perf(model): tune llama.cpp for cpu-basic

Set n_threads=2 when SPACE_ID is present to match 2 vCPU on CPU Basic.
Update docstrings to reflect CPU Basic deployment instead of ZeroGPU.

Co-authored-by: Cursor <cursoragent@cursor.com>

src/model_loader.py

@@ -1,9 +1,7 @@
 """Singleton model loader for GGUF models.
 
 Keeps one model loaded at a time. Supports dual-mode deployment:
-- HF Spaces (ZeroGPU): CPU-only inference (n_gpu_layers=0) because
-  ZeroGPU's CUDA emulation is PyTorch-specific and llama.cpp links
-  against the system CUDA runtime which is not available.
+- HF Spaces (CPU Basic): CPU-only inference (n_gpu_layers=0, n_threads=2)
 - Local: Full GPU offload via CUDA when libcudart.so.12 is present.
 """
 
@@ -98,8 +96,7 @@ class _ModelManager:
     def _detect_gpu_layers() -> int:
         """Detect whether to offload layers to GPU.
 
-        On HF Spaces the CUDA runtime is unavailable to llama.cpp
-        (ZeroGPU only exposes CUDA through PyTorch emulation).
+        On HF Spaces (CPU Basic) there is no CUDA — always CPU.
         Locally, probe for libcudart.so.12 to confirm real CUDA.
         """
         if os.environ.get("SPACE_ID"):
@@ -145,18 +142,23 @@ class _ModelManager:
             chat_handler = Qwen25VLChatHandler(clip_model_path=str(mmproj_path))
 
         n_gpu_layers = self._detect_gpu_layers()
+        n_threads = 2 if os.environ.get("SPACE_ID") else None
         logger.info(
-            "Loading model: %s (handler: %s, gpu_layers: %s)",
-            config.model_file, config.handler_type, n_gpu_layers,
+            "Loading model: %s (handler: %s, gpu_layers: %s, threads: %s)",
+            config.model_file, config.handler_type, n_gpu_layers, n_threads or "default",
         )
 
-        self._llm = Llama(
-            model_path=str(model_path),
-            chat_handler=chat_handler,
-            n_gpu_layers=n_gpu_layers,
-            n_ctx=config.n_ctx,
-            verbose=False,
-        )
+        llama_kwargs: dict = {
+            "model_path": str(model_path),
+            "chat_handler": chat_handler,
+            "n_gpu_layers": n_gpu_layers,
+            "n_ctx": config.n_ctx,
+            "verbose": False,
+        }
+        if n_threads is not None:
+            llama_kwargs["n_threads"] = n_threads
+
+        self._llm = Llama(**llama_kwargs)
         self._current_config = config
         logger.info("Model loaded successfully")
         return self._llm