Aritra Chatterjee
Aritra Chatterjee
> However, you raise a very good point around plugins for these tools, as well as the complexity we now have to deal with around configuration options. I'll give it...
Looking at the indent of the first line of the visual selection solves my problem. But there might be other use cases, that require different handling.
IMO, latexindent should be considered a fixer, not a linter!
@Ryoo72 did you figure this out?
@DarkLight1337 I'm quite sure the issue is due to the wrong `image_token_id`. According to the tokenizer the `image_token_id` is `64003` but it is `64000` inside the `merge_vision_embeddings` function.
The tokenizer loaded by vllm seems to have 3 extra tokens [, , at positions 64000-64002 causing the issue.
Loading the tokenizer using transformers lead to the same issue. This indicates a mismatch between the `config` and the `tokenizer.model` in the original checkpoint. ``` from transformers import AutoTokenizer tokenizer...
I have double checked the command, code and output. Versions are as follows: ``` trl: 0.9.4 transformers: 4.41.2 torch: 2.3.1 cuda: 12.4 python: 3.10.14 ```
@qgallouedec any update on this? I created a new environment with the latest version of try and transformers but still facing the same issue.