llama-cpp-bindings 0.6.0

//! Safe wrapper around `llama_sampler`.

use std::borrow::Borrow;
use std::ffi::{CString, c_char};
use std::fmt::{Debug, Formatter};

use crate::context::LlamaContext;
use crate::ffi_error_reader::read_and_free_cpp_error;
use crate::model::LlamaModel;
use crate::token::LlamaToken;
use crate::token::data_array::LlamaTokenDataArray;
use crate::token::logit_bias::LlamaLogitBias;
use crate::{GrammarError, SampleError, SamplerAcceptError, SamplingError};

fn check_sampler_accept_status(
    status: llama_cpp_bindings_sys::llama_rs_status,
    error_ptr: *mut c_char,
) -> Result<(), SamplerAcceptError> {
    match status {
        llama_cpp_bindings_sys::LLAMA_RS_STATUS_OK => Ok(()),
        llama_cpp_bindings_sys::LLAMA_RS_STATUS_INVALID_ARGUMENT => {
            Err(SamplerAcceptError::InvalidArgument)
        }
        _ => Err(SamplerAcceptError::CppException(unsafe {
            read_and_free_cpp_error(error_ptr)
        })),
    }
}

fn check_sampler_not_null(
    sampler: *mut llama_cpp_bindings_sys::llama_sampler,
    error_ptr: *mut c_char,
) -> Result<LlamaSampler, GrammarError> {
    if sampler.is_null() {
        Err(GrammarError::NullGrammar(unsafe {
            read_and_free_cpp_error(error_ptr)
        }))
    } else {
        Ok(LlamaSampler { sampler })
    }
}

fn checked_u32_as_i32(value: u32) -> Result<i32, GrammarError> {
    i32::try_from(value).map_err(|convert_error| {
        GrammarError::IntegerOverflow(format!("value exceeds i32::MAX: {convert_error}"))
    })
}

fn checked_usize_as_i32_sampling(value: usize) -> Result<i32, SamplingError> {
    i32::try_from(value).map_err(|convert_error| {
        SamplingError::IntegerOverflow(format!("value exceeds i32::MAX: {convert_error}"))
    })
}

/// A safe wrapper around `llama_sampler`.
pub struct LlamaSampler {
    /// Raw pointer to the underlying `llama_sampler`.
    pub sampler: *mut llama_cpp_bindings_sys::llama_sampler,
}

impl Debug for LlamaSampler {
    fn fmt(&self, f: &mut Formatter<'_>) -> std::fmt::Result {
        f.debug_struct("LlamaSamplerChain").finish()
    }
}

impl LlamaSampler {
    /// Sample and accept a token from the idx-th output of the last evaluation.
    ///
    /// # Errors
    ///
    /// Returns [`SampleError`] if the C++ sampler throws an exception or if the index is invalid.
    pub fn sample(&mut self, ctx: &LlamaContext, idx: i32) -> Result<LlamaToken, SampleError> {
        let mut token: i32 = -1;
        let mut error_ptr: *mut c_char = std::ptr::null_mut();

        let status = unsafe {
            llama_cpp_bindings_sys::llama_rs_sampler_sample(
                self.sampler,
                ctx.context.as_ptr(),
                idx,
                &raw mut token,
                &raw mut error_ptr,
            )
        };

        match status {
            llama_cpp_bindings_sys::LLAMA_RS_STATUS_OK => Ok(LlamaToken(token)),
            llama_cpp_bindings_sys::LLAMA_RS_STATUS_INVALID_ARGUMENT => {
                Err(SampleError::InvalidArgument)
            }
            _ => Err(SampleError::CppException(unsafe {
                read_and_free_cpp_error(error_ptr)
            })),
        }
    }

    /// Applies this sampler to a [`LlamaTokenDataArray`].
    pub fn apply(&self, data_array: &mut LlamaTokenDataArray) {
        data_array.apply_sampler(self);
    }

    /// Accepts a token from the sampler, possibly updating the internal state of certain samplers
    /// (e.g. grammar, repetition, etc.)
    ///
    /// # Errors
    /// Returns [`SamplerAcceptError`] if the underlying sampler rejects the token.
    pub fn accept(&mut self, token: LlamaToken) -> Result<(), SamplerAcceptError> {
        self.try_accept(token)
    }

    /// Accepts several tokens from the sampler or context, possibly updating the internal state of
    /// certain samplers (e.g. grammar, repetition, etc.)
    ///
    /// # Errors
    /// Returns [`SamplerAcceptError`] if the underlying sampler rejects any token.
    pub fn accept_many(
        &mut self,
        tokens: impl IntoIterator<Item = impl Borrow<LlamaToken>>,
    ) -> Result<(), SamplerAcceptError> {
        for token in tokens {
            self.try_accept(*token.borrow())?;
        }

