Memory requirement changes after converting a model using create_long_model() function
- Ngôn ngữ chính
- Python
- Star
- 2.2k
- Fork
- 285
- Chỉ số merge pull request
- Không có pull request nào được merge trong 30 ngày
Mô tả
I have tried converting `roberta-base`, `distilroberta-base` to `longformer` with the `create_long_model()` function in the given notebook `convert_model_to_long.ipynb`
The problem I'm facing is that it produces CUDA out of memory issue when I try to finetune it for token classification. I can't even fit a single batch with the converted model while I can fit more than one batch with `allenai/longformer-base-4096`, `roberta-base` with Tesla T4 which has 16GB memory. I also tried fp16 precision and gradient checkpointing, but the converted model always gives CUDA OOM issue regardless of the size.
Any hint where to look for solving this issue?
I'm trying to train a token classifier with huggingface's [run_ner.py](https://github.com/huggingface/transformers/tree/master/examples/token-classification) script.
*My problem is somewhat similar to https://github.com/allenai/longformer/issues/81
Hướng dẫn đóng góp
Chưa lập chỉ mục được hướng dẫn đóng góp cho kho mã nguồn này
Đánh giá
Issue này chưa được đánh giá.