feat: update llama.cpp to b8665

.cargo/config.toml

@@ -1,6 +1,6 @@
 [env]
 # refer: https://stackoverflow.com/questions/43577885/is-there-a-cargo-environment-variable-for-the-workspace-directory
 CARGO_WORKSPACE_DIR = { value = "", relative = true }
-LLAMA_CPP_TAG = "b8233"
+LLAMA_CPP_TAG = "b8665"
 # CUDA 13.1 requires C++17
 NVCC_PREPEND_FLAGS = "-std=c++17"

koharu-llm/build.rs

@@ -155,9 +155,25 @@ fn ensure_source_tree(out_dir: &Path, llama_cpp_tag: &str) -> Result<PathBuf> {
         fs::File::open(&tarball_path).context("failed to reopen downloaded source tarball")?;
     let decoder = GzDecoder::new(tarball);
     let mut archive = Archive::new(decoder);
-    archive
-        .unpack(&cache_dir)
-        .context("failed to extract llama.cpp source tarball")?;
+    // Extract selectively: skip benchmarks and other non-essential paths that
+    // contain very long filenames which exceed Windows' 260-char path limit.
+    for entry in archive
+        .entries()
+        .context("failed to read tarball entries")?
+    {
+        let mut entry = entry.context("failed to read tarball entry")?;
+        let path_str = entry
+            .path()
+            .context("failed to read entry path")?
+            .to_string_lossy()
+            .into_owned();
+        if path_str.contains("/benches/") || path_str.contains("/bench/") {
+            continue;
+        }
+        entry
+            .unpack_in(&cache_dir)
+            .with_context(|| format!("failed to unpack `{path_str}`"))?;
+    }
 
     if !source_root.join("include/llama.h").exists() {
         bail!(
@@ -230,7 +246,9 @@ fn generate_loader(
     spec: LoaderSpec<'_>,
 ) -> Result<()> {
     let builder = spec.function_patterns.iter().fold(
-        base_builder(header_path, include_dirs),
+        base_builder(header_path, include_dirs)
+            // Block functions using C FILE* which has no Rust equivalent in bindgen.
+            .blocklist_function("llama_model_load_from_file_ptr"),
         |builder, pattern| builder.allowlist_function(pattern),
     );
 

koharu-llm/src/safe/mtmd.rs

@@ -217,8 +217,8 @@ impl MtmdContext {
     /// Get audio bitrate in Hz (e.g., 16000 for Whisper).
     /// Returns None if audio is not supported.
     #[must_use]
-    pub fn get_audio_bitrate(&self) -> Option<u32> {
-        let rate = unsafe { crate::sys::mtmd_get_audio_bitrate(self.context.as_ptr()) };
+    pub fn get_audio_sample_rate(&self) -> Option<u32> {
+        let rate = unsafe { crate::sys::mtmd_get_audio_sample_rate(self.context.as_ptr()) };
         (rate > 0).then_some(rate.unsigned_abs())
     }
 

koharu-runtime/src/llama.rs

@@ -56,15 +56,16 @@ impl LlamaDistribution {
         }
     }
 
-    fn assets(self) -> &'static [&'static str] {
+    fn assets(self) -> Vec<String> {
+        let tag = LLAMA_CPP_TAG;
         match self {
-            Self::WindowsCuda13X64 => &[
-                "llama-b8233-bin-win-cuda-13.1-x64.zip",
-                "cudart-llama-bin-win-cuda-13.1-x64.zip",
+            Self::WindowsCuda13X64 => vec![
+                format!("llama-{tag}-bin-win-cuda-13.1-x64.zip"),
+                "cudart-llama-bin-win-cuda-13.1-x64.zip".to_string(),
             ],
-            Self::WindowsVulkanX64 => &["llama-b8233-bin-win-vulkan-x64.zip"],
-            Self::LinuxVulkanX64 => &["llama-b8233-bin-ubuntu-vulkan-x64.tar.gz"],
-            Self::MacosArm64 => &["llama-b8233-bin-macos-arm64.tar.gz"],
+            Self::WindowsVulkanX64 => vec![format!("llama-{tag}-bin-win-vulkan-x64.zip")],
+            Self::LinuxVulkanX64 => vec![format!("llama-{tag}-bin-ubuntu-vulkan-x64.tar.gz")],
+            Self::MacosArm64 => vec![format!("llama-{tag}-bin-macos-arm64.tar.gz")],
         }
     }
 
@@ -200,7 +201,7 @@ pub(crate) async fn ensure_ready(runtime: &Runtime) -> Result<()> {
     if !install.is_current() {
         install.reset()?;
 
-        for asset in distribution.assets() {
+        for asset in &distribution.assets() {
             let url = format!("{RELEASE_BASE_URL}/{LLAMA_CPP_TAG}/{asset}");
             let archive = runtime
                 .downloads()