[BUG]: DefaultSamplingPipeline - strange behavior at high temperature

[PioneerMNDR]

Description

I decided to try the popular configuration min_p = 0.1 and temp = 1.5 or higher.
I get the following result:

I used the example LLama.Examples/Examples/LLama3ChatSession.cs
To show the incorrect behavior.
The only thing I changed
var chatHistory = new ChatHistory();
and

   var inferenceParams = new InferenceParams
  {
      SamplingPipeline = new DefaultSamplingPipeline
      {
          Temperature = 1.5f,
          MinP=0.1f,
                            
      },

      MaxTokens = 100, // keep generating tokens until the anti prompt is encountered
      AntiPrompts = [model.Tokens.EndOfTurnToken!] // model specific end of turn string
  };

In my project I use BatchedExecutor with the correct formatting "Promt template" and Anti promts, and I get exactly the same result. I also changed the sampling order of ProcessTokenDataArray and it did not change anything. I tested it on CUDA and Vulkan. I noticed a pattern in that the first 20-30 tokens are correct, and then chaos begins.

In LM Studio and Kobold CPP I set the temperature even higher, and Min p even lower, but everything worked fine there

Reproduction Steps

1. Use DefaultSamplingPipeline
2. Set temperature higher than 1.2
3. Set min_p = 0.1 or higher

Environment & Configuration

• Operating system: Win10
• .NET runtime version: 8.0.4
• LLamaSharp version: 0.16.0
• CUDA version (if you are using cuda backend): 12
• CPU & GPU device: RTX 3050 8gb and i5-12400
• Model: L3-8B-Stheno-v3.2-Q6_K.gguf

Known Workarounds

No response

[martindevans]

If possible, could you try adding some breakpoints/logging into the calls here. These are basically the lowest level calls, directly into llama.cpp

In particular looking for two things:

• Are the values you set actually getting passed through correctly? Just to make sure there's not something overwriting the values you've set.
• Are the other calls all being made with default values? Maybe try commenting them out just to be extra sure!

On a sidenote, the next version of LLamaSharp will be completely replacing sampling, because there has been a major redesign of the API on the llama.cpp side recently.

[PioneerMNDR]

These are the values ​​that are passed if they are not defined:

For the sake of purity of the experiment, I decided to comment out other samplers:

Nothing has changed(1):

Nothing has changed(2):

It feels like when he gets a high temperature he forgets EOS and starts hallucinating.
I decided to conduct an experiment. If I turn min_p up to 1, the model will always respond the same, regardless of temperature:

ex num	pic
Ex1
Ex2
Ex3

The experiment shows that min_p sampler works. And I really don't understand what the problem is. True, I reduced min_p to 0.01, and he still told the joke about the bicycle