bug: TensorRT-LLM frequency_penalty Parameter in model.yml Only Functions Correctly with Value 1, Produces Gibberish for other values

[Van-QA]

Describe the bug
The frequency_penalty parameter in the model.yml file only functions correctly when its value is set to 1. The model generates gibberish responses when the value is set to 0 or 0.5.

To Reproduce
Steps to reproduce the behavior:

1. Go to the model.yml file.
2. Set the frequency_penalty parameter to 0.
3. Run the model and observe the responses.
4. Repeat steps 2 and 3 with frequency_penalty set to 0.5.
5. Set the frequency_penalty parameter to 1 and observe the responses.

Expected behavior
The model should produce coherent and contextually appropriate responses regardless of the frequency_penalty value, whether set to 0, 0.5, or 1.

Screenshots

Additional context
This issue affects the consistency and reliability of the model's output, impacting its usability in applications that rely on varied frequency_penalty settings for diverse response generation.

[imtuyethan]

is this bug urgent?

[vansangpfiev]

Moved to sprint 24 to handle tensorrt-llm engine

[gabrielle-ong]

Deprecated due to TensorRT-LLM not supporting Desktop
Parent issue: #1742