fix: remove unused other_mimi instance to save ~200MB GPU memory

[ThanhNguyxn]

Summary

Fixes #46 - Remove unused other_mimi MimiModel instance that wastes ~200MB GPU memory

Problem Analysis

After careful code review, the other_mimi instance in both server.py and offline.py:

1. Is instantiated identically to the primary mimi:

   mimi = loaders.get_mimi(args.mimi_weight, args.device)
   other_mimi = loaders.get_mimi(args.mimi_weight, args.device)  # Same weights, same device

2. Has its outputs discarded (assigned to _):

   codes = mimi.encode(chunk)        # ← Used
   _ = other_mimi.encode(chunk)      # ← Discarded
   
   main_pcm = mimi.decode(tokens[:, 1:9])  # ← Used
   _ = other_mimi.decode(tokens[:, 1:9])   # ← Discarded

3. Consumes significant resources:

   • ~200MB GPU memory for the duplicate model
   • Extra compute cycles on every encode/decode call
   • Maintains separate streaming state that's never read

Investigation

I verified there are no hidden side effects:

• other_mimi.encode() and other_mimi.decode() return values are never used
• The streaming state of other_mimi is reset but never read
• No inter-model communication exists between mimi and other_mimi
• The two instances share no state (each is independently loaded)

Changes

server.py

• Removed other_mimi from ServerState dataclass fields
• Removed other_mimi from __init__ parameters
• Removed other_mimi.streaming_forever(1) call
• Removed other_mimi.encode(chunk) in warmup() and opus_loop()
• Removed other_mimi.decode(tokens) in warmup() and opus_loop()
• Removed other_mimi.reset_streaming() call
• Removed other_mimi instantiation in main()
• Removed other_mimi from ServerState() constructor call

offline.py

• Removed other_mimi parameter from warmup() function
• Removed other_mimi parameter from decode_tokens_to_pcm() function
• Removed other_mimi.encode() call in warmup()
• Removed other_mimi.decode() calls in warmup() and decode_tokens_to_pcm()
• Removed other_mimi instantiation in run_inference()
• Removed other_mimi.streaming_forever(1) call
• Removed other_mimi.reset_streaming() call
• Updated function calls to not pass other_mimi

Memory Impact

Metric	Before	After	Savings
MimiModel instances	2	1	~200MB GPU RAM
encode() calls/frame	2	1	50% compute
decode() calls/frame	2	1	50% compute

Backward Compatibility

✅ Fully backward compatible - no API changes, no behavioral changes

Testing Recommendation

# Server mode
python -m moshi.server --ssl "$SSL_DIR"

# Offline mode
python -m moshi.offline --input-wav input.wav --output-wav output.wav \
  --output-text output.json --voice-prompt NATM1.pt

Both should work identically but with ~200MB less GPU memory usage.