TensorRT-LLMs/docs/source/torch/features/sampling.md

Sampling

The PyTorch backend supports most of the sampling features that are supported on the C++ backend, such as temperature, top-k and top-p sampling, stop words, bad words, penalty, context and generation logits, and log probs.

The following example prepares two identical prompts which will give different results due to the sampling parameters chosen:

from tensorrt_llm import LLM
llm = LLM(model='nvidia/Llama-3.1-8B-Instruct-FP8')
sampling_params = SamplingParams(
        temperature=1.0,
        top_k=8,
        top_p=0.5,
    )
llm.generate(["Hello, my name is",
            "Hello, my name is"], sampling_params)

When using speculative decoders such as MTP or Eagle-3 the subset of sampling options available is more restricted.