CUDA error when feeding more frames to PLM

Hi,

Thank you for your amazing work. I'm trying to modify and test the PLM by feeding more frames to the model. I manually set `config.data.max_video_frames = 48` and run
```
python apps/plm/generate.py --ckpt facebook/Perception-LM-8B --media_type video --media_path /shared/nas2/yaox11/ruisen/test_data/vqo/0cf71e76-0ed9-4fb9-875f-be49528105b7_797-857.mp4 --question "Describe this video."
```
and the code ran into the following error:

```
File "/shared/nas2/yaox11/ruisen/perception_models/apps/plm/generate.py", line 577, in main
generation, loglikelihood, greedy = generator.generate(prompts)
^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/shared/nas2/yaox11/miniconda3/envs/plm/lib/python3.12/site-packages/torch/utils/_contextlib.py", line 116, in decorate_context
return func(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^
File "/shared/nas2/yaox11/ruisen/perception_models/apps/plm/generate.py", line 428, in generate
prompt_logits = self.prefill(
^^^^^^^^^^^^^
File "/shared/nas2/yaox11/ruisen/perception_models/apps/plm/generate.py", line 337, in prefill
prefill_out = self.model.forward(
^^^^^^^^^^^^^^^^^^^
File "/shared/nas2/yaox11/ruisen/perception_models/apps/plm/transformer.py", line 175, in forward
h = super().forward(h, tok_idx=tok_idx, mask=mask, attn_impl=attn_impl)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/shared/nas2/yaox11/ruisen/perception_models/core/transformer.py", line 629, in forward
h = layer(h, freq_cis, tok_idx=tok_idx, mask=mask, attn_impl=attn_impl)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/shared/nas2/yaox11/miniconda3/envs/plm/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/shared/nas2/yaox11/miniconda3/envs/plm/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1747, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/shared/nas2/yaox11/ruisen/perception_models/core/transformer.py", line 580, in forward
self.attention_norm(x),
^^^^^^^^^^^^^^^^^^^^^^
File "/shared/nas2/yaox11/miniconda3/envs/plm/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/shared/nas2/yaox11/miniconda3/envs/plm/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1747, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/shared/nas2/yaox11/ruisen/perception_models/core/transformer.py", line 330, in forward
output = self._norm(x.float())
^^^^^^^^^^^^^^^^^^^^^
File "/shared/nas2/yaox11/ruisen/perception_models/core/transformer.py", line 326, in _norm
return x * torch.rsqrt((x * x).mean(-1, keepdim=True) + self.eps)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: CUDA error: device-side assert triggered
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1
Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.
```
I suspect this to be related to the maximum sequence length allowed for PLM. I wonder is there a way to fix this?