fix: resolve TypeError when max_tokens=None in generate() and CLI

[voidborne-d]

Summary

Fixes #281.

When cli.py omits --max-tokens (default None), it passes max_tokens=None to Dia.generate(), which overrides the hardcoded default of 3072. The generation loop then compares dec_step < None, raising:

TypeError: '<' not supported between instances of 'int' and 'NoneType'

Root Cause

generate() declared max_tokens: int = 3072, but callers (including cli.py) pass explicit None which overrides the default. Two comparisons depend on max_tokens being an integer:

while dec_step < max_tokens:       # line 696
is_max_len = current_step_idx >= max_tokens - max_delay_pattern  # line 721

Note that _prepare_generation() already handles None correctly via max_generation_length or config.decoder_config.max_position_embeddings, so only the generate() function itself crashes.

Changes

dia/model.py

• Change generate() signature: max_tokens: int = 3072 → max_tokens: int | None = None
• Add resolution at function entry: if max_tokens is None: max_tokens = self.config.decoder_config.max_position_embeddings
• This is consistent with how _prepare_generation already handles None
• Existing callers passing explicit int values are unaffected

cli.py

• Only include max_tokens in generate() kwargs when user explicitly sets --max-tokens, avoiding the None override entirely
• Import and use DEFAULT_SAMPLE_RATE from dia.model instead of hardcoded 44100 to keep the sample rate in sync with the model constant

tests/test_generate_max_tokens.py

• 15 regression tests organized into 6 test classes:
  • TestGenerateSignature (2): signature accepts None, default is None
  • TestNoneResolution (4): guard exists, resolves to config value, happens before loop, preserves explicit int
  • TestCLI (4): default is None, guards against passing None, imports DEFAULT_SAMPLE_RATE, no hardcoded 44100
  • TestPrepareGeneration (1): accepts None default
  • TestConstants (2): DEFAULT_SAMPLE_RATE defined, save_audio() uses it
  • TestIssue281Reproduction (2): confirms int < None raises TypeError, verifies no unguarded comparisons
• Tests parse source code directly (no torch/torchaudio required) so they run in lightweight CI environments

Testing

$ python3 -m pytest tests/test_generate_max_tokens.py -v
15 passed in 0.08s