推理失败，卡Loading rmvpe model,assets/rmvpe/rmvpe.pt

ubuntu 24.04 4GPU 2080TI 22G 2680 V4 128G
2406版本
所要配置文件均为默认

几个问题：
1、训练一键启动只能完成第一步，下面的步骤需要手动启动。

2、步骤3，训练时，3-4张 GPU 报错，只能两张或单张运行
报提示：/mnt/disk2t/RVC/infer/modules/train/train.py:429: FutureWarning: torch.cuda.amp.autocast(args...) is deprecated. Please use torch.amp.autocast('cuda', args...) instead.

3、推理时
2024-11-03 20:15:01 | INFO | infer.modules.vc.modules | Get sid: boy1.pth
2024-11-03 20:15:01 | INFO | infer.modules.vc.modules | Loading: assets/weights/boy1.pth
2024-11-03 20:15:02 | INFO | infer.modules.vc.modules | Select index: logs/liu/added_IVF370_Flat_nprobe_1_boy1_v2.index
2024-11-03 20:15:14 | INFO | fairseq.tasks.hubert_pretraining | current directory is /mnt/*/RVC
2024-11-03 20:15:14 | INFO | fairseq.tasks.hubert_pretraining | HubertPretrainingTask Config {'_name': 'hubert_pretraining', 'data': 'metadata', 'fine_tuning': False, 'labels': ['km'], 'label_dir': 'label', 'label_rate': 50.0, 'sample_rate': 16000, 'normalize': False, 'enable_padding': False, 'max_keep_size': None, 'max_sample_size': 250000, 'min_sample_size': 32000, 'single_target': False, 'random_crop': True, 'pad_audio': False}
2024-11-03 20:15:14 | INFO | fairseq.models.hubert.hubert | HubertModel Config: {'_name': 'hubert', 'label_rate': 50.0, 'extractor_mode': default, 'encoder_layers': 12, 'encoder_embed_dim': 768, 'encoder_ffn_embed_dim': 3072, 'encoder_attention_heads': 12, 'activation_fn': gelu, 'layer_type': transformer, 'dropout': 0.1, 'attention_dropout': 0.1, 'activation_dropout': 0.0, 'encoder_layerdrop': 0.05, 'dropout_input': 0.1, 'dropout_features': 0.1, 'final_dim': 256, 'untie_final_proj': True, 'layer_norm_first': False, 'conv_feature_layers': '[(512,10,5)] + [(512,3,2)] * 4 + [(512,2,2)] * 2', 'conv_bias': False, 'logit_temp': 0.1, 'target_glu': False, 'feature_grad_mult': 0.1, 'mask_length': 10, 'mask_prob': 0.8, 'mask_selection': static, 'mask_other': 0.0, 'no_mask_overlap': False, 'mask_min_space': 1, 'mask_channel_length': 10, 'mask_channel_prob': 0.0, 'mask_channel_selection': static, 'mask_channel_other': 0.0, 'no_mask_channel_overlap': False, 'mask_channel_min_space': 1, 'conv_pos': 128, 'conv_pos_groups': 16, 'latent_temp': [2.0, 0.5, 0.999995], 'skip_masked': False, 'skip_nomask': False, 'checkpoint_activations': False, 'required_seq_len_multiple': 2, 'depthwise_conv_kernel_size': 31, 'attn_type': '', 'pos_enc_type': 'abs', 'fp16': False}
2024-11-03 20:15:20 | INFO | infer.modules.vc.pipeline | Loading rmvpe model,assets/rmvpe/rmvpe.pt

卡住不动

经查卡在 codedataset.py 循环8次不动了

        for t in opt_ts:
            print('跑到这了！！')
            t = t // self.window * self.window
            if if_f0 == 1:
                audio_opt.append(
                    self.vc(
                        model,
                        net_g,
                        sid,
                        audio_pad[s : t + self.t_pad2 + self.window],
                        pitch[:, s // self.window : (t + self.t_pad2) // self.window],
                        pitchf[:, s // self.window : (t + self.t_pad2) // self.window],
                        times,
                        index,
                        big_npy,
                        index_rate,
                        version,
                        protect,
                    )[self.t_pad_tgt : -self.t_pad_tgt]
                )
            else:
                audio_opt.append(
                    self.vc(
                        model,
                        net_g,
                        sid,
                        audio_pad[s : t + self.t_pad2 + self.window],
                        None,
                        None,
                        times,
                        index,
                        big_npy,
                        index_rate,
                        version,
                        protect,
                    )[self.t_pad_tgt : -self.t_pad_tgt]
                )
            s = t

Start by reproducing the inference hang on Ubuntu 24.04 and inspect infer/modules/vc/pipeline.py together with the reported codedataset.py loop. Also review train.py around line 429 for the multi-GPU training warning and startup behavior. Done means the reported inference loading stall and the described training failures have clear, reproducible outcomes and fixes.