Expose hidden states, plain-text tokenization and typed GGUF reads, and drop Windows support

.github/actions/install-build-dependencies/action.yml

@@ -1,5 +1,5 @@
 name: install-build-dependencies
-description: Install OS-specific system packages needed to build llama-cpp-bindings (CMake, libclang, GNU make).
+description: Install the Linux system packages needed to build llama-cpp-bindings (CMake, libclang).
 
 runs:
   using: composite
@@ -8,15 +8,3 @@ runs:
       if: runner.os == 'Linux'
       shell: bash
       run: sudo apt-get update && sudo apt-get install -y cmake libclang-dev
-
-    - name: install windows build dependencies
-      if: runner.os == 'Windows'
-      shell: bash
-      run: choco install -y make
-
-    - name: set windows libclang path
-      if: runner.os == 'Windows'
-      shell: bash
-      env:
-        WINDOWS_LIBCLANG_PATH: 'C:\Program Files\LLVM\bin'
-      run: echo "LIBCLANG_PATH=$WINDOWS_LIBCLANG_PATH" >> "$GITHUB_ENV"

.github/workflows/unit-tests.yml

@@ -29,7 +29,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        os: [ubuntu-latest, windows-latest, macos-latest]
+        os: [ubuntu-latest, macos-latest]
     defaults:
       run:
         shell: bash

llama-cpp-bindings-build/src/android_ndk.rs

@@ -152,8 +152,6 @@ const fn detect_host_tag() -> Result<&'static str, AndroidNdkDetectionError> {
         Ok("darwin-x86_64")
     } else if cfg!(target_os = "linux") {
         Ok("linux-x86_64")
-    } else if cfg!(target_os = "windows") {
-        Ok("windows-x86_64")
     } else {
         Err(AndroidNdkDetectionError::UnsupportedHostPlatform)
     }

llama-cpp-bindings-build/src/bindgen_config.rs

@@ -154,10 +154,6 @@ pub fn generate_bindings(
         builder = configure_android_bindgen(builder, ndk, target_triple);
     }
 
-    if target_os.is_msvc() {
-        builder = configure_msvc_bindgen(builder, target_triple)?;
-    }
-
     let bindings = builder.generate().map_err(BuildError::Bindgen)?;
 
     callbacks.verify_every_privatized_field_was_found()?;
@@ -180,6 +176,7 @@ fn create_base_builder(llama_src: &Path, callbacks: BindingCallbacks) -> bindgen
         .derive_partialeq(true)
         .allowlist_function("ggml_.*")
         .allowlist_type("ggml_.*")
+        .allowlist_var("GGML_MAX_DIMS")
         .allowlist_function("gguf_.*")
         .allowlist_type("gguf_.*")
         .allowlist_function("llama_.*")
@@ -192,7 +189,6 @@ fn create_base_builder(llama_src: &Path, callbacks: BindingCallbacks) -> bindgen
         .blocklist_function("llama_model_load_from_file_ptr")
         .blocklist_type("FILE")
         .blocklist_type("_IO_.*")
-        .blocklist_type("_iobuf")
         .prepend_enum_name(false);
 
     for function in DEPRECATED_FUNCTIONS {
@@ -228,41 +224,3 @@ fn configure_android_bindgen(
 
     builder.clang_arg(format!("--target={target_triple}"))
 }
-
-fn configure_msvc_bindgen(
-    mut builder: bindgen::Builder,
-    target_triple: &str,
-) -> Result<bindgen::Builder, BuildError> {
-    let compiler = cc::Build::new()
-        .try_get_compiler()
-        .map_err(BuildError::NativeCompiler)?;
-
-    let msvc_include_paths = compiler
-        .env()
-        .iter()
-        .find(|(key, _)| key.eq_ignore_ascii_case("INCLUDE"))
-        .map(|(_, value)| value.clone());
-
-    if let Some(include_paths) = msvc_include_paths {
-        for include_path in include_paths
-            .to_string_lossy()
-            .split(';')
-            .filter(|path| !path.is_empty())
-        {
-            builder = builder.clang_arg("-isystem").clang_arg(include_path);
-            debug_log!("Added MSVC include path: {}", include_path);
-        }
-    }
-
-    builder = builder
-        .clang_arg(format!("--target={target_triple}"))
-        .clang_arg("-fms-compatibility")
-        .clang_arg("-fms-extensions");
-
-    debug_log!(
-        "Configured bindgen with MSVC toolchain for target: {}",
-        target_triple
-    );
-
-    Ok(builder)
-}

llama-cpp-bindings-build/src/cmake_config.rs

@@ -31,9 +31,6 @@ pub fn configure_and_build(context: &BuildContext) -> Result<PathBuf, BuildError
     configure_system_ggml(&mut config)?;
     let backends_dir = configure_dynamic_backends(&mut config, &context.cmake_dir)?;
 
