fix(engine): offload Gemma to GPU by setting n_gpu_layers=999 (#833)

LlamaModelParams::default() sets n_gpu_layers=0, so even with --features
rocm the model ran entirely on CPU at ~19 t/s. Setting n_gpu_layers to a
large sentinel value asks llama.cpp to offload every layer the model
has; llama.cpp clamps to the real count (27 for Gemma 2 2B). Observed
throughput jumps from 19 t/s to 74 t/s on an RX 9070 once the ROCm
binary is also compiled for gfx1201 (see tooling commit).

Also adds server/sr-voice/.gitignore so locally-built binaries don't
sneak into the worktree. Release binaries ship out-of-tree per #850.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

server/sr-voice/.gitignore

@@ -0,0 +1,4 @@
+# Built binaries — platform-specific, rebuilt via distrobox + cargo.
+# Per #850, release binaries will ship as CI artifacts, not in the repo.
+/bin/
+/target/

server/sr-voice/src/inference.rs

@@ -39,11 +39,19 @@ pub struct InferenceEngine {
 
 impl InferenceEngine {
     /// Load a GGUF model from disk.
+    ///
+    /// Offloads all layers to the GPU via ROCm. The binary is built with
+    /// llama-cpp-rs + ROCm support (see Makefile `build-sr-voice` target),
+    /// but `LlamaModelParams::default()` sets `n_gpu_layers = 0`, which
+    /// runs the entire model on CPU at ~10× lower throughput. Setting
+    /// `n_gpu_layers` to a large sentinel value (999) asks llama.cpp to
+    /// offload every layer the model has; it clamps to the real count.
+    /// For Gemma 2 2B (27 layers) this fully GPU-offloads the model.
     pub fn load(config: &InferenceConfig) -> Result<Self, VoiceError> {
         let backend =
             LlamaBackend::init().map_err(|e| VoiceError::ModelLoadFailed(e.to_string()))?;
 
-        let model_params = LlamaModelParams::default();
+        let model_params = LlamaModelParams::default().with_n_gpu_layers(999);
         let model = LlamaModel::load_from_file(
             &backend,
             Path::new(&config.model_path),