### Describe the bug

[Karliz24]

Describe the bug

When fine-tuning XTTS_v2 on LJSpeech following the script (recipes/ljspeech/xtts_v2/train_gpt_xtts.py), the loss is always NaN! So terrible.
I try to reduce the learning rate, change the batch size, change the DDP config, at last it doesn't work!

To Reproduce

Expected behavior

The loss is expected to be normal, but now it is NaN.

Logs

> Training Environment:
 | > Backend: Accelerate
 | > Mixed precision: False
 | > Precision: float32
 | > Current device: 0
 | > Num. of GPUs: 1
 | > Num. of CPUs: 28
 | > Num. of Torch Threads: 1
 | > Torch seed: 1
 | > Torch CUDNN: True
 | > Torch CUDNN deterministic: False
 | > Torch CUDNN benchmark: False
 | > Torch TF32 MatMul: False
 > Start Tensorboard: tensorboard --logdir=/home/xj_data/liuwenrui/model/TTS/examples/checkpoints/xtts_ljspeech/GPT_XTTS_LJSpeech_FT-September-05-2024_07+56AM-0000000

 > Model has 518442047 parameters

 > EPOCH: 0/1000
 --> /home/xj_data/liuwenrui/model/TTS/examples/checkpoints/xtts_ljspeech/GPT_XTTS_LJSpeech_FT-September-05-2024_07+56AM-0000000
 > Filtering invalid eval samples!!
 > Total eval samples after filtering: 131

 > EVALUATION 

 | > Synthesizing test sentences.
The attention mask is not set and cannot be inferred from input because pad token is same as eos token. As a consequence, you may observe unexpected behavior. Please pass your input's `attention_mask` to obtain reliable results.

  --> EVAL PERFORMANCE
     | > avg_loader_time: 0.1370311975479126 (+0)
     | > avg_loss_text_ce: nan (+0)
     | > avg_loss_mel_ce: nan (+0)
     | > avg_loss: nan (+0)


 > EPOCH: 1/1000
 --> /home/xj_data/liuwenrui/model/TTS/examples/checkpoints/xtts_ljspeech/GPT_XTTS_LJSpeech_FT-September-05-2024_07+56AM-0000000
 > Sampling by language: dict_keys(['en'])

 > TRAINING (2024-09-05 07:57:04) 

   --> TIME: 2024-09-05 07:57:06 -- STEP: 0/406 -- GLOBAL_STEP: 0
     | > loss_text_ce: nan  (nan)
     | > loss_mel_ce: nan  (nan)
     | > loss: nan  (nan)
     | > current_lr: 5e-06 
     | > step_time: 0.6947  (0.6947054862976074)
     | > loader_time: 0.9726  (0.9726102352142334)


   --> TIME: 2024-09-05 07:57:34 -- STEP: 50/406 -- GLOBAL_STEP: 50
     | > loss_text_ce: nan  (nan)
     | > loss_mel_ce: nan  (nan)
     | > loss: nan  (nan)
     | > current_lr: 5e-06 
     | > step_time: 0.5593  (0.5495537853240967)
     | > loader_time: 0.0208  (0.025657353401184083)


   --> TIME: 2024-09-05 07:58:03 -- STEP: 100/406 -- GLOBAL_STEP: 100
     | > loss_text_ce: nan  (nan)
     | > loss_mel_ce: nan  (nan)
     | > loss: nan  (nan)
     | > current_lr: 5e-06 
     | > step_time: 0.5421  (0.5507096624374387)
     | > loader_time: 0.027  (0.024191355705261236)


   --> TIME: 2024-09-05 07:58:32 -- STEP: 150/406 -- GLOBAL_STEP: 150
     | > loss_text_ce: nan  (nan)
     | > loss_mel_ce: nan  (nan)
     | > loss: nan  (nan)
     | > current_lr: 5e-06 
     | > step_time: 0.5438  (0.5505051151911414)
     | > loader_time: 0.0248  (0.023627797762552902)


   --> TIME: 2024-09-05 07:59:01 -- STEP: 200/406 -- GLOBAL_STEP: 200
     | > loss_text_ce: nan  (nan)
     | > loss_mel_ce: nan  (nan)
     | > loss: nan  (nan)
     | > current_lr: 5e-06 
     | > step_time: 0.5429  (0.5503179728984828)
     | > loader_time: 0.0192  (0.023366824388504036)


