Due diligence

• I have done my due diligence in trying to find the answer myself.

Topic

The PyTorch implementation

Question

Hi,

While reviewing the codebase, I noticed there's a second MimiModel instance (other_mimi) instantiated in both server.py and offline.py:

other_mimi = loaders.get_mimi(mimi_path, device=device)
other_mimi.set_num_codebooks(lm.dep_q)

This instance:

1. Processes the same input audio as the primary mimi via other_mimi.encode(chunk)
2. The encoded output is never used (no variable assignment, not passed to the LM)
3. Consumes ~200MB additional GPU memory

Looking at handle_connection:

other_mimi.encode(chunk)  # Result discarded
codes = mimi.encode(chunk)  # This one is actually used

Questions:

1. What was the intended purpose of other_mimi?
2. Was this meant for a separate voice prompt encoding path with isolated streaming state?
3. Is this dead code that can be safely removed, or is there planned functionality?

Any clarification would be appreciated. We're working on optimizing memory usage for deployment on consumer GPUs.

Thanks!