        Ok(())
    }

    /// Accepts several tokens from the sampler or context, possibly updating the internal state of
    /// certain samplers (e.g. grammar, repetition, etc.)
    ///
    /// # Errors
    /// Returns [`SamplerAcceptError`] if the underlying sampler rejects any token.
    pub fn with_tokens(
        mut self,
        tokens: impl IntoIterator<Item = impl Borrow<LlamaToken>>,
    ) -> Result<Self, SamplerAcceptError> {
        self.accept_many(tokens)?;

        Ok(self)
    }

    /// Try accepting a token from the sampler. Returns an error if the sampler throws.
    ///
    /// # Errors
    /// Returns an error if the underlying sampler rejects the token.
    pub fn try_accept(&mut self, token: LlamaToken) -> Result<(), SamplerAcceptError> {
        let mut error_ptr: *mut c_char = std::ptr::null_mut();

        let status = unsafe {
            llama_cpp_bindings_sys::llama_rs_sampler_accept(
                self.sampler,
                token.0,
                &raw mut error_ptr,
            )
        };

        check_sampler_accept_status(status, error_ptr)
    }

    /// Resets the internal state of the sampler.
    ///
    /// This can be useful when you want to start fresh with a sampler without creating a new instance.
    pub fn reset(&mut self) {
        unsafe {
            llama_cpp_bindings_sys::llama_sampler_reset(self.sampler);
        }
    }

    /// Gets the random seed used by this sampler.
    ///
    /// Returns:
    /// - For random samplers (dist, mirostat, `mirostat_v2)`: returns their current seed
    /// - For sampler chains: returns the first non-default seed found in reverse order
    /// - For all other samplers: returns 0xFFFFFFFF
    #[must_use]
    pub fn get_seed(&self) -> u32 {
        unsafe { llama_cpp_bindings_sys::llama_sampler_get_seed(self.sampler) }
    }

    /// Combines a list of samplers into a single sampler that applies each component sampler one
    /// after another.
    ///
    /// If you are using a chain to select a token, the chain should always end with one of
    /// [`LlamaSampler::greedy`], [`LlamaSampler::dist`], [`LlamaSampler::mirostat`], and
    /// [`LlamaSampler::mirostat_v2`].
    #[must_use]
    pub fn chain(samplers: impl IntoIterator<Item = Self>, no_perf: bool) -> Self {
        unsafe {
            let chain = llama_cpp_bindings_sys::llama_sampler_chain_init(
                llama_cpp_bindings_sys::llama_sampler_chain_params { no_perf },
            );

            for sampler in samplers {
                llama_cpp_bindings_sys::llama_sampler_chain_add(chain, sampler.sampler);
                std::mem::forget(sampler);
            }

            Self { sampler: chain }
        }
    }

    /// Same as [`Self::chain`] with `no_perf = false`.
    ///
    /// # Example
    /// ```rust
    /// use llama_cpp_bindings::token::{
    ///    LlamaToken,
    ///    data::LlamaTokenData,
    ///    data_array::LlamaTokenDataArray
    /// };
    /// use llama_cpp_bindings::sampling::LlamaSampler;
    /// use llama_cpp_bindings::llama_backend::LlamaBackend;
    /// let backend = LlamaBackend::init().unwrap();
    ///
    /// let mut data_array = LlamaTokenDataArray::new(vec![
    ///     LlamaTokenData::new(LlamaToken(0), 0., 0.),
    ///     LlamaTokenData::new(LlamaToken(1), 1., 0.),
    ///     LlamaTokenData::new(LlamaToken(2), 2., 0.),
    /// ], false);
    ///
    /// data_array.apply_sampler(&mut LlamaSampler::chain_simple([
    ///     LlamaSampler::temp(0.5),
    ///     LlamaSampler::greedy(),
    /// ]));
    ///
    /// assert_eq!(data_array.data[0].logit(), 0.);
    /// assert_eq!(data_array.data[1].logit(), 2.);
    /// assert_eq!(data_array.data[2].logit(), 4.);
    ///
    /// assert_eq!(data_array.data.len(), 3);
    /// assert_eq!(data_array.selected_token(), Some(LlamaToken(2)));
    /// ```
    #[must_use]
    pub fn chain_simple(samplers: impl IntoIterator<Item = Self>) -> Self {
        Self::chain(samplers, false)
    }

