Operating system
Linux
Hardware
GPU with CUDA
Description
Awesome work！
However, I encountered a problem when training with Chinese corpus: the codebook and K values did not match.
[rank0]: File "/home/conda/envs/real_moshi/lib/python3.12/site-packages/moshi/models/lm.py", line 338, in forward
[rank0]: assert K == self.num_codebooks, (K, self.num_codebooks)
[rank0]: ^^^^^^^^^^^^^^^^^^^^^^^
[rank0]: AssertionError: (9, 17)
I made some changes to the train.py file.
Around line 141, I made the following changes.
mimi = checkpoint_info.get_mimi(device="cuda")
# 重要：Moshi 7B 模型期望 n_q=16 个音频 codebooks
# 必须手动设置 Mimi 使用正确的 codebooks 数量以匹配模型
mimi.set_num_codebooks(lm_config["n_q"])
main_logger_info(f"Mimi 设置为使用 {mimi.num_codebooks} 个 codebooks")
mimi.eval()
Everything seems to be working now.
Extra information
nothing
Environment
Fill in the following information on your system.
- Operating system version:
- Python version:
- PyTorch version:
- CUDA version (run
python -c 'import torch; print(torch.version.cuda)'):
- GPU model and memory:
Operating system
Linux
Hardware
GPU with CUDA
Description
Awesome work！
However, I encountered a problem when training with Chinese corpus: the codebook and K values did not match.
I made some changes to the train.py file.
Around line 141, I made the following changes.
Everything seems to be working now.
Extra information
nothing
Environment
Fill in the following information on your system.
python -c 'import torch; print(torch.version.cuda)'):