Hello, thank you very much for your work.
Regarding your answer in the issue:
We mainly modify the forward fucntion in modeling_qwen2_5_vl.py to support training with latent tokens.
Mirage/transformers/src/transformers/models/qwen2_5_vl/modeling_qwen2_5_vl.py
Line 1961 in e71ebc0
Also the generate function in geneation/utils.py for decoding with latent token during inference.
Mirage/transformers/src/transformers/generation/utils.py
Line 3197 in e71ebc0
Note that there are some hard-coded changes in our modified transformers library, so we recommend installing our version for proper reproduction of the results. We plan to release a cleaner version of code in the future!
We would like to ask whether, in the version of Transformers you provided, the only modifications are in these two places.
Hello, thank you very much for your work.
Regarding your answer in the issue:
We would like to ask whether, in the version of Transformers you provided, the only modifications are in these two places.