chore: update llama.cpp to b6482 (3d4053f); bump crate versions to 0.1.123

Author: admiralakber

Cargo.lock

@@ -1,7 +1,7 @@
 [package]
 name = "llama-cpp-2"
 description = "llama.cpp bindings for Rust"
-version = "0.1.122"
+version = "0.1.123"
 edition = "2021"
 license = "MIT OR Apache-2.0"
 repository = "https://github.com/utilityai/llama-cpp-rs"
@@ -10,7 +10,7 @@ repository = "https://github.com/utilityai/llama-cpp-rs"
 
 [dependencies]
 enumflags2 = "0.7.12"
-llama-cpp-sys-2 = { path = "../llama-cpp-sys-2", version = "0.1.113" }
+llama-cpp-sys-2 = { path = "../llama-cpp-sys-2", version = "0.1.123" }
 thiserror = { workspace = true }
 tracing = { workspace = true }
 tracing-core = { workspace = true }
@@ -35,7 +35,7 @@ mtmd = ["llama-cpp-sys-2/mtmd"]
 
 
 [target.'cfg(all(target_os = "macos", any(target_arch = "aarch64", target_arch = "arm64")))'.dependencies]
-llama-cpp-sys-2 = { path = "../llama-cpp-sys-2", version = "0.1.113", features = [
+llama-cpp-sys-2 = { path = "../llama-cpp-sys-2", version = "0.1.123", features = [
     "metal",
 ] }
 

llama-cpp-2/Cargo.toml

@@ -28,7 +28,8 @@ impl LlamaContext<'_> {
     /// * `dest` - The sequence id to copy the cache to.
     /// * `size` - The size of the cache to copy.
     pub fn copy_cache(&mut self, src: i32, dest: i32, size: i32) {
-        unsafe { llama_cpp_sys_2::llama_kv_self_seq_cp(self.context.as_ptr(), src, dest, 0, size) }
+        let mem = unsafe { llama_cpp_sys_2::llama_get_memory(self.context.as_ptr()) };
+        unsafe { llama_cpp_sys_2::llama_memory_seq_cp(mem, src, dest, 0, size) }
     }
 
     /// Copy the cache from one sequence to another.
@@ -57,9 +58,8 @@ impl LlamaContext<'_> {
         let p1 = p1
             .map_or(Ok(-1), i32::try_from)
             .map_err(KvCacheConversionError::P1TooLarge)?;
-        unsafe {
-            llama_cpp_sys_2::llama_kv_self_seq_cp(self.context.as_ptr(), src, dest, p0, p1);
-        }
+        let mem = unsafe { llama_cpp_sys_2::llama_get_memory(self.context.as_ptr()) };
+        unsafe { llama_cpp_sys_2::llama_memory_seq_cp(mem, src, dest, p0, p1) };
         Ok(())
     }
 
@@ -92,18 +92,16 @@ impl LlamaContext<'_> {
         let p1 = p1
             .map_or(Ok(-1), i32::try_from)
             .map_err(KvCacheConversionError::P1TooLarge)?;
-        Ok(unsafe { llama_cpp_sys_2::llama_kv_self_seq_rm(self.context.as_ptr(), src, p0, p1) })
+        let mem = unsafe { llama_cpp_sys_2::llama_get_memory(self.context.as_ptr()) };
+        Ok(unsafe { llama_cpp_sys_2::llama_memory_seq_rm(mem, src, p0, p1) })
     }
 
-    /// Returns the number of used KV cells (i.e. have at least one sequence assigned to them)
-    #[must_use]
-    pub fn get_kv_cache_used_cells(&self) -> i32 {
-        unsafe { llama_cpp_sys_2::llama_kv_self_used_cells(self.context.as_ptr()) }
-    }
 
     /// Clear the KV cache
     pub fn clear_kv_cache(&mut self) {
-        unsafe { llama_cpp_sys_2::llama_kv_self_clear(self.context.as_ptr()) }
+        let mem = unsafe { llama_cpp_sys_2::llama_get_memory(self.context.as_ptr()) };
+        // clear both metadata and data buffers to match previous semantics
+        unsafe { llama_cpp_sys_2::llama_memory_clear(mem, true) }
     }
 
     /// Removes all tokens that do not belong to the specified sequence
@@ -112,7 +110,8 @@ impl LlamaContext<'_> {
     ///
     /// * `seq_id` - The sequence id to keep
     pub fn llama_kv_cache_seq_keep(&mut self, seq_id: i32) {
-        unsafe { llama_cpp_sys_2::llama_kv_self_seq_keep(self.context.as_ptr(), seq_id) }
+        let mem = unsafe { llama_cpp_sys_2::llama_get_memory(self.context.as_ptr()) };
+        unsafe { llama_cpp_sys_2::llama_memory_seq_keep(mem, seq_id) }
     }
 
