Upgrade vendored llama.cpp to b10068 and bump to 0.11.0

Cargo.toml

@@ -15,7 +15,7 @@ members = [
 
 [workspace.package]
 edition = "2024"
-version = "0.10.0"
+version = "0.11.0"
 license = "Apache-2.0"
 repository = "https://github.com/intentee/llama-cpp-bindings"
 
@@ -31,15 +31,15 @@ glob = "=0.3.3"
 hf-hub = "=0.5.0"
 inventory = "=0.3.24"
 libtest-mimic = "=0.8.2"
-llama-cpp-bindings = { path = "llama-cpp-bindings", version = "=0.10.0" }
-llama-cpp-bindings-build = { path = "llama-cpp-bindings-build", version = "=0.10.0" }
-llama-cpp-bindings-sys = { path = "llama-cpp-bindings-sys", version = "=0.10.0" }
-llama-cpp-bindings-types = { path = "llama-cpp-bindings-types", version = "=0.10.0" }
-llama-cpp-error-recorder = { path = "llama-cpp-error-recorder", version = "=0.10.0" }
-llama-cpp-gbnf = { path = "llama-cpp-gbnf", version = "=0.10.0" }
-llama-cpp-log-decoder = { path = "llama-cpp-log-decoder", version = "=0.10.0" }
-llama-cpp-test-harness = { path = "llama-cpp-test-harness", version = "=0.10.0" }
-llama-cpp-test-harness-macros = { path = "llama-cpp-test-harness-macros", version = "=0.10.0" }
+llama-cpp-bindings = { path = "llama-cpp-bindings", version = "=0.11.0" }
+llama-cpp-bindings-build = { path = "llama-cpp-bindings-build", version = "=0.11.0" }
+llama-cpp-bindings-sys = { path = "llama-cpp-bindings-sys", version = "=0.11.0" }
+llama-cpp-bindings-types = { path = "llama-cpp-bindings-types", version = "=0.11.0" }
+llama-cpp-error-recorder = { path = "llama-cpp-error-recorder", version = "=0.11.0" }
+llama-cpp-gbnf = { path = "llama-cpp-gbnf", version = "=0.11.0" }
+llama-cpp-log-decoder = { path = "llama-cpp-log-decoder", version = "=0.11.0" }
+llama-cpp-test-harness = { path = "llama-cpp-test-harness", version = "=0.11.0" }
+llama-cpp-test-harness-macros = { path = "llama-cpp-test-harness-macros", version = "=0.11.0" }
 llguidance = "=1.7.0"
 log = "=0.4.29"
 nom = "=8.0.0"

llama-cpp-bindings-sys/Cargo.toml

@@ -37,10 +37,6 @@ include = [
     "/llama.cpp/convert_hf_to_gguf.py", # Yes, it's required
     "/llama.cpp/common/build-info.cpp.in",
 
-    "/llama.cpp/ggml/src/ggml-cuda.cu",
-    "/llama.cpp/ggml/src/ggml-metal.m",
-    "/llama.cpp/ggml/src/ggml-metal.metal",
-
     "/llama.cpp/include/llama.h",
     "/llama.cpp/include/llama-cpp.h",
 
@@ -50,9 +46,6 @@ include = [
     "/llama.cpp/ggml/src/ggml-vulkan/**/*",
     "/llama.cpp/ggml/src/ggml-hip/**/*",
 
-    "/llama.cpp/ggml/src/llamafile/sgemm.h",
-    "/llama.cpp/ggml/src/llamafile/sgemm.cpp",
-
     "/llama.cpp/pocs",
     "/llama.cpp/vendor",
 

llama-cpp-bindings-tests/tests/multimodal_image_and_audio.rs

@@ -149,10 +149,9 @@ fn image_and_audio_together(fixture: &LlamaFixture<'_>) -> Result<()> {
         "model should generate a description from combined image and audio input"
     );
     assert!(
-        description.contains("sheep"),
-        "the gemma-4 vision encoder recognizes the image animals as \"sheep\" (a borderline \
-         llama/sheep call the b9585 clip-encoder update tipped); the assertion tracks the \
-         model's actual recognition so it still proves the image reached the output; \
+        description.contains("alpaca"),
+        "the gemma-4 vision encoder recognizes the image animals as \"alpaca\"; the assertion \
+         tracks the model's actual recognition so it still proves the image reached the output; \
          got: {description:?}"
     );
     assert!(

llama-cpp-bindings/src/context/kv_cache_type.rs

@@ -43,6 +43,9 @@ pub enum KvCacheType {
     TQ1_0,
     TQ2_0,
     MXFP4,
+    NVFP4,
+    Q1_0,
+    Q2_0,
 }
 