    /// Updates the logits `l_i' = l_i/t`. When `t <= 0.0`, the maximum logit is kept at its original
    /// value, the rest are set to -inf
    ///
    /// # Example:
    /// ```rust
    /// use llama_cpp_bindings::token::{
    ///    LlamaToken,
    ///    data::LlamaTokenData,
    ///    data_array::LlamaTokenDataArray
    /// };
    /// use llama_cpp_bindings::sampling::LlamaSampler;
    ///
    /// let mut data_array = LlamaTokenDataArray::new(vec![
    ///     LlamaTokenData::new(LlamaToken(0), 0., 0.),
    ///     LlamaTokenData::new(LlamaToken(1), 1., 0.),
    ///     LlamaTokenData::new(LlamaToken(2), 2., 0.),
    /// ], false);
    ///
    /// data_array.apply_sampler(&mut LlamaSampler::temp(0.5));
    ///
    /// assert_eq!(data_array.data[0].logit(), 0.);
    /// assert_eq!(data_array.data[1].logit(), 2.);
    /// assert_eq!(data_array.data[2].logit(), 4.);
    /// ```
    #[must_use]
    pub fn temp(t: f32) -> Self {
        let sampler = unsafe { llama_cpp_bindings_sys::llama_sampler_init_temp(t) };
        Self { sampler }
    }

    /// Dynamic temperature implementation (a.k.a. entropy) described in the paper
    /// <https://arxiv.org/abs/2309.02772>.
    #[must_use]
    pub fn temp_ext(t: f32, delta: f32, exponent: f32) -> Self {
        let sampler =
            unsafe { llama_cpp_bindings_sys::llama_sampler_init_temp_ext(t, delta, exponent) };
        Self { sampler }
    }

    /// Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration"
    /// <https://arxiv.org/abs/1904.09751>
    ///
    /// # Example:
    /// ```rust
    /// use llama_cpp_bindings::token::{
    ///    LlamaToken,
    ///    data::LlamaTokenData,
    ///    data_array::LlamaTokenDataArray
    /// };
    /// use llama_cpp_bindings::sampling::LlamaSampler;
    ///
    /// let mut data_array = LlamaTokenDataArray::new(vec![
    ///     LlamaTokenData::new(LlamaToken(0), 0., 0.),
    ///     LlamaTokenData::new(LlamaToken(1), 1., 0.),
    ///     LlamaTokenData::new(LlamaToken(2), 2., 0.),
    ///     LlamaTokenData::new(LlamaToken(3), 3., 0.),
    /// ], false);
    ///
    /// data_array.apply_sampler(&mut LlamaSampler::top_k(2));
    ///
    /// assert_eq!(data_array.data.len(), 2);
    /// assert_eq!(data_array.data[0].id(), LlamaToken(3));
    /// assert_eq!(data_array.data[1].id(), LlamaToken(2));
    /// ```
    #[must_use]
    pub fn top_k(k: i32) -> Self {
        let sampler = unsafe { llama_cpp_bindings_sys::llama_sampler_init_top_k(k) };
        Self { sampler }
    }

    /// Top-nσ sampling as described in academic paper "Top-nσ: Not All Logits Are You Need"
    /// <https://arxiv.org/pdf/2411.07641>
    ///
    /// This method filters logits by selecting only those within *n* standard deviations of the mean.
    ///
    /// # Parameters
    /// - `n`: Number of standard deviations from the mean to include in sampling
    ///
    /// # Example
    /// ```rust
    /// use llama_cpp_bindings::sampling::LlamaSampler;
    /// use llama_cpp_bindings::token::{
    ///     LlamaToken,
    ///     data::LlamaTokenData,
    ///     data_array::LlamaTokenDataArray
    /// };
    ///
    /// let mut data_array = LlamaTokenDataArray::new(vec![
    ///     LlamaTokenData::new(LlamaToken(0), 0.0, 0.0),
    ///     LlamaTokenData::new(LlamaToken(1), 1.0, 0.0),
    ///     LlamaTokenData::new(LlamaToken(2), 2.0, 0.0),
    /// ], false);
    ///
    /// data_array.apply_sampler(&mut LlamaSampler::top_n_sigma(2.0));
    /// ```
    #[must_use]
    pub fn top_n_sigma(n: f32) -> Self {
        let sampler = unsafe { llama_cpp_bindings_sys::llama_sampler_init_top_n_sigma(n) };
        Self { sampler }
    }

    /// Locally Typical Sampling implementation described in the paper <https://arxiv.org/abs/2202.00666>.
    #[must_use]
    pub fn typical(p: f32, min_keep: usize) -> Self {
        let sampler = unsafe { llama_cpp_bindings_sys::llama_sampler_init_typical(p, min_keep) };
        Self { sampler }
    }

