Skip to content

win11下用MOSS-TTS Local Transformer v1.5 微调出现的错误 #193

Description

@freenewfree

尊敬的项目组:
你们好,今天在用贵项目组的MOSS-TTS Local Transformer v1.5 在win下进行微调,运行完数据预处理,进行最终微调训练时,出现如下错误:

D:\ai\MOSS-TTS\moss_tts_local_v1.5\finetuning>"D:\ai\MOSS-TTS\py312\Scripts\accelerate.exe" launch --mixed_precision bf16 sft.py --model-path "D:\ai\MOSS-TTS\autodl-tmp\models\MOSS-TTS-Local-Transformer-v1.5" --codec-path "D:\ai\MOSS-TTS\autodl-tmp\models\MOSS-Audio-Tokenizer-v2" --codec-weight-dtype fp32 --codec-compute-dtype bf16 --train-jsonl "D:\ai\MOSS-TTS\autodl-tmp\data1.5\train_with_codes.jsonl" --output-dir "D:\ai\MOSS-TTS\autodl-tmp\output\moss_tts_local_v1_5_sft" --per-device-batch-size 1 --gradient-accumulation-steps 8 --learning-rate 1.0e-5 --warmup-ratio 0.03 --lr-scheduler-type constant --num-epochs 3 --mixed-precision bf16 --channelwise-loss-weight 1,2.6666666666666665,2.6666666666666665,2.6666666666666665,2.6666666666666665,2.6666666666666665,2.6666666666666665,2.6666666666666665,2.6666666666666665,2.6666666666666665,2.6666666666666665,2.6666666666666665,2.6666666666666665 --gradient-checkpointing --attn-implementation flash_attention_2 --skip-nonfinite-batches
The following values were not passed to accelerate launch and had defaults used instead:
--num_processes was set to a value of 1
--num_machines was set to a value of 1
--dynamo_backend was set to a value of 'no'
To avoid this warning pass in values for each of the problematic parameters or run accelerate config.
[2026-06-27 23:22:05] [sft] global_batch_size=1 x 1 x 8 = 8
[2026-06-27 23:22:05] [sft] distributed_type=DistributedType.NO num_processes=1 using_pre_sharded_files=False train_files=1 local_train_files=1 local_train_records=71
The tokenizer you are loading from 'D:\ai\MOSS-TTS\autodl-tmp\models\MOSS-TTS-Local-Transformer-v1.5' with an incorrect regex pattern: https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503/discussions/84#69121093e8b480e709447d5e. This will lead to incorrect tokenization. You should set the fix_mistral_regex=True flag when loading this tokenizer to fix this issue.
Loading weights: 100%|████████████████| 438/438 [00:00<00:00, 4335.92it/s, Materializing param=transformer.norm.weight]
[2026-06-27 23:22:06] [sft] attn=flash_attention_2 model_dtype=torch.bfloat16 n_vq=12 channelwise_loss_weight=[1.0, 2.6666666666666665, 2.6666666666666665, 2.6666666666666665, 2.6666666666666665, 2.6666666666666665, 2.6666666666666665, 2.6666666666666665, 2.6666666666666665, 2.6666666666666665, 2.6666666666666665, 2.6666666666666665, 2.6666666666666665] gradient_checkpointing=True gradient_checkpointing_impl=module_wrapper gradient_checkpointing_scope=global gradient_checkpointing_use_reentrant=False
[2026-06-27 23:22:09] [sft] scheduler=constant warmup_steps=1 local_micro_batches_per_epoch=71 micro_batches_per_epoch=71 dropped_local_micro_batches_per_epoch=0 optimizer_steps_per_epoch=9 max_train_steps=27
[2026-06-27 23:23:14] [sft] warning: Non-finite gradient norm at epoch=0 global_step=0: nan. first_nonfinite_grad=transformer.embed_tokens.weight bad=388956160/388956160 dtype=torch.bfloat16 shape=(151936, 2560); record_ids=['29', '40', '65', '70', '1', '33', '18', '56']; skipped
[2026-06-27 23:24:20] [sft] warning: Non-finite gradient norm at epoch=0 global_step=0: nan. first_nonfinite_grad=transformer.embed_tokens.weight bad=388956160/388956160 dtype=torch.bfloat16 shape=(151936, 2560); record_ids=['67', '28', '25', '3', '23', '6', '34', '14']; skipped
[2026-06-27 23:25:27] [sft] warning: Non-finite gradient norm at epoch=0 global_step=0: nan. first_nonfinite_grad=transformer.embed_tokens.weight bad=388956160/388956160 dtype=torch.bfloat16 shape=(151936, 2560); record_ids=['32', '66', '31', '41', '8', '51', '13', '22']; skipped
[2026-06-27 23:26:24] [sft] warning: Non-finite gradient norm at epoch=0 global_step=0: nan. first_nonfinite_grad=transformer.embed_tokens.weight bad=388956160/388956160 dtype=torch.bfloat16 shape=(151936, 2560); record_ids=['44', '45', '61', '68', '4', '20', '64', '50']; skipped
[2026-06-27 23:27:22] [sft] warning: Non-finite gradient norm at epoch=0 global_step=0: nan. first_nonfinite_grad=transformer.embed_tokens.weight bad=388956160/388956160 dtype=torch.bfloat16 shape=(151936, 2560); record_ids=['57', '2', '0', '11', '15', '19', '36', '38']; skipped

