Gate setup/llama subprocess paths behind provider check; allow zerogpu_transformers

app.py

@@ -798,7 +798,7 @@ def api_update_settings(payload: dict[str, Any]) -> dict[str, Any]:
         updates["PHANTOM_GRID_LLAMACPP_GPU_LAYERS"] = gpu_layers
 
     provider = updates.get("PHANTOM_GRID_LLM_PROVIDER", load_settings().llm_provider)
-    if provider not in {"minicpm_omni", "llama_cpp_server", "external_llama_cpp_server"}:
+    if provider not in {"minicpm_omni", "llama_cpp_server", "external_llama_cpp_server", "zerogpu_transformers"}:
         raise HTTPException(status_code=400, detail="Choose a supported AI backend.")
     if provider == "llama_cpp_server":
         model_path = Path(updates.get("PHANTOM_GRID_LLAMACPP_MODEL_PATH") or str(load_settings().llamacpp_model_path or "")).expanduser()
@@ -858,6 +858,17 @@ def api_llama_action(action: str, payload: dict[str, Any]) -> dict[str, Any]:
         api_update_settings(payload)
     settings = load_settings()
     normalized = action.strip().lower()
+    # ZeroGPU: no subprocess to start/stop/restart. Report status; ignore lifecycle verbs.
+    if settings.llm_provider == "zerogpu_transformers":
+        llama_status, omni_status = _service_statuses(settings)
+        event = "ZeroGPU backend runs in-process." if normalized in {"start", "restart", "stop"} else None
+        return {
+            "ok": True,
+            "event": event,
+            "llama": llama_status,
+            "omni": omni_status,
+            "settings": _settings_payload(settings),
+        }
     if settings.llm_provider == "external_llama_cpp_server" and normalized in {"start", "restart", "stop"}:
         llama_status, omni_status = _service_statuses(settings)
         return {
@@ -929,6 +940,24 @@ def api_omni_models() -> dict[str, Any]:
 
 def api_setup_status() -> dict[str, Any]:
     global _SETUP_PROCESS
+    # ZeroGPU provider: the model is loaded in-process by llm/zerogpu_backend.py
+    # at import. There is no local runtime to install, no subprocess to manage,
+    # and no cmake/llama-server dependency. Report ready so the UI's auto-poll
+    # doesn't fire /api/setup/start (which would spawn the cmake provisioner).
+    if load_settings().llm_provider == "zerogpu_transformers":
+        health = OmniClient(load_settings()).health()
+        ready = bool(health.get("ready"))
+        return {
+            "ok": True,
+            "state": "ready" if ready else "installing",
+            "stage": "ready" if ready else "service",
+            "message": "ZeroGPU model is ready." if ready else "ZeroGPU model is loading...",
+            "progress": 100 if ready else 80,
+            "files_ready": True,
+            "service_ready": ready,
+            "installing": False,
+            "updated_at": None,
+        }
     paths = _local_runtime_paths()
     scan = scan_minicpm_models(paths["models"])
     files_ready = (
@@ -983,6 +1012,10 @@ def api_setup_status() -> dict[str, Any]:
 
 def api_setup_start(payload: dict[str, Any] | None = None) -> dict[str, Any]:
     global _SETUP_PROCESS
+    # ZeroGPU provider: nothing to install or launch — the in-process backend
+    # was loaded at import. The setup-start request is a no-op.
+    if load_settings().llm_provider == "zerogpu_transformers":
+        return {**api_setup_status(), "event": "ZeroGPU runtime is in-process; no setup needed.", "ok": True}
     # Persist any picker choices before kicking off setup so the provisioner
     # and the launcher both see the chosen model/GPU/context.
     if payload:
@@ -1504,6 +1537,9 @@ def _llama_status(settings, health: dict[str, Any] | None = None) -> dict[str, A
 
 def _start_llama_process(settings) -> dict[str, Any]:
     global _LLAMA_PROCESS
+    # The ZeroGPU backend is in-process — there is no subprocess to spawn.
+    if settings.llm_provider == "zerogpu_transformers":
+        return {"ok": True, "event": "ZeroGPU backend runs in-process; no llama subprocess is needed."}
     if _LLAMA_PROCESS is not None and _LLAMA_PROCESS.poll() is None:
         return {"ok": True, "event": f"The selected AI backend is already managed as PID {_LLAMA_PROCESS.pid}."}
     if settings.llm_provider == "external_llama_cpp_server":