    /// Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration"
    /// <https://arxiv.org/abs/1904.09751>
    #[must_use]
    pub fn top_p(p: f32, min_keep: usize) -> Self {
        let sampler = unsafe { llama_cpp_bindings_sys::llama_sampler_init_top_p(p, min_keep) };
        Self { sampler }
    }

    /// Minimum P sampling as described in <https://github.com/ggerganov/llama.cpp/pull/3841>
    #[must_use]
    pub fn min_p(p: f32, min_keep: usize) -> Self {
        let sampler = unsafe { llama_cpp_bindings_sys::llama_sampler_init_min_p(p, min_keep) };
        Self { sampler }
    }

    /// XTC sampler as described in <https://github.com/oobabooga/text-generation-webui/pull/6335>
    #[must_use]
    pub fn xtc(p: f32, t: f32, min_keep: usize, seed: u32) -> Self {
        let sampler =
            unsafe { llama_cpp_bindings_sys::llama_sampler_init_xtc(p, t, min_keep, seed) };
        Self { sampler }
    }

    /// Grammar sampler
    ///
    /// # Errors
    /// Returns an error if the grammar is invalid or the sampler cannot be initialized.
    pub fn grammar(
        model: &LlamaModel,
        grammar_str: &str,
        grammar_root: &str,
    ) -> Result<Self, GrammarError> {
        let (grammar_str, grammar_root) =
            Self::sanitize_grammar_strings(grammar_str, grammar_root)?;
        let mut error_ptr: *mut c_char = std::ptr::null_mut();

        let sampler = unsafe {
            llama_cpp_bindings_sys::llama_rs_sampler_init_grammar(
                model.vocab_ptr(),
                grammar_str.as_ptr(),
                grammar_root.as_ptr(),
                &raw mut error_ptr,
            )
        };

        check_sampler_not_null(sampler, error_ptr)
    }

    /// Lazy grammar sampler, introduced in <https://github.com/ggerganov/llama.cpp/pull/9639>
    ///
    /// This sampler enforces grammar rules only when specific trigger words or tokens are encountered.
    ///
    /// # Errors
    /// Returns an error if the grammar or trigger words are invalid.
    pub fn grammar_lazy(
        model: &LlamaModel,
        grammar_str: &str,
        grammar_root: &str,
        trigger_words: impl IntoIterator<Item = impl AsRef<[u8]>>,
        trigger_tokens: &[LlamaToken],
    ) -> Result<Self, GrammarError> {
        let (grammar_str, grammar_root) =
            Self::sanitize_grammar_strings(grammar_str, grammar_root)?;
        let trigger_words = Self::sanitize_trigger_words(trigger_words)?;
        let mut error_ptr: *mut c_char = std::ptr::null_mut();

        let mut trigger_word_ptrs: Vec<*const c_char> =
            trigger_words.iter().map(|cs| cs.as_ptr()).collect();

        let sampler = unsafe {
            llama_cpp_bindings_sys::llama_rs_sampler_init_grammar_lazy(
                model.vocab_ptr(),
                grammar_str.as_ptr(),
                grammar_root.as_ptr(),
                trigger_word_ptrs.as_mut_ptr(),
                trigger_word_ptrs.len(),
                trigger_tokens.as_ptr().cast(),
                trigger_tokens.len(),
                &raw mut error_ptr,
            )
        };

        check_sampler_not_null(sampler, error_ptr)
    }

    /// Lazy grammar sampler using regex trigger patterns.
    ///
    /// Trigger patterns are regular expressions matched from the start of the
    /// generation output. The grammar sampler will be fed content starting from
    /// the first match group.
    ///
    /// # Errors
    /// Returns an error if the grammar or trigger patterns are invalid.
    pub fn grammar_lazy_patterns(
        model: &LlamaModel,
        grammar_str: &str,
        grammar_root: &str,
        trigger_patterns: &[String],
        trigger_tokens: &[LlamaToken],
    ) -> Result<Self, GrammarError> {
        let (grammar_str, grammar_root) =
            Self::sanitize_grammar_strings(grammar_str, grammar_root)?;
        let trigger_patterns = Self::sanitize_trigger_patterns(trigger_patterns)?;
        let mut error_ptr: *mut c_char = std::ptr::null_mut();

        let mut trigger_pattern_ptrs: Vec<*const c_char> =
            trigger_patterns.iter().map(|cs| cs.as_ptr()).collect();

        let sampler = unsafe {
            llama_cpp_bindings_sys::llama_rs_sampler_init_grammar_lazy_patterns(
                model.vocab_ptr(),
                grammar_str.as_ptr(),
                grammar_root.as_ptr(),
                trigger_pattern_ptrs.as_mut_ptr(),
                trigger_pattern_ptrs.len(),
                trigger_tokens.as_ptr().cast(),
                trigger_tokens.len(),
                &raw mut error_ptr,
            )
        };

