Add generation params to inference engines

[oelachqar]

Add Generation Parameters to Inference Engines

This PR introduces a set of missing generation parameters, and updates all the current inference engines to support them, or if not supported by the engine, logs a warning that the parameter will be ignored.

The following parameters are added: temperature, top_p, frequency_penalty, presence_penalty, stop sequences, logit_bias, and min_p.

Towards OPE-328

Changes

• Updated GenerationParams class with new parameters
• Implemented parameter support in AnthropicInferenceEngine, LlamaCppInferenceEngine, NativeTextInferenceEngine, RemoteInferenceEngine, and VLLMInferenceEngine

Usage

Setting Generation Parameters

from oumi.core.configs.params.generation_params import GenerationParams

params = GenerationParams(
    max_new_tokens=100,
    temperature=0.7,
    top_p=0.9,
    frequency_penalty=0.1,
    presence_penalty=0.1,
    stop=["END"],
    logit_bias={50256: -100},  # Decrease likelihood of EOS token
    min_p=0.05
)

Using Parameters with an Inference Engine

from oumi.inference.llama_cpp_inference_engine import LlamaCppInferenceEngine

engine = LlamaCppInferenceEngine(model_params, generation_params=params)
response = engine.generate(conversation)

[linear]

OPE-328

[taenin]

If you want to simplify the warning logic, we could implement a get_supported_params method in the baseInferenceEngine that by default returns an empty set. This method is called by infer() and will print warnings where relevant. Each derivative class could simply implement that method and get alerting as needed. Food for thought though, not completely necessary as we only have a few engines.

Additionally, we should update our unit tests to verify that the new parameters are passed appropriately

[oelachqar]

If you want to simplify the warning logic, we could implement a get_supported_params method in the baseInferenceEngine that by default returns an empty set. This method is called by infer() and will print warnings where relevant. Each derivative class could simply implement that method and get alerting as needed. Food for thought though, not completely necessary as we only have a few engines.

Additionally, we should update our unit tests to verify that the new parameters are passed appropriately

That's a great suggestion -- logged OPE-546 to address as a follow-up. Update this PR to include unit tests for the generation params