Add sampling API back to LlamaTokenDataArray; Add DRY and XTC Samplers

examples/simple/src/main.rs

@@ -17,7 +17,7 @@ use llama_cpp_2::model::params::kv_overrides::ParamOverrideValue;
 use llama_cpp_2::model::params::LlamaModelParams;
 use llama_cpp_2::model::LlamaModel;
 use llama_cpp_2::model::{AddBos, Special};
-use llama_cpp_2::sampling::params::LlamaSamplerChainParams;
+use llama_cpp_2::sampling::params::LlamaSamplerParams;
 use llama_cpp_2::sampling::LlamaSampler;
 
 use std::ffi::CString;
@@ -246,10 +246,12 @@ either reduce n_len or increase n_ctx"
     // The `Decoder`
     let mut decoder = encoding_rs::UTF_8.new_decoder();
 
-    let sampler_params = LlamaSamplerChainParams::default();
-    let mut sampler = LlamaSampler::new(sampler_params)?
-        .add_dist(seed.unwrap_or(1234))
-        .add_greedy();
+    let mut sampler = LlamaSampler::new(LlamaSamplerParams::chain(&[
+        LlamaSamplerParams::Dist {
+            seed: seed.unwrap_or(1234),
+        },
+        LlamaSamplerParams::Greedy,
+    ]));
 
     while n_cur <= n_len {
         // sample the next token

llama-cpp-2/src/context.rs

@@ -9,6 +9,7 @@ use crate::llama_batch::LlamaBatch;
 use crate::model::{LlamaLoraAdapter, LlamaModel};
 use crate::timing::LlamaTimings;
 use crate::token::data::LlamaTokenData;
+use crate::token::data_array::LlamaTokenDataArray;
 use crate::token::LlamaToken;
 use crate::{
     DecodeError, EmbeddingsError, EncodeError, LlamaLoraAdapterRemoveError,
@@ -202,6 +203,21 @@ impl<'model> LlamaContext<'model> {
         })
     }
 
+    /// Get the token data array for the last token in the context.
+    ///
+    /// This is a convience method that implements:
+    /// ```ignore
+    /// LlamaTokenDataArray::from_iter(ctx.candidates(), false)
+    /// ```
+    ///
+    /// # Panics
+    ///
+    /// - underlying logits data is null
+    #[must_use]
+    pub fn token_data_array(&self) -> LlamaTokenDataArray {
+        LlamaTokenDataArray::from_iter(self.candidates(), false)
+    }
+
     /// Token logits obtained from the last call to `decode()`.
     /// The logits for which `batch.logits[i] != 0` are stored contiguously
     /// in the order they have appeared in the batch.
@@ -217,6 +233,7 @@ impl<'model> LlamaContext<'model> {
     ///
     /// - `n_vocab` does not fit into a usize
     /// - token data returned is null
+    #[must_use]
     pub fn get_logits(&self) -> &[f32] {
         let data = unsafe { llama_cpp_sys_2::llama_get_logits(self.context.as_ptr()) };
         assert!(!data.is_null(), "logits data for last token is null");
@@ -237,6 +254,21 @@ impl<'model> LlamaContext<'model> {
         })
     }
 
+    /// Get the token data array for the ith token in the context.
+    ///
+    /// This is a convience method that implements:
+    /// ```ignore
+    /// LlamaTokenDataArray::from_iter(ctx.candidates_ith(i), false)
+    /// ```
+    ///
+    /// # Panics
+    ///
+    /// - logit `i` is not initialized.
+    #[must_use]
+    pub fn token_data_array_ith(&self, i: i32) -> LlamaTokenDataArray {
+        LlamaTokenDataArray::from_iter(self.candidates_ith(i), false)
+    }
+
     /// Get the logits for the ith token in the context.
     ///
     /// # Panics

llama-cpp-2/src/lib.rs

@@ -195,14 +195,6 @@ pub enum LlamaLoraAdapterRemoveError {
     ErrorResult(i32),
 }
 
-/// An error that can occur when initializing a sampler.
-#[derive(Debug, Eq, PartialEq, thiserror::Error)]
-pub enum LlamaSamplerError {
-    /// llama.cpp returned null
-    #[error("null reference from llama.cpp")]
-    NullReturn,
-}
-
 /// get the time (in microseconds) according to llama.cpp
 /// ```
 /// # use llama_cpp_2::llama_time_us;