        check_sampler_not_null(sampler, error_ptr)
    }

    /// `LLGuidance` sampler for constrained decoding.
    ///
    /// Uses the `llguidance` and `toktrie` Rust crates to enforce grammar constraints
    /// during token sampling. Supports JSON schema, regex, Lark, and other grammar types.
    ///
    /// # Errors
    ///
    /// Returns [`GrammarError`] if the grammar is invalid or the sampler cannot be initialized.
    pub fn llguidance(
        model: &LlamaModel,
        grammar_kind: &str,
        grammar_data: &str,
    ) -> Result<Self, GrammarError> {
        crate::llguidance_sampler::create_llg_sampler(model, grammar_kind, grammar_data)
    }

    fn sanitize_grammar_strings(
        grammar_str: &str,
        grammar_root: &str,
    ) -> Result<(CString, CString), GrammarError> {
        if !grammar_str.contains(grammar_root) {
            return Err(GrammarError::RootNotFound);
        }

        let grammar = CString::new(grammar_str).map_err(GrammarError::GrammarNullBytes)?;
        let root = CString::new(grammar_root).map_err(GrammarError::GrammarNullBytes)?;

        Ok((grammar, root))
    }

    fn sanitize_trigger_words(
        trigger_words: impl IntoIterator<Item = impl AsRef<[u8]>>,
    ) -> Result<Vec<CString>, GrammarError> {
        trigger_words
            .into_iter()
            .map(|word| CString::new(word.as_ref()).map_err(GrammarError::TriggerWordNullBytes))
            .collect()
    }

    fn sanitize_trigger_patterns(
        trigger_patterns: &[String],
    ) -> Result<Vec<CString>, GrammarError> {
        trigger_patterns
            .iter()
            .map(|pattern| CString::new(pattern.as_str()).map_err(GrammarError::GrammarNullBytes))
            .collect()
    }

    /// DRY sampler, designed by p-e-w, as described in:
    /// <https://github.com/oobabooga/text-generation-webui/pull/5677>, porting Koboldcpp
    /// implementation authored by pi6am: <https://github.com/LostRuins/koboldcpp/pull/982>
    ///
    /// # Errors
    /// Returns an error if any string in `seq_breakers` contains null bytes.
    pub fn dry(
        model: &LlamaModel,
        multiplier: f32,
        base: f32,
        allowed_length: i32,
        penalty_last_n: i32,
        seq_breakers: impl IntoIterator<Item = impl AsRef<[u8]>>,
    ) -> Result<Self, GrammarError> {
        let seq_breakers: Vec<CString> = seq_breakers
            .into_iter()
            .map(|seq_breaker| CString::new(seq_breaker.as_ref()))
            .collect::<Result<Vec<_>, _>>()?;
        let mut seq_breaker_pointers: Vec<*const c_char> = seq_breakers
            .iter()
            .map(|seq_breaker| seq_breaker.as_ptr())
            .collect();

        let n_ctx_train_value = model.n_ctx_train().map_err(|convert_error| {
            GrammarError::IntegerOverflow(format!(
                "n_ctx_train does not fit into u32: {convert_error}"
            ))
        })?;
        let n_ctx_train = checked_u32_as_i32(n_ctx_train_value)?;
        let sampler = unsafe {
            llama_cpp_bindings_sys::llama_sampler_init_dry(
                model.vocab_ptr(),
                n_ctx_train,
                multiplier,
                base,
                allowed_length,
                penalty_last_n,
                seq_breaker_pointers.as_mut_ptr(),
                seq_breaker_pointers.len(),
            )
        };

        Ok(Self { sampler })
    }

    /// Penalizes tokens for being present in the context.
    ///
    /// Parameters:
    /// - ``penalty_last_n``: last n tokens to penalize (0 = disable penalty, -1 = context size)
    /// - ``penalty_repeat``: 1.0 = disabled
    /// - ``penalty_freq``: 0.0 = disabled
    /// - ``penalty_present``: 0.0 = disabled
    #[must_use]
    pub fn penalties(
        penalty_last_n: i32,
        penalty_repeat: f32,
        penalty_freq: f32,
        penalty_present: f32,
    ) -> Self {
        let sampler = unsafe {
            llama_cpp_bindings_sys::llama_sampler_init_penalties(
                penalty_last_n,
                penalty_repeat,
                penalty_freq,
                penalty_present,
            )
        };
        Self { sampler }
    }