 impl From<KvCacheType> for llama_cpp_bindings_sys::ggml_type {
@@ -81,6 +84,9 @@ impl From<KvCacheType> for llama_cpp_bindings_sys::ggml_type {
             KvCacheType::TQ1_0 => llama_cpp_bindings_sys::GGML_TYPE_TQ1_0,
             KvCacheType::TQ2_0 => llama_cpp_bindings_sys::GGML_TYPE_TQ2_0,
             KvCacheType::MXFP4 => llama_cpp_bindings_sys::GGML_TYPE_MXFP4,
+            KvCacheType::NVFP4 => llama_cpp_bindings_sys::GGML_TYPE_NVFP4,
+            KvCacheType::Q1_0 => llama_cpp_bindings_sys::GGML_TYPE_Q1_0,
+            KvCacheType::Q2_0 => llama_cpp_bindings_sys::GGML_TYPE_Q2_0,
         }
     }
 }
@@ -120,15 +126,44 @@ impl From<llama_cpp_bindings_sys::ggml_type> for KvCacheType {
             x if x == llama_cpp_bindings_sys::GGML_TYPE_TQ1_0 => Self::TQ1_0,
             x if x == llama_cpp_bindings_sys::GGML_TYPE_TQ2_0 => Self::TQ2_0,
             x if x == llama_cpp_bindings_sys::GGML_TYPE_MXFP4 => Self::MXFP4,
+            x if x == llama_cpp_bindings_sys::GGML_TYPE_NVFP4 => Self::NVFP4,
+            x if x == llama_cpp_bindings_sys::GGML_TYPE_Q1_0 => Self::Q1_0,
+            x if x == llama_cpp_bindings_sys::GGML_TYPE_Q2_0 => Self::Q2_0,
             _ => Self::Unknown(value),
         }
     }
 }
 
 #[cfg(test)]
 mod tests {
+    use std::ffi::CStr;
+
     use super::KvCacheType;
 
+    fn is_retired_upstream_slot(raw: llama_cpp_bindings_sys::ggml_type) -> bool {
+        unsafe { llama_cpp_bindings_sys::ggml_blck_size(raw) == 0 }
+    }
+
+    #[test]
+    fn every_upstream_ggml_type_maps_to_a_known_variant() {
+        for raw in 0..llama_cpp_bindings_sys::GGML_TYPE_COUNT {
+            if is_retired_upstream_slot(raw) {
+                continue;
+            }
+
+            let upstream_name =
+                unsafe { CStr::from_ptr(llama_cpp_bindings_sys::ggml_type_name(raw)) }
+                    .to_string_lossy();
+
+            assert_ne!(
+                KvCacheType::from(raw),
+                KvCacheType::Unknown(raw),
+                "ggml type {raw} (\"{upstream_name}\") has no matching KvCacheType variant, so it \
+                 silently degrades to Unknown; add the variant to keep the mirror complete"
+            );
+        }
+    }
+
     #[test]
     fn kv_cache_type_unknown_preserves_raw_value() {
         let unknown_raw: llama_cpp_bindings_sys::ggml_type = 99999;
@@ -176,6 +211,9 @@ mod tests {
             KvCacheType::TQ1_0,
             KvCacheType::TQ2_0,
             KvCacheType::MXFP4,
+            KvCacheType::NVFP4,
+            KvCacheType::Q1_0,
+            KvCacheType::Q2_0,
         ];
 
         for variant in all_variants {

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

@@ -397,7 +397,18 @@ impl Default for LlamaModelParams {
 mod tests {
     use crate::model::split_mode::LlamaSplitMode;
 
-    use super::LlamaModelParams;
+    use super::{LLAMA_CPP_MAX_DEVICES, LlamaModelParams};
+
+    #[test]
+    fn device_capacity_is_not_smaller_than_the_upstream_limit() {
+        let upstream_limit = crate::max_devices();
+        assert!(
+            LLAMA_CPP_MAX_DEVICES >= upstream_limit,
+            "LLAMA_CPP_MAX_DEVICES ({LLAMA_CPP_MAX_DEVICES}) backs a fixed-size device array and is \
+             clamped against llama_max_devices() ({upstream_limit}); once upstream reports more \
+             devices than the array can hold, the extra devices are dropped without any error"
+        );
+    }
 
     #[test]
     fn default_params_have_expected_values() {

llama-cpp-bindings/src/mtmd/mtmd_context.rs

@@ -178,6 +178,7 @@ impl MtmdContext {
         let text_cstring = CString::new(text.text)?;
         let input_text = llama_cpp_bindings_sys::mtmd_input_text {
             text: text_cstring.as_ptr(),
+            text_len: text_cstring.as_bytes().len(),
             add_special: text.add_special,
             parse_special: text.parse_special,
         };