11:31 +08:00
parent 5f8581d05e
commit 6cafe5582c
4 changed files with 6 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
pyproject.toml
View File

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
[project] [project]
name = "qwen-tts" name = "qwen-tts"
version = "0.1.0" version = "0.1.1"
description = "Qwen-TTS python package" description = "Qwen-TTS python package"
readme = "README.md" readme = "README.md"
requires-python = ">=3.9" requires-python = ">=3.9"

3
qwen_tts/core/tokenizer_12hz/modeling_qwen3_tts_tokenizer_v2.py
View File

@@ -1009,10 +1009,11 @@ class Qwen3TTSTokenizerV2Model(Qwen3TTSTokenizerV2PreTrainedModel):
""" """
return_dict = return_dict if return_dict is not None else self.config.return_dict return_dict = return_dict if return_dict is not None else self.config.return_dict
audio_lengths = (audio_codes[..., 0] > -1).sum(1) * self.decode_upsample_rate
audio_codes = torch.clamp(audio_codes, min=0)
audio_values = self.decoder.chunked_decode(audio_codes.transpose(1, 2)).squeeze(1) audio_values = self.decoder.chunked_decode(audio_codes.transpose(1, 2)).squeeze(1)
audio_lengths = (audio_codes[..., 0] > 0).sum(1) * self.decode_upsample_rate
audio_values = [a[:l] for a, l in zip(audio_values, audio_lengths)] audio_values = [a[:l] for a, l in zip(audio_values, audio_lengths)]
if not return_dict: if not return_dict:

3
qwen_tts/core/tokenizer_25hz/modeling_qwen3_tts_tokenizer_v1.py
View File

@@ -1509,12 +1509,13 @@ class Qwen3TTSTokenizerV1Model(Qwen3TTSTokenizerV1PreTrainedModel):
""" """
return_dict = return_dict if return_dict is not None else self.config.return_dict return_dict = return_dict if return_dict is not None else self.config.return_dict
audio_lengths = (audio_codes > -1).sum(1) * self.decode_upsample_rate
audio_codes = torch.clamp(audio_codes, min=0)
audio_values = self.decoder(code=audio_codes, audio_values = self.decoder(code=audio_codes,
reference_mel=ref_mels, reference_mel=ref_mels,
conditioning=xvectors) conditioning=xvectors)
audio_lengths = (audio_codes > 0).sum(1) * self.decode_upsample_rate
audio_values = [a[:l] for a, l in zip(audio_values, audio_lengths)] audio_values = [a[:l] for a, l in zip(audio_values, audio_lengths)]
if not return_dict: if not return_dict:

2
qwen_tts/inference/qwen3_tts_tokenizer.py
View File

@@ -326,7 +326,7 @@ class Qwen3TTSTokenizer:
else: else:
# List[Tensor/np] # List[Tensor/np]
audio_codes_list = [_to_tensor(c, dtype=torch.long) for c in audio_codes_list] audio_codes_list = [_to_tensor(c, dtype=torch.long) for c in audio_codes_list]
audio_codes_padded = pad_sequence(audio_codes_list, batch_first=True, padding_value=0).to(self.device) audio_codes_padded = pad_sequence(audio_codes_list, batch_first=True, padding_value=-1).to(self.device)
with torch.inference_mode(): with torch.inference_mode():
if model_type == "qwen3_tts_tokenizer_25hz": if model_type == "qwen3_tts_tokenizer_25hz":