   --> TIME: 2024-09-05 07:59:29 -- STEP: 250/406 -- GLOBAL_STEP: 250
     | > loss_text_ce: nan  (nan)
     | > loss_mel_ce: nan  (nan)
     | > loss: nan  (nan)
     | > current_lr: 5e-06 
     | > step_time: 0.5364  (0.5505284156799314)
     | > loader_time: 0.024  (0.023049613952636726)


   --> TIME: 2024-09-05 07:59:58 -- STEP: 300/406 -- GLOBAL_STEP: 300
     | > loss_text_ce: nan  (nan)
     | > loss_mel_ce: nan  (nan)
     | > loss: nan  (nan)
     | > current_lr: 5e-06 
     | > step_time: 0.5382  (0.5508304643630979)
     | > loader_time: 0.0228  (0.022726009686787927)


   --> TIME: 2024-09-05 08:00:27 -- STEP: 350/406 -- GLOBAL_STEP: 350
     | > loss_text_ce: nan  (nan)
     | > loss_mel_ce: nan  (nan)
     | > loss: nan  (nan)
     | > current_lr: 5e-06 
     | > step_time: 0.5357  (0.5509397418158392)
     | > loader_time: 0.0224  (0.022749708720615953)


   --> TIME: 2024-09-05 08:00:56 -- STEP: 400/406 -- GLOBAL_STEP: 400
     | > loss_text_ce: nan  (nan)
     | > loss_mel_ce: nan  (nan)
     | > loss: nan  (nan)
     | > current_lr: 5e-06 
     | > step_time: 0.5422  (0.5508994698524474)
     | > loader_time: 0.0199  (0.022739327549934404)


 > EVALUATION 

 | > Synthesizing test sentences.

  --> EVAL PERFORMANCE
     | > avg_loader_time: 0.15962785482406616 (+0.022596657276153564)
     | > avg_loss_text_ce: nan (+nan)
     | > avg_loss_mel_ce: nan (+nan)
     | > avg_loss: nan (+nan)


 > EPOCH: 2/1000
 --> /home/xj_data/liuwenrui/model/TTS/examples/checkpoints/xtts_ljspeech/GPT_XTTS_LJSpeech_FT-September-05-2024_07+56AM-0000000

 > TRAINING (2024-09-05 08:01:04) 

   --> TIME: 2024-09-05 08:01:31 -- STEP: 44/406 -- GLOBAL_STEP: 450
     | > loss_text_ce: nan  (nan)
     | > loss_mel_ce: nan  (nan)
     | > loss: nan  (nan)
     | > current_lr: 5e-06 
     | > step_time: 0.5423  (0.5685172839598219)
     | > loader_time: 0.0218  (0.02611354806206443)


   --> TIME: 2024-09-05 08:02:00 -- STEP: 94/406 -- GLOBAL_STEP: 500
     | > loss_text_ce: nan  (nan)
     | > loss_mel_ce: nan  (nan)
     | > loss: nan  (nan)
     | > current_lr: 5e-06 
     | > step_time: 0.5477  (0.5601014101758914)
     | > loader_time: 0.0206  (0.024027073636968085)


   --> TIME: 2024-09-05 08:02:29 -- STEP: 144/406 -- GLOBAL_STEP: 550
     | > loss_text_ce: nan  (nan)
     | > loss_mel_ce: nan  (nan)
     | > loss: nan  (nan)
     | > current_lr: 5e-06 
     | > step_time: 0.5499  (0.5568753977616625)
     | > loader_time: 0.0221  (0.023584759897655915)


   --> TIME: 2024-09-05 08:02:58 -- STEP: 194/406 -- GLOBAL_STEP: 600
     | > loss_text_ce: nan  (nan)
     | > loss_mel_ce: nan  (nan)
     | > loss: nan  (nan)
     | > current_lr: 5e-06 
     | > step_time: 0.5477  (0.5556149875994809)
     | > loader_time: 0.0204  (0.023164422241682858)

Environment

{
    "CUDA": {
        "GPU": [
            "NVIDIA A100-SXM4-80GB",
            "NVIDIA A100-SXM4-80GB"
        ],
        "available": true,
        "version": "11.8"
    },
    "Packages": {
        "PyTorch_debug": false,
        "PyTorch_version": "2.1.2+cu118",
        "TTS": "0.22.0",
        "numpy": "1.26.3"
    },
    "System": {
        "OS": "Linux",
        "architecture": [
            "64bit",
            "ELF"
        ],
        "processor": "x86_64",
        "python": "3.10.14",
        "version": "#1 SMP Mon Jul 22 15:34:17 CST 2024"
    }
}

Additional context

So terrible experience. I try to train it for 3 days, and the loss is always NaN

Publicación original de @r666ay en coqui-ai/TTS#3988