-    config.static_crt(context.static_crt);
-    configure_msvc_exception_handling(&mut config, context.target_os);
-    configure_msvc_config_flags(&mut config, context.target_os, &context.profile);
     config
         .out_dir(&context.cmake_dir)
         .profile(&context.profile)
@@ -145,46 +142,6 @@ fn map_cpu_feature_to_ggml(feature: &str) -> Option<&'static str> {
     }
 }
 
-const fn msvc_exception_handling_flag(target_os: TargetOs) -> Option<&'static str> {
-    if target_os.is_msvc() {
-        Some("/EHsc")
-    } else {
-        None
-    }
-}
-
-fn configure_msvc_exception_handling(config: &mut Config, target_os: TargetOs) {
-    let Some(flag) = msvc_exception_handling_flag(target_os) else {
-        return;
-    };
-
-    config.cxxflag(flag);
-}
-
-fn msvc_config_flags(target_os: TargetOs, profile: &str) -> Option<&'static str> {
-    if !target_os.is_msvc() {
-        return None;
-    }
-
-    match profile {
-        "Debug" => Some("/Ob0 /Od /RTC1"),
-        "MinSizeRel" => Some("/O1 /Ob1 /DNDEBUG"),
-        "Release" => Some("/O2 /Ob2 /DNDEBUG"),
-        "RelWithDebInfo" => Some("/O2 /Ob1 /DNDEBUG"),
-        _ => None,
-    }
-}
-
-fn configure_msvc_config_flags(config: &mut Config, target_os: TargetOs, profile: &str) {
-    let Some(flags) = msvc_config_flags(target_os, profile) else {
-        return;
-    };
-    let config_suffix = profile.to_uppercase();
-
-    config.define(format!("CMAKE_C_FLAGS_{config_suffix}"), flags);
-    config.define(format!("CMAKE_CXX_FLAGS_{config_suffix}"), flags);
-}
-
 fn configure_shared_libs(config: &mut Config, build_shared_libs: bool) {
     config.define(
         "BUILD_SHARED_LIBS",
@@ -207,7 +164,7 @@ fn configure_platform_specific(
                 configure_android_cmake(config, ndk, target_triple);
             }
         }
-        _ => {}
+        TargetOs::Linux => {}
     }
 }
 
@@ -261,16 +218,6 @@ fn configure_gpu_backends(config: &mut Config, target_os: TargetOs) -> Result<()
 
 fn configure_vulkan_linking(target_os: TargetOs) -> Result<(), BuildError> {
     match target_os {
-        TargetOs::Windows(_) => {
-            let vulkan_path = env::var("VULKAN_SDK").map_err(|source| BuildError::Environment {
-                name: "VULKAN_SDK",
-                source,
-            })?;
-            let vulkan_lib_path = Path::new(&vulkan_path).join("Lib");
-
-            println!("cargo:rustc-link-search={}", vulkan_lib_path.display());
-            println!("cargo:rustc-link-lib=vulkan-1");
-        }
         TargetOs::Linux => {
             match env::var("VULKAN_SDK") {
                 Ok(vulkan_path) => {
@@ -289,7 +236,7 @@ fn configure_vulkan_linking(target_os: TargetOs) -> Result<(), BuildError> {
 
             println!("cargo:rustc-link-lib=vulkan");
         }
-        _ => (),
+        TargetOs::Apple(_) | TargetOs::Android => {}
     }
 
     Ok(())
@@ -315,67 +262,6 @@ fn configure_system_ggml(config: &mut Config) -> Result<(), BuildError> {
     Ok(())
 }
 
