Issue with Seq2Seq model

[kr-ramesh]

🐛 Bug

Please reproduce using our template Colab and post here the link

https://colab.research.google.com/drive/1Eu0rxSdbdJbZUBlgJ4wxR5QXJ732bc90?usp=sharing (Some parts of the code are redundant, apologies for that)

To Reproduce

⚠️ We cannot help you without you sharing reproducible code. Do not ignore this part :)
Steps to reproduce the behavior:

1. Run the cells in order
2. Model is currently set to full fine-tuning. Error does not appear with LoRA, so it appears to be with certain layers
3. Memory issues might occur in Google Colab

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
[<ipython-input-1-78cd311674a4>](https://localhost:8080/#) in <cell line: 191>()
    200                   task_type = dict_task_types[task_type],
    201                   dataset_name = dataset_name)
--> 202         obj.finetune_model()

3 frames
[<ipython-input-1-78cd311674a4>](https://localhost:8080/#) in finetune_model(self)
    164               total_loss += loss.detach().float()
    165               loss.backward()
--> 166               optimizer.step()
    167               lr_scheduler.step()
    168               optimizer.zero_grad()

[/usr/local/lib/python3.10/dist-packages/opacus/optimizers/optimizer.py](https://localhost:8080/#) in step(self, closure)
    551             with torch.enable_grad():
    552                 closure()
--> 553         if self.pre_step():
    554             return self.original_optimizer.step()
    555         else:

[/usr/local/lib/python3.10/dist-packages/opacus/optimizers/optimizer.py](https://localhost:8080/#) in pre_step(self, closure)
    536         if self.grad_samples is None or len(self.grad_samples) == 0:
    537             return True
--> 538         self.clip_and_accumulate()
    539         if self._check_skip_next_step():
    540             self._is_last_step_skipped = True

[/usr/local/lib/python3.10/dist-packages/opacus/optimizers/optimizer.py](https://localhost:8080/#) in clip_and_accumulate(self)
    442                 g.reshape(len(g), -1).norm(2, dim=-1) for g in self.grad_samples
    443             ]
--> 444             per_sample_norms = torch.stack(per_param_norms, dim=1).norm(2, dim=1)
    445             per_sample_clip_factor = (
    446                 self.max_grad_norm / (per_sample_norms + 1e-6)

RuntimeError: stack expects each tensor to be equal size, but got [1] at entry 0 and [384] at entry 5

Expected behavior

No gradient mismatch should be there. A deeper dive into this seems to indicate a gradient mismatch, and it appears that during the computation of the activations, the mismatch appears (this might have something to do with the forward hooks). The batch size is in the first dimension for all data points. Error seems to be appearing because of relative_attention_bias related weights in the model - on freezing it, training proceeds as it should (there is a permute function transformation applied to the relative_attention_bias layer that is then stored in another variable, which might be leading to this - refer to modeling_t5.py in the huggingface transformers library).

Environment

Please copy and paste the output from our
environment collection script
(or fill out the checklist below manually).

You can get the script and run it with:

wget https://raw.githubusercontent.com/pytorch/pytorch/master/torch/utils/collect_env.py
# For security purposes, please check the contents of collect_env.py before running it.
python collect_env.py

• PyTorch Version (e.g., 1.0): 2.5.0+cu121
• OS (e.g., Linux): Ubuntu 22.04.3 LTS (x86_64)
• How you installed PyTorch (conda, pip, source): N/A
• Build command you used (if compiling from source): N/A
• Python version: 3.10.12
• CUDA/cuDNN version: 12.1
• GPU models and configuration: T4
• Any other relevant information: Ran this on system with A6000, older version of cuda, etc. Encountered the same error.

Additional context

[HuanyuZhang]

Good catch. I believe the root cause is as what you describe: the same name but different shapes of values fools the hooks, causing this gradient mismatch.

Need to think more on the general solution to avoid this.