     #[allow(clippy::doc_markdown)]
@@ -146,9 +145,8 @@ impl LlamaContext<'_> {
         let p1 = p1
             .map_or(Ok(-1), i32::try_from)
             .map_err(KvCacheConversionError::P1TooLarge)?;
-        unsafe {
-            llama_cpp_sys_2::llama_kv_self_seq_add(self.context.as_ptr(), seq_id, p0, p1, delta);
-        }
+        let mem = unsafe { llama_cpp_sys_2::llama_get_memory(self.context.as_ptr()) };
+        unsafe { llama_cpp_sys_2::llama_memory_seq_add(mem, seq_id, p0, p1, delta) };
         Ok(())
     }
 
@@ -183,7 +181,8 @@ impl LlamaContext<'_> {
             .map_or(Ok(-1), i32::try_from)
             .map_err(KvCacheConversionError::P1TooLarge)?;
         let d = c_int::from(d.get());
-        unsafe { llama_cpp_sys_2::llama_kv_self_seq_div(self.context.as_ptr(), seq_id, p0, p1, d) }
+        let mem = unsafe { llama_cpp_sys_2::llama_get_memory(self.context.as_ptr()) };
+        unsafe { llama_cpp_sys_2::llama_memory_seq_div(mem, seq_id, p0, p1, d) }
         Ok(())
     }
 
@@ -194,19 +193,7 @@ impl LlamaContext<'_> {
     /// * `seq_id` - The sequence id to get the max position for
     #[must_use]
     pub fn kv_cache_seq_pos_max(&self, seq_id: i32) -> i32 {
-        unsafe { llama_cpp_sys_2::llama_kv_self_seq_pos_max(self.context.as_ptr(), seq_id) }
-    }
-
-    /// Defragment the KV cache
-    /// This will be applied:
-    ///   - lazily on next [`LlamaContext::decode`]
-    ///   - explicitly with [`Self::kv_cache_update`]
-    pub fn kv_cache_defrag(&mut self) {
-        unsafe { llama_cpp_sys_2::llama_kv_self_defrag(self.context.as_ptr()) }
-    }
-
-    /// Apply the KV cache updates (such as K-shifts, defragmentation, etc.)
-    pub fn kv_cache_update(&mut self) {
-        unsafe { llama_cpp_sys_2::llama_kv_self_update(self.context.as_ptr()) }
+        let mem = unsafe { llama_cpp_sys_2::llama_get_memory(self.context.as_ptr()) };
+        unsafe { llama_cpp_sys_2::llama_memory_seq_pos_max(mem, seq_id) }
     }
 }

llama-cpp-2/src/context/kv_cache.rs

@@ -335,34 +335,17 @@ impl LlamaContextParams {
         self.context_params.n_ubatch
     }
 
-    /// Set the `flash_attention` parameter
-    ///
-    /// # Examples
-    ///
-    /// ```rust
-    /// use llama_cpp_2::context::params::LlamaContextParams;
-    /// let params = LlamaContextParams::default()
-    ///     .with_flash_attention(true);
-    /// assert_eq!(params.flash_attention(), true);
-    /// ```
+    /// Set the flash attention policy using llama.cpp enum
     #[must_use]
-    pub fn with_flash_attention(mut self, enabled: bool) -> Self {
-        self.context_params.flash_attn = enabled;
+    pub fn with_flash_attention_policy(mut self, policy: llama_cpp_sys_2::llama_flash_attn_type) -> Self {
+        self.context_params.flash_attn_type = policy;
         self
     }
 
-    /// Get the `flash_attention` parameter
-    ///
-    /// # Examples
-    ///
-    /// ```rust
-    /// use llama_cpp_2::context::params::LlamaContextParams;
-    /// let params = LlamaContextParams::default();
-    /// assert_eq!(params.flash_attention(), false);
-    /// ```
+    /// Get the flash attention policy
     #[must_use]
-    pub fn flash_attention(&self) -> bool {
-        self.context_params.flash_attn
+    pub fn flash_attention_policy(&self) -> llama_cpp_sys_2::llama_flash_attn_type {
+        self.context_params.flash_attn_type
     }
 
     /// Set the `offload_kqv` parameter to control offloading KV cache & KQV ops to GPU

llama-cpp-2/src/context/params.rs

@@ -180,9 +180,6 @@ impl LlamaModelParams {
 /// ```
 /// # use llama_cpp_2::model::params::LlamaModelParams;
 /// let params = LlamaModelParams::default();
-/// #[cfg(not(target_os = "macos"))]
-/// assert_eq!(params.n_gpu_layers(), 0, "n_gpu_layers should be 0");
-/// #[cfg(target_os = "macos")]
 /// assert_eq!(params.n_gpu_layers(), 999, "n_gpu_layers should be 999");
 /// assert_eq!(params.main_gpu(), 0, "main_gpu should be 0");
 /// assert_eq!(params.vocab_only(), false, "vocab_only should be false");

llama-cpp-2/src/model/params.rs

@@ -1,7 +1,7 @@
 [package]
 name = "llama-cpp-sys-2"
 description = "Low Level Bindings to llama.cpp"
-version = "0.1.122"
+version = "0.1.123"
 edition = "2021"
 license = "MIT OR Apache-2.0"
 repository = "https://github.com/utilityai/llama-cpp-rs"