-#[cfg(test)]
-mod msvc_exception_handling_tests {
-    use crate::target_os::TargetOs;
-    use crate::windows_variant::WindowsVariant;
-
-    use super::msvc_exception_handling_flag;
-
-    #[test]
-    fn msvc_targets_compile_llama_cpp_with_unwind_semantics() {
-        assert_eq!(
-            msvc_exception_handling_flag(TargetOs::Windows(WindowsVariant::Msvc)),
-            Some("/EHsc")
-        );
-    }
-
-    #[test]
-    fn targets_without_msvc_keep_their_toolchain_default_exception_handling() {
-        assert_eq!(msvc_exception_handling_flag(TargetOs::Linux), None);
-        assert_eq!(
-            msvc_exception_handling_flag(TargetOs::Windows(WindowsVariant::Other)),
-            None
-        );
-    }
-}
-
-#[cfg(test)]
-mod msvc_config_flag_tests {
-    use crate::target_os::TargetOs;
-    use crate::windows_variant::WindowsVariant;
-
-    use super::msvc_config_flags;
-
-    #[test]
-    fn every_msvc_configuration_keeps_llama_cpp_assertions_compiled_out() {
-        let msvc = TargetOs::Windows(WindowsVariant::Msvc);
-
-        assert_eq!(msvc_config_flags(msvc, "Debug"), Some("/Ob0 /Od /RTC1"));
-        assert_eq!(
-            msvc_config_flags(msvc, "MinSizeRel"),
-            Some("/O1 /Ob1 /DNDEBUG")
-        );
-        assert_eq!(
-            msvc_config_flags(msvc, "Release"),
-            Some("/O2 /Ob2 /DNDEBUG")
-        );
-        assert_eq!(
-            msvc_config_flags(msvc, "RelWithDebInfo"),
-            Some("/O2 /Ob1 /DNDEBUG")
-        );
-    }
-
-    #[test]
-    fn targets_and_profiles_without_msvc_defaults_keep_the_flags_cmake_chose() {
-        assert_eq!(msvc_config_flags(TargetOs::Linux, "Release"), None);
-        assert_eq!(
-            msvc_config_flags(TargetOs::Windows(WindowsVariant::Msvc), "Fastest"),
-            None
-        );
-    }
-}
-
 #[cfg(test)]
 mod cpu_feature_mapping_tests {
     use super::map_cpu_feature_to_ggml;

llama-cpp-bindings-build/src/cpp_wrapper.rs

@@ -20,11 +20,6 @@ pub fn compile_cpp_wrappers(target_os: TargetOs) -> Result<(), BuildError> {
         build.file(source);
     }
 
-    if target_os.is_msvc() {
-        build.flag(format!("/std:{CPP_STANDARD}"));
-        build.flag("/EHsc");
-    }
-
     if target_os.is_android() && cfg!(feature = "static-stdcxx") {
         build.cpp_link_stdlib(None);
     }

llama-cpp-bindings-build/src/lib.rs

@@ -11,7 +11,6 @@ mod library_linking;
 mod native_library;
 mod rebuild_tracking;
 mod target_os;
-mod windows_variant;
 
 use std::env;
 use std::path::{Path, PathBuf};
@@ -42,8 +41,6 @@ pub enum BuildError {
     },
     #[error("generated bindings could not be written: {0}")]
     BindingsWrite(#[source] std::io::Error),
-    #[error("native compiler setup failed: {0}")]
-    NativeCompiler(#[source] cc::Error),
     #[error("native wrapper compilation failed: {0}")]
     NativeWrapper(#[source] cc::Error),
     #[error("filesystem operation failed for {path}: {source}")]
@@ -88,7 +85,6 @@ pub struct BuildContext {
     pub cargo_cfg_target_env: String,
     pub build_shared_libs: bool,
     pub profile: String,
-    pub static_crt: bool,
     pub android_ndk: Option<AndroidNdk>,
 }
 
@@ -97,20 +93,17 @@ impl BuildContext {
         let target_triple = required_env("TARGET")?;
         let cargo_cfg_target_os = required_env("CARGO_CFG_TARGET_OS")?;
         let cargo_cfg_target_env = optional_env("CARGO_CFG_TARGET_ENV")?.unwrap_or_default();
-        let target_os = TargetOs::from_cargo_cfg(&cargo_cfg_target_os, &cargo_cfg_target_env)
-            .ok_or_else(|| BuildError::UnsupportedTargetOs {
+        let target_os = TargetOs::from_cargo_cfg(&cargo_cfg_target_os).ok_or_else(|| {
+            BuildError::UnsupportedTargetOs {
                 cargo_cfg_target_os: cargo_cfg_target_os.clone(),
-            })?;
+            }
+        })?;
         let out_dir = PathBuf::from(required_env("OUT_DIR")?);
         let manifest_dir = required_env("CARGO_MANIFEST_DIR")?;
         let llama_src = Path::new(&manifest_dir).join("llama.cpp");
 
         let build_shared_libs = cfg!(feature = "dynamic-link");
         let profile = native_profile(&required_env("PROFILE")?);
-        let static_crt = optional_env("CARGO_CFG_TARGET_FEATURE")?
-            .unwrap_or_default()
-            .split(',')
-            .any(|feature| feature == "crt-static");
 
         let cargo_cfg_target_arch = required_env("CARGO_CFG_TARGET_ARCH")?;
         let android_ndk = if target_os.is_android() {
@@ -137,7 +130,6 @@ impl BuildContext {
             cargo_cfg_target_env,
             build_shared_libs,
             profile,
-            static_crt,
             android_ndk,
         })
     }