Add Linux Vulkan support for local inference (#9038)

Signed-off-by: jh-block <jhugo@block.xyz>

crates/goose-cli/Cargo.toml

@@ -74,6 +74,7 @@ code-mode = ["goose/code-mode"]
 local-inference = ["goose/local-inference"]
 aws-providers = ["goose/aws-providers"]
 cuda = ["goose/cuda", "local-inference"]
+vulkan = ["goose/vulkan", "local-inference"]
 telemetry = ["goose/telemetry"]
 otel = ["goose/otel"]
 # disables the update command

crates/goose-server/Cargo.toml

@@ -17,6 +17,7 @@ code-mode = ["goose/code-mode"]
 local-inference = ["goose/local-inference"]
 aws-providers = ["goose/aws-providers"]
 cuda = ["goose/cuda", "local-inference"]
+vulkan = ["goose/vulkan", "local-inference"]
 telemetry = ["goose/telemetry"]
 otel = ["goose/otel"]
 rustls-tls = [

crates/goose/Cargo.toml

@@ -37,6 +37,7 @@ aws-providers = [
     "dep:aws-sdk-sagemakerruntime",
 ]
 cuda = ["local-inference", "candle-core/cuda", "candle-nn/cuda", "llama-cpp-2/cuda"]
+vulkan = ["local-inference", "llama-cpp-2/vulkan"]
 rustls-tls = [
     "reqwest/rustls",
     "rmcp/reqwest",

crates/goose/src/providers/local_inference.rs

@@ -86,6 +86,7 @@ impl InferenceRuntime {
             }
         };
         llama_cpp_2::send_logs_to_tracing(LogOptions::default());
+        log_inference_backend_devices();
         let runtime = Arc::new(Self {
             models: StdMutex::new(HashMap::new()),
             backend,
@@ -156,20 +157,55 @@ pub fn resolve_model_path(model_id: &str) -> Option<ResolvedModelPaths> {
     None
 }
 
+fn is_accelerator_device(device_type: LlamaBackendDeviceType) -> bool {
+    matches!(
+        device_type,
+        LlamaBackendDeviceType::Gpu
+            | LlamaBackendDeviceType::IntegratedGpu
+            | LlamaBackendDeviceType::Accelerator
+    )
+}
+
+fn is_non_cpu_device(device_type: LlamaBackendDeviceType) -> bool {
+    !matches!(device_type, LlamaBackendDeviceType::Cpu)
+}
+
+fn log_inference_backend_devices() {
+    let devices = list_llama_ggml_backend_devices();
+    let non_cpu_devices: Vec<_> = devices
+        .iter()
+        .filter(|device| is_non_cpu_device(device.device_type))
+        .collect();
+
+    if non_cpu_devices.is_empty() {
+        tracing::info!(
+            device_count = devices.len(),
+            "No non-CPU llama.cpp backend devices detected for local inference"
+        );
+        return;
+    }
+
+    for device in non_cpu_devices {
+        tracing::info!(
+            index = device.index,
+            backend = %device.backend,
+            name = %device.name,
+            description = %device.description,
+            device_type = ?device.device_type,
+            memory_total_bytes = device.memory_total as u64,
+            memory_free_bytes = device.memory_free as u64,
+            "Non-CPU llama.cpp backend device detected for local inference"
+        );
+    }
+}
+
 pub fn available_inference_memory_bytes(runtime: &InferenceRuntime) -> u64 {
     let _ = &runtime.backend;
     let devices = list_llama_ggml_backend_devices();
 
     let accel_memory = devices
         .iter()
-        .filter(|d| {
-            matches!(
-                d.device_type,
-                LlamaBackendDeviceType::Gpu
-                    | LlamaBackendDeviceType::IntegratedGpu
-                    | LlamaBackendDeviceType::Accelerator
-            )
-        })
+        .filter(|d| is_accelerator_device(d.device_type))
         .map(|d| d.memory_free as u64)
         .max()
         .unwrap_or(0);