    /// Mirostat 1.0 algorithm described in the paper <https://arxiv.org/abs/2007.14966>. Uses tokens instead of words.
    ///
    /// # Parameters:
    /// - ``n_vocab``: [`LlamaModel::n_vocab`]
    /// - ``seed``: Seed to initialize random generation with.
    /// - ``tau``: The target cross-entropy (or surprise) value you want to achieve for the
    ///   generated text. A higher value corresponds to more surprising or less predictable text,
    ///   while a lower value corresponds to less surprising or more predictable text.
    /// - ``eta``: The learning rate used to update `mu` based on the error between the target and
    ///   observed surprisal of the sampled word. A larger learning rate will cause `mu` to be
    ///   updated more quickly, while a smaller learning rate will result in slower updates.
    /// - ``m``: The number of tokens considered in the estimation of `s_hat`. This is an arbitrary
    ///   value that is used to calculate `s_hat`, which in turn helps to calculate the value of `k`.
    ///   In the paper, they use `m = 100`, but you can experiment with different values to see how
    ///   it affects the performance of the algorithm.
    #[must_use]
    pub fn mirostat(n_vocab: i32, seed: u32, tau: f32, eta: f32, m: i32) -> Self {
        let sampler = unsafe {
            llama_cpp_bindings_sys::llama_sampler_init_mirostat(n_vocab, seed, tau, eta, m)
        };
        Self { sampler }
    }

    /// Mirostat 2.0 algorithm described in the paper <https://arxiv.org/abs/2007.14966>. Uses tokens instead of words.
    ///
    /// # Parameters:
    /// - ``seed``: Seed to initialize random generation with.
    /// - ``tau``: The target cross-entropy (or surprise) value you want to achieve for the
    ///   generated text. A higher value corresponds to more surprising or less predictable text,
    ///   while a lower value corresponds to less surprising or more predictable text.
    /// - ``eta``: The learning rate used to update `mu` based on the error between the target and
    ///   observed surprisal of the sampled word. A larger learning rate will cause `mu` to be
    ///   updated more quickly, while a smaller learning rate will result in slower updates.
    #[must_use]
    pub fn mirostat_v2(seed: u32, tau: f32, eta: f32) -> Self {
        let sampler =
            unsafe { llama_cpp_bindings_sys::llama_sampler_init_mirostat_v2(seed, tau, eta) };
        Self { sampler }
    }

    /// Selects a token at random based on each token's probabilities
    #[must_use]
    pub fn dist(seed: u32) -> Self {
        let sampler = unsafe { llama_cpp_bindings_sys::llama_sampler_init_dist(seed) };
        Self { sampler }
    }

    /// Selects the most likely token
    ///
    /// # Example:
    /// ```rust
    /// use llama_cpp_bindings::token::{
    ///    LlamaToken,
    ///    data::LlamaTokenData,
    ///    data_array::LlamaTokenDataArray
    /// };
    /// use llama_cpp_bindings::sampling::LlamaSampler;
    ///
    /// let mut data_array = LlamaTokenDataArray::new(vec![
    ///     LlamaTokenData::new(LlamaToken(0), 0., 0.),
    ///     LlamaTokenData::new(LlamaToken(1), 1., 0.),
    /// ], false);
    ///
    /// data_array.apply_sampler(&mut LlamaSampler::greedy());
    ///
    /// assert_eq!(data_array.data.len(), 2);
    /// assert_eq!(data_array.selected_token(), Some(LlamaToken(1)));
    /// ```
    #[must_use]
    pub fn greedy() -> Self {
        let sampler = unsafe { llama_cpp_bindings_sys::llama_sampler_init_greedy() };
        Self { sampler }
    }

    /// Creates a sampler that applies bias values to specific tokens during sampling.
    ///
    /// # Parameters
    /// - ``n_vocab``: [`LlamaModel::n_vocab`]
    /// - ``biases``: Slice of [`LlamaLogitBias`] values specifying token-bias pairs
    ///
    /// # Errors
    /// Returns [`SamplingError::IntegerOverflow`] if `biases.len()` exceeds `i32::MAX`.
    ///
    /// # Example
    /// ```rust
    /// use llama_cpp_bindings::token::{LlamaToken, logit_bias::LlamaLogitBias};
    /// use llama_cpp_bindings::sampling::LlamaSampler;
    ///
    /// let biases = vec![
    ///     LlamaLogitBias::new(LlamaToken(1), 1.5),  // Increase probability of token 1
    ///     LlamaLogitBias::new(LlamaToken(2), -1.0), // Decrease probability of token 2
    /// ];
    ///
    /// // Assuming vocab_size of 32000
    /// let sampler = LlamaSampler::logit_bias(32000, &biases).unwrap();
    /// ```
    pub fn logit_bias(n_vocab: i32, biases: &[LlamaLogitBias]) -> Result<Self, SamplingError> {
        let bias_count = checked_usize_as_i32_sampling(biases.len())?;
        let data = biases
            .as_ptr()
            .cast::<llama_cpp_bindings_sys::llama_logit_bias>();

        let sampler = unsafe {
            llama_cpp_bindings_sys::llama_sampler_init_logit_bias(n_vocab, bias_count, data)
        };

