sampling: add Top-nσ sampler (#11223)

* initial sampling changes:

* completed top nsigma sampler implementation

* apply parameter to only llama-cli

* updated readme

* added tests and fixed nsigma impl

* cleaned up pr

* format

* format

* format

* removed commented tests

* cleanup pr and remove explicit floats

* added top-k sampler to improve performance

* changed sigma to float

* fixed string format to float

* Update src/llama-sampling.cpp

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* Update common/sampling.cpp

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* Update src/llama-sampling.cpp

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* Update src/llama-sampling.cpp

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* Update src/llama-sampling.cpp

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* Update src/llama-sampling.cpp

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* added llama_sampler_init

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

examples/main/README.md

@@ -265,6 +265,14 @@ Being experimental and unique, XTC is disabled by default. The recommended combi
 
 Example usage: `--xtc-probability 0.5 --xtc-threshold 0.1`
 
+### Top-nσ Sampling
+
+-   `--top-nsigma N`: Limit the next token selection to a subset of tokens with pre-softmax logits that are within n * σ less than the max logit (default: -1, -1 = disabled).
+
+Top-nσ sampling is a text generation method that selects tokens based on a statistical threshold in pre-softmax logits. It works by only sampling from tokens with logits that are within n * σ of the maximum logit. This method helps maintain a stable sampling space regardless of temperature scaling, allowing it to perform well on reasoning tasks even in high temperatures. Without complex probability manipulation, it efficiently filters tokens directly on the pre-softmax logits. A higher value for top-nsigma (e.g., 5) will take more noisy tokens into consideration, while a lower value (e.g., 1) will focous on the more informative region of the sampling space.
+
+Example usage: `--top-nsigma 1`
+
 ### Logit Bias
 
 -   `-l TOKEN_ID(+/-)BIAS, --logit-bias TOKEN_ID(+/-)BIAS`: Modify the likelihood of a token appearing in the generated text completion.