Samplers should know know their own SamplingParameters

[dgerlanc]

What behavior of the library made you think about the improvement?

Since we already have classes for samplers, we should encapsulate the known behaviors of the samplers within that class.

How would you like it to behave?

Currently, the SamplingParameters are processed in conditional statement at line 431 of outlines/generate/api.py

I think we should move the generation of the sampling parameters to the individual sampler classes.

The new code in SequenceGeneratorAdapter would be:

def __init__(self, model, logits_processor, sampler):
    self.model = model
    self.logits_processor = logits_processor
    self.sampling_params = sampler.sampling_params()

And each of these classes would get a sampling_parameters property or method.

• MultinomialSampler
• GreedySampler
• BeamSearchSampler

This will allow us to add new samplers without modifying the code in SequenceGeneratorAdapter