Don't re-use logits processors in SequenceGeneratorAdapter, copy them

[lapp0]

Fixes #1109

Problem

Logits processors can't be reused across multiple generation runs and must be copied. SequenceGeneratorAdapter wasn't respecting this requirement. The result is that in generator = generate.choice(...), the generator can only be used once.

During inference, logits processors are called with a sequence of input_ids + output_ids. Since structured generation only applies to output_ids, we track the length of input_ids on the first call and treat subsequent tokens as output_ids. However, this approach fails when the input_ids sequence changes.

Solution

copy the logits processors for each SequenceGeneratorAdapter.__call__(...), ensuring they can correctly determine the start of output_ids.

Further Work

Logits processors requires more documentation. Especially since vLLM contributors have discussed replacing Outlines logits processors implementation with import outlines.processors.