        Ok(Self { sampler })
    }
}

impl Drop for LlamaSampler {
    fn drop(&mut self) {
        unsafe {
            llama_cpp_bindings_sys::llama_sampler_free(self.sampler);
        }
    }
}

#[cfg(test)]
mod tests {
    use super::LlamaSampler;
    use crate::GrammarError;

    #[test]
    fn sanitize_grammar_strings_valid() {
        let result = LlamaSampler::sanitize_grammar_strings("root ::= \"hello\"", "root");

        assert!(result.is_ok());
    }

    #[test]
    fn sanitize_grammar_strings_root_not_found() {
        let result = LlamaSampler::sanitize_grammar_strings("expr ::= \"hello\"", "root");

        assert_eq!(result.err(), Some(GrammarError::RootNotFound));
    }

    #[test]
    fn sanitize_grammar_strings_null_byte_in_grammar() {
        let result = LlamaSampler::sanitize_grammar_strings("root ::= \"\0\"", "root");

        assert!(matches!(
            result.err(),
            Some(GrammarError::GrammarNullBytes(_))
        ));
    }

    #[test]
    fn sanitize_grammar_strings_null_byte_in_root() {
        let result = LlamaSampler::sanitize_grammar_strings("ro\0ot ::= \"hello\"", "ro\0ot");

        assert!(matches!(
            result.err(),
            Some(GrammarError::GrammarNullBytes(_))
        ));
    }

    #[test]
    fn sanitize_trigger_words_valid() {
        let words: Vec<&[u8]> = vec![b"hello", b"world"];
        let result = LlamaSampler::sanitize_trigger_words(words);

        assert!(result.is_ok());
        assert_eq!(result.expect("valid trigger words").len(), 2);
    }

    #[test]
    fn sanitize_trigger_words_empty_list() {
        let words: Vec<&[u8]> = vec![];
        let result = LlamaSampler::sanitize_trigger_words(words);

        assert!(result.is_ok());
        assert!(result.expect("valid trigger words").is_empty());
    }

    #[test]
    fn sanitize_trigger_words_null_byte() {
        let words: Vec<&[u8]> = vec![b"hel\0lo"];
        let result = LlamaSampler::sanitize_trigger_words(words);

        assert!(matches!(
            result.err(),
            Some(GrammarError::TriggerWordNullBytes(_))
        ));
    }

    #[test]
    fn sanitize_trigger_patterns_valid() {
        let patterns = vec!["^hello$".to_string(), "world.*".to_string()];
        let result = LlamaSampler::sanitize_trigger_patterns(&patterns);

        assert!(result.is_ok());
        assert_eq!(result.expect("valid trigger patterns").len(), 2);
    }

    #[test]
    fn sanitize_trigger_patterns_empty_list() {
        let patterns: Vec<String> = vec![];
        let result = LlamaSampler::sanitize_trigger_patterns(&patterns);

        assert!(result.is_ok());
        assert!(result.expect("valid trigger patterns").is_empty());
    }

    #[test]
    fn sanitize_trigger_patterns_null_byte() {
        let patterns = vec!["hel\0lo".to_string()];
        let result = LlamaSampler::sanitize_trigger_patterns(&patterns);

        assert!(matches!(
            result.err(),
            Some(GrammarError::GrammarNullBytes(_))
        ));
    }

    #[test]
    fn apply_modifies_data_array() {
        use crate::token::LlamaToken;
        use crate::token::data::LlamaTokenData;
        use crate::token::data_array::LlamaTokenDataArray;

        let sampler = LlamaSampler::greedy();
        let mut data_array = LlamaTokenDataArray::new(
            vec![
                LlamaTokenData::new(LlamaToken::new(0), 1.0, 0.0),
                LlamaTokenData::new(LlamaToken::new(1), 5.0, 0.0),
            ],
            false,
        );

        sampler.apply(&mut data_array);

        assert_eq!(data_array.selected_token(), Some(LlamaToken::new(1)));
    }

    #[test]
    fn accept_succeeds() {
        let mut sampler = LlamaSampler::chain_simple([
            LlamaSampler::penalties(64, 1.1, 0.0, 0.0),
            LlamaSampler::greedy(),
        ]);

        sampler
            .accept(crate::token::LlamaToken::new(1))
            .expect("test: accept should succeed");
    }

