feat(local): Qwen3.5 V2 hybrid runtime support

Cargo.toml

@@ -82,10 +82,10 @@ urlencoding = "2.1"
 version-compare = "0.2.1"
 
 [target.'cfg(target_os = "macos")'.dependencies]
-tvm-runtime = { git = "https://github.com/brekkylab/tvm-runtime-rs", rev = "ee8c5626e5309827b00af9e46111fe23b0307308", features = ["metal"] }
+tvm-runtime = { git = "https://github.com/brekkylab/tvm-runtime-rs", rev = "cea927a", features = ["metal"] }
 
 [target.'cfg(any(target_os = "linux", target_os = "windows"))'.dependencies]
-tvm-runtime = { git = "https://github.com/brekkylab/tvm-runtime-rs", rev = "ee8c5626e5309827b00af9e46111fe23b0307308", features = ["vulkan"] }
+tvm-runtime = { git = "https://github.com/brekkylab/tvm-runtime-rs", rev = "cea927a", features = ["vulkan"] }
 
 [target.'cfg(not(target_arch = "wasm32"))'.dependencies]
 ailoy-faiss-sys = { path = "./crates/faiss-sys" }
@@ -96,7 +96,7 @@ parking_lot = "0.12.4"
 rmcp = { version = "0.11.0", features = ["client", "reqwest", "transport-child-process", "transport-streamable-http-client", "transport-streamable-http-client-reqwest"] }
 tokenizers = { version = "0.22.2", default-features = false, features = ["onig"] }
 tokio = { version = "1.0", default-features = false, features = ["macros", "rt-multi-thread", "sync"] }
-tvm-ffi = { git = "https://github.com/brekkylab/tvm-runtime-rs", rev = "ee8c5626e5309827b00af9e46111fe23b0307308" }
+tvm-ffi = { git = "https://github.com/brekkylab/tvm-runtime-rs", rev = "cea927a" }
 uuid = { version = "1.18.0", features = ["v4"] }
 
 [target.'cfg(target_arch = "wasm32")'.dependencies]

build.rs

@@ -29,6 +29,26 @@ fn build_native() {
     if std::env::var_os("CARGO_FEATURE_NODEJS").is_some() {
         napi_build::setup();
     }
+
+    // Link libmlc_llm_module so its TVM_FFI_STATIC_INIT_BLOCKs run at process
+    // load. We deliberately link the *_module variant rather than libmlc_llm:
+    // mlc-llm's Python package loads libmlc_llm_module via tvm.ffi.load_module,
+    // so linking the same dylib keeps a single GlobalFunctionTable in process
+    // when both ailoy and mlc-llm are imported together. Linking libmlc_llm
+    // directly would put two copies of every mlc.json_ffi.* function in the
+    // global registry and the second registration aborts at process load.
+    println!("cargo:rerun-if-env-changed=MLC_LLM_LIB_DIR");
+    if let Ok(mlc_dir) = std::env::var("MLC_LLM_LIB_DIR") {
+        println!("cargo:rustc-link-search=native={}", mlc_dir);
+        println!("cargo:rustc-link-lib=dylib=mlc_llm_module");
+    }
+
+    // Make the resulting cdylib portable: tell the linker to look for runtime
+    // dependencies right next to itself (`@loader_path`). Combined with the
+    // build/install step that copies libmlc_llm_module / libtvm{,_runtime,_ffi
+    // [_testing]} into bindings/python/ailoy/, this gives ailoy a self-
+    // contained dylib closure with no hard-coded venv paths.
+    println!("cargo:rustc-link-arg=-Wl,-rpath,@loader_path");
 }
 
 fn build_wasm() {

src/model/local/chat_template.rs

@@ -61,8 +61,25 @@ impl ChatTemplate {
             ThinkEffort::High => "high",
             _ => "",
         };
+        // HuggingFace-style chat templates address message body as `message.content`
+        // (singular), but our Message struct serializes the field as `contents`.
+        // Re-serialize each message into a serde_json::Value so we can also expose
+        // `content` as an alias, otherwise jinja sees `content == undefined` and
+        // emits empty user/system bodies which collapses the model.
+        let messages_for_template: Vec<serde_json::Value> = messages
+            .iter()
+            .map(|m| {
+                let mut v = serde_json::to_value(m).unwrap_or(serde_json::Value::Null);
+                if let Some(obj) = v.as_object_mut() {
+                    if let Some(c) = obj.get("contents").cloned() {
+                        obj.insert("content".to_string(), c);
+                    }
+                }
+                v
+            })
+            .collect();
         let ctx = context!(
-            messages => messages,
+            messages => messages_for_template,
             tools => if !tools.is_empty() { Some(tools) } else { None::<_> },
             documents => if !documents.is_empty() { Some(documents) } else { None::<_> },
             add_generation_prompt => add_generation_prompt,