| 289 | return self.audio_tokenizer.data_from_latent(token_data)[..., -(token_data.shape[1]*2000):] |
| 290 | |
| 291 | def init_cache(self, bsize, device, dtype, length:int=None): |
| 292 | cache_shape = [self.c.stack_config.layers, length or self.c.stack_config.seq_len, 2, self.kv_heads, self.head_dim] |
| 293 | self.cache = T.full((bsize, *cache_shape), CACHE_FILL_VALUE, device=device, dtype=dtype).transpose(0, 1) |
| 294 | self.resynthesizer.init_cache(bsize, device, dtype, length) |
| 295 | self.use_audio_cache = True |
| 296 | |
| 297 | def deinit_cache(self): |
| 298 | self.cache = [None] * len(self.layers) |