LoRA Training Taking Longer Than Expected Compared to ViTs or Other Models

[cod3r0k]

Self Checks

• This template is only for bug reports. For questions, please visit Discussions.
• I have thoroughly reviewed the project documentation (installation, training, inference) but couldn't find information to solve my problem. English 中文 日本語 Portuguese (Brazil)
• I have searched for existing issues, including closed ones. Search issues
• I confirm that I am using English to submit this report (我已阅读并同意 Language Policy).
• [FOR CHINESE USERS] 请务必使用英文提交 Issue，否则会被关闭。谢谢！:）
• Please do not modify this template and fill in all required fields.

Cloud or Self Hosted

Self Hosted (Source)

Environment Details

ubuntu

Steps to Reproduce

I ran training with LoRA.
I don't know why it takes longer compared to ViTs or other models.

Does that make sense? How long should it typically take, for example, on LJSpeech with a T4 GPU?

LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
[2025-01-24 20:32:52,673][fish_speech.models.text2semantic.lit_module][INFO] - [rank: 0] Set weight decay: 0 for 432 parameters
[2025-01-24 20:32:52,674][fish_speech.models.text2semantic.lit_module][INFO] - [rank: 0] Set weight decay: 0.0 for 61 parameters

   | Name                      | Type              | Params | Mode 
-------------------------------------------------------------------------
0  | model                     | DualARTransformer | 644 M  | train
1  | model.embeddings          | Embedding         | 105 M  | train
2  | model.codebook_embeddings | Embedding         | 8.5 M  | train
3  | model.layers              | ModuleList        | 362 M  | train
4  | model.norm                | RMSNorm           | 1.0 K  | train
5  | model.output              | Linear            | 105 M  | train
6  | model.fast_project_in     | Identity          | 0      | train
7  | model.fast_embeddings     | Embedding         | 1.1 M  | train
8  | model.fast_layers         | ModuleList        | 60.4 M | train
9  | model.fast_norm           | RMSNorm           | 1.0 K  | train
10 | model.fast_output         | Linear            | 1.1 M  | train
-------------------------------------------------------------------------
6.2 M     Trainable params
637 M     Non-trainable params
644 M     Total params
2,576.370 Total estimated model params size (MB)
433       Modules in train mode
0         Modules in eval mode
Sanity Checking: |          | 0/? [00:00<?, ?it/s][2025-01-24 20:32:52,851][fish_speech.datasets.semantic][INFO] - [rank: 0] Reading 1 / 1 files
[2025-01-24 20:32:53,264][fish_speech.datasets.semantic][INFO] - [rank: 0] Read total 1 groups of data
[2025-01-24 20:33:05,749][fish_speech.datasets.semantic][INFO] - [rank: 0] Reading 1 / 1 files
[2025-01-24 20:33:06,059][fish_speech.datasets.semantic][INFO] - [rank: 0] Read total 1 groups of data
Epoch 0: |          | 100/? [26:13<00:00,  0.06it/s, v_num=0, train/loss=8.250, train/top_5_accuracy=0.412]
Validation: |          | 0/? [00:00<?, ?it/s]
Validation:   0% 0/10 [00:00<?, ?it/s]       
Validation DataLoader 0:   0% 0/10 [00:00<?, ?it/s]
Validation DataLoader 0: 100% 10/10 [00:47<00:00,  4.80s/it]
Epoch 0: |          | 200/? [53:16<00:00,  0.06it/s, v_num=0, train/loss=7.880, train/top_5_accuracy=0.438, val/loss=8.060, val/top_5_accuracy=0.423]
Validation: |          | 0/? [00:00<?, ?it/s]
Validation:   0% 0/10 [00:00<?, ?it/s]       
Validation DataLoader 0:   0% 0/10 [00:00<?, ?it/s]
Validation DataLoader 0: 100% 10/10 [00:47<00:00,  4.78s/it]
Epoch 0: |          | 300/? [1:20:17<00:00,  0.06it/s, v_num=0, train/loss=7.940, train/top_5_accuracy=0.418, val/loss=8.060, val/top_5_accuracy=0.423]
Validation: |          | 0/? [00:00<?, ?it/s]
Validation:   0% 0/10 [00:00<?, ?it/s]       
Validation DataLoader 0:   0% 0/10 [00:00<?, ?it/s]
Validation DataLoader 0: 100% 10/10 [00:47<00:00,  4.79s/it]
Epoch 0: |          | 380/? [1:42:08<00:00,  0.06it/s, v_num=0, train/loss=7.880, train/top_5_accuracy=0.430, val/loss=7.890, val/top_5_accuracy=0.433]

✔️ Expected Behavior

No response

❌ Actual Behavior

No response

[Stardust-minus]

It's normal to have a longer training time if you finetune a larger model.

[cod3r0k]

How long it takes? @Stardust-minus

[Stardust-minus]

Varies depends on which GPU you use.

[cod3r0k]

I use the A100, and its utilization is at 100%, but it takes a long time compared to my previous experience with VITS. The steps increase very slowly. After multiple days, it is still at epoch 0 and is only making minimal progress with the steps, which I think is very inefficient. Additionally, it consumes a huge amount of RAM (~120GB). Why is this happening? @Stardust-minus

[cod3r0k]

After 22 hours, it doesn’t progress to a better state. Accuracy fluctuates between 46 and 47 percent.

Epoch 0: | | 11613/? [28:10:18<00:00, 0.11it/s, v_num=0, train/loss=7.000, train/top_5_accuracy=0.475, val/loss=6.770, val/top_5_accuracy=0.472]