ient_checkpointing=True gradient_checkpointing_impl=module_wrapper gradient_checkpointing_scope=global gradient_checkpointing_use_reentrant=False
[2026-06-27 23:32:45] [sft] scheduler=constant warmup_steps=0 local_micro_batches_per_epoch=71 micro_batches_per_epoch=71 dropped_local_micro_batches_per_epoch=0 optimizer_steps_per_epoch=9 max_train_steps=27
Traceback (most recent call last):
File "D:\ai\MOSS-TTS\moss_tts_local_v1.5\finetuning\sft.py", line 1173, in
main()
File "D:\ai\MOSS-TTS\moss_tts_local_v1.5\finetuning\sft.py", line 1107, in main
raise FloatingPointError(message)
FloatingPointError: Non-finite gradient norm at epoch=0 global_step=0: nan. first_nonfinite_grad=transformer.embed_tokens.weight bad=388956160/388956160 dtype=torch.bfloat16 shape=(151936, 2560); record_ids=['29', '40', '65', '70', '1', '33', '18', '56']
Traceback (most recent call last):
File "", line 198, in _run_module_as_main
File "", line 88, in run_code
File "D:\ai\MOSS-TTS\py312\Scripts\accelerate.exe_main
.py", line 5, in
File "D:\ai\MOSS-TTS\py312\Lib\site-packages\accelerate\commands\accelerate_cli.py", line 50, in main
args.func(args)
File "D:\ai\MOSS-TTS\py312\Lib\site-packages\accelerate\commands\launch.py", line 1407, in launch_command
simple_launcher(args)
File "D:\ai\MOSS-TTS\py312\Lib\site-packages\accelerate\commands\launch.py", line 993, in simple_launcher
raise subprocess.CalledProcessError(returncode=process.returncode, cmd=cmd)
subprocess.CalledProcessError: Command '['D:\ai\MOSS-TTS\py312\python.exe', 'sft.py', '--model-path', 'D:\ai\MOSS-TTS\autodl-tmp\models\MOSS-TTS-Local-Transformer-v1.5', '--codec-path', 'D:\ai\MOSS-TTS\autodl-tmp\models\MOSS-Audio-Tokenizer-v2', '--codec-weight-dtype', 'fp32', '--codec-compute-dtype', 'bf16', '--train-jsonl', 'D:\ai\MOSS-TTS\autodl-tmp\data1.5\train_with_codes.jsonl', '--output-dir', 'D:\ai\MOSS-TTS\autodl-tmp\output\moss_tts_local_v1_5_sft', '--per-device-batch-size', '1', '--gradient-accumulation-steps', '8', '--learning-rate', '2.0e-5', '--warmup-ratio', '0', '--lr-scheduler-type', 'constant', '--num-epochs', '3', '--mixed-precision', 'bf16', '--channelwise-loss-weight', '1,32', '--gradient-checkpointing', '--attn-implementation', 'flash_attention_2']' returned non-zero exit status 1.

这是在AI的帮助下,进行的微调参数训练的,训练音频数据(AU转为标准的48KHZ 2声道,16bit,1536kbps)从中文改为纯数字,文本内容也把阿拉伯数字都改为一、二、三之类的,包括特殊符号都去除,只留下,。?、等常用标点符号,但都是出现上面的错误,请问是1.5不支持在win下微调吗?还是代码刚发布,还有bug需要调整呢,望回复为盼!感激🙏

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions