From 64cb31a6c3ac9f16de541ad661b42154753e2bec Mon Sep 17 00:00:00 2001 From: kj863257rc Date: Thu, 18 Sep 2025 13:59:45 +0800 Subject: [PATCH] Update infer_v2.py: solve the problem of persistent cache buildup (#382) * Update infer_v2.py clear old cache * Update infer_v2.py: solve the problem of persistent cache buildup clear old cache --- indextts/infer_v2.py | 9 +++++++++ 1 file changed, 9 insertions(+) diff --git a/indextts/infer_v2.py b/indextts/infer_v2.py index b39090b..d3aaa7d 100644 --- a/indextts/infer_v2.py +++ b/indextts/infer_v2.py @@ -370,6 +370,12 @@ class IndexTTS2: # 如果参考音频改变了，才需要重新生成, 提升速度 if self.cache_spk_cond is None or self.cache_spk_audio_prompt != spk_audio_prompt: + if self.cache_spk_cond is not None: + self.cache_spk_cond = None + self.cache_s2mel_style = None + self.cache_s2mel_prompt = None + self.cache_mel = None + torch.cuda.empty_cache() audio,sr = self._load_and_cut_audio(spk_audio_prompt,15,verbose) audio_22k = torchaudio.transforms.Resample(sr, 22050)(audio) audio_16k = torchaudio.transforms.Resample(sr, 16000)(audio) @@ -421,6 +427,9 @@ class IndexTTS2: emovec_mat = emovec_mat.unsqueeze(0) if self.cache_emo_cond is None or self.cache_emo_audio_prompt != emo_audio_prompt: + if self.cache_emo_cond is not None: + self.cache_emo_cond = None + torch.cuda.empty_cache() emo_audio, _ = self._load_and_cut_audio(emo_audio_prompt,15,verbose,sr=16000) emo_inputs = self.extract_features(emo_audio, sampling_rate=16000, return_tensors="pt") emo_input_features = emo_inputs["input_features"]