    #[test]
    fn try_accept_succeeds_on_penalties_sampler() {
        let mut sampler = LlamaSampler::chain_simple([
            LlamaSampler::penalties(64, 1.1, 0.0, 0.0),
            LlamaSampler::greedy(),
        ]);

        let result = sampler.try_accept(crate::token::LlamaToken::new(42));

        assert!(result.is_ok());
    }

    #[test]
    fn accept_many_multiple_tokens() {
        use crate::token::LlamaToken;

        let mut sampler = LlamaSampler::chain_simple([
            LlamaSampler::penalties(64, 1.1, 0.0, 0.0),
            LlamaSampler::greedy(),
        ]);

        sampler
            .accept_many([LlamaToken::new(1), LlamaToken::new(2), LlamaToken::new(3)])
            .expect("test: accept_many should succeed");
    }

    #[test]
    fn with_tokens_builder_pattern() {
        use crate::token::LlamaToken;

        let _sampler = LlamaSampler::chain_simple([
            LlamaSampler::penalties(64, 1.1, 0.0, 0.0),
            LlamaSampler::greedy(),
        ])
        .with_tokens([LlamaToken::new(10), LlamaToken::new(20)])
        .expect("test: with_tokens should succeed");
    }

    #[test]
    fn all_sampler_constructors() {
        use crate::token::LlamaToken;
        use crate::token::logit_bias::LlamaLogitBias;

        let _temp = LlamaSampler::temp(0.8);
        let _temp_ext = LlamaSampler::temp_ext(0.8, 0.1, 1.0);
        let _top_k = LlamaSampler::top_k(40);
        let _top_n_sigma = LlamaSampler::top_n_sigma(2.0);
        let _top_p = LlamaSampler::top_p(0.9, 1);
        let _min_p = LlamaSampler::min_p(0.05, 1);
        let _typical = LlamaSampler::typical(0.9, 1);
        let _xtc = LlamaSampler::xtc(0.1, 0.5, 1, 42);
        let _dist = LlamaSampler::dist(42);
        let _mirostat = LlamaSampler::mirostat(32000, 42, 5.0, 0.1, 100);
        let _mirostat_v2 = LlamaSampler::mirostat_v2(42, 5.0, 0.1);
        let biases = vec![LlamaLogitBias::new(LlamaToken::new(0), -100.0)];
        let _logit_bias = LlamaSampler::logit_bias(32000, &biases);
        let _chain = LlamaSampler::chain([LlamaSampler::greedy()], true);
    }

    #[test]
    fn reset_and_get_seed() {
        let mut sampler = LlamaSampler::dist(42);
        sampler.reset();
        let _seed = sampler.get_seed();
    }

    #[test]
    fn debug_formatting() {
        let sampler = LlamaSampler::greedy();
        let debug_output = format!("{sampler:?}");
        assert!(debug_output.contains("LlamaSampler"));
    }

    #[test]
    fn checked_u32_as_i32_overflow() {
        let result = super::checked_u32_as_i32(u32::MAX);
        assert!(result.is_err());
    }

    #[test]
    fn checked_usize_as_i32_sampling_overflow() {
        let result = super::checked_usize_as_i32_sampling(usize::MAX);
        assert!(result.is_err());
    }

    #[test]
    fn check_sampler_accept_status_ok() {
        let result = super::check_sampler_accept_status(
            llama_cpp_bindings_sys::LLAMA_RS_STATUS_OK,
            std::ptr::null_mut(),
        );

        assert!(result.is_ok());
    }

    #[test]
    fn check_sampler_accept_status_invalid_argument() {
        let result = super::check_sampler_accept_status(
            llama_cpp_bindings_sys::LLAMA_RS_STATUS_INVALID_ARGUMENT,
            std::ptr::null_mut(),
        );

        assert!(matches!(
            result,
            Err(crate::SamplerAcceptError::InvalidArgument)
        ));
    }

    #[test]
    fn check_sampler_accept_status_exception() {
        let result = super::check_sampler_accept_status(
            llama_cpp_bindings_sys::LLAMA_RS_STATUS_EXCEPTION,
            std::ptr::null_mut(),
        );

        assert!(matches!(
            result,
            Err(crate::SamplerAcceptError::CppException(_))
        ));
    }

    #[test]
    fn check_sampler_not_null_returns_error() {
        let result = super::check_sampler_not_null(std::ptr::null_mut(), std::ptr::null_mut());

        assert!(